doi
stringlengths 28
28
| title
stringlengths 19
311
| abstract
stringlengths 217
5.08k
| plain language summary
stringlengths 115
4.83k
| article
stringlengths 3.87k
161k
|
|---|---|---|---|---|
10.1371/journal.pntd.0002370 | Non-peptidic Cruzain Inhibitors with Trypanocidal Activity Discovered by Virtual Screening and In Vitro Assay | A multi-step cascade strategy using integrated ligand- and target-based virtual screening methods was developed to select a small number of compounds from the ZINC database to be evaluated for trypanocidal activity. Winnowing the database to 23 selected compounds, 12 non-covalent binding cruzain inhibitors with affinity values (Ki) in the low micromolar range (3–60 µM) acting through a competitive inhibition mechanism were identified. This mechanism has been confirmed by determining the binding mode of the cruzain inhibitor Nequimed176 through X-ray crystallographic studies. Cruzain, a validated therapeutic target for new chemotherapy for Chagas disease, also shares high similarity with the mammalian homolog cathepsin L. Because increased activity of cathepsin L is related to invasive properties and has been linked to metastatic cancer cells, cruzain inhibitors from the same library were assayed against it. Affinity values were in a similar range (4–80 µM), yielding poor selectivity towards cruzain but raising the possibility of investigating such inhibitors for their effect on cell proliferation. In order to select the most promising enzyme inhibitors retaining trypanocidal activity for structure-activity relationship (SAR) studies, the most potent cruzain inhibitors were assayed against T. cruzi-infected cells. Two compounds were found to have trypanocidal activity. Using compound Nequimed42 as precursor, an SAR was established in which the 2-acetamidothiophene-3-carboxamide group was identified as essential for enzyme and parasite inhibition activities. The IC50 value for compound Nequimed42 acting against the trypomastigote form of the Tulahuen lacZ strain was found to be 10.6±0.1 µM, tenfold lower than that obtained for benznidazole, which was taken as positive control. In addition, by employing the strategy of molecular simplification, a smaller compound derived from Nequimed42 with a ligand efficiency (LE) of 0.33 kcal mol−1 atom−1 (compound Nequimed176) is highlighted as a novel non-peptidic, non-covalent cruzain inhibitor as a trypanocidal agent candidate for optimization.
| Chagas disease (American trypanosomiasis) is a parasitic infection that kills millions of mostly poverty-stricken people in Latin America. In recent years it has also spread to nonendemic countries – the United States, Canada, Europe, Australia and Japan – as a result of immigration. The only available drugs for its treatment were introduced more than forty years ago, have low efficacy, and cause various severe side effects. This dire public health situation has prompted us to search for new small molecules to act as drug candidates to treat Chagas disease. The T. cruzi enzyme cruzain, a key biological catalyst used by the protozoan to digest host proteins, is a validated drug target for Chagas disease. By combining in silico molecular design, X-ray crystallography and biological screening, we found a new class of non-covalent small molecules that inhibit cruzain in low micromolar concentrations.
| Chagas disease, widespread in Latin America, is caused by the kinetoplastid protozoan parasite Trypanosoma cruzi. Despite efforts to reduce the transmission of the parasite by controlling the hematophagous triatomine insect vector, the World Health Organization estimates that 10 million people are infected worldwide, with another 25 million at risk. Most cases are in Latin America, where Chagas disease is endemic, but it is also found in Canada, the United States, Europe (mainly in Spain and Portugal), Japan and Australia [1]–[3].
T. cruzi's complex life cycle involves two replicative forms: the epimastigote, in the gut of the insect vector, and the amastigote, an intracellular form in the infected mammal. The two infective non-replicative forms are the metacyclic trypomastigote in the insect vector and the bloodstream trypomastigote released from infected cells into the blood of the mammal [4].
Chagas disease has an acute phase and a chronic latent phase. The acute phase, which occurs shortly after infection, lasts for a few weeks or months, whereas the chronic phase develops over many years. The acute phase may not be noticed because it is symptom-free or exhibits only mild symptoms that are not unique to Chagas disease. These can include fever, fatigue, body aches, headache, and rash, loss of appetite, diarrhea and vomiting. Chronic phase symptoms appear between 10 and 20 years after infection and affect the heart, nervous and digestive systems.
The treatment of Chagas disease involves the front-line drugs nifurtimox and benznidazole. These two old drugs are effective at curing the infection mostly in the acute phase, with successful cure up to 80%, but are almost ineffective in chronically infected patients [5]. Moreover, due to its collateral effects nifurtimox is no longer available in most Latin American countries. In addition to the severe side effects of the available chemotherapy, drug-resistance has been observed in some trypanosome strains. Thus, the discovery of new, safer and more effective drugs to treat Chagas disease is of utmost importance.
Cruzain (a recombinant form of cruzipain, EC 3.4.22.51) has excellent pre-clinical validation evidence as a druggable target. Cruzain belongs to the family of cysteine proteases (papain-like enzymes known as clan CA) and is closely related to cathepsins L and S, which are also associated with other pathologies in humans [6]. It is the major cysteine protease in T. cruzi and is essential for the development and survival of the parasite within the host cells. Numerous protease inhibitors with different scaffolds and catalytic mechanisms show activity against the parasite in culture and animal models of the disease [7]. Clan CA cysteine proteases are effectively inhibited by several classes of peptide inhibitors including transition state-based, reversible and irreversible inhibitors [8]. Examples of reversible transition state-based inhibitors are peptide aldehydes, α-diketones, α-ketoesters, α-ketoamides and α-ketoacids [9]. Clan CA proteases are also irreversibly inhibited by peptidyldiazomethyl ketones, fluoromethyl ketones, peptide epoxides (E-64, E-64-c, E-64-d) and vinyl sulfones [10]. Recently, non-covalent inhibitors have been discovered through high-throughput screening (HTS) platforms and, despite their lower potency relative to previously reported covalent compounds, they represent important breakthroughs in the development of non-peptidic compounds with drug-like features [11], [12].
A promising molecular class acting with antiparasitic activity can be found in vinyl sulfones. In pre-clinical trials, the inhibitor K11777 (Figure 1A) has been shown to be non-mutagenic, well tolerated, to have an acceptable pharmacokinetic profile and demonstrated efficacy in models of acute and chronic Chagas disease both in mice and dogs [13]. Additional studies of vinyl sulfone compounds have led to the identification of an arginine variant of K11777, named WRR-483 (Figure 1B) with remarkable biological properties [14].
The aim of this study was to identify new molecular classes of cruzain inhibitors by focusing on non-peptidic non-covalent ligands. To this end, we have carried out virtual screening of the ZINC Database [15], a free database of commercially-available compounds for virtual screening, utilizing ligand- and target-based virtual screening methods [16], [17], followed by enzymatic assays, X-ray crystallography and SAR studies of the most promising hits. Of nine cruzain inhibitors, five show trypanocidal activity against the trypomastigote infective form of the Tulahuen lacZ strain. We also expect that a newly identified fragment of the 2-acetamidothiophene-3-carboxamide class can advance the search for new non-covalent cruzain inhibitors.
A variety of methods are available to virtually screen small organic compound databases. A multi-step cascade strategy using integrated ligand- and target-based virtual screening methods was applied as illustrated in Figure 2.
The Ethics Committee on Animal Experimentation of the Faculty of Medicine of Ribeirão Preto – University of Sao Paulo approved the cytotoxicity assays (approval no. 076/2010). This Committee adheres to Conselho Nacional de Controle de Experimentação Animal – CONCEA, created by Brazilian Law number 11794 of 8 October 2008. Assays were run according to the guidelines of the Ministry of Science, Technology and Innovation of Brazil.
The compound collection enrichment process began by the filtering ca. 8.5 million structures from the ZINC Database, resulting in a sub-library containing ca. 3.5 million structures. Since the sub-library still encompassed a large number of structures, two fast virtual screening methods based on ligand and receptor were employed in order to enrich it. The agreement between ligand- and target-based virtual screening methods was then used as the criterion for the selection of an enriched, focused sub-library. Based on the predicted Tanimoto similarity metric and docking score of known inhibitors, a set of thresholds was established for the selection of untested compounds. The values adopted as thresholds and the numbers of compounds resulting from this analysis are summarized in Table 2.
The analysis of cruzain-inhibitor complexes available in the PDB database reveals the importance of residues Gly66 and Leu67, located between subsites S2 and S3, to the molecular recognition of ligands. The amidic nitrogen of Gly66 along with the amidic oxygen of Asp158 frequently interact via hydrogen bonding with amides from known cruzain inhibitors. Furthermore, Leu67 is able to accommodate hydrophobic groups that occupy the S2 portion of the active site, thus contributing to adequate shape-matching of small molecules in this cleft. Hence, we have used these important interactions to set Gly66 and Leu67 as constraints for pose selection in all docking experiments. Based on these constraints and established thresholds, the molecular docking using the FRED program performed well at discriminating between active and inactive compounds when using the SHAPEGAUSS score function. Although the binding energy predicted by the molecular docking programs is not rigorously calculated, it was used as preliminary filter for lessening the number of molecules in the compound library.
As can be seen in Table 2, the thresholds established from available information of known ligands and employed as criteria for selecting compounds in the virtual screening (VS) experiments resulted in a significant enrichment of the compound collection. Since compounds retrieved according to the metric threshold adopted for each program may differ, we used this lack of agreement for the removal of more compounds from the enriched collection. Only those compounds retrieved by both programs that had a docking score or similarity index above the defined threshold were selected to join the smaller final sub-collection. Thus, the integration of target- and ligand-based methods resulted in the selection of a series of compounds with high potential to inhibit the enzyme cruzain and with diverse chemical features to allow the identification of new scaffolds to target this enzyme.
The reduction of the number of molecules in the virtual library allowed us to apply the consensus score strategy including pKi values calculated by the HQSAR (Hologram Quantitative Structure Activity Relationship) model and Glide Extra Precision scoring function (Glide XP), which requires a higher computational cost compared to the ROCS and FRED programs. Combined ranking was calculated using the scaled rank-by-number approach as described in [25]; the top 5% compounds were submitted to visual inspection.
The criteria used for selection of compounds based on visual inspection were: (i) the occupation of the site, mainly in subsites S1, S2 and S3, (ii) hydrogen bonding, emphasizing the Gly66 H-bonding interaction set up as constraint in molecular docking and (iii) chemical structure diversity, where we prioritized compounds with similar shape to known inhibitors. Figure 3A shows the binding mode of K11777 inhibitor to the enzyme cruzain in the crystal structure conformation, which was also used as reference in the 3D similarity search; Figure 3B to Figure 3D are the docking poses predicted by Glide XP for three compounds selected for in vitro assays. As can be seen, the occupation of the active site by these compounds is similar to that of the co-crystalized inhibitor, appropriately filling the focused pockets and fulfilling the requirements imposed through H-bonding constraint. [43]
After visual inspection, 23 compounds were selected for in vitro assays against the enzyme cruzain. First, the activity of compounds was determined through measurement of IC50 values in the presence of Triton X-100 (0.01% v/v) in order to avoid artifactual aggregate-based inhibition. Afterwards, those compounds that showed activities had their mechanism of enzyme inhibition and affinity constants determined. Dose-response curves were used for measuring the IC50 values and Michaelis-Menten curves in the presence and absence of inhibitors in three different concentrations, which allowed the determination of the mechanism of inhibition and affinity constants. Since the identified active compounds showed similar curves, representative curves are shown in Figure 4A and 4B for Neq30 of Table S1 (see Supporting Information). A similar approach for the discovery of cruzain inhibitors is described elsewhere by Ferreira et al. (see reference [44]). Using the docking strategy the authors were able to find one inhibitor out of 17 screened compounds that displayed a Ki of 32 µM, via a competitive mechanism of cruzain inhibition. Our consensus ligand-based virtual screening (LBVS) and target-based virtual screening (TBVS) approaches also using the Lineweaver-Burk plots in a similar fashion confirmed the competitive mode of action of 12 out of 23 compounds whose average IC50 value is 40.3 µM (with 3 compounds in the range of 3.5 µM) – see Table S1.
The dose-response plot in Figure 4A shows the inflexion of the semi-log curve fit with a pIC50 of 5.12 µM, which corresponds to the potency exhibited by Neq30 listed in Table S1. As shown in Figure 4B the increase in the concentration of inhibitor causes a decrease in the affinity of substrate but no change in the maximum velocity, thus indicating that the inhibitor competes for the same site as substrate Z-FR-MCA.
Cruzain shares high similarity with the mammalian homolog cathepsin L. Thus, it is of interest to evaluate selectivity [45] of selected compounds on both these enzymes, since promising trypanocidal scaffolds that act on cruzain could also be relevant to target cathepsin L, which has been studied as target for the treatment of cancer. Therefore, these compounds were tested against both enzymes. The 2D molecular structure representations of assayed compounds, IC50 and Ki values are shown in Table S1 of the Supporting Information.
The multi-step virtual screening protocol designed to identify cruzain inhibitors was validated by the discovery of 12 hits with affinity to the enzymes ranging from 3.7 to 89 µM among only 23 compounds assayed in the in vitro enzymatic assay. Recent studies report the search for cruzain inhibitors in the ZINC database using virtual screening methods. Ferreira et al. selected 17 compounds using molecular docking and found one active compound with an IC50 of 77 µM, which was further optimized by SAR to yield an IC50 of 200 nM [44]. Using a ligand-based virtual screening approach, Malvezzi et al. found one compound with low micromolar potency among 19 selected by a pharmacophore model [46]. By comparison, the number of hits we found using virtual screening methods was higher, which can be attributed to the integrated TBVS and LBVS consensual strategy. Consensus strategies and multi-step cascade protocols have been shown to achieve higher hit-rates when compared to LBVS or TBVS used separately [25], [47]–[49]. In order to confirm the activity of the scaffold found we carried out an initial SAR using the approach called SAR by catalog. This is a particularly powerful (and generally accessible) approach to the initial development of fragments [50]–[52], which assists in progressing through scaffolds with different R-groups.
The establishment of peptide SAR is often the first step in defining a critical, minimum sequence SAR for modulation of a particular target. The convenient and rapid synthesis of peptide analogs facilitates the identification of peptides having attractive biological properties. However, peptidomimetics possess well-recognized burdens as potential drugs, including susceptibility to enzymatic or chemical hydrolysis of peptide bonds and the metabolism of amino acid side chains, which influences the investigator in favor of discovering non-peptidic scaffolds with drug-like properties for cruzain inhibitors. Therefore, among the most potent compounds, there are new non-peptidic scaffolds including the compounds Neq24, Neq25, Neq30 and Neq42, for instance. Thus, before hit optimization, we evaluated trypanocidal activity of the six most potent compounds with the aim of selecting the most promising for molecular optimization. This assay was carried out at a single dose of 250 µM using the same procedure described above. Figure 5 shows the results for the two active compounds Neq42 and Neq37.
Compound Neq42, which represents a new scaffold for non-peptidic cruzain inhibitors, was able to kill the T. cruzi parasite and displayed similar activity to benznidazole, which was used as positive control in the assays. For this reason, we have chosen this compound as a reference to investigate its SAR.
Due to the moderate molecular complexity of the compound Neq42, a structure simplification strategy was adopted for SAR investigation taking as reference not only its potency but also its ligand efficiency (LE), a way of normalizing the potency and MW of a ligand to provide a useful comparison between compounds with a range of MWs and activities. The higher the LE of a hit, the better its chance of being optimized to a potent drug-like compound [53], [54].
The structure of Neq42 was resolved into two main scaffolds: the first containing the 2-acetamidothiophene-3-carboxamide moiety, and the other the triazole ring moiety substituted in positions 1 and 2 with the benzyl and piperidine, respectively, as can be seen in Figure 6. Based on the predicted mode of binding obtained by molecular docking, a series of structures was selected from commercial databases to investigate the SAR of this compound (Figure 6). When the 2-acetamidothiophene-3-carboxamidegroup was maintained, the piperidine and benzyl groups were removed to assess their contribution to the potency. On the other hand, when the triazole moiety was kept, three substances were selected by replacing the 2-acetamidothiophene-3-carboxamide group to a phenyl group (hydrophobic), a nitrile (which might covalently bind to the catalytic cysteine) and a cyclic sulfone (hydrophilic). The 2D structures and biological activity of the selected compounds are summarized in Figure 6.
The 2-acetamidothiophene-3-carboxamide is probably the one responsible for the activity presented by compound Neq42. As can be observed by the comparison between the active and inactive series shown in Figure 6, molecules lacking this moiety completely lose their activity against the cruzain enzyme. Nevertheless, potencies of compounds Neq165, Neq176, Neq177 are still in a similar order of magnitude as compound Neq42, notwithstanding a significant MW lowering that results in an increased LE for the analogs (0.33 kcal mol−1 atom−1 for compound Neq176). This observation signals that the piperidine and benzyl groups give rise only to a minor contribution to the potency, since removal of both groups (compound Neq176) resulted only in a slight decrease in potency, but with a significant LE improvement to 0.33 kcal mol−1 atom−1. The compound Neq172, which is a combination derived from compounds Neq38 and Neq42, resulted in smaller LE and potency, once again evincing the importance of 2-acetamidothiophene-3-carboxamide for the activity. Not only tailored LEs were substantially increased but also a new non-peptidic scaffold, which is an excellent starting point for optimization, was identified in agreement with the proposed binding mode shown in Figure 7. [43]
In order to evaluate the mode of binding (MOB) for Neq176, the crystallographic structure for cruzain-Neq176 complex was determined to 2.62 Å resolution. The protein crystallized at P43212 space group with five copies of cruzain in the asymmetric unit, with the compound Neq176 bound in three of the copies (chains A, B and C) in a similar mode of binding. The Neq176 binds in the S2 and S3 pockets, with the 2-acetamidothiophene-3-carboxamide group making two hydrogen bond interactions with the Nα of Gly66 (3.15 Å) and with the alpha oxygen of Asp161 (3.08 Å). The N(2) atom of the inhibitor also makes an important H-bond interaction with the alpha oxygen of Gly66 (3.09 Å). The inhibitor is stabilized at the S3 site by the interaction of the N(4) atom from the 1,2,4-triazole group with the γ oxygen of Ser61 (2.67 Å). Crystal structure showing the binding mode (MOB) of Neq176 with cruzain wild type can be found in Figure 7. Similar hydrogen bonding pattern of interactions was previously recognized for a covalent inhibitor (PDB entry 3IUT) [55]. Nonetheless, our X-ray structural data analysis confirm that the fragment Neq176, albeit within the active site, is not covalently bound to native cruzain since the carbonyl center group in the thiophene moiety is 4.20 Å distant from Sγ of the Cys25. The coordinates and structure factors of the protein were deposited in Protein Data Bank (PDB) with the ID code 4KLB.
An important feature in these compounds is related to the mechanism of inhibition they disclose, which are competitive and reversible. The validation of these cruzain enzyme inhibitors, along with their innovative molecular class pinpoint these compounds as promising for the development of new trypanocidal agents. In order to further strengthen the molecular basis of these findings, active compounds shown in Figure 6 were assayed against cultures of T. cruzi parasite following the protocol described above. Results are summarized in Table 3.
As can be seen from Table 3, compound Neq42 analogs presented trypanocidal activity at micromolar range validating this molecular class not only as enzyme inhibitors, but also as trypanocidal agents (pIC50 ca. 3.8 on average for compounds Neq165, Neq172, Neq176, and Neq177, and 4.9 for compound Neq42 versus 4.2 for benznidazole). Furthermore, as observed for the activity against cruzain enzyme, the decoration in the triazole ring does not significantly contribute to the trypanocidal activity among the actives. This suggests that major contribution for activity arises from the 2-acetamidothiophene-3-carboxamide group, which put forward the idea that triazole replacement might be a crucial element for designing new compounds with improved potency.
Compounds Neq177 and Neq176 are active cruzain and T. cruzi inhibitors that are cytotoxic only at concentrations above 250 µM, whilst compounds Neq42 and Neq172 are cytotoxic at IC50 of 50 and 23 µM, respectively (Table 3). These values were construed by tailoring down the chemical complexity of compound Neq42, which yielded the percentage of cell death caused by compounds Neq176 and Neq177 as evaluated against the cultured mouse spleen cells to dramatically drop and to show a significant decrease in cytotoxicity values. Thus, although compounds Neq176 and Neq177 are the least cytotoxic trypanocidal agents toward the studied cells, their potencies toward T. cruzi decreased along with decreased potencies against cruzain. Compound Neq42, on the other hand, has the higher potency in this series against cruzain, paralleling its higher trypanocidal potency, but with increased cytotoxicity. Nevertheless, it is noteworthy that the cytotoxicity ratio (IC50(cyto)/IC50)(T. cruzi) for compound Neq42 and benznidazole is within the same magnitude range values (5 and 8, respectively), and thus these structure-activity and-toxicity relationships enlarge lead optimization opportunities.
In addition, it is known that compounds bearing the 2-acetamidothiophene-3-carboxamide moiety are also found to be inhibitors of IKKβ kinase phosphorylation of IκB, thus being inhibitors of NF-κB activation [56]. Cruzain knockout with inhibitors are also lethal for T. cruzi via the signaling factor NF-κB P65, which is colocalized with cruzain on the cell surface of the intracellular wild T. cruzi [57]. Hence, we may envisage that our compounds will be of interest in the search for new drug candidates acting on inflammatory cardiomyopathy that is a hallmark on Chagas disease, but acting as non-covalent inhibitors.
Chagas disease is a neglected trypanosomiasis with enormous social and economic impact in most countries of Latin America. It is of utmost importance to develop new and more effective drugs with fewer side effects than the currently available chemotherapy. Hitherto, significant efforts have been made focusing on cruzain enzyme as a promising target and compound K11777, a cruzain inhibitor set to enter clinical studies as a new antichagasic drug. Here, we successfully used integrated in silico and in vitro approaches, with X-ray crystallography as an orthogonal tool, to discover new non-peptidic hits with trypanocidal activity against cruzain. Thus, we identified new trypanocidal agents that bear the 2-acetamidothiophene-3-carboxamide as the group responsible for enzyme inhibition and trypanocidal activity. The 2-acetamidothiophene-3-carboxamide binds non-covalently to cruzain, does not violate the rule of five and actually is a fragment with proper ligand efficiency (0.33 kcal mol−1 atom−1), with a low molecular mass (283.3 g mol−1) and CLogP of 0.7, properties that illuminate the way ahead for maneuvering toward a lead-like molecule.
In summary, we present a new hit, 2-acetamidothiophene-3-carboxamide, that non-covalently inhibits cruzain, has trypanocidal activity and manageable structure-activity and structure-toxicity relationships. We anticipate that this compound will advance the lead optimization process for Chagas disease chemotherapy.
Detailed information on cruzain EC 3.4.22.51 (also known as cruzipain) can be found at http://www.brenda-enzymes.org/php/result_flat.php4?ecno=3.4.22.51&Suchword=&organism%5B%5D=Trypanosomacruzi&show_tm=0 and also at https://www.ebi.ac.uk/chembldb/target/inspect/CHEMBL3563. For Cathepsin L, this site is also of interest: https://www.ebi.ac.uk/chembldb/target/inspect/CHEMBL3837.
|
10.1371/journal.pcbi.1004921 | Conservation in Mammals of Genes Associated with Aggression-Related Behavioral Phenotypes in Honey Bees | The emerging field of sociogenomics explores the relations between social behavior and genome structure and function. An important question is the extent to which associations between social behavior and gene expression are conserved among the Metazoa. Prior experimental work in an invertebrate model of social behavior, the honey bee, revealed distinct brain gene expression patterns in African and European honey bees, and within European honey bees with different behavioral phenotypes. The present work is a computational study of these previous findings in which we analyze, by orthology determination, the extent to which genes that are socially regulated in honey bees are conserved across the Metazoa. We found that the differentially expressed gene sets associated with alarm pheromone response, the difference between old and young bees, and the colony influence on soldier bees, are enriched in widely conserved genes, indicating that these differences have genomic bases shared with many other metazoans. By contrast, the sets of differentially expressed genes associated with the differences between African and European forager and guard bees are depleted in widely conserved genes, indicating that the genomic basis for this social behavior is relatively specific to honey bees. For the alarm pheromone response gene set, we found a particularly high degree of conservation with mammals, even though the alarm pheromone itself is bee-specific. Gene Ontology identification of human orthologs to the strongly conserved honey bee genes associated with the alarm pheromone response shows overrepresentation of protein metabolism, regulation of protein complex formation, and protein folding, perhaps associated with remodeling of critical neural circuits in response to alarm pheromone. We hypothesize that such remodeling may be an adaptation of social animals to process and respond appropriately to the complex patterns of conspecific communication essential for social organization.
| Sociogenomics explores the relationship between social behavior and the genome. An important issue is the extent to which results from social insects can be used to understand social behavior in other animals. We address this question through computational studies of previously published experimental data on patterns of brain gene expression in honey bees in response to particular environmental conditions and stimuli. We found that for one particular stimulus, response to alarm pheromone, the set of honey bee genes differentially expressed in the brain contains disproportionately large numbers of genes also found in mammals, including humans. This enrichment in orthologous genes suggests surprisingly strong similarities in socially responsive genetic circuits common to honey bees and mammals. A large number of the human counterparts of these genes are important for regulating protein folding, protein metabolism, and regulation of protein complex formation, perhaps reflecting changes in macromolecular complexes involved in remodeling critical neural circuits in response to the alarm pheromone. Noting that alarm pheromone is a component of the honey bee’s communication system, we hypothesize that such rapid remodeling may be an adaptation in the brain cells of social animals to deal with the complex patterns of conspecific signaling essential for social organization.
| Social behavior, like phenotypes of any level of complexity, is regulated by the activity of genomic networks and resulting gene expression. At the same time that specific examples of genes influencing behavior were being discovered empirically[1,2], the field of systems biology was developing[3]. The essence of systems biology is to use computation and genomic technologies to enable detailed observation at the sequence level of the dynamics of cell, tissue, and organism responses to specific challenges. The power of systems biology is that it enables comprehensive dynamic patterns of transcription, translation, post-translational modification, and functioning of gene products to be observed and analyzed. These approaches provide fertile ground for the development of testable hypotheses and ultimately confident inferences about the relationship between the genome and phenome (the sum total of the organism’s phenotypic traits), even when the phenome is based on complex patterns of gene interactions. The systems approach has catalyzed the development of the fields of evo-devo[4] and, more recently, sociogenomics [1]. Sociogenomics focuses on how genes influence social behavior [2] and how environmental attributes—especially those related to the social environment—influence genome activity [5].
Evo-devo has led to new insights into the molecular basis for the evolution of morphological novelties, molecular mechanisms underlying the development of morphology in the individual, and how development responds to the environment on a genomic level. Specifically, it has shown that the major (but not only) driver in evolution of form has been changes in expression patterns of functionally conserved genes [6] Synergistically, sociogenomics seeks to provide insights into the evolution of social behavior, the genomic mechanisms underlying social behavior in an individual and a species, and how social behavior is influenced by the environment at the genomic level [1]. Similar to the evolution of biological form, the evolution of a vertebrate social decision-making network has been shown to be largely (but again not entirely) by variations in conserved genes and networks [7].
One approach to sociogenomics is hypothesis-driven. In this approach, researchers begin with a hypothesis about the role of a gene or a group of genes in social behavior based on prior knowledge of the function or activity of those genes. As an example of this approach, O’Tuathaigh et al [8] observed that the knockout of the mouse ortholog of the human schizophrenia gene neuregulin 1 disrupted social novelty behavior, but left spatial learning and working memory processes intact. This gene has close homologs throughout the vertebrates, putative orthologs in arthropods, and significantly similar homologs annotated as coding for cell wall anchoring proteins in some bacteria.
By contrast, systems biology studies often begin with no hypothesis (except the fundamental one that social behavior has genomic bases) and scan comprehensively to see what correlations emerge. As an example of this approach, Cummings et al [9] identified differential gene expression patterns in the response of female swordtail fish to different classes of conspecifics (attractive males, unattractive males, other females). This broad systems approach was extended across multiple species in a study in which molecular orthology and comparative brain morphology were used to identify social behavior networks in vertebrates [10]. This work nicely illustrates the above-mentioned convergence of sociogenomics and evo-devo.
The studies cited above highlight the fact that understanding the genomic correlates of human social behavior requires us to use a variety of model organisms, in part because of the invasive nature of many experimental protocols. Ebstein et al [2] observed that “Human beings are an incredibly social species and along with eusocial insects engage in the largest cooperative living groups in the planet’s history.” This leads to the question: to what extent are there relevant genomic correlates between eusocial insects and humans, given that the last common ancestor of eusocial insects and humans lived approximately 670 million years ago [11] and almost certainly was very different in appearance from either an insect or a vertebrate. It may be that both eusocial insect and human social traits are elaborations and modifications of underlying patterns that were present in a common ancestor, even if the elaborations occurred independently[12]. As a corollary to this view, some species in the lineages leading to both insects and chordates would have lost or inhibited expression of these traits, while other species such as the eusocial insects and humans would have continued to express them and use them as a set of building blocks for social behavior. To the extent this is true, comparative genomics of eusocial insect social behavior and human social behavior may yield insights into some of the most fundamental aspects of the genomics of social behavior. This would be an example of the general principle that conserved elements between species separated by great evolutionary distance are likely to be universal building blocks of common aspects of the species’ phenomes [6].
Among the eusocial insects, the honey bee is a valuable model organism. Many experiments have linked brain gene expression patterns to social behavioral characteristics and environmental stimuli, and the honey bee genome has been sequenced [13]. In addition, individual members of a honey bee colony have well-defined social roles. It is known that the division of labor within the hive is based on both genetic differences between individual honey bees and also on environmental influences that include visual, tactile, and chemical signals that colony members send to each other, as well as environmental influences external to the colony [13]. However, the interplay between these factors is poorly defined with respect to variation in particular genes or regulatory domains in the genome. There are statistically detectable hereditary tendencies for particular honey bees to play particular social roles, but the individual bee’s social role is determined by the interactions between both social and environmental factors, as well as heredity. Understanding this complex interplay of internal and external factors is central to sociogenomics.
One way to make a connection between honey bee and human sociogenomics is by inference of genetic orthology. Unfortunately, orthology is of necessity not verifiable in the same fashion as other techniques of bioinformatics, since it involves theoretical reconstruction of an evolutionary history that cannot be experimentally replicated. Thus, there is no reliable validation set on which to test a method. Different reasonable ways of creating orthologies may give significantly different results [14]. Whether one makes a liberal or conservative interpretation of orthological relationships produced by a particular method depends on the context, in particular whether one is concerned about contamination by false positive identifications of orthologs, or more concerned about loss of information by false negatives. In the present paper, we use a new application of orthology to test the hypothesis that the social behavior of honey bees and other metazoans, including humans, has common fundamental genomic building blocks.
This paper seeks to explore the degree of relevant sequence conservation between honey bees and humans. Our starting point is the data set from Alaux et al [15], who used microarrays to analyze differential brain gene expression patterns exhibited by individual honey bees of different genetic backgrounds, engaged in different social roles and in different colony environments. African and European honey bees are subspecies of the Western honey bee, Apis mellifera, and they differ from each other in their hive defense behavior in a number of ways that have been summarized as a social behavioral counterpart to variations of threshold and intensity of the “flight or fight” response seen in vertebrate organisms; African bees are much more aggressive than European bees [16]. In general, different phenotypes may arise from either differences in gene function or from different patterns of gene expression [17]. In the African and European honey bees it is presumed that the different phenotypes are largely the result of different patterns of gene expression, and differences in the expression of hundreds of genes in the brain have been reported [15]. Bees in Alaux et al were raised in a cross-fostered experimental design to examine the influences of genetic background and social environment on brain gene expression.
We analyzed the above-cited [15] data sets to explore the following two questions: 1) to what extent are the differentially expressed genes associated with social behavior in the honey bee conserved across the Metazoa; and 2) through analysis of the highly conserved genes, is it possible to infer that there are likely to be gene co-expression patterns associated with social behavior that are common to a wide range of metazoans, including humans?
We examined eight sets of social behavior-related differentially expressed genes from Alaux et al [15]. They are described in Table 1 and S3 Table.
For each of the honey bee genes on the microarray, we interrogated the InParanoid database of orthologous genes [18] to ascertain how many orthologs each gene had within a set of organisms including yeast plus 53 metazoa. The results are plotted in Fig 1 in the form of a histogram that shows what fraction of the genes had 1, 2, …. 54 orthologs. Just under 5% of the genes had no orthologs in the InParanoid set; within this data set they were unique to the honey bee. Of the 54 species we compared to the honey bee, 17 were insects. The position of the first peak in the distribution (at 15 orthologs) is due to genes that were largely conserved in insects and were uncommon in other metazoan lineages. The position of the second peak (at 50 orthologs) was due to genes that are broadly conserved across the metazoa. There were 1631 honey bee genes in the InParanoid database that were not included on the microarray. Approximately one third of those 1631 genes not included were unique to the honey bee (Fig 2). Over 50% of the 1631 genes had fewer than four orthologs in the set of 54 species analyzed. This relative lack of conservation of the excluded genes is largely a function of how the microarray was designed [15] Since the major conclusions of this paper were based on orthology to other metazoa, and since the genes excluded from the analysis had relatively few such orthologs, the conclusions are unlikely to be significantly affected by the exclusion of these genes.
Table 2 provides the overall summary of the results. At the 0.05 significance level (based on Benjamini-corrected p-values), three of the sets were selectively enriched in genes conserved across the Metazoa: the Alarm_Pheromone set, the Old_vs_Young set, and the Soldier_CG (colony genotype) set. By the same standard of significance, the Guard_CG, Guard_WG (worker genotype), and the Forager_WG sets were significantly depleted in highly conserved genes (i.e., the Benjamini-corrected p-value was over 0.95).
We examined the conservation pattern with each of the species used in the analysis, via a heat map, for the eight data sets (Fig 3). These analyses were based on the InParanoid orthology database (Fig 3A) and the OrthoMCL orthology database, which contained a smaller number of species (Fig 3B). A relatively high degree of conservation was distributed across a wide range of metazoans for Old_vs_Young, Alarm_Pheromone, and Soldier_CG sets. For Soldier_CG and Old_vs_Young, the most significant conservation was within the insect group. For the Alarm_Pheromone set, on the other hand, Fig 3A and 3B indicate that the greatest degree of conservation was in mammals. Another way of visualizing the greater degree of conservation in mammals is in Fig 3C, which shows box-and-whisker plots of the distribution of p-values for orthology enrichment of the honeybee Alarm Pheromone set for the honeybee’s closest relatives (arthropods) and for human’s closest relatives, the highly social placental mammals. For both the InParanoid and the OrthoMCL databases, the degree of conservation clearly tends higher (lower p-value) for the mammals than for the arthropods. To test the statistical significance of the greater conservation of the Alarm Pheromone set in placental mammals we applied the Kolgomorov-Smirnov (KS) test, which is a standard method for assessing the significance of the difference between two unbinned distributions [19]. Fig 3D and 3E show the KS comparison cumulative fraction plots for the arthropod/placental mammal p-value distributions from the Alarm Pheromone gene set using the InParanoid and the OrthoMCL orthology databases, respectively. In these plots the horizontal axis represents the range of p-values for orthology enrichment and the vertical axis represents the fraction of species in each class whose p-values are below a particular level. The critical features of each graph are the quantity D, representing the maximum different between the plots for the two distributions, and a corresponding P (The likelihood that the difference between the distributions arose by chance, which is a function of D and of the number of values in the two distributions; see Press et al, 1992 [19], for exact expression for computing P). For the InParanoid set, the value of D is 0.75, meaning that the lowest quartile of the p-values for the arthropods is within the range of the p-values for the placental mammals, while the upper 75% of the arthropod p-values is larger than any of the placental mammals. The value of P (the likelihood that this discrepancy between the distributions arose by chance) is .001. In Fig 3E, which shows the comparison cumulative fraction plots for the distributions as derived from the OrthoMCL data base, the value of D is 1, because there is no overlap between the distributions. The largest p-value of any of the placental mammals is smaller than the smallest p-value for any of the arthropods. Therefore the value of P is vanishingly small. Based on these statistics, we confidently conclude that the genes differentially expressed in the honey bee in response to the alarm pheromone are systematically enriched in orthologs to genes in placental mammals.
This finding suggests that, of all the gene sets analyzed, the set differentially expressed in response to the alarm pheromone stimulus was most likely to include genes from genomic networks common to honey bees and mammals. The analyzed gene expression data and the results of the orthology searches are provided in spreadsheet form in S3 Table.
In order to be conservative in our assignment of orthologs (minimize false positives, even at the expense of incurring false negatives) we chose for detailed further analysis the set of 145 genes that were differentially expressed in the alarm pheromone response and conserved in all the Eutheria (placental mammals) species (altogether 10 species in InParanoid, ranging from B.taurus to H. sapiens) considered in this study. The p-value for over-representation of orthologs of placental mammals in this set was actually smaller than 1e-6 (see Methods), which constitutes a correlation effectively impossible to have occurred by chance. Similarly, the most significantly conserved genes for all the insect species in the Old_vs_Young set were identified by a correlation effectively impossible to have occurred by chance (also with a p-value smaller than 1e-6). A larger set of genes (conserved in mouse and human but not necessarily in all 10 eutherian species) was also analyzed, as was a smaller set of genes conserved in all the vertebrates. Generally, the mouse-and-human set showed very similar GO enrichment patterns to the eutherian set, while the all-vertebrate set showed far fewer enriched ontology classes. Results of this analysis are provided in supplementary material.
In each of the three classes of bees (soldier, forager, guard) where we have both a CG gene set (differential gene expression between bees raised in predominantly African and European colonies) and a WG gene set (differential gene expression between genetically African and genetically European honey bee), we compared the degree of enrichment in orthologs with other metazoans. There was greater enrichment in orthologs in the CG set than in the WG set (p = .043 for guards, p < .0005 for foragers, p < .0005 for soldiers). The soldier cg-wg orthology is especially interesting for two reasons. Firstly the overall degree of orthology is much greater for the soldiers than for the foragers or guards. Secondly the most dramatic behavioral difference between the African and European bees is the behavior of the soldiers. The degree of difference between the soldier cg and wg orthologies is visualized in Fig 3F, which shows the cumulative fractional difference of the two distributions of p-values for pairwise orthology enrichment between the honey bee and the other 54 organisms represented in the analysis. It is important to note that the behavioral phenotype of the soldiers corresponds mainly to the phenotype of the colony in which they were raised. The cross-fostered soldiers are phenotypically much like the other soldiers in their colony, but differ in gene expression patterns. Our finding speaks to the general issue of the interaction between nature and nurture in defining social behavior, suggesting that if we wish to draw inferences for other metazoans from the different behavior of African and European honey bees, we must consider how the colony socializes individual bees. At the genomic level, this suggests that the overall genetic composition of African and European colonies (which would presumably be reflected in the nurturing environment in the hive, but is beyond the scope of the current study) is perhaps more important than the genetic differences of individual bees for understanding the broader comparative relevance of strain differences in behavior. Note also that the pattern of orthology enrichment across the metazoa is quite different for the soldier cg set than for the alarm pheromone set. Whereas the alarm pheromone set showed enriched orthology particularly for the highly social placental mammals, the orthology enrichment for the soldier cg set is higher for closer relatives to the honey bee, most notably the arthropods—most of whom are not eusocial. To summarize the orthology results:
We used the DAVID suite of programs to identify Gene Ontology (GO) categories that were over-represented in the 145 alarm pheromone-responsive genes mentioned above, relative to their overall incidence in the human genome (131 of these 145 genes” human orthologs have Entrez annotations), at p-values of 0.01 and 0.05 (Benjamini-corrected for multiple hypothesis assumption). For better comparison, we performed three separate GO analyses: 1) for all these 131 genes, 2) for the 73 up-regulated genes 3) for the 58 down-regulated genes. The results are summarized in Figs 4 and 5 and in Tables 3 and 4. Tables 3 and 4 provide the names of the enriched GO categories, together with the p-value for their enrichment. The GO output analysis output, upon which Figs 4 and 5 and Tables 3 and 4 are based, is shown in spreadsheet form in S2 Table. The analysis below is based specifically on the gene set that was conserved among all the Eutheria. We also did the analysis on a larger set of genes conserved in the mouse and human but not necessarily in all the Eutheria. The results of that analysis was practically identical to the Eutheria-conserved set, so the verbal analysis below applies to that set as well.
Several of the same GO categories appeared in the results of more than one of the three analyses (up-regulated, down-regulated, all differentially expressed). Three GO categories were enriched in all three of the analyses, all three in the “Cellular Component” category (Table 4). They are: GO:0005737 (cytoplasm—“All of the contents of a cell excluding the plasma membrane and nucleus, but including other subcellular structures”), GO:0005622 (intracellular—“The living contents of a cell; the matter contained within (but not including) the plasma membrane, usually taken to exclude large vacuoles and masses of secretory or ingested material. In eukaryotes it includes the nucleus and cytoplasm.”), and GO:0044424 (intracellular part—essentially the same definition as GO:0005622 and with the same parent term, GO:0044464 (cell part). (Definitions in quotation marks are from EBI QuickGO[20]). The enrichment of these three terms in all of the three categories of differentially expressed genes means that few of the differentially expressed gene products reside in the plasma membrane, and both up-regulated and down-regulated genes were enriched for gene products found in other parts of the cell. The rest of the Cellular Component categories provided more specificity with respect to the locations of up-regulated and down-regulated genes.
Examination of enrichment in “Biological Processes” categories revealed several insights (Fig 4). There was a strongly enriched GO category under “cellular component organization or biogenesis” [node 54]—“macromolecular complex subunit organization” [node38] (Benjamini p-value = 0.0096). This enrichment suggests that the human pattern orthologous to the expression pattern of the honey bee alarm pheromone response involves protein complex organization and biogenesis. This GO term was not significantly enriched for down-regulated genes. 2) “Cellular metabolic process” [node10] was also an enriched GO term (Benjamini p-value = 0.016). This suggests that the human pattern orthologous to the expression pattern of the honey bee alarm pheromone response involves modulation of metabolism. 3) More specialized categories within the “response to stimulus” GO term were “response to stress” [node 31] and “response to unfolded protein” [node 29]. Taken together, these enrichments suggest that the human response pattern that is orthologous to the honey bee alarm pheromone response also involves responses to chemical and possibly other stimuli. It is plausible that the response to unfolded protein seen in this section of the tree was related to protein metabolism and biogenesis, and the protein complex assembly that was simultaneously being up-regulated during the overall organism response as indicated in other parts of the tree. “Protein folding” [node 15] was also enriched.
Gene Ontology analysis for molecular function revealed that that all the enriched GO terms fall under one general category—“binding” (22). GO analysis for cellular component (Fig 5 and Table 4) revealed the enrichment pattern included multiple cell components—cytoplasm, nucleus, mitochondria (mostly significant for down-regulated genes) and other organelles, protein, and possibly other macromolecular complexes. This was consistent with the biological processes and the molecular functions enriched in our analyses, which are localized in in a variety of cell components.
Since the members of the gene set from which these inferences are derived were conserved across the eutherians, it is plausible that the inferences are valid for eutherians in general. However, it should be reiterated that the results described in this section do not refer to the totality of either the honey bee alarm pheromone response or of a complete network in humans and other vertebrates. Rather, they refer to components of the honey bee alarm pheromone response network that are widely conserved in eutherians and have a well-defined GO classification in humans. These components were presumably present and possibly part of an interacting network in the last common ancestor of the human and the honey bee about 670 million years ago. Both the honey bee alarm pheromone network and networks in eutherians that share these components will undoubtedly have other different non-shared components particular to their respective classes of organism.
Tables 5 and 6 show genes in our analysis set annotated with enriched GO biological process terms that have been implicated in neural and behavioral disorders, and those biological process terms with which they are associated that are also included in the list of enriched terms for the complete alarm pheromone set. This list was constructed by manual inspection of literature and OMIM databases, so is not comprehensive. The results of a GO analysis for this set of 25 genes is given in S4 Table, showing all biological process terms enriched to a p-value of 0.05 or better. The overwhelming majority of the enriched biological processes relate to metabolism in a way that would pertain to many different types of cells in addition to brain cells. Protein folding and organization of macromolecular complexes also appear as enriched categories. These genes are selected for both a specific brain response in the honey bee and also for broad conservation in the placental mammals. The interesting feature of this analysis is the convergence of three factors: 1) implication in human mental disease, 2) differential expression in the honeybee in response to a conspecific language element (the alarm pheromone) and 3) broad conservation across the placental mammals. It appears at least in part that several varieties of mental illness are based on issues related to evolutionarily deeply rooted and broadly conserved genes, as opposed to being solely related to genes specific to human cognition and behavior, or even specific to brain or neural function.
This study was designed to examine the plausibility of the premise that the genomic networks underlying a response to a stimulus for social behavior (alarm pheromone response in honey bees) might have counterparts conserved in mammals, even though mammals do not use this particular alarm pheromone and the last common ancestor between honey bees and mammals lived approximately 670 million years ago [11]. Based on results from two different orthology databases, we found that the honey bee genes differentially expressed in response to alarm pheromone were more strongly conserved in orthologs to mammals than in orthologs to other metazoans, including those more closely related to the honey bee (nonsocial insects). We hypothesize that these orthologous sets are conserved remnants of a network responding to conspecific signals that first emerged in a common ancestor of insects and vertebrates and has been selectively conserved in social metazoa.
The reader will have noted that the experimental context of this paper was done on material from whole brains. For processing of conspecific signals such as spoken or written language in humans, many imaging studies show that several different regions of the brain are simultaneously activated. We therefore believe that whole brain studies such as ours are useful in revealing underlying commonalities of mechanism, but should be complemented by region-specific analyses.
It should be noted that this particular study deals only with those parts of the putative conserved network that are differentially expressed in response to the external signal. There may be other genes that are part of the network, but are present at relatively steady levels. This may be the reason for the conspicuous lack of genes for plasma membrane proteins in the “cellular component” category of enriched GO classes found in this study (Table 4). Plasma membrane proteins must be involved in any response to external signals, but their role in mediating between extracellular stimuli and intracellular response does not necessarily require either up- or down-regulation in immediate response to the alarm pheromone stimulus.
By contrast, genes in our study whose products reside in the nucleus were upregulated, genes in the mitochondria and other organelles were downregulated, and significant numbers of genes in the remainder of the cell were differentially regulated in both directions. Our results indicate that alarm pheromone exposure triggers significant physical remodeling of intracellular molecular signaling machinery. At the core of sociality is the ability to transmit and respond to complicated signals from conspecifics [21]. This is widely thought to involve the ability of nervous systems to rapidly increase the activity of some cellular networks and reduce the activity of others in response to these signals [22]. Our results suggest that there is another level of complexity enabled by the ability to remodel macromolecular interaction networks within cells in response to a transient signal from conspecifics, such as alarm pheromone. This remodeling would allow for changes in responses to subsequent signals, i.e., for stimuli experienced presently to enable individuals to “predict” the future. Since our results are based on enrichment of orthologous genes between honey bees and mammals, this hypothesis implies the original development of this remodeling ability in an ancient common ancestor of mammals and insects.
Based on these results we offer the following speculation about possible mechanisms for macromolecular remodeling within brain cells and organismic sociality. The time scales for protein folding, for binding reactions, and for assembly of macromolecular complexes from pre-existing elements, can be fractions of a second, so these processes can take place rapidly enough to be consistent with the time scale of the alarm pheromone response. However, transcription and translation of genes will take many seconds or minutes [23]. The necessarily faster time scale for the alarm pheromone response suggests involvement of a more rapid remodeling process, perhaps involving microRNAs, which have for several years been postulated to play a role in synaptic plasticity [24]. The recently developed CLIP-seq technology [25] permits comprehensive identification of microRNA binding sites in a variety of tissues, including the brain [26]. Thus it should be possible in the future to explore this speculation and experimentally characterize the roles of specific microRNA in brain remodeling in response to conspecific signals.
Perhaps one aspect of the dichotomy between highly social and solitary animals is in the ability of the individual brain cells in social animals to remodel their interaction networks in response to signals from conspecifics. This ability would not come without a tradeoff, since maximal speed of response would be achieved by activating existing hard-wired networks. Thus evolutionary niches have persisted for both highly social and less social animals, with less social animals optimized for speed of response to all stimuli by activating hard-wired circuits, while highly social animals have developed the ability to remodel molecular circuits in response to signals from conspecifics—a process which results in necessarily slower response. This may also apply to the evolution of the most complex form of conspecific communication–human language. In this view the corresponding circuits underlying honey bee chemical language and human auditory language would be “phenologs”; that is, varying phenomes based on orthologous genes [27].
The p-values in Table 2 for the average number of metazoan orthologs for each data set were computed as follows: For each experimental data set, random sets of matching size were sampled from the 7462 honey bee genes that were present in InParanoid database [18] and spotted on the array, and the average number of orthologs per gene was calculated for each random set. This random sampling was repeated one million times and the number of random sets with average ortholog number equal to or larger than the experimental set was counted. The count divided by 106 gave us the p-value for the average ortholog number of the test set. S1 and S2 Figs show how the p-values of the average ortholog number of Forager_CG and Alarm_Pheromone sets were calculated. The p-values for the total number of orthologs of each set for each species (Fig 3) were computed similarly.
For calculating the p-value for the CG-WG difference, the KS-test p-values for the CG-WG difference for Soldier, Forager and Guard (0.026, .122, and .612 respectively) were combined using Fisher’s method [28].
For calculating the p-value for over-representation of orthologs of placental mammals in the Alarm_Pheromone set and over-representation of orthologs of insects in the Old_vs_Young set, p-values in each species (S1 Table) were also combined using Fisher’s method.
In presenting and discussing the results, we use the term “conserved” to be measured by the number of orthologs that a particular sequence has; i.e., the more orthologs a gene or protein has in other species, the more “conserved” the gene is.
Enrichment of the conserved gene sets in particular Gene Ontology categories was determined using the functional annotation tool in the Database for Annotation, Visualization, and Integrated Discovery (DAVID)[29]. All parameters are default except that we use GO_TERM_*_ALL instead of GO_*_FAT. Extra functional analyses (of various qualities) were also included: OMIM_Disease [30], COG_Ontology [31], SP_PIR_Keywords [32], Up_Seq_Feature [33], BBID [34], BioCarta [35], Kegg_Pathway [36], Interpro Domains [37], Pir_Superfamily [38], and Smart [39].
The raw Gene Ontology results of “Eutheria-conserved”,Alarm_Pheromone genes are listed in S2 Table. Figures of Gene Ontology trees (Figs 4 and 5) were generated by Python scripts and Cytoscape [40]. Benjamini-Hochberg corrected p-values provided by DAVID are used for indication of significance [29].
Scientific references about the relationship between behavior/neural functions and genes associated with significant GO terms were identified with GeneCard and manual search with Google Scholar, using keywords “behavior”,”disease”,”neural”, and”aggression”.
First, honey bee genes that showed up on the microarray studied in Alaux et al [15] were selected. This was done based on the annotation file of the Honey Bee Oligonucleotide Microarray [15]. Out of many available methods [14] of defining orthologs, two were chosen, InParanoid [18] and OrthoMCL [41]. InParanoid has the most extensive coverage of the honey bee proteome and other proteomes of completed genomes in searchable ortholog databases. Out of all these “microarray-present” honey bee genes, we identified those that are also present in InParanoid. This was done by mapping the BeeBase IDs (which are the IDs used in the data set from Alaux et al [15]) to NCBI Refseq IDs (which are the IDs used in InParanoid for honey bee). 7462 of these “microarray-present” honey bee genes are present in InParanoid. At the time of the analysis, there were 100 eukaryotic species in InParanoid with 54 of them (including Apis mellifera) being metazoan species. With S. cerevisiae added as a control, the data set used for our analysis had 55 species, which we interrogated for orthology with the 7462 InParanoid honey bee proteins.
|
10.1371/journal.pcbi.1006395 | Balance of mechanical forces drives endothelial gap formation and may facilitate cancer and immune-cell extravasation | The formation of gaps in the endothelium is a crucial process underlying both cancer and immune cell extravasation, contributing to the functioning of the immune system during infection, the unfavorable development of chronic inflammation and tumor metastasis. Here, we present a stochastic-mechanical multiscale model of an endothelial cell monolayer and show that the dynamic nature of the endothelium leads to spontaneous gap formation, even without intervention from the transmigrating cells. These gaps preferentially appear at the vertices between three endothelial cells, as opposed to the border between two cells. We quantify the frequency and lifetime of these gaps, and validate our predictions experimentally. Interestingly, we find experimentally that cancer cells also preferentially extravasate at vertices, even when they first arrest on borders. This suggests that extravasating cells, rather than initially signaling to the endothelium, might exploit the autonomously forming gaps in the endothelium to initiate transmigration.
| Transmigration of immune cells into and out of the blood vessels is a crucial process for the functioning of the immune system during infections and acute inflammations, and aberrant transmigration may contribute to chronic inflammations. Likewise, cancer metastasis critically depends on intra-and extravasation of cancer cells through the endothelium. While much research investigated the role of immune or cancer cells in signaling to the endothelium, facilitating effective transmigration, and some work uncovered a role of passive mechanical properties such as stiffness during transmigration, little is known about the active role the endothelium itself plays during such processes. Our computational model, together with new data, highlights the dynamic nature of endothelial cells, leading to gap formations through mechanical processes within the endothelium, without influence of cancer or immune cells. Thus, our results highlight the need to take the active mechanics of the endothelium into account when devising strategies to overcome the adverse effects of endothelial gap formation during inflammation or cancer.
| Immune and cancer cells alike are characterized by their ability to migrate within the vasculature and then to leave the vasculature into different tissues. These processes are crucial for a functioning immune system to fight acute infections [1] or participate in wound healing [2]. However, chronic inflammation or tumor metastases are ultimately also initiated by extravasating immune or cancer cells, respectively [3, 4, 5]. Hence, while extravasation is critical to cure communicable diseases, it is also a critical contributor to virtually all non-communicable disease, ranging from cancer to asthma, atherosclerosis, rheumatoid arthritis and heart diseases [6, 4, 7].
Much of the research on extravasation (often termed diapedesis in the context of immune cells) has focused on the role of the extravasating cell during this process, and how it interacts with the endothelial cells of the vasculature through which it is transmigrating. First, the extravasating cell needs to arrest in the vasculature. This may occur through single cells or clusters getting physically stuck in small capillaries, through the formation of adhesions, or both [8, 9, 10, 11, 12]. Such adhesion is mediated by molecules including P-and E-selectin, ICAM, VCAM or integrins [13]. The actual process of transmigration can occur through a single endothelial cell (transcellular extravasation) or, more commonly, in between two or more endothelial cells (paracellular extravasation) [14, 1].
During paracellular extravasation, it was investigated how the extravasating cell signals to the endothelial cells, leading to weakening of VE-cadherin-mediated cell-cell junctions and subsequently gap formation, through which the cells can transmigrate [8, 1]. Gap formation may, for instance, be stimulated by thrombin [15]. As such, molecular signaling events are firmly established as important contributors to extravasation of immune cells.
However, on a fundamental level, all the processes involved in extravasation are mechanical processes. Transmigration, like other forms of cell migration, involves the generation of mechanical forces through the actomyosin cytoskeleton [16]. Moreover, the mechanical properties of the endothelium provide passive mechanical resistance [16]. For instance, increased endothelial cell and junctional stiffness will reduce paracellular extravasation rates [17, 14]. Interestingly, recent research established that active mechanical properties of the endothelial cells are also critical during endothelial gap formation [18, 19, 20], and the rearrangements of cytoskeletal structures are associated with changes in barrier function. For instance, a rich actin cortex parallel to cell-cell borders is associated with stabilized VE-cadherin junctions and thus tight barriers [21, 22], whereas actomyosin stress fibers pulling radially on junctions can lead to junctional remodeling [18, 23]. Additionally, actin-rich pores can actively contract to prevent leakage during extravasation [24]. However, there is still a lack in mechanistic and systems level understanding of the different roles of active and passive mechanical properties of the endothelium.
Mathematical multiscale models are powerful tools to investigate the interplay of different physical drivers of biological processes. Many different approaches have been employed to model and understand the dynamics of epithelial monolayers. Agent based models, where individual cells are explicitly taken into account, include center based models (CBM) [25], vertex models [26, 27] and deformable models (DFM) [28, 29]. However, these models do not explicitly model cell-cell adhesion dynamics in a way that leads to the experimentally observed gap formation in monolayers of endothelial cells, and can thus not easily be employed to study this problem so crucial for cancer and immune transmigration.
In this paper, we introduce a mathematical multiscale model of the mechanics of an endothelial monolayer where each endothelial cell contains contractile actin structures that may contract radially or in parallel to the plasma membrane. Then, cells are tethered to neighboring cells by cell-cell junctions that can dynamically form and break in a force-dependent manner. We employ this model to investigate the mechanisms of gap formation in an endothelial monolayer. Interestingly, we find that gaps open dynamically in the absence of any extravasating cells. These gaps form preferentially at the vertices where three or more endothelial cells meet, as opposed to the borders in between two cells. This is in line with our experimental data obtained in-vitro from quantifying gap formation of monolayers of human umbilical vein endothelial cells (HUVECs) seeded on glass. We quantify the frequency of gap openings as well as the duration of gap openings, obtaining good agreement between numerical predictions and experiments. Moreover, through multi-dimensional parameter studies, the mathematical model is able to give us insights into the physical and molecular drivers of the gap formation and gap dynamics. The model predicts that active and passive mechanical forces play an important role in the initial gap formation and in controlling size and lifetime of gaps once they initially formed. The catch bond nature of the cell-cell adhesion complexes as well as the force-dependent reinforcement of adhesion clusters may both stabilize junctions in response to forces acting on them. However, while the catch bonds ultimately weaken when forces are increased beyond the maximal lifetime of a single molecular bond, the force-dependent reinforcement will increase adhesion strength with increasing force [30, 18]. While the catch bond nature and the force dependence of the adhesion clustering processes both crucially influence gap opening frequencies, we find that gap lifetime and gap size are even more sensitive to the passive mechanical properties of the cell. Increased stiffness of the membrane/cortex and, even more notably, of the actin stress fibers will reduce lifetime and size, since the cells will then increasingly resist opening gaps through counteracting forces. On the other hand, we find that changes in bending stiffness of the membrane/cortex may have gap promoting or inhibiting effects.
Our model predictions of gap opening frequency and lifetime at both cell vertices and borders are validated by experiments observing such gaps in endothelial monolayers in the absence of any extravasating cell. The results thus challenge the paradigm that all extravasating cells primarily cause gap opening through interactions with the endothelium [1, 8, 31]. We then show experimentally that extravasating cancer cells indeed primarily extravasate at vertices, in line with similar observations for neutrophils [32]. Moreover, we show that cancer cells prefer to extravasate at vertices even when they initially attached to the endothelium at two-cell borders. This suggests that, even though extravasating cells can actively interact with the endothelium during transmigration, as shown in earlier studies, they may also take advantage of the autonomous occurrence of a gap, as predicted and verified to occur in our model. In summary, our work highlights the importance of taking the dynamic and autonomous mechanical properties of the endothelium into account when trying to understand gap formation and extravasation.
We present a novel model of an endothelial cell (EC) monolayer that incorporates different intracellular mechanical structures and dynamical cell-cell adhesions. The intracellular mechanical state is determined by radial contractile actin stress fibers and the cell membrane together with the actin cortex. For simplicity, we combined membrane and cortex into single viscoelastic elements, composed of an elastic spring and a viscous damper, that we refer to, from now on, as membrane elements. The radial stress fibers are also modeled by viscoelastic elements with different mechanical properties from the membrane, similar to a model of epithelial cells [28] (see Fig 1A). Neighboring cells may form cell-cell adhesions at adjacent nodes, and the resulting adhesion bond is modeled through a spring. The passive mechanical properties of the monolayer are thus modeled through a network of connected elastic and viscoelastic elements, similar to models of epithelial sheets [29, 28]. Since we are interested in studying the opening dynamics of gaps in the endothelial barrier, we explicitly simulate the dynamical binding of adhesion complexes. Contractions represent myosin motor activity that is known to exhibit randomness [33], so we employ Monte-Carlo simulations to estimate the occurrence of such forces as well as that of protrusive forces due to actin polymerization. The forces are then redistributed across the network of connected viscoelastic elements. Cell-cell adhesion complexes that mechanically link neighboring cells can dynamically bind and unbind in a force-dependent manner. The adhesion complexes in the model provide an effective description of both bonds of cell-cell adhesion molecules (such as VE-cadherin) and bonds of these adhesion molecules to the cytoskeleton. Cadherins and adhesion-cytoskeleton bonds are known to increase their binding strength in response to smaller forces, before they ultimately rupture [34]. This catch-bond type behavior is included in our model, and unbinding is thus simulated through a force-dependent Monte-Carlo simulation. Moreover, the number of VE-cadherins in an adhesion complex is modeled through a force-dependent adhesion clustering mechanism, as described in [18, 23, 35, 36, 37]. A more detailed description of the mathematical model and its numerical implementation is given in S1 Text.
We employ our endothelial monolayer model to explore the dynamics of endothelial cell junctions. We predict the frequency, size and duration of gaps, as well as the preferred geometrical locations of the gap formation, and compare the predictions with our experimental measurements. The parameters used in the simulations are detailed in S1 Table. After comparing our predictions with the experimental results, we perform sensitivity analyses to investigate how cell mechanical properties, cell-cell adhesion characteristics and myosin generated forces regulate the formation, lifetime and size of gaps in the endothelium.
Here we present a summary of the major parameters of the model that had a significant impact on our model behavior, and were consequently thoroughly investigated through sensitivity analysis in the remainder of this paper. Table 1 lists all these parameters, and for a complete list and discussion see the Supporting Information. The main parameters investigated are related to cell mechanical properties, adhesion properties or myosin force generated processes.
Cell mechanical properties are dictated by stress fiber stiffness (Ksf), membrane stiffness (Kmemb) and bending stiffness (incorporated through a rotational spring constant, Kbend). Stress fiber stiffness controls the rigidity of the interior of the cell, whereas membrane stiffness controls the rigidity of the membrane and the adjacent actin cortex. Bending stiffness acts on the membrane nodes depending on the relative orientation between the edges connecting at a given node.
Adhesion properties are controlled by the mechanical properties of the adhesion complexes and their binding and unbinding rates. Adhesion complex mechanics are modeled by linear springs, controlled by their stiffness constant, K a d h 0. The binding rate depends on distance and can be controlled by the adhesion complex density, ρadh. We then model the reinforcement of a bond that is already formed by the additional recruitment of adhesive proteins into the bond. Reinforcement is force dependent and can be controlled by the binding rate constant for adhesion reinforcement, k r e i n f 0. Unbinding follows a catch bond behavior. The catch bond unbinding curve can be modified through two rate coefficients: k s 0, which represents a slip bond, and k c 0, which is the additional parameter characterizing the initial increase in the bond lifetime with force (see S12 Fig).
Then, the model includes contractile forces due to myosin motor activity, and protrusive forces that may arise due to actin polymerization. These forces can be directed radially (following the stress fibers direction) or in a tangential direction (following membrane segments). In the sensitivity analysis we have varied the magnitude of contraction forces in the radial direction (FRadial) and in the tangential direction (Fcortex).
Fig 1B and 1C and S1 Movie show typical simulations of the monolayer dynamics of the computational model. We observe that gaps open preferentially at vertices, i.e. the intersections of three or more cells, as opposed to the border between two cells. We have quantified this by counting the total number of gaps formed as well as their lifetime at borders and vertices of the cell in the center of the monolayer, and showed that our model predictions are in line with the experimental observations (Fig 1G and 1H). These experiments were performed by seeding HUVEC cells on glass, where they formed a continuous monolayer. The gaps were experimentally quantified through inspection of visible gaps within the VE-cadherin-GFP signal in the monolayer (arrows in Fig 1E and 1F). Controls simultaneously showing VE-cadherin-GFP and CD31 staining show that the VE-cadherin gaps are also visible in the CD31 staining, indicating that the VE-cadherin gaps correspond to real physical gaps between two or more cells S7 Fig (see Methods for further details of the experimental setup and quantification). Vertices are points where more than two cells exert forces and where tangential force components naturally propagate to. Therefore, it is expected that stress concentrates at the three cell vertex rather than at the two cells borders, and the simulations confirm this hypothesis (Supplementary S9 Fig and S3 Movie). The forces on adhesion clusters at the vertices are thus more likely to exceed the corresponding force of maximal lifetime of the bonds, as will be discussed in more detail below.
We study how variations in the mechanical properties of the cells, the cell-cell adhesion complexes or force variations affect the rate of gap formation. Fig 2A and 2B show how passive mechanical properties of the cell affect both the frequency (Fig 2A) and the location of the gap openings (Fig 2B). Increasing stiffness of either the membrane or the stress fibers provokes a decrement of the gap generation frequency (Fig 2A and S4 and S5 Movies). This is intuitive, since increasing stiffness stabilizes the movements of cells and makes the monolayer less dynamic. On the other hand, the location of the gap openings (i.e. whether they occur at a vertex or border) is critically affected by membrane stiffness at low values, until it stabilizes for intermediate and high membrane stiffness. In contrast, stress fiber stiffness affects gap location for very high stiffness, where gaps are almost fully prevented from opening at the borders (Fig 2B). Interestingly, increasing bending stiffness first increases gap generation up to a maximum point, before it leads to a decrease in gap opening frequency (Fig 2A). For small to intermediate bending stiffness, the frequency of gap openings increases, since bending stiffness is critical for effective force propagation between neighboring adhesion sites at vertices. When a single adhesion complex ruptures, bending stiffness leads to increased forces on neighboring adhesion complexes. After a peak in gap opening frequency at intermediate bending stiffness, a drop in the gap formation is observed for higher bending stiffness. This is caused by the resulting stabilization of the existing gaps at vertices. This high bending stiffness opposes sharp corners of the membrane at vertices and thus favors stable gaps that are permanently open, implying no new gaps are formed (S7 Movie). On the other hand, at cell borders, a high bending stiffness implies that if a single adhesion cluster is ruptured, the forces on it are redistributed across many neighboring adhesion sites and this stabilizes the borders (Fig 2B).
Turning to the role of cell-cell adhesion complex properties, our model shows that as the junctions become more stable, gaps open less frequently (Fig 2D). To increase cell-cell junction stability, we increase the mechanical stiffness of individual adhesion bonds, or the density of adhesion molecules. These results are in line with previous experimental work [14], which reported that more stable cell-cell junctions result in fewer transmigrating cells. While the total number of gaps at either vertex or border decreases with increasing cell-cell adhesion complex stiffness or cell-cell adhesion density available for binding, we see that there are no significant differences between gaps generated at the vertex and gaps generated at the borders (Fig 2E).
Fig 2G and 2H show the impact of changing the cortical and radial forces, where the total force is kept constant (when the radial force decreases, the cortical force is increased by the same magnitude). This is biologically relevant since cells are known to shift their cytoskeletal compartments in a context dependent manner [38]. In fact, cell monolayers subjected to shear flow have been reported to increase cortical actin while decreasing stress fibers [14]. Endothelial cells in particular, are known to exhibit both radial and tangential stress fibers with a different effect on gap opening dynamics [39]. As the force shifts from radial to cortical forces, total gap formation fluctuates with a slight increase as cortical forces increase (Fig 2G). For high cortical forces, the gaps also clearly tend to localize more at the vertices (Fig 2H). This is because contractions parallel to the membrane result in force concentrations at the vertices. For very high cortex forces, the typical stresses on adhesion clusters at the vertices may thus be higher than the force where the lifetime of catch bonds peaks (Supplementary S12 Fig), explaining the small increase in the number of gaps formed (Fig 2G). On the other hand, we will later show that these gaps formed at high cortical forces are typically small and have a short lifetime, limiting their potential for extravasation (see Fig 3I and 3J).
To take into account that molecular or physical perturbations may simultaneously affect multiple parameters, we now study how variations of pairs of these parameters at the same time may influence the monolayer integrity and the localization of the gap formation. Although, we have previously seen in Fig 2A that membrane and stress fiber stiffness have a similar effect on the gap opening frequency, in Fig 2C we can observe how the effect of varying stress fiber stiffness is clearly predominant over the effect of varying membrane stiffness. Fig 2F shows the impact of varying cell-cell adhesion stiffness and cell-cell adhesion complex density available for binding. Interestingly, there is a synergy between both parameters on regulating gap opening frequency, as evident through the curved shape of the levels of equal gap opening frequency (Fig 2F). In Fig 2I we show the combined role of cortex and radial forces, thus not keeping total force fixed as in Fig 2G and 2H. This confirms that total force is the main driver of gap opening frequency, as opposed to a redistribution of forces between cortex and stress fibers (Fig 2I).
The lifetime and size of a gap are physical parameters that may also limit a cancer or immune cell’s potential to extravasate through the monolayer. Here, we show how the lifetime and size of a gap are influenced by cell mechanical and junction properties, without the presence of extravasating cells (Fig 3). We observe that membrane stiffness has a marginal influence on the life time of the gap, whereas increasing stress fiber stiffness clearly reduces the time that a gap is open and the gap size (Fig 3A and 3B). Indeed, higher stress fiber stiffness will result in mechanical resistance to an opening gap and thus inhibit the propagation of the defect in the cell-cell junctions, leading to a stabilization of the monolayer (see S4 and S5 Movies). The dominance of stress fiber stiffness over membrane stiffness in regulating lifetime and size remains valid in a broad range of parameter values (Fig 3C and 3D).
Interestingly, increasing bending stiffness to high values may increase gap lifetime (Fig 3A). This is because higher bending stiffness will resist deviations from straight membranes. Thus, at straight borders, higher bending stiffness will resist gap openings whereas at vertices with high curvature, cells are more likely to adapt their shape resisting high curvature, thus favoring opened gaps. The dynamics of the monolayer for low bending stiffness is shown in S6 Movie.
Fig 3E and 3G show that adhesion complex stiffness and density at low values do not have a big impact on lifetime, however as they increase, lifetime starts to decrease. Both stiffness and density have a similar effect, since the total stiffness of an adhesion complex depends on both density and single bond stiffness (Eq. S10). Higher stiffness of the adhesion complex leads to more passive mechanical resistance to gap openings, and this effect dominates for high stiffness. The level of noise due to repeats of our MC simulations is higher for these adhesion parameters than for the parameters determining cell mechanics. Likewise, for the gap size, the stabilizing effect of both adhesion complex stiffness and density dominates and leads to a reduction in gap size (Fig 3F and 3H). However, the effect of increasing the density is slightly stronger than that of increasing single bond stiffness. This is because the density affects not only adhesion complex stiffness (Eq. S10), but also the rate of forming new adhesion complexes (Eq. S9) and the rate of reinforcing existing bonds (Eq. S11). These effects together thus synergize to stabilize gaps and prevent them from growing too large.
Earlier, we have shown that a shift in the force (from radial to cortical) produces an increment in gap formation (Fig 2G). Fig 3I and 3J show that this shift in the force reduces gap lifetime and size. This indicates that, although the frequency of opening is increased, these gaps are smaller and last shorter in time which may reduce paracellular extravasation, as suggested in previous experimental work [14]. Combined changes of cortical and radial force show that although both kinds of forces are needed to increases gap size and lifetime, the impact of radial forces is clearly predominant over the impact of cortex forces (Fig 3K and 3L). This is intuitive, since radial forces clearly separate cell borders generating bigger gaps and make them harder to close, whereas cortical forces distribute forces to vertex regions. This does not provoke large cell deformations, which is reflected in the low impact on the gap size and lifetime observed.
In Fig 4A–4C we show the impact of varying the catch-bond unbinding parameter k c 0 that shifts the location of the peak of maximal lifetime of a single catch bond, while we maintain the actual maximum value through simultaneously shifting the slip-bond unbinding parameter k s 0 (Eq. S12 and S12 Fig). We observe that for a pure slip bond (corresponding to k c 0 = 0), gaps occur at a higher rates than for small nonzero values of k c 0. Increasing k c 0 further leads to a minimum in gap opening frequency, from which the frequency increases again. This minimum corresponds to a maximum of stability, where forces on the adhesion complexes are similar in magnitude to the peak of stability of the catch bond. Consequently, shifting the location of that peak even further towards higher forces (by increasing k c 0 even further) means we destabilize the catch bonds again. Note that the gap lifetime and size of gaps are much less influenced by the location of the catch bond maximum than the gap opening frequency.
In Supplementary S11 Fig, we show histograms of the forces on adhesions comparing the number of bound clutches, the number of unbinding events, and the ratio of unbound to total bonds for slip bonds (k c 0 = 0) to the catch bond with reference values (k c 0 = 0 . 27 s - 1). We see that adhesions in the catch bond case bear and disengage at higher forces than for the slip bond case, confirming that the typical forces on bonds are of such magnitude that the catch bond nature stabilizes the junctions.
In Fig 4D–4F we modify the reinforcement binding rate k r e i n f 0 to check the influence of the reinforcement. This is different from the previous analysis where the adhesion complex density available for binding was changed, since now the binding probability based on distance is not affected (Eq. S9). However, we see the same trend of increasing stability with increasing k r e i n f 0 (Fig 4D), in line with the result obtained from varying cadherin density (Fig 2D), suggesting that binding is mainly regulated by this reinforcement process. Similar to the catch bond, we see that adhesion reinforcement is less important in determining gap size or lifetime (Fig 4E and 4F) than in regulating gap opening frequency.
We have shown that both the magnitude of forces and the cytoskeletal compartment that generates the forces (stress fibers or cortex) affect gap opening frequency, size and and lifetime. Besides these broad compartments, many other biological and physical parameters affect how forces ultimately act on cell-cell junctions: Forces may act in a directed manner due to larger parallel actin bundles and synchronous myosin activation, e.g. initiated through waves of activators [15], or may act more randomly [33]. We test variations in force applications through parameters that affect the transition time when forces change (t T r a n s i t i o n F o r c e), through spatial force distributions and through the velocity at which forces are modified. In Fig 4G we observe how increasing the force transition time t T r a n s i t i o n F o r c e slowly reduces the gap opening. This is due to the fact that a slower, persistent application of forces leads to a redistribution of the forces through rearrangement and remodeling of the cell. It is consistent with experimental works that showed that force fluctuations influence gap opening dynamics [15].
Then, distributing the same radial forces over several adjacent stress fibers reduces gap opening frequency (Fig 4H). More spatially distributed forces are less capable of damaging cell-cell junctions than localized peak forces, since such high peak forces are required to overcome the catch bond maximal lifetime. Likewise, high peak forces lead to longer lifetime and larger size of the resulting gaps (Supplementary S13C and S13D Fig).
Next we observe the effect of force persistence in time. We vary the force recalculation time parameter (equally for all forces) in Fig 4I. Results show that the time that forces are applied does not have a big influence on gap formation. This suggests that cells are able to adapt to forces in longer time scales and therefore it is not the time that forces are applied what regulates gap formation, but the transitions of force fluctuations and their spatial distribution.
To demonstrate that the geometry of the gap opening dynamics is physiologically relevant, we quantified the characteristics of extravasating cancer cells through monolayers of HUVECs, as shown in S9 Movie. Here, a tumor cell is seen transmigrating through an endothelial monolayer at a tricellular junction as delineated by VE-cadherin GFP, followed by gap-closure after the tumor cell has completely cleared the barrier. Fig 5A shows the ratio of tumor cells that extravasated at vertices, relative to borders. We see that tumor cells preferentially extravasate at the vertices, in line with the previously observed increased frequency of gaps opening there (Fig 1G) and similar observations of extravasating neutrophils [32]. Moreover, even if cancer cells initially arrest at the border between two endothelial cells, they are much more likely to extravasate at a vertex at later points in time, rather than at the border where they initially attached to, perhaps first through migration on the surface of the endothelium and subsequent preferential attachment to points of exposed basement membrane as a result of inherent EC junctional dynamics (Fig 5B). This could suggest that in addition to the possibility of cancer cells actively signaling to open gaps in the endothelium, endothelial barrier dynamics itself can also present the cancer cells with opportunities to begin the transmigration process.
The computational model presented in this paper allowed us to study how gaps in an endothelial monolayer initially open, grow, stabilize and finally close, and we identified which physical properties dominantly regulate each stage.
The model simulates a cell monolayer in two dimensions. Adhesion between cells is simulated through binding or unbinding of adhesion complexes located on adjacent cells. These adhesion complexes are dynamically engaging and disengaging as the myosin generated forces cause cell deformations. Because of cell-cell adhesion rupture, gaps between the cells are formed. To perform our simulations, the model is based on a number of assumptions that simplified the model. First of all, due to the typically small height (about 3μm) of endothelial cells [40], we neglected the third dimension perpendicular to the monolayer. However, disruptions of the spatio-temporal dynamics of adhesion molecules and cytoskeletal organization in the third dimensions are likely to impact gap formation. Incorporating such effects into our model would consequently require a 3D model of a cell with more detailed descriptions of the subcellular mechanics. However, the purpose of our model was to demonstrate the broad impact of subcellular mechanical structures on gap formation. For this reason, we modeled the cells in two dimensions and included only radial stress fibers and contractile actin fibers parallel to the membrane. This was motivated by experiments that indicated different roles of these actin structures on gap formation [23]. Each discrete adhesion complex is simulated as one cluster to simulate recruitment of proteins such as vinculin or talin, without increasing the total number of components in the simulation. Myosin generated forces included in the model are assumed to occur only in the direction of the stress fibers or membrane.
Supplementary S14 Fig summarizes some of our key conclusions: By comparing our reference case with extreme variations of very low stress fiber or membrane stiffness, we see that both passive mechanical properties and adhesion complex properties are important in controlling gap opening frequency (Supplementary S14A Fig). On the other hand, the lifetime and especially size of gaps increases significantly with lower stress fiber or membrane stiffness, since the softer cells are more likely to deform and adapt in response to the opened gap (Supplementary S14B and S14C Fig). We also verify that stress fiber stiffness influence is stronger than membrane stiffness influence. In contrast, properties of the cell-cell adhesions strongly affect the frequency of the gap openings, but less so their lifetime or size. Indeed, decreasing the density of adhesion bonds or the adhesion stiffness strongly increase the frequency of forming gaps (Supplementary S14A Fig), while only marginally affecting the size or lifetime of the gaps (Supplementary S14B and S14C Fig). These data thus summarizes our biological model where adhesion properties control the initial formation of gaps, while cell mechanical properties are critical in limiting the size and duration of opened gaps.
Our results that gaps open more frequently at vertices than at borders were true over wide ranges of parameters (Fig 2B, 2E and 2H). Only extremely small bending stiffness led to similar frequencies of gaps at vertices and at borders (Fig 2B). These results also show that earlier experimental observations, where neutrophils were found to extravasate preferentially at endothelial cell vertices, [32], can be explained through the mechanical dynamics of the endothelial monolayer alone. Consequently, this may be a general mechanism for extravasating cells, and we found a similar behavior with extravasating cancer cells (Fig 5). This finding is complementary to the extensive literature that suggests that chemical or mechanical signaling of extravasating immune or cancer cells to the endothelium facilitates extravasation [1, 8, 31]. There are many potential hypotheses why both the autonomous dynamics of the endothelial monolayer and the bidirectional signaling with the extravasating cells may play a role during extravasation: It may be that initial autonomously forming gaps are important for extravasating cells to sense a gap and they consequently signal to widen the gap or to keep it open. The gap sizes that the model predicts are of the order of magnitude of a few microns, which is enough for extravasating cells to protrude through the gap. Our previous study indicates that tumor cells can squeeze significantly when transmigrating through artificial gaps [16], so the autonomous gaps may be of sufficient size for complete transmigration. Nevertheless, endothelial gaps may widen during transmigration, so crosstalk between the transmigrating cell and the endothelium likely remains an important factor contributing to the likelihood and speed of extravasation. Then, whether bidirectional signaling or autonomous gap formation dominates the process may be cell type specific. For instance, it is still a major research question why certain cancer cells preferentially metastasize to certain organs [41]. We may speculate that not only the signaling of the specific primary tumor cells with an organ-specific type of endothelial cells influences the likelihood of extravasation [8]. Also, the mechanical properties of the endothelium of the target organs will likely play a major role. Our flexible modeling framework was tested with a HUVECs monolayer, yet, by changing the physical parameters of the model, it may be quickly adapted to other endothelia.
Besides testing our model with different endothelial cells, some other important steps towards validating our model conclusions in vivo will be to test our model with more realistic three dimensional microvasculature with blood flow, embedded in extracellular matrix and surrounded by supporting cells such as pericytes, fibroblasts or, for brain, astrocytes [41, 1]. Such real, in vivo microvasculature consists of vessels that are curved and exposed to shear stresses due to the flow. That, in turn, may be affected by extravasating cells that may obstruct blood flow. Similarly, matrix stiffness was shown to affect endothelial monolayer integrity [42, 43]. Some complications in validating our results in vivo involve the lack of available in vitro cultures that are required to provide high throughput, microscopy resolution and level of experimental control that is lacking in vivo, making direct comparison of computational models to in vivo experiments unfeasible. However, the recent rapid progress in developing more complex and organ specific in vitro assays of 3D microvasculature will make such validations feasible in the near future [44, 45, 46].
Our model is based on a number of simplifications. We do not consider the effect of extracellular matrix and substrate stiffness properties on monolayer integrity, despite the known effect of these properties on cell mechanics. It is important to remark that cells on glass may behave very differently than in vivo endothelial vessels. Our model also does not include the effect of fluid pressure or tangential stress due to fluid flow. Pressure and blood flow would induce additional forces over the monolayer that could affect gap generation processes. For example, it was observed [14] that tangential flow could induce the strengthening of cell-cell junctions, therefore reducing paracellular extravasation.
Modeling such complex environments presents a great challenge to both in vitro and in silico models. It is therefore essential to justify assumptions that can reduce this complexity and make the model development feasible. Here, we have assumed that the mechanics of the inside of a cell is determined by a fixed number of stress fibers, although it is known that inside the cell there are different polymer structures such as microtubules and intermediate filaments. Moreover, actin filaments are not fixed in time but appear and disappear depending on their stability and polymerization rates. To simulate all of this with high accuracy would require a completely different model in which the computational cost that would exceed current capabilities. For the purpose of this project, we focused on incorporating essential cell mechanical structures that have been implicated in the regulation of gap formation, and modeled a fixed number of stress fiber similar to other works [28, 29]. Similarly, we have simulated adhesion complexes as discrete elements that can bind two membrane points of neighboring cells. In real cells, adhesion complexes between cells are formed by a great variety of proteins such as VE-cadherins, α-catenin, talin or vinculin. While the spatio-temporal dynamics of each of these adhesion molecules likely influences gap formation, no computational model can currently explain their precise organization in adhesion complexes and their resulting effect on gap formation. Consequently, our model included an effective term that describes the force dependent recruitment of adhesions, as observed in different experimental studies [23, 20].
Moreover, there are also challenges to the mathematical modeling of complex 3D microvasculature. Modeling of epithelial sheets in 3D has proved challenging, with some recent interesting progress after decades of mainly focusing on epithelial monolayers in 2D [47, 48, 49]. These models are based on frameworks such as vertex models, where the dynamics of each cell is incorporated into the dynamics of vertices between cells. There are many other modeling frameworks that can capture different aspects of the complex cell behavior, such as cell based models [50], immersed boundary models [51] or subcellular element models [52, 53]. These modeling frameworks are, however, not directly suitable to predict the formation of gaps at either vertices or borders. Given these challenges, is was paramount to establish a 2D mathematical model of an endothelial monolayer that was validated with novel experiments and that was able to lead to insights into the mechanisms of endothelial gap formation.
Human umbilical chord vein cells (HUVECs) were transduced with VE-cadherin-GFP using methods described previously [45]. HUVECs at P7-10 were seeded onto 35 mm glass bottom Mattek dishes (at 3 × 105 cells/dish), which had been plasma treated for 30 seconds previously. Cells were allowed to grow to confluence (beyond 100%) in EGM-2MV (Lonza) for 3 days before imaging. Dishes were imaged on an Olympus FV1000 confocal microscope with magnifications of 30X (oil immersion), under an environmental chamber set at 37C and 5% CO2. The chamber was equilibrated for ∼ 30 min prior to the start of image acquisition. For time-lapse videos of junctional dynamics, z-stacks of 40μm (4μm steps) were taken at intervals of 3 minutes.
Time-lapse images were appended and analyzed manually on ImageJ. A single unique junctional disruption is defined as a vertex or border with an observed gap of greater or equal than 3μm, and are preceded and proceeded at some point in time with a closure (no visible gap in fluorescence greater than 0.6μm). The number of junctional disruption events was counted for each border and vertex of an image over a total time period of 2 hours. Vertices and borders belonging to the same cells were still considered to be unique.
Tumor cells were suspended in EGM-2MV (Lonza) and a concentration of 15,000 cells/mL, and 1mL of the suspension was gently added to each HUVEC monolayer. Cells were allowed to settle first for ∼10 minutes before acquisition of t = 0 images. For quantification of extravasation, z-stacks were taken at 3μm steps at an endpoint of 6 hours to image the entirety of the tumor cell and endothelial monolayer. Any tumor cell that has breached the endothelial layer as evidenced by protrusion extension across and beneath the endothelial layer was considered as “extravasated”. Delineation of the endothelial barrier is visualized via CD31 staining (Biolegend, Cat # 303103) for 30 min in EGM-2MV at 37C and 5% CO2 prior to imaging.
|
10.1371/journal.ppat.1006663 | Effects of glutamate and ivermectin on single glutamate-gated chloride channels of the parasitic nematode H. contortus | Ivermectin (IVM) is a widely-used anthelmintic that works by binding to and activating glutamate-gated chloride channel receptors (GluClRs) in nematodes. The resulting chloride flux inhibits the pharyngeal muscle cells and motor neurons of nematodes, causing death by paralysis or starvation. IVM resistance is an emerging problem in many pest species, necessitating the development of novel drugs. However, drug optimisation requires a quantitative understanding of GluClR activation and modulation mechanisms. Here we investigated the biophysical properties of homomeric α (avr-14b) GluClRs from the parasitic nematode, H. contortus, in the presence of glutamate and IVM. The receptor proved to be highly responsive to low nanomolar concentrations of both compounds. Analysis of single receptor activations demonstrated that the GluClR oscillates between multiple functional states upon the binding of either ligand. The G36’A mutation in the third transmembrane domain, which was previously thought to hinder access of IVM to its binding site, was found to decrease the duration of active periods and increase receptor desensitisation. On an ensemble macropatch level the mutation gave rise to enhanced current decay and desensitisation rates. Because these responses were common to both glutamate and IVM, and were observed under conditions where agonist binding sites were likely saturated, we infer that G36’A affects the intrinsic properties of the receptor with no specific effect on IVM binding mechanisms. These unexpected results provide new insights into the activation and modulatory mechanisms of the H. contortus GluClRs and provide a mechanistic framework upon which the actions of drugs can be reliably interpreted.
| IVM is a gold standard anti-parasitic drug that is used extensively to control invertebrate parasites pest species. The drug targets the glutamate-gated chloride channel receptor (GluClR) found on neurons and muscle cells of these organisms, causing paralysis and death. However, IVM resistance is becoming a serious problem in human and animal health, as well as human food production. We provide the first comprehensive investigation of the functional properties of the GluClR of H. contortus, which is a major parasite in grazing animals, such as sheep and goats. We compared glutamate and IVM induced activity of the wild-type and a mutant GluClR, G36’A, that markedly reduces IVM sensitivity in wild populations of pests. Our data demonstrate that the mutation reduces IVM sensitivity by altering the functional properties of the GluClR rather than specifically affecting the binding of IVM, even though the mutation occurs at the IVM binding site. This study provides a mechanistic framework upon which the actions of new candidate anthelmintic drugs can be interpreted.
| Pentameric ligand gated ion channels (pLGICs) are membrane-bound receptors that facilitate the diffusion of ions across cell membranes in response to the binding of agonists. The glutamate-gated chloride channel receptor (GluClR), first identified in arthropods, such as insects and crustaceans [1–3], is an anion-selective pLGIC found at neuronal and neuromuscular inhibitory synapses [4]. GluClRs are also present in other major metazoan phyla, including platyhelminths and nematodes [4], but have not yet been identified in vertebrates. GluClRs can exist as homo- or hetero-pentamers [5]. Crystal structures of the homomeric GluClR from the nematode, C. elegans, have recently been determined in ligand-bound [6] and apo [7] states.
The mechanism of agonist activation has been studied extensively in vertebrate pLGIC members, such as the glycine (GlyR) [8], acetylcholine (AChR) [9–11], serotonin (5-HT3R) [12] and GABAA (GABAAR) [13–15] receptors, as well as ELIC, which is a bacterial pLGIC [16]. A detailed study of the biophysical properties of GluClR activation has not been undertaken, even though GluClRs constitute a major group of pLGICs, many organisms that express them are serious parasitic pests, or vectors for disease transmission and they are a major target for anthelminthic drugs. For instance, O. volvulus and W. bancrofti are nematodes that cause river blindness (onchocerciasis) and elephantiasis (lymphatic filariasis), respectively, in humans. Another nematode, H. contortus [17] is a serious pathogen in ruminant agricultural animals such as cattle, sheep and goats. The sea lice (arthropod) species C. rogercresseyi [18] and L. salmonis [19], ravage salmon and trout farming industries worldwide. The cereal cyst nematode H. avenae devastates broad acre cereal crops across temperate wheat-producing regions of the world [20, 21]. A. gambiae is the mosquito vector that transmits malaria in over 90% of world-wide cases [22]. Finally, the flatworm blood fluke, S. mansoni, inflicts schistosomiasis (associated with serious systemic morbidities) on hundreds of millions of people in underdeveloped communities [23].
Macrocyclic lactones (MLs) such as ivermectin (IVM), moxidectin, abamectin and emamectin are widely used to control all of these, as well as many other, nematode and arthropod pests [24]. IVM works by activating GluClRs in pharyngeal muscle cells and motor neurons of these organisms, thereby causing death by flaccid paralysis or starvation [25]. Unfortunately, however, IVM resistance is emerging as a serious problem in many pest species [21, 26–29] prompting the need for new generation treatments.
Functional and crystallographic studies have recently delineated the binding pocket of IVM and identified potential residues that IVM interacts with [6, 30, 31]. The main structure of the pocket is formed by first (TM1) and third (TM3) transmembrane domains of adjacent receptor subunits, at the level of the upper leaflet of the cell membrane [6]. Site-directed mutagenesis of transmembrane domains has identified critical residues that drastically affect IVM potency in the avr-14b subunit of H. contortus, such as TM3-G36’ [30] and TM1-P230 [31], and in the α subunit of C. elegans, such as TM1-L279 and TM1-F276 [32, 33]. The glutamate binding site and TM3 domain are also sites that harbour mutations in wild ML-resistant strains of C. elegans [27], whereas ML resistance in wild isolates of pest species have been attributed to mutations at, TM3-30’ in P. xylostella [34] and TM3-36’ in T. urticae [35, 36]. Of particular note, a Gly at the 36’ position is thought to be essential for exquisite IVM [30] and abamectin [34] sensitivity, and larger substitutions at this location were proposed to reduce ML sensitivity by hindering access to its binding site [31, 34, 37]. However, the effects of these mutations are generally evaluated using functional assays that lack the resolution needed to distinguish discrete functional states in the activation process. A detailed mechanistic understanding of how wild-type and mutated receptors respond to glutamate is a prerequisite to understanding how IVM and other modulating ligands affect the receptor. This aim is best achieved through the study of single channel currents mediated by individual receptors [38]. Without a quantitative understanding of activation and modulation mechanisms of the receptor, attempts to design drugs with higher potency and selectivity for the receptor would be intractable.
Four GluClR subunits have been identified in H. contortus [α3A (avr-14a), α3B (avr-14b), β and α], all of which express on motor neuron commissures [39, 40]. However, the native stoichiometric combinations of these subunits is unknown [4]. Here we investigated homomeric receptors comprising the avr-14b subunit, which is also expressed in pharyngeal neurons [39, 40]. We will refer to this subunit as α (avr-14b). In heterologous expression systems, homomeric receptors comprising either α (avr-14b) [30, 41] or α subunits form high affinity IVM binding sites, whereas the β subunit homomers do not [42].
In this study we investigated the biophysical properties of the homomeric α (avr-14b) GluClR from H. contortus as: 1) H. contortus is a major parasitic pest of domestic ruminant animals, 2) IVM is used widely to control H. contortus, 3) IVM resistance has emerged as a major problem in this species [43], and 4) GluClRs comprising or containing the α (avr-14b) subunit are most likely the major biological IVM target in this species [4, 25]. Here we sought to quantify the activation properties of this receptor in the presence of glutamate and IVM, and to explore the mechanism by which the TM3-G36’A mutation reduces IVM sensitivity to a level that is similar to vertebrate GlyRs and GABAARs [30, 44–46].
Single receptor currents (Fig 1A) were measured and plotted as a function of voltage (Fig 1B) to determine the single channel conductance of the receptor. Using Eq 1 and a mean current amplitude of 1.80 ± 0.03 pA (n = 7, at ‒70 mV), the estimated single channel conductance of the homomeric GluClR was 22.9 ± 0.3 pS. The i-V was nearly linear (Fig 1B). The slight inward rectification and relatively small conductance of the homomeric GluClR is very similar to that determined for ternary GABAARs containing, α, β and γ subunits [13, 14]. A recent study has also estimated the current amplitude of the heteromeric GluClR of C. elegans at ‒ 90 mV to be ~ 1.9 pA [32].
Single channel activity was recorded in the presence of a broad range of L-glutamate concentrations (10 mM– 5 nM) to determine the receptor’s sensitivity to glutamate, the active durations of single receptors, the total time spent in conducting configurations (PO) and the shut and open dwell characteristics within each active period. Continuous sweeps of single channel activity, recorded from a patch in the presence of 200 μM glutamate is shown in Fig 1C. At this and higher concentrations the activity of single receptors occurred as clearly defined periods of variable duration, termed ‘activations’, where the receptor oscillated between conducting and non-conducting configurations. These active periods were interrupted by relatively long intervals of inactivity where the receptor adopted desensitised states. These states are distinct from ligand-bound shut states both structurally [47, 48] and functionally [49]. With few exceptions, desensitised states are much longer-lived than shut states. Mean dwell times of the shut and open durations within activations were generated by plotting histograms and fitting these to mixtures of exponentials (Fig 1D). The shut dwell data were best described by two exponential components, whereas the open dwell histogram was best fit to three exponential components. The dwell time constants were similar when the receptors were exposed to 10 mM and 1 mM glutamate, but at 200 μM the time constant of the longer shut component increased (S1 Table).
Reducing the glutamate concentration to 30 μM resulted in similar single channel activity (Fig 1E). The number of components and the time constants of both dwell histograms (Fig 1F), appeared little changed, except for a further increase in the time constant of the longer shut component (S1 Table). In addition, the mean duration of the active periods appeared to become shorter as glutamate concentration decreased (S2 Table).
At low glutamate concentrations there appeared to be a transition from mostly tightly grouped to loosely grouped periods of activity and isolated open-shut events. For example, 2 μM glutamate elicited activity that comprised a mixture of isolated open-shut events and activations consisting of openings and shuttings in rapid succession, as with the higher glutamate concentrations. However, these latter more complex activations were more likely to occur in shorter bursts (Fig 2A). The dwell histograms also exhibited two shut and three open components with similar time constants to those for the higher concentrations of glutamate, but the time constant of the longer shut component continued to increase and the fraction of the longest open time constant diminished (Fig 2B, S1 Table). 30 nM glutamate elicited activations that occurred as bursts of loosely spaced openings and brief open-shut events (Fig 2C). Moreover, long stretches of record corresponding to receptor desensitisation were mostly absent. The dwell histograms revealed changes to both shut and open components. Here both shut components increased and the longest open component disappeared (Fig 2D, S1 Table).
As 30 nM glutamate was effective at eliciting single channel activity the concentration was lowered even further, to 5 nM. Remarkably, even at this concentration GluClRs were activated. Most of the activity comprised simple shut-open events, but the occasional activation of loosely spaced openings was also apparent. In contrast, in the absence of glutamate, receptor openings were extremely rare, brief and essentially negligible (Fig 3A). These data show that 1) the homomeric GluClRs are exquisitely sensitive to glutamate and 2) from 2 μM glutamate and below, the activity becomes increasingly simple and brief, likely reflecting an effect consistent with agonist dissociation from partially liganded receptors. The dwell histograms derived from 30 nM and 5 nM glutamate showed distinct differences compared to those of higher concentrations. Here both the longer and briefer shut components increased (Fig 3B) and the third, longest open component was absent, whereas the remaining two open components decreased (Fig 3C, S1 Table). The invariant open dwell components for concentrations ≥ 2 μM glutamate are consistent with an optimal degree of ligation for receptor activation, as has been shown for the GlyR [8]. In contrast, the decrease in the remaining two open component time constants at nanomolar concentrations of glutamate is consistent with sub-optimal activation of receptors. We infer that at 30 nM and 5 nM each receptor is bound to fewer ligand molecules than at the higher concentrations, giving rise to openings with briefer lifetimes [50]. We also infer that the steadily increasing longer shut component at ≤ 200 μM glutamate is additional evidence that the receptors are able to activate without all glutamate binding sites being occupied [8].
The total time spent in open states was also compared across glutamate concentrations. This analysis demonstrates that PO increases as a function of glutamate concentration (Fig 3D). Consistent with the inference that the receptors are highly sensitive to glutamate, the PO at glutamate concentrations ≥ 10 μM were all > 0.90. PO showed a significant decrease at 2 μM glutamate and dropped to 0.21 at 30 nM and 0.14 at 5 nM (Fig 3D, S2 Table). A Hill fit to the PO plot revealed a maximum of 0.99, an EC50 of 70 nM and a Hill coefficient of 0.82. The mean duration of activations was also plotted and showed that active durations declined from ~500 ms to ~330 ms between 10 mM and 30 μM glutamate. Fitting the data to a Hill equation produced an EC50 of 31 μM, a Hill coefficient of 0.56, a maximum duration of 500 ms and a minimum of ~80 ms (Fig 3E).
The activation properties of homomeric GluClRs were also investigated at an ensemble current level using rapid solution exchange [14, 51] of glutamate onto macropatches. As GluClRs are located at inhibitory synapses, these experiments were carried out to mimic synaptic activation conditions by determining the response of many (~20–100) receptors to rapid glutamate exposure. By avoiding the distorting effects of receptor desensitisation encountered with slower agonist application methods, rapid solution exchange techniques also establish a more accurate ligand concentration ‒ peak current relationship. Peak current was achieved by rapidly applying glutamate for either 50 ms (5 mM– 20 μM) or 500 ms (10 μM– 0.5 μM, Fig 4A). These data were fitted to a Hill equation, yielding an EC50 for glutamate of 43 μM and a Hill slope of 0.8 (Fig 4B). Whole-cell experiments on the same GluClR produced an EC50 for glutamate of ~15 μM and a Hill slope of ~1.7 [30]. The 2–3 fold difference in EC50 and Hill slope between whole-cell and macropatch data are consistent with open and desensitised states, which have a higher affinity for ligand, having made a significant contribution to the whole-cell data.
Similar experiments were carried out to determine the relationship between the activation rate of the current and agonist concentration. Normalised examples of these recordings are illustrated in Fig 4C and the group data are summarised in Fig 4D. A Hill fit to this plot produced an EC50 of 0.95 mM and a Hill slope of 1.0. The upper asymptote of the activation plot was ~9000 s−1, representing the maximum activation rate [51, 52], whereas the lower level was ~10 s−1.
Homomeric GluClRs containing the G36’A mutation exhibit a markedly reduced IVM sensitivity (EC50) when recorded in whole-cell configuration [30]. However, any changes to the intrinsic properties of the receptor conferred by the mutation have yet to be examined in mechanistic detail.
G36’A-containing receptors were first examined on a single channel level. Applied glutamate elicited a similar current amplitude to wild-type receptors, suggesting the G36’A had no appreciable effect on channel conductance (Fig 5A). A current amplitude of 1.81 ± 0.02 pA (n = 7) for the mutant at ‒70 mV yielded a conductance of 23.0 ± 0.2 pF, if it is assumed that under the same recording conditions the reversal potential is similar to that for wild-type.
However, moderate to high (30 μM– 10 mM) concentrations of glutamate revealed two distinct types of activations in the mutant receptor (Fig 5A), whereas the same concentration of glutamate elicited homogeneous activations in the wild-type receptor (Fig 5B). The two activation modes mediated by the mutant GluClR were quantified on the basis of PO and duration only for 1 mM glutamate. The analysis revealed a very low PO activation mode of 0.14 ± 0.02 (n = 6) and mean active periods of 1333 ± 222 ms duration and a higher, more wild-type like mode with a PO of 0.87 ± 0.05 and a mean active duration of 309 ± 62 ms (Fig 5C). In contrast, 1 mM glutamate produced a single PO of 0.99 (S2 Table) in the wild-type receptor consistent with fewer shuttings within each activation (Fig 5D). The two activation modes observed in the mutant receptors were pooled for further analysis for all concentrations where they were apparent so as to determine the net effect of the mutation on PO and active durations, and facilitate a more direct comparison to wild-type receptors. The dwell histograms for the mutant receptor at 1 mM glutamate required two shut and three open components (Fig 5E), but the longer shut component was substantially increased compared to wild-type receptors (Fig 5F) and the two longest open components were reduced (S1 Table).
Two distinct gating modes were also observed at a moderate (30 μM) glutamate concentration (Fig 6A), but were difficult to distinguish at a low (2 μM) concentration because the activations became too brief and simple (Fig 6B). Over the concentration range tested, the mean duration of activations of the mutant receptor were considerably shorter than wild-type with a maximum mean duration of 200 ms, as was the mean PO, which peaked at 0.73 (Fig 6C, S2 Table). Fig 6D summarises the dwell component data over the glutamate concentrations that were tested on the mutant receptors. Consistent differences to wild-type receptors include an increase in the long shut component and briefer open components (Fig 6E). At 2 μM glutamate only one shut component and two open components were resolvable (Fig 6F, S1 Table).
The briefer active periods exhibited by the G36’A mutant receptors is indicative of accelerated ensemble current deactivation [14, 53] and desensitisation [49]. To investigate whether receptor desensitisation was affected by the G36’A mutation, the long quiescent periods corresponding to desensitisation in single channel records were quantified, then corrected for channel number [49]. Sample recordings for wild-type and mutant receptors are shown in Fig 7A and 7B, respectively. A saturating concentration of glutamate (10 or 1 mM) was first rapidly applied onto each patch, ensuring that all the receptors in the patch were activated, after which constant agonist perfusion was maintained over the patch for the remainder of the recording. Clearly defined steps corresponding to the single channel amplitude (~2 pA) became apparent as all the receptors desensitised back to baseline. The number of steps was then taken as an estimate of the total number of receptors contained in each recorded patch. Only patches expressing 1–10 steps (channels) were accepted for analysis.
The long desensitised periods were estimated by plotting shut dwell histograms for the entire record, as is illustrated in Fig 7C and 7D. The shut events could be divided into two broad components. The briefer component corresponded to shut events within active periods. This component could be subdivided into briefer components (as in Figs 1, 2, 4 and 5). The longer component represented the mean desensitised lifetime, and it was this component that was corrected for channel number. This method of analysis produced a mean desensitised lifetime for wild-type receptors of 91 s (n = 5), and was used to determine a re-sensitisation transition rate constant (ω, Fig 7E) of 0.011 s‒1. Similarly, the desensitisation rate constant (δ) was estimated from the mean duration of the active periods at saturating glutamate concentrations (500 ms, Fig 3B) to be 2.00 s‒1 [49]. This produced an equilibrium constant (δ/ω) for desensitisation of 182 for wild-type receptors. A similar analysis for the G36’A mutant receptor produced a mean desensitised lifetime of 125 s (n = 8) and an ω of 0.008 s‒1. However, a more significant change was estimated for the mean duration of active periods for the mutant, which saturated at 200 ms (S2 Table), producing a δ value of 5.00 s‒1 and an equilibrium constant of 625. Thus, mutant receptors desensitised ~3.4 times more rapidly than wild-type receptors. From this analysis it can be inferred that the G36’A mutation increases the likelihood of the receptors entering desensitised states.
To determine if the estimates of receptor desensitisation reflected current decay and desensitisation in ensemble currents, macropatches expressing wild-type or G36’A mutant GluClRs were exposed to a saturating (3 mM) concentration of glutamate for either ~1 ms or 500 ms. In response to a ~1 ms application, the deactivation phase of macropatch currents mediated by wild-type GluClRs was adequately described by two standard exponential functions with a weighted time constant of 67 ± 4 ms (Fig 8A and 8B). The individual time constants (and fractions) are tabulated in Table 1. To allow comparison with other, better characterised pLGICs, heteromeric α1β GlyRs and α5β3γ2L GABAARs were tested under similar conditions. A ~1 ms pulse of 3 mM glycine applied to α1β GlyRs elicited macropatch currents that also exhibited a two component decay phase with a weighted mean of time constant of 22 ± 3 ms (Fig 8A). In contrast, a ~1 ms pulse of 3 mM GABA applied to α5β3γ2L GABAARs activated macropatch currents that decayed considerably more slowly than those of GluClRs. They also deactivated with two components, with a weighted time constant of 275 ± 2 ms (Fig 8A and 8B, Table 1). 3 mM glutamate was also rapidly applied for ~1 ms onto patches expressing the G36’A mutant GluClR (Fig 8C). The weighted time constant from a two component fit was 11 ± 1 ms, which was 2-fold faster than those of the α1β GlyR and over 6-fold faster than the wild-type GluClR (Fig 8B).
The activation phase of the currents was also measured by fitting 10–100% of the rising phase of the current to Eq 2. The measurements, summarised in Fig 8D, demonstrate that currents mediated by all receptors tested activate with similar time constants, which ranged between 0.1 ‒ 0.2 ms. In contrast to the deactivation kinetics, the G36’A mutation had no significant effect on the ability of the receptor to activate upon exposure to glutamate.
Ensemble desensitisation was examined by rapidly applying 3 mM glutamate onto macropatches for a duration of 500 ms (Fig 8E). Wild-type GluClRs desensitized with single time constant of 492 ± 38 ms, whereas the G36’A mutant receptor required two exponential functions to adequately describe the desensitisation phase of the current (Table 1). The weighted desensitisation time constant for the mutant receptor was 252 ± 26 ms (Fig 8F). We infer that the number of components that were needed to describe single receptor and ensemble desensitisation is related to modal activation in the mutant receptor. Consistent with this inference, estimates of the mean active durations for both receptors at saturating glutamate match very closely with the time constants of ensemble desensitisation (Tables 1 and S2). Overall, these data demonstrate that the G36’A mutation abbreviates single channel active periods, which manifest as accelerated deactivation and desensitisation in ensemble currents. These alterations to the intrinsic activation properties of the receptor are likely the underlying reasons for the order of magnitude rightward shift in the whole-cell concentration-response relationship for glutamate, reducing its sensitivity (EC50) from 15 μM to 154 μM. However, studies have also revealed a parallel shift in IVM sensitivity (EC50), from 40 nM to 1.2 μM in the G36’A-containing receptor [30, 45].
IVM is both a direct agonist and a potentiator of glutamate responses at the GluClR. In our final set of experiments we wished to test the hypothesis that the changes to the functional properties of the receptors conferred by the G36’A substitution gives rise to the reduced sensitivity to IVM, as it does for glutamate. To test this idea, we recorded single channel currents in the presence of 5 nM IVM alone (direct activation) or in 5 nM IVM + 2 μM glutamate (potentiation). In both experiment types the receptors opened to an amplitude of 1.8 pA (e.g., Fig 9A), suggesting that the presence of IVM had little effect on the permeation pathway.
Wild-type receptors exhibited a substantial degree of potentiation and direct activation by IVM. However, the recordings also revealed that these experiments were not ‘steady state’. We confined our analysis the steady-state phase of both experiments types (direct activation and potentiation). When membrane patches expressing wild-type receptors were exposed to 5 nM IVM alone, no receptor activity was apparent for the first 41 ± 4 ms. After this initial silent period the activations were initially well separated, but increased in duration for 47 ± 25 s, after which the active durations reached a steady-state equilibrium of almost continuous activity (Fig 9A) of all the receptors present in each patch (between 1–4 receptors). The mean active duration of the receptors at steady-state was 9.5 ± 2.6 s and had a PO of 0.65 ± 0.07 (Table 2). The shut intervals were best described by three components whereas the open interval histograms required four exponential components for fitting (S3 Table). The presence of additional shut and open components suggests that IVM alone induces activity of greater complexity or exposes state lifetimes that are not easily resolvable when glutamate is present. Receptor desensitisation by IVM alone had a mean lifetime of 536 ± 140 ms (ω = 1.87 s‒1). A mean active duration of 9.5 ± 2.6 s (δ = 0.105 s‒1) yielded an equilibrium constant of 0.06.
Direct activation of G36’A mutated receptors by 5 nM IVM produced a similar lag time before equilibrium was reached (Fig 9C). At equilibrium the receptors were active for a mean duration of 46 ± 8 ms and a PO of 0.85 ± 0.01 (Fig 9D, Table 2). These active periods were much briefer than wild-type receptors when activated by IVM directly (9.5 s). Dwell histograms revealed two shut and three open components, which is less complex than wild-type (S3 Table). Moreover, the mutated receptors desensitised for a mean of 2004 ± 268 ms, yielding a desensitisation equilibrium constant of 41.7 s‒1 (ω = 0.499 s‒1 and δ = 20.8 s‒1). The mean active durations and PO data are summarised as bar plots in Fig 9E and 9F, respectively.
Wild-type receptors activated rapidly upon exposure to 5 nM IVM and 2 μM glutamate. For the first 66 ± 18 s after commencement of the recording, the active periods were well-separatedincreasing in duration over time (Fig 10A), until an apparent steady-state equilibrium was reached (Fig 10B).
An analysis of the active durations, PO and the dwell time components at equilibrium produced a mean active duration of 15.9 ± 4.2 s and a PO of 0.88 ± 0.02 (Table 2) for potentiation of wild-type currents. The dwell histograms were best described by three shut and three open components (S3 Table). The time constants for the first two shut and open components were similar to those in the presence of low concentrations of glutamate (S2 Table). In contrast, the longest open component was at ~ 200 ms and represented about 40% of the total open intervals. To estimate receptor desensitisation, long stretches of record consisting of single receptor activity were analysed to obtain the main shut component that separated discrete active periods (Fig 10B). This shut component produced a short-lived mean, after correcting for channel number, of 223 ± 36 ms and thus an ω of 4.48 s‒1. Using a mean active duration of 15.9 s (δ = 0.063 s‒1), the calculated equilibrium constant for receptor desensitisation was 0.01 in the presence of IVM and glutamate.
Similar experiments were carried out for the G36’A-containing mutant. Here too the active periods initially increased in duration (Fig 10C), before equilibrating to steady-state activity (Fig 10D). However, steady-state activity was not near continuous, as was observed for the wild-type receptors. Instead, individual receptors were active for a mean duration of 113 ± 25 ms and had a PO of 0.60 ± 0.04 (Table 2). Receptor desensitisation was also unlike that of wild-type receptors. The mean shut lifetime for long stretches of record was 2381 ± 657 ms (ω of 0.420 s‒1). This yielded an equilibrium constant for desensitisation of 21.1. As for glutamate-gated activity, IVM produced briefer active periods and induced greater desensitisation in the G36’A GluClRs than wild-type. The summary of the mean active durations and PO is provided in Fig 10E and 10F, respectively. The POs were significantly different between direct activation and potentiation for both wild-type and mutant receptors. However, because direct activation by IVM of the mutant receptors produced brief, simple activations, the PO determined for this activity was relatively high.
In summary, IVM acted as an agonist and potentiated currents in the presence of glutamate at wild-type and G36’A mutated GluClRs to elicit significantly longer active periods and markedly reduce receptor desensitisation. In addition, the sparse activity at the start of the recordings, which equilibrated to steady-state activity implies that, as with glutamate, additional binding of IVM molecules to each receptor saturates receptor activation.
The two broad aims of this study were firstly, to examine the activation properties of GluClRs expressed by a parasitic species in the presence of its physiological agonist and secondly, to explore the mechanism of IVM sensitivity. To achieve the first aim glutamate-gated currents were examined over a wide concentration range on single receptor and ensemble levels. The conductance of the receptor channel was determined to be ~23 pS, which is close to that of GABAARs that comprise α, β and γ subunits [13, 54]. Upon binding to glutamate, wild-type GluClRs activated rapidly (~9000 s‒1, Fig 4), comparable with the rate of other pLGICs, including the G36’A mutated GluClRs (Fig 8). The experiments also revealed that wild-type GluClRs were highly responsive even at low nanomolar concentrations of glutamate and exhibited active durations and an open probability that was concentration-dependent. These parameters saturated at ~500 ms and 0.99, respectively (Fig 3).
Dwell interval analysis of active periods demonstrated that the receptors have multiple components, indicating that each receptor oscillates between multiple functional states during receptor activation [8, 11, 13]. The pattern of dwell components also indicated that at ≥ 2 μM glutamate an optimal number of bound glutamate molecules achieves efficient receptor activation. This is similar to GlyR activation, whereby fitting data to postulated kinetic schemes it was deduced that three bound glycine molecules are sufficient for optimal activation [8]. The decrease in open dwell times at nanomolar concentrations of glutamate clearly showed that at these concentrations fewer glutamate molecules were bound on average to each receptor [50].
The G36’D and G36’E mutations have been identified in the ML-resistant agricultural pest mite T. urticae [35, 36]. These mutations occur on different subunit isoforms, suggesting that heteromeric GluClRs containing different substitutions to G36’ could work either individually or synergistically to reduce ML sensitivity [36, 55]. The G36’E mutation is particularly effective at reducing ML sensitivity on its own and homomeric receptors expressed in oocytes demonstrate complete insensitivity to two MLs (abamectin and milbemycine A4) [55]. Our data suggest that the G36’A mutation gives rise to significant functional changes, such as a reduced active duration and an increase in desensitisation of single receptors, which manifest as faster current decay and reduced sensitivity to glutamate and IVM. Whether these functional changes are also present in G36’D or E has yet to be determined. However, given that both substitutions contribute large side groups that are likely negatively charged, it is likely that these too would affect the activation properties of the receptors. The physico-chemical properties of aspartate and glutamate may also restrict access of IVM to its binding site.
We chose to study the G36’A mutation because it dramatically decreases IVM sensitivity [30] and is located on the TM3 domain where crystallographic data show that it contributes one side of the IVM binding site [6] (Fig 11). Given its location, it is tempting to hypothesise that the G36’A substitution reduces IVM sensitivity simply by disrupting the binding of IVM. However, the mutation also decreases the EC50 of glutamate [30], which binds at an extracellular domain site that is over 3 nm from the site of the mutation. Another mechanism that could reconcile the parallel decrease in glutamate and IVM sensitivities is that the actions of both ligands reveal changes to the intrinsic activation properties of the receptor conferred by the mutation. Distinguishing between these two possibilities is critical to understanding the mechanism of action of IVM. This is of particular importance given that IVM resistance in H. contortus and other pest species is an emerging concern [21, 26, 29].
To help distinguish between these two possibilities we analysed glutamate- and IVM-gated currents in wild-type and G36’A mutated receptors. Clear evidence that the G36’A mutation markedly compromised receptor activation was gleaned in the presence of glutamate alone. The mutation gave rise to two distinct and stable modes of activation; one that was similar to wild-type, and another with a much reduced PO (Figs 5 and 6). The wild-type like mode was briefer than the activations mediated by wild-type receptors over the glutamate concentrations tested and both modes had lower POs than wild-type. When analysed together, the net effect of these modes produced a maximum mean active duration of ~200 ms and a PO of 0.70 (Fig 6). These parameters underlie the reduced sensitivity to glutamate observed in the G36’A mutated receptors. Indeed, where 2 μM glutamate elicited robust activity in wild-type receptors, it produced only sparse, simple activity in the mutant. These results led to the hypothesis that the mutation impaired receptor desensitisation and ensemble current decay. This was tested at the single receptor level (Fig 7) and in macropatches (Fig 8). The single receptor experiments yielded desensitisation equilibrium constants of ~180 and ~625 for wild-type and mutant receptors, respectively, representing a 3.4-fold greater likelihood of adopting desensitised states in the mutant. Ensemble deactivation and desensitisation rates were also much abbreviated in the mutant, producing mean time constants that corresponded well to the mean active durations of single receptors (Fig 8). It is notable that other pLGICs, such as α1β GlyRs and α1β2γ2 GABAARs have a similar sensitivity to IVM [30, 44] and also exhibit similar rates of current decay [14, 51, 53] to the G36’A mutant. Moreover, pLGICs that exhibit low IVM sensitivity also contribute non-G36’-containing TM3 domains to their IVM binding sites [30, 37, 44].
IVM acted as a ligand on its own and synergistically with glutamate to enhance currents elicited by glutamate. It did not affect the single channel conductance even though it binds at a site within the transmembrane segments and is predicted to interact with the pore-lining TM2 domain [6]. At GABAARs it has been demonstrated that 10 nM IVM alone lengthens the durations of single receptor currents without changing single channel conductance [46].
2 μM glutamate applied alone at wild-type GluClRs gave rise to a mean active duration of ~150 ms. When 5 nM IVM and 2 μM glutamate were applied together, current potentiation manifested as prolonged active durations with a mean of ~16 s, representing a two order of magnitude increase. The same combination of IVM and glutamate at G36’A mutated receptors produced active durations with a mean of 113 ms (Fig 10), compared to a mean of ~11 ms elicited by 2 μM glutamate alone. This also represents an order of magnitude change, but the absolute durations were much briefer than in wild-type receptors. A similar pattern was observed between wild-type and G36’A receptors in the presence of 5 nM IVM alone. The mean duration of active periods for wild-type receptors was ~9.5 s, whereas that for the mutant was a mere 48 ms (Fig 9). As for glutamate-gated currents, the active periods of the G36’A mutated receptor were much briefer when activated by IVM alone or in conjunction with glutamate compared to wild-type receptors.
Receptor desensitisation in the presence of IVM was estimated by fitting shut histograms to long periods of record that contained successive single receptor activations (Figs 9 and 10). Receptor saturation, where all the receptors in each patch became active, was then used to count active receptors and correct for the desensitisation time constant. This analysis revealed that desensitisation was nearly abolished at wild-type receptors, especially in the presence of IVM and glutamate. The mean lifetimes of desensitised states were between ~220 ms and ~540 ms and yielded equilibrium constants of ~0.01 for IVM plus glutamate and ~0.06 for IVM alone, respectively. IVM alone induced a mean desensitisation lifetime of 2002 ms and an equilibrium constant of ~42 in the mutant receptors. This represents a significant increase in desensitisation compared to wild-type receptors under the same recording conditions. These data demonstrate that in the presence of each agonist alone and when they are co-applied, the G36’A mutated receptors exhibited briefer active periods and enhanced desensitisation compared to wild-type. Our data strongly support the inference that the loss of sensitivity reported for both agonists [30] is due to the same mechanistic process, and not fundamentally related to IVM binding interactions at the 36’ position. Although we cannot categorically rule out an IVM binding effect our data show that the wild-type and the G36’A mutant receptors are similarly affected even when receptor activation is at saturation throughout the recording. These conditions also correspond to ligand saturation where occupancy of receptors in unbound states is negligible. That this is the case for glutamate (Figs 5 and 6) and IVM (Figs 9 and 10) strongly suggests that both agonists are less efficacious at activating the mutant receptors.
A notable difference between the actions of glutamate and IVM was that the onset and equilibration of currents in the presence of IVM were much slower than observed for glutamate. A lag time of over ~1‒1.5 minutes was apparent between the initiation of channel activity and the time when activations equilibrated to a constant mean duration for both mutant and wild-type receptors. Indeed, no activity was seen when IVM was applied alone for the first minute or so. Diffusion limited binding rates, calculated for ligands that encounter receptor binding sites directly from aqueous solution, including ligands of similar dimensions to IVM are in the range of ~5‒7 x 109 M‒1s‒1 [56]. For instance, the upper limit of the diffusion rate for a small aqueous molecule like glutamate is ~109 M‒1s‒1 [57]. The binding energy and correlated structural changes at binding sites can reduce these values by about two orders of magnitude (~106‒108 M‒1s‒1)[56]. These diffusion rates are far too high to account for the lag time observed in the recordings, suggesting the existence of other rate-limiting factors [58]. Structural evidence indicates that IVM binds to an inter-subunit cavity in the upper leaflet of the lipid bilayer [6], as do other highly lipophilic ligands such as neurosteroids [59] and anaesthetics [60]. The IVM binding pocket in GluClRs is likely to be partly occupied by lipid, requiring its displacement by IVM for access to the pocket [6, 7]. Due to its lipophilic nature, IVM is believed to partition into cell membranes [61] where it reaches a high local concentration, consistent with persistent whole-cell currents [30]. Thus, much of the ‘binding energy’ of IVM could be derived from the nonspecific free energy of membrane partitioning, giving rise to a high apparent affinity, whereas the actual ligand-channel interaction could be relatively weak [62]. Our data are consistent with IVM partitioning in the lipid membrane and diffusing to its binding pocket [63], where its concentration would increase to produce current saturation over the course of several minutes in patches of membrane. The increase in the active durations of individual receptors over the initial phase of the recordings and the emergence of a long open time constant at saturation also suggests that multiple IVM molecules bind to each receptor over course of the experiment to produce saturation. Heteromeric α1β2γ2 GABAARs have also been shown to bind multiple IVM molecules, to produce interface-specific potentiation and direct current activation [44].
It has been suggested that the flexible ‘hinge’ function of glycine residues found within K+ channels [64, 65] and pLGICs [66] can serve to isolate protein segments, or even entire domains, from surrounding protein conformational changes during channel activation [65]. According to their respective high resolution molecular structures, the TM3 domain backbones of the α1 GlyR (which contains an endogenous A36’ residue) and the GluClR are closely aligned in the shut state (Fig 11A). However, upon IVM binding, the GlyR TM3 undergoes a larger displacement (Fig 11B). This differential displacement is also observed when the TM3 domains corresponding to shut and IVM-bound states are overlaid from the same receptor (Fig 11C and 11D). This strongly suggests that the A36’ residue confers structural rigidity to the TM3. The structural comparisons in Fig 11 illustrate that the G36’ acts to minimise deformation of the TM3 between state transitions during the conformational activation ‘wave’ of pLGICs [67]. Because the G36’A mutation causes briefer active durations and an increased likelihood of adopting glutamate and IVM-induced desensitised states we conclude that the alanine destabilises open states via reduced backbone flexibility and a larger TM3 displacement. Functional studies have established that pLGIC activation and desensitization are mediated by structurally distinct sets of conformational changes at the both extracellular-transmembrane domain interface [48, 49] and at the intracellular end of the pore [47]. The difference in IVM-induced TM3 displacement in the wild-type and G36’A mutant GluClRs will cause TM3 to interact differentially with one or both of these regions, and could thus explain the differential effect of the mutation on desensitization.
The H. contortus α (avr-14b) GluClR is an important biological target for IVM, although IVM resistance is emerging as a problem in this pest species. Here we quantified the effects of glutamate and IVM on these receptors with the aim of understanding the structural and functional bases of their modulatory effects. We found the receptor to be highly responsive to low nanomolar concentrations of both ligands. Dwell interval analysis of active periods demonstrated that the receptor oscillates between multiple functional states during activation by either ligand. However, we also observed that the duration of activations increased with increasing ligation of receptors by either ligand. The G36’A mutation, which was previously thought to hinder access of IVM to its binding site on the receptor, was found to decrease the duration of active periods and increase receptor desensitisation. On an ensemble macropatch level these changes gave rise to enhanced current decay and desensitisation rates. There are two main reasons why we consider these effects are due to impaired channel gating and not impaired IVM binding. First, the impairment to gating was quantitatively similar for the two ligands which bind to structurally distinct sites, and second, the impairment was observed at saturating concentrations of either ligand, thus ruling out a contribution to gating from binding and unbinding events. We infer that G36’A affects the intrinsic properties of the receptor with no specific effect on IVM binding. These results provide new insights into the activation and modulatory mechanism of the GluClR and provide a mechanistic framework upon which the actions of new candidate anthelmintic drugs can be reliably interpreted.
HEK AD293 cells (ATCC cell lines, VA USA) were seeded onto poly-D-lysine coated glass coverslips and transfected with cDNAs encoding the GluClR subunit avr-14B (pcDNA 3.1+) of H. contortus using a calcium phosphate-DNA co-precipitate method. The cDNA encoding the CD4 surface antigen was also added to the transfection mixture and acted as a marker of transfected cells. Cells were used for experiments 2–3 days after transfection. The point mutation, TM3-G36’A, was incorporated into the subunit using the QuickChange site-directed mutagenesis method. Successful incorporation of mutation was confirmed by sequencing the mutated DNA.
All experiments were carried out at room temperature (21–24°C). Single-channel and macropatch currents were recorded from outside-out excised patches at a clamped potential of −70 mV, unless indicated otherwise. The patches were continuously perfused via a gravity-fed double-barrelled glass tube. Out of one barrel flowed an extracellular bath solution containing (in mM), 140 NaCl, 5 KCl, 1 MgCl2, 2 CaCl2, 10 HEPES, and 10 D-glucose and titrated to pH 7.4. The adjacent barrel contained agonist dissolved in this extracellular solution. Glass electrodes were pulled from borosilicate glass (G150F-3; Warner Instruments), coated with a silicone elastomer (Sylgard-184; Dow Corning) and heat-polished to a final tip resistance of 4‒15 MΩ when filled with an intracellular solution containing (in mM) 145 CsCl, 2 MgCl2, 2 CaCl2, 10 HEPES, and 5 EGTA, pH 7.4. Stock solutions of L-glutamate were also pH-adjusted to 7.4 with NaOH. A 10 mM stock of IVM (Sigma-Aldrich) was dissolved in 100% DMSO and kept frozen at ‒20°C. Fresh working stocks of IVM at 5 nM were prepared by dissolving the appropriate quantity directly in extracellular solution. 100% DMSO when dissolved in extracellular solution alone at the same concentration as is present in working solutions containing 5 nM IVM had no effect on patches excised from cells transfected with GluClRs or from untransfected cells.
Excised patches were directly perfused with extracellular solution by placing them in front of one barrel of the double-barrelled glass tube. Single channel currents were elicited by exposing the patch continuously to agonist containing solution, flowing through the adjacent barrel. 1–2 glutamate concentrations were applied to most patches for single receptor experiments. A ~1 minute wash with agonist-free extracellular solution was applied between each glutamate application. Because IVM does not readily wash out, either 2 μM glutamate + 5 nM IVM or 5 nM IVM alone were applied to a given patch. Macropatch currents were elicited by lateral translation of the tube from the agonist free to agonist containing barrel using a piezo-electric stepper (Siskiyou). This achieved rapid solution exchange (<1 ms). Currents were recorded using an Axopatch 200B amplifier (Molecular Devices), filtered at 5 kHz and digitized at 20 kHz using Clampex (pClamp 10 suite, Molecular Devices) via a Digidata 1440A digitizer.
The experiments that were carried out can be broadly divided into 1) single receptor currents at steady-state and 2) ensemble currents, which are phasic. The two types or experiments are complimentary and provide different data. Single channel recordings yield information on receptor conductance and functional state complexity (eg. active durations and dwell histograms). The fast application (~1 ms) ensemble measurements mimic synaptic currents.
Single-channel current amplitudes were measured in Clampfit. In current-voltage (i-V) experiments, the amplitude was measured at voltages of, ±70 mV, ±35 mV, ±15 mV and 0 mV. The data were fit to a polynomial function in Sigmaplot (Systat Software) and the reversal potential was read directly from the plots. Single-channel conductance (γ) was calculated from the single-channel amplitude (i) using Ohm’s law:
γ=iVhold−Vljp−Vrev
Eq 1
Where Vhold is the holding potential (−70mV), Vljp is the liquid junction potential and Vrev is the reversal potential. Vljp was calculated to be 4.7 mV for the solutions used in the experiments [68]. We confined our analysis to the largest, main conductance level. QuB software was used to analyse the kinetic properties of GluClR activations. Segments of single-channel activity separated by long periods of baseline were selected by eye and idealized into noise-free open and shut events using a temporal resolution of 70 μs. Idealized data were initially fit with a simple activation scheme in which open and shut states were added to a central shut state. This fit was used to determine the critical time (tcrit), which was taken from the shut interval durations and used to divide the idealized segments into clusters (or bursts at < 2 μM glycine) of single receptor activity. Clusters and bursts will be referred to as activations. tcrit applied to single channel records of wild-type activity varied between 5–30 ms for concentrations ≥ 2 μM and 120–180 ms for 30 nM and 5 nM glutamate. Activation mode analysis for G36’A-containing receptors at 1 mM glutamate required tcrit times of 180–200 ms (low PO) or 15–50 ms (high PO). Pooled data obtained from G36’A-containing receptors were defined using tcrit times of 20–50 ms. Finally, IVM-induced single channel currents were defined using tcrit values of 50–120 ms for both wild-type and mutant receptors. This analysis yielded mean cluster durations and intra-activation open probabilities (PO). All data are presented as mean ± SEM of between 3 and 16 patches. The shut periods that correspond to receptor desensitisation were estimated by generating shut histograms for long stretches of record (several minutes) that exhibited single receptor activations. Receptor desensitisation was modelled as a single transition from an open conducting state (ARo) to a desensitised state (ARd). Where A is the agonist, R is the receptor and the superscripts denote open (o) or desensitised (d). δ denotes the desensitisation rate constant whereas the re-sensitisation (or re-activation) rate constant is denoted by ω. The equilibrium constant for desensitisation is δ/ω.
Macropatch currents were analysed by fitting the onset phase of the current to a single exponential of the form:
I(t)=Imax(1−e−kobst)
Eq 2
Where I(t) is the current at time t, Imax is the maximum current amplitude and kobs is the pseudo-first order rate constant for current activation. The decay phase of the macropatch currents were fit to two standard exponential functions.
Data are presented as mean ± SEM. Power analysis of our data sets for IVM revealed power levels of 0.9‒1.0. Two-tailed, unpaired t-tests were used to compare wild-type and mutant current parameters in the presence of IVM and p < 0.01 was taken as the significance threshold.
The alignments of TM3 domains of the GluClR and GlyR were done using the Internal Coordinate Mechanics software (ICM-Pro Molsoft LLC, San Diego, CA). The α-carbon atoms of the N-terminal residues from TM3-29’ to TM3-36’ were superimposed and used as a fixed reference. The displacement between α-carbon atoms at position TM3-56’ were then measured between two given TM3 domains The structures used for this analysis were, the GluClR in a non-conducting conformation (PDB, 4TNV [7]), the GluClR in complex with IVM (PDB, 3RHW [6]), the α1 GlyR in a non-conducting conformation in complex with strychnine (PDB, 3JAD [69]), and the structure of the α1 GlyR in complex with IVM (PDB, 3JAF [69]). The final representations were created using the Pymol Molecular Graphics System, Version 1.3.
|
10.1371/journal.pgen.1000939 | DNA Adenine Methylation Is Required to Replicate Both Vibrio cholerae Chromosomes Once per Cell Cycle | DNA adenine methylation is widely used to control many DNA transactions, including replication. In Escherichia coli, methylation serves to silence newly synthesized (hemimethylated) sister origins. SeqA, a protein that binds to hemimethylated DNA, mediates the silencing, and this is necessary to restrict replication to once per cell cycle. The methylation, however, is not essential for replication initiation per se but appeared so when the origins (oriI and oriII) of the two Vibrio cholerae chromosomes were used to drive plasmid replication in E. coli. Here we show that, as in the case of E. coli, methylation is not essential for oriI when it drives chromosomal replication and is needed for once-per-cell-cycle replication in a SeqA-dependent fashion. We found that oriII also needs SeqA for once-per-cell-cycle replication and, additionally, full methylation for efficient initiator binding. The requirement for initiator binding might suffice to make methylation an essential function in V. cholerae. The structure of oriII suggests that it originated from a plasmid, but unlike plasmids, oriII makes use of methylation for once-per-cell-cycle replication, the norm for chromosomal but not plasmid replication.
| Bacteria usually have one chromosome but can have extrachromosomal replicons, called plasmids. Although normally dispensable, plasmids can confer adaptive advantage to cells in stressful environments. Bacteria can also have multiple chromosomes, each carrying essential genes, as in eukaryotes. In all organisms, chromosomes duplicate once before the cells divide so that the daughter cells can receive equal genetic dowry, but this is not usually the case with bacterial plasmids. Vibrio cholerae, the causative agent for the disease cholera, has a typical bacterial chromosome like the chromosome of the well-studied bacterium Escherichia coli and has a second chromosome with many signatures indicating its origin from a plasmid. Here we show that, in spite of the distinct nature of the two chromosomes, they both duplicate once per cell cycle, and they both require DNA adenine methylation for this purpose. Our study suggests that once-per-cell-cycle replication is a necessary feature of a chromosome in multichromosome bacteria, and provides a paradigm of how methylation could endow extrachromosomal replicons with the capacity to duplicate like chromosomes.
| The regulatory potential of canonical DNA sequences can be greatly expanded by epigenetic modifications. Methylation is the most common modification of DNA and is widely used to control many cellular processes [1]. In bacteria, DNA methylation is restricted to adenine and cytosine residues [2], and can facilitate or interfere with DNA-protein interactions, thereby modulating various DNA transactions [3]. Such transactions include gene expression, DNA restriction, DNA mismatch repair, and chromosome replication and segregation [4], [5].
Most of our knowledge regarding the role of methylation in chromosome replication comes from studies in Caulobacter crescentus and Escherichia coli. In C. crescentus, initiation of DNA replication requires the adenines of the GANTC sequences in the origin of replication to be methylated on both the top and bottom strands by the methylase CcrM. How the methylation helps the origin function is not known, although methylation lowers DNA stability [6], [7] and thereby could facilitate origin-opening, an essential step in the replication initiation process. It is also possible that the methylation changes DNA structure to facilitate protein-DNA interactions at the origin [8]. Irrespective of the mechanism, methylation not only controls the timing of initiation but also restricts initiation to once per cell cycle [9]. Following initiation, the hemimethylated sister origins cannot be reused in the same cell cycle, as the CcrM methylase is not synthesized until the end of the replication cycle.
In E. coli, the methylase is called Dam and acts on the adenines of GATC sequences, which are particularly frequent in the origin of replication, oriC. In this bacterium also the methylation most likely helps in origin-opening [8], [10] but plays a more definite role in restricting the initiation to once per cell cycle [11]. In E. coli, immediate reinitiation is prevented, not by delaying the synthesis of the methylase, but by preventing its action through sequestration of hemimethylated sister origins by a hemimethylation-specific DNA binding protein, SeqA [12]. Sequestration renders DNA unavailable to the methylase. The sequestration also allows initiation synchrony whereby the multiple origins that E. coli maintains during rapid growth fire nearly simultaneously. It is believed that the sequestration process continues at least until all the origins have fired. This happens in a narrow window of time giving rise to the initiation synchrony phenotype [13]. In the absence of Dam, the newly replicated origins, without their hemimethylation marks, remain indistinguishable from the unreplicated ones. The choice of origin for replication being random, once-per-cell-cycle initiation from each origin is no longer guaranteed. As a result, in dam mutants, the initiation becomes asynchronous and cells can have origins that do not fire at all or fire more than once in the same cell cycle. The consequences are the same in seqA mutants, because without sequestration, replicated origins also remain competent for reinitiation.
The lack of discrimination between replicated and unreplicated origins can lead to origin incompatibility [14]. If extra copies of oriC are introduced as plasmids into wild type (WT) E. coli, the plasmid copies do not compete with the chromosomal oriC because of sequestration of newly replicated origins. Without sequestration, in dam or seqA mutants, the plasmid copies remain available for reinitiation, and under selection they can block the growth of cells in which the chromosomal origins did not get a chance to fire. Sequestration-deficient strains are therefore not easily transformed with oriC plasmids [15]. Thus, although not normally required, Dam or SeqA can be essential in a competitive situation.
Vibrio cholerae has two chromosomes (chrI and chrII). The origin of chrI (oriI) shares 58% identity with the E. coli oriC, and both have similarly high densities of GATC sites. The origin of chrII (oriII) also has a high density of GATC sites but has a second feature of a major class of plasmids: repeated initiator-binding sites (iterons) [16]. The dam gene is also essential for V. cholerae, although the reason has remained unknown [17]. Our interest in the role of methylation in V. cholerae chromosomal replication stems from the fact that although the bacterium is a close relative of E. coli, plasmids with either oriI or oriII could transform WT E. coli, but not when it lacked Dam [18]. It remained unclear whether the failure to recover transformants in the case of oriI is because the origin could not function or because of competition (incompatibility) with the closely related chromosomal oriC [14], [18]. Incompatibility is unlikely the case of oriII, since it has little similarity to oriC. Moreover, while oriI and oriC are regulated by the DnaA initiator protein, oriII is regulated by its own specific initiator, RctB [19]. The reason for the Dam requirement of oriII could thus be for the functioning of the origin itself.
Here we show that oriC can be replaced by oriI in the E. coli chromosome, and in this chromosomal context oriI functions without requiring Dam or SeqA. Incompatibility with the chromosomal oriC thus remains a satisfactory explanation of the earlier finding of a Dam requirement for oriI plasmids [18]. For oriII, Dam but not SeqA appears to be required as only fully methylated oriII DNA, but not hemi- or un-methylated DNA, could bind efficiently to the oriII-specific initiator RctB in vitro. Since the binding of RctB is a prerequisite for oriII function, this provides an explanation for why Dam is essential for V. cholerae, chrII being indispensable. Finally, we show that SeqA is necessary to restrict initiation to once per cell cycle for both oriI and oriII, as is the norm for chromosomal origins. Although chrII is believed to have originated from a plasmid, our findings of the methylation requirement for its initiation and cell-cycle specific regulation are unprecedented in studies of plasmids [20], [21]. It appears that a plasmid origin acquired methylation to function as a chromosomal origin, thus providing a novel example of origin evolution in bacteria.
The E. coli origin of replication, oriC, does not require dam and seqA to initiate replication. In contrast, plasmids driven by oriC are highly deficient in transformation of dam mutants [11]. This is believed to be due to irreversible sequestration of hemimethylated plasmid origins by the SeqA protein after the first round of replication [22]. Indeed, seqA and dam seqA strains can be transformed by oriC plasmids, although the efficiency is lower compared to WT due to incompatibility with the chromosomal copy of the origin [15]. The requirement of dam thus is not intrinsic to oriC function and appears so only in the plasmid context. The dam requirement of V. cholerae oriI has so far been studied only in the plasmid context. However, in contrast to oriC plasmids, oriI plasmids not only failed to transform an E. coli dam mutant but also a seqA or a dam seqA mutant, raising the possibility that the genes could be essential for oriI [11], [18]. We confirmed the plasmid results using E. coli MG1655 (BR1703) and its dam (CVC1415), seqA (BR1704) and dam seqA (CVC1424) mutant derivatives. As before, the dam, seqA and dam seqA mutants could not be transformed with an oriI plasmid, and only the dam mutant could not be transformed with the oriC plasmid (Figure 1A). We suggest below that the oriI plasmid possibly replicated in the absence of dam or seqA, which competed out replication from the chromosomal oriC and led to inviability of the transformants.
To avoid plasmid-mediated competition (incompatibility), we studied oriI by placing it in the E. coli chromosome. Using the Red recombineering system, we replaced the minimal oriC region with the corresponding oriI region (Materials and Methods). The resultant strain, MG1655ΔoriC::oriI-zeo (CVC1400, Table 1; hereafter called MG1655ΔoriC::oriI), could be made dam minus by P1 transduction, using dam-16::aph (CVC1383) as the source of the mutant dam allele [23]. We could also replace the oriC region of MG1655ΔseqA10 with ΔoriC::oriI by P1 transduction. The viability of dam, seqA or dam seqA mutant derivatives of MG1655ΔoriC::oriI (CVC1401, CVC1416 and CVC1425, respectively) indicates that oriI does not require Dam and SeqA for functioning in E. coli.
To understand why oriI and oriC behave similarly in the chromosomal context but differently in the plasmid context, we repeated the transformation experiments using MG1655ΔoriC::oriI cells as the host. The oriI plasmid could now transform the seqA and the dam seqA derivatives of MG1655ΔoriC::oriI efficiently but not the dam derivative (Figure 1A and 1B). The failure to transform the dam derivative can be attributed to permanent sequestration. In contrast to oriI, oriC not only failed to transform the dam derivative but also the damseqA derivative of MG1655ΔoriC::oriI. The results can be understood assuming initiation from oriI to be more efficient than from oriC. Most likely, the weaker oriC failed to compete with oriI in the chromosome (incompatibility) that led to inviability of the transformants. It is known in E. coli that incompatibility problems can be aggravated when the incoming and recipient origins have unequal efficiencies [24].
oriI and oriC were further analyzed using flow cytometry [25]. Replication initiation and cell division were blocked by antibiotics rifampicin and cephalexin, respectively, but sufficient time was allowed after drug addition to complete replication elongation (replication run-out). This method provides a measure of the fraction of the population that already initiated replication at the time of drug addition. In LB, after the replication run-out, MG1655 cells were distributed mostly into two populations, one with four and the other with eight full chromosomes (Figure 2A). This indicates that cells were born with four origins and they all fired synchronously once, giving rise to the eight chromosome peak. In the dam and seqA mutants, cells had a widely varying number of chromosomes indicating asynchronous initiation (Figure 2C and 2E) [22], [26]. There were also cells with more than eight chromosomes indicating that initiation was no longer restricted to once per cell cycle. In the engineered strain, MG1655ΔoriC::oriI, replication initiation was synchronous (Figure 2B) but not in its dam or seqA derivatives (Figure 2D and 2F). The requirements of dam and seqA for synchronous and once-per-cell-cycle initiation are thus maintained when oriI replaces oriC.
Compared to the WT, replication initiation was less frequent in dam mutants but more frequent in seqA mutants in the case of both the origins. As is oriC, Dam seems to be playing a positive role and SeqA a negative role in replication initiation from oriI.
It was reported earlier, and we confirmed, that oriII plasmids can not transform an E. coli dam mutant but can transform a seqA mutant [18]. The oriII plasmids also failed to transform the dam seqA mutant, indicating that irreversible sequestration cannot account for the dam requirement. The oriII function could not be tested in the chromosomal context, as was done for oriI, because attempts to replace oriC with oriII failed. In any event, incompatibility between oriC and oriII appears to be an unlikely explanation for the dam requirement, as the structure and control elements of the two origins are different [19]. We show below that the reason for the dam requirement could be for binding of oriII to its specific initiator RctB.
A distinguishing feature of oriII is that its putative RctB binding sites, called 11- and 12-mers, all contain a GATC site. This prompted us to test whether methylation of the sites might be important for RctB binding (Figure 3A). We first tested binding to the six tandem 12-mers within the minimal oriII by an electrophoretic mobility shift assay. Purified RctB bound efficiently to the 12-mer fragment, when it was fully methylated (Figure 3B). The binding was nearly saturated because most of the DNA molecules were maximally retarded. Binding to hemimethylated DNA, where either the top or the bottom strand carried the methylation marks, and to unmethylated DNA was significantly less. In these cases, most of the bound species appeared as a smear, indicative of weaker binding. The binding improved when the DNA samples were remethylated using Dam in vitro (Figure 3B). The binding of RctB to the three 11-mers or to a pair of 12- and 11-mers in the negative-control region of oriII was also efficient when the sites were fully methylated (Figure S1A and S1B). Mutating GATC sites to GATG in the 11- or the 12-mer abolished the binding (Figure S1C). These results indicate that full methylation can significantly improve the affinity of RctB to the 11- and 12-mers.
To confirm these results in vivo, RctB binding to a plasmid with the six 12-mers was studied in MG1655 or its dam derivative by chromatin immunoprecipitation (ChIP), and the immunoprecipitated DNA analyzed by quantitative PCR. Compared to the vector, the plasmid with the 12-mers was preferentially enriched by immunoprecipitation when the DNA samples were from WT cells (Figure 3D). No significant enrichment was obtained when the DNA samples were from the dam mutant. These results show the importance of methylation for efficient RctB binding in vivo, and therefore, for replication of chrII.
To test how well the results obtained in vitro and in E. coli reproduce in the native host, the dam gene of V. cholerae was deleted in the presence of a complementing plasmid, pTS-PBADdam (pGD93, Table 2). The replication of this plasmid is temperature sensitive and the cloned V. cholerae dam gene is under the control of an arabinose-inducible and glucose-repressible promoter, PBAD. On LB plates, under the permissive condition (30°C and in the presence of arabinose), the Δdam/pTS-PBADdam strain grew as well as the WT but under the restrictive condition (42°C and in the presence of glucose), single colonies were barely visible (Figure 4A). In LB broth, under the restrictive condition, the mutant grew slower than the WT (with generation times of 27 min and 22 min, respectively), and the growth plateaued to an OD of 0.53 only (Figure 4B). Moreover, the number of viable cells in the mutant culture was only 0.02% of the number of viable WT cells, when initially similar cultures of both were grown for seven and a half hours under restrictive conditions (Figure 4C). The viable cells in the mutant all retained the dam complementing plasmid without selection for it. The results thus appear consistent with an earlier report that Dam is essential for V. cholerae [17].
Under the condition of dam depletion, we expected that initiation at oriII would decrease more than initiation at oriI. This was tested by determining the relative replication efficiencies of the two chromosomes in exponentially growing cells by qPCR. We quantified the amount of DNA at the two origins and the two termini to obtain the ratios oriI/oriII, oriI/terI, oriII/terII and terI/terII. Under the restrictive condition, there was a significant increase (4-fold) in the value of oriI/oriII and of terI/terII, while the values of oriI/terI and oriII/terII remained unchanged (Figure 4D). These results are consistent with our expectation that compared to chrI, replication of chrII is more dependent on Dam.
The hemimethylation period, the time to remethylate a GATC site after passage of the replication fork, is particularly prolonged at oriC because of the presence of high density of GATC sites within the origin [12]. The prevalence of high density of GATC sites in both oriI and oriII (Figure 5A) prompted us to examine their hemimethylation period, as was done using asynchronous exponential cultures [27], [28].
We examined the hemimethylation period of a GATC site within the origin and, for comparison, another site external to the origin (about 300 kb away) for each of the chromosomes. In oriI, the GATC site chosen is between DnaA boxes R3 and R4, and in oriII, it is between the fourth and the fifth 12-mers (arrows, Figure 5A). Total genomic DNA was extracted and digested with restriction enzymes whose recognition sequences overlap a GATC site and whose cleavage is inhibited when the site is fully methylated but not in one of the two hemimethylated sister sites, generated by passage of the replication fork (Figure 5B). The fraction of hemimethylated (cut) DNA at each of the origin sites was significantly higher than at the external sites (Figure 5C). The values were 11±3% and 56±8% for oriI and oriII, respectively, while at the external markers they were 4±0.8% and 8±3%, respectively (Figure 5D). The results indicate that as in E. coli, the hemimethylation period is prolonged at the two V. cholerae origins but the duration of the period can be significantly different for the two.
From the E. coli paradigm, we expected that SeqA would be required to prolong the hemimethylation periods at both the origins [22]. To test for the requirement, a partial in-frame deletion of seqA was made where the deleted region was substituted with a zeocin drug-resistance cassette, maintaining the seqA reading frame (Figure S2A). The resulting gene was called ΔseqAP and the strain CVC1410. Replication run-out experiments indicated that initiation of one or both the chromosomes has become asynchronous (Figure S2B), and in this respect, V. cholerae appears to be similar to E. coli (Figure 2A and 2E) [22].
For the GATC site tested in oriI, the fraction of hemimethylated DNA increased from 11% in WT to 68% in ΔseqAP (Figure 6A and 6C). Providing Dam or SeqA from a plasmid in the ΔseqAP background decreased the fraction of hemimethylated DNA. The decrease by providing excess of Dam was expected because it converts hemimethylated DNA to fully methylated DNA. The increase in the absence of SeqA and decrease in its presence were unexpected, if SeqA were responsible for prolonging the period. The seqA plasmid did not change the period significantly in the WT background (Figure S3). The results indicate that it is the absence of SeqA that causes the increase of hemimethylated oriI DNA, a result opposite to that found for oriC [29]. The behavior of oriII was similar to that of oriC: The fraction of hemimethylated DNA decreased from 75% in WT to 17% in ΔseqAP (Figure 6B and 6C). Thus seqA effects can be opposite in different origins at specific GATC sites. It remains to be seen whether the results are site-specific or true for the entire origins.
The opposite response of the GATC sites tested in oriI and oriII was also seen in a V. cholerae mutant where seqA was completely deleted (ΔseqAT, CVC2003; Figure S4). oriI also responded opposite to oriC in E. coli (Figure 7). While the percent of hemimethylated DNA at oriC dropped from 13% in MG1655 to 9% in MG1655ΔseqA10, the values at oriI increased from 9% in MG1655ΔoriC::oriI to 25% in its ΔseqA10 derivative. These results suggest that the opposite behavior of oriI and oriC upon seqA deletion is intrinsic to the sequence context of the GATC sites tested in the two origins rather than the sequestration machinery of the two bacteria. Thus depending upon the context, SeqA can both shorten and prolong the hemimethylation period of a GATC site.
Although a role of SeqA in restraining replication initiation in V. cholerae was suggested by the flow cytometry results (Figure S2B), they did not allow us to distinguish whether one or both the chromosomes were affected. We used fluorescence microscopy to follow replication initiation of the two chromosomes individually. The numbers and positions of oriI and oriII were determined in WT and ΔseqAP strains of V. cholerae by the GFP-P1ParB/parS system [30], [31]. For oriI in WT, 94% of the cells had two to four foci and the rest one or three foci, indicating synchronous and once-per-cell-cycle initiation (Figure 8A and 8E). In contrast, only 45% of ΔseqAP cells showed this pattern (Figure 8B and 8E). The remaining cells had five to nine foci. The significant increase in the number of cells with odd numbers of foci and more than four foci indicates that initiation is no longer synchronous and no longer limited to once per cell cycle in the absence of SeqA.
The regulation of chrII initiation was also affected. While 100% of the cells in the presence of SeqA showed one to two foci (Figure 8C and 8E), this was true for 83% of the ΔseqAP cells (Figure 8D and 8E). The remaining cells showed three to six foci. SeqA thus contributes to synchronous and once-per-cell-cycle initiation of both the chromosomes.
Here we have addressed the role of DNA adenine methylation in replication of the two V. cholerae chromosomes. In bacterial replication, adenine methylation can contribute by regulating gene expression, by helping origin opening, and by regulating initiation so that it occurs only once per cell cycle. From the regulatory point of view, the major contribution of methylation is the marking of promoters/origins so that unreplicated DNA can be distinguished from the replicated ones. Newly replicated DNA is uniquely marked with hemimethylated sites that lend themselves to regulation in various ways. In E. coli, the newly replicated initiator (dnaA) promoter and the origin (oriC) are silenced (sequestered) by the SeqA protein, which prevents their reuse for a significant period of the cell cycle. In C. crecsentus, the hemimethylated origin and the initiator promoter are also less active but the mechanisms remain unclear. In V. cholerae, we show that full methylation of oriII promotes initiator binding, providing a new role of the marks in replication initiation, and that SeqA is required for once-per-cell-cycle replication from both the origins (oriI and oriII), as in the case of oriC. By contributing to both initiation and its regulation, methylation thus serves two fundamental requirements for genome maintenance in V. cholerae. A comparison of oriII to plasmid origins also allowed us to address how a plasmid origin could have evolved to drive a chromosome in a cell-cycle specific fashion. We elaborate on these issues below.
Our work started by questioning the essentiality of Dam and SeqA for functioning of oriI since a similar origin, oriC, can do without them [18]. We find that the requirements are not real for oriI but were imposed due to the use of plasmids to check the origin function. When we replaced oriC in the E. coli chromosome with oriI, making it the only origin in the cell, both the dam and seqA genes could be deleted (Figure 1). Thus for the functioning of oriI and oriC, methylation is not essential but it improves chromosomal replication initiation and its control (Figure 2A–2D), including the ability to tolerate extra copies of the origin in trans (Figure 1). In bacteria such as Bacillus subtilis that are naturally devoid of the methylation system, ori plasmids can exert an inhibitory effect (incompatibility) on chromosomal replication [32]. Methylation thus can help bacterial survival in a competitive situation.
Dam plays a previously unrecognized role for oriII. It significantly promotes binding of the chrII-specific initiator, RctB, to the origin, thus possibly serving an essential function (Figure 3). Origin methylation is known to be essential for replication of C. crescentus chromosome, and of plasmids P1 and ColV-K30 in E. coli [33], [34], [35]. The reason is not clear in these cases, but unlikely to be for initiator binding. The initiator binding sites in these systems lack the sequences required for methylation. In contrast, RctB binding sites have an internal Dam recognition site, and methylation of the sites is required for initiator binding (Figure 3 and Figure S1). Thus, for oriII, the mechanism whereby methylation could be essential for its function and, therefore, for the bacterial survival is clear.
We show that seqA is not an essential gene in V. cholerae by obtaining viable seqA deletion mutants of V. cholerae. Although earlier studies suggested the gene to be essential, the finding that both oriI and oriII could function without SeqA in E. coli encouraged us to attempt isolation of the deletion mutants [18], [28]. In a deletion mutant, the number of both oriI and oriII per cell was found to be greater than in the WT (Figure 8). The overreplication indicates a breakdown of once-per-cell-cycle replication and reveals that SeqA is a negative regulator of replication. The latter was also concluded when the role of SeqA was studied by SeqA overproduction [28]. There was also an increase in the number of cells with odd number of origins for both the chromosomes, indicating loss of initiation synchrony. Thus, SeqA appears to contribute to both once-per-cell-cycle replication and initiation synchrony.
An unexpected finding of this study is that the hemimethylation period of oriI and oriC changed in opposite ways upon seqA deletion: for oriC it decreased whereas for oriI it increased (Figure 6 and Figure 7). The decrease in the case of oriC is expected since SeqA is believed to be the key factor that prolongs the period [22]. A significant increase of the period without requiring SeqA shows that there are other ways to prolong the period, and that SeqA can play an opposite role of shortening the period. The opposite roles of SeqA were seen in isogenic strains of both V. cholerae and E. coli, suggesting that the reason cannot be due to species-specific factors (Figure 6 and Figure 7). The period also changed in opposite ways for oriI and oriII in the same seqA mutants of V. cholerae. SeqA thus has the capacity to both increase and decrease the duration of the period.
SeqA binding to DNA is favored in GATC-dense areas [36], [37]. The density of GATC sites around the diagnostic GATC site happens to be quite different in the three origins. In particular, the diagnostic site in oriI is present in a relatively isolated position (Figure 7A). It is possible that the results therein might be site-specific and not representative of the entire origin.
Proteins other than SeqA that interact with origins can also explain the differences in the hemimethylated periods of the origins. DnaA is known to compete with SeqA for binding to some of the sites in oriC [38], and can significantly prolong the period even without SeqA [37]. Thus, DnaA is a likely candidate for prolonging the period for oriI in the absence of SeqA.
Upon seqA deletion, although the hemimethylation period changed oppositely for oriI and oriII, both the chromosomes over-replicated (Figure 8). The prolongation of the period thus may not always be diagnostic of the role of SeqA in the negative regulation of replication. As stated above, competition with DnaA for oriI binding could be another way for SeqA to exert its negative regulatory role [38]. The correlation of the prolongation of the period and the strength of negative regulation was also poor in the case of oriII. Although, the period reduced drastically in a seqA mutant, the corresponding relaxation of replication was modest (Figure 8). In oriII, the negative control is mediated primarily by limiting RctB, which apparently makes the contribution of sequestration to regulation less significant [19].
ChrII has many plasmid-like features including the organization of its origin. Plasmids generally initiate their replication randomly in the cell cycle and control it independently of the chromosome [16], [20], [31], [39]. Plasmid copy number can vary among individual cells due to replication error and unequal segregation. To maintain the mean copy number, plasmids adjust for fluctuations in copy number by replicating more in cells that receive fewer copies than the mean, and replicating less in cells with more copies than the mean. Thus, once-per-cell-cycle replication is not suited for the maintenance of plasmid copy number. We show here that unlike plasmids, chrII replicates once per cell cycle, like other bacterial chromosomes. The high density of GATC sites of oriII is not typical for plasmid origins but is a conserved feature of all sequenced strains of the family Vibrionaceae [18]. It appears that the involvement of methylation has rendered functioning of a plasmid-like origin similar to that of a chromosomal origin.
Why does initiation need to be cell-cycle specific for the chromosome? Completion of cell division demands that the septum forming area be cleared of DNA [40]. Plasmids are generally small and have correspondingly short replication elongation periods. Incompletely replicated plasmids are unlikely to cause steric hindrance to cell division for a significant period, unlike incompletely replicated chromosomes [41]. If chrII were to initiate replication randomly in the cell cycle like the plasmids, late-initiating chrII would likely delay cell division and create heterogeneity in cell generation times. V. cholerae ΔseqA cells did form elongated cells, indicative of a cell division defect (our unpublished results). One reason for this could be steric hindrance to cell division from late-initiating chrII. We suggest that a chromosome replicating from an origin with a plasmid provenance is subject to selection pressure to make the initiation cell-cycle specific, and the acquisition of methylation sites could allow that.
Understanding the role of methylation can also be important for another reason. It has been suggested that one of the common conspicuous features of the two origins being the high density of GATC sites, their methylation could be a mechanism to coordinate the replication between the two chromosomes [18]. Methylation is essential for the viability of bacteria with multiple chromosomes such as Rhizobium meliloti [42], Brucella abortus [43] and Agrobacterium tumefasciens [44] in addition to V. cholerae [17]. Although there is no evidence yet for direct communication among the chromosomes for replication initiation in any system, it is possible that in these bacteria methylation could be coordinating the replication to the cell cycle, as is does for V. cholerae and possibly other members of the family of Vibrionaceae.
Bacterial strains and plasmids used in this study are listed in Table 1 and Table 2, respectively. Primers are listed in Text S1. E. coli and V. cholerae were grown in LB (10 g tryptone +5 g yeast extract +5 g NaCl per liter, pH adjusted with NaOH to ∼7) or M63 medium (KH2PO4 3 g + K2HPO4 7 g + (NH4)2SO4 2 g + FeSO4 0.5 mg + MgSO4.7H2O 0.25 g, pH adjusted with KOH to ∼7) supplemented with 2 mM MgSO4, 0.1 mM CaCl2, 0.01% thiamine and 0.2% glucose, and additionally 0.1% casamino acids when desired. Antibiotics were used at the following concentrations: ampicillin, 100 µg/ml; chloramphenicol, 25 µg/ml for E. coli, 5 µg/ml for V. cholerae; erythromycin, 20 µg/ml; kanamycin, 25 µg/ml; spectinomycin, 50µg/ml; tetracycline, 15 µg/ml; and zeocin, 25 µg/ml. Diaminopimelic acid (DAP) was used at 0.8 mM, L-arabinose at 2 or 0.2 mg/ml, IPTG at 100 µM and thymidine at 0.3 mM.
To replace oriC (coordinates 3923756–3924022) with oriI (coordinates 2961130–364), the latter was amplified from DNA of CVC209 by PCR using primers GD113 and GD114. The PCR product was digested with EcoRI and BamHI, and ligated to similarly digested pEM7-Zeo. The resulting plasmid, pGD83, was digested with SacI and BamHI, and the fragment containing the oriI-zeo region was ligated to a similarly digested vector, pSW23, generating pGD79. The oriI-zeo region of pGD79 was amplified with primers GD124 and GD125, and the product used to replace oriC of CVC1394 by the mini-λ Red recombineering method [45]. The mini-λ prophage was eliminated from the strain by a 30°C to 42°C temperature shift. The resultant strain was called MG1655ΔoriC::oriI-zeo (CVC1400), and the replacement was confirmed by sequencing of the origin region. The genomic DNA of the dam mutant derivative (CVC1401) was confirmed for the absence of adenine methylation by its resistance to DpnI but not to MboI and BfuCI restriction enzymes (data not shown).
Cultures of E. coli were grown in LB to OD600≈0.2 and processed for flow cytometry after replication run-out in the presence of rifampicin (150 µg/ml) and cephalexin (10 µg/ml) for three hours as described [46]. The peak fluorescence intensity of an overnight grown E. coli culture in M63 + 0.2% glucose medium (without casamino acids) was taken to represent one genome equivalent.
A fragment with six 12-mers was obtained from pGD61 by digestion with XhoI and NotI.
Fragments with three 11-mers and a pair of 12- and 11-mers were obtained from pTVC86 and pTVC88, respectively, by digestion with XhoI and BamHI. For methylated and unmethylated fragments, the plasmids were from a dam+ (BR2699) and a dam− (CVC1060) strain, respectively. The fragments were gel-purified, dephosphorylated with Shrimp Alkaline Phosphatase (USB Corporation), and end-labeled with 50 µCi [γ-32P]ATP (PerkinElmer) by using 30 units of T4 polynucleotide kinase (New England Biolabs) and purified through ProbeQuant G-50 micro columns (GE Healthcare). To obtain hemimethylated DNA, oligonucleotide primers, TVC64 and TVC138 (Sigma-Genosys), were end-labeled and purified as above. The labeled primers were then used for PCR one at a time with methylated DNA as template for one cycle to obtain two populations of hemimethylated DNA, one with methylation on the top strand and the other on the bottom strand. The binding reactions were essentially as described [47].
A partial deletion of seqA was made by deleting codons 51 to 140 and substituting the deleted region with a zeocin cassette maintaining the seqA reading frame as follows. The seqA gene was amplified from CVC209 by PCR with primers GD87 and GD88. The product was digested with EcoRI and cloned in similarly digested vector, pSW4426T. The resultant plasmid was used as template for PCR with primers GD91 and GD92 to amplify the 5′ end of seqA, the plasmid backbone and the 3′end of seqA. After digestion with MfeI, a site of which was present within GD91 and GD92 primers, the PCR product was ligated to the zeocin cassette. The cassette was obtained from pEM7-Zeo by PCR, using primers GD89 and GD90 and digested with EcoRI before ligating to the MfeI fragment. The resulting plasmid, pGD70, containing the ΔseqAP::zeo allele was used to replace seqA of CVC209 by the allele-exchange method [48]. The resulting ΔseqAP::zeo mutant (CVC1410) grew slower than the WT. In LB at 37°C, the doubling times of the mutant was 32±2 min as opposed to 19±2 min for the WT. The ΔseqAP::zeo allele is called hereafter ΔseqAP.
The entire seqA ORF was also deleted and substituted with the zeocin cassette as follows. First, a kilobase region located downstream the stop codon of seqA was amplified by PCR with primers GD228 and GD229, the product digested with EcoRI and BamHI and cloned in a derivate of pEM7-Zeo (pGD111), previously digested with the same enzymes, generating pGD113. pGD111 is essentially same as pEM7-Zeo except that the multi-cloning site upstream of zeo is modified to include KpnI and NdeI restriction sites. Next, a kilobase region located upstream of the start codon of seqA was amplified by PCR with primers GD230 and GD231, the product digested with KpnI and NdeI and cloned in pGD113, previously digested with the same enzymes, generating pGD114. The flanking regions of seqA, now flanking the zeocin cassette, was amplified by PCR with primers GD257 and GD258 and the linear product was introduced by natural transformation in a hapR+ Δdns derivative of N16961 (CVC1121) essentially as described [49], [50]. The transformants were selected for zeocin resistance and checked for the replacement of the seqA gene by the zeocin cassette by PCR and DNA sequencing. The resulting ΔseqAT::zeo mutant (CVC2003) grew as slow as the ΔseqAP::zeo mutant with a doubling time of 32±2 min. The ΔseqAT::zeo allele is called hereafter ΔseqAT.
A complete deletion of the dam ORF and its substitution with a zeocin cassette was obtained by the allele-exchange method in the presence of a complementing plasmid, pGD93. The replication of the plasmid was thermo-sensitive and it carried the V. cholerae dam under the PBAD promoter. pGD93 was made as follows: the dam gene was amplified by PCR with primers GD72 and GD73, and the product after digestion with EcoRI and KpnI was cloned in pBAD24, previously digested with the same enzymes, generating pGD55. Next, the NdeI-HindIII fragment from pGD55 containing the dam gene was cloned in pKOBEGA, previously digested by NdeI and HindIII, generating the pGD93. For allele-exchange, a kilobase region located downstream of the stop codon of dam was amplified by PCR with primers GD261 and GD262, and the product after digestion with EcoRI and BamHI was cloned in a derivative of pEM7-zeo, previously digested with the same enzymes, generating pGD117. A 700 bp region located upstream of the start codon of dam was amplified by PCR with primers GD263 and GD264, and the product after digestion with KpnI and NdeI was cloned in pGD117, previously digested with the same enzymes, generating pGD118. The zeocin cassette with the flanking regions of dam was amplified by PCR with primers GD268 and GD269, and the product cloned as a blunt end fragment in pSW23, previously digested with SmaI, generating the pGD120. The plasmid was digested with SacI and SalI, and the fragment with the zeocin cassette was cloned into pDS132, previously digested also with the same enzymes. The resulting plasmid, pGD121, was used to replace dam of CVC209/pGD93. The resulting strain, CVC2023, was confirmed for the replacement of dam by the zeocin cassette by PCR and by DNA sequencing. To deplete Dam, single colonies grown in the presence of ampicillin (to select pGD93) and arabinose (to express dam) were used to inoculate LB without any drug but containing glucose (to repress dam expression) and the cultures were grown at 42°C (to stop plasmid replication).
Genomic DNA was isolated from cells of log phase cultures (OD600≈0.3), using the Genelute Bacterial Genomic DNA kit (Sigma). For analyzing chrI and E. coli DNA, 1 µg of DNA was digested 2 hours with 7.5 or 15 units of HphI (New England Biolabs) at 37°C, and the products resolved in a 1.5% agarose gel. For chrII, the conditions were similar except that TaqαI was used at 65°C. The origin probes were prepared by PCR using primers GD36 and GD37 for oriI, GD40 and GD41 for oriII, GD67 and GD68 for oriC, and GD150 and GD151 for ΔoriC::oriI. The primers for external markers on the three chromosomes were GD38 and GD39, GD42 and GD43, and GD128 and GD129, respectively. The probes for ori and the external markers were made radioactive using the RediPrimeII random primer labeling kit (GE Healthcare) and [α-32P] dCTP (PerkingElmer) and mixed separately for the two chromosomes. The band intensities were recorded and quantified as described earlier [46].
Marker frequency was determined by qPCR using a PTC-200 Peltier Thermal Cycler (MJ Research) and a LightCycler 480 SYBR Green I Master (Roche). Genomic DNA was prepared from log phase cultures in LB with Genelute Bacterial Genomic DNA kit (Sigma), and 313 pg was used in each reaction as template. The primers were used at 0.3 µM each. They were proximal to either oriI (GD136 and GD137) or oriII (GD156 and GD157) or terI (GD142 and GD143) or terII (GD140 and GD141) region of the two chromosomes, and were identical to those described [39]. The primer pairs were such that they produced ∼100 to 130 bp fragments in all cases. Cp (crossing point) values were determined and used for calculating the oriI/oriII, oriI/terI, oriII/terII and terI/terII ratios. The ratios were normalized to those of a culture grown to stationary phase in supplemented M63 medium (without casamino acids). Mean ratios were obtained from DNA prepared from three cultures, each grown from independent colonies, and each DNA was analyzed in triplicate.
The method was modified from the one described by Lin and Grossman [51]. Briefly, cultures at OD600nm = 0.3 were treated with 1% formaldehyde at room temperature for 30 min. After cell lysis and sonication, RctB complexes were precipitated with antibody against RctB (IP DNA) and Dynabeads-Protein G magnetic beads (Invitrogen), followed by stringent washings (see Text S1 for the detailed ChIP protocol). After reversal of the cross-links by incubation at 65°C overnight, the samples were treated by protease K (Sigma) and then purified with a PCR purification Kit (Qiagen). To quantify the enrichment of RctB binding sites in the IP DNA, 5 µl of 1∶100 dilution of the IP DNA was used to perform locus-specific real-time qPCR with primers GD218 and GD219, specific to the vector backbone of the plasmid carrying the RctB binding sites, and primers GD191 and GD192, specific to a gene in the E. coli genome that served as a reference, as described [52].
|
10.1371/journal.pntd.0004568 | Cost-Effectiveness of Antivenoms for Snakebite Envenoming in 16 Countries in West Africa | Snakebite poisoning is a significant medical problem in agricultural societies in Sub Saharan Africa. Antivenom (AV) is the standard treatment, and we assessed the cost-effectiveness of making it available in 16 countries in West Africa.
We determined the cost-effectiveness of AV based on a decision-tree model from a public payer perspective. Specific AVs included in the model were Antivipmyn, FAV Afrique, EchiTab-G and EchiTab-Plus. We derived inputs from the literature which included: type of snakes causing bites (carpet viper (Echis species)/non-carpet viper), AV effectiveness against death, mortality without AV, probability of Early Adverse Reactions (EAR), likelihood of death from EAR, average age at envenomation in years, anticipated remaining life span and likelihood of amputation. Costs incurred by the victims include: costs of confirming and evaluating envenomation, AV acquisition, routine care, AV transportation logistics, hospital admission and related transportation costs, management of AV EAR compared to the alternative of free snakebite care with ineffective or no AV. Incremental Cost Effectiveness Ratios (ICERs) were assessed as the cost per death averted and the cost per Disability-Adjusted-Life-Years (DALY) averted. Probabilistic Sensitivity Analyses (PSA) using Monte Carlo simulations were used to obtain 95% Confidence Intervals of ICERs.
The cost/death averted for the 16 countries of interest ranged from $1,997 in Guinea Bissau to $6,205 for Liberia and Sierra Leone. The cost/DALY averted ranged from $83 (95% Confidence Interval: $36-$240) for Benin Republic to $281 ($159–457) for Sierra-Leone. In all cases, the base-case cost/DALY averted estimate fell below the commonly accepted threshold of one time per capita GDP, suggesting that AV is highly cost-effective for the treatment of snakebite in all 16 WA countries. The findings were consistent even with variations of inputs in 1—way sensitivity analyses. In addition, the PSA showed that in the majority of iterations ranging from 97.3% in Liberia to 100% in Cameroun, Guinea Bissau, Mali, Nigeria and Senegal, our model results yielded an ICER that fell below the threshold of one time per capita GDP, thus, indicating a high degree of confidence in our results.
Therapy for SBE with AV in countries of WA is highly cost-effective at commonly accepted thresholds. Broadening access to effective AVs in rural communities in West Africa is a priority.
| Antivenom is the main intervention against snakebite poisoning but is relatively scarce, unaffordable and the situation has been compounded further by the recent cessation of production of effective antivenoms and marketing of inappropriate products. Given this crisis, we assessed the cost effectiveness of providing antivenoms in West Africa by comparing costs associated with antivenom treatment against their health benefits in decreasing mortality. In the most comprehensive analyses ever conducted, it was observed the incremental cost effectiveness ratio of providing antivenom ranged from $1,997 in Guinea Bissau to $6,205 for Liberia and Sierra-Leone per death averted while cost per Disability Adjusted Life Year (DALY) averted ranged from $83 for Benin Republic to $281 for Sierra-Leone. There is probability of 97.3–100% that antivenoms are very cost-effective in the analyses. These demonstrate antivenom is highly cost-effective and compares favorably to other commonly funded healthcare interventions. Providing and broadening antivenom access throughout areas at risk in rural West Africa should be prioritized given the considerable reduction in deaths and DALYs that could be derived at a relatively small cost.
| Snakebite poisoning is a significant cause of death and disability in rural West Africa [1,2,3,4,5,6,7]. The exact burden of snakebite is difficult to ascertain and is often undereported. A study by Jean-Philippe Chippaux reported an estimate of over 314, 000 envenomations, 7300 mortality and nearly 6000 amputations occurring yearly in sub-Saharan Africa (SSA) [7]. However, even in West Africa alone, a range of 1504 to 18,654 annual mortality from snakebite envenoming has been made [8]. This is further compounded by the variability in snakebite incidence with estimates of as high as 500 bites per 100,000 persons per year in parts of northern Nigeria [9].
Vipers (Echis ocellatus, E. leucogaster and E. jogeri) are a major cause of snakebite envenoming throughout the sub-region mainly in Benin republic, Burkina Faso, Cameroun, Chad, Gambia, Ghana, Mali, Niger, Nigeria, Togo and Senegal [1,2,3,4,5,6,7]. In the sub-region, envenoming from snakes other than vipers mostly results from African spitting cobras (Naja nigricollis, N. katiensis), puff-adder (Bitis arietans), mambas (Dendroaspis viridis, D. polylepis), burrowing asps or stiletto snakes (Atractaspis species), night adders (Causus maculatus, C. rhombeatus, C. resimus, C. lichtensteinii) and very rarely boomslang (Dispholidus typus). Joger’s carpet viper (E. jogeri) is confined to Mali. Romane’s carpet viper (Echis leucogaster) and Egyptian cobras (Naja haje and N. senegalensis) are causes of snakebite envenoming in the Sahelian and drier parts of West Africa while the forest cobra (Naja melanoleuca) and the Gaboon viper (Bitis gabonica) cause occasional bites in the rain forest and South-eastern parts of the sub-region [1,5,7].
In West Africa, carpet vipers may account for as many as two thirds of all snakebite envenoming although their range is limited to the savannah region [1,9,10,11]. Envenoming from carpet vipers leads to swelling and tissue damage at the site of bite, local and systematic bleeding, anaemia and shock. Often death results from cerebral haemorrhage, bleeding elsewhere or haemorrhagic shock [1,10,11]. The bleeding abnormality results from a prothrombin activating metalloprotease “Ecarin” and a FX activating component, an anticoagulant, platelet activator/inhibitor and haemorrhagins in the snake’s venom [1,10,11]. Non-clotting blood detected by the 20minute Whole Blood Clotting Test [20WBCT] virtually confirms carpet viper envenoming in the northern third of Africa (roughly north of the equator) and is utilized to assess adequacy of treatment [1,10,11]. Most non-carpet viper bites lead to local swelling and tissue damage. The colubrids, boomslangs and twig snake (Thelotornis kirtlandii), are back fanged snakes that rarely envenom but can cause severe bleeding and acute kidney injury. Neurotoxic features may result from Naja haje, Naja melanoleuca and Dendroaspis spp bites with deaths often resulting from respiratory muscle paralysis [12]. The risk of death from snakebites other than viper envenoming is lower [9,13,14,15,16,17], but cobra spits may lead to blindness and bites to cancerous ulcers, abortions, scarring, arthrodeses, contractures and psychological impairment leading to permanent disability and productivity loss following hospitalization and incapacitation [7,18,19,20,21]. Cessation of bleeding abnormalities and restoration of clotting following administration of effective antivenom usually occurs promptly in carpet viper envenoming. Antivenom is efficacious in decreasing the likelihood of dying and is the main treatment for snakebite envenoming [1,11,22,23]. However, its administration is associated with early adverse reactions (EAR) which rarely results in fatality.[24,25,26]. Specific interventions may be required to either prevent EAR with administration of premedication prior to antivenom or to treat it once developed following antivenom administration [25,26]. Antivenoms are formulated as either liquid agents that needs to conveyed and stored at low temperature with a life span of about three years [27,28] or as freeze dried substances that are more stable with extended shelf life. Both types of formulations have been produced for the sub-region [6,27,28]. The average cost per treatment of antivenom was reported as US$124 (range US$55–$640) although a median price of US$153 was also reported for Sub-Saharan Africa [29,30,31]. The few effective antivenoms in the sub-region generally have been scarce, locally unaffordable and inaccessible where they are most needed. Partly for these reasons antivenom utilization has drastically declined to a very small fraction of indicated need. The situation has been compounded further by the recent announcement by Sanofi-Pasteur that production and distribution of FAV Afrique, currently the most widely distributed and most dependable antivenom in the sub-region, will be discontinued by 2016. Its loss will exacerbate an already serious public health crisis and makes the management of snakebite even more challenging [32]. It is therefore extremely important within the context of other competing public health priorities to assess the health economics of antivenoms to guide policy. Before the recent publication of our work focusing on Nigeria [33], few economic evaluations of preliminary nature had been conducted on antivenoms [34,35]. Here, we evaluated the cost-effectiveness of antivenom utility in the treatment of snakebite envenomation by computing incremental cost effectiveness ratios (ICERs) of the cost per death averted and the cost per DALY averted by adapting a previously published model for Nigeria to 16 countries in WA. We performed the analysis from healthcare system perspective to provide policy makers with evidence towards broadening access to antivenoms given their importance in preventing loss of lives and limbs among poor vulnerable communities in West Africa.
A decision analytic model (Fig 1) was adapted to estimate health outcomes and costs associated with the availability and use of geographically appropriate and effective antivenoms for snakebite poisoning in West Africa [33]. Details of the model structure are described elsewhere [33]. Briefly, the model assessed the availability of effective antivenoms relative to no availability in the decision node. The model differentiated snakebite envenoming by carpet viper and non-carpet viper and distinction was made on the basis of the 20WBCT in the treatment arm of the model. Evidence of coagulopathy would lead to the administration of mono-specific antivenom that neutralizes carpet viper venom only, whereas absence of coagulopathy triggers the administration of a polyspecific antivenom that neutralizes venoms from several snakes, including the carpet viper. In the first chance node, the model included EARs associated with antivenom use, which are more likely to occur with polyspecific rather than the monospecific antivenom [23,27,28,36,37]. Symptoms of EAR were diverse and death could happen in about 1% of cases [24,25,26]. Survivors of snakebite may recover completely or remain with significant sequaela (e.g. amputation) that was considered in the model. Treatment outcomes were converted into DALYs on the basis of local life expectancy. Tree Age Pro Suite Healthcare 2014 software was used for analyses.
The cost/death averted for the 16 countries of interest varied. It was as low as $1,997 in Guinea Bissau to as high as $6,205 in Liberia and Sierra Leone. The cost/DALY averted ranged from a low of $83 (95% Confidence Interval: $36-$240) for Benin Republic to a high of $281 ($159–457) for Sierra-Leone. In all cases, the base-case cost/DALY averted estimate fell below the commonly accepted threshold of one time per capita GDP, suggesting that AV is highly cost-effective for the treatment of snakebite in all 16 WA countries [51,52].
The findings from the analyses were also consistent to variations of inputs in 1-way sensitivity and scenario analyses as depicted (Table 3 and Fig 2). The individual countries’ model results were most sensitive to effectiveness of antivenom in decreasing mortality, natural (unattended) mortality, costs of antivenoms and types of snake causing envenoming (Fig 2). Results were not sensitive to antivenom associated EAR or the cost of managing it. Varying the cost of antivenom from $125 to two times for victims who may require two doses, i.e. $306, still yielded ICER estimates that remain cost-effective. The ICERs rose when the frequency of snakebite envenomation due to saw-scaled viper was reduced to 0% except in Benin and Guinea Conakry where Antivipmyn antivenom is used and is effective even against elapids (Table 3) [42].
Moreover, the ICER ranged from $97.26 in Benin to high levels of $13,964.26 in Liberia and $15,278.99 in Sierra Leone even in the worst case scenario where (poly-specific) antivenoms have nil effectiveness (0%) against bites from snakes other than carpet viper. These estimates fall outside the cost-effectiveness thresholds in Liberia and Sierra Leone largely because non-carpet viper accounts for 99% of SBE. Applying a modest reduction of 40% on the probability of EAR with the use of adrenaline premedication [25,26,33] gave a cost per DALY averted slightly lower than base-case ICERs. Similarly, the ICERs were only very slightly altered even when more serious or more frequent disabilities were substituted in the model. This was demonstrated with venom-induced-blindness (0.01%) or Post-Traumatic-Stress-Disorder (20%) with disability-weights of 0.552 and 0.105 respectively [18,21,48].
Furthermore, our PSA confirms the model findings remain consistent to concurrent variation of all model inputs, as the ICERs with their respective 95% confidence limits are far less than the cost-effectiveness thresholds (Table 3). It showed that in majority of simulations (97.3% in Liberia to 100% in Cameroun, Guinea Bissau, Mali, Nigeria and Senegal (Fig 3)) our model results yielded an ICER that fell below the threshold of one time per capita GDP, thus, indicating a high degree of confidence in our results [51,52].
Economic modeling is very useful in determining the best ways to utilize resources to optimally manage medical conditions where there are competing priorities and limited resources [49]. This is the first extensive assessment of the cost-effectiveness of expanding antivenom access in the 16 countries in West Africa. We find that the cost/death averted for the 16 countries of interest ranged from $1,997 in Guinea Bissau to $6,205 for Liberia and Sierra Leone. The cost/DALY averted ranged from $83 (95% Confidence Interval: $36-$240) for Benin Republic to $281 ($159–457) for Sierra-Leone. The ICER point estimate is <$100/DALY averted in 7 countries, $100-$200/DALY averted in 7 countries and <$300/DALY averted in 2 countries. The results show that snakebite antivenoms are highly cost-effective in West Africa, as our findings are far less than the one time per capita GDP threshold [51,52]. While it will be worthwhile to repeat the analysis for similar geographic and socioeconomic settings, most of the model inputs such as the antivenom efficacy and cost would largely be similar across many African countries with similar GDPs where snakebite envenoming occurs. An exception may be the prevalence of carpet viper (Echis ocellatus) envenoming, a snake that is confined to West Africa extending eastwards only as far as Chad and Central Africa Republic. However, even in areas without carpet viper bites (0%) our results demonstrate that antivenoms remain highly cost-effective.
The combination of effective and relatively inexpensive antivenoms and the utilization of a cheap, dependable and simple point-of-care test have been instrumental to our results. Antivenom effectiveness is loosely inversely related to the ICER per DALY saved (Fig 2) [33]. With discontinuation of production of geographically appropriate effective antivenoms and marketing of inappropriate ineffective products the cost per DALY saved will substantially soar [32,39,40]. The 20WBCT is discriminatory and will be useful following envenoming from other snakes in the rest of Africa that result in coagulopathy e.g., other species of carpet viper (Echis leucogaster, Echis pyramidum, Echis jogeri, Echis coloratus) and boomslang. In the context of differing circumstances where multiple types of snakes with varying manifestations of envenoming or lack of reliable cheap differentiating test, antivenoms may not be as cost-effective. Nevertheless, where patients come along with the dead snakes, decision on antivenom choice is feasible [23] and in such cases the differentiating test becomes not so useful. In the face of competing health needs and constrained-resources, it would be helpful to contextualize our findings in the light of other healthcare initiatives. The cost-effectiveness of Human Immunodeficiency Virus treatment used in similar resource-constrained settings as first-line, second-line or for protecting negative partners among discordant partners ranged from US$530 to US$1037 per year of life gained [33,55,56,57]. However, these estimates are still higher than the highest ICERs obtained for two countries in this analysis, i.e., Liberia and Sierra-Leone with $257/DALY averted and $281/DALY averted respectively. The antivenom cost effectiveness is comparable to what obtains in other healthcare programmes. For example, the cost/DALY averted obtained in this study ranged from $100 to $200 for 7 countries (see Table 3) and is comparable to the cost effectiveness of rotavirus vaccines in other developing countries in Africa and Asia [33,58,59]. Similarly, the cost/DALY averted in the remaining 7 countries was <$100/DALY averted and is similar to what obtains for preventing Human Papilloma Virus and pneumococcal infections with vaccines in West Africa [33,60,61].
We estimated a cost/DALY averted ranging from $73 in Benin to $247 in Sierra Leone if the cost input of antivenom is reduced to $125 per dose as obtained in Mali [30]. Doubling the cost of antivenom for patients requiring more than a dose still yielded ICERs that remain cost effective ranging from $136/DALY averted in Benin to $464/DALY averted in Sierra Leone.
In 8 of the 16 countries with neither indigenous antivenom effectiveness data nor data from adjacent countries, we used efficacy derived from a meta-analysis with data inputs from other countries in the sub-region [22, 38].
The study has a number of limitations. First, the effectiveness of antivenom was derived from observational studies rather than RCTs. Definitive placebo controlled trials of antivenom are considered unethical and RCTs are unlikely to be conducted in the absence of a suitable comparator to antivenom. However, to reduce bias, improve data quality and validity of estimates, three investigators independently searched both English and French literature and extracted data using a checklist for consistency. Secondly, we applied the estimated protection conferred by antivenoms against carpet viper envenoming to the antivenoms used for other than carpet viper envenoming (except in Benin and Guinea Conakry with specific estimates), though this assumption was subsequently dropped in a scenario analysis where antivenoms were assumed to be ineffective (0%) against non-carpet viper envenoming. Thirdly, the analysis mainly considered amputation as the major disability to the exclusion of other anecdotal complications [7, 20, 43, 44]. Fourthly, in our model, antivenoms only conferred protection against death—being a more objective and valid outcome. Fifthly, other benefits of antivenom (e.g. speedier recovery) were not included in the model. Sixth, we used separate EAR risk inputs for EchiTab-G and EchiTab-Plus respectively [23] but could not discern their respective efficacy against mortality so the combined estimate of the two was used in Burkina Faso and Nigeria. Lastly, we did not include the costs incurred for logistics in conveying and preserving antivenoms as we assumed facilities for existing health programmes will be utilized.
The findings from the cost effectiveness analysis demonstrate that providing and broadening antivenom access throughout areas at risk in rural West Africa should be prioritized given the considerable reduction in deaths and disabilities that could be derived at a relatively small cost.
|
10.1371/journal.pgen.1000151 | Off-Target Effects of Psychoactive Drugs Revealed by Genome-Wide Assays in Yeast | To better understand off-target effects of widely prescribed psychoactive drugs, we performed a comprehensive series of chemogenomic screens using the budding yeast Saccharomyces cerevisiae as a model system. Because the known human targets of these drugs do not exist in yeast, we could employ the yeast gene deletion collections and parallel fitness profiling to explore potential off-target effects in a genome-wide manner. Among 214 tested, documented psychoactive drugs, we identified 81 compounds that inhibited wild-type yeast growth and were thus selected for genome-wide fitness profiling. Many of these drugs had a propensity to affect multiple cellular functions. The sensitivity profiles of half of the analyzed drugs were enriched for core cellular processes such as secretion, protein folding, RNA processing, and chromatin structure. Interestingly, fluoxetine (Prozac) interfered with establishment of cell polarity, cyproheptadine (Periactin) targeted essential genes with chromatin-remodeling roles, while paroxetine (Paxil) interfered with essential RNA metabolism genes, suggesting potential secondary drug targets. We also found that the more recently developed atypical antipsychotic clozapine (Clozaril) had no fewer off-target effects in yeast than the typical antipsychotics haloperidol (Haldol) and pimozide (Orap). Our results suggest that model organism pharmacogenetic studies provide a rational foundation for understanding the off-target effects of clinically important psychoactive agents and suggest a rational means both for devising compound derivatives with fewer side effects and for tailoring drug treatment to individual patient genotypes.
| Neuropsychiatric disorders such as depression and psychosis affect one-quarter of all individuals during their lifetime, and despite efforts to improve the selectivity of psychoactive drugs, all are associated with side effects. Drug efficacy and tolerance are known to be linked to an individual's genetic profile, but little is known about the nature of this correlation due, in part, to the current emphasis on screening compounds against targets in vitro. Here we present a comprehensive, genome-wide effort to understand drug effects on the cellular level using an unbiased genome-wide assay to determine the importance of every yeast gene for tolerance to 81 psychoactive drugs. We found that these medications perturbed many evolutionarily conserved genes and cellular pathways, such as those required for vesicle transport, establishment of cell polarity, and chromosome biology. The 500,000 drug–gene measurements obtained in this study increase our understanding of the mechanism of action of psychoactive drugs. Specifically, this study provides a framework to assess the next generation of psychoactive agents and to guide personalized medicine approaches that associate genotype and phenotype.
| Neuropsychiatric disorders will effect 25% of all individuals at some point in their lives, with devastating social and economic consequences [1]. This constellation of diseases encompasses schizophrenia, depression, age-related memory and cognition decline, and the degeneration of neuromuscular function. Most prescribed psychoactive drugs are thought to primarily target neurotransmission pathways in the central nervous system, and thereby cause changes in perception, mood, consciousness, and behavior. Many of these therapeutics have been developed using in vitro assays and, as such, may have other unknown targets and unanticipated cellular effects in vivo. For example, side effects of antipsychotic drugs include tremors, hypotension, impotence, lethargy, and seizures [2]. In an effort to improve efficacy and to reduce side effects, new generations of drugs have been developed; among these are the so-called atypical antipsychotics such as clozapine. While clozapine is linked to a reduced risk of neuromuscular side effects, it is associated with new side effects such as life-threatening agranulocytosis in up to 1% of patients [3], and, less frequently, fatal myocarditis [4]–[6]. As such, the therapeutic benefit of this and other new atypical drugs remains open to debate. For example, a comprehensive meta-regression analysis that compared both typical and atypical drugs concluded that atypical antipsychotics were neither more effective nor better tolerated than conventional agents [7]. Other classes of psychoactive drugs, such as the antidepressants, also cause numerous undesirable side effects and the broad usage of these medications have been questioned [8].
Surrogate genetics is an effective approach to interrogate heterologous gene function or drug mechanism of action using simpler model organisms [9],[10]. The budding yeast Saccharomyces cerevisae has previously been used to help elucidate the basis of some psychiatric disorders [11]–[18]. For example, the expression in yeast of mutant and wildtype forms of the Huntington's disease gene revealed important factors regulating the toxicity of protein aggregates [11],[15],[19], and a genome-wide suppressor screen in yeast uncovered kynurenine 3-monooxygenase as a potential new therapeutic target for the treatment of Huntington's disease [13]. In other studies, expression in yeast of the alpha-synclein gene associated with Parkinson's disease yielded a network of interacting genes that modulate cellular toxicity [11],[15],[19].
Recently, the genome-wide collection of yeast gene deletion strains has been used to generate genetic profiles of drug sensitivity and resistance [20]–[26]. These profiles have uncovered unexpected mechanisms of action for well-known drugs, such as for the anti-metabolite 5-fluorouracil in perturbation of rRNA processing [22],[25] and for the anti-cancer agent tamoxifen in calcium homeostasis [26].
To better understand potential off-target effects of FDA-approved psychoactive drugs and their analogs, we profiled 214 psychoactive compounds in quantitative wildtype yeast growth assays and generated genome-wide deletion sensitivity profiles for the 81 drugs that caused overt growth defects. The sensitivity profiles for 49 of these drugs were overrepresented for core cellular functions such as chromatin organization, establishment of cell polarity, and membrane organization and biogenesis. Our results provide a rational foundation for personalized drug approaches and for understanding unwanted side effects in clinically important psychoactive agents.
To ask if psychoactive compounds can inhibit wildtype budding yeast growth, we challenged yeast with 76 high-purity psychoactives representing 16 ligand categories that encompass a broad spectrum of treatments for neurological disorders (see Figure 1 for workflow and Table S1 for drug information). Despite the fact that yeast lacks the established neuronal targets of these compounds, 17/76 (22%) drugs inhibited the growth of wildtype yeast (when tested at 200 µM) and are hereafter referred to as “bioactive”. This observation shows that in addition to their reported targets, many of these compounds also have secondary mechanisms of action. In fact, over half of the 16 tested ligand classes included compounds that were bioactive (Figure 2A). Among these, serotonin uptake inhibitors were most effective; four of five tested molecules in this class inhibited yeast growth (Figure 2A). Because our assay depends on growth inhibition in order to observe any effects on specific deletion strains, we proceeded with the 17 bioactive compounds and determined a drug dose that inhibited wildtype growth by ∼15% (Figure 2B, Table S1). In our previous genome-wide studies this level of inhibition best captured the ability to identify the known drug target while minimizing the number of generally sensitive strains [22],[23]. Applying this drug dose, we subjected the bioactive compounds to genome-wide parallel fitness profiling. In this technique, pools of deletion strains are grown competitively for several generations in the presence of a sub-lethal concentration of drug, and genomic DNA is extracted. After PCR-amplification of the unique molecular barcodes incorporated into each gene deletion cassette, the relative role of each gene for growth in the presence of drug is determined by hybridization of the PCR products to a DNA microarray carrying the barcode complements [27]–[29]. The relative abundance of sequence tags in the drug experiments is compared to control experiments and fitness ratios and z-scores are calculated (see Materials and Methods). We used two pools of diploid strains: i) heterozygous deletion strains deleted for one copy of the essential genes (1158 strains), which often identifies compound targets through HaploInsufficiency Profiling (HIP) [22],[30], and ii) homozygous deletion strains deleted for both copies of non-essential genes (4768 strains); this HOmozygous Profiling (HOP) assay identifies genes that buffer the drug target pathway [24].
Using this combination of the HIP and HOP assays we found that only a few deletion strains (∼5) exhibited significant sensitivity to most of the 17 bioactive compounds (Figure 2C). In contrast, several deletion strains (∼50) were scored as sensitive for the α1-adrenoceptor antagonist SR 59230A and the three selective serotonin re-uptake inhibitors fluoxetine (Prozac), clomipramine, and fluvoxamine (Figure 2C). Given this unexpected potency of the serotonergic drugs in our yeast assays, we extended our investigation to encompass pharmacologically related agents and screened two commercially available drug libraries encompassing 95 serotonergic and 55 dopaminergic compounds. These drug libraries contained the four FDA-approved serotonergics sertraline (Zoloft), fluoxetine (Prozac), paroxetine (Paxil), and cyproheptadine (Periactin), and the four FDA-approved dopaminergics bromocriptine (Parlodel), clozapine (Clozaril), haloperidol (Haldol), and pimozide (Orap). Based on our initial results, we anticipated a high rate of bioactivity on yeast for these two drug classes. Indeed, 66/150 (44%) of the serotonergic and dopaminergic drugs were bioactive, a significant difference compared to the 22% of the initially screened drugs that represented the 16 different ligand sets (p<10−7).
The high prevalence of bioactivity in yeast prompted us to ask if any particular psychoactive drug attribute correlated with the ability of these compounds to inhibit wildtype yeast growth. We first performed structural clustering of all ∼220 screened psychoactive compounds using chemical fingerprints in Pipeline Pilot (Accelyrs, San Diego). As more than half of the resulting clusters contained both active and inactive drugs, chemical structure was not predictive of drug action on wildtype yeast growth for this selection of compounds (data not shown). We next asked if any physiochemical properties, as predicted from the structures, were linked to drug activity. The parameters we tested included the number of H-bond donors and acceptors, molecular weight, and hydrophobicity as measured by AlogP (the octanol-water partition coefficient). These measures are important descriptors used in the empirical parameter set known as Lipinski's Rule of Five [31]. In addition to the Lipinski descriptors, we tested six other parameters relevant to drug activity: van der Waals surface area, molecular surface area, molecular solubility, logD (the octanol-water distribution coefficient; a combination of logP and pKa), number of rings and number of rotatable bonds. Principal component analysis revealed that a partition coefficient of AlogP>3 was best able to predict drug activity (p<4.9e-13, for details see Materials and Methods) as shown in Figure 3. A molecular weight of >260g/mole was also indicative of an active compound (p<3.4e-05, Figure 3). If there is a correlation between human side-effects and conserved cellular pathways scored using our surrogate yeast system, it is possible that an additional study could help predict such effects based on structural features.
To systematically interrogate compound mechanisms of action, we subjected the 66 bioactive serotonergic and dopaminergic compounds to genome-wide fitness assays using the approach described above (Figure 1). Combined with the initial set of 17 bioactive drugs, we screened a total of 81 unique drugs (two drugs occurred in duplicate in the chemical libraries), eight of which are used therapeutically (Table 1). Fitness ratios and z-scores for all deletion strains are provided in Tables S2 and S3, respectively (raw data are available at ArrayExpress, EMBL-EBI, accession number E-MTAB-34). The genome-wide fitness profiles were reproducible as the average correlation coefficient for the five replicated compounds was 0.83, which is similar to the average correlation coefficient of 0.72 reported in a previous large-scale fitness study [23]. As an unbiased control, we calculated the average correlation coefficient between all possible random drug pairs in our assay. As expected, this value (0.44) was lower than the average correlation coefficient for duplicates, but well above the previously noted average correlation of zero for unrelated compounds (Maureen Hillenmeyer, unpublished data). In agreement with this, two-dimensional hierarchical clustering [32] did not separate the dopaminergic and serotonergic profiles into two distinct groups, but clearly separated drugs from these two classes from most other compounds profiled (Figure 4). Further indicating the general similarities between dopaminergic and serotonergic drugs in our yeast screen, 25% of the significantly sensitive strains (r>2, z>3, see Materials and Methods) scored in both drug categories (Table S4).
To ask which cellular functions and pathways were required for resistance to the tested drugs, we performed functional enrichment tests using Gene Ontology (GO) annotations specifically focusing on sensitive strains in the i) essential heterozygous, ii) homozygous or iii) both collections (see Materials and Methods). 32 drug sensitivity profiles were not enriched for any GO Process but the remaining 49 profiled drugs (60.5%) interfered with 106 different processes (multiple-testing corrected p-value<0.0001, Table S5). For visual clarity, we collapsed these 106 processes down to 22 (Table S6). The drug sensitivity profiles obtained with the combined set of heterozygous and homozygous strains were enriched for the highest number of condensed GO processes (119 processes, purple color in Figure 5), while 12 processes were uniquely enriched among sensitive homozygous deletion strains (blue color in Figure 5). These processes likely reflect drug detoxification mechanisms (e.g. “vesicle transport” and “response to drug”) or other processes required for resistance to compound by an unknown mechanism (e.g. “amino acid biosynthesis and metabolism”). Two processes were uniquely scored for essential genes (red color in Figure 5) and are further discussed below.
Investigating the general nature of our enrichment profiles, we found that the most frequently enriched processes across all drugs and genetic backgrounds were vesicle transport, protein localization, and telomere biology (Figure 5). Genes functioning in cell morphogenesis, establishment of cell polarity, cell cycle, amino acid biosynthesis, chromatin organization, RNA metabolism, and membrane organization were also needed for resistance to several (>5) of the psychoactive drugs. A few GO Processes were unique to a single drug: protein glycosylation (A77636), methylation (SB 216641), cell wall organization and biogenesis (GR 127935), and membrane lipid metabolic process (pimozide). In the subsequent sections we focus on the analysis of the FDA-approved drugs and summarize the most notable enrichments for these drugs in Table 2. First, we discuss identified buffering pathways and drug detoxification mechanisms. Next, we concentrate on potential new drug targets identified for the therapeutically used psychoactive drugs.
Vesicle transport was the most commonly overrepresented process among genes required for resistance to psychoactive drugs (Figure 5) suggesting that uncompromised vesicle transport function is a general requirement for psychoactive drug detoxification. The enrichment of cellular transport genes was especially pronounced in response to clozapine treatment, where 9 of the 10 most required genes belonged to this category (Table 3). Protein sorting and localization accounted for the second most frequently enriched process (Figure 5). Deletion of vesicle trafficking and protein localization genes often resulted in very severe phenotypes (bright yellow in Figure 4). Gene products with protein localization roles include those involved in selecting cargo proteins for endosome-to-Golgi retrieval (e.g. Vps29), and those involved in sorting proteins in the vacuole (e.g. Pep8). Interestingly, the fitness profiles obtained with certain vesicle transport and protein localization deletions clustered with those obtained with strains deleted for genes functioning in actin filament organization/stabilization (arc18Δ, tpm1Δ, vrp1Δ,), mRNA degradation (lsm1Δ), and stabilization of membrane amino acid transporters (npr1Δ) (Figure 4, left text panel). A second, large group of strains mainly deleted for genes functioning in vesicle transport and protein localization exhibited similar phenotypes across the 81 drugs as ckb1Δ and ckb2Δ, which are deleted for genes functioning in regulation of transcription and mitotic cell cycle (Figure 4, right panel).
Most of the drug sensitivity profiles were enriched for both protein localization and telomere biology (Figure 5). The apparent “linking” of these enrichments could be attributed to genes that are, in fact, involved in both these processes. Examples of such genes function in the three Endosomal Sorting Complexes Required for Transport, more specifically in ESCRT I (VPS28, STP22), in ESCRT II (SNF8 and VPS25), and in ESCRT III (SNF7). These genes are, in addition, associated with telomere defects [33],[34].
Because the more recently developed atypical antipsychotic drugs are still associated with side effects and their benefits are currently debated, we compared the phenotypic profiles of the atypical antipsychotic clozapine to two traditional antipsychotics, reasoning that if atypical drugs are more specific, they would exhibit fewer off-target effects in yeast. In contrast to this expectation, the atypical antipsychotic clozapine exhibited a similar number of significantly sensitive (r>2, z>3, see Materials and Methods) deletion strains (26) as the typical antipsychotic drugs pimozide (29) and haloperidol (20). Comparing the fitness profiles of clozapine with the typical antipsychotics pimozide and haloperidol, we found that each drug was associated with unique functional enrichment profiles: clozapine for telomere biology and protein localization, pimozide for membrane lipid metabolic processes, and haloperidol for aromatic amino acid biosynthesis and metabolism (Figure 5). In contrast, vesicle transport was enriched in all three drug sensitivity profiles. The more detailed GO processes behind the condensed process vesicle transport were vesicle-mediated transport for all three drugs and, in addition, secretory pathway, secretion, post-Golgi vesicle-mediated transport and Golgi vesicle transport for haloperidol and clozapine (Tables S5 and S6). The distinct fitness profiles are consistent with the structural differences that exist between these drugs (Figure S1). For example, clozapine has substructures (piperazine and diazepine) that do not exist in pimozide and haloperidol, and haloperidol contains two benzene rings while pimozide has three.
Compared to the other investigated therapeutics, the fitness profile in the anti-Parkinson drug bromocriptine pointed to a single potential off-target mechanism of action for this drug. The only overrepresented function among sensitive strains was amino acid biosynthesis and metabolism (Figure 5) and the most sensitive strains were deleted for the aromatic biosynthesis genes TRP3, TRP4, TRP1, ARO1, TRP2, and ARO2. In addition to bromocriptine, six other dopaminergic drugs also interfered with amino acid biosynthesis and metabolism (Figure 5). The sensitivity profiles of all these seven drugs shared the enrichment for the detailed GO process aromatic compound metabolic process (Tables S5 and S6) due to the sensitive phenotype of 13 strains in total. Among them, strains deleted for TRP1, TRP2, TRP3, TRP4, TRP5, ARO2, and ARO3 were scored in all 7 drugs and strains deleted for ARO1 and ARO7 in 6 drugs. Besides the notable enrichment for genes involved in aromatic compound metabolism, the sensitivity of strains missing other genes also contributed to the observed GO process enrichment. Such genes included the folic acid (vitamin B9) biosynthesis gene FOL2, the panthothenate (vitamin B5, precursor of coenzyme A) biosynthesis gene FMS1, and the protein kinase GCN2, which induces amino acid biosynthesis genes in yeast in response to starvation and, in addition, restricts intake of diet lacking essential amino acids in rats [35].
The sensitivity profile of the typical antipsychotic pimozide showed a unique enrichment for membrane lipid metabolic processes not seen for any of the other 80 profiled drugs (Figure 5). In pimozide, the MCD4-deletion strain had the strongest phenotype and was 21-fold depleted compared to the control (Table 3). MCD4 is highly conserved among eukaryotes and functions in glycosyl-phosphatidylinositol (GPI) anchor synthesis. Because MCD4 is an essential gene, it may represent an additional, clinically relevant drug target for pimozide. The inositol-lipid-mediated signaling gene PIK1 and the spingholipid-mediated signaling gene YPK1 were also among the ten most required genes for resistance to pimozide (Table 3). They clustered with a group of other strains deleted for genes involved in lipid biology (Figure 4), such as the de novo lipid synthesis genes PAH1 and SUR4.
Eight drugs, among them the antidepressant fluoxetine, were enriched for the condensed term establishment of cell polarity (purple or blue color in Figure 5). In total, 51 genes were assigned to the detailed GO process establishment and/or maintenance of cell polarity and caused a sensitive phenotype when deleted (Tables S5 and S6). Many of these genes scored in the majority of the drugs, for example all four members (CKA1, CKA2, CKB1, and CKB2) of the casein kinase II-holoenzyme complex, and TPM1, the major isoform of tropomyosin which directs polarized cell growth and organelle distribution. For the seven drugs where the enrichment for establishment and/or maintenance of cell polarity was scored using sensitive homozygous and essential heterozygous strains (purple color in Figure 5), six essential members (EXO70, SEC3, SEC6, SEC8, SEC10 and SEC15) of the exocyst complex, which determines where secretory vesicles dock and fuse, were scored in all drugs except fluoxetine.
Drug targets are often encoded by essential genes, thus essential genes scored in our assay may represent important additional targets of psychoactive compounds that may be useful in the development of therapeutics for other applications. In a given heterozygous strain, the reduced gene copy number of a potential drug target leads to a reduced level of the corresponding protein. When this strain is grown in the presence of a drug targeting the heterozygous locus, the result is a further decrease in “functional” dosage due to the drug binding to the protein target. If this protein is important for growth, the result will be drug sensitivity [22]. In our functional enrichment tests, two processes were uniquely overrepresented among sensitive essential genes (red color in Figure 5): mitotic and meiotic cell cycle for fluorophenyl-methoxytropane and chromatin organization for cyproheptadine. Examples of targeted essential genes in cyproheptadine treatment include chromatin-remodeling genes (ARP4, ARP7, ARP9), genes in the multisubunit (NuA4) histone acetyltransferase complex (EPL1, ESA1, SWC4), and RSC4 and RSC6 in the RSC Chromatin remodeling complex.
Although not revealed as a functional enrichment among sensitive strains deleted for essential genes, most of the other FDA-approved drugs also have potential secondary drug targets as infered by the presence of essential genes among the ten most required genes for drug resistance (Table 3). As judged by the high number of sensitive strains deleted for essential genes in paroxetine treatment (10 strains) and sertraline treatment (9 strains), these selective serotonin re-uptake inhibitors are particularly rich in potential secondary drug targets. Essential genes required for resistance to the FDA-approved drugs include those involved in RNA processing, transcription and translation, genes functioning in the protein folding chaperonin complex, and the chromatin-remodeling/DNA repair gene ARP4 (bold in Table 3). Deletion of ARP4 resulted in some of the most sensitive phenotypes when cells were treated with cyproheptadine, sertraline, or with haloperidol (Table 3). ARP4 has a close human homolog, ACTL6B, which encodes a subunit of the BAF (BRG1/brm-associated factor) complex in mammals, functionally related to the SWI/SNF complex in S. cerevisiae. The SWI/SNF complex is thought to facilitate transcriptional activation by antagonizing chromatin-mediated transcriptional repression [36]. Another example of an essential gene required for drug resistance in several FDA-approved drugs is GSP1, which functions in RNA-processing (Table 3). The mammalian homolog of Gsp1, Ran (BlastP E-value<E-261) is, as in yeast, a nuclear GTP-binding protein.
Interestingly, the fitness profile of the ARP4-deleted strain was very similar to the strains deleted for the cytosolic chaperonin subunits CCT5, CCT8 and TCP1 (Figure 4). The chaperonin complex is involved in protein folding (primarily of actin and tubulin) and cytoskeleton organization [37]. In our fitness assays, seven of eight CCT-strains scored as significantly sensitive in many of the probed psychoactive drugs. Some (CCT3, CCT4, CCT7 and CCT8) were even among the top-ten required genes for resistance to cyproheptadine, fluoxetine, paroxetine, and sertraline (Table 3). Furthermore, several deletion strains with uncharacterized functions had similar fitness profiles as the chaperonins CCT5, CCT8 and TCP1 (Figure 4). Among them were TVP23 and YIP5 which both localize to the late Golgi, YEL048C which is synthetic lethal with GCS1 (involved in ER to Golgi transport), APM2 (homologous to medium chain of mammalian clathrin-associated protein complex involved in vesicle transport) and SWH1 (similar to mammalian oxysterol-binding protein, localized to Golgi and nucleus-vacuole junction).
To test if our findings in yeast might reflect drug action in human cells, we looked at the proportion of scored genes with human homologs. Among the strains significantly sensitive to at least one psychoactive compound, 58.4% were deleted for a gene with a close human homolog (BlastP E-value<E-6), as compared to 45.0% for all analyzed deletion mutants regardless of whether they had a fitness defect or not. To test if strains deleted for genes involved in core cellular processes are more sensitive in general, we compared our results obtained with the 81 psychoactive compounds to 81 randomly chosen chemically diverse compounds (see Materials and Methods). We found that a similar proportion of genes with close human homologs (59.7%) were scored for strains that were significantly sensitive to at least one of these diverse chemicals. Despite this similarity in proportion of sensitive strains with human homologs in the two datasets, conserved genes were scored much more frequently (in >10% of the compounds) in the psychoactive drug set than in the random drug set. In fact, considering only genes deleted in frequently scored strains, 64.1% of the psychoactive drugs had close human homologs (BlastP E-value<E-6) while the corresponding proportion for the structurally diverse drug set was significantly (p<0.006) lower (45.4%) and similar as the fraction of human homologs for multi-drug resistance genes (47.1%) in a recently published study [23]. This difference points to a significant enrichment of frequently scored sensitive strains with human homologs for the psychoactive drugs. Among the strains sensitive to the highest number of psychoactive compounds, seven of eight had close human homologs: NEO1, SAC1, PIK1, VPS29, PEP8, ARP4 and VPS35. The majority of these genes are involved in vesicle transport, which was the most frequently enriched function among strains sensitive to psychoactive drugs. Thus, the specific psychoactive drug detoxification mechanisms identified in yeast are likely to be of importance in humans treated with psychoactives.
Many psychoactive drugs are associated with adverse secondary effects in humans yet the mechanisms that underlie these off-target effects are poorly understood. To address mechanisms of drug action in a systematic manner, we profiled the genome-wide collection of budding yeast deletion strains for sensitivity to a broad spectrum of psychoactive compounds, of which dopaminergic and serotonergic drugs had a high bioactivity. Among 214 tested compounds, we uncovered 81 drugs that conferred a measurable growth defect on wildtype yeast. An appropriate dose of these active compounds was applied to the pooled heterozygous and homozygous yeast deletion sets to identify genes whose function is required for optimal growth in the presence of drug. Fifteen percent of all yeast strains (deleted for non-dubious ORFs) exhibited significant sensitivity (r>2, z>3) to these 81 psychoactive compounds and more than half of the drugs interacted with core cellular functions. Several clinically important drugs, such as fluoxetine, cyproheptadine, and clozapine were linked to diverse cellular processes. This observation may explain both the diversity of side effects observed in human patients and the therapeutic variability associated with these drugs. That is, polymorphisms in any of the conserved processes affected by a given drug are a likely source of the individual variation in response to drug. For instance, the response to the frequently prescribed antipsychotic clozapine is highly variable between individuals as the same dose can have markedly different efficacy and/or side effects in different patients [38]. Genes functioning in vesicle transport, protein localization, telomere biology, and catabolic processes were required for clozapine resistance in yeast. In another example, fluoxetine is associated with side effects such as seizures, nausea, sleepiness, anxiety, and serious allergic reactions. This antidepressant affects numerous cellular processes including establishment of cell polarity, protein localization, and cytoskeleton organization and biogenesis. Given the limited number of FDA-approved drugs within the set of 81 compounds analyzed here and the overlapping side effects associated with these drugs, it is not yet possible to correlate any single side effect to a particular perturbed pathway.
The most frequently scored sensitivity for the 81 profiled antipsychotic drugs was due to loss of secretory pathway function, likely indicating the importance of vesicle transport (e.g. to the vacuole) for drug detoxification. The lysosome (the mammalian vacuole equivalent) is known as the major site of degradation of both exogenous and endogenous molecules. For FDA-approved drugs, the requirement for vesicle transport genes was reflected in the frequent sensitivity of the neo1 deletion strain as the most sensitive strain in six FDA-approved drugs. Neo1 is an essential, highly conserved type 4 P-type ATPase involved in intracellular membrane- and protein-trafficking. Members of this family of P-type ATPases are implicated in the translocation of phospholipids from the outer to the inner leaflet of membrane bilayers. Our data suggested that interference with membrane structure and transport through inhibition of Neo1 is an additional, unwanted mechanism of action for clozapine, cyproheptadine, fluoxetine, paroxetine, sertraline and haloperidol, and their drug analogs. The importance in humans of functional 4 P-type ATPases is well documented as hereditary cholestasis, caused by defects in biliary epithelial transporters, has been directly linked to mutations in a 4 P-type ATPase gene [39].
In addition to the frequently observed requirement for uncompromised vesicle transport for drug detoxification, several drug sensitivity profiles were enriched for more specific processes. Within the FDA-approved drug group, the antidepressant paroxetine was unique in targeting RNA processing genes, pimozide interfered with membrane lipid metabolic processes, cyproheptadine preferentially targeted essential genes with chromatin remodelling functions, and fluoxetine interfered with establishment of cell polarity. Furthermore, seven dopaminergic compounds including the anti-Parkinson drug bromocriptine resulted in sensitivity of strains deleted in aromatic amino acid biosynthetic genes. This sensitivity may be a result of that dopaminergic drugs block aromatic amino acid uptake in yeast, requiring yeast to activate the corresponding biosynthetic pathways. Given the fact that aromatic amino acids are precursors to dopamine and serotonin, this was an interesting observation suggesting that the levels of intracellular precursors may be important in the response to certain psychoactive drugs.
Interestingly, interference with members of the chaperonin complex resulted in some of the most severe phenotypes. Seven of eight CCT-strains scored as significantly sensitive in several psychoactive drugs, among them CCT5. The human homolog of this gene is associated with hereditary neuropathy [40]. Although it is unclear how mutated CCT5 causes this disease, it has been postulated that its mutation leads to accumulation of misfolded cytoskeletal proteins, leading to defective assembly of actin into microfilaments resulting in neuronal apoptosis [41]. In our yeast screens, CCT5 was needed for resistance to eight different compounds (cyproheptadine, paroxetine, fluoxetine, indatraline, MDL72222, CY208-243, 2-Chloro-11-(4-methylpiperazino)-dibenz[b,f]oxepin, N-Desmethyl-clozapine, and 3-alpha-[(4-Chlorophenyl)-phenylmethoxy]-tropane. We conclude that interference with tubulin and actin folding is an important, secondary mechanism of action of these compounds.
As an example of how the information from our yeast assays may lead to testable drug-gene interaction hypotheses in humans, we found that the levels of the yeast strain heterozygous for ACC1 was eleven-fold reduced in ritanserin as compared to the control, indicating that the acetyl-CoA carboxylase Acc1 may be a secondary target of ritanserin. Like its yeast counterpart, the human homolog ACACA is required for de novo biosynthesis of long-chain fatty acids and its activity drops during fasting [42]. Because increased appetite is a reported side-effect during ritanserin treatment [43], it is tempting to speculate that biochemically mimicking fasting would increase appetite.
These studies raise several important issues for further consideration. Understanding the mechanisms that underlie adverse effects of clinically approved drugs is crucial for the development of next generation therapeutics with improved selectivity and efficacy. Moreover, knowledge of patient polymorphisms in off-target pathways may allow adverse effects of any given drug to be preempted by personalized pharmacogenomic strategies. It is also conceivable that some of the observed secondary drug effects are critical for therapeutic benefit.
In summary, a number of cellular processes were associated with sensitivity to the dopaminergic and serotonergic classes of psychoactive compounds. This points to additional, previously uncharacterized mechanisms of action for these drugs in humans and suggests follow-up experiments aimed at understanding a drug's mechanism of action on a genome-wide level. Our results suggest that model organism pharmacogenetics can be used as a comprehensive and unbiased tool in initial studies aiming at unraveling secondary effects and mechanisms of action for therapeutic compounds and their analogs. A more rigorous understanding of the complete mechanism of drug action in humans would be beneficial in the development of a new generation of better tolerated psychoactive drugs, and in personalized medicine.
High purity compounds for genome-wide fitness profiles were obtained from Tocris BioScience (http://www.tocris.com) as ligand sets and as the serotonergic (#1732) and dopaminergic (#1718) collections. In total, these drug collections comprised 226 drugs, 12 of which overlapped between the collections. A complete list of drugs, catalogue numbers, solvents, and concentration used in the genome-wide screens is provided in Table S1.
For genome-wide fitness profiles, the complete sets of ∼4700 homozygous deletion strains and ∼1100 essential heterozygous deletion strains in the BY4743 and BY4744 backgrounds (MATa/α his3Δ1/his3Δ1 leu2Δ0/leu2Δ0 lys2Δ0/LYS2 MET15/met15Δ0 ura3Δ0 /ura3Δ0 ORF::kanMX4) were used [29],[44]. A strain in the same genetic background with YDL227C replaced by kanMX4 was used as the wildtype control for drug titration curves. Strains were stored in 7% DMSO at -80°C. Because all experiments were performed in rich media (YPD [45], without antibiotics), it is unlikely that the presence of auxotrophies had a major effect on our results, however, we cannot rule out that the disruption of the corresponding pathways in yeast may, in some cases, alter our findings. Beginning from an initial maximal concentration of 200 µM, the degree of growth inhibition was determined by exposing wildtype cells to a serial dilution of compound until only a slight inhibition (∼15%) of wildtype growth was observed (see Figure S1). Cells were inoculated at an OD600 of 0.0625 in serial dilutions of drug and grown in a Tecan GENios microplate reader (Tecan Systems Inc., San Jose, USA) at 30°C with orbital shaking. Optical density measurements (OD600) were taken every 15 minutes until the cultures were saturated, and doubling time (D) was calculated as described [46]. Fitness assays using pooled deletion strains were performed as described [47] with the following modifications: i) after growth, 350 µl from each of two independent cultures of the 5-generation homozygous pool and 350 µl from the 20-generation heterozygous essential pool were combined, thereby allowing for approximately equal representation of barcodes for PCR reactions and hybridization to the same DNA chip using the unique barcodes incorporated in each of these strains. ii) for amplification of the tags, ∼0.2 µg genomic DNA was combined with a 1 µM mix of either up- or down-tags and 82% (v/v) Platinum High Fidelity PCR Supermix (Invitrogen # 11306-016) containing anti-Taq DNA polymerase antibody, Pyrococcus species GB-D thermostable polymerase, recombinant Taq DNA polymerase, Mg2+, and dNTPs, iii) extension temperature was 68°, iv) extension was for 2 min except for a final 10 min extension v) 34 cycles of amplification were performed, vi) after 10-16 h, the hybridization mix was removed from Affymetrix Gene Chips, replaced with Wash A (6x SSPE, 0.01% Tween), and chips were stained and washed using GeneChip Fluidics Station 450 (Affymetrix) according to the GeneFlex_Sv3_450 protocol with one additional wash A cycle before the staining.
Intensity values for the probes on the chip were extracted using the GeneChip Operating Software (Affymetrix). Quantile normalization, outlier omission, fitness defect ratio (denoted as “r”) and z-score (denoted as “z”) calculations were performed as previously described [47],[48]. In short, fitness defect ratios were calculated for each deletion strain as the log2 of the ratio between the mean signal intensities of the control and the drug chips. The larger the ratio, the more depleted (sensitive) is the strain as compared to control condition without the drug. To include the variance in the control experiments, we also calculated z-scores for each gene by dividing the difference in mean intensity across the control chips and treatment with the mean standard deviation of the signal intensities for the given gene across all 18 control chips [48]. The larger the z-score, the more likely it is that a given strain is significantly depleted from the pool. In our analysis, we scored a deletion strain as significantly sensitive using a threshold for both z-score and log2 intensity ratio. A threshold of z>3 was selected based on our earlier observations that above this limit, 100% of 186 deletion strains detected as sensitive by microarray could be confirmed using individual growth assays [24]. This stringent threshold was chosen to minimize the number of false positives. In addition, we added a further requirement that a sensitive strain should display at least a fourfold depletion (r>2, i.e. log2>2) compared to the control condition. This criterion was added to avoid including z-scores which were artificially high due to a low standard deviation in the control chips. Due to the way the screens were performed (at low drug concentration, i.e. an IC15) and analyzed [22],[24] we have focused on sensitive strains in this work, as opposed to apparently resistant strains. Two-dimensional hierarchical clustering of the fitness ratios was performed using Pearson correlation [32] and the data was visualised using the MultiExperiment Viewer from the TM4 microarray software suite (http://www.tm4.org/index.html).
In each of the 81 profiled drugs, sensitive deletion strains were tested for Gene Ontology Functional enrichment using the standard hypergeometric test provided by the GoStats Bioconductor modules for R [49]. For each drug, we performed three independent functional enrichment tests using i) sensitive heterozygous strains deleted for essential genes (z>2), ii) sensitive homozygous strains (z>2), and iii) all sensitive strains in the given drug with z>2. As the global control set, we used all yeast ORFs in the corresponding deletion background with chip intensity values above background. The background was determined as the average value of all unused tags on the chip (∼3600 tags×5 copies = 18000 values) +2 standard deviations of the background tags. Obtained p-values were corrected for multiple testing by multiplying by the number of identified terms. Adjusted p-values<0.0001 were considered significant. GO processes linked to less than 20 or more than 300 genes in our background set were excluded from our tests. Two-dimensional hierarchical clustering of overrepresented GO processes was performed using binary data [50]. To test the robustness of our functional enrichment tests, we repeated the same analysis using each of the following thresholds: z>3, r>2, r>3 and found consistent functional enrichment profiles.
In the calculations of the proportion sensitive strains deleted for genes with close human homologs (Blastp E-value<E-6), we used a set of 81 recently profiled (our unpublished data) compounds with potency against wildtype yeast. These compounds represent structurally diverse chemicals derived from Chemical Diversity Labs, Inc. repository of >500,000 compounds.
Structure data files were obtained from Tocris and Pubchem for all compounds and Babel canonical smile strings were generated. In the chemical structure clustering, extended connectivity fingerprints based on functional classes in Pipeline Pilot were used [51]. In the physiochemical property clustering, ten descriptors representing important properties for potential drug candidates were calculated after salts were stripped, using Frowns and Openeye cheminformatic libraries [31],[52]. PCA was used to find the strongest properties that separated active from non-active compounds. The revealed properties ALogP and molecular weight were validated to see how they correlated with the pattern of the other eight descriptor loadings. The non parametric Wilcoxon rank sum test supported LogP (p-value 4.91e-13) and MW (p-value 3.42e-05) as significant representative properties.
All supplementary data can also be downloaded from our webpage, http://chemogenomics.med.utoronto.ca/Supplemental/psychoactives/.
|
10.1371/journal.ppat.1005780 | Ecological Contexts of Index Cases and Spillover Events of Different Ebolaviruses | Ebola virus disease afflicts both human and animal populations and is caused by four ebolaviruses. These different ebolaviruses may have distinct reservoir hosts and ecological contexts that determine how, where, and when different ebolavirus spillover events occur. Understanding these virus-specific relationships is important for preventing transmission of ebolaviruses from wildlife to humans. We examine the ecological contexts surrounding 34 human index case infections of ebolaviruses from 1976–2014. Determining possible sources of spillover from wildlife, characterizing the environment of each event, and creating ecological niche models to estimate habitats suitable for spillover, we find that index case infections of two ebolaviruses, Ebola virus and Sudan virus, have occurred under different ecological contexts. The index cases of Ebola virus infection are more associated with tropical evergreen broadleaf forests and consuming bushmeat than the cases of Sudan virus. Given these differences, we emphasize caution when generalizing across different ebolaviruses and that location and virus-specific ecological knowledge will be essential to unravelling how human and animal behavior lead to the emergence of Ebola virus disease.
| Multiple Ebola virus disease outbreaks have occurred over the past 40 years, yet we still do not know the geographical distributions, definitive host species, and suitable habitats for animal-to-human transmission of different ebolaviruses. Each Ebola virus disease outbreak has started with at least one transmission event from a wildlife host to a human, also known as a spillover event. While researchers have studied these events in regards to Ebola virus disease outbreaks, many studies neglect that there are multiple ebolaviruses and that these viruses may differ in their spillover events. We characterize the specific ecological contexts of different ebolavirus spillover events based on recorded index case infections. Comparing the environmental contexts of these cases and using ecological niche modelling, we find that two ebolaviruses have different suitable habitats for spillover. The different habitats and contexts of the two ebolaviruses involved in the majority of outbreaks, Ebola virus and Sudan virus, indicate that we must further investigate virus-specific differences in ebolaviruses and their hosts.
| From the first recognized outbreak of Ebola virus disease (EVD) in 1976 to the recent outbreak beginning in 2013, our knowledge about the molecular biology and epidemiology of viruses belonging to the genus Ebolavirus has increased dramatically. Yet after nearly 40 years of research, we still have a limited understanding of the ecology and evolution of these viruses outside the context of outbreaks in humans [1]. One limitation in understanding ebolavirus ecology has been identifying the reservoir hosts in which ebolaviruses persist in nature, while another obstacle has been determining what causes the sporadic transmission of ebolaviruses from their natural reservoir into other animals and humans, leading to subsequent human-to-human transmission and outbreaks of EVD. These initial episodes of animal-to-human transmission are called spillover events, and knowing when, where, and under what environmental conditions ebolavirus spillovers occur could reveal underlying relationships in ebolavirus ecology. Additionally, identifying host-pathogen interactions of ebolaviruses with their natural reservoir and spillover hosts, as well as the interactions of these hosts with humans, could help researchers improve preemptive measures for transmission from wildlife as well as answer fundamental questions about virus ecology.
The genus Ebolavirus belongs to the family Filoviridae along with the genera Cuevavirus and Marburgvirus. Five species of viruses have been established in the genus Ebolavirus: Zaire ebolavirus, Bundibugyo ebolavirus, Sudan ebolavirus, Taï Forest ebolavirus, and Reston ebolavirus. The viruses belonging to these species are known as Ebola virus (EBOV), Bundibugyo virus (BDBV), Sudan virus (SUDV), Taï Forest virus (TAFV), and Reston virus (RESTV), respectively [2]. These different ebolaviruses are genetically distinct, with SUDV and RESTV being the most divergent from the other ebolaviruses [3]. Factors influencing the speciation of ebolaviruses and how ebolavirus speciation relates to reservoir host evolution and ecology remain enigmatic.
No ebolavirus has ever been isolated from a putative reservoir species. In addition, of all the ebolaviruses, only EBOV has had its RNA detected in potential reservoir hosts, 3 species of African fruit bats [4]. Duikers (Cephalophus species), gorillas (Gorilla gorilla), chimpanzees (Pan troglodytes), and various rodents have also tested positive for EBOV RNA [5,6]. It is widely suspected that these are probably incidental hosts that are indirectly infected by bats [7]. While serological evidence exists for RESTV in Asian bats and TAFV was found in a deceased chimpanzee, both SUDV and BDBV have yet to be identified via serology or PCR in any wildlife [7]. Therefore, we do not know the definitive reservoir host species for any ebolavirus or what factors influence ebolavirus transmission from wildlife into human populations.
While the difficulty of detecting ebolaviruses in wildlife reservoirs hinders the identification of reservoir hosts and the determination of their enzootic cycles, examining the ecology of ebolaviruses at the human-animal interface could yield insights about potential animal hosts as well as the ecological conditions that drive the emergence of these pathogens. EBOV, SUDV, BDBV, and TAFV are known to cause EVD in humans. Antibodies to RESTV have been detected in humans in the Philippines [8]; however no RESTV spillover events in humans have been documented, and it is assumed that RESTV is nonpathogenic in humans [9]. Therefore, outbreaks of EVD in human populations have enabled researchers to characterize the other four ebolaviruses according to their locality, case fatality, and epidemiology [10], as well as understand human-to-human transmission [11]. However, the individual spillover events that lead to outbreaks in humans have not been as well characterized. While few ebolavirus spillover events have been confirmed, there are reported index cases, the first human cases to be clinically described or laboratory confirmed in a chain of transmission. The central estimate for the incubation period until onset of EVD is 5.3–12.7 days for EBOV, 3.35–12 days for SUDV, and 6.3–7 days for BDBV [12]. Therefore, these index cases provide an approximation of roughly where and when spillover events have occurred. Since only EBOV has been sparsely detected in wildlife, examining these spillover events via index case reports is currently the only way that we can consistently compare the ecologies of multiple ebolaviruses.
In order to specify the ecological contexts of ebolavirus spillover events, one must first define the habitats where spillover events occur. Ecological niche models (ENMs) can be used to qualitatively compare the habitats where different species occur and identify regions of habitat suitability [13]. This toolset is increasingly being used to predict the ecological niches of viruses. For example, cases of human monkeypox disease have been used to model the ecological niches of monkeypox virus [14, 15]. Instead of using species occurrences and predicting fundamental ecological niches, we can use the locations of ebolavirus index cases and their associated spillover events from wildlife into humans to determine suitable habitats for ebolavirus spillover. Comparing the suitable spillover habitats of different ebolaviruses allows us to further compare the ecological contexts of multiple ebolaviruses and determine virus-specific factors of spillover.
Here we characterize the habitat and context of all known ebolavirus index case infections and associated spillover events into humans from 1976–2014 to investigate species and location specific ecological relationships. We use an ENM modeling approach that is optimized for small sample sizes to compare the habitats of spillover events of different ebolaviruses. In doing so, we find that distinct ebolaviruses spill over into humans under specific ecological contexts and are associated with different habitats.
We identified a total of 34 index cases and the associated spillover events of four ebolaviruses (24 EBOV, 7 SUDV, 2 BDBV, and 1 TAFV) (Table 1). We hereafter refer to both these index cases and their associated spillover events as “spillover events.” Spillover events of viruses from each species occurred in distinct geographic locations (Fig 1), while 1 SUDV and 4 EBOV spillover events occurred in the same location as a previous event.
EBOV spillover events have occurred at latitudes ranging from -5.3°-8.6° throughout the year during both wet and dry seasons. SUDV spillover events were more spatially clustered at .64°-4.6° and 6/7 (86%) occurred during the wet season. Two SUDV spillovers occurred in the same location during the same season within 3 years from each other. The two BDBV events occurred at .77°-2.7° during the wet season, and the TAFV event also occurred during the wet season.
We used the locations of ebolavirus spillover events and environmental covariates to create ecological niche models, which identified habitats similar to those where different ebolaviruses have spilled over into humans. Suitable habitats for EBOV and SUDV spillover events within Africa are shown in Fig 2. These models were made under the assumption that EBOV and SUDV spillovers from wildlife do not occur outside of mainland Africa. Additional models were made to compare the habitats of EBOV and SUDV spillover events within a global context (Fig A in S1 Text). Due to the limited sample size of TAFV and BDBV spillover events, we could not create models for these species that were statistically significant. The minimum training presence threshold was chosen to create the binary maps because it was more liberal than the 10 percentile training presence threshold. The minimum training presence threshold is established by the lowest habitat suitability in the training data set; therefore, all indicated regions have ecological conditions that at minimum match those in the least suitable confirmed location of spillover.
The models for EBOV and SUDV at the minimum training presence threshold successfully predicted spillover event locations at high rates, 85% (17/20) and 66% (4/6) respectively. A P value of 3e-05 was calculated for the SUDV models and 3e-19 for the EBOV models, indicating that both models were statistically significant at predicting distribution of spillover events compared to random.
The models of EBOV and SUDV spillovers at the minimum training presence threshold overlapped in approximately 12% of their total area. No SUDV spillover events were within the model of EBOV, and only three EBOV events occurred within the model of SUDV. Of the original 20 environmental covariates, 9 were determined to be important in contributing to the models of both ebolaviruses (Table 2).
SUDV spillover events occurred at a significantly higher mean elevation than those of EBOV (p = 0.004). The 95% CI for the difference in means of SUDV and EBOV events was 196–671 m. EBOV events are also more associated with evergreen broadleaf forest compared to other land cover types than SUDV events (p = 0.0007), whereas SUDV events are more associated with woodland (p = 0.0078).
The long-term monthly mean rainfall and temperature varied between SUDV and EBOV locations (Fig 3). Comparing SUDV and EBOV spillover locations, there was no significant difference in the mean temperature (p = 0.18) or rainfall (p = 0.95) of the actual month when an event occurred.
The suspected animal sources of all known spillover events of viruses from different Ebolavirus species are shown in Table 1. The geographic distributions of these animals from the IUCN Red List [16] in relation to EBOV and SUDV ENMs are in Figs B-D in S1 Text. EBOV spillover events were more associated with bushmeat contact than SUDV spillover events (p = 0.012). Chimpanzees, gorillas, duikers, monkeys, and fruit bats were all suspected sources of spillover for EBOV. Only one SUDV spillover event could be potentially linked to the bushmeat of a baboon, a species not found to be associated with EBOV spillover. In the majority of SUDV spillover events, no possible animal source could be identified. Two SUDV spillover events were linked to the same factory containing insectivorous bats and rodents.
The recent outbreak of EVD has inspired research across many fields, so it is critical to communicate that multiple ebolaviruses can cause EVD outbreaks and could have distinct ecological relationships. Our findings demonstrate that the spillover events of different ebolaviruses do occur within specific ecological contexts and habitats. EBOV spillovers have occurred within or on the edges of tropical evergreen broadleaf forests, and index patients are often hunters, villagers, or outdoor workers who have come into contact with animals such as bats, primates, and duikers. In contrast, SUDV spillover events occur at higher elevations, are more associated with woodlands, and have cryptic animal sources of spillover. In order to study the ecological contexts of TAFV and BDBV, more information on their possible animal reservoirs and spillover events are necessary.
Our study provides an approximation of the different ecological contexts of the index cases and spillover events of four ebolaviruses known to cause EVD in humans. We compared the contexts of EBOV and SUDV spillover events as well as qualitatively estimated areas of habitat suitability for spillover of these viruses. We did not attempt to determine the fundamental ecological niches of different ebolaviruses and Ebolavirus species. Instead, we compared the ecology of ebolaviruses based on the contexts of index cases and associated spillover events. Our models indicate habitats that are similar to those where index cases have occurred. A biogeographical study with additional assumptions about ebolaviruses and their hosts could quantitatively compare whether the differences in EBOV and SUDV spillover event locations are due to differences in niche conservatism or differences in the distributions of spillover events.
One limitation to our approach is that spillover event locations are unlikely to be independent and index case reports may be subject to geographical and temporal variation in reporting bias. Therefore, our models could exclude actual habitats of these viruses. For instance, more EBOV spillover events have been detected in the western Congo Basin than in the central Congo Basin. Thus, parts of the central Congo Basin are excluded from our models despite that this region includes putative EBOV host ranges (Figs B-D in S1 Text). Our models may also include habitats that are ecologically suitable based on spillover events but are unlikely to contain the viruses. For example, we identified habitats in Southern Africa that have suitable ecological conditions but are geographically isolated from where ebolaviruses have so far been detected.
A further limitation to our study is that we used index case reports to approximate when spillover events occurred and analyzed the mean environmental data from the month of symptom onset or date of death, if symptom onset date was unavailable. Therefore, the variable duration of illness and incubation periods for ebolaviruses may influence our estimations of spillover event locations and results about seasonality. Additionally, historical monthly precipitation data was limited in spillover locations, so we used coarse resolution data to classify seasons and compare the precipitation within the month of a spillover event. Long-term monthly mean and bioclimatic data were available at much higher resolutions and were used for our other analyses. Until more ebolavirus spillover events are confirmed, our study provides an approximation of the ecological conditions of spillover events of ebolaviruses.
Despite these limitations, our study further supports that researchers cannot generalize about the ecological contexts of different ebolaviruses. Previous authors have used climatic data and ENMs to make inferences about spatial and temporal relationships of ebolaviruses. One group used NVDI models and Landsat data and found that the 1994–1996 EVD outbreaks occurred in tropical forest and were associated with climate changes from drier to wetter conditions [17]. In contrast, another group found that 1994–2002 EVD outbreaks were associated with drier conditions at the end of the rainy season [18], while another study found EVD outbreaks to be associated with lower temperatures and higher humidity [19]. These studies did not differentiate between different ebolaviruses, which may explain their discrepancies. Considering the different contexts of SUDV and EBOV spillover events, we found no associations between the temperature or precipitation during the month of a spillover event, and spillovers of both ebolaviruses occurred in both wet and dry seasons.
We also find that generalizing across ebolaviruses when making ENMs can miss virus-specific relationships. For instance, one group used the occurrence data of the 3 species of Old World fruit bats that were positive for EBOV RNA, EBOV infections in wildlife, and the locations of outbreaks associated with multiple ebolaviruses to create predictive risk maps for EVD [20]. This approach assumes that the reservoir species for different ebolaviruses are the same and that the spillover events of these viruses occur under the same ecological conditions. However, another group used the occurrence data from 12 EVD outbreaks in the period of 1976–2002 to create ENMs and found that eliminating SUDV occurrence data from the other ebolaviruses created a prediction that did not include the distribution of SUDV [21]. Using a modeling approach that was optimized for small sample sizes and spillover locations from 1976–2014, we corroborate this observation that EBOV and SUDV are associated with different habitats and may need to be considered separately in further ecological modeling.
The habitats suitable for SUDV and EBOV spillovers correspond with the serological evidence of these viruses in humans and wildlife. Our models showed that habitats in West Africa and Central Africa were suitable for EBOV spillover, while East Africa and parts of Central Africa were more suitable for SUDV spillover. Serological surveys in humans or animals have found antibodies to EBOV or SUDV in the majority of the countries identified by our models (Table 3). However, the cross-reactivity of different ebolavirus antibodies in these assays makes it difficult to distinguish the type of ebolavirus infection. As more seroepidemiological surveys are done and diagnostics improve, we can use this information to make more informed conclusions about where different ebolaviruses are found in humans and animals.
Understanding the ecological contexts of ebolavirus spillover events also allows us to infer about the potential geographic distributions of these viruses and their respective hosts. Our models support that there is ample suitable habitat for EBOV and SUDV spillover. The recent discovery of Lloviu virus, a related filovirus, in insectivorous Miniopterus schreibersii bats in Europe [22], the detection of filovirus RNA and antibodies in Rousettus leschenaultii in China and Bangladesh [23, 24], the circulation of RESTV in Southeast Asia [25, 26] and the recent emergence of EBOV in humans in West Africa suggest the possible circulation for filoviruses far beyond the areas with recorded EVD outbreaks. The lack of recorded spillover events in areas with suitable ecological conditions could therefore be due to the absence of pathogenic filoviruses and their respective hosts, lack of recognition of spillover events, absence of ecological and anthropogenic factors driving specific spillover events, or a combination of these factors.
Considering the different environments in which SUDV and EBOV spillovers have occurred, we can form two hypotheses about their distributions and reservoir hosts: SUDV and EBOV occupy different host species (potentially multiple species) with different habitats or SUDV and EBOV persist in the same species that is able to occupy multiple habitats. We can investigate these potential host relationships through comparing our models with the distributions of animal species that have been previously associated with ebolaviruses. The suitable habitats that we determined for EBOV and SUDV spillovers are shared among some potential bat hosts and are specific to others. The 3 species of fruit bats that were positive for EBOV RNA, Hypsignathus monstrosus, Myonycteris torquata, and Epopmops franqueti, have geographic ranges that overlap more closely with the tropical forests where EBOV spillovers have occurred, but the eastern boundaries of these species occur near SUDV spillover events as well [27] (Fig B in S1 Text). Additional African bat species have been identified as potential reservoirs for EBOV based on serology [7], but again the cross-reactivity of these assays makes it difficult to make associations with particular ebolaviruses. Of the distributions of serologically positive bat species (Fig C in S1 Text), those of Micropteropus pusillus and Mops condylurus best match the woodland habitat associated with SUDV [28, 29]. M. condylurus belongs to the family Molossidae, whereas the other potential ebolavirus hosts belong to Pteropodidae. Moreover, bats belonging to the same genus (M. trevori) were found in the textile factory in Nzara where at least two independent spillover events of SUDV occurred [30] and have a geographic distribution within the SUDV habitat (Fig D in S1 Text). Perhaps the evolutionary and ecological differences between molossid and pteropid bats could explain the divergence between SUDV and EBOV.
The hypothesis that different ebolaviruses may have different host species, and therefore different habitats suitable for spillover, is supported by in vitro and in vivo experiments. In vitro studies have shown that the receptor NPC1 influences filovirus susceptibility in different bat species [31]. These studies may be useful in determining whether particular bat species are capable reservoirs for different ebolaviruses. In addition, experimental infection studies showed efficient replication of Marburg virus, but limited replication of the five ebolaviruses in Rousettus aegypticus [32, 33], the reservoir host for Marburg virus. These findings highlight the potential for a single filovirus-single reservoir host species relationship, which may be why EBOV and SUDV spillovers occur in different habitats.
Different relationships of ebolaviruses with secondary hosts and regional human-animal interfaces could also explain the differing contexts of EBOV and SUDV events. The majority of the EBOV spillover cases came from infected primates, whereas the sources of SUDV were unidentified. Additionally, there have been no documented outbreaks of EVD in chimpanzees in East Africa near the habitat of SUDV, indicating that reservoir hosts of SUDV may not come into contact with wild apes. For example, western lowland gorillas and chimpanzees share Ficus spp. as a food source with the potential EBOV reservoir bat species H. monstrosus [34, 35], and such an epizootic link may not exist for SUDV and its reservoir host. Furthermore, other forest-dwelling animals, such as the bay duiker (Cephalophus dorsalis), are only associated with EBOV spillovers and have been positive for EBOV RNA, while the woodland savannah-inhabiting baboon (Papio sp.) has only been associated with an SUDV spillover. The different animal species that are associated with these two viruses and their spillovers further supports that these ebolaviruses may have different ecologies.
Lastly, our study also provides more evidence about the evolution of ebolaviruses. It has been previously noted that there is remarkably little genetic diversity between both spatially and temporally separated strains of the same ebolavirus [36]. We found relatively large and contiguous areas of suitable habitat for both EBOV and SUDV spillover, which might explain why genetically similar viruses can circulate across large distances. Meanwhile, isolates of EBOV and SUDV differ by more than 40% in their genomes on the nucleotide level [37], which could be explained by the small overlap in their spillover habitats and possible geographical isolation of their host species. Current phylogenetic trees place SUDV in a different clade than EBOV, and it is possible that geographic isolation led to this speciation, potentially due to the Albertine Rift [36], which is near the eastern border of the EBOV habitats in our models. More extensive sampling of ebolaviruses in wildlife and rapid identification of index cases will increase our understanding of ebolavirus ecology and evolution, as well as potentially guide preemptive control strategies.
Overall, we show that ecological contexts of ebolavirus spillover events are virus-specific, relating to particular habitats, animal distributions, and human activities. Therefore, researchers must be careful about generalizing about ebolaviruses and their ecologies. Uncovering nuances in virus ecology will require further explorations of the human-animal interfaces that lead to viral spillover and collaborations across disciplines.
The geographic coordinates and identities of index cases were determined from the original literature describing EVD outbreaks and case reports (Table 1), using the Centers for Disease Control and Prevention’s EVD chronology as a guide [38], as well as a database of 22 EVD outbreaks [39]. Index cases were defined as patients who were the first to exhibit symptoms of EVD in an epidemic chain and had no previous contact with EVD patients. Additionally, index cases often had direct contact with wildlife prior to becoming symptomatic for EVD. The index case dates were determined by the date of symptom onset for the index patient. If this date was unavailable, the date of death of the index patient was used. In four cases only the month of an index case could be determined.
Additional index cases were identified by considering separate epidemic chains of transmission. Within the EVD outbreaks in the Republic of the Congo and Gabon are multiple spillover events, characterized by separate virus strains and epidemic chains of transmission [40, 41]. Index case patient demographics were determined from the literature. Research studies and case reports were examined to link index patients to potential sources of spillover, all of which were circumstantial.
The majority of the coordinates that we determined for spillover event locations corresponded to the index point locations in a recently created EVD database, which contains details about some spillover events [39]; however we identified additional spillover events that were not described in the database. We included the locations of index cases in Meliandou (Guinea) and Boende (DRC) that were not included in the database. We also used adjusted locations for the villages of Mwembe and Nakisamata (S1 Text).
Climate and terrain data were used to construct the ENMs. Layers of rasterized climate data of 19 bioclimatic variables as well as elevation came from the WorldClim database [42], which averages values from 1950–2000 at a spatial resolution of 30 arc seconds. We point sampled the elevation as well as long-term monthly mean rainfall and temperature at spillover locations from the WorldClim dataset in QGIS [43].
To look for seasonal relationships, we gathered the monthly mean precipitation during the month and year at the location where a spillover event occurred. We used the GPCP Version 2.2 Combined Precipitation Data Set provided by the NOAA/OAR/ESRL PSD, Boulder, Colorado, USA (available at http://www.esrl.noaa.gov/psd/). The GPCP data set has a spatial coverage of 2.5° latitude X 2.5° longitude, and uses a combination of satellite and gauge data to calculate mm precipitation per day [44]. Monthly values in the dataset are from 1979-October 2014. Therefore, we could not obtain precise monthly precipitation data for the 3 spillover events prior to 1979, so we did not include these events in analyses of within season rainfall, and we used the long-term monthly mean precipitation for classifying them into wet or dry seasons.
Temperature data for the specific month of a spillover event was gathered from the GHCN CAMS Gridded 2m Temperature (Land) dataset also provided by the NOAA/OAR/ESRL PSD, which has a resolution of .5° latitude X .5° longitude and contains monthly mean land surface temperatures from 1948 to October 2012 [45].
Vegetation and land cover were determined by mapping spillover event locations on raster maps from the Global Landcover facility (available at http://glcfapp.glcf.umd.edu). We used a global map with a spatial resolution of 225 seconds and fourteen land cover classes developed from NOAA-AVHRR satellite images from 1981–1994. We point sampled each location in QGIS to determine the land cover classification at that geographic location.
The ENMs were built using Maximum Entropy Species Distribution Modeling (MaxEnt), version 3.3.3k [46]. The MaxEnt program applies a machine learning method to estimate the distribution of a species under maximum entropy in geographic space using environmental factors as covariates and presence-only data as inputs [47]. We chose MaxEnt over other ENMs because it is robust with small numbers of occurrences and presence-only data [48].
Models were built to determine suitable habitat for ebolaviruses using spillover events as presence-only inputs. We used the 20 environmental covariates clipped to mainland Africa for our models and analyses because TAFV, BDBV, SUDV and EBOV spillovers have only occurred within mainland Africa. We also created models with a global environmental extent for comparison. Because our aim was to characterize the environments where different ebolavirus spillover events have occurred, we did not make assumptions about sampling or the density of the population. Instead of designing the models to provide probabilistic output, we used our models as indices of habitat suitability [13].
MaxEnt can use a subset of presence points to train the model, while reserving a subset to test the predictive strength of the model. Iteratively leaving out a single occurrence point, training the model, and then testing whether that point is included in the model, works well for determining the predictive ability of a model with a small sample size [49]. Therefore, we used a leave-one-out cross-validation method for each species, in which for a sample size n of spillover locations for each species, we divided the data into n equal size folds and kept one fold out to test the model. We repeated this process n times and then averaged the models for each species.
In addition to these sampling changes, the MaxEnt models were run on the default parameters with the cumulative output and the jackknife approach for comparing environmental covariates. The cumulative output reflects habitat suitability, where the probability of occurrence in each cell is the sum of the probability in that cell as well as all other cells with lesser or equal probability [49]. The minimum training presence and the 10 percentile training presence were compared as thresholds to determine which regions were suitable or unsuitable for the respective species. To test whether the models were statistically significant at predicting presence locations compared to random, we created a program for the statistical test described by Pearson et al. 2006 [49] (available at: https://github.com/AndrewJudson/jackknife). We calculated the percent overlap of the models by dividing the area of overlap by the total area of both models. Traditional niche overlap statistics such as Schoener’s D and I were not calculated because these assume a probability distribution for the species [50], whereas our models predicted habitat suitability.
Reducing the number of environmental covariates in ENMs enables researchers to determine which covariates are driving the model. We chose to use the same covariates across the different models so that we could compare the models with each other. We used a hierarchical approach and correlation matrix to remove covariates from the initial 20 that did not contribute to the models of either ebolavirus and were highly correlated with each other. We removed covariates as long as there were no changes from the original models. For analyses and mapping, we used the models with all 20 covariates.
All statistical tests to determine whether the spillover events of a particular ebolavirus were more associated with certain ecological conditions were done using Fisher’s exact tests. In order to compare the differences in mean elevation, temperature, or precipitation at spillover locations, two-tailed Welch’s t-tests were used. All statistical analyses were performed in R [51], and a significance level of p < 0.05 was used.
|
10.1371/journal.pcbi.1003830 | Baseline CD4+ T Cell Counts Correlates with HIV-1 Synonymous Rate in HLA-B*5701 Subjects with Different Risk of Disease Progression | HLA-B*5701 is the host factor most strongly associated with slow HIV-1 disease progression, although risk of progression may vary among patients carrying this allele. The interplay between HIV-1 evolutionary rate variation and risk of progression to AIDS in HLA-B*5701 subjects was studied using longitudinal viral sequences from high-risk progressors (HRPs) and low-risk progressors (LRPs). Posterior distributions of HIV-1 genealogies assuming a Bayesian relaxed molecular clock were used to estimate the absolute rates of nonsynonymous and synonymous substitutions for different set of branches. Rates of viral evolution, as well as in vitro viral replication capacity assessed using a novel phenotypic assay, were correlated with various clinical parameters. HIV-1 synonymous substitution rates were significantly lower in LRPs than HRPs, especially for sets of internal branches. The viral population infecting LRPs was also characterized by a slower increase in synonymous divergence over time. This pattern did not correlate to differences in viral fitness, as measured by in vitro replication capacity, nor could be explained by differences among subjects in T cell activation or selection pressure. Interestingly, a significant inverse correlation was found between baseline CD4+ T cell counts and mean HIV-1 synonymous rate (which is proportional to the viral replication rate) along branches representing viral lineages successfully propagating through time up to the last sampled time point. The observed lower replication rate in HLA-B*5701 subjects with higher baseline CD4+ T cell counts provides a potential model to explain differences in risk of disease progression among individuals carrying this allele.
| The clinical course of HIV-1 infection is characterized by considerable variability in the rate of progression to acquired immunodeficiency syndrome (AIDS) among patients with different genetic background. The human leukocyte antigen (HLA) B*5701 is the host factor most strongly associated with slow HIV-1 disease progression. However, the risk of progression to AIDS also varies among patients carrying this specific allele. To gain a better understanding of the interplay between HIV-1 evolutionary rate variation and risk of disease progression, we followed untreated HLA-B*5701 subjects from early infection up to the onset of AIDS. The analysis of longitudinal viral sequences with advanced computational biology techniques based on coalescent Bayesian methods showed a highly significant association between lower synonymous substitution rates and higher baseline CD4+ T cell counts in HLA-B*5701 subjects. The finding provides a potential model to explain differences in risk of disease progression among individuals carrying this allele and might have translational impact on clinical practice, since synonymous rates, which are proportional to in vivo viral replication rates, could be used as a novel evolutionary marker of disease progression.
| The clinical course of HIV-1 infection is characterized by considerable variability in the rate of disease progression among patients with different genetic background [1]–[3]. It has been shown that the likelihood of progressing to AIDS for subjects with baseline viral load (VL) around or lower than 10,000 copies/mL is dependent on baseline CD4+ T cell counts [4]. Subjects with baseline CD4+ T cell counts <750 cells/mm3 are at significantly higher risk for progression to AIDS (high-risk progressors, HRPs) than those with CD4+ T cell counts >750 cells/mm3 (low-risk progressors, LRPs). There is also evidence that HIV-1 genome controls virulence; however, the mechanisms underlying differential risk of progression to AIDS are not fully understood and likely involve both viral dynamics and host immune system [5].
CD8+ T cell responses play an important protective role in HIV-1 infection. HIV-1 replication in vivo is temporally associated with the appearance of CD8+ T lymphocyte responses [6], and the rate of disease progression is dependent on human leukocyte antigen (HLA) class I alleles [7], [8]. HLA-B*5701 is the host factor most strongly associated with slow HIV-1 disease progression [1], [9] and, in subjects with this allele, the CD8+ T cell responses target several epitopes in the gag p24 gene [10]–[12]. This often results in the evolution of viral variants that escape CD8+ T cell responses [13], [14], although there is evidence that some escape mutations in HLA-B*5701-restricted epitopes in p24 might occur at the expense of viral fitness [15]–[17].
HLA-B*5701 subjects with detectable viral load are ideal patients to investigate how the interaction between on-going viral intra-host evolution and immune system relates to risk of disease progression. It has recently been shown that HLA-B*5701 LRP subjects have a larger fraction of polyfunctional cells – i.e. cells producing two or more immune mediators (such as gamma interferon, interleukin-2, macrophage inflammatory protein 1 β, and Perforin) in response to specific HLA-B*5701-restricted epitopes in p24 – than HRPs [5]. At the same time, the study found that HIV-1 evolutionary rate is lower in LRPs compared to HRPs [5]. However, the exact mechanism, evolutionary meaning and clinical implications of these observations are still unclear. The rate of evolution estimated by molecular clock analysis of longitudinally sampled viral sequences is a compounded parameter, which depends on different factors, such as viral mutation (error) rate per generation, generation time (i.e. the viral replication rate), as well as the interplay between neutral genetic drift and positive or purifying selection.
In molecular adaptation studies, investigating the ratio of nonsynonymous and synonymous substitutions (dN/dS) has often proved to be useful [18], although evaluating the absolute nonsynonymous and synonymous substitutions rates separately can sometimes provide greater insights [19]–[21]. In HIV-1 intra-host evolution, for example, differences in synonymous substitution rates may reflect differences in mutation rate or generation time (i.e. viral replication rate), while different nonsynonymous rates may be linked to changes in selective pressure and effective population size [21]. Lemey et al. (2007) showed that HIV-1 disease progression seems to be predicted by synonymous substitution rates, which are indicative of the underlying viral replication dynamics [20]. By using a different method, Lee et al. (2008) also showed that the rate of intra-host HIV-1 evolution was not constant, but rather slowed down at a rate correlated with the rate of CD4+ T cell decline [19]. However, these studies were performed on patients of unknown HLA type, which makes it difficult to assess the potential impact of the host immune response on viral evolution and disease progression. The mechanism relating evolutionary rates and disease progression may also involve factors such as replication capacity [16] of the infecting virus or T cell activation [22]–[25]. Moreover, virus generation times and the ability of the viral strains to replicate in different environments could be affected by the virus population dynamics in latently infected cells [26], [27].
The present work focuses on a cohort of six untreated HIV-1 infected subjects, all carrying the HLA-B*5701 allele, followed longitudinally from early infection up to seven years. Bayesian molecular clock estimates, based HIV-1 gag p24 sequences, were analyzed in combination with in vitro viral replication capacity and immune activation data. The integration of experimental data with coalescent-based estimates allowed to develop, for the first time, a possible explanation for the correlation between HIV-1 in vivo replication rate and different risk of disease progression in HLA-B*5701 subjects.
Analyses were performed using longitudinal gag p24 sequence data from six HIV-1 infected subjects (P1-P6) carrying the HLA-B*5701 allele [5]. The subjects had different risk of progression toward AIDS based on CD4+ T cell count at baseline 10–11 weeks post infection (wpi) [4]. Three of these subjects (P1-P3) were classified as high-risk progressors (HRPs) and three (P4-P6) as low-risk progressors (LRPs) [5]. The average baseline viral load (VL) was 4,250 copies/mL for HRPs and 4,229 copies/mL for LRPs, while average baseline CD4+ T cell count was 458 cells/mm3 for HRPs and 1,129 cells/mm3 for LRPs.
The presence of molecular clock signal in each data set was first investigated by regression between root-to-tip divergence and sampling date on ML trees, which showed high correlation (r2>0.6) for each data set. HIV-1 evolutionary rate, estimated by molecular clock analysis of longitudinally sampled viral sequences, has been shown to be lower in LRPs than in HRPs [5]. However, rate estimates can be biased due to potential differences in internal and external branches of the phylogenetic tree. HIV-1 high mutation rate is expected to lead to a considerable number of deleterious mutations in the viral population, such that the most recent mutations segregating on external branches of HIV-1 phylogenies are likely to be transient [20], [28], [29], [30]. Deleterious mutations are rapidly purified and their inclusion can bias nucleotide divergence and evolutionary rate estimates, while mutation along internal branches are usually fixed. Internal and external branches were, thus, defined for 200 trees randomly sampled from the posterior distribution of HIV-1 gag p24 genealogies inferred with a Bayesian framework under a relaxed molecular clock (Figure 1), and mean evolutionary rates were estimated separately for each branch subset (Figure 2). In longitudinally sampled genealogies it is also possible to define the subset of branches connecting the root node with the most recent common ancestor of the sequences sampled at the last time point, which represent the surviving viral population successfully propagating over time through sequential bottlenecks driven by either positive selection or neutral genetic drift [28]. However, in HIV-1 intra-host genealogies, the last sampled sequences may not be monophyletic and different sets of backbone branches can be defined by a simple rotation around an internal branch. Therefore, a weighted average of the evolutionary rate was also calculated for the rates estimated along all the possible backbone paths of a genealogy (one example of such paths is shown by the branches highlighted in orange in Figure 1). For most patients, evolutionary rates in internal branches and backbone paths of the viral genealogies were very similar, while rates for external branches were higher (for all patients except P6) due, as expected, to an increased amount of deleterious mutations. However, the difference in mean substitution rates between HRPs and LRPs was still significant (p<0.05) in each analysis.
Gag p24 evolutionary rate differences between HRPs and LRPs were investigated in more detail by disentangling nonsynonymous (dN) and synonymous (dS) rates. Absolute dN and dS rates for all, internal, and external branches, as well as average rates for the backbone paths of the viral genealogies were estimated for each patient. The virus infecting HRPs displayed significantly higher dN rates along internal branches (Mann-Whitney U-test, p = 0.024) compared to the LRPs (Figure 3). There was also a trend toward higher dN rates in HRPs compared to LRPs when backbone paths (Figure 3), external or all branches (Figure S1) of the viral genealogies were analyzed. HIV-1 dS rates were significantly higher in HRPs than LRPs (Figure 3) for both internal branches (p = 0.024) and backbone paths (p = 0.024). A significant difference (p = 0.024) between the two groups of patients was also observed when external or all branches were analyzed (Figure S1).
Similarly, plots of HIV-1 dN and dS divergence over time within each patient were estimated for all and internal branches and along backbone paths. For internal branches, as well as along backbone paths, the virus infecting HRPs displayed a faster accumulation of dN substitutions over time compared to the LRPs for all patients except P5 (Figure 4). Analogous results were observed for all branches (Figure S2). There was a clear separation in dS divergence over time between the two groups of patients. The viruses infecting HRPs showed, overall, a higher number of dS substitutions over time for both internal branches and backbone paths compared to LRPs (Figure 4). Interestingly, during the first year (up to 400 days) of the infection, the accumulation of dS substitutions seemed to happen at the same rate between the two groups of patients. After the first year, the two groups began to diverge, and the virus populations in LRPs appeared to accumulate dS substitutions more slowly than those in HRPs.
The observed difference in dN and dS substitution patterns between the two groups of patients could be due to strong site-to-site rate variation, which has the potential to bias the estimates [31]. To investigate this possibility within the p24 gene we estimated the coefficient of variation (CoV) of substitution rates across dS and dN sites. As expected, the analysis revealed significant across-site variation in viral dS for all data sets, although the values were lower compared to the ones estimated for the HIV-1 env data set analyzed in Lemey et al. [20]. For both dN and dS, CoVs were similar among all six patients and there was no significant difference between the HRPs and LRPs (Table 1). Therefore, it is likely that the presence of significant rate heterogeneity across dS and dN sites equally affected HIV-1 evolutionary rate estimates in both groups of patients and does not account for the observed differences.
Differences in mean dS between HRPs and LRPs could also be the result of different levels of purifying selection. For each branch set (all, internal, external and backbone paths) of the viral genealogies, dN/dS ratios were calculated and compared. In general, correlation of dN versus dS rates was weak (0.17–0.39), and slopes were <1, indicating signal for purifying selection rather than neutral genetic drift or positive selection (Table S1). No significant difference was observed between LRPs and HRPs (Table S2) indicating that differential purifying selection was also an unlikely explanation for the observed substitution patterns.
Since dS substitutions are neutral or nearly neutral [32], HIV-1 dS is expected to be proportional to the virus replication rate [20]. The higher dS in HRPs may be the consequence of an infection with fitter viral variants characterized by faster replication. We examined viral replication capacity (RC) for all six HLA-B*5701 subjects (1–4 time points) by using a Phenosense Gag-Pro assay (see Methods). As expected, a trend toward increased RC over time was observed in all patients likely linked to the progressive fixation of fitter variants driven by ongoing selection [33]–[35]. However, no differences in RC measurements (10–332 wpi) were apparent between LRPs and HRPs (Table 2), suggesting that the observed difference in dS may not be related to fitter (high replicative) variants infecting the HRPs.
Finally, CD38 expression on CD4+ and CD8+ T cells was measured at baseline (13–17 wpi). Two HRPs (P1 and P2) displayed the highest values of CD38 expression on CD4+ T cells. Differences in CD38 expression on CD4+ or CD8+ T cells between the two groups of patients were not statistically significant (Table S3). Therefore, T cell activation during early infection is also an unlikely explanation for the observed difference in dS between HRPs and LRPs.
In order to identify other potential mechanisms behind the observed differences in viral evolutionary rate, correlation between mean dN or dS for different branch sets (all, internal, external and backbone paths) and clinical parameters for each patient (baseline CD4 count, baseline VL, CD4 slope, VL slope and baseline T cell activation) were investigated (Table S4). The only strong correlation found (r2 = 0.9) was between weighted average of dS estimated along possible backbone paths and baseline CD4+ T cell counts (Figure 5). In particular, higher baseline CD4+ T cell counts (10–11 wpi) were correlated with lower dS, indicative of lower replication rates and longer viral generation times in the LRPs compared to HRPs. The inverse correlation was highly significant (p = 0.002), even after Bonferroni correction (p = 0.09).
The present work investigated in depth the relationship between dS and dN viral evolutionary rate and risk of disease progression in HIV-1-infected subjects carrying the HLA-B*5701 allele. A recent study carried out on the same cohort has shown that HRPs have significantly lower polyfunctional CD8+ T cell responses, as well as higher viral evolutionary rate than LRPs [5]. The study also noticed that dS and dN changes, calculated by pairwise comparisons, were higher in HRPs than LRPs. However, the exact mechanism driving HIV-1 faster evolutionary rate in HRPs and its clinical implications remained obscure. Herein, absolute rates of dS and dN substitutions were estimated by Bayesian molecular clock analysis from longitudinally sampled HIV-1 gag p24 sequences along different branches of the viral genealogies. The evolutionary analyses and the comparison of dN and dS with various clinical, immunological and virological parameters resulted in three major and novel findings, which provide a potential mechanism for the initial observations reported in Norstrom et al. (2012). First, it was shown that the virus infecting LRPs exhibited significantly lower dN and dS divergence over time compared to HRPs. Second, no significant difference in site-to-site variation of dS or in dN/dS ratios along different branches of the HIV-1 genealogies was observed between the two groups of patients. This indicates that differences in rate heterogeneity across synonymous sites or purifying selection were unlikely to be the cause of lower viral dS in LRPs. Third, the analysis detected a strong inverse correlation between HIV-1 dS, which is directly proportional to the virus replication rate [33], [36], and baseline (10–11 wpi) CD4+ T cell count.
Changes in absolute HIV-1 dN and dS rates have been investigated previously [34], by analyzing env sequences from the Shankarappa et al. (1999) data set – nine HIV-1 infected patients followed longitudinally from the time of seroconversion [34]. However, seven out of nine patients in that data set received antiretroviral treatment during the follow-up time, their HLA-type was unknown and no data on in vitro viral RC or immune activation were available, making it difficult to disentangle the different factors that may have contributed to the interplay between viral evolution and disease progression. On the other hand, the present study examined untreated HLA-B*5701 subjects with different risk of disease progression, where HIV-1 evolutionary patterns could be compared to in vitro viral RC data, using a novel Gag-pro phenotypic assay, as well as immune activation data and a number of clinically relevant parameters.
Using the Shankarappa data set [34], Lemey et al. (2007) provided some evidence that slow HIV-1 disease progression can be predicted by lower dS rates, which are indicative of the underlying viral replication dynamics. In agreement with our finding, they did not detect significant differences in rate heterogeneity across synonymous sites between patients with different rates of disease progression. Rate heterogeneity, however, was generally higher than the one estimated for the viruses infecting the subjects enrolled in the present study, which likely reflects the higher diversity in env gp120 compared to gag p24 region. Lemey et al. (2007) also suggested that the slower replication dynamics of HIV-1 in patients with slow disease progression could depend on the state of immune activation of the host. Indeed, T cell activation (defined as the expression of CD38 on the T cells) is one of the strong predictors of progression to AIDS [22], [23], [25], [37], [38]. Nevertheless, within the HLA-B*5701 subjects studied herein no significant differences in T cell activation was observed. In addition, our analysis showed no differences in viral RC between the HRPs and LRPs. A trend toward increasing RC over time was observed in viruses sampled from patients with RC longitudinal data available, which is expected as a result of the continuous emergence and fixation of fitter viral variants over time [33]. HRPs displayed significantly higher dN rate along internal branches of the viral genealogies compared to the LRPs, which may be indicative of a higher rate of adaptation in the HRPs. It is important to notice, however, that the small sample size of our cohort requires a certain caution before drawing firm conclusions and no samples from earlier than 11 wpi were available to compare whether RC differed between HRPs and LRPs during primary infection. Moreover, the RC assay tested only one part of the viral genome that may not fully capture viral replication capacity. Yet, the data suggest that neither T cell activation nor an initial infection with fitter viral variants would explain the difference in dS substitution patterns between HRPs or LRPs carrying the HLA-B*5701 allele.
An intriguing alternative can be hypothesized by considering the highly significant inverse correlation between baseline CD4+ T cell count and average dS rates along branches representing lineages effectively propagating through time. The finding suggests a mechanistic link between CD4+ T cell count and the virus replication rate [33], [36], by indicating that HLA-B*5701 subjects with CD4+ T cell counts >750 cells/mm3 within the first 10–11 weeks of the infection will keep HIV-1 replication under better control during the subsequent years. This observation is in agreement with earlier results showing that subjects with a stronger immune system during early infection exhibit more constrained viral evolution, probably linked to a more robust HLA-B*5701-specific CD8+ T cell response [5]. In other words, the higher polyfunctional responses observed in these subjects [5] coupled with a larger number of CD4+ T cells during early infection may ultimately result in an overall slower in vivo replication rate of the virus. There is evidence that emergence of escape mutations in p24, as a consequence of CD8+ T cell responses, can negatively affect viral fitness [16], and thereby be indirectly responsible for control of viral replication, longer generation times, and lower risk to progress to AIDS.
Replication rates can also depend on the ability of the viral strains to replicate in different environments [16]. Differences in the contribution of latent HIV-1 reservoirs, such as resting memory CD4+ T cells, to the circulating virus population can impact mean generation times and replication rates even though the may produce only a fraction of circulating viruses [26], [27]. Further work will be necessary to clarify the relationship between HIV-1 generation times and replication dynamics in different viral reservoirs. Even though we included more HLA-B*5701 patients than in previous studies, our sample size remains small and conclusions need to be taken with caution. Regardless, our findings provide, for the first time, a possible evolutionary mechanism for different risk of disease progression in HLA-B*5701 subjects. They indicate that subjects who maintain high CD4+ T cell counts in early infection are more likely to control HIV-1 replication for an extended time and that synonymous substitution rates, which are proportional to in vivo replication rates, could be used as a novel evolutionary marker of disease progression.
The University of California, San Francisco (UCSF) Committee on Human Research, the Regional Ethical Council in Stockholm, Sweden (2008/1099-31), and the University of Florida review board approved this study. All patients provided written informed consent and all clinical investigations were conducted according to the principles expressed in the Declaration of Helsinki.
The study included six untreated HIV-1 subtype B infected men (P1-P6), all carrying the HLA allele B*5701, from the OPTIONS cohort [39]. Patients were enrolled within six months of HIV seroconversion and followed longitudinally. Patients' details have been described in a previous study [5]. Briefly, five of them were men who have sex with men (MSM) and one (P5) was an injecting drug user (IDU).
Each patient-specific data set included HIV-1 gag p24 sequences obtained by single genome sequencing of longitudinal plasma samples as previously described [5], [40]. GenBank accession numbers for the sequences analyzed in this study are: JX234575-JX235332. All sequences included in the present study were non-recombinant, as previously described [5]. The presence of molecular clock signal in each patient data set was investigated by regression between root-to-tip divergence and sampling date using ML likelihood trees inferred with the best fitting nucleotide substitution model, chosen by a hierarchical likelihood ratio test, as previously described [5]. For each data set, the Markov chain Monte Carlo (MCMC) sampler implemented in BEAST 1.7 [41], was used to obtain a posterior distribution of trees under a relaxed molecular clock model with the best fitting population prior (according to estimated Bayes Factors) [5], [42]. The selected population prior for each data set was the Bayesian skyline plot. The approach to infer synonymous and nonsynonymous substitution rates, and to explore how these rates change through time, is an empirical extension of the coalescent-based Bayesian relaxed clock models [20], [42]. Briefly, a subsample of 200 trees was randomly selected from posterior distribution and used to re-estimate branch lengths proportional to either nonsynonymous or synonymous substitutions according to the method described in Lemey et al [20]. For each clock-like genealogy, the rate of absolute nonsynonymous and synonymous substitutions was estimated including all branches in the genealogy, as well as internal and external branches only. The weighted average of the evolutionary rate was also calculated for the rates estimated along all the possible backbone paths of a genealogy (weighted by the number of branches along each path). Backbone paths represent lineages propagating (i.e. effectively surviving) from root node to sequences sampled at the last time point through sequential population bottlenecks. For each data set, HIV-1 among-site nonsynonymous and synonymous rate variation was analyzed by comparing two nested models with the likelihood ratio test: the constant rate variation model, which assumes that neither nonsynonymous nor synonymous rates vary across sites, and the dual random effects likelihood (REL) model, where site-specific nonsynonymous or synonymous rates are drawn from independent general discrete distributions with three rate categories [31], [43]. The dual REL model estimates the coefficient of variation (CoV), defined as standard deviation/mean for nonsynonymous or synonymous rates across sites. A large CoV and a low p-value for the test comparing the dual model with the null hypothesis (constant model) that CoV = 0 indicate significant rate variation from codon to codon in the alignment.
Viral replication capacity (RC) was measured in vitro using the PhenoSense Gag-Pro assay [44]. Sequences of gag and protease genes were amplified from patients' plasma by RT-PCR and transferred into a resistance test vector (RTV) containing a luciferase reporter gene. Transfections of HEK293 cells with patient-derived gag-pro RTVs and an amphotropic murine leukemia virus envelope expression vector were performed to generate pseudovirus stocks for infection of HEK293 cells. Gag-pro mediated RC was determined by measuring the viral infectivity (luciferse activity) of patient-derived pseudoviruses relative to NL3-4, the reference control, and expressed as a percentage. Immune activation data were also available for all subjects. The proportion of CD4+ and CD8+ T cells expressing CD38 was measured at 13–17 weeks post infection (wpi), as previously described [23].
A one-tail Mann-Whitney U-test was carried out with an online calculator using the normal approximation (http://elegans.som.vcu.edu/leon/stats/utest.html) to assess whether substitution rates in HRPs were significantly higher than in LRPs. Slopes of CD4+ T cell counts and viral load (VL) were obtained from least squares regression of log-transformed CD4 counts and VL over time (years). Model coefficients were back transformed and converted from proportions to percentage effect by subtracting one and multiplying by 100 to obtain individual estimates of percent change over time. Mean nonsynonymous and synonymous rates for different set of branches (all, internal, external and backbone paths) were compared to the corresponding clinical parameters of each patient (baseline CD4 count, baseline VL, CD4 slope, VL slope and baseline T cell activation) using Pearson's linear correlation to calculate the associated t-values and assess significance. All p-values obtained from applying any test statistic multiple times were adjusted with the Bonferroni correction.
|
10.1371/journal.pcbi.1004211 | Optimising and Communicating Options for the Control of Invasive Plant Disease When There Is Epidemiological Uncertainty | Although local eradication is routinely attempted following introduction of disease into a new region, failure is commonplace. Epidemiological principles governing the design of successful control are not well-understood. We analyse factors underlying the effectiveness of reactive eradication of localised outbreaks of invading plant disease, using citrus canker in Florida as a case study, although our results are largely generic, and apply to other plant pathogens (as we show via our second case study, citrus greening). We demonstrate how to optimise control via removal of hosts surrounding detected infection (i.e. localised culling) using a spatially-explicit, stochastic epidemiological model. We show how to define optimal culling strategies that take account of stochasticity in disease spread, and how the effectiveness of disease control depends on epidemiological parameters determining pathogen infectivity, symptom emergence and spread, the initial level of infection, and the logistics and implementation of detection and control. We also consider how optimal culling strategies are conditioned on the levels of risk acceptance/aversion of decision makers, and show how to extend the analyses to account for potential larger-scale impacts of a small-scale outbreak. Control of local outbreaks by culling can be very effective, particularly when started quickly, but the optimum strategy and its performance are strongly dependent on epidemiological parameters (particularly those controlling dispersal and the extent of any cryptic infection, i.e. infectious hosts prior to symptoms), the logistics of detection and control, and the level of local and global risk that is deemed to be acceptable. A version of the model we developed to illustrate our methodology and results to an audience of stakeholders, including policy makers, regulators and growers, is available online as an interactive, user-friendly interface at http://www.webidemics.com/. This version of our model allows the complex epidemiological principles that underlie our results to be communicated to a non-specialist audience.
| Increases in global trade and travel suggest outbreaks of plant disease caused by invasive pathogens will increase in frequency. We use mathematical modelling to show how control of such disease outbreaks can be optimised. Although our methods and analyses are generic, we use the attempted eradication of citrus canker from Florida (1996–2006) as a case study, and focus upon the performance of reactive culling (i.e. removal of all host plants within a certain distance of detected infection). We show how the cull radius can be optimised, even when there is significant cryptic infection (i.e. infection without visible symptoms). The inherent randomness of disease transmission implies a control strategy can lead to a number of outcomes: the optimal strategy therefore depends on the level of risk that is tolerable. We also consider balancing local vs. global impacts of disease. We show how it can be optimal to control initial outbreaks very extensively, even though this would lead to many local removals, since timely local eradication reduces the risk of a devastating large-scale epidemic. Our model is available as an interactive, user-friendly interface at http://www.webidemics.com/, intended to illustrate the sometimes counter-intuitive epidemiological principles that underlie successful disease control.
| Impacts of invading pathogens can be extremely severe, and so understanding how controls can be optimised is imperative [1]. We focus here on plant disease, motivated by the serious and potentially irreparable ecological damage that can follow introductions of plant pathogens into natural host populations [2], and the obvious food security and economic implications of epidemics in crops [3–5]. Increased global trade and travel mean the risk of introduction of exotic pathogens can only reasonably be expected to increase [6], which in turn indicates control of invasive plant disease is likely to remain important for many years to come.
We target optimising reactive eradication of small-scale outbreaks of an invading plant pathogen [7–10] occurring in regions extending from 1-10km. We concentrate on how cryptic infection (i.e. infectivity without symptoms), the inherent stochasticity of epidemics, and uncertainties in the parameters controlling disease spread affect the performance of control via local removal of plant hosts in the vicinity of detected infection. Effectively controlling in this fashion is extremely challenging, requiring an estimate of how far the epidemic has spread ahead of visibly infected regions. Nevertheless, recent modelling studies of a number of plant pathogens have shown how, in principle, control by local removal of susceptible hosts can be effective in managing plant disease [11–13], even when disease spreads non-locally via complex contact networks [14–18]. A consensus has begun to emerge that this type of control can be successful, albeit with the beguilingly simple proviso that there is a need to “match the scale of control with the intrinsic scale of the epidemic” [19]. The obvious problem is that the appropriate scale is very difficult to define, and depends in a complex fashion on the interplay between the epidemiology of the plant-pathogen interaction, the spatial distribution of susceptible hosts, the implementation and logistics of detection and control, and the current state of the epidemic. The quantitative detail of how these factors affect the nature of the optimal control strategy and how effectively it performs is extremely complex, and general principles remain ill-understood. However, identifying such general principles is clearly relevant to the control of all plant diseases.
Here we use mathematical modelling to investigate epidemiological principles underlying successful control. We consider a range of strategies for management of a newly invading plant pathogen, and identify optimal control scenarios that minimise the “epidemic impact”; we define this to be the total of both the number of hosts lost to disease and healthy hosts removed by control. We show the importance of allowing for the inherent stochasticity of epidemics in comparing control scenarios by analysing the empirical distribution of epidemic impacts for fixed values of epidemiological and control parameters. In particular, we show how the optimal control strategy changes when we take account of different levels of risk aversion [20]. We further analyse the effects of critical epidemiological and logistic factors on disease control. The following are considered: the rate and spatial scale of disease spread, the initial level of infection, the ability to detect disease, the length of time infection remains cryptic, the frequency of surveying and any notice period or other delay before removal. We also show how accounting for the risk of export of inoculum outside of the region of immediate interest leads to optimum control strategies that differ from those derived by focusing solely on local impacts [21,22]. Finally we consider how control is adversely affected if pathogen spread occurs via a thick-tailed dispersal kernel, although we demonstrate that even then an optimal control strategy can still be defined.
Although the issues we investigate are generic, we illustrate our work in the context of an economically important system for which eradication by reactive culling was attempted: citrus canker (caused by the bacterium Xanthomonas citri pv. citri) on commercial and residential citrus in Florida. The United States government spent over one billion dollars on survey, control and compensation costs during an eradication campaign that ran from 1996 to 2006 [9] (see S1 Text for further details). Aside from the attraction of choosing a prominent and controversial real-world example to frame our analyses, a major motivation for using this system as our case-study is an extremely detailed data set on disease spread in five uncontrolled sites in the Miami region, originally collected by the United States Department of Agriculture [23]. These data have allowed parsimonious, stochastic, spatially-explicit, epidemiological models to be fitted to the spread of citrus canker that track the disease status of individual host plants [11,12,24,25], and here we use a flexible extension of these models to analyse the effectiveness of the control scenarios we consider. We also take advantage of recent work fitting this type of model to huanglongbing disease (also known as HLB or citrus greening) [26], a disease of citrus vectored by psyllids and that is caused by Candidatus Liberibacter spp. bacteria, in order to demonstrate the flexibility of our methods and the generality of the underlying principles.
A principal challenge in using epidemiological models to inform policy is rendering the assumptions and outputs of models in forms that can be readily understood and interrogated by policy makers [27]. What is needed is a tool to allow epidemiological and disease control scenarios to be explored by regulatory decision makers. This would not only help ensure appropriate action is taken, but would also help ensure factors affecting the success (and the risk of failure) of a preferred control strategy are understood by those who have to make and justify decisions. Epidemiologists have typically adopted a “black-box” approach, in which the analytical process is hidden and where only the resulting control recommendations are delivered. This lack of transparency makes it difficult and in some cases impossible for stakeholders affected by control to question the scientific basis of decision making, leading to controversy and even to less effective control [28]. Accordingly we introduce a user-friendly interface to our model, enabling the effects on control performance of changes to disease spread and control parameters to be explored, which as we demonstrate here can include applying the model to different pathosystems. By allowing for an ensemble of simulation runs with identical parameters, this ‘front-end’ also allows stochastic variability to be visualised. The front-end is available online at http://www.webidemics.com/ (Webidemics is a backronym: (WEB)-based (I)nteractive (D)emonstration of (E)pidemiological (M)odelling (I)nforming (C)ontrol (S)trategies). The Webidemics interface demonstrates the challenges inherent in optimising control strategies that account for cryptic infection, stochasticity and uncertainty in parameter values.
Our model readily accommodates an arbitrarily complex host landscape, from regular geometrical patterns typical of agricultural and horticultural crops, to random and clustered patterns typical of plant hosts in natural environments. With that generality in mind, we illustrate the approach for a random host landscape typical of urban Florida, with 2000 trees randomly distributed on a 2km x 2km square at a plausible density for dooryard citrus (i.e. trees in residential gardens) [12,25]. We note our Webidemics interface also accommodates a citrus grove (the colloquial term for what is also referred to as a “citrus orchard”), with 2016 hosts in two adjacent blocks, planted in rows 10m apart and with a 5m within-row host spacing, reflecting standard practice in the U.S. citrus industry. For brevity all results presented in this paper for citrus canker correspond to the random host landscape (note, however, that our application of the model to HLB considers disease spread through the citrus grove host landscape).
We use a spatially-explicit, stochastic, individual-based, compartmental model to represent the spread of a plant pathogen through a population of N hosts (Fig 1a). Hosts are categorised by disease status: (S)usceptible hosts are uninfected; (E)xposed hosts are latently infected, and so are neither symptomatic nor infectious; (C)ryptic hosts are infectious but still asymptomatic; (D)etectable hosts are symptomatic but not infectious; (I)nfected hosts are both infectious and symptomatic; and (R)emoved hosts are epidemiologically inert, either because of disease-induced death or because the host has been removed by any control effort. The Webidemics interface allows the timing of the cryptic and detectable classes to be exchanged, with visible symptoms either preceding or following the onset of infectiousness, and therefore can represent any of the SECI, SEDI, SECIR and SEDIR epidemic models [19]. Here, motivated by the biology of citrus canker [24,29,30], we concentrate exclusively on the SECI and SECIR variants of the model, in which detectable symptoms strictly follow infectiousness. This is the case in which control is most difficult, since it is hampered by invisible cryptic infection. We note that, although citrus canker does not itself directly kill host plants, accounting for control requires there to be a removed compartment in the model, as does allowing the user of our Webidemics interface to apply the model to host-pathosystems for which there is disease-induced host death.
In the absence of control, the E to I and C/D to I transitions occur at fixed rates γ and σ, and so waiting times in these compartments are exponentially distributed, with means 1/γ and 1/σ, respectively (see Table 1). This assumption could of course readily be relaxed to allow for other distributions of waiting times [31]. In the SECIR and SEDIR variants of the model, the rate of disease-induced death is μ, again with an exponentially distributed transition time (mean 1/μ). For any given host plant, the rate of the S to E (susceptible to exposed) transition depends on the status of other hosts and the suitability of the environment for infection. In particular, if host i is susceptible at time t then it becomes latently infected (i.e. transitions to the E compartment) at rate
ϕi(t)=ω(t)(ε+β∑jK(dji;α)),
(1)
where the summation runs over all infectious hosts, j, and where host j is at distance dji from host i. The underlying maximal rates of primary and secondary infection are ε and β, respectively, and ω(t)≤1 parameterises any time-variation in environmental suitability for infection (see below).
The dispersal kernel, K(d;α), reflects the probability that an infectious host causes infection of a susceptible host at distance d, and is governed by scale parameter α. To allow robustness to the form of dispersal to be explored, we consider two contrasting kernels: the thin-tailed exponential kernel, K(d;α) = A exp (–d/α), and the thick-tailed Cauchy kernel, K(d;α) = A/(1+(d/α)2). In each case the normalising constant A is set via (1/N)∑i∑j≠iK(dij;α)=1.
We assume that the host landscape is surveyed for disease at regular intervals Ts, starting at time t = T0, and on each survey any symptomatic (i.e. class D or I) hosts are independently detected with probability p. Detected hosts are flagged for subsequent removal, together with any other hosts (irrespective of disease status) within a pre-determined distance L of each detected focus. In practice, removal actually occurs after a variable time delay, and we assume this is normally distributed with specified mean (Tc) and standard deviation (σc). This allows for logistic delay(s) in control, including notice periods to allow for legal challenges, or delays in deployment of requisite equipment and/or manpower. A truncated normal distribution is used to ensure that all delays are positive and so that removal occurs strictly after detection. Removed hosts are not replanted in our model, in keeping with the original practice for citrus canker in Florida.
We model environmental suitability for pathogen spread via a discrete-time Markov chain with two states, (S)uitable and (U)nsuitable (Fig 1b). This controls ω(t) in (Equation 1), with ω(t) = ωs = 1 for state S and ω(t) = ωu ≤ 1 for state U. The value ωu is therefore a measure of the relative unsuitability of state U. A probabilistic transition between states potentially occurs every Tw units of time. The probability of entering either state then depends only upon the current state, with p(U | S) = η and p(S | U) = ρ. At equilibrium, the probability of the environment being suitable (i.e. in state S) is π=ρη+ρ, and so the mean value of ω(t) is π + (1–π)ωu. The initial state is chosen randomly according to π, ensuring that the equilibrium properties of the chain control its statistics.
Markov chains offer a parsimonious approximation to environmental dynamics in a number of contexts [32], and a similar two-state formalism has previously been used to model infection rates of plant pathogens [33]. An advantage of a Markov chain model is that it simply requires information on threshold conditions, for example temperature and humidity, that favour or inhibit infection, and these are likely to be known for many plant pathogens or can at least be quantified relatively easily. Although we acknowledge more extensive information concerning these drivers is in fact already well-known for citrus canker [29], in general this obviates the need to derive costly functional forms for relationships between propagule production and environmental driving variables. By default, however, in illustrating the use of the model we restrict our attention to the case in which ωu = ωs = 1.
Here we use the citrus canker system as a case study to illustrate general principles underlying the effectiveness of control. Parameters can readily be adapted via our front-end interface to reflect other pathosystems (cf. S2 Text, which describes the application of our model to HLB, and S1–S3 Figs., which show the results). As defaults we therefore use illustrative parameters informed by the biology [29,30] and adapted from previous models of citrus canker [11,12,24,25] (Table 1) to drive our mathematical model. The host population is surveyed every 90 days [30], symptomatic hosts are detected with probability 0.8, and host removal occurs exactly 60 days after detection [29]. The default cull radius is set to be 75m; this default radius was chosen to emphasise the range of outcomes that is possible for a single control strategy, even when all parameters remain fixed. Epidemics are seeded with two exposed hosts at t = 0 (a different pair of hosts for each realisation). The average latent period is 10d [29] and symptoms take an average of 110d to emerge following infection [11]. The dispersal kernel is exponential, with mean dispersal distance of 40m (i.e. α = 20m), and we take the rate of secondary infection to be β = 0.03d-1 (cf. the values of α = 37m and β = 0.036d-1 as used in the analyses of Cook et al. and Parnell et al. [11,12,24], after accounting for our normalisation of the dispersal kernel). We selected default dispersal scale and infection rate parameters that lead to slightly slower and more spatially-restricted spread in comparison with those in previous analyses. This allows us to present extensive sensitivity analyses to parameters that would be expected to make control more difficult (e.g. long cryptic periods, lengthy delays between detection and tree removal), without optimal controls degenerating to immediate removal of the entire population at the time of the first control for our rather small population of interest. However, we demonstrate the robustness of our results to this slight alteration of the parameters in S3 Text and S4–S6 Figs., in which we repeat a selection of the analyses using exactly the parameters of Cook et al. [24] as a baseline.
In explaining and discussing the practical use of this model with stakeholders, including policy makers, regulators and growers, it became apparent that providing a user-friendly version of the model for presentation was important to allow the inferences to be understood and visualised by non-specialists. We therefore developed an online interface to the model, allowing the results of either a single run or a small ensemble of runs to be explored, and also allowing for the alteration of parameters controlling disease spread and/or control. This Webidemics front-end runs in commonly-used web browsers via the freely-available Adobe Flash Player plug-in. It is an interface to a ‘back-end’ program that runs on a central web server; this is written in C, and is the component that actually performs the model simulations presented in this paper. Implementation of the back-end via Gillespie’s algorithm [34] allows the extensive replication (many millions of independent runs) that underlies the analytical results we present. It also allows the user of the front-end to obtain results from an entire ensemble of hundreds of replicate epidemics within a reasonable time. Parameters and results are passed between the front- and back-ends via a Perl CGI (“Common Gateway Interface”) wrapper program hosted on the web server.
To examine how control performance depends on the cull radius, we performed 10000 replicate simulations for epidemics spreading according to the default parameters (cf. Table 1), at control radii ranging from L = 0m to L = 500m (Fig 2a). The epidemic impact, κE (the total number of hosts lost to disease or control by the time of eradication) is highly sensitive to the cull radius, L. At small L, the region of cryptic infection surrounding detected trees is underestimated and the disease spreads widely; at large L, many healthy trees are unnecessarily removed. At intermediate cull radii, performance improves markedly, and an optimum radius can be uniquely determined if the objective is phrased in terms of minimising an average of the epidemic impact, κE. For example, median κE is minimised at cull radius L = 159m, with a median of 132 hosts removed from the total of 2000. The radius L = 159m is therefore optimal in the sense of previous modelling studies [11,12,24], and the “intrinsic scale of the epidemic” sensu Gilligan [35] has therefore been identified at this radius.
However, focusing solely on average performance ignores elements of the response of epidemic impact (κE) to the cull radius (L) that may be of practical significance. The distributions shown in the inset to Fig 2 for a selection of radii (L = 50m, 75m, 100m, 150m, 200m and 400m) allow the following pair of provisos to this naïve optimum to be identified.
In practice optimisation of any control strategy would be driven by parameters estimated for pathogen spread and epidemiology, and these would be subject to error and/or uncertainty. This tends to make the sensitivity of epidemic impact (κE) to changes in cull radius identified in i) (above) unworkable, since an optimum strategy (distribution D) that skirts so close to failure (distribution C or more dramatically B) would almost certainly be difficult to recommend in practice. A pragmatic choice would therefore be to focus on a higher percentile of the distribution of κE when prescribing the control, with the particular percentile selected corresponding to the risk-aversion of the decision-maker. This approach can be formalised, by explicitly considering a risk of failure that is deemed to be acceptable (Fig 2b). In particular, given a notion of an acceptable level of risk (e.g. at most a 10% chance of κE corresponding to the loss of more than Ω = 20% of hosts) a range of acceptable cull radii can readily be determined. A workable strategy would then be to select a cull radius near the centre of this range. Such a combination of criteria would lead to a prescribed cull radius of around 225m here (for the default parameters, p(Failure at risk = 20%) < 0.1 for 122m < L < 329m).
For simplicity, we revert to using the median epidemic impact to summarise the efficacy of intervention, although we note the criterion could be readily adjusted as described above to suit different degrees of risk aversion. The optimal cull radius L is surprisingly unresponsive to the initial level of infection, E0 (Fig 2c); the “correct” radius is virtually unaffected by the epidemic size when control starts. However, the corresponding epidemic impact increases very rapidly, and, for example, when only 2.5% of hosts are initially infected (uniformly at random), approximately 80% of hosts would eventually have to be removed before the pathogen was eradicated, even when controlling optimally. This confirms and quantifies the intuition that it is important to act quickly when confronted by a new outbreak, particularly if the initial infections are not clustered in space. The optimum cull radius and the performance when controlling optimally also depend strongly on the scale of pathogen dispersal (α) and the rate of secondary infection (β) (Fig 2d and 2e).
We illustrate typical decisions that must be made by policy-makers by focusing on the effect on control performance of a selection of four parameters that may be changed during the course of an eradication scheme (Fig 3). These are the average cryptic period, (1/σ), which may vary depending upon environmental conditions; the probability of detecting a symptomatic host (p), which may vary depending on the experience of the teams of observers; and the interval between successive surveys (Ts) and the delay period before culling actually occurs (Tc), which are both controlled by the availability of resources. As 1/σ, Ts or Tc increase, or as p decreases, effective control becomes more challenging, and so the optimal cull radius and epidemic impact at this radius both increase. Particularly striking is that extreme changes to the probability of detection and average cryptic period are required for performance to degrade significantly. The influence of changes to either of these is mitigated by averaging over a large number of hosts: only a single host must become detectable or be detected for control to be initiated locally, and the resulting cull then affects many nearby hosts simultaneously. However, because the survey interval and the notice period both affect all hosts equally, more modest changes affect the success of control to a greater extent.
Susceptible hosts will almost always be present outside the area of first detection. However, we have focused on control performance on a small landscape; in this sense we have considered only the “local” impact of the epidemic. In practice “global” impacts (i.e. on all plants that could possibly become infected, irrespective of location) would also need attention in designing control strategies.
A possible proxy for the risk to the area outside the region of immediate interest is the time taken to eradicate the local epidemic (“epidemic time”, τE), since this sets the duration of possible export of inoculum (Fig 4a). Surprisingly the epidemic time (τE) initially increases as the cull radius (L) increases, at least for L below the optimum that minimises the local epidemic impact. While such controls fail to keep up with the region of cryptic infection surrounding detected hosts and so do not effectively control the epidemic, a proportion of infected plants is detected and removed on each round of surveying, and this causes the epidemic to spread more slowly (because infected hosts are being removed and there are fewer susceptible hosts to infect). Slower spread then allows the pathogen to persist for longer, since it takes more time for the infection and eventual removal of hosts. For larger control radii, however, we note the epidemic time can be very small; the pathogen is eradicated very quickly, generally within one or two rounds of detection and control.
The appropriate balance between local and global impacts is necessarily a pragmatic choice to be made by the decision maker, and it is impossible for us to be too prescriptive. We account for this need for flexibility by introducing a tuneable composite measure of global “epidemic cost”, ΨE, intended to balance the epidemic impact and epidemic time, allowing for different weightings of local vs. global impacts. In particular, we define the normalised epidemic cost via
ψE=(1-η)κ^E+ητ^E,
where κ^E and τ^E are simply the epidemic impact, κE, and the epidemic time, τE, normalised to a [0,1] scale (this is done by dividing by the maximum of the median values over all cull radii; for κE this is 1980 hosts at L = 0m, whereas for τE it is 3200d at L = 63m). The dimensionless weighting parameter η then controls the relative importance assigned to global impacts. In particular, taking η = 0 means only local impacts would be considered, with normalised epidemic cost ψE= κ^E (i.e. the normalised epidemic impact), whereas η = 1 would entirely focus on impacts outside the region of immediate interest, with ψE= τ^E (i.e. the normalised epidemic time). As η is increased, the cull radius that minimises ΨE is increased (Fig 4b): the larger weighting given to global impacts means that it becomes increasingly optimal to control very aggressively to eradicate the local epidemic as quickly as possible, despite the large number of local removals that would then be required.
The epidemic time, τE, may be of particular significance to policy makers, since the duration of a control programme will be an important determinant of public opinion. However, the time taken to eradicate the pathogen is of course not the only way of characterising the risk of pathogen spread outside the area that is actively being controlled. The area under the disease progress curve
AE=∫t=0τE(C(t)+I(t))dt,
quantifies the total amount of inoculum that would be exported over the course of the entire epidemic (Fig 4c). This can be used to calculate the probability of at least one escape to the region that is not being controlled, pE, via
pE=1−exp(−λAE),
where λ is a measure of the degree of connectivity between the local and non-local populations of host plants. In principle the connectivity (λ) could be determined for any particular landscape structure, although as we show here, the response of pE to the cull radius (L) is robust to extremely wide variations in the value of λ (Fig 4d). Assuming that a single escape from the region under active control would be sufficient to initiate a global epidemic, we can then define a variant measure of epidemic cost via
ζE= (1-δ) κ^E + δ p^E,
where δ controls the importance assigned to local vs. global impacts (i.e. δ plays the same role in the definition ζE of as does η in the definition of ΨE), and where p^E is pE normalised to a [0,1] scale (by dividing by the maximum value of pE, which occurs at L = 0). The response of ζE to the weighting parameter δ is similar to the response of ΨE to η (cf. Fig 4e, the response to different values of δ when λ = 10-5d-1), and the conclusion that the optimal radius increases with increasing the weighting of global impacts, δ, is robust to all values of the connectivity, λ, we consider, over a range of orders of magnitude (Fig 4f). Again, very extensive controls in the region under active management become optimal when the possible global impacts of disease are judged to be important.
Neri et al. [25] recently fitted a model of the type we use here to the dataset on the spread of citrus canker in Miami that we described in the Introduction (note this is also the dataset used by Cook et al. [24]). In common with Cook et al. [24], Neri et al. [25] found that an exponential dispersal kernel was best-supported by the data. However these authors found only a small difference in model goodness of fit between the exponential and Cauchy kernels. Neri et al. [25] suggest this partial lack of identifiability is driven by the effect of continual primary infection (i.e. infection from outside the study site, within which disease spread was mapped). At large distances from infected hosts, the small and slowly decreasing probability of infection that would be associated with Cauchy dispersal is very difficult to distinguish from the small and effectively unchanging probability that would follow an exponential kernel combined with a constant background rate of primary infection caused by fat-tailed dispersal from one or more distant sources of inoculum, or by anthropomorphic introduction of inoculum on implements, clothing or cuttings. The study sites were relatively small (<10km2) uncontrolled regions embedded within a large ongoing epidemic. A non-zero rate of primary infection from outside was necessary to fit the spread data in these sites, since there was significant ingress of infection from outside each site. Here, since we specifically target an isolated outbreak of emerging plant disease, far from any large source of inoculum, primary infection is not required in the analyses in the current paper. We accordingly set the rate of primary infection to zero throughout our analyses, and default to using an exponential dispersal kernel, for consistency with the fitting of Neri et al. [25] and Cook et al. [24], together with the previous analyses of Parnell et al. [11,12].
It is well known that exponential dispersal leads to epidemics characterised by wavelike spread, whereas thick-tailed dispersal (exemplified here by the Cauchy kernel) implies continual production of distant secondary foci with no well-defined epidemic front [36,37]. The striking difference in epidemic pattern suggests that effective control via local removal of hosts should be more difficult when there is Cauchy dispersal. For purposes of comparison, we therefore test the effect of fat-tailed dispersal on our analyses
At low infection rates, effective control remains possible even with thick-tailed dispersal. For the Cauchy kernel with scale parameter α = 20m and infection rate β = 0.007d-1, good control can be achieved using a cull radius L = 100m (Fig 5a), although the long tail of the epidemic impact distribution (Inset B) reveals large epidemics are possible even when controlling optimally. However, successful control requires a small infection rate, and only moderate increases to the infection rate β greatly increase both the minimum median epidemic impact κE and the cull radius, L, at which this optimum κE is achieved (Fig 5b). What responses of the median κE to L and β do not reveal, however, is how quickly the chance of a large epidemic increases as the infection rate goes up. For an infection rate β of only 0.01d-1 there is a significant risk of failure of control, as indicated by the extremely variable distribution of epidemic impact for all control radii, even near the optimum cull radius L ~ 300m (Fig 5c). Indeed it is impossible to select a range of radii that leads to at most a 10% chance of losing less than Ω = 20% of hosts for this value of β (see Fig 5d and contrast with Fig 2c), reiterating the relative difficulty of control when there is fat-tailed dispersal. We also note that fat-tailed dispersal would be expected to increase the degree of connectivity between the local and non-local populations of host plants, λ, were we to repeat the analysis associated with Fig 4 using such a dispersal kernel. Taken together these observations suggest that successful control by culling may be difficult for plant pathogens that spread via windborne propagules which are tolerant to desiccation. These include powdery mildews and rust pathogens. Effective control of these pathogens is likely to require large local cull radius and could be expected to have relatively high risks of failure.
The Webidemics interface (http://www.webidemics.com/) shows either the results of a single realisation of the model (Fig 6a) or summarises an entire ensemble of replicate simulations performed using identical parameters (Fig 6b). When results from a single run are displayed, an animation of disease progress is shown, with hosts colour coded by epidemiological type (i.e. S, E, C or D, I, R), and with any hosts set to be inaccessible for detection denoted by a black cross. Any hosts in compartment R that were removed by disease-induced death are distinguished from those culled by control. The same colour coding is used for the graph showing the number of hosts in each class.
In the single run screen, the graph shows the time-dependence in the number of hosts in each class for the realisation being shown. However, in the ensemble of runs screen, the graph tracks the time-dependence of the average number of hosts in each compartment. The left hand panel then shows animated histograms of the numbers of hosts in each compartment over time. Clicking on any bar in a histogram switches back to the single run view, displaying a (randomly chosen) realisation from within the original ensemble that had a number of hosts within the range of the chosen histogram entry at the relevant time.
Epidemiological, climatic and/or control parameters may be set on either screen, using the three buttons at the bottom right: clicking any of these buttons reveals a pop-up panel allowing parameter values to be set (Fig 6c). When parameters are altered, the displayed results are not updated until a call is made to the back-end to actually run new simulation(s) with the new parameters (this is done by clicking the “Run New Simulation”/“Run New Ensemble” button). Changes to parameters that have not yet been followed by a call to the back-end and so are not reflected in the current results are indicated by a colour change of the button from grey to red. Further details of the user-interface are available via its help facility, which includes a full description showing how to use model in practice, designed for first time users.
The default cull radius of 75m as used in the front-end was chosen to emphasise how stochasticity can affect the effectiveness of a single control scenario when all parameters are fixed. We show here the results from this scenario, using fixed default epidemiological and control parameters (see Table 1), as seen by the user of our front-end interface. Control efficacy is extremely variable (Fig 7a and 7b). The realisation shown in Fig 7a leads to fewer than 10% of hosts (159 from 2000) removed before eradication at 790 days. However the simulation run shown in Fig 7b reveals the risk of a far greater epidemic impact, despite identical parameters controlling disease spread and control. A small proportion of asymptomatic but infectious trees escape control on each round of removal, and this leads to widespread disease. Nearly 90% of hosts (1743 out of 2000) are eventually removed before the pathogen is fully controlled at t = 4300d. Similar behaviour is easily observable via the interface (cf. Fig 7c, which shows typical histograms summarising the final state of 500 runs using the default parameterisation).
We present a novel stochastic analysis of the control of plant disease, focusing on how the performance of reactive control by localised culling can be optimised. We have also introduced Webidemics, an interactive online tool designed to communicate principles affecting effective control to an audience of stakeholders, including policy makers, regulators, growers and scientists. Default parameterisation of the underlying epidemiological model targets the spread of citrus canker in Florida by adapting the parameters of previous modelling studies, although these defaults can readily be altered to represent other model parameterisations or even pathosystems by the user of our front-end interface. The analysis and the user-friendly interface address and illustrate the challenges posed by cryptic infection, stochasticity and uncertainty in parameter values, and demonstrate how these factors must be accounted for in designing successful disease control strategies.
Our key result is to verify that it is indeed possible to optimise control via targeted host removal by matching the “intrinsic scale of the epidemic” [35], selecting a cull radius that minimises the epidemic impact, κE (i.e. the total number of hosts lost to disease or control before the pathogen is eradicated). However, given particular parameters controlling disease spread and the logistics of control, we have shown how the cull radius that would be selected depends on the percentile of the epidemic impact distribution that is to be optimised over, and therefore on the risk-aversion of the decision-maker. Since costs of disease are typically greater than costs of detection and control, under-control can be more harmful than over-control [22]. This is reflected in the sharp increase in epidemic impact, with even small decreases in the cull radius below the optimum (Fig 2), a pattern that is largely unresponsive to the values of the parameters (Fig 3). The pattern also holds for other baseline sets of parameters (S3 Text) and also when applying the model to other pathosystems (S2 Text). We have also confirmed the intuition that control should start quickly to be successful; if even a small proportion of hosts is infected when intervention commences, a large epidemic impact appears unavoidable, particularly if the initial infections are not spatially-clustered (Fig 2c).
We investigated how the nature of the optimum control strategy is conditioned on a selection of parameters controlling the host-pathogen interaction and the logistics of control. Our results show how success of control depends strongly on the rate and spatial scale of pathogen spread (Fig 2d and 2e). We also showed how factors that act differently at the level of the individual host can have less severe effects as they are altered than those that affect all individuals equally, as a consequence of the former being averaged over the entire population (Fig 3). While control by local culling can remain viable when there is thick-tailed dispersal, control is only then successful for low infection rates (Fig 5). Even moderate rates of spread due to higher infection rates (β) lead to too many new disease foci when there is fat-tailed dispersal, and, in turn, this means that control is often unsuccessful, with significant risk of large epidemic impacts. This high risk of failure remains hidden if only the median epidemic impact is considered (Fig 5c), re-emphasising the need to examine the full distribution of outcomes when assessing the efficacy of control.
Our purpose here was to illustrate and test some key principles underlying the success of control. For this we selected a biologically-plausible set of parameters for citrus canker that allowed extensive sensitivity analyses to factors that are likely to affect the success of control. Our initial analyses focused on the optimal cull radius to eradicate disease in local outbreaks following a small number of primary infections in a spatially restricted (2km x 2km) urban host landscape. Our analyses show that the magnitude of the optimal cullradius depends upon the degree of risk aversion to failure of control, calculated from percentiles for the probability of a given epidemic impact that accounts for the cost of disease and the cost of control. Values for our default parameterisation for citrus canker varied from 104m for the 5th percentile of local epidemic impact to 194m for the 95th percentile, with an optimal radius of 159m when optimising over the median impact (Fig 2). However, while the responses of epidemic impact to changes in the cull radius, and of the optimum cull radius to different levels of risk aversion, are both robust, the basic estimate of the optimal cull radius can vary widely (from ~100m to ~500m), depending on changes to epidemiological and logistical parameters (Fig 3) and the underlying parameterisation selected for the model (S3 Text).
All of these figures relate to local control and are typically smaller than the cull radius of 571.9m used by USDA for statewide control of the citrus canker epidemic in Florida from 2002 [29]. One driver of this difference is that our default host landscape uses planting densities typical of residential citrus, whereas extended regions of commercial citrus would lead to faster spread due to higher planting densities. Moreover uncooperative landowners meant that in practice certain trees were inaccessible for pathogen survey, and legal challenges sometimes led to extremely long delays before other trees could be cut down. While both of these may be investigated via the front-end, our default parameters arguably downplay these effects (e.g. a fixed notice period of 60 days, when in practice legal challenges could lead to delays of many months or even years). More significantly, however, these initial estimates do not take account of the risk of infection spreading from the prescribed region of interest to surrounding regions. As we have shown here, to advance from optimisation at local to statewide scales in fact requires proper consideration of the balance between local and global impact of the epidemic (Fig 4).
Work targeting animal epidemics suggests optimal control strategies depend strongly on how local and global impacts are balanced [21,22]. We examined this by introducing two variants of the “epidemic cost”, that account for possible infections outside the area of interest via the proxies of local time to eradication or the probability of pathogen escape, but that are flexible enough to allow for different weightings of local vs. global priorities. Strategies giving a high weighting to global performance required extensive controls at the local level (Fig 4b, 4e and 4f). This is because the potential for spread of disease to create a new focus of infection elsewhere is judged to be so harmful that it becomes optimal to use a rather draconian policy in the region under active control, even at the cost of many local removals.
While we motivate our analyses using citrus canker, our work can readily be placed in a broader context. The recent emergence of citrus greening or huanglongbing (HLB, caused by Candidatus Liberibacter spp. bacteria), potentially an even more devastating disease [38], puts control of citrus pathogens firmly back on the scientific and political agenda in the United States. HLB is vectored by psyllids, and although it would be reasonable to assume dispersal of infective vectors declines monotonically with distance from infected plants, it might be expected that changes in psyllid populations over time would add extra complexity to disease dynamics. However, recent work has shown how our underlying model focusing only on disease status and representing disease spread via a time-independent and spatially-isotropic dispersal kernel can be applied to this pathosystem with no change to the fundamental model structure [26]. We recreated a selection of our results for the control of HLB in a citrus grove in S2 Text. We emphasise that using this type of model for HLB means that the activity and population dynamics of the psyllid population are not tracked explicitly, but instead that these factors are included in the dispersal and infection rate parameters of the model. The results concerning principles for control were qualitatively unchanged, although of course the exact detail of the optimum radius and epidemic impact were different reflecting a different pathogen and host topology.
Similar models are increasingly used for other plant pathogens at both small [11–13,18,24–26,39–42] and large spatial scales [10,16,36,43–47]. An emerging epidemic to which models are already being applied is sudden oak death (caused by Phytophthora ramorum) in the United Kingdom [48]: predictions from a larger-scale stochastic compartmental model are already informing the extent of felling of commercial larch [49]. The United Kingdom government’s response to Chalara ash dieback is also based on predictions from this type of model [50]. Models with static hosts have also been applied to pathogens of agricultural animals, most notably for epidemics of foot and mouth disease.
Our particular focus has been a simple control strategy, in which all hosts within a certain distance of detected hosts are removed, and where this distance is fixed in advance. While this corresponds to the approach most often taken in practice, and has definite advantages in terms of ease of implementation and transparency to those affected by control, recent modelling work has examined more elaborate strategies. In particular te Beest et al. [51] consider a complex and time-varying control strategy for an animal disease epidemic spreading through a set of farms that takes account of heterogeneity in the potential risk according to farms’ position and the current state of the epidemic. Our results are also conditioned on the metric used to define the epidemic impact. More complex notions of cost are possible; an obvious extension, for example, would be to include the cost of detection [14,15,18]. Although our model allows for fluctuations in environmental conditions and we allow the user to set parameters causing the pathogen to be affected by the environment via the front-end, we have not focused on these effects in this paper. We have also not accounted for the additional and significant difficulty in control of novel invasive pathogens for which the parameters controlling spread are themselves ill-characterised. Nor have we considered the effects of any spatial patterning of the host population in terms of, for example, systematic differences in host quality or differential resistance, although we note these spatial patterns would need to be well-characterised in order to be used in the model. We suggest that exploration of these issues, together with assessing and optimising the performance of control scenarios on spatially-extended host landscapes and particularly when there is thick-tailed dispersal, are important challenges. Further work is also underway to augment our theoretical work with interactive user-friendly front-end interfaces, after our extremely positive experiences in using the Webidemics interface to explain and present the ideas underlying our results to an audience of non-specialists.
|
10.1371/journal.pcbi.1001113 | Consensus-Phenotype Integration of Transcriptomic and Metabolomic Data Implies a Role for Metabolism in the Chemosensitivity of Tumour Cells | Using transcriptomic and metabolomic measurements from the NCI60 cell line panel,
together with a novel approach to integration of molecular profile data, we show
that the biochemical pathways associated with tumour cell chemosensitivity to
platinum-based drugs are highly coincident, i.e. they describe a consensus
phenotype. Direct integration of metabolome and transcriptome data at the point
of pathway analysis improved the detection of consensus pathways by 76%,
and revealed associations between platinum sensitivity and several metabolic
pathways that were not visible from transcriptome analysis alone. These pathways
included the TCA cycle and pyruvate metabolism, lipoprotein uptake and
nucleotide synthesis by both salvage and de novo pathways. Extending the
approach across a wide panel of chemotherapeutics, we confirmed the specificity
of the metabolic pathway associations to platinum sensitivity. We conclude that
metabolic phenotyping could play a role in predicting response to platinum
chemotherapy and that consensus-phenotype integration of molecular profiling
data is a powerful and versatile tool for both biomarker discovery and for
exploring the complex relationships between biological pathways and drug
response.
| Resistance to chemotherapy drugs in cancer sufferers is very common. Using a
panel of 59 cell lines obtained from different types of cancer we study the
links between the genes and metabolites measured in these cells and the
resistance the cells show to common cancer drugs containing platinum. In order
to combine the information given by the genes and metabolites we introduce a new
pathway-based approach, which allows us to explore synergy between the different
types of data. We then extend the procedure to look at a wider panel of drugs
and show that the pathways we found were associated with platinum are not just
the pathways which are frequently selected for a large number of drugs. Given
the increasing use of multiple sets of measurements (genes, metabolites,
proteins etc.) in biological studies, we demonstrate a powerful, yet
straightforward method for dealing with the resulting large datasets and
integrating their knowledge. We believe that this work could contribute to
developing a personalised medicine approach to treating tumours, where the
genetic and metabolic changes in the tumour are measured and then used for
prediction of the optimal treatment regime.
| In the quest to understand complex biological systems at multiple levels of
biological organization, the need arises to combine knowledge from experiments of
different types to create a full picture of a system's behavior. Modern
molecular profiling (“omics”) methods, such as transcriptomics,
proteomics and metabolomics allow one to build up a global picture of system
characteristics, and to search for interactions and coordinated behavior between the
different levels. While each level can be studied separately, greater statistical
and explanatory power can be gained by integrating this knowledge into a single
coherent model of the system. This is currently one of the greatest challenges in
systems biology.
Inter-omic data integration can be performed at different levels [1], the simplest of
which is conceptual integration. At this level, each omics data set is analysed
separately and a coherent biological rationale is constructed which explains
phenomena observed in the separate molecular profiles. For example, changes in
levels of both enzyme transcripts and metabolites from the same pathway could be
explained by the hypothesis of differential regulation of that pathway. However this
subjective approach can lead to plausible biological explanations that arise through
spurious statistical associations and conversely some potentially novel mechanisms
may be overlooked. The statistical level of integration is more objective. In this
approach, links between data sets are made using rigorous statistical procedures
such as correlation, regression or more sophisticated techniques. To date, much
inter-omic data integration has been performed at the conceptual level [2], [3], [4] while various
methods have been proposed and demonstrated for statistical integration [5], [6], [7], [8], [9].
Many researchers have found that interpretation of omics data at the level of
individual molecular entities can be difficult and have opted for an analysis at the
pathway or functional level [10]. This is mainly because particular changes in biochemical
pathways, associated with phenotypic conditions such as disease can often arise from
a range of different alterations in a pathway. A common method for performing
pathway-level analysis on single omic data is over-representation (OR) analysis
[11], [12], in which a set
of molecular elements (e.g. genes) that are differentially expressed or correlated
with the phenotype of interest are first selected. The set is then compared against
molecular sets defined a priori (e.g. genes in established
pathways) to identify those sets that show greater overlap with the
phenotype-associated genes than would be expected by chance. The final list of
significantly over-represented or ‘enriched’ sets/pathways is used to
aid biological interpretation of the data. As well as performing OR with genes,
Metabolite Set Enrichment Analysis (MSEA) [13] and other metabolite
over-representation techniques [14] have also been developed. In this work we contrast the
application of the OR analysis approach to transcript and metabolite data
individually to the alternative of considering them simultaneously, using
established pathways to guide an integrated analysis of the two data sets.
In addition to the inter-omic integration of metabolomic and transcriptomic data, our
approach involves a further type of data integration that we call
consensus-phenotype integration. In this approach, several examples of the same
phenotype, achieved in different ways, are used within the experimental design. For
example, one may study a particular mechanism of toxicity via the use of different
chemical treatments that have a similar mode of action. One can thus identify
features that are central to the phenotype in question across different types of
“omics” data, as opposed to features that are specific to a single
instance of the phenotype being studied.
In this work, we aim to elucidate mechanisms of drug sensitivity through the use of
inter-omic statistical data integration using drug sensitivity, transcriptomic and
metabolomic data from the NCI60 cell line panel [15]. The NCI60 is a panel of tumor
derived cell lines corresponding to diverse tissue types, which has been subject to
extensive molecular phenotypic and pharmacological characterization. We used
baseline (untreated) metabolic and transcriptional profiles readily available for 58
lines as well as growth inhibition data from an array of 118 drugs [15], [16]. We correlate
growth inhibition to the molecular profiles to identify pathways related to drug
sensitivity. We first focus on platinum sensitivity as it is a well-defined
phenotype, linked to a well-investigated mode of action that has important clinical
implications. Many chemotherapeutic regimes are based on platinum compounds, and
resistance to these drugs is a major obstacle in successful treatment of some
cancers. The mechanisms that cause variation in response to therapy are not well
understood, and the ability to predict sensitivity from a baseline profile of the
tumor would help to improve therapy selection and thereby potentially reduce patient
morbidity and mortality. We then expand our analysis to a larger set of 118 drugs to
investigate whether the method is able to associate drugs with similar modes of
action. We show that statistical integration conducted through a joint analysis of
the data gives specific advantages in terms of sensitivity and confidence of pathway
associations.
Figure 1 shows a schematic
overview of our data analysis strategy. Whole genome gene expression
(transcriptomic), metabolomic, and drug sensitivity data were obtained for the NCI60
tumor cell line panel. The transcriptomics data was derived using the U133
Affymetrix chip; in total 44928 probesets were measured, equating to 17150 gene
products mapping to distinct UniProt identifiers, each measured across 58 cell lines
[17]. The
metabolomic data consisted of measurements of the total intracellular abundance of
154 uniquely identified metabolites across all 58 cell lines [18], including lipid compounds
(e.g. cholesterol), glycolytic intermediates (e.g. glucose-6-phosphate), nucleic
acid metabolites (e.g. adenine, uracil, hypoxanthine) and amino acids (e.g.
glutamate, taurine). The full list along with our assigned KEGG IDs can be obtained
in Table S4.
We used drug sensitivity data (GI50 values indicating the concentration
of the drug which inhibited cell growth by 50%) [15], [16] initially for four
platinum-based chemotherapeutics, cisplatin, carboplatin, tetraplatin and
iproplatin. Data for a fifth platinum drug (diaminocyclohexyl-Pt(II)) was available,
and was used at a later stage as a test compound to validate our findings.
For each drug we ranked all probe sets by their absolute Pearson correlation
(|r|) to the −log(GI50)
values across all cell lines. Setting the false discovery rate (FDR) [19] at
60% we then selected genes considered to be significantly associated to
chemosensitivity. A high FDR was tolerated at this stage of the analysis to
ensure that subsequent pathway analysis was adequately powered. Repeating this
process for the metabolite data we obtained separate panels of genes and
metabolites that were deemed to be associated with the sensitivity to each drug
(see Table
S1). In total 3, 33, 37 and 92 metabolites and 915, 1620, 5035 &
6533 genes were identified as associated with sensitivity to carboplatin,
cisplain, iproplatin and tetraplatin treatment respectively.
To assess which pathways characterized the drug sensitivity phenotype we then
performed OR analysis with pathways from the ConsensusPathDB [20]. The
ConsensusPathDB collates pathways from several public databases of protein
interactions, signaling and metabolic pathways as well as gene regulation in
humans. We restricted our analysis to sources covering biochemical reactions:
KEGG [21],
Reactome [22], Netpath (http://www.netpath.org),
Biocarta (http://www.biocarta.com), HumanCyc [23] and the pathway interaction
database (PID) [24]. The use of multiple databases reduces bias by
enhancing coverage. At the time of analysis the ConsensusPathDB contained 1875
pathways from the selected sources, of which 1651 contain at least one gene and
581 contain at least one metabolite measured in the NCI60 data (excluding the
highly prevalent ‘currency’ metabolites phosphate, diphosphate and
NADP+). OR analysis of the phenotype-associated gene panels indicated that
63, 74, 233 and 242 pathways were associated with cisplatin, carboplatin,
iproplatin and tetraplatin sensitivity respectively (p<0.05). The equivalent
analysis for metabolite panels indicated that 24, 13, 4, & 5 pathways were
associated with these phenotypes.
To highlight pathways relevant to general platinum sensitivity, as opposed to
particular platinum compounds, we looked for pathways that were associated with
more than one drug response phenotype (‘consensus-phenotype
integration’; Figure 2 A
& B). Within the gene transcript analysis (Figure 2A), the drugs appeared to divide into
two pairs that shared many pathways in common. Iproplatin and tetraplatin were
most similar, sharing 143 (133+4+4+2) of the 330
(75+5+5+4+143+92+3+1+2), ie. 43%
of the pathways associated with either drug (Figure 2A). Carboplatin and cisplatin also
show a high level of similarity (32 of 103 pathways, 31%). In the
metabolic analysis (Figure
2B) the similarity between iproplatin and tetraplatin was much lower,
while carboplatin and cisplatin retained a high level of similarity with 7 of
the 30 pathways being shared (23%). The gene analysis highlights many
more pathways than metabolite analysis due to both the higher number of pathways
with sufficient numbers of quantified transcripts and the limited number of
quantified and identified metabolites.
We next combined the transcriptomic and metabolomic data into a joint inter-omic
OR analysis (Figure 2C) by
estimating the joint probability of association of each pathway with the drug
sensitivity phenotype assuming independence between the probability of
association from the gene and metabolite data separately (see Methods). 35 pathways were found to be
significant for at least one drug in the joint analysis that did not feature in
either of the separate analyses of gene expression or metabolite levels. To
confirm the significance of the increase in pathway detection after integration
of the metabolic and transcriptomic data, we estimated the null distribution of
the joint analysis probabilities by permuting the gene analysis pathway
probabilities relative to the metabolite analysis pathway probabilities. For
carboplatin only 3 of the 100 permutations produced more pathways than the real
data and for cisplatin no permutations produced as many pathways as the real
data. However, for iproplatin and tetraplatin, the number of pathways detected
was not significantly enhanced by the joint OR analysis, suggesting that the
combined analysis may be most advantageous when the numbers of significantly
associated genes or metabolites are relatively low.
To examine the significance of the numbers of pathways in the joint OR analysis
that were commonly associated to the effect of multiple drug treatments, two
null models were generated. Null model I assumed that genes and metabolites
identified as significantly associated to a phenotype were randomly selected
whereas null model II correspondingly assumed that pathways are selected
randomly. Table 1
summarizes the pathway coincidence between the output of joint OR analysis
across the four platinum drugs for these two null models compared to the real
data and reports the associated FDR in each analysis. We observed that by
setting our criterion of significant association between a pathway and platinum
sensitivity at requiring a majority of the drugs to be associated with that
pathway (i.e. at least 3/4) we achieved acceptable FDRs of 0.2% (null
model I) and 16.9% (null model II, the most extreme scenario).
Using the majority overlap criterion we compared the number of pathways
consistently associated with platinum sensitivity between the individual and
joint analyses (Figure 2D).
The joint OR analysis identified all pathways highlighted by the individual
–omic OR analyses combined (17 in total), but also indicated a further 13
pathways that were consistently associated (+76%). No pathways were
found to be common between both the separate gene and metabolite analyses.
Overall 30 pathways met the majority criterion of association with sensitivity at
least 3 platinum drugs and hence general platinum cytotoxicity (Table 2; Figure 2C & D). All the
databases used to compile the ConsensusPathDB contributed pathways to the final
selected consensus pathways, highlighting the value of the ConsensusPathDB
strategy in pathway analysis. While this subset of pathways included those with
established relationships to platinum sensitivity and general chemosensitivity,
such as DNA repair and Akt regulation of nuclear transcription, there were also
several pathways related to metabolic processes not previously reported as
determinants of platinum sensitivity. These included nucleotide metabolism,
fatty acid, triglyceride and lipid metabolism.
The added value of the inter-omic OR analysis prior to consensus phenotype
integration can be more clearly discerned at the individual pathway level. Figure 3 is a network
representation of the base excision repair (BER) pathway from Reactome and
depicts both the detected entities and the drugs with which each detected entity
is associated. While the majority of entities were significantly associated to
the effect of at least one of the four platinum agents, there was significant
variation in the pattern of association and no gene or metabolite was
significantly associated to all four treatments. Accordingly the pathway was
only significantly associated to tetraplatin and iproplatin sensitivity using
the transcriptome data alone, and to carboplatin and cisplatin sensitivity using
the metabolite data in isolation. Using the joint OR analysis the BER pathway
was significantly associated to the sensitivity to all four platinum compounds
(Table 2) and the
evidence for association with each drug was increased, due to the added
information from the alternative data type. Of the 12 pathways for which
inter-omic OR analysis improved the consensus between the drugs, 10 refer to
metabolic processes.
In order to validate and to test the generalisability of our findings we then
examined GI50 data from a test compound, diaminocyclohexyl-Pt(II).
After conducting the same inter-omic OR analysis as described previously, we
observed that the effects of this compound on the NCI60 panel was associated
with 5 of the 6 pathways common to all 4 other platinum drugs along with a
further 12 pathways from Table
2 and 90% (220/245) of the pathways associated with
diaminocyclohexyl-Pt(II) were Associated with at least one of the other platinum
drugs. In particular there were 138/152 pathways commonly associated between
diaminocyclohexyl-Pt(II), iproplatin and tetraplatin sensitivity. Since OR
analysis makes no distinction between positive and negative molecule/sensitivity
correlations, we also examined the direction of associations between the
metabolites detected in the consensus pathways and the GI50 of all
platinum drugs (Table 3).
In total, a panel of 22 metabolites were associated with the consensus metabolic
pathways from analysis of the four training compounds. While there was variation
in the metabolites associated with specific treatments, where a significant
association was observed the direction of correlation was consistent across the
training set. The GI50 values of our test compound,
diaminocyclohexyl-Pt(II), was significantly correlated to 19/22 metabolites in
this panel, with complete consistency in the direction of association with the
training set data.
To explore more broadly the relationships between chemosensitivity and biological
pathways across a range of agents, and to ascertain the specificity of the
consensus phenotype analysis for platinum sensitivity pathways, inter-omic OR
analysis was performed using GI50 data for all 118 compounds
available within the NCI 60 dataset. In total 1262 pathways were significantly
associated with the drug sensitivity of at least one compound, while 82
compounds gave at least one significant pathway. Figure 4 shows the clustered heat-map of the
binary association matrix in which each element is set to one if a pathway is
significantly associated with sensitivity to a given drug and zero otherwise
(see Table
S2). Significant clustering of the drugs according to mode of action
is visible. For example the dihhryofolate reductase inhibitor methotrexate
co-clustered with related compounds aminopterin, trimetrexate, and
Baker's-soluble-antifolate (triazinate) (Figure 4, blue asterisks); the sensitivities
to all four compounds were associated with 91 common pathways. While one might
expect structural analogues such as these to produce a similar pattern of
sensitivity and hence similar pathway associations, structurally unrelated
compounds that share a common molecular target also co-clustered in certain
instances. One interesting observation was the similarity in pathway
association, reflected by common membership of a cluster, of several structural
analogues, the anthracycline-based compounds, (doxorubicin, zorubicin,
danorubicin hydrochloride and deoxydoxorubicin) with the podophyllotoxin-based
etoposide and teniposide (Figure
4, green asterisks). The four anthracyclines in the cluster share a
large proportion of associated pathways: 63 of the 401 pathways associated with
any of the anthracyclines are commonly associated to the effect of all four
compounds, and of these, 60 are also associated to the chemosensitivity to
either teniposide or etoposide. Etoposide and its derivatives directly inhibit
topoisomerase II activity, followed by induction of DNA strand breaks and
selective cytotoxicity in tumour cells [25] whereas anthracyclines
intercalate DNA, indirectly inhibiting the progression of topoisomerase II and
blocking replication [26]. Thus, the inter-omic pathway analysis is apparently
able to associate chemosensitivity phenotypes on the basis of a common
pathophysiological link independent of whether the key molecular targets are
affected directly, or indirectly by an upstream process.
While such mechanistic relationships were readily observable, the most prominent
division between the compounds, visible as the two largest clusters in Figure 4, appeared to be
separating on the overall frequency of pathways associated with
chemosensitivity, with the top cluster in the diagram possessing on average 2.95
times the number of positive associations of the lower cluster. While each of
the five platinum compounds in the dataset were most similar in pathway
associations to another platinum compound, they were separated across the two
largest clusters with cisplatin and carboplatin forming one group and
tetraplatin, iroplatin and diaminocyclohexyl-Pt II another. This separation
agreed with the low numbers of common chemosensitivity pathways between members
of these two groups in earlier analyses (Figure 2). Thus, the clustering structure did
not describe associations common across the platinum compounds, illustrating the
difficulty of using clustering approaches alone to identify pathways that may
determine class-specific chemosensitivity and the advantages of the consensus
phenotype approach.
To assess the specificity of the identified consensus platinum-sensitivity
pathways we compared these to the most frequently associated pathways in the
global inter-omic OR analysis (Table S2). Of the 54 (top 50 including ties)
most frequently associated pathways (Table S3), just seven intersect with pathways
identified by consensus phenotype integration, mostly related to
immunoregulatory processes (“T-cell receptor” – Netpath;
“B-Cell receptor” – Netpath; “Rho GTPase cycle”
– Reactome; “lCK and FYN tyrosine kinases in initiation of TCR
activation” – BioCarta; “AMB2 integrin signalling”
– PID; “Immunoregulatory interactions between a Lymphoid and a
non-Lymphoid cell” – Reactome; and “TCR signalling in naive
CD8 T cells” – PID). Hence the remaining 23/30 consensus
platinum-sensitivity pathways, dominated by metabolic processes, are not
associated with sensitivity to a wide range of chemotherapeutic agents and are
more likely to be specific to platinum sensitivity.
Our results show that an inter-omic, consensus phenotype approach to integration of
molecular profiles can reveal a cellular metabolic phenotype robustly associated
with platinum chemosensitivity across the NCI-60 cell line panel. Many of the
specific aspects of this phenotype are consistent with the perturbations described
across many studies of tumour cell metabolism, and several of these have been
associated with the development, or likely acquisition, of drug resistance
phenotypes. The classic hallmark of tumour cell metabolism is the Warburg effect: an
increase in glucose uptake and glycolysis to lactate even in normal oxygen
conditions. In addition to the Warburg effect tumour cells are frequently reported
as exhibiting higher rates of glutaminolysis, fatty acid and lipid metabolism, and
nucleotide synthesis [27]. Our observations from the NCI-60 molecular profiles
suggest a positive correlation between all of these phenotypes and platinum
chemosensitivity.
Figure 5 summarises some of the
key correlations observed between gene transcription, metabolite levels and platinum
sensitivity from the consensus pathways indicated by our analysis. The relatively
higher levels of citrate and phosphoenolpyruvate (PEP), observed in more sensitive
cell lines (Figure 5A), are
consistent with low TCA cycle activity (via product inhibition) and increased
diversion of glycolytic intermediates into anabolic pathways such as the pentose
phosphate which feeds nucleotide synthesis [28]. Under these conditions tumour
cells increase the uptake of glutamine and its conversion to oxaloacetate via
glutamate and 2-oxoglutarate (2-OG) in order to replace TCA cycle intermediates and
NADPH [29].
Both glutamate and 2-OG levels were also higher in more sensitive cell lines. Thus
more ‘Warburg–like’ cells appear more sensitive to platinum
treatment than less metabolically transformed lines. The selection of TCA cycle and
pyruvate metabolism as a sensitivity pathway in our analysis is likely to reflect
these associations.
The dependency of tumour cells on glycolysis for synthetic intermediates could be
exploited in platinum chemotherapy; for example the clinically-relevant glycolysis
inhibitor 2-deoxy-glucose (2-DG) has been shown to enhance cisplatin cytotoxicity in
head and neck cancer cells [30]. Interestingly, this synergy appeared to be mediated in
part via oxidative stress, a process that would lead to DNA lesions (e.g.
8-oxo-2′-deoxyguanosine) requiring base excision repair (BER) which was one of
the key consensus sensitivity pathways selected by our analysis (Figure 3). While it is clear that
nucleotide excision repair (NER) capacity is linked to cisplatin resistance [31], [32], [33]; it is becoming
evident that BER is also important in the effect of cisplatin derived drugs [34]. Cross-linking
of DNA via platinum derived drugs can increase the production of free radicals by
disrupting the cellular redox balance [35]. We suggest that the
association of the BER pathway with four platinum drugs observed in the present
study is related to increased ROS production and not adduct formation (repaired by
NER). Intracellular levels of ROS seem vital to the cytotoxic effect of the platinum
derived drugs, further evidenced by the fact that oxaliplatin (a later generation of
Pt drug) is highly cytotoxic but forms less platinum-DNA adducts compared to equal
amounts of cisplatin [35].
A particularly high degree of coordination between gene transcript and metabolite
levels was observed in nucleotide metabolism, revealing a robust association between
increased nucleotide synthesis, both de novo and via recovery of
catabolic intermediates, and tumour cell Pt sensitivity (Figure 5B). For example, in the de
novo pathway, we observed a positive correlation between levels of dUTP
(a precursor to dTMP), expression of dUTP pyrophosphatase (DUT
r = 0.38), expression of thymidylate synthase (TYSY
r = 0.27) and platinum sensitivity. dUTP has to be hydrolysed
to dUMP by DUT to prevent the incorporation of uracils into DNA and suppression of
DUT has been shown to sensitize cells to other chemotherapeutics such as pyrimidine
anti-metabolites [36].
Increased expression of nucleotide salvage pathway enzymes (e.g. uracil
phosphoribosyl transferase or UPP (r = 0.20),
hypoxanthine-guanine phosphoribosyl transferase or HPRT,
r = 0.27) in sensitive cell lines was accompanied by decreases
in several intermediates of purine and pyrimidine catabolism (namely guanine,
guanosine, hypoxanthine, inosine, uracil, uridine and urea) and increase in CMP, the
nucleotide product of HPRT. Kowalski et al. [37] have shown clear links
between inactivation of salvage pathway enzymes such as HGPRT or loss of feedback
inhibition to AMP and GMP de novo synthesis and cisplatin
resistance in yeast. Interestingly in the same study the addition of low
concentrations of extracellular purines also abolished cisplatin cytotoxicity; thus
the metabolome may have a causal influence on platinum sensitivity and not just
represent epiphenomena that is a passive consequence of aberrant cell division.
Our pathway analysis also predicts that lipid metabolism has a direct impact on
chemosensitivity. We observed lower cholesterol, glycerol, and hexadecanoic acid
(palmitate) in more sensitive cell lines, together with negative correlations
between expression of apolipoprotein E (APOE; mean
R = −0.21), LDL receptor (LDLR; mean
R = −0.27) and platinum sensitivity (Figure 5C). All these observations are consistent
with a hypothesis that increased uptake of lipoproteins and constituent
triglycerides, fatty acids and cholesterols can confer resistance to platinum, a
phenomenon previously shown in drug resistant leukemic cell lines [38]. A related
pathway highlighted as associated with sensitivity was phosphatidylcholine
biosynthesis. We observed a positive correlation between choline kinase (CK,
r = −0.28, correlation to −log(GI50)) expression
and resistance to platinum. Recent work by Shah et al. [39] in breast cancer cells have shown
that CK regulates pro-survival MAPk and PI3K/Akt signaling via phosphatidic acid,
and that overexpression leads to drug resistance.
While previous pathway analysis was conducted on gene expression profiles alone from
the NCI60 dataset [40], the use of correlation analysis and the combination of
metabolite and gene transcription measures in our study provides an unprecedented
level of detail into the contribution of metabolic pathways to drug sensitivity.
Using gene set enrichment analysis (GSEA), Reidel et al. [40] suggested that, in addition to
a number of cell signaling and survival networks, methionine metabolism may
contribute to chemotherapeutic resistance to multiple agents, while fatty acid and
β-alanine metabolism were specifically associated with platinum-resistance. In
the context of fatty acid metabolism we show here that lipid uptake and processing
may in fact be the driving factor in this association. It is also interesting to
note that although we did not observe over-representation of β-alanine and
methionine metabolic pathways, both β-alanine and S-adenosylmethionine levels
were significantly positively correlated to platinum sensitivity, adding functional
evidence in support of these earlier findings.
At present, our study is one of very few that presents a strategy for simultaneous
interpretation of gene expression data, metabolic profiles and physiological
endpoints using biological pathway analysis, and has several advantages over other
approaches. Multivariate analysis using pattern recognition algorithms such as PCA,
[41], PLS
[42] and
Kohonen Networks (Self-Organising maps) [43], have been shown to be useful in
revealing novel associations between “-omics” datasets, but fail to take
into account prior biological knowledge relevant to the phenomenon at hand - a
feature which is clearly present in pathway-based techniques. Gene and metabolite
coregulation at the pathway level has been previously studied using OR analysis
[44], [45]. Transcripts
significantly correlated to metabolite levels were examined for over-representation
of Gene Ontology terms [46] or pathways (defined by MapMan BINS). Bradley and
Gibons' work reveals a degree of coordination present between transcriptional
and metabolic measurements at a pathway level, a necessary prerequisite for our
approach to be successful. Importantly none of these examples use a function
physiological endpoint (cytotoxicity) as driver in pathway selection, leading to a
consensus phenotype description of the phenomenon of interest. We show here that
such an approach is critical in reducing false positive selection of pathways.
All OR techniques share the limitation that they rely on a database containing
pre-defined pathways, and therefore cannot identify novel pathways or functional
modules. In this work we have tried to overcome this limitation somewhat through
deconstructing the pathways which were significantly associated and then
functionally interpreting the elements of the pathways which showed significant
associations (Figure 5).
However, even this requires that the elements of the process are sufficiently
grouped in existing pathways to allow for those pathways to be significantly
associated.
Ultimately, a systems biology approach, such as the inter-omic pathway analysis
presented in our study, could assist the development of anti-resistance
chemotherapeutic strategies, and better individualization of treatment, i.e.
personalized medicine. Using gene expression models (GEMs) based on cytotoxicity in
the NCI-60 panel, Williams et al. [47] were able to stratify tumour response and/or patient
survival in seven independent cohorts of patients with breast, bladder and ovarian
cancer. Crucially, the in vitro derived GEMs outperformed those
derived directly from in vivo data. Recently it has also been shown
that pre-treatment metabolic profiles can be used to predict the metabolic fate or
effect of drugs in rodents [48], healthy humans [49], [50] and breast cancer patients
[51]. Given that
the metabolic phenotype of cancer is already the basis of imaging techniques such as
FDG-PET that are currently used to detect early responses to therapy, there is
potentially great value in combing such pharmaco-metabonomic studies with other
characterization of the patient or tumour genome and it is our belief that the
integration of molecular profile data yields more than the sum of its parts. It
remains to be seen if the combination of “-omics” data provides a
competitive advantage over targeted biomarker studies for prognosis and prediction
of drug response in oncology. It remains to be seen if the combination of
“-omics” data provides a competitive advantage over targeted biomarker
studies for prognosis and prediction of drug response in oncology. Several major
challenges to such approaches and translation from in vitro studies, in particular
tumour heterogeneity, require further study. However, irrespective of biomarker
development, the knowledge that chemotherapeutic sensitivity is in part determined
by the metabolic phenotype suggests that metabolic enzymes may be potential targets
in oncology for both drug naive and chemoresistant patients.
The NCI60 data was downloaded from http://dtp.nci.nih.gov/mtargets/download.html on 27th August
2008. For this work three datasets were used: metabolite levels, gene expression
levels and drug sensitivities. The metabolite data consists of measurements of
352 metabolites, 154 identified, across 58 cell lines, performed by Metabolon
Inc. [52].
The transcriptomics data was obtained using the U133 Affymetrix chip by
Genelogic [17]. 44928 probesets were measured, equating to 17150
genes mapping to distinct UniProt identifiers, measured across the same 58 cell
lines. The drug resistance data was selected from the 118 ‘mechanism of
action’ drugs data [15], [16]. Each compound was profiled in between 2 and 1176
independent experiments in a 48-hour sulforhodamine B assay. The values used are
the −log(GI50), where GI50 is the dosage of the drug
which inhibits the growth of the cells by 50%. . GI50 values
were averaged across the replicates for each cell line, thus increasing the
robustness of the primary phenotypic endpoint.
Pathways were derived from the ConsensusPathDB [20] which assimilates
pathways from a range of public databases (see Results). Gene IDs are mapped to UniProt [53] protein IDs. For
metabolites, where available KEGG [21] compound IDs were used,
else, ChEBI [54] IDs were used.
Pearson correlations were calculated between all transcript/metabolite levels and
−log(GI50) values for each drug. Transcripts/metabolites
significant below a false discovery rate threshold of 60% were retained
in each test set for OR analysis. Each UniProt identifier in the ConsensusPathDB
pathways can be mapped to zero or more gene identifiers on the U133 chip. The
background estimate (m in equation 1) for OR analysis was adjusted to reflect
the following: 1) Where ConsensusPathDB proteins could not be mapped to any gene
identifiers, these were ignored; and 2) where ConsensusPathDB proteins mapped to
multiple probesets measuring genes in the transcript data, the number of
probesets was used. In addition, several metabolites, often referred to as
“currency metabolites”, which appear in many pathways and do not
provide specificity were removed before analysis. The currency metabolites
removed were phosphate, diphosphate and NADP+. Thus, given the
transcriptomic and metabolomic data, an “effective size”,
Ni, could be defined for each pathway,
I, in terms of genes and metabolites, The effective pathway
size may be larger or smaller than the actual number of proteins/metabolites in
the pathway. Pathway significance was calculated using the hypergeometric
distribution,(1)where K is the number of genes or
metabolites associated with the drug and ki is the number of genes or
metabolites from the pathway. P<0.05 was used as the criterion defining
significance of pathway enrichment.
We used the pathway p-values pi from the individual
analyses to combine the data. If there were no transcripts or no metabolites
measured for pathway i, we set
pi = 1 for that data type.
Since the transcript/metabolite data were generated from separate experiments.
We thus assumed independence of the pathway associations from the different data
sets. We thus computed the joint probability pJi of
association of pathway i with the drug sensitivity phenotype as
pJi = pGi
pMi where pGi and
pMi denote the probability of association
from the individual gene and metabolite data separately.
Null model 1 was generated by creating random gene and metabolite lists of
matching size to those observed for each of the drugs. Standard OR analysis was
then performed and then numbers of overlapping pathways were recorded. 100 sets
of random lists were generated and the mean number of pathways common to
different numbers of drugs were recorded in table 1.
Null model II assumes that the pathways are selected at random, and so taking the
numbers of pathways selected for each drug, the exact probability of a pathway
being selected for n drugs was calculated. To do this we
calculated the probability of a pathway being selected at random from the full
list of pathways, given the number of pathways selected. By calculating this for
each of the drugs we have.
Additionally, we examined the added information given by the joint analysis. For
each drug the lists of p-values from the metabolite and transcript analyses were
randomly permuted 100 times before combination (randomizing the pathway
association between the two sets). The number of times in which more pathways
were significant (p<0.05) for the permuted lists than in the real data was
recorded.
Cumulative false discovery rates for all models were calculated by dividing the
‘expected’ number of pathways as given by the null model, by the
actual (cumulative) number of pathways found in the real data in at least
n drugs.
|
10.1371/journal.pgen.1004324 | Genetic Interactions Involving Five or More Genes Contribute to a Complex Trait in Yeast | Recent research suggests that genetic interactions involving more than two loci may influence a number of complex traits. How these ‘higher-order’ interactions arise at the genetic and molecular levels remains an open question. To provide insights into this problem, we dissected a colony morphology phenotype that segregates in a yeast cross and results from synthetic higher-order interactions. Using backcrossing and selective sequencing of progeny, we found five loci that collectively produce the trait. We fine-mapped these loci to 22 genes in total and identified a single gene at each locus that caused loss of the phenotype when deleted. Complementation tests or allele replacements provided support for functional variation in these genes, and revealed that pre-existing genetic variants and a spontaneous mutation interact to cause the trait. The causal genes have diverse functions in endocytosis (END3), oxidative stress response (TRR1), RAS-cAMP signalling (IRA2), and transcriptional regulation of multicellular growth (FLO8 and MSS11), and for the most part have not previously been shown to exhibit functional relationships. Further efforts uncovered two additional loci that together can complement the non-causal allele of END3, suggesting that multiple genotypes in the cross can specify the same phenotype. Our work sheds light on the complex genetic and molecular architecture of higher-order interactions, and raises questions about the broader contribution of such interactions to heritable trait variation.
| Although it is well known that interactions among genetic variants contribute to many complex traits, the forms of these interactions have not been fully characterized. Most work on this problem to date has focused on relatively simple cases involving two or three loci. However, higher-order interactions involving larger numbers of loci can also occur, and may have significant effects on the relationship between genotype and phenotype. In this paper, we dissect a colony morphology trait that segregates in a cross of two yeast strains and is caused by genetic interactions among five or more loci. Our work demonstrates that higher-order interactions can have major phenotypic effects, and provides novel insights into the genetic and molecular basis of these interactions.
| Understanding the genetic basis of complex traits is critical for advancing medicine, evolutionary biology, and agriculture [1], [2]. A challenge to progress in this area is that genetic variants can interact, resulting in unexpected phenotypic consequences [3]–[7]. Most of our knowledge about these genetic interactions in natural systems comes from studies focused on two-locus interactions where at least one of the loci exhibits a measurable effect on its own (e.g., [8]). However, evidence suggests that genetic interactions involving three or more loci also occur [9], [10], and that loci participating in such interactions may not individually have detectable effects [11]. Determining how these higher-order interactions arise and influence phenotypic variation could help solve the ‘missing heritability’ problem faced by geneticists studying humans and model species [12].
In this paper, we describe the genetic basis of a complex trait that is influenced by higher-order interactions. We identified this phenotype, a dramatic change in the morphology of Saccharomyces cerevisiae colonies, in a cross of haploid derivatives of the lab strain BY4716 and the clinical isolate 322134S (hereafter ‘BY’ and ‘3S’, respectively). The colony morphology trait in the BY×3S cross is similar to phenotypes described in other yeast isolates and crosses (e.g., [13]–[19]). Thus, by comprehensively determining the genetic basis of colony morphology variation among BY×3S offspring, we not only generate novel insights into how higher-order interactions contribute to phenotypic variation, but also provide new information regarding the genetic basis of a frequently studied model complex trait.
Although both BY and 3S, as well as most of their haploid offspring, form smooth colonies (Figure 1A–C), ∼2% of their progeny exhibited rough colonies when we examined 250 segregants (Figure 1D). Previous work has shown that such heritable variation in colony morphology in S. cerevisiae can arise due to naturally occurring polymorphisms or spontaneous mutations at chromosomal loci [13], [14], [18], [19], aneuploidies [17], and prions [15]. Unlike chromosomal loci, which should show stable inheritance across generations, aneuploidies and prions can be gained or lost, resulting in phenotypic switching. Multiple lines of evidence suggest that chromosomal loci are the primary cause of rough morphology in the BY×3S cross. Neither BY nor 3S exhibits rough morphology, indicating that the phenotype likely requires a combination of alleles from both of these strains. Consistent with this statement, we found that the frequency of rough morphology increased to 12.5% and 21.2% among recombinant haploid progeny obtained by backcrossing a rough segregant to BY and 3S, respectively (Tables S1 and S2; Methods). The higher frequency of rough segregants in backcrosses is expected if alleles from both parents contribute to the trait, as fewer causative alleles should segregate in the backcrosses than in the original cross. Further supporting the argument that our observations of rough morphology were due to chromosomal loci instead of transient factors, we found no evidence for chromosome-scale aneuploidies or phenotypic switching in the backcrossed segregant (Figure S1; Methods).
To identify loci that contribute to rough morphology, we generated thousands of random spores from the aforementioned backcrosses and used low-coverage whole genome sequencing to selectively genotype individuals that showed the phenotype (Methods). We obtained 92 and 88 rough segregants from the BY and 3S backcrosses, respectively. Using these data, we detected five genomic loci that were strongly enriched among these individuals but not among control segregants (Figure S2): three on Chromosomes IV, V, and XV inherited from 3S (Figure 2A), and two on Chromosomes XIII and XIV inherited from BY (Figure 2B). All of these loci, except the one on Chromosome XIV, were fixed among individuals with rough morphology.
We attempted to determine causal genes underlying each of the five loci. Our initial resolution of the loci was between 4 and 14 genes (Figure 2C–G; Table S3; Methods). To decrease the number of candidate genes, we performed targeted genotyping on 19 additional backcross segregants, as well as 8 multi-locus introgression strains that had been subjected to 6 rounds of backcrossing with selection for the rough phenotype (Figure S3; Methods). This additional stage of genetic mapping refined the loci to between 2 and 9 genes per locus, and 22 genes in total (Figure 2C–G; Table S4, S5, S6). We deleted each of the 20 remaining non-essential candidate genes from one of the multi-locus introgression strains (Methods). Across these deletions, a single gene at each locus showed an effect on the phenotype: TRR1 (Chromosome IV), FLO8 (Chromosome V), MSS11 (Chromosome XIII), END3 (Chromosome XIV), and IRA2 (Chromosome XV) (Figure 2C–G). Because the two remaining candidate genes—AVO1 and TOP2—were essential, we examined them using an alternative strategy that suggested they do not contribute to the observed colony morphology variation (Text S1).
We used complementation tests to determine whether the five identified genes possess functional variation (Methods). Each haploid deletion strain was mated to three rough and three smooth haploid backcross progeny (Methods). These matings were designed to produce diploids that were homozygous for the required alleles at four of the causal loci and hemizygous for the fifth causal locus. For END3, FLO8, MSS11, and TRR1, the experiments provided support that the parental alleles differ in their effects. All matings of deletion strains to smooth backcross progeny produced smooth hemizygotes. Further, either two (in the cases of FLO8 and MSS11) or three (in the cases of TRR1 and END3) of the matings of deletion strains to rough backcross progeny produced rough hemizygotes (Figure 3A). However, for IRA2, the two possible hemizygotes showed no phenotypic difference, with both exhibiting smooth morphology (Figure 3A). IRA2 has been reported to show haploinsufficiency in growth rate experiments [20], and this haploinsufficiency may also explain some of our reciprocal hemizygosity results for this gene.
To provide stronger support for IRA2's role in the trait, we performed allele replacements of IRA2 in a smooth backcross segregant that carried the non-causal allele of IRA2, as well as the causal alleles of END3, FLO8, MSS11, and TRR1 (Methods). While transformations with the IRA23S allele had no phenotypic effect, we found that transformations with the IRA2 allele from the rough segregant that had been backcrossed resulted in a change from smooth to rough morphology (Figure 3B). Sequencing of IRA2 from 3S and the rough segregant revealed a single difference between the two alleles: a frameshift mutation that truncates the protein by 117 amino acids (hereafter referred to as IRA23S-Δ2933; Text S2). IRA2 is known to be hypermutable and spontaneous mutations in this gene have been shown to influence a variety of multicellular growth phenotypes [19], [21]. However, our results demonstrate that the effects of spontaneous mutations in IRA2 can depend on an individual's genotype at a number of additional genes. We also checked for IRA23S-Δ2933 in the four other rough individuals that we found in our original BY×3S mapping population. Three of these rough segregants possessed the frameshift mutation, suggesting that IRA23S-Δ2933 probably arose during the outgrowth of the BY/3S diploid prior to its sporulation.
Previous work by other groups identified functional polymorphisms in END3 and FLO8 that also segregate in our cross [22], [23]. BY has a premature stop mutation in FLO8 that prevents it from undergoing many forms of multicellular growth [22]. As for END3, a missense polymorphism in this gene contributes to variability in high temperature growth in a cross of the clinical isolate YJM789 and S288c, the progenitor of BY [23]. Of relevance to our study, this variant in END3 has effects that are strongly dependent on genetic background [24]. With respect to TRR1, the Saccharomyces Genome Resequencing Project [25] and our own sequencing data indicate that the BY and 3S alleles of this gene differ by a single nucleotide, which is a synonymous SNP in the 52nd codon of the gene: BY has an ATC codon and 3S has an ATT codon. Although both of these codons are recognized by the same isoleucine tRNA, the ATT codon is preferred by a nearly two-to-one ratio throughout the yeast genome, suggesting that the SNP might have an effect on translational efficiency. Only lab-derived S. cerevisiae strains carry the ATC allele that confers smooth morphology, while all other sequenced S. cerevisiae and S. paradoxus strains harbor the ATT allele that is likely involved in rough morphology. Work to determine the functional variant(s) in MSS11, which possesses a number of coding and noncoding polymorphisms that could have effects, is ongoing (Table S7).
The causal genes encode proteins with diverse cellular functions: End3 plays a role in clathrin-mediated endocytosis [26], [27], Flo8 and Mss11 are transcription factors that regulate cell-cell adhesion and multicellular phenotypes in S. cerevisiae [28], [29], Ira2 is a negative regulator of the RAS-cAMP pathway [30], and Trr1 is an enzyme involved in oxidative stress response [31], [32]. Flo8 and Mss11 physically interact [33], and IRA2 and MSS11 show a genetic interaction when both are knocked out [34]. To our knowledge, none of the other pairs of identified genes have been reported to interact at the biochemical, genetic, physical, or regulatory levels. To assess whether Flo8 and Mss11 might directly regulate the expression of the other genes, we examined existing data from calling card analyses, a technique that identifies genomic sites bound by transcription factors [16]. These results indicated that Flo8 and Mss11 are unlikely to bind the promoters of END3, IRA2, and TRR1, although admittedly the study involved a different strain than our cross parents.
After identifying causal genes at the five loci, we analyzed the effects of these genes in more detail by genotyping them in a panel of phenotyped segregants from dissected backcross tetrads (Methods). Every individual with rough morphology possessed the 3S allele of FLO8 and TRR1, the BY allele of MSS11, and IRA23S-Δ2933 (Figures 4A–B and S4A–B; Tables S8 and S9). Although most individuals with rough morphology carried END3BY, a small fraction of individuals with END33S also showed the trait (Figures 4B and S4C; Table S9), indicating that alleles at additional loci complement END33S.
We more deeply investigated the genetic basis of rough morphology among individuals with END33S. First, we used a gene knockout strategy to check whether END33S is necessary for these individuals to exhibit rough morphology (Methods). end33SΔ strains were smooth (Figure S5), suggesting that the alternate genetic architecture for rough morphology requires END33S. Second, we tried to identify loci that complement END33S. Four rough END33S progeny were present in our sequenced mapping population from the 3S backcross. Among these segregants, we detected 11 previously unidentified genomic regions where individuals shared the same genotype (Figure S6; Table S10; Methods). We were able to reduce this set to four candidate loci on Chromosomes VII, XI, XII, and XV by genotyping additional backcross progeny (Table S11; Methods). To determine which of the four loci have causal roles in rough morphology, we mated a relevant backcross segregant to 3S and analyzed a panel of 51 second-generation backcross progeny (Table S12; Methods). The BY alleles at the Chromosome VII and XV loci were fixed among the 39 individuals with rough morphology, while the other two loci showed no evidence of playing a role in the trait (Figure 4C; Table S12). Given only individuals that carried BY alleles at both the Chromosome VII and XV loci exhibited rough morphology, it is likely that these loci genetically interact to complement END33S.
Our findings indicate that the segregant used for backcrossing carried more than one set of interacting alleles that can specify rough morphology (Figure 4D). Identifying the causal genes and genetic variants underlying the Chromosome VII and XV loci can thus shed light on how these different genotypes produce the same trait. However, our ongoing efforts to clone the causal factors at these loci are limited by the crude resolution of the present data (each locus is presently resolved to >60 kilobases; Table S10). We note that initial gene deletion experiments focused on 18 candidates (Table S13), including LAS17 and YAP1802, whose cognate proteins functionally interact with End3 [35], [36], have been unsuccessful. Moving forward, we plan to determine the genes that underlie the Chromosome VII and XV loci, and characterize their relationship with END3.
In summary, we have demonstrated that sets of five or more genetic variants can synthetically interact to produce major phenotypic effects. Alleles involved in these higher-order interactions may either be polymorphisms that segregate in natural populations or spontaneous mutations. Our results also illustrate that rather than functioning in a single biochemical pathway, protein complex, or regulatory circuit, the genes involved in higher-order interactions can play roles in a number of cellular processes. This finding implies that characterizing higher-order interactions using data from screens and annotations focused solely on reference genomes may be a challenge, and highlights how genetic variation can serve as a tool for detecting previously unidentified functional relationships among genes. Further, we have shown that multiple sets of alleles can interact to produce the same phenotypic effect. Additional work is necessary to determine how this latter finding is mediated at the molecular and systems levels. Overall, our study suggests that characterizing the larger-scale contribution of higher-order interactions to phenotypic variation is a necessary step in improving our basic understanding of the genotype-phenotype map.
All phenotyping experiments were performed on agar plates containing yeast extract and peptone (YP) with 2% ethanol as the carbon source (YPE). Prior to phenotyping, strains were grown up in liquid YP with 2% dextrose (YPD). Stationary-phase cultures were manually pinned onto YPE and allowed to grow for five days at 30°C, and were then imaged using a standard digital camera.
Sequencing data from the rough segregant used in backcross experiments was examined at the chromosome-scale for evidence of aneuploidy. Average per base coverage of each chromosome was computed in R and compared to the genome-wide average. This segregant was also plated at low density on a large number of YPE plates. We screened tens of thousands of colonies for instances of phenotypic switching and observed no cases where an individual converted from rough to smooth morphology.
Strains used in this paper contained the Synthetic Genetic Array marker system [37], which allowed us to easily generate large numbers of recombinant MATa progeny. All segregants discussed in the paper were MATa can1Δ::STE2pr-SpHIS5 his3Δ and all backcrosses involved mating these individuals to either a BY or a 3S strain that was MATα his3Δ. In these crosses, strains with opposite mating types were mixed together on a YPD plate and incubated for four hours at 30°C. Zygotes were then obtained by microdissection. To generate segregants, diploids were sporulated at room temperature using the protocol described by Guthrie and Fink [38]. Once sporulation had completed, spore cultures were digested with β-glucuronidase and then plated onto yeast nitrogen base (YNB) plates containing canavanine, as described previously [39]. Spores were plated at a density of roughly 100 to 200 colonies per plate.
Whole genome sequencing libraries were prepared using the Illumina Nextera kit, with each of the backcross segregants barcoded with a unique sequence tag. The libraries were mixed together in equimolar fractions and sequenced on an Illumina HiSeq machine by the Beijing Genomics Institute using 100 base pair (bp) ×100 bp reads. Sequencing reads were then mapped to the S. cerevisiae reference genome using the Burrows-Wheeler Aligner (BWA) [40]. We used data from 36,756 high confidence SNPs that had been identified based on comparison of Illumina sequence data for 3S to the BY genome. Similar to Andolfatto et al. [41], we employed Hidden Markov Models (HMMs) to determine the haplotypes of the segregants based on the sequencing data. We computed the fraction of reads at each SNP that came from BY and used the vector of these fractions in HMMs that were implemented chromosome-by-chromosome in the HMM() package of the R statistical programming environment. Any segregants producing data that showed evidence of contamination, diploidy, or aneuploidy were excluded from genetic mapping and downstream analyses. Four and eight such individuals were left out of the BY and 3S mapping populations, respectively.
Genotypes inferred from the HMM were used in genetic mapping analyses. At each position in the genome, we determined the fraction of individuals that carried the allele from the parent not used in the backcross. We scanned the genome for alleles from the non-backcross parent that were detected in a large fraction of segregants. We report loci where these alleles were at 95% frequency or higher. To determine intervals in which causal genes were located, we identified the smallest region that was bounded by recombination breakpoints among individuals from a backcross that shared the same allele at a peak.
Backcross diploids were sporulated and digested in β-glucuronidase to permit tetrad dissection. Standard microdissection techniques were used to isolate tetrads and separate individual spores.
Haploid multi-locus introgression strains were constructed using six rounds of recurrent backcrossing with phenotypic selection, starting from the same segregant used in our backcross mapping experiment. Eight of these strains were generated, with four made by recurrently backcrossing to 3S and four made by recurrently backcrossing to BY. We also used a subset of individuals from the tetrad dissections that showed rough morphology. To conduct the fine-mapping, we typed these individuals at a number of markers in each interval using PCR and restriction digestion, or Sanger sequencing.
All genes within causal loci were deleted using the CORE cassette, in the same manner described by Storici et al. [42]. Homology tails matching the 60 bases immediately up- and downstream of each gene were attached to the CORE cassette through PCR and introduced into cells using the Lithium Acetate method [43]. Selection for G418 resistance was used to screen for integration of the CORE cassette; correct integration was then checked using PCR. All deletions were performed in a haploid multi-locus introgression strain. To perform complementation tests, deletion strains were mated to multiple dissected segregants that carried either the causal or non-causal allele of the deleted gene, as well as the causal alleles at the four other involved genes. The same phenotyping methods described above were employed for these strains. To generate allele replacement strains for IRA2, a smooth segregant with the non-causal allele of IRA2 and the causal alleles at the other four loci was transformed using a modified form of adaptamer mediated allele replacement [44]. Transformations were conducted with two partially overlapping PCR products—a full-length amplicon of IRA2 that was tailed at the 3′ end with the 5′ portion of the kanMX cassette and a copy of the kanMX cassette that was tailed on the 3′ end with part of the intergenic region downstream of IRA2. Knock-ins were identified using selection on G418 and verified by Sanger sequencing.
Sequenced strains from the backcross to 3S were partitioned based on their genotype at END3. We then screened these individuals for sites where they all carried BY alleles. A group of additional rough segregants with END33S that had been obtained during tetrad dissections were genotyped by PCR amplification and restriction digestion of markers across each of the new loci. One of these additional backcross segregants was mated to 3S, and a panel of rough progeny from this second-generation backcross were typed at the remaining candidate loci.
|
10.1371/journal.pbio.1001283 | Presynaptically Localized Cyclic GMP-Dependent Protein Kinase 1 Is a Key Determinant of Spinal Synaptic Potentiation and Pain Hypersensitivity | Synaptic long-term potentiation (LTP) at spinal neurons directly communicating pain-specific inputs from the periphery to the brain has been proposed to serve as a trigger for pain hypersensitivity in pathological states. Previous studies have functionally implicated the NMDA receptor-NO pathway and the downstream second messenger, cGMP, in these processes. Because cGMP can broadly influence diverse ion-channels, kinases, and phosphodiesterases, pre- as well as post-synaptically, the precise identity of cGMP targets mediating spinal LTP, their mechanisms of action, and their locus in the spinal circuitry are still unclear. Here, we found that Protein Kinase G1 (PKG-I) localized presynaptically in nociceptor terminals plays an essential role in the expression of spinal LTP. Using the Cre-lox P system, we generated nociceptor-specific knockout mice lacking PKG-I specifically in presynaptic terminals of nociceptors in the spinal cord, but not in post-synaptic neurons or elsewhere (SNS-PKG-I−/− mice). Patch clamp recordings showed that activity-induced LTP at identified synapses between nociceptors and spinal neurons projecting to the periaqueductal grey (PAG) was completely abolished in SNS-PKG-I−/− mice, although basal synaptic transmission was not affected. Analyses of synaptic failure rates and paired-pulse ratios indicated a role for presynaptic PKG-I in regulating the probability of neurotransmitter release. Inositol 1,4,5-triphosphate receptor 1 and myosin light chain kinase were recruited as key phosphorylation targets of presynaptic PKG-I in nociceptive neurons. Finally, behavioural analyses in vivo showed marked defects in SNS-PKG-I−/− mice in several models of activity-induced nociceptive hypersensitivity, and pharmacological studies identified a clear contribution of PKG-I expressed in spinal terminals of nociceptors. Our results thus indicate that presynaptic mechanisms involving an increase in release probability from nociceptors are operational in the expression of synaptic LTP on spinal-PAG projection neurons and that PKG-I localized in presynaptic nociceptor terminals plays an essential role in this process to regulate pain sensitivity.
| Pain is an important physiological function that protects our body from harm. Pain-sensing neurons, called nociceptors, transduce harmful stimuli into electrical signals and transmit this information to the brain via the spinal cord. When nociceptors are persistently activated, such as after injury, the connections they make with neurons in the spinal cord are altered in a process called synaptic long-term potentiation (LTP). In this study, we examine the molecular and cellular mechanisms of LTP at synapses from nociceptors onto spinal neurons. We use multiple experimental approaches in mice, from genetic to behavioural, to show that this form of LTP involves presynaptic events that unfold in nociceptors when they are repetitively activated. In particular, an enzyme activated by the second messenger cGMP, referred to as Protein Kinase G-I, phosphorylates presynaptic proteins and increases the release of neurotransmitters from nociceptor endings in the spinal cord. When we genetically silence Protein Kinase G-I or block its activation in nociceptors, inflammatory pain is markedly reduced at the behavioural level. These results clarify basic mechanisms of pathological pain and pave the way for new therapeutic approaches.
| Plasticity in peripheral nociceptors and their synapses with spinal neurons has been proposed as a cellular basis for the development and maintenance of pain hypersensitivity following peripheral inflammation or nerve injury [1]–[3]. Activation of nociceptive nerve afferents at frequencies relevant to pathological pain states can trigger long-term potentiation (LTP) at spinal synapses between nociceptor terminals and spinal neurons projecting nociceptive information to the brain [4],[5]. Importantly, this form of synaptic plasticity can be evoked by asynchronous activation of nociceptors in vivo [5], occurs in humans [6], and is functionally associated with a sensation of exaggerated pain [5],[6]. Although there is evidence for a requirement of post-synaptic calcium-dependent mechanisms in the induction of LTP at this synapse [5], the precise mechanisms underlying the expression of spinal LTP are not entirely clear [7].
Synaptic LTP evoked by natural, asynchronous low-rate discharges in C-nociceptors on spino-PAG neurons was recently shown to constitute a very fitting correlate of spinal amplification phenomena underlying inflammatory pain [5],[7]. This form of synaptic change has been reported to involve activation of NMDA receptors, NO release, and synthesis of cGMP [5],[7]. However, which of the diverse targets of cGMP come into play at this synapse and how they mechanistically bring about long-lasting changes in the transfer of nociceptive information between the nociceptors and spinal neurons projecting to the brain is not understood so far. Furthermore, very little is known about exactly how neural circuits involved in pain processing are modulated by cGMP and which cellular and molecular processes underlie these changes.
Studies on several different biological systems have shown that cGMP regulates multiple cellular targets, including diverse cGMP-gated ion channels, such as cyclic nucleotide-gated (CNG) and hyperpolarization-activated cyclic nucleotide-gated (HCN) channels, the cGMP-dependent protein kinases, PKG-I/cGK-I and PKG-II/cGK-II, as well as diverse phosphodiesterases (PDEs) [8],[9]. Nearly all of these molecular targets of cGMP are expressed in nociceptive pathways and may potentially contribute to the key role of cGMP in synaptic potentiation in the spinal cord. Amongst these targets, PKG-I has emerged as a key mediator of cGMP functions in smooth muscle and platelet function [8]. The α-isoform of PKG-I has been reported to be expressed very highly in the primary sensory neurons in the dorsal root ganglia (DRG) over developmental [10] and adult stages [11], and several regions in the brain and the spinal cord also express PKG-I [12],[13]. Pharmacological and genetic studies in global, constitutive mutant mice have linked PKG-I to the development of the nociceptive circuitry as well as to spinal mechanisms of hyperalgesia [14].
Based upon this background, this study was designed with two goals in mind. First, it addressed the potential involvement of presynaptic mechanisms in the expression of synaptic potentiation on spinal projection neurons, which has not been explored or described previously. Second, it aimed to explore a potential role for PKG-I localized presynaptically in the spinal terminals of nociceptors in spinal potentiation and to clarify cellular and molecular mechanisms underlying these processes. We reasoned that the use of a conditional, region-specific gene deletion strategy to specifically manipulate presynaptic mechanisms might constitute an unambiguous approach towards addressing the above questions. Our results show that spinal synaptic potentiation triggered by nociceptor activation is associated with a long-lasting change in the probability of neurotransmitter release from spinal terminals of nociceptors. Using viable, developmentally normal transgenic mice lacking the PKG-I specifically in nociceptors with preserved expression in spinal neurons, brain, and all other organs, we demonstrate here that PKG-I localised in nociceptor terminals constitutes a key mediator of synaptic LTP and that its activation is functionally associated with pain hypersensitivity in vivo.
Mice lacking PKG-I specifically in a primary nociceptor-specific manner (SNS-PKG-I−/−) were generated via Cre/loxP-mediated recombination by mating mice carrying the floxed prkg1 allele (PKG-Ifl/fl) [15] with a mouse line expressing Cre recombinase under control of the Nav1.8 promoter (SNS-Cre) [16]. We have previously demonstrated that SNS-Cre mice enable gene recombination commencing at birth selectively in nociceptive (Nav1.8-expressing) sensory neurons, without affecting gene expression in the spinal cord, brain, or any other organs in the body [16],[17]. An anti-PKG-I antibody [18] yielded specific staining in wild-type dorsal root ganglia (DRG), but not in those from global PKG-I−/− mice [19], thereby revealing Cre/loxP-mediated deletion of PKG-I in DRG of SNS-PKG-I−/− mice (Figure 1A). Quantitative size-frequency analysis revealed that a majority of DRG neurons expressing PKG-I in wild-type mice are small-diameter neurons, which show a near complete loss of PKG-I expression in SNS-PKG-I−/− mice (Figure 1B; p<0.001). In contrast, a few large-diameter neurons showed low levels of anti-PKG-I immunoreactivity in DRGs of PKG-Ifl/fl, which was entirely retained in SNS-PKG-I−/− mice (Figure 1A and B). Confocal analysis of dual immunofluorescence experiments revealed PKG-I immunoreactivity in nearly all Isolectin-B4 (IB4)-labelled non-peptidergic nociceptors and substance P-expressing peptidergic nociceptors in PKG-Ifl/fl mice, both of which are selectively lost in SNS-PKG-I−/− mice (typical examples in Figure 1C and quantitative summary in Figure 1D). In contrast, large-diameter neurofilament-200-immunoreactive neurons entirely retained PKG-I expression in the SNS-PKG-I−/− mice (Figure 1C,D). Taken together, these results show that PKG-I is normally expressed in nearly all nociceptors and is selectively lost from these neurons, but not from tactile-responsive and proprioceptive DRG neurons, in SNS-PKG-I−/− mice.
We found that PKG-I expression is entirely unaltered in the brains of SNS-PKG-I−/− mice (an example of expression in cerebellar purkinje neurons is shown in Figure 1E, right panel), whereas global PKG-I−/− mice demonstrated a complete loss of anti-PKG-I immunoreactivity (Figure 1E, right panel). In the spinal cord of SNS-PKG-I−/− mice, anti-PKG-I immunoreactivity was decreased selectively in the superficial dorsal laminae, which represent termination zones of the nociceptive afferents, as would be expected from SNS-Cre-mediated gene deletion in nociceptors (Figure 1E, left panel). In contrast, neurons in the spinal cord entirely maintained immunoreactivity for PKG-I and appeared particularly conspicuous (arrowheads in Figure 1E, left panel) due to the loss of PKG-I labelling in afferent terminals in SNS-PKG-I−/− mice. Furthermore, anti-Cre immunohistochemistry as well as Western blot analysis with anti-PKG-I antibody confirmed that SNS-PKG-I−/− mice show a DRG-specific loss of PKG-I while retaining expression in the spinal cord and brain (Figure S1).
In contrast to global PKG-I−/− mice, which typically demonstrate lethality in the first few weeks of life, SNS-PKG-I−/− mice were normal, fertile, and showed a normal life expectancy. In contrast to defects reported in global PKG-I−/− mice [10], SNS-PKG-I−/− mice showed normal early targeting of TrkA-expressing primary afferents arising from the DRG (arrowheads in Figure 1F, upper panels) in the developing spinal dorsal horn at embryonic day 13 (E13). Unlike global PKG-I−/− mice [10], SNS-PKG-I−/− mice did not show defects in T-branching of DiI-labelled primary afferents in the spinal cord over embryonic developmental stages (arrows in Figure 1F, lower panels). Similarly, central and peripheral patterning of peptidergic or non-peptidergic nociceptors was normal in adult SNS-PKG-I−/−mice, as revealed by immunostaining for substance P and binding to IB4, respectively, in the spinal dorsal horns and skin (Figure S2A). Because peptidergic mechanisms have been suggested to play an important role in spinal LTP [5], we ascertained that SNS-PKG-I−/− mice are not different from control mice with respect to the abundance of substance P in the spinal circuitry. Control and knockout mice exhibited the same prevalence of substance P-immunoreactive cells within DRG (33%±5% versus 29%±3%, respectively), which were not significantly different from each other (p>0.05, Student's t test). Moreover, the level of substance P immunoreactivity was similar in the superficial spinal dorsal horn across genotypes (mean intensities in PKG-Ifl/fl mice and SNS-PKG-I−/− mice were 50±3 and 51±3 arbitrary units, respectively). Importantly, confocal microscopy revealed normal density of synapses between substance P-containing nociceptive afferents and PSD-95-positive puncta (representing postsynaptic aspects of glutamatergic synapses) in the spinal dorsal horns of SNS-PKG-I−/− mice as compared to PKG-Ifl/fl mice (examples and quantification in Figure S2B). Finally, we addressed the internalization of NK1 receptors on spinal lamina I neurons following peripheral nociceptive stimulation in vivo, which has been demonstrated to be a clear indicator of nociceptive activity-induced synaptic release of substance P [20]. As shown in Figure S2C, application of a 52°C heat stimulus for 20 s to the plantar paw surface led to internalization of NK1 receptors in lamina I neurons of L3/L4 segments to a similar extent in SNS-PKG-I−/− and PKG-Ifl/fl mice (quantification in Figure S2D). Unlike global PKG-I−/− mice [14], SNS-PKG-I−/− mice showed a normal lamination of the spinal cord over early postnatal stages (Figure S2E). Thus, the multiple developmental defects in the patterning of sensory afferents and spinal lamination that have been reported in global PKG-I−/− mice were not observed in SNS-PKG-I−/− mice.
To address activity-dependent plasticity at spinal synapses, we recorded C-fiber-evoked synaptic LTP on spinal lamina I neurons projecting to the periaqueductal grey (PAG), which were retrogradely labelled upon stereotactic injection of DiI in the PAG (the experimental scheme is shown in Figure 2A and an example of a labelled cell is shown in Figure S3A) [5]. In spinal-PAG projection neurons of wild-type mice, a conditioning low frequency stimulation of 2 Hz for 2 min produced synaptic LTP of monosynaptic C-fiber evoked EPSCs by more than 200% at 30 min (Figure 2B). LTP at these synapses was preserved in the presence of strychnine and gabazine, which block glycinergic and GABAergic inhibitory neurotransmission, respectively (Figure 2B). Similar results were obtained upon using another standard blocker of GABAergic neurotransmission, bicuculline, in combination with strychnine (Figure S3B). Hence, LTP does not manifest due to primary afferent depolarization mediated by presynaptic GABA receptors or disinhibition of the postsynaptic neuron. To test whether LTP requires a postsynaptic function of PKG-I, we dialyzed standard PKG-I inhibitors, such as the non-permeant peptide inhibitor RKRARKE [21],[22] or KT5823 [23], into spinal neurons via the patch pipette. These manipulations did not affect the magnitude or duration of C-fiber-evoked LTP at spino-PAG synapses (Figure 2C and Figure S3C), suggesting that PKG-I localized postsynaptically in spino-PAG projection neurons does not play a role in LTP at this synapse.
To assess the role of PKG-I localized presynaptically in spinal nociceptor terminals, we then analysed PKG-Ifl/fl mice and SNS-PKG-I−/− mice. In spinal-PAG projection neurons of PKG-Ifl/fl mice, a conditioning low frequency stimulation of 2 Hz for 2 min produced LTP with a magnitude of more than 200% at 30 min and more than 300% by 60 min (typical examples of time course and EPSC traces are given in Figure 2D,E). Prior to the conditioning stimulus, baseline values of C-fiber-evoked EPSCs stayed constant over the period of recording in both genotypes (Figure 2D,E). In striking contrast to PKG-Ifl/fl mice, the conditioning stimulus did not evoke LTP in spinal-PAG projection neurons in SNS-PKG-I−/− mice (Figure 2D,E; see Figure 2F and 2G for quantitative summary at 30 min post-conditioning stimulus; p<0.001; at least 13 neurons from each genotype were tested).
For a clear interpretation of these data, it is imperative to address how basal nociceptive transmission at these synapses is affected in SNS-PKG-I−/− mice. Analysis of EPSC magnitude evoked by the first and the last pulse of the conditioning train revealed short-term depression of evoked EPSCs during the conditioning train, which was equivalent in PKG-Ifl/fl mice and SNS-PKG-I−/− mice (Figure 3A; p = 0.95), showing that the conditioning stimulus was equally effective in mice from both groups. Furthermore, basal C-fiber-evoked EPSCs were comparable between PKG-Ifl/fl and SNS-PKG-I−/− mice. We also established detailed input-output curves representing the relationship between the intensity of dorsal root stimulation and evoked EPSCs in the absence of a conditioning stimulus and found no differences between PKG-Ifl/fl and SNS-PKG-I−/− mice (Figure 3B; p = 0.74). Furthermore, the intensity of dorsal root stimulation required to elicit an action potential in post-synaptic spinal-PAG projection neurons was identical in PKG-Ifl/fl and SNS-PKG-I−/− mice (an example is shown in Figure 3C). The intact nature of responsiveness in spinal-PAG projection neurons was demonstrated by similarities in their activation profiles upon current injection in SNS-PKG-I−/− mice and their PKG-Ifl/fl littermates (an example is shown in Figure 3D). In all of the above experiments, quantitative analyses revealed that resting membrane potential, action potential width, delay after stimulation artefact, action potential threshold, delay for generation of first action potential (latency to first AP), as well as amplitude of after hyperpolarisation (AHP) were similar in SNS-PKG-I−/− mice and PKG-Ifl/fl mice (Figure 3E; p>0.05 in all cases, Student's t test). Finally, as an additional indicator of the number of fibers activated during electrical stimulation, we recorded fiber volleys in input/output measurements. Recording C-fiber volleys in the L4 and L5 dorsal roots derived from PKG-Ifl/fl and SNS-PKG-I−/− mice revealed typical responses, which increased in amplitude with increasing stimulus intensity (representative traces are shown in Figure 3F). The amplitudes of C-fiber volley responses were not significantly different between PKG-Ifl/fl and SNS-PKG-I−/− mice (stimulus-response curves are shown in Figure 3G; p>0.05; n = 16 per genotype), demonstrating directly that the number of fibers activated upon electrical stimulation was comparable between genotypes and could therefore not explain the failure to evoke synaptic potentiation in SNS-PKG-I−/− mice. These comprehensive analyses show that a presynaptic loss of PKG-I was specifically linked to a failure of activity-dependent potentiation of transmission at synapses between nociceptors and spinal-PAG projection neurons, but not to modulation of basal synaptic transmission.
Following our observation that a specific presynaptic alteration in PKG-I expression perturbed spinal LTP, we then addressed potential contributions of presynaptic mechanisms at synapses between nociceptors and spinal projection neurons [7],[24]. By recording miniature EPSCs in spinal-PAG projection neurons of control mice, we observed that quantal content varies largely, which is expected since these spinal projection neurons receive multisynaptic inputs from primary afferents as well as spinal interneurons. Furthermore, inputs arising from spinal interneurons are not expected to change in SNS-PKG-I−/− mice since the molecular perturbation is specific to nociceptor terminals. Thus, the net contribution of C-fibers to the population of mEPSCs is difficult to assess because it is unclear which fraction of mEPSCs can be attributed to C-fibers, which then makes detection of potentially small presynaptic changes highly unlikely when performing mini-analysis. To study synaptic events which could be clearly assigned to activation of presynaptic primary afferent fibers alone, we employed a protocol of minimal stimulation, setting the dorsal root stimulation parameters such that a synaptic failure rate of approximately 60% was achieved in recording solution containing 1 mM Ca2+ and 5 mM Mg2+. The failure rate remained constant over a period of at least 30 min upon repetitive test stimulation in the absence of a conditioning stimulus (55.9%±3.9% pre- and 57.1%±8.1% at 30 min). However, upon application of the conditioning stimulus, minimal stimulation using the same parameters evoked a decrease in the frequency of synaptic failures within a few minutes in slices derived from PKG-Ifl/fl mice, indicating a change in the probability of neurotransmitter release; decrease in synaptic failures was accompanied by a corresponding rise in the magnitude of C-fiber-evoked EPSCs (see Figure 4A for typical example and Figure 4C for quantitative summary of C-fiber-evoked EPSCs recorded every 15 s; n = 5; p = 0.04). In contrast, the rate of synaptic failures did not change significantly following conditioning stimulus in slices derived from SNS-PKG-I−/− mice (see Figure 4B and 4C; n = 7; p = 0.68). EPSC values at the end of the recording were not significantly elevated as compared to basal values in SNS-PKG-I−/− mice (25±1.7 pA before and 21.1±1.2 pA at 30 min after the conditioning stimulus; p = 0.14).
In analogy to studies on hippocampal circuits, although the above evidence for a decrease in failure rate is indicative of increased presynaptic release probability, it has also been linked to unsilencing of silent synapses [25]. Therefore, to further consolidate presynaptic mechanisms, we focussed on the analysis of paired-pulse facilitation (PPF), which represents a short-lasting increase in the second evoked EPSP when it follows shortly after the first and is well accepted as an indication of presynaptic mechanisms of long-term potentiation in the hippocampus [26]. In hippocampal CA1 neurons, PPF can increase as well as decrease in conjunction with LTP in a manner inversely proportional to the PPF prior to the conditioning stimulus [26]. Indeed, we obtained similar results in recordings at spinal synapses between C-fibers and spinal-PAG projection neurons. In spinal slices derived from mice of both genotypes, we found evidence for PPF as well as paired-pulse depression (PPD) prior to the LTP-inducing conditioning stimulus (typical traces are shown in Figure 4D). Whereas a majority of neurons derived from PKG-Ifl/fl mice demonstrated a clear change in PPF or PPD following conditioning stimulus, neurons derived from SNS-PKG-I−/− mice did not (see examples in Figure 4D). We then plotted the paired-pulse ratio (PPR) of the entire cohort of recorded neurons at 30 min after conditioning stimulation as a function of the basal PPR recorded prior to the conditioning stimulus (Figure 4E,F). This analysis revealed that neurons in PKG-Ifl/fl mice with larger basal values of PPR prior to the conditioning stimulus (indicated by filled round symbols in Figure 4E) consistently showed a decrease in PPR after the conditioning stimulus (Figure 4D), which is indicative of an increase in the probability of release (PR, Figure 4E). This drop in PPF following conditioning stimulation did not come about or was reduced in neurons from SNS-PKG-I−/− mice (filled round symbols in Figure 4F). A smaller cohort of synapses in PKG-Ifl/fl mice showed an increase in PPF after conditioning stimulation, but this was restricted to neurons with a low magnitude of PPF prior to the conditioning stimulus (i.e., a PPR of about 1.1–1.2, filled square symbols in Figure 4E) and a low expression of LTP (Figure S4A). Again, this change was not observed in the corresponding cohort of neurons in SNS-PKG-I−/− mice (filled square symbols in Figure 4F, Figure S4B). Conversely, in PKG-Ifl/flmice, higher magnitudes of LTP (i.e., between 150% and 350%, indicated by black frame in Figure S4A) were consistently associated with a decrease in PPR, which is indicative of an increase in release probability. Neither LTP nor consistent changes in PPR were observed in SNS-PKG-I−/− mice (Figure S4B). In conclusion, the failure rate analysis and PPR analysis strongly support the inference that the expression of LTP at spino-PAG synapses comes about via presynaptic mechanisms involving an increase in release probability via PKG-I.
In an effort to understand the underlying molecular mechanisms, we then addressed potential substrates for the kinase activity of PKG-I in nociceptors. In particular, we reviewed known substrates of PKG-I in other biological systems and focussed on those for which we hypothesized a role in synaptic transmission. We first set up an assay system for testing involvement of PKG-I substrates in the DRG selectively upon persistent nociceptive stimulation in vivo, using Vasodilator-stimulated phosphoprotein (VASP), a classical target of PKG-I, as an indicator of PKG-I activity [8],[27]. Lysates of L4-L5 DRGs from naïve PKG-Ifl/fl and SNS-PKG-I−/− mice showed comparable levels of VASP expression (Figure S5A, basal). Within minutes after persistent nociceptive stimulation via injection of formalin in the hindpaw, L4-L5 DRGs from PKG-Ifl/fl mice showed a striking phosphorylation of VASP at Serine 239 (typical examples in Figure S5A and summary in Figure S5B,C; see Text S1 for details; p = 0.03 as compared to basal). This was markedly reduced in formalin-injected SNS-PKG-I−/− mice (Figure S5A,B; p = 0.18 as compared to basal SNS-PKG-I−/− mice and 0.03 as compared to formalin-injected PKG-Ifl/fl mice). These results show that persistent activation of nociceptors leads to rapid signalling via PKG-I in DRG neurons in vivo, which is lost in SNS-PKG-I−/− mice, as expected.
Using this assay system, we then addressed another key target of PKG-I, which has been mainly studied so far mechanistically in smooth muscle cells. Dephosphorylation of myosin light chains (MLC) [28] via PKG-I-dependent phosphorylation and activation of myosin light chain phosphatase in smooth muscle cells is a decisive mechanism underlying NO-mediated vasodilation [8]. Following formalin injection in the paw, we observed a strong phosphorylation of MLC in L4-L5 DRGs from PKG-Ifl/fl mice, which was found to be lacking in formalin-treated SNS-PKG-I−/− mice (Figure 5A and 5B). These differences did not arise due to differences in expression levels of MLC between SNS-PKG-I−/− mice and PKG-Ifl/fl mice (Figure 5B; Figure S5C). This finding was unexpected because it suggests a role for PKG-I in increasing MLC phosphorylation in DRG neurons, which is contrary to the classical role ascribed to PKG-I in MLC dephosphorylation. We reasoned that if our findings hold true, synthesis of cGMP ought to be a critical intermediate step in activity-dependent MLC phosphorylation in DRG neurons. Indeed, in mice pre-treated with an inhibitor of the soluble guanylyl cyclase, ODQ, and a pan-inhibitor of membrane-bound guanylyl cyclases, LY83583, via intrathecal application, formalin-induced MLC phosphorylation in L4-L5 DRGs was strongly reduced (see examples in Figure 5C; quantitative summary from three experiments is given below the Western blot). Immunohistochemistry revealed that nociceptor activation-induced increase in MLC phosphorylation occurred in the spinal termination zone of nociceptors (lamina I and II) as well as in spinal neurons (Figure 5D).
Interestingly, synaptic potentiation induced by a conditioning stimulus on spinal-PAG projection neurons was abolished in the presence of ML-7, an inhibitor of MLCK (Figure 5E and Figure 5F; p = 0.004 as compared to vehicle-treated control slices). Furthermore, consistent with our observations in SNS-PKG-I−/− mice, inhibition of MLC phosphorylation did not affect basal transmission at this synapse (Figure 5G; p = 0.761). Thus, the PKG-I target, pMLC, is functionally linked to potentiation of synaptic transmission in nociceptive laminae.
Ikeda et al. [5] have reported that inhibition of IP3R activation blocks conditioning stimulus-induced synaptic potentiation at synapses between nociceptors and spinal-PAG projection neurons. This is particularly interesting because IP3R1 contains a PKG-I-recognition motif at serine 1755 and has been reported to be phosphorylated by PKG-I in vitro, putatively leading to gain of function [29],[30]. We observed that IP3R1 is indeed a target of PKG-I in nociceptors and is functionally associated with modulation of calcium release from intracellular stores. In immunoprecipitation experiments from L4-L5 DRGs, formalin-injected PKG-Ifl/fl mice demonstrated highly enhanced serine 1755 phosphorylation of IP3R1 over the basal state; this effect was markedly reduced in DRGs obtained from formalin-injected SNS-PKG-I−/− mice (Figure 6A,B), although expression levels of IP3R1 were comparable between SNS-PKG-I−/− mice and PKG-Ifl/fl mice (Figure S5D).
PKG-I-mediated phosphorylation of serine 1755 in IP3R1 has been suggested to positively modulate IP3R1 activity in heterologous test systems [30]. We observed that this function of PKG-I indeed plays an important role in modulating calcium release from intracellular stores in nociceptive neurons of the DRG. We performed Fura-2-based calcium imaging on dissociated DRG neurons derived from PKG-Ifl/fl and SNS-PKG-I−/− mice using Fluro488-conjugated Isolectin B4 (IB4-Fluro488) for live identification of small-diameter nociceptive neurons (neurons dually labelled with Fura-2 and IB4-Fluor488 are indicated by arrowheads in Figure 6C). The baseline values of the Fura2 ratios (F340/F380) were not significantly different between control mice (1.049±0.010) and SNS-PKG-I−/− mice (1.040±0.008) (p>0.05; Student's t test). Stimulation of calcium release via activation of Gq/11-phospholipase C-IP3R1 pathway by addition of ligands, such as bradykinin (BK) and the P2Y-receptor ligand, UTP, led to typical increases in the ratio of Fura-2 fluorescence at 340/380 nm in neurons from PKG-Ifl/fl mice, which were markedly reduced in neurons from SNS-PKG-I−/− mice (see Figure 6D for typical examples and Figure 6E for quantitative summary; p<0.001 with respect to BK and UTP). In contrast, Fura2-labelled neurons with large-diameter somata, which were IB4-negative, did not show differences in calcium responses between PKG-Ifl/fl mice and SNS-PKG-I−/− mice (an example is indicated by arrow in Figure 6C and quantitative summary is given in Figure 6E). In contrast, rapid calcium influx caused by KCl-induced depolarisation or capsaicin-evoked influx of calcium via TRPV channels was comparable in DRG neurons derived from SNS-PKG-I−/− mice and PKG-Ifl/fl mice (Figure 6D,E; p>0.05), showing thereby that a loss of PKG-I in nociceptive neurons is specifically linked to defects in IP3R-mediated calcium release from intracellular stores.
Taken together, these biochemical and functional experiments suggest that following persistent nociceptive stimulation, PKG-I mediates potentiation of IP3R1 activity and MLC phosphorylation in sensory neurons, which is functionally linked to synaptic LTP at synapses between C-nociceptors and spinal-PAG projection neurons.
We then went on to address whether these findings bear relevance to pain-related behaviour in vivo and found a functional role for PKG-I and its substrates in behavioural paradigms for spinal sensitization. As a test system, we studied the Phase II of formalin-induced nocifensive behavioural responses, which are manifest at 10–60 min after intraplantar formalin injection, for two reasons: one, this represents a widely used paradigm for studying central changes in pain processing caused by a persistent activation of nociceptors [31], and two, intraplantar formalin induces synaptic LTP on spinal projection neurons with a matching time-course [5]. Formalin-induced phase II responses were significantly reduced upon intrathecal pretreatment with 2-APB or ML-7 to the lumbar spinal cord (Figure 7A; p<0.01 for 2-APB and ML-7 in comparison to vehicle control, respectively), implicating involvement of IP3R function and MLC phosphorylation, respectively. Similarly, SNS-PKG-I−/− mice showed markedly reduced phase II responses than PKG-Ifl/fl mice (Figure 7B; p<0.001 as compared to PKG-Ifl/fl mice).
Basal withdrawal thresholds and response latencies to acute application of paw pressure (e.g., as tested with a dynamic aesthesiometer) (Figure S6A, left panel) or thermal stimuli (e.g., a radiant infrared heat ramp) (Figure S6A, right panel), respectively, to the paw surface were found to be similar across SNS-PKG-I−/− mice and their control littermates (p>0.05). Furthermore, motor performance on a Rotarod was unaffected in SNS-PKG-I−/− mice (Figure S6B; p = 0.20). We have previously shown in details that SNS-Cre mice show no alterations in the processing of acute pain or chronic inflammatory or neuropathic pain [6],[32].
In the context of studying disease-induced pain hypersensitivity, we first focussed on a model of inflammatory pain which is associated with primary hyperalgesia in the inflamed area and ongoing nociceptive inputs from the periphery throughout the time of testing, namely unilateral hindpaw inflammation induced by injection of Complete Freund's Adjuvant (CFA) [32],[33]. CFA injection produced similar levels of edema in SNS-PKG-I−/− and PKG-Ifl/fl mice (Figure S6C) and hypersensitivity to graded von Frey mechanical stimuli (Figure 7C,D) or to plantar heat (Figure 7E) applied to the ipsilateral paw was assessed at 6, 12, 24, 48, and 96 h thereafter. Following inflammation, PKG-Ifl/fl mice demonstrated the characteristic leftward and upward shift in the stimulus-response curve over basal curves reflecting mechanical hypersensitivity (black squares in Figure 7C). In contrast, SNS-PKG-I−/− mice demonstrated a less marked deviation from baseline behaviour upon CFA-induced inflammation (red squares in Figure 7C). Furthermore, the relative drop in response thresholds to von Frey hairs (defined here as minimum force required to elicit 40% response frequency) in the inflamed state over basal (pre-CFA) state occurred to a significantly lesser extent in SNS-PKG-I−/− mice as compared to PKG-Ifl/fl mice (left panel in Figure 7D; p<0.05 at all time points tested). Finally, SNS-PKG-I−/− mice showed a significantly lower magnitude of thermal hyperalgesia than PKG-Ifl/fl mice at 6 h after CFA and did not show hyperalgesia at all from 12 h onwards after CFA injection, whereas PKG-Ifl/fl mice continued to show thermal hyperalgesia all the way up to the latest time point tested, namely 96 h following CFA injection (Figure 7E; p<0.01 between PKG-Ifl/fl and SNS-PKG-I−/− mice at all time points tested). We infer from the above that the development of primary hyperalgesia and mechanical allodynia following somatic inflammation is impaired by a loss of PKG-I in nociceptors.
Although perturbation of spinal LTP may have contributed to the above phenotype in SNS-PKG-I−/− mice, it is conceivable that a peripheral role for PKG-I in nociceptors may at least partially account for changes in primary hyperalgesia. To address functional changes in nociceptor sensitivity in the inflamed tissue, we utilised the skin-nerve preparation [34] to study the electrophysiological properties of identified polymodal C-fibres and Aδ-mechanoceptors (AM) in the saphenous nerve. The excitability of mechanoreceptive C-fibers and AM-fibers showed a small, but significant, increase following paw inflammation in PKG-Ifl/fl mice (see Figure 7F for typical examples), but not in SNS-PKG-I−/− mice (Figure 7F). These data indicate defects in the development of peripheral sensitization in nociceptors of SNS-PKG-I−/− mice, which could contribute to a reduction in primary hyperalgesia; however, they are unlikely to account for the marked defects in mechanical allodynia observed following inflammation in SNS-PKG-I−/− mice.
To explore central contributions, we utilised two models of aberrant pain which are triggered initially by peripheral inputs but do not require ongoing nociceptor activity in the periphery for maintenance. For example, capsaicin injection in the skin activates C-fibers and evokes hyperalgesia in the area of the flare (primary hyperalegsia) as well as outside of the flare (secondary hyperalgesia). In PKG-Ifl/fl mice, we observed that injection of capsaicin in the skin of the lower thigh produced a marked allodynia at the hindpaw plantar surface, which was clearly excluded from the area capsaicin-induced flare (see shift in von Frey response frequency in Figure 8A; black symbols). SNS-PGK-I−/− mice showed markedly reduced secondary hypersensitivity with capsaicin as compared to PKG-Ifl/fl mice (red symbols in Figure 8A). Moreover, a capsaicin-induced drop in mechanical threshold (allodynia) was markedly reduced in SNS-PGK-I−/− mice as compared to PKG-Ifl/fl mice (Figure 8B). It is well accepted that capsaicin-induced secondary mechanical hypersensitivity reflects C-fiber-evoked central amplification processes and can last for several hours, long after nociceptor responses to capsaicin have ceased owing to desensitisation of TRP channels [35]. Nevertheless, to rule out a potential contribution of ongoing peripheral inputs to the above-described phenotypic differences, we performed experiments in which nerve conduction was blocked with lidocaine in the peripheral dermatome in which capsaicin was injected in wild-type mice. As expected, lidocaine-induced nerve blockade prior to capsaicin injection blocked the induction of capsaicin-induced mechanical hypersensitivity (Figure 8C); in contrast, when lidocaine was injected 15 min after capsaicin, mechanical hypersensitivity developed normally (Figure 8C), indicating that beyond the initial trigger, capsaicin-induced mechanical hypersensitivity is independent of ongoing input from peripheral nociceptors. In further experiments, we addressed the peripheral and central contributions of PKG-I. Pharmacological inhibition of PKG-I with KT5823 injected prior to injection of capsaicin in the same dermatome in wild-type mice did not block the development of capsaicin-induced mechanical hypersensitivity (Figure 8D); in contrast, when KT5823 was injected intrathecally prior to peripheral capsaicin injection, the induction of mechanical hypersensitivity was markedly inhibited (Figure 8E), indicating a role for central PKG-I, but not peripherally expressed PKG-I. To further delineate the origin of the central (spinal) locus of PKG-I function, we undertook similar experiments in PKG-Ifl/fl mice and SNS-PGK-I−/− mice. Interestingly, intrathecally administered PKG-I inhibitor blocked the development of capsaicin-induced mechanical hypersensitivity in PKG-Ifl/fl mice and did not lower mechanical sensitivity any further in SNS-PGK-I−/− mice (Figure 8F), demonstrating thereby the presynaptic locus of its action.
In the muscle pain model by Sluka and colleagues [36], two consecutive injections of dilute acidic saline in the flank muscle lead to secondary mechanical hypersensitivity in the ipsilateral and contralateral paws, which lasts for several weeks. The initial peripheral insult (i.e., flank muscle) is spatially distinct from the area of application of nociceptive stimuli (paw surface), ruling out a contribution of peripheral paw sensitization to the behavioural phenotype. Secondly, it has been shown in details previously that the secondary hyperalgesia in the paw lasts for several months after muscle injection, is not associated with any persistent inflammation or injury to the muscle tissue, is independent of peripheral inputs, and is thus central in origin [36]. Upon testing at 24 h after the induction of muscle pain, PKG-Ifl/fl mice demonstrated a pronounced leftward and upward shift in the stimulus-response curve to von Frey hairs applied to the plantar paw surface (black squares in Figure 9A, middle panel), which was still evident 3 wk later in the ipsilateral (Figure 9A, right panel); this changes come about in the paw ipsilateral to the injected flank muscle (upper panels in Figure 9A) as well as in the contralateral paw (lower panels in Figure 9A). In contrast, SNS-PKG-I−/− mice did not show significant deviations (red squares in Figure 9A). Analysis of paw-withdrawal thresholds to graded pressure also consistently revealed that muscle injection-induced drop in paw mechanical thresholds at the paws was significantly lesser in SNS-PKG-I−/− mice than in control littermates at all time points tested (Figure 9B). In conclusion, these analyses support an essential role for presynaptic PKG-I in nociceptor terminals in central mechanisms of secondary mechanical hypersensitivity.
Finally, we asked whether PKG-I expressed in nociceptors constitutes an important target of the NMDA-NOS-cGMP pathway. Consistent with previous studies [37], intrathecally administered NMDA produced a rapid facilitation of the paw withdrawal reflex (Figure 10A). Importantly, in striking contrast to PKG-Ifl/fl mice (black symbols), SNS-PKG-I−/− mice completely failed to develop hyperalgesia following intrathecal NMDA delivery (red symbols, upper panel in Figure 10A). Similar results were obtained upon delivery of an NO donor, NOC-12, to the spinal cord via intrathecal catheters (upper panel in Figure 10B). Furthermore, intrathecal delivery of NMDA and NOC-12 produced a facilitation of the tail flick reflex in PKG-Ifl/fl mice, but not in SNS-PKG-I−/− mice (lower panels in Figure 10A,B), showing thereby that PKG-I is critically required for the pro-nociceptive functions of the NMDA and NO.
Because soluble guanylyl cyclases (sGC) represent a key molecular link between NO and activation of PKG-I, the above results imply that NO activates sGC in spinal presynaptic terminals of nociceptors. While some studies report a lack of sGC expression in DRG neurons [38], others reported expression in a population of small diameter DRG neurons and in primary afferents [39],[40]. Here, we carried out mRNA in situ hybridisation using riboprobes recognising the beta subunit of sGC on mouse DRG sections and observed distinct, specific signals over the soma of several large and small-diameter DRG neurons (arrowheads and arrows in Figure 10C, respectively). Furthermore, the satellite cells surrounding DRG neurons showed dense signals (red arrows in Figure 10C). Sense control probes did not yield any appreciable signals (Figure 10C). These results indicate that sGC mRNA is expressed in sensory neurons of the DRG.
In addition to sGC enzymes, which are directly activated upon NO, the membrane-bound guanylyl cyclases (mGC; Npr family) also contribute to cGMP production in some organs (e.g., in the cardiovascular system) [41]. Stimulated by recent reports on expression of mGCs in DRG neurons [42], we administered a cocktail of natriuretic peptides (ANP, BNP, and CNP) intrathecally and observed marked hyperalgesia within 15 min after delivery, which lasted for about 45–50 min in wild-type mice (unpublished data) and PKG-Ifl/fl mice (Figure 10D). Interestingly, natriuretic peptide-induced hyperalgesia was also entirely abrogated in SNS-PKG-I−/− mice (Figure 10D). These results suggest that the NMDA-NOS-soluble guanylyl cyclase-cGMP pathway as well as the natriuretic peptide-mGC-cGMP pathways converge upon PKG-I expressed in spinal terminals of nociceptors to modulate nociceptive processing in the spinal cord.
Finally, we undertook experiments to test whether PKG-I expression alone or some downstream factor perturbed by an early loss of PKG-I is responsible for the deficits in pain hypersensitivity in SNS-PKG-I−/− mice. We constructed chimeric Adeno-associated virions of the serotypes AAV1 and AAV2 expressing an C-terminally GFP-tagged version of the murine PKG-I cDNA [43]. Injection in unilateral L3 and L4 DRGs in vivo led to a broad expression in the DRG. AAV1/2 chimeric virions expressing GFP alone served as controls. PKG-Ifl/fl mice and SNS-PKG-I−/− mice expressing GFP-tagged PKG-I or GFP alone showed normal basal sensitivity to graded von Frey stimuli (Figure 11B). Upon peripheral injection of capsaicin, PKG-Ifl/fl mice expressing GFP-tagged PKG-I showed a small increase in mechanical hypersensitivity than PKG-Ifl/fl mice expressing GFP, which was only statistically significant at some intensities of mechanical stimuli (Figure 11C). As expected, SNS-PKG-I−/− mice overexpressing GFP in DRG showed markedly reduced mechanical hypersensitivity with capsaicin than PKG-Ifl/fl mice overexpressing GFP. Importantly, overexpression of GFP-tagged PKG-I fully restored mechanical hypersensitivity in SNS-PKG-I−/− mice (Figure 11D). This indicates that expression of PKG-I is both necessary and sufficient for inducing centrally maintained hypersensitivity upon persistent peripheral activation of C-fibers.
In contrast to the intensively studied forms of LTP in the hippocampus, very few studies have addressed cellular and molecular mechanisms of LTP at spinal synapses regulating the flow of nociceptive information from the periphery towards the brain [7],[24]. Here we observed that nociceptive activity-driven LTP at synapses between nociceptive terminals and spinal neurons projecting nociceptive inputs to the PAG requires presynaptic mechanisms for its full expression. Furthermore, our results indicate that this function is mediated by cGMP acting via PKG-I. We base our inferences on three main observations: (1) A specific loss of PKG-I in presynaptic, but not post-synaptic, compartments of this synapse abolished C-fiber-evoked LTP without altering basal neurotransmission at this synapse; (2) LTP was temporally accompanied by a decrease in the rate of synaptic failures in a presynaptic-PKG-I-dependent manner; and (3) LTP was associated with a change in the PPR, which did not take place when PKG-I was deleted presynaptically in nociceptor terminals. Importantly, higher magnitudes of LTP were consistently associated with a decrease in PPF, and thereby with an increase in release probability.
Previous studies have shown that the NMDA receptor-NO-cGMP pathway is important in the induction of spinal LTP [4],[5], and it has been assumed that this pathway comes into play in the post-synaptic compartment. However, all of the above signal transducers are also expressed presynaptically in afferent terminals in the spinal dorsal horn [44]. Thus, pre- and post-synaptic contributions to spinal LTP have not been worked out so far. There is evidence for a requirement for post-synaptic Ca2+ change for the induction of the LTP (i.e., in experiments with BAPTA in recording pipette; [5]). Taken together with our results, this suggests that a calcium-dependent postsynaptic mechanism may be required for the induction of LTP (e.g., via NMDA receptor-dependent generation of NO); in contrast, a presynaptic change involving cGMP- and PKG-I-dependent increase in neurotransmitter release may mediate the expression of LTP at synapses between nociceptors and spinal-PAG projection neurons.
Mechanistically, this may come about via involvement of multiple phosphorylation targets of PKG-I. While some targets have been identified in heterologous systems, very little is known about the nature and functional role of PKG-I targets in vivo. Here, we identified and validated two primary targets in DRG neurons, namely the IP3R1 and MLC, and observed that PKG-I modulates intracellular calcium release as well as MLC phosphorylation differently in DRG neurons as compared to other biological systems, such as the smooth muscle. For example, in some biological systems, PKG-I has been reported to negatively modulate calcium signals via its interaction with IRAG [18]. However, IRAG interacts selectively with the beta-isoform of PKG-I, which is barely expressed in the DRG, but not with PKG-I-alpha, the predominant form found in DRG neurons [10]. Furthermore, our observations that PKG-I potentiates calcium release induced by typical mediators of nociceptive sensitization, such as bradykinin, in identified nociceptive neurons and that repetitive activation of nociceptors in vivo leads to PKG-I-mediated phosphorylation of IP3R1 at serine 1755, a site associated with positive functional modulation, implicate PKG-I as a positive modulator of calcium signalling in nociceptive neurons. In light of electrophysiological analyses reported here, this raises the possibility that calcium released from IP3R1-gated stores may participate in modulating presynaptic function. Although a few studies at hippocampal synapses have proposed an involvement of calcium stores in modulation of presynaptic release [45],[46], underlying cellular mechanisms are not known. Results of this study suggest that activation of presynaptic PKG-I may constitute the molecular link between synaptic activity and the elevation of resting levels of calcium in presynaptic terminals, thereby potentiating synaptic transmission via an increase in release probability. Furthermore, we observed that MLC was phosphorylated in a nociceptive activity-dependent manner and that PKG-I is required for MLC phosphorylation in the DRG. Although our electrophysiological data implicate phosphorylated MLC in LTP at spinal synapses, not much can be inferred about downstream mechanisms at this stage. At central synapses, MLC phosphorylation was initially implicated in vesicle transport and in regulation of vesicular pools [47]; however, these inferences could not be corroborated in detailed subsequent analyses [48]. Taken together, more detailed analyses overcoming current technical hindrances in studying mobilization of vesicular pools in the complex circuitry of the DRG and spinal dorsal horn will be required to understand mechanisms underlying PKG-I-mediated modulation of release probability at this synapse.
The possibility of functionally linking synaptic changes described here to changes in nociceptive behaviour simultaneously represents a good opportunity and a major challenge. As the first parameter to test this relationship, we focused on the phase II responses in the intraplantar formalin test, which has been attributed to spinal nociceptive sensitization triggered by an initial barrage of C-fiber inputs [31]. Indeed, we observed that presynaptic loss of PKG-I as well as functional perturbation of MLCK and IP3R, its substrates involved in LTP, inhibited phase II behavioural responses. However, a contribution of central mechanisms could not be inferred from the formalin data due to several reasons: Although MLCK/IP3R inhibitors were administered spinally, the genetically induced loss of PKG-I in SNS mice occurred throughout the nociceptor, including peripheral terminals. Moreover, there is still some ongoing activation of peripheral nociceptors in the phase II of the formalin response [31]. Indeed, our electrophysiological analyses in the CFA inflammatory pain model suggested that peripheral PKG-I may contribute to primary hyperalgesia.
Therefore, we focused on pain models in which nociceptive hypersensitivity is triggered by peripheral nociceptors, but maintained via central mechanisms that outlast and are independent of peripheral inputs. One of these is the chronic muscle pain model in which injections of dilute acidic saline in the gastronemius muscle evokes a long-lasting secondary mechanical hyperalgesia in the ipsilateral and contralateral paws, which lasts for several weeks, is unrelated to muscle damage and is not maintained by continued primary afferent input from the site of injury, as shown by experiments involving dorsal rhizotomy and lidocaine injections in the muscle [36]. Similarly, we addressed capsaicin-induced secondary mechanical hyperalgesia outside of the primary flare, which albeit triggered by C-fiber inputs, is maintained via mechanisms of central origin as indicated by previous studies [35] and our analyses. In both models, we found marked defects in central hypersensitivity in SNS-PKG-I−/− mice. Our results indicated that capsaicin-evoked mechanical hypersensitivity is neither dependent on peripheral PKG-I function nor does it require ongoing peripheral nociceptor sensitisation. Moreover, they revealed that PKG-I expressed in central terminals of nociceptors plays a decisive role in the induction of mechanical hypersensitivity after persistent C-fiber stimulation via capsaicin. Finally, reinstating PKG-I expression in the DRG in adult SNS-PKG mice fully restored capsaicin-evoked mechanical hypersensitivity, indicating that PKG-I directly, and not some factor affected by a genetic loss of PKG-I, is a functional determinant of C-fiber-evoked mechanical hypersensitivity.
In summary, this study shows that PKG-I expressed in nociceptors terminals is the principal target of cGMP at nociceptive synapses. Furthermore, it suggests that PKG-I-mediated presynaptic facilitation and LTP in spinal projection neurons is functionally involved in activity-dependent centrally mediated nociceptive hypersensitivity.
Additional details on methods are provided in Text S1.
Homozygous mice carrying the flox allele of the mouse prkg1 gene, which encodes the cGMP dependent kinase 1 (PKG-Ifl/fl) [15], have been described previously in detail. PKG-Ifl/fl mice were crossed with SNS-Cre mice [16] to obtain PKG-Ifl/fl;SNS-Cre+ mice (referred to as SNS-PKG-I−/− mice in this article) and PKG-Ifl/fl mice (control littermates). Mice were crossed into the C57BL6 background for more than 8 generations. Mice lacking PKG-I globally (PKG-I−/− mice) have been described before. Only littermates were used in all experiments to control for background effects.
Mice (14–18 d old) were anesthetized with a mixture of Dormitor, Dormicum, and Fentanyl, and stereotactic injections of DiI into the PAG were carried out (see Text S1 for details). After 2 to 3 d, transverse 350–450 µm thick spinal cord slices with dorsal roots attached were obtained and whole cell patch clamp recordings of identified DiI-positive neurons were performed as described in Text S1. Test pulses of 0.1 ms with intensity of 3 mA were given at 30 s intervals to the dorsal root via a suction electrode. For studying the site of expression of synaptic potentiation, we used minimal stimulation in conditions of low release probability (in mM: NaCl 127; KCl, 1.8; KH2PO4, 1.2; Ca2+ 1.0; Mg2+, 5; NaHCO3, 26; glucose, 15; oxygenated with 95% O2, 5% CO2; pH 7.4). Dorsal root was stimulated at intensity of threshold to evoke EPSCs on DiI-labelled spino-PAG projection neuron. Under these conditions, the failure rate was 60.9%±6.3% (n = 10). To induce synaptic potentiation, low frequency stimulation (conditioning stimulus, 2 Hz for 2 min) was applied to dorsal root as a conditioning stimulus with the same intensity as the test stimulus [5]. The recording mode during conditioning stimulation was the same as that before and after conditioning stimulation. Neurons are voltage clamped at −70 mV. Because a suction electrode was utilized to stimulate the whole root, a suprathreshold stimulus was required to fully recruit C-fibers in the root [5]. Synaptic strength was quantified by assessing the peak amplitudes of EPSCs. The mean amplitude of 4–5 EPSCs evoked by test stimuli prior to conditioning stimulation served as a control. Significant changes from control were assessed by measuring the peak amplitudes of five consecutive EPSCs every 5 min after conditioning stimulation. Additional details are given in Text S1. In some experiments, blockers of inhibitory neurotransmission, such as Gabazine (10 µM) and Strychnine (1 µM), were added to the bath.
In a subset of animals, paired-pulse stimuli with an inter-stimulus interval of 110 ms (0.1 ms pulse duration, 3 mA intensity, every 30 s) were used (see Text S1 for details). Paired-pulse ratio (facilitation or depression) of C-fiber-evoked EPSC was calculated as the amplitude of the second C-eEPSC divided by that of the first C-eEPSC in a pair. In a subset of experiments, PKG-I inhibitors such as KT5823 (10 µM) or RKRARKE (250 µM) were infused post-synaptically via the patch pipette.
The following antibodies were used for Western blots and biochemical analyses: anti-IP3R1, anti-pS1755 IP3R1 (kind gift from Prof. Richard Wojcikiewicz), anti-VASP (Alexis Biochemical), anti-MLC, anti-alpha tubulin (Sigma), anti-pSer239 VASP, anti-pThr18/pSer19 MLC (Cell Signaling technology), anti-PKG-I antibody [18], secondary HRP labelled anti-rabbit (Sigma Aldrich), or anti-mouse (GK Healthcare UK Ltd.).
The following antibodies were used for immunohistochemistry: phospho-ERK1/2 antibody (Cell signalling), anti-Fos antibody (Chemicon), anti-Isolectin B4 antibody (vector laboratories), anti-Calcitonin gene related peptide antibody (Immunostar), anti-Neurofilament 200 antibody (Chemicon), anti-Substance P antibody (Chemicon), anti-PKG-I antibody [18], anti-PSD-95 antibody (a gift from M. Watanabe) and anti-TrkA antibody (a kind gift from Prof. L. F. Reichardt), and anti-cre antibody (Novagen).
The soluble guanylyl cyclase inhibitor, ODQ (25 mg/kg body weight; sigma Aldrich), and the membrane guanylyl cyclase inhibitor, LY83583 (12.5 mg/kg body weight; sigma Aldrich), were dissolved in 50% DMSO and injected in a volume of 250 µl intraperitoneally. The following drugs were administered intrathecally in vivo: an inhibitor of MLCK (ML-7; 15 nmol Alexis Biochemical, dissolved in 5% DMSO), an inhibitor of IP3R (2-APB; 2 nmol; Calbiochem), NMDA (100 fmol; Sigma Aldrich), the NO donor, NOC-12 (17 nmol; Sigma Aldrich), atrial natriuretic peptide, brain natriuretic peptide and c-type natriuretic peptide (rANP1-28, mBNP45, and hCNP1-22; 330 pmol of each natriuretic peptide; American peptide company, Inc., USA), and the PKG-I inhibitor KT5283 (200 pmoles). See Text S1 for details on intrathecal delivery. Mice were allowed to recover for 2 d after surgery, and only animals showing complete lack of motor abnormalities were used for further experiments. 5 µl of drugs were applied followed by flushing of the catheter with 10 µl of 0.9% saline. The following drugs were administered peripherally in the vicinity of the paw in experiments pertaining to capsaicin-induced mechanical hypersensitivity: KT5823 (200 pmoles) and lidocaine (10 µl of 2%).
A total of 17 PKG-Ifl/fl and 15 SNS-PKG-I−/− mice were used in the electrophysiological recordings of nerve activity. An ex vivo skin-nerve preparation was used to study the properties of mechanosensitive C- and A-δ afferent fibres which innervate the skin in the inflamed area 24 h following CFA inoculation (20 µl) as described previously (see Text S1 for details).
All animal use procedures were in accordance with ethical guidelines imposed by the local governing body (Regierungspräsidium Karlsruhe, Germany). All behavioural measurements were done in awake, unrestrained, age-matched mice of both sexes that were more than 3 mo old by individuals who were blinded to the genotype of the mice being analyzed (see Text S1 for details).
The open reading frame of mouse PKG-I fused C-terminally with GFP [43] or EGFP alone was cloned in an AAV expression construct, and chimeric AAV1/2 virions were generated using standard protocols. Virions were diluted 1∶2 with 20% mannitol and injected unilaterally into L3 and L4 DRGs (1 µl per DRG, or approx. 107 transfection units per DRG) in deeply anesthetized mice as described in detail previously [49]. Mice were tested in behavioural tests 2 wk after injection. At the end of the experiment, mice were perfused as described above and expression of GFP was confirmed via fluorescence analysis.
All data are presented as mean ± standard error of the mean (S.E.M.). For comparisons of multiple groups, analysis of variance (ANOVA) for random measures was performed followed by post hoc Fisher's test to determine statistically significant differences. When comparing two groups that were studied in parallel, Student's t test was employed. Unless otherwise specified, the p values shown in the figures and text are derived from ANOVA and post hoc Fisher's test. p<0.05 was considered significant.
|
10.1371/journal.ppat.1006248 | Synergy between the classical and alternative pathways of complement is essential for conferring effective protection against the pandemic influenza A(H1N1) 2009 virus infection | The pandemic influenza A(H1N1) 2009 virus caused significant morbidity and mortality worldwide thus necessitating the need to understand the host factors that influence its control. Previously, the complement system has been shown to provide protection during the seasonal influenza virus infection, however, the role of individual complement pathways is not yet clear. Here, we have dissected the role of intact complement as well as of its individual activation pathways during the pandemic influenza virus infection using mouse strains deficient in various complement components. We show that the virus infection in C3-/- mice results in increased viral load and 100% mortality, which can be reversed by adoptive transfer of naïve wild-type (WT) splenocytes, purified splenic B cells, or passive transfer of immune sera from WT, but not C3-/- mice. Blocking of C3a and/or C5a receptor signaling in WT mice using receptor antagonists and use of C3aR-/- and C5aR-/- mice showed significant mortality after blocking/ablation of C3aR, with little or no effect after blocking/ablation of C5aR. Intriguingly, deficiency of C4 and FB in mice resulted in only partial mortality (24%-32%) suggesting a necessary cross-talk between the classical/lectin and alternative pathways for providing effective protection. In vitro virus neutralization experiments performed to probe the cross-talk between the various pathways indicated that activation of the classical and alternative pathways in concert, owing to coating of viral surface by antibodies, is needed for its efficient neutralization. Examination of the virus-specific complement-binding antibodies in virus positive subjects showed that their levels vary among individuals. Together these results indicate that cooperation between the classical and alternative pathways not only result in efficient direct neutralization of the pandemic influenza virus, but also lead to the optimum generation of C3a, which when sensed by the immune cells along with the antigen culminates in generation of effective protective immune responses.
| The pandemic influenza A(H1N1) 2009 virus is now circulating seasonally and causing a significant disease burden worldwide. Hence, it is important to delineate the immune components required for protection against its infection. Here we demonstrate that presence of intact complement is essential for clearing the pandemic influenza virus infection, wherein complement synthesized by B cells plays a major role. Further, we show that activation of the classical as well as alternative pathways is a requisite for efficient neutralization of the virus as well as the optimum generation of C3a, which is necessary for boosting the protective immune responses. Our results thus reveal that deficiencies of components of the classical and alternative pathways enhance the susceptibility to and severity of the pandemic influenza virus infection.
| Influenza viruses, the members of the family Orthomyxoviridae, are globally important acute respiratory pathogens known to cause significant morbidity and mortality [1–3]. These negative sense RNA viruses possess a segmented genome encompassing a constellation of genes, which facilitate genetic reassortment upon infection with more than one influenza A virus strains [4–6]. This may result in generation of a novel virus with the ability to cause pandemics in humans [3]. One such example of a reassortant is the emergence of the novel swine-originated influenza A(H1N1) 2009 virus [A(H1N1)pdm09] that caused the first pandemic of the 21st century infecting a large population and about 200,000 confirmed human deaths worldwide within the first 12 months [7–9]. The virus possesses a unique genetic combination of swine, avian and human influenza virus genes encompassing assemblage of genes such as hemagglutinin (HA), nucleoprotein (NP) and non-structural (NS) segments from the classical swine H1N1 lineage, polymerase basic 1 (PB1), polymerase basic 2 (PB2) and polymerase A (PA) from the North American H3N2 triple reassortant swine lineage, and the neuraminidase (NA) and matrix (M) segments from the Eurasian swine virus lineage [10]. The A(H1N1)pdm09 virus was first detected in India in May 2009 [11] and since then there have been outbreaks in many parts of the country. Genetic characterization of the pandemic virus isolates from India indicated that they clustered with the globally circulating prototype strain [11] with indigenous transmission in the country [12]. Within a short time the A(H1N1)pdm09 virus became one of the predominant subtypes in the country co-circulating with the seasonal influenza virus strains [13]. Genetic and antigenic characterization of this virus suggested it to be distinct from the seasonal human influenza A(H1N1) strain [14]. Thus, identification of factors influencing the pathogenesis and control of the pandemic influenza virus is essential.
The complement system is one of the critical barriers against viruses [15–17]. It has the ability to recognize and eliminate viruses directly as well as indirectly [18]. The direct mechanisms include neutralization owing to aggregation, opsonisation, lysis and promotion of phagocytosis via complement receptors [15,19], whereas the indirect mechanisms involve modulation of the adaptive immunity owing to interaction between the complement components and constituents of the adaptive immune responses [20,21]. In particular, complement has been shown to play an essential role in augmentation of virus-specific B cell and T cell responses [21–25].
Studies in the animal models have demonstrated that the complement system plays an important role in providing protection against the seasonal influenza virus infection [16,26,27], which is expected to be due to its ability to control the virus by neutralization, and enhancement of the protective immune responses. In support of this argument, the classical pathway (CP) has been shown to neutralize the seasonal influenza viruses directly with the help of natural [28] and induced antibodies [29]. And the deficiency of C3 has been shown to be associated with reduced priming of T cells in the lung-draining lymph nodes (dLN) and recruitment of effector T cells into the lung [16]. Recently, this priming defect was attributed to decreased dendritic cell (DC)-mediated viral antigen transport to the dLN, and signalling through C3aR/C5aR was found to be crucial for DC migration from the lung to dLN [27]. Which complement pathway(s) contribute towards the augmentation of these in vivo protective immune responses however is still not clearly understood.
The novel 2009 pandemic influenza H1N1 virus has been shown to activate complement [30], but whether complement is capable of neutralizing this virus and what role the individual complement pathways play in its neutralization, and in controlling the infection has not yet been studied. In the present study, we therefore have asked what role intact complement (using C3-/- mice) and its individual complement pathways (using C4-/- and factor B-/- mice) play in controlling the pandemic influenza virus infection, and whether the pandemic influenza H1N1 virus is susceptible to neutralization by all the complement pathways. Our data show that deficiency of intact complement results in heightened vulnerability to the pandemic influenza virus infection in mice leading to complete mortality, and that synergy between the classical and alternative pathways is necessary for efficient protection.
The role of the individual complement pathways during influenza virus infection is not clear. Thus to address this, we examined the relative in vivo contribution of the individual complement pathways in providing protection against the A(H1N1)pdm09 virus infection. All the three pathways converge at C3 activation step. Hence to understand the role of intact complement, C3-/- mice were infected by inoculating a sub-lethal dose of the virus by the intranasal route. Infection in C3-/- mice showed severe illness with significant weight loss leading to 100% mortality by day 11 post-infection (p.i.) (Fig 1A & 1B). However, WT mice showed only 10% weight loss at the peak of infection, and all mice fully recovered at day 12 p.i (Fig 1A & 1B), strongly establishing that complement plays a protective role during the pandemic influenza virus infection.
Next, to determine the contribution of the individual pathways, we infected C4-/- mice [deficient in classical pathway (CP)/lectin pathway (LP)] and FB-/- mice (deficient in alternative pathway; AP) and monitored them for weight loss and mortality. Results showed significant weight loss in both the knockout strains compared to the WT mice (Fig 1A) with 32% mortality in C4-/- and 24% mortality in FB-/- mice (Fig 1B). Together, these data suggest that the CP/LP and AP are capable of providing a certain degree of protection owing to the activation of each in the absence of the other, however, cooperativity between these pathways is needed to provide complete protection against the influenza A(H1N1)pdm09 virus infection.
To determine whether the increased mortality observed above in the complement deficient mice infected with the A(H1N1)pdm09 virus is associated with increased pulmonary pathology in these mice (C3-/-, C4-/- and FB-/-), we performed histopathological analysis of the lungs collected at days 4 and 7 p.i. Mock infected mice showed normal lung architecture with intact cellular details of alveoli, bronchi and blood vessels at both these time points (S1 Fig). At day 4 p.i. (Fig 2A), virus infected WT mice showed minimal pathological changes with congested blood vessels and mild emphysematous changes of alveolar parenchyma and normal bronchiolar epithelium. However, at day 7 p.i. (Fig 2B), the virus infected WT mice showed mild-to-moderate pathological changes that included foci of haemorrhage in lung parenchyma/alveoli, peri-bronchial and peri-vascular infiltration of inflammatory mononuclear cells (MNCs) comprising mainly of lymphocytes, alveolar macrophages (AMs), neutrophils and few histiocytes. Other changes include exudation, emphysema, and interstitial deposits of inflammatory cells with alveolitis. The histopathological changes at both the time points were also apparent in C3-/- and FB-/- infected mice, but the overall degree of infiltration of inflammatory MNCs was higher in these mice compared to the WT mice. In addition, multiple foci of degenerative changes including loss of bronchiolar epithelium (sloughing) and bronchiolar hyperplasia, with edematous changes and consolidation of alveolar parenchyma were also observed in these mice. The C4-/- mice also showed similar changes, but the degree of inflammation was more similar to the WT mice (Fig 2A & 2B).
We next assessed the presence of infectious virus in the lungs of C3-/-, C4-/- and FB-/- mice infected with A(H1N1)pdm09 virus and compared that to the infectious virus titer in the lungs of infected WT mice (Fig 3). The viral load was found to be higher during the initial phase of the infection i.e., at day 4 p.i., in complement deficient (C3-/-, C4-/- and FB-/-) as well as WT mice. The load however decreased significantly at day 7 p.i. in the WT mice (p < 0.012), but not in the C3-/- mice suggesting that these mice are unable to clear the virus efficiently. The infectious virus titer also remained high in the C4-/- and FB-/- mice, but was relatively lower compared to that found in the C3-/- mice (Fig 3). Collectively, these data suggest that individual complement pathways are capable of clearing the virus to a certain extent, but their summative association is required for the efficient clearance.
Development of the antibody response against influenza hemagglutinin (HA) is linked with immune protection [31]. Thus, next we asked whether the antibody response to HA is reduced in the pandemic influenza virus infected complement deficient mice (C3-/-, C4-/- and FB-/-) compared to the WT mice. Examination of the HA-inhibition (HAI) antibody titers in the WT mice showed lower titers at day 4 p.i., and this significantly increased at day 7 p.i. (p < 0.05, Fig 4A). This marked increase in the HAI antibody titer however was not observed in C3-/- and C4-/- mice, suggesting that the loss of immune protection in these mice could, in part, be due to decrease in the antibody titer. This raised the question: Are these protective antibodies IgM or IgG? Measurement of the pandemic influenza A virus HA-specific IgM and IgG in sera of WT and C3-/- mice 7 days p.i. showed that IgM, but not IgG levels are significantly lower in C3-/- mice compared to the WT mice (Fig 4B). Consequently, protection seems to be primarily mediated by IgM. Interestingly, the HAI antibody titers in FB-/- mice at day 7 p.i. were similar to those in the WT mice, suggesting that immune factors other than the HAI antibodies ought to be responsible for the partial loss in immune protection seen in these mice (Fig 1B).
Because C3-/- mice showed significantly decreased HAI antibody titer compared to the WT mice, we determined whether the loss of protection against the influenza A(H1N1)pdm09 virus infection in C3-/- mice is indeed, in part, due to the decreased antibody response. Thus, immune sera collected at day 7 p.i. from the WT or C3-/- mice was heat-inactivated, and injected intraperitoneally (i.p.) in naïve C3-/- mice, which were then challenged by intranasal (i.n.) route with the pandemic influenza virus. The immune sera from the WT mice significantly prevented the weight loss (Fig 5A) and mortality (Fig 5B). However, the immune sera from the C3-/- mice failed to prevent the weight loss and only slightly delayed mortality (Fig 5). Together, these results suggest that impaired virus-specific antibody response, in part, is responsible for the severe viral infection and mortality in C3-/- mice.
It was clear from the above data that C3 plays a pivotal role in enhancing the protective humoral response against the pandemic influenza virus infection. The major source of C3 is liver, but extra-hepatic C3 has been shown to play a key role in enhancing the humoral response to peripheral viral infection [32]. Therefore, this raised the question whether liver-derived C3 or extra-hepatic C3 is more critical in providing the protection. To distinguish the contribution of C3 produced from liver cells and leukocytes, we adoptively transferred splenocytes either from naïve WT or from naïve C3-/- mice into C3-/- mice, and challenged these mice with the A(H1N1)pdm09 virus (Fig 6A). Results showed that the C3-/- mice that received WT splenocytes showed recovery in weight and increased survival (62.5%; p < 0.001) whereas those that received C3-/- splenocytes failed to prevent weight loss and caused complete mortality (Fig 6B). These data therefore point out a major role of extra-hepatic C3 in providing protection against the A(H1N1)pdm09 virus. Next, we determined whether splenic T cells or B cells can serve as a source of C3 and correct the C3-/- phenotype as both are known to synthesize C3 [33]. We thus purified T and B cells from the spleen of naïve WT or C3-/- mice (S3 Fig) and adoptively transferred into C3-/- mice which were then challenged with the A(H1N1)pdm09 virus (Fig 6C). We observed that WT B cells, but not WT T cells could partially protect the C3-/- mice against the pandemic influenza virus infection suggesting that C3 produced by B cells contribute significantly to the generation of protective immune response against the virus.
Complement activation results in the generation of immuno-modulatory peptides termed C3a and C5a. These peptides have been shown to modulate the expression of costimulatory molecules on dendritic cells by signaling through their receptors C3aR and C5aR and thereby influence T-cell priming [34]. Their role has also been studied in seasonal influenza virus infection. Signaling through C3aR and C5aR has been shown to be critical for migration of dendritic cells to the draining lymph node in response to influenza virus infection [27]. In view of the above, we asked, is C3aR and C5aR signaling vital during the pandemic influenza virus infection? We thus blocked the C3aR and/or C5aR signalling in the WT mice by pharmacologically targeting these receptors with specific antagonists (C3aRA and C5aRA), and monitored the severity of infection. Thus, WT mice challenged with A(H1N1)pdm09 virus received a daily i.p. injection of C3aRA or C5aRA or both (1mg/kg of body weight) from day -2 to 10 p.i. (Fig 7A). Results showed significant weight loss and mortality (60%) in mice that received both C3aRA and C5aRA compared to the vehicle treated mice. Mice treated with C5aRA alone however recovered well with only 10% mortality. Notably, mice that received only C3aRA also showed significant weight loss with 57% mortality (Fig 7B & 7C). These data suggest that C3aR-mediated signalling plays a predominant role in generating a protective immune response to the pandemic influenza virus, with signalling through C5aR playing a relatively minor role.
To further strengthen our conclusion that C3aR plays a principal role during pandemic influenza virus infection, we also performed infection experiment in C3aR-/- and C5aR-/- mice available on BALB/c background. Although earlier studies have shown that A(H1N1)pdm09 virus infects C57BL/6 and BALB/c mice equally well [35], we included C3-/- mice on BALB/c background to ascertain whether mouse strain differences affect the susceptibility to the virus infection. As expected, the C3-/- mice displayed sustained weight loss with 100% mortality; only 10% mortality was observed in WT BALB/c mice (Fig 7D and 7E). Further, conforming to our receptor antagonist data (Fig 7B & 7C), C3aR-/- mice showed heightened susceptibility to the pandemic influenza virus infection, while C5aR-/- mice showed complete recovery. If anything, C5aR-/- mice recovered better than the WT mice.
Our in vivo data clearly established the role of complement in controlling the pandemic influenza virus infection in mice. Thus, next to determine whether the virus is susceptible to direct neutralization by complement we performed in vitro virus neutralization assays. To measure the classical pathway (CP)-mediated neutralization, the virus was incubated with normal mouse plasma (as a source of complement) in the presence of virus-specific mouse antibodies and Ca++-Mg++. And to examine the alternative pathway (AP)-mediated neutralization, the virus was incubated with normal mouse plasma in the presence of Mg-EGTA which allows selective activation of the AP. Our results showed that the virus was efficiently neutralized by CP (Fig 8A), but not by AP (Fig 8B). Though these results were in agreement with enhanced infection seen in C4-/- mice, they were in sharp contrast to the enhanced infection observed in FB-/- mice (Fig 1). We thus tested the possibility whether the pandemic virus becomes susceptible to AP-mediated neutralization when coated by antibodies. It is clear from the results presented in Fig 8C that the virus indeed becomes susceptible to neutralization by AP when coated by antibodies. Together these results suggest that the virus becomes susceptible to complement either by CP or by AP only when coated by antibodies.
We observed in vivo synergy between the CP and AP of complement for providing protection against the pandemic influenza virus infection. To address whether such synergy also operates during in vitro neutralization of the virus, we tested if the CP-mediated deposition of C3b onto the viral surface leads to activation of the AP loop and thereby augment neutralization. Examination of the CP-mediated neutralization of the virus using FB-/- mouse plasma showed significant reduction in the extent of neutralization compared to the normal mouse plasma (Fig 8D) suggesting that synergy does exist between the classical and alternative pathways during the pandemic virus neutralization.
Earlier studies on seasonal influenza virus have demonstrated that deposition of the early complement component (e.g., C3b and C4b) followed by aggregation of the virus is enough for neutralization, and activation of the terminal pathway leading to MAC-mediated lysis is not essential [28]. We thus also looked into the requirement of the terminal pathway for CP-mediated neutralization of the pandemic influenza virus. We observed that C5-deficient plasma was able to neutralize the virus as efficiently as the C5-sufficient plasma (Fig 8E). It is thus evident that MAC-mediated lysis is not necessary for the neutralization of the pandemic influenza virus.
Influenza A(H1N1)pdm09 virus is a human pathogen, hence next we sought to establish the susceptibility of this virus to the human CP- and AP-mediated neutralization. The neutralization assays employed were essentially similar to those described above for the mouse complement, except that the human sera (a source of complement) and virus-specific human antibodies (isolated from the pandemic influenza virus infected subjects) were used. The results obtained were essentially similar to that observed using the mouse complement: the virus showed susceptibility to CP-mediated neutralization, but resistance to AP-mediated neutralization (Fig 9A & 9B). And as expected, coating of virus with antibody resulted in AP-mediated neutralization (Fig 9C).
We next asked whether seasonal influenza A virus, A/Perth/16/2009(H3N2), is also resistant to AP-mediated neutralization. Our results showed that unlike the pandemic influenza virus, seasonal influenza virus is susceptible to AP-mediated neutralization (Fig 9D). These results intrigued us to investigate whether the difference in AP-mediated neutralization of the above two viruses is due to the difference in the ability of their surfaces to allow C3b deposition. Activation of C3 near their surfaces using purified complement components confirmed our presumption: C3b deposition was efficient on the H3N2 viral surface, but not on the pandemic influenza virus unless coated by antibodies (Fig 9E). Thus, there exists distinct difference in the susceptibility of these two viruses towards AP-mediated neutralization.
Earlier mannose-binding lectin (MBL) has been shown to bind to various strains of the seasonal influenza viruses [36], but the LP does not seem to neutralize these viruses [28]. Therefore, we also examined if the pandemic influenza virus is susceptible to the LP-mediated neutralization using C1q-deficient human serum that lacks CP, but contains intact LP. Interestingly, C1q-deficient sera failed to neutralize the virus suggesting that LP does not play a role in neutralizing the pandemic influenza virus (Fig 9F).
The observed susceptibility of the A(H1N1)pdm09 virus to the human CP portrayed the importance of the virus-specific complement-binding antibodies in controlling the virus. This raised the question whether such antibody response is robust in humans during the pandemic influenza virus infection. We therefore looked for generation of such antibodies in the infected individuals. A total of 21 human serum samples collected from the individuals, which were positive for the presence of antibodies against the A(H1N1)pdm09 virus, were heat-inactivated and screened for the presence of complement binding antibodies by measuring their ability to enhance virus neutralization in the presence of complement. The sera from the infected individuals demonstrated varied antibody-mediated complement-enhanced neutralization potential (Fig 10A). Based on the results, the sera were grouped as: i) showing substantial antibody-mediated complement-enhanced neutralization (>50%; Group I), ii) showing moderate antibody-mediated complement-enhanced neutralization (30%-50%; Group II), and iii) showing negligible antibody-mediated complement-enhanced neutralization (<30%; Group III) (Fig 10A).
Next, to determine which class/subclass of the virus-specific antibody is responsible for supporting the complement-enhanced neutralization of the virus, we looked at the dependence of antibody-mediated complement-enhanced neutralization on the class/subclass of pandemic influenza virus-specific antibody (S2 Fig). We observed a significant negative correlation between the IgG titer and the virus titer that remained after the neutralization (r = - 0.501, p = 0.02) and between the IgG1 titer and the virus titer that remained after the neutralization (r = -0.628; p < 0.002). These correlations suggest that IgG1 antibodies present in the sera of A(H1N1)pdm09 virus-positive individuals are capable of supporting the complement-enhanced neutralization (Figs 10B & S2).
The complement system has been shown to impact the control of influenza virus infections [16,27,28], though the contribution of all the respective pathways are not yet fully determined. In the present study, we have investigated which complement pathways are triggered by the influenza A(H1N1)pdm09 virus and their impact in controlling the infection. Our data establish that synergy between the CP and AP is the key to effective control of the pandemic influenza virus infection. We demonstrate that C3 deficiency results in complete lethality, while C4 or FB deficiency results only in partial lethality (24% - 32%), suggesting in vivo cross-talk between the CP and AP. Such cross-talk was also apparent during in vitro virus neutralization wherein triggering of the AP loop following the activation of CP resulted in more efficient neutralization of the virus.
Complement deficiencies (C3-/-, C4-/- or FB-/-) in mice resulted in increased susceptibility to the pandemic influenza virus infection leading to partial or complete lethality. To look into the possible cause, we examined the effect of the deficiencies on the histopathological changes and viral clearance in the lung. We observed an overall higher degree of inflammation in C3-/- and FB-/- mice compared to the WT mice, but not in C4-/- mice i.e., the lethality in mice did not correlate well with exacerbated inflammatory changes. The viral clearance in the lung however showed a better correlation with lethality: WT mice effectively cleared the virus from the lung by day 7 p.i., but the same was not observed in C3-/- mice. The viral load also remained high in C4-/- and FB-/- mice lungs, but was lower compared to that observed in C3-/- mice. Intriguingly, initial virus load did not differ between the WT and complement deficient mice suggesting that complement-mediated protection seen in WT mice was a late phenomenon which occurred after day 5 p.i.
It is now well established that there is generation of strain-specific antibody response to the surface antigens of influenza viruses such as HA [37] and NA [38] during infection. It is also known that antibodies to HA in particular are key to protection against influenza [39,40] and these HA specific antibodies bind to complement [41]. Additionally, it is also documented that these antibodies are generated about 5 days post-infection [41]. We thus asked whether generation of antibodies against HA antigen is affected in complement deficient mice, and whether individual pathways differ in inducing anti-HA response. Our results demonstrate that C3 deficiency results in a significant decrease in HA-specific antibodies (particularly of IgM subclass) suggesting that generation of HA antibodies can be modulated by complement. It is also notable that apart from C3-/- mice, reduced HA antibody response was also observed in C4-/-, but not in FB-/- mice, suggesting that the CP/LP, but not the AP is critical for generation of HA response. Similar pathway-specific difference in the antibody responses was observed earlier during West Nile virus infection [42].
Though the above data showed a correlation between the C3 deficiency and reduction in response to HA, it did not establish the causal relationship between reduction in response to HA and higher lethality in C3-/- mice. Thus, to establish such a relationship, we performed rescue experiment with immune sera from A(H1N1)pdm09 virus infected WT and C3-/- mice collected at 7 days p.i., a time point at which significant difference in HA response was observed between the WT and C3-/- mice. Injection of WT immune sera to C3-/- mice infected with the virus provided considerable protection (lethality was reduced from 100% to 33%), while injection of C3-/- immune sera failed to provide any protection. These observations ascertain that decreased antibody response in C3-/- mice, at least in part, is responsible for the reduced immune protection in these mice resulting in lethality.
Importance of C3 and C4 in humoral immune responses has been investigated earlier using C3 and C4 deficient mice. These mice showed impaired ability to form germinal centres and develop antigen-specific antibody responses [43]. A similar phenotype was also observed in the C3d receptor (CD21/CD35) knockout mice [44,45]. Importantly, it was also shown that C3 synthesized in the splenic lymphoid compartment is capable of reconstituting the impaired humoral responses in C3-/- mice [46]. The mechanisms involved in C3-mediated germinal center responses include: i) activation of naïve B cells owing to co-ligation of B cell receptor and CD21 (a part of CD19-CD21-CD81 complex) by C3d (a C3 fragment) coupled to the antigen [47,48] and ii) trapping and retention of C3d coupled antigen on follicular dendritic cells expressing CD21/CD35 [49]. The major source of C3 in the splenic lymphoid compartment was thought to be macrophages [46]. In the present study, we show that C3 synthesized by B cells can also partially correct the C3-/- phenotype. These data therefore suggest that apart from macrophages, C3 produced by B cells also play a significant role in initiating the germinal center responses and mounting protective immunity.
Previously it has been shown that C3 deficiency results in enhanced viral spread and impaired recruitment of virus-specific CD4+ and CD8+ effector T cells in the lung during seasonal influenza infection owing to attenuated T cell priming [16]. Interestingly, such defect was not observed in CD21/CD35-deficient mice suggesting the involvement of other C3 receptor(s) [16]. In a later study, impaired priming of T cells during seasonal influenza virus infection in C3-/- mice was linked to reduced migration of dendritic cells (DCs) from the lung to the draining lymph nodes as a result of lack of direct signaling through C3a and C5a receptors expressed on the lung DCs [27]. In the present study, we observed a high degree of lethality in the WT mice infected with the pandemic influenza virus after blocking the C3aR signaling with little effect after blocking the C5a signaling using receptor antagonists. Similar results were also obtained when we employed C3aR-/- and C5aR-/- mice to study the relative importance of these two receptors during the pandemic influenza virus infection. Thus, our data suggest that primarily the C3aR-mediated signaling is critical for optimum priming of CD4+ and CD8+ T cells. The CD4+ T cells then provide help for optimal generation of neutralizing antibodies, and these antibodies along with CD8+ T cell effector function provide protection against the pandemic influenza virus infection.
It is apparent from the above discussion that complement activation and synergy between the various complement pathways is vital for effective control of in vivo pandemic influenza virus infection. To illustrate how various complement pathways are triggered during the influenza virus infection and why synergy is critical, we performed a series of in vitro experiments. Data obtained revealed a number of features of the virus-complement interactions.
First, the pandemic virus triggered complement only when they were coated by antibodies. No neutralization of the virus was observed by the CP or AP when it was incubated with the sera or plasma alone, but neutralization was apparent by both the pathways when the virus was coated with antibodies. Extrapolation of these results to in vivo situation would mean that complement is unable to control the pandemic influenza virus until the appearance of virus-specific antibodies. This proposition gains support from the fact that complement deficient mice do not show any difference in the virus titer compared to the WT mice at 4 days post-infection, a time when antibody response to HA is negligible.
Secondly, the pandemic influenza virus surface is not amenable to C3b deposition. It is well known that deposition of C3b onto the target surface is key to activation of the AP [50], which then culminates in virus neutralization [51]. Direct activation of C3 near the viral surface showed significant deposition of C3b onto the seasonal influenza A(H3N2) virus, but not on the pandemic influenza virus, unless coated by antibodies, which correlated with the neutralization of these viruses. These data thus indicate that modification of the pandemic influenza virus surface by antibodies is necessary for C3b deposition and thereby neutralization. A possible mechanistic explanation for the difference in C3b deposition on the pandemic and the seasonal H3N2 influenza viruses is as below. C3b is deposited onto the activator surface due to its ability to covalently attach to the acceptor molecules on that surface via ester or amide linkages. This attachment of C3b however is not a nonspecific reaction as C3b displays strong preferences for certain carbohydrates and high degree of specificity for particular sites on the acceptor molecules [50,52–54]. For example, Thr144 and Ser1217 have been identified as the major attachment sites for C3b on IgG1 and C4b, respectively [52–54]. HA and NA are the two major surface glycoproteins of the influenza A viruses that project out from its outer surface. It is therefore likely that C3b attachment onto the influenza viruses takes place due to covalent linkage of C3b to HA and/or NA. Since C3b displays specificity for specific residues on the acceptor molecules, it is conceivable that it attaches to specific residues on these molecules. The HA and NA of the pandemic and seasonal H3N2 influenza viruses belong to different antigenic subtypes, therefore, we propose that C3b acceptor site(s) have been altered in the pandemic influenza virus and thus C3b is only attached when its surface is modified by IgG, which is an effective C3b acceptor.
Thirdly, simultaneous activation of the CP and AP support enhanced neutralization of the pandemic virus. Activation of the CP is known to trigger the AP loop resulting in additional deposition of C3b onto the target surface. Here we showed that such amplification of the AP loop as a result of CP activation leads to enhanced neutralization of the pandemic virus. Thus, synergy between the CP and AP result in augmented complement activation ensuing more neutralization of the virus owing to C3b deposition.
In summary, our data reveal the importance of cross-talk between the CP and AP that provides sufficient trigger (C3b deposition and C3a production) required for efficient protection against the pandemic influenza virus infection. Based on our present findings and from the earlier studies, we propose the following model for complement-mediated protection during the pandemic influenza H1N1 2009 virus infection (Fig 11) wherein: i) recognition of the pandemic influenza virus by antibodies triggers the activation of CP as well as AP leading to C3b deposition and direct neutralization of the virus to a certain extent, ii) the complement activation fragments C3d, as well as C3a generated as a result of complement activation enhance B cell responses and the effector CD4+ and CD8+ T cell responses, respectively and iii) the effector CD8+ T cells and antibodies then efficiently contain the virus infection.
All the animal experiments performed in this study were approved by the Institutional Animal Ethics Committees of the National Centre for Cell Science, Pune (NCCS) and the National Institute of Virology (NIV), Pune. Use of human serum samples from A(H1N1)pdm09 virus positive and negative individuals for the study was approved by the Institutional Ethical Committees of NCCS and NIV. All adult subjects provided informed written consent; for child participants informed written consent was provided by the parent/guardian of the child.
The pandemic influenza A(H1N1)pdm09 virus (A/India/Jln_NIV9436/2009; GenBank accession nos: HM204573, HM241701-07 [12]) and influenza A(H3N2)09 virus (A/India/NIV33041/2009) utilized in this study were isolated in 2009 at NIV, Pune. The virus stocks were prepared by propagating the virus in the embryonated chicken eggs. In brief, the virus was inoculated into the allantoic cavity of 10-day-old chicken embryos. The allantoic fluid was then harvested 72 hr post inoculation and the virus titer were determined by hemagglutination (HA) assay. Samples having HA titers of ≥64 were pooled together to make the virus stock and stored at -80°C in aliquots. The infectious virus titer (50% tissue culture infectious dose, TCID50) of the stock was determined as described below and calculated by employing the Reed and Muench method [55]. For ELISA, the virus was inactivated using beta-propiolactone (β–PL) [56] and purified over 10%-50% sucrose gradient.
Wild type BALB/c and C57BL/6 mice, C3-/- and C4-/- on C57BL/6 background, and C3aR-/- and C5aR1-/- mice on BALB/c background were purchased from The Jackson Laboratory, USA. Factor B-/- (FB-/-) mice on C57BL/6 background and C3-/- mice on BALB/c background were a kind gift from Prof. Marina Botto, Department of Medicine, Imperial College of London. C5-sufficient B10.D2/nSnJ and C5-deficient B10.D2/oSnJ mice were also procured from The Jackson Laboratory, USA. All the mice were bred and housed in a barrier-maintained animal facility of NCCS. Animal experiments were performed in the Biosafety Level 3 (BSL-3) laboratory at NIV, and the mice were housed in HEPA-filtered, negative pressure, individually ventilated cages.
Madin-Darby Canine Kidney (MDCK) cell line used for TCID50 determination was obtained from the Centers for Disease Control and Prevention, USA and maintained in Dulbecco’s modified Eagles medium (DMEM) containing 10% FCS, 200 U/ml penicillin and 0.2 mg/ml streptomycin at 37°C in a 5% CO2 atmosphere with 95% humidity. Human IgG required for the virus neutralization assays was purified from pooled serum samples of A(H1N1)pdm09 virus positive individuals by caprylic acid precipitation method. Mouse IgG needed for the virus neutralization assays was purified from the pooled sera of WT mice infected with the A(H1N1)pdm09 virus using Hi-Trap Protein A column (GE Healthcare Bio-Sciences, Sweden). Homogeneity of the antibodies was assured by SDS-PAGE analysis. Normal human serum (NHS) used as a source of complement was obtained from subjects negative for the A(H1N1)pdm09 antibodies. C1q depleted human serum was purchased from Complement Technologies, USA. Normal (WT), C3-/-, C4-/-, FB-/- and C5-sufficient and C5-deficient mouse plasma used as a source of complement were obtained from the respective naïve mice by collecting the blood in 20 mM EDTA. Heat-inactivated human serum and mouse plasma were prepared by inactivating them at 56°C for 30 min. Immune mouse serum utilized for rescue of the A(H1N1)pdm09 virus infected C3-/- mice was prepared by drawing the blood from C57BL/6 mice infected with 450 TCID50 of the virus at indicated time; serum having HAI antibody titers ≥320 were pooled together and used for the rescue study.
A group of female mice (~8 week old) were lightly anesthetized and infected i.n. with 450 TCID50 of the egg grown A(H1N1)pdm09 virus in 50 μl of allantoic fluid diluted in phosphate-buffered saline (PBS). Mock infection was performed with the allantoic fluid from uninfected eggs diluted in PBS. After infection, the mice were weighed individually each day and observed for signs of illness and mortality. Mice with severe infection and body weight loss of ≥ 30% of their initial body weight were sacrificed; these mice were also considered for calculation of mortality. To determine the virus titers and for histopathology, the mice were sacrificed on days 4 and 7 p.i., and their lungs and blood were harvested under sterile conditions. One part of the lung was utilized for determining the virus titer by TCID50 assay, while the other part was fixed in 10% formalin saline for histopathology. Blood was allowed to clot to collect serum for determination of the antibody response using HAI assay and ELISA.
For antagonist studies, wild type mice received daily intraperitoneal injection (i. p.) of either C3a receptor antagonist (C3aRA; SB 290157; GenoMechanix, USA) or C5a receptor antagonist (C5aRA; Ac-Phe-[Orn-Pro-dCha-Trp-Arg]; GenoMechanix, USA) or both (1mg/kg of body weight in 1.16% DMSO) from day -2 to 10 post-infection.
For rescue studies with immune sera, C3-/- mice received two i.p. injections of pooled immune sera (250μl each) at day -1 and 6 post-infection. The immune sera was collected at day 7 post-infection from 20 WT mice infected with A(H1N1)pdm09 virus.
For adoptive transfer studies, C3-/- mice received (i.v.) either total splenocytes (1 x 106 cells/mouse) or purified splenic B or T cells (1 x 106 cells/mouse) isolated from naïve WT or C3-/- mice. After 7 days of adoptive transfer, these mice were infected intranasally with the pandemic influenza virus (450 TCID50) and monitored for weight loss and survival. For splenocyte preparation and purification of T and B cells, spleens from WT or C3-/- mice were harvested and minced to prepare single cell suspensions and RBCs were removed using ammonium-chloride-potassium (ACK) lysis buffer. These splenocytes were stained with Alexa Fluor 647 conjugated anti-mouse CD3ε (Clone 145-2C11; Biolegend, San Diego, CA) and FITC conjugated anti-mouse CD19 (clone eBio1D3; eBioscience, San Diego, CA) on ice for 30 minutes. The cells were then washed with PBS and passed through 70 μm pore containing cell strainer. Purification of T cells (CD3ε+CD19- cells) and B cells (CD19+CD3ε- cells) was performed using BD FACS ARIA III sorter (BD Biosciences, San Jose, CA). The purity of these sorted cells was >96% (S3 Fig).
For virus titration, lungs were harvested under sterile conditions, weighed, and homogenized in viral transport medium (Hank’s balanced salt solution containing 250 μg/ml of gentamicin, 2000 U/ml of penicillin and 200 μg/ml streptomycin) to make 10% lung suspension. This suspension was then centrifuged and the supernatant obtained was used for quantitating the virus by TCID50 assay. In brief, the lung samples (25 μl) were serially diluted (3-fold) in a 96-well microtiter plate and incubated at 37°C in 5% CO2 for 1 hr. Thereafter, 100 μl of MDCK cells (1.5 x 104 cells/well) were added to each well and the plate was kept at 37°C for 18 hrs in 5% CO2. The infected cells were then fixed with 80% cold acetone and wells were blocked with 1% milk in PBS-T (PBS containing 0.1% tween 20). To detect viral antigens, wells were washed three times with PBS-T and incubated at 22°C for 1 hr each with 100 μl rabbit anti-influenza A antibody (1:3000 dilution; raised at NIV) followed by a 1:2000 dilution of peroxidase-conjugated goat anti-rabbit IgG antibody (Sigma). Both the above incubation steps were followed by five washes with PBS-T. Finally, the optical density was measured at 490 nm after adding ortho-phenylene diamine (OPD) and stopping the reaction with 1 N H2SO4. The virus titer (TCID50) was calculated using the Reed and Muench method [55].
For histopathology, the lung tissues were fixed in 10% neutral buffered formalin. These tissues were then embedded in paraffin, cut into 4μm thick sections and stained (with hematoxylin and eosin) and analyzed at the Veterinary College Core Facility, Krantisinh Nana Patil College of Veterinary Science, Shirwal.
The HAI assay was performed to determine the anti-HA antibody response in the infected mice. In brief, 50 μl of RDE (receptor destroying enzyme; Denka Siken, Japan) treated mice serum (pre-diluted, 1:10) was serially diluted (2-fold) in a V-bottom 96 well microtiter plate (Tarsons, India). The diluted serum was then mixed with 25μl of 8 HA units of β-PL inactivated virus and incubated at room temperature for 30 min. Thereafter, 50μl of 0.5% turkey RBC was added to each well and the plate was kept at 4°C for 30 min before reading. The reciprocal of the last dilution at which the antibodies were still able to inhibit the virus-mediated RBC agglutination was considered as the HAI antibody titer.
For detection of antibodies against influenza A(H1N1)pdm09 virus in human serum, microtiter plates (Greiner Bio-One) were coated overnight at 4°C with β-PL inactivated virus (250 ng/well). The wells were then blocked with 5% milk, washed thrice and incubated with the diluted sera (1:1600 for IgG; 1:400 for IgM; 1:1600 for IgG1, IgG2, IgG3 and IgG4) for 1 hr at room temperature. To detect the bound IgG antibodies, wells were washed three times with PBS-T and incubated with mouse anti-human IgG, IgG1, IgG2, IgG3 or IgG4 (1:1000 dilution; Sigma Aldrich, USA), followed by rabbit anti-mouse HRP conjugate (1:500 dilution; BioRad) and ABTS. IgM antibodies were detected with goat anti-human IgM HRP (1:250 dilution; Sigma Aldrich, USA). Incubation with class/subclass specific antibodies was for 1 hr at room temperature and that for the conjugate was for 30 min at room temperature; each of the incubation steps, except ABTS, was followed by three washes with PBS-T. The optical density was read at 415 nm.
Measurement of HA-specific IgM and IgG in sera of A(H1N1)pdm09 virus infected mice was performed as below. The microtiter plates (Greiner) were coated with pandemic influenza virus HA protein [200 ng/well; rHA of the H1N1 (A/California/7/2009); ThermoFisher Scientific, Waltham, MA] by keeping the plates overnight at 4°C; control wells were coated with similar amount of BSA. The wells were then blocked by adding 5% milk, washed once with PBS-T and incubated with 1:50 diluted heat inactivated mouse sera for 1 hr at room temperature. Thereafter, the wells were washed three times with PBS-T and incubated with goat anti-mouse IgM HRP conjugate (1:1000 dilution; Sigma Aldrich) or rabbit anti-mouse IgG HRP conjugate (1:1000 dilution; Sigma Aldrich) for 30 min at room temperature. The unbound conjugate was then washed, ABTS was added, and optical density was measured at 415 nm. Mouse sera and conjugates were diluted in PBS-T containing 0.5% milk and 0.5% BSA.
Virus neutralization was studied using the human as well as the mouse complement. To determine the human CP-mediated neutralization of the virus, 32 μl of the A(H1N1)pdm09 virus (500 TCID50) was mixed with 12 μl of human IgG (36 μg) purified from influenza A(H1N1)pdm09 virus positive subjects and 16 μl or 32 μl serum from the pandemic virus negative individuals as a source of complement in the presence of 0.15 mM Ca++ and 0.5 mM Mg++ in a total volume of 320 μl DMEM containing 1% BSA. The reaction mixture was then incubated at 37°C for 1 hr, diluted in DMEM, and assayed for virus titer by determining TCID50. For the human AP-mediated neutralization of the virus, 70 μl of the A(H1N1)pdm09 virus (850 TCID50) was mixed with 70 μl, 210 μl or 350 μl of serum from the pandemic virus negative individuals as a source of complement in the presence of 5 mM Mg-EGTA in a total volume of 700 μl DMEM containing 1% BSA. As for CP, the reaction mixture was then incubated at 37°C for 1 hr, diluted in DMEM, and assayed for virus titer by determining TCID50. Virus neutralization by the mouse CP and AP of complement was performed in a manner essentially similar to that described above for the human CP and AP with the exception that indicated amount of mouse plasma instead of sera was used as a source of complement and excess of Ca++/Mg++ and Mg++ was added to neutralize EDTA present in the plasma for measuring CP- and AP-mediated neutralization, respectively.
For the AP loop studies, CP-mediated neutralization was performed with FB-/- mice plasma. To study the role of antibodies in AP activation, virus was first incubated with virus specific antibodies followed by AP-mediated neutralization with human serum and mice plasma respectively, as mentioned.
ELISA plates were coated with 300 ng/well of virus [A(H1N1)pdm09 or A(H3N2] overnight at 4°C in PBS. The wells were then blocked with 5% skimmed milk, washed once, and reactions were set up on ice by adding three different concentrations of AP components (Rxn#1: 6 μg C3, 1.5 μg FB and 60 ng FD; Rxn#2: 4 μg C3, 1 μg FB and 40 ng FD; Rxn#3: 2 μg C3, 0.5 μg FB and 20 ng FD) in the presence (5 μg, 2.5 μg or 1.25 μg) or absence of antibody in a total volume of 100 μl GVB containing 5 mM Mg-EGTA. Thereafter, the plate was incubated at 30°C for 30 min and washed three times. C3b deposited on the viral surface was detected by adding anti-human C3 HRP conjugate (1:1000 dilution, ICN; 45 min incubation at room temperature) followed by ABTS. The optical density was read at 415 nm.
A total of 21 human serum samples from the A(H1N1)pdm09 virus positive individuals were screened for the presence of complement binding antibodies. In brief, sera of the infected individuals were heat inactivated at 56°C for 30 min and utilized for virus neutralization in the presence or absence of human complement. For neutralization, 850 TCID50 of the A(H1N1)pdm09 virus in 70 μl was incubated with 10 μl of heat inactivated serum from the virus positive subject and 35 μl of virus negative human serum as a source of complement in the presence of 0.15 mM Ca++ and 0.5 mM Mg++ in a total volume of 700 μl at 37°C for 1 hr. The reaction mixture was then titrated for infectious virus by TCID50 as described above.
One-way ANOVA followed by Tukey’s Post-Hoc test was performed to compare different groups. Comparison between two groups was made using Mann-Whitney Rank Sum test. The percent survival of mice in different groups was plotted as Kaplan-Meier plots and analyzed using Log-Rank test. Results of in vivo studies are presented as mean ± SEM, while results of in vitro studies are presented as mean ± SD. All the statistical analyses were performed using PASW Statistics 18 software (New York, USA).
|
10.1371/journal.pgen.0030207 | Activation of Inflammation/NF-κB Signaling in Infants Born to Arsenic-Exposed Mothers | The long-term health outcome of prenatal exposure to arsenic has been associated with increased mortality in human populations. In this study, the extent to which maternal arsenic exposure impacts gene expression in the newborn was addressed. We monitored gene expression profiles in a population of newborns whose mothers experienced varying levels of arsenic exposure during pregnancy. Through the application of machine learning–based two-class prediction algorithms, we identified expression signatures from babies born to arsenic-unexposed and -exposed mothers that were highly predictive of prenatal arsenic exposure in a subsequent test population. Furthermore, 11 transcripts were identified that captured the maximal predictive capacity to classify prenatal arsenic exposure. Network analysis of the arsenic-modulated transcripts identified the activation of extensive molecular networks that are indicative of stress, inflammation, metal exposure, and apoptosis in the newborn. Exposure to arsenic is an important health hazard both in the United States and around the world, and is associated with increased risk for several types of cancer and other chronic diseases. These studies clearly demonstrate the robust impact of a mother's arsenic consumption on fetal gene expression as evidenced by transcript levels in newborn cord blood.
| Arsenic is an environmental pollutant and known human carcinogen. Chronic exposure to arsenic-contaminated water is an important public health hazard around the world, including the United States, with millions exposed to drinking water with levels that far exceed World Health Organization (WHO) guidelines. Given the implications of prenatal exposure on human health and the known public health hazard of chronic arsenic exposure, this study was aimed at establishing the extent to which maternal arsenic exposure in a human population affects newborn gene expression. The authors show that prenatal arsenic exposure in a human population results in alarming gene expression changes in newborn babies. The gene expression changes monitored in babies born to mothers exposed to arsenic during pregnancy are highly predictive of prenatal arsenic exposure in a subsequent test population. The study establishes a subset of just 11 transcripts that captured maximal predictive capability that could prove promising as genetic biomarkers of prenatal arsenic exposure. Pathway analysis of the genome-wide response in the babies exposed to arsenic in utero indicates robust activation of an integrated network of pathways involving NF-κB, inflammation, cell proliferation, stress, and apoptosis. This study contributes to our understanding of biological responses to arsenic exposure.
| Arsenic is a ubiquitous environmental pollutant and a known human carcinogen [1]. Chronic arsenic exposure is an important public health hazard around the world, with millions of people exposed to drinking water with levels far exceeding the guideline of 10 μg/l established by the WHO. Exposure to arsenic-contaminated drinking water is alarmingly high in many countries, most notably Bangladesh, where >25 million people are chronically exposed to extreme arsenic levels. Arsenic contamination is also a significant health concern in the United States, with numerous public water supplies measuring above the WHO limit [2].
Epidemiological studies indicate that chronic arsenic exposure in drinking water is associated with increased risk of skin, bladder, lung, liver, and kidney cancer [1]; in 1987, arsenic was classified as a Group 1 carcinogen by the International Agency for Research on Cancer. Although the mechanism of arsenic-induced carcinogenesis is not clearly established, it has been attributed to genotoxicity associated with reactive oxygen species [3]. Arsenic is also implicated in other human diseases such as vascular disorders, peripheral neuropathy, bronchiecstasis, and diabetes [1].
The long-term health consequences of prenatal arsenic exposure in human populations are pronounced, with increased mortality rates caused by prenatal and early childhood exposures [4]. The detrimental health impact of prenatal arsenic exposure has also been shown in rodent models where in utero arsenic exposure resulted in a striking carcinogenic response (5-fold increase in hepatocellular carcinomas) among offspring; in utero arsenic exposure also changed the expression of genes involved in cell proliferation, stress, and cell–cell communication that are evident even when the offspring reach adulthood. These results have profound implications suggesting that in utero arsenic exposure may result in epigenetic changes that persist through the life of the organism, ultimately impacting health status. A landmark study in mouse models shows that, indeed, in utero exposures via the maternal diet can cause permanent gene expression changes in the offspring that affect susceptibility to disease in the adult [7].
Given the implications of prenatal exposure on human health and the known public health hazard of chronic arsenic exposure, we set out to establish the extent to which maternal arsenic exposure in a human population impacts newborn gene expression. Additionally, these studies were aimed at understanding exactly how arsenic affects biological systems and identifying genes that could be used as predictors, and therefore potential biomarkers, of prenatal arsenic exposure.
Our study was based in the Ron Pibul and Bangkok districts of Thailand (Figure S1). The first case of arsenicosis (arsenic poisoning) in Thailand was reported in 1987 from the Ron Pibul district [8]. Rather than natural leaching of arsenic from geologic sources, Ron Pibul arsenic contamination is attributed to tin mining that took place from the 1960s to the 1980s. Arsenic concentrations in groundwater and shallow wells have been classified at a mean level of 503.5 μg/l, about 50 times higher than WHO guidelines [9].
Using a population of arsenic-exposed and -unexposed mothers (as defined by WHO standards of chronic exposure to ∼10 μg/l arsenic), we set out to identify gene expression changes in the cord blood of newborns significantly associated with the extent of prenatal arsenic exposure. Cord blood is derived almost exclusively from the fetus; therefore, gene expression changes assessed in cord blood are representative of the newborn [10]. For this study, exposure classification was based on arsenic concentration in the mother's toenails, as this is representative of long-term arsenic accumulation [11,12]. Toenail samples were taken from a population of 32 volunteer subjects to quantify arsenic exposure in the mothers. A level of 0.5 μg/g toenail arsenic corresponds to chronic consumption of water with ∼10 μg/l (see Materials and Methods), which is the official WHO maximum recommended concentration of arsenic in drinking water [11,12]. For the purposes of this study, women with toenail arsenic levels of <0.5 μg/g were considered unexposed, and women with toenail levels of ≥0.5 μg/g were considered exposed. The levels of toenail arsenic across the 32 pregnant women ranged from 0.1 to 68.63 μg/g (Figure 1A). Given the paucity of available unexposed newborn cord blood from Ron Pibul, the experimental design required additional utilization of unexposed newborn cord blood samples from Bangkok.
We set out to determine whether gene expression changes in a set of infants born to arsenic-exposed women versus unexposed women (as judged by WHO guidelines) could be used to predict arsenic exposure in a test population. For these analyses, two-class prediction was employed, where a training population was used to derive gene sets that were then tested as predictors of exposure in a separate population. The analyses were carried out in two phases: (i) where the training population was selected at random and the analyst “blinded” to arsenic exposure level in the test population and (ii) where all arsenic exposure levels of the population were revealed and used to define new training populations.
The first training population comprised 13 newborn subjects selected at random from the 32 newborns (Figure 1A). Specifically, RNA was extracted from cord blood of newborns 1–13, and hybridized to whole human genome arrays (Materials and Methods). To identify genes whose expression was associated with prenatal arsenic exposure, we used an approach that combined differential expression testing between the populations, plus a positive or inverse correlation of expression with increasing arsenic exposure (Materials and Methods). From the 13 newborn subjects, we identified the first expression signature (first gene set, Figure 1B) composed of 170 genes (Table S1) that differentiated the unexposed newborns (subjects 1–6) from the arsenic-exposed newborns (subjects 7–13). This prenatal arsenic exposure expression signature of 170 genes was then used to predict prenatal exposure in the remaining population of 19 newborns (subjects 14–32). The percent accuracy of class prediction was determined post-analysis by revealing the arsenic exposure of the test population to the analyst. Expression of these 170 genes accurately predicted prenatal arsenic exposure in 15 of 19 (79%) of the newborns (Figure 1B).
When the arsenic levels of the entire population were revealed, it became apparent that the first training population was composed of newborns with a wide range of exposure levels distributed over almost the entire range (Figure 1B). We hypothesized that a training population based on extreme exposures might yield higher predictive capacity. To assess this, arsenic-associated genes were identified using newborns at the extremes of arsenic exposure (i.e., the lowest versus the highest exposures) as the second training population (Figure 1A, second training population). Six newborns comprised the low-exposure population (subjects 1, 14, 15, 2, 16, and 3), and six newborns comprised the high-exposure population (subjects 29, 30, 12, 13, 31, and 32) (Figure 1A). As with the first gene set, differential expression testing and correlation analysis identified an expression signature, this time composed of 38 genes (Table S2) that differentiated infants born to mothers with very low and very high arsenic exposure levels (Figure 1A). These 38 genes were used to predict arsenic exposure in the remaining test population of 20 newborns. Even though the gene set was smaller (38 versus 170), prediction was just as high as that of the first gene set, with prenatal arsenic exposure accurately predicted in 16 of 20 (80%) of the newborns (Figure 1B, second test population).
We next determined whether a training population derived from a combination of all of the training samples used to generate the first and second gene set would yield an expression signature with higher predictive capacity. This third training population was composed of nine unexposed newborns and 11 exposed newborns (Figure 1A). Differential expression testing and correlation analysis identified an expression signature of 11 genes (Figure 1B) that could predict prenatal arsenic exposure in 10 of 12 (83% accuracy) of the remaining newborn test population (Figure 1B). It is noteworthy that with only 11 genes, the power of prediction is as high as the first and second gene sets.
Many of the genes in the third gene set were represented in the gene sets derived from the first and second training populations. Specifically, five of the 11 were identified in the first gene set and all 11 were present in the second gene set (Table 1). Given the high predictive capacity of these 11 genes, we hypothesize that these are key genes involved in the prenatal response of babies to arsenic and represent potential biomarkers of arsenic exposure. The potential arsenic biomarker set is composed of transcripts for the CXL1, DUSP1, EGR-1, IER2, JUNB, MIRN21, OSM, PTGS2, RNF149, SFRS5, and SOC3 genes (Table 1). The dose response of expression level of each of the identified biomarkers is evident when plotted versus arsenic exposure across the population (Figure S2). Furthermore, to substantiate the association of the expression of the biomarkers with arsenic exposure, a multivariate model was employed (Materials and Methods). The model was employed to determine significance of association of expression with two factors: (i) arsenic exposure and (ii) geographic source of samples (Materials and Methods). Geographic source was determined to be a nonsignificant factor for the expression level of the biomarkers (p = 0.11), whereas arsenic exposure was determined to be a highly significant factor (p = 1.3 × 10−9). Furthermore, for the set of biomarkers, the two factors of arsenic exposure and geographic source were not associated (p = 0.77).
Notably, associated molecular functions for the 11 gene products include stress response and cell cycle regulation. The zinc finger DNA binding transcription factor EGR-1 (early growth response 1) is related to cell proliferation and is induced by mitogens such as EGF [13]. EGR-1 regulates both proinflammatory cytokine activation and p53 transcription [14,15]. Not surprisingly, as EGR-1 is known to activate cytokines, such signaling molecules are present in the arsenic biomarker gene set; namely, OSM (oncostatin M), a member of the interleukin-6 (IL-6) family of cytokines known to control cell cycle progression [16], CXL1 (chemokine ligand 1), and SOC (suppressor of cytokine signaling 3). Additionally, DUSP1 (dual specificity phosphatase 1) is involved in cell cycle regulation and is known to modulate cytokine expression [17,18]. An inflammation-activated acute phase response is indicated by the presence of the JUNB transcription factor, and IER2 (immediate early response 2) transcripts in the biomarker set.
For a more global assessment of the impact of prenatal arsenic exposure on fetal gene expression, all biological pathways modulated in response to arsenic exposure were identified by studying the ontology of all the genes differentially expressed between the exposed and unexposed newborns across the entire population. For these analyses, the entire newborn population was used (the fourth population, Figure 1A) to define the fourth gene set that was differentially expressed between the two populations: the 21 newborns whose mothers were exposed to arsenic and the 11 newborns whose mothers were unexposed. It should be noted that for this analysis of global changes between the populations, the requirement for correlation with increasing arsenic exposure was not imposed (Materials and Methods). This analysis identified 447 genes differentially expressed between the two populations of newborns, of which 404 (90%) were upregulated (Figure 2A; Table S3). Gene ontology enrichment analysis was performed to classify the genes modulated by prenatal arsenic exposure (Materials and Methods). This analysis identified ten gene ontology categories that were significantly enriched in the list of 447 genes (Table 2). Among the gene ontology categories that are significantly enriched are immune and inflammatory response (p < 0.001) (Table 2).
As an alternative approach to determine if groups of genes with common function are differentially expressed between the two newborn populations (arsenic exposed or unexposed), we have employed the knowledge-based Gene Set Enrichment Analysis (GSEA) (Materials and Methods). GSEA identified significant enrichment (false discovery rate [FDR] q-value < 0.01) of ten expression signatures with common biological function that are differentially expressed between the unexposed and exposed newborns. The groups of genes include three that represent stress-response signatures and three that represent tumor/cancer signatures (Table 3). The GSEA results also highlight that genes associated with estrogen receptor signaling are differentially expressed between the unexposed and exposed newborn populations (Table 3).
We next determined whether known molecular interactions exist among the proteins encoded by the arsenic modulated transcripts. Of the 447 arsenic modulated transcripts, 285 gene products were identified in the Ingenuity knowledge base and overlayed with known human molecular interactions (Materials and Methods). Among these proteins, we identified the presence of a large arsenic-modulated interacting network of proteins (Figure 2B). Specifically, we identified a large interacting network comprised of 105 human proteins encoded by arsenic-modulated transcripts (indicated as red and green nodes) (Figure 2B; Table S4). The probability of finding 105 arsenic-modulated transcripts that encode for a protein network of this size by chance is p < 10−55. Of the 105 proteins, 96 (91%) had transcripts that were upregulated in response to arsenic exposure.
Further analysis identified three highly significant (p < 10−55) sub-networks embedded within the large interacting network (Figure 3A–3C). The first sub-network centers around the nuclear transcription factor NF-κB and the pro-inflammatory interleukin 1 family member IL1-β (Figure 3A). This network integrates two members of the potential biomarkers; namely, SOC3 and CXCL1 (Figure 3A). Note that transcripts for all proteins directly associated with NF-κB in this sub-network are upregulated in infants born to arsenic-exposed mothers (Figure 3A).
The second sub-network integrates biomarker member DUSP1 with two stress-activated transcription factors; namely, signal transducer and activator of transcription (STAT1) and hypoxia inducible factor-1 α (HIF-1α) (Figure 3B). Transcripts for both STAT1 and HIF-1α were upregulated in infants with arsenic-exposed mothers (Figure 3B). STAT1 is involved in cytokine signal transduction and is known to be activated by arsenic [19]. HIF-1α activation and resultant tumorigenesis has been linked to chronic arsenic exposure [20].
The third sub-network integrates four of the 11 potential arsenic biomarkers; namely, EGR-1, OSM, PTGS2, and JUNB (Figure 3C). These arsenic biomarker gene products are highly integrated with proteins known to be involved in cell cycle regulation, including JUN and FOS, as well as stress-response proteins such as interleukin-8 (IL-8) (Figure 3C). An overlay of molecular processes represented in this sub-network highlights the finding that prenatal arsenic exposure modulates numerous biological processes including stress response, signal transduction, cell adhesion, and transcription (Figure 3C).
Using network analyses, we also established that there are known molecular interactions among the 11 potential arsenic biomarker genes. Eight of the 11 biomarker gene products (exclusive of SFRS5, MIRN21, and RNF149) are highly integrated with tumor necrosis factor-α (TNF-α), another proinflammatory cytokine (Figure 3D). TNF-α is involved in the control of both cell proliferation and apoptosis [21]. Here, we identify TNF-α activation in newborn cord blood upon exposure to prenatal arsenic.
In an effort to uncover potential regulatory mechanisms underlying the transcription of the arsenic-modulated gene sets, we performed transcription factor binding site analysis within the promoters of the arsenic-modulated genes (Materials and Methods). Promoter region comparisons for the arsenic-modulated genes identified significant enrichment (p < 0.05) for two transcription factor binding sites across all four gene sets. Specifically, binding sites for NF-κB and serum response factor (SRF) are enriched in all four arsenic-modulated gene sets (Table 4). Moreover, metal response element binding sites (MREs) for the metal-responsive transcription factor-1 (MTF1) are enriched in three of the four gene sets (sets 1, 3, and 4) (Table 4). The MTF1 binding site enrichment was highest for the third gene set with five of the 11 genes containing the MRE element (Figure 3D). Notably, the enrichment for MTF1 in the second gene set only narrowly misses the enrichment p < 0.05 cutoff, at p = 0.054 (Table 4). MTF1 was shown to be activated upon arsenic exposure in animal models [23,24]. It is noteworthy that gene targets for a known arsenic-inducible transcription factor are found among the transcripts modulated in the cord blood of infants born to arsenic exposed mothers.
As the unexposed samples utilized in this study were obtained from two different locations and could confound expression testing, we have used an alternative approach to substantiate the identified arsenic-induced pathways. Differential expression testing was performed between the cord blood of exposed and unexposed newborns from Ron Pibul (Materials and Methods). These analyses identified 321 genes that were differentially expressed between the arsenic-unexposed and -exposed newborns (Table S5). Notably, a direct comparison of gene expression changes identified considerable overlap between the transcripts differentially expressed between the newborns from Ron Pibul and transcripts differentially expressed across the whole population (fourth gene set) (Table S5).
To identify the biological pathways modulated by prenatal arsenic exposure, the proteins encoded by the 321 transcripts were analyzed for significant enrichment of molecular networks (Materials and Methods). Three highly significant protein sub-networks (p < 10−30) were identified (Figure S3). As with the network findings from the entire population of newborns, the networks identified here integrate proteins known to be involved in cell cycle regulation including JUN, as well as stress-response proteins such as interleukin-8 (IL-8), the pro-inflammatory interleukin 1 family member IL1-β, and hypoxia inducible factor-1 α (HIF-1α) (Figure S3). Furthermore, the NF-κB protein is integrated into the sub-networks and found to be activated in the cord blood of newborns exposed to arsenic within the Ron Pibul population (Figure S3).
Finally, our analyses included comparisons of the gene expression changes identified in this study with arsenic-induced gene expression changes reported in the literature in mouse models as well as a separate arsenic-exposed human population. Our results were compared with (i) expression changes in livers of mice treated with arsenic [24], (ii) expression changes identified in arsenic-induced tumors resulting from in utero exposures to arsenic in mice [6], and (iii) expression changes in blood from a human population from Taiwan exposed to arsenic [25]. These comparisons identify overlap of similarly modulated transcripts in response to arsenic exposure that include: BCL6 (B-cell CLL/lymphoma 6), CD14 (CD14 antigen), CXCL1 (chemokine ligand 1), EGR1 (early growth response 1), FOS (v-fos FBJ murine osteosarcoma), FOSB (FBJ murine osteosarcoma viral oncogene homolog B), GADD45B (growth arrest and DNA damage inducible beta), IFNGR1 (interferon gamma receptor 1), IL1B (interleukin 1 beta), IL1R1 (interleukin 1 receptor 1), JUN (v-jun sarcoma virus oncogene), MAPK6 (mitogen-activated protein kinase 6), MT1X (metallothionein 1X), RAD23B (RAD23 homolog B), and TOP1 (topoisomerase DNA 1) (Tables S3 and S5). These results highlight the modulation of stress related transcripts in both mice (acute and in utero exposures) and a separate adult human population in response to arsenic exposure.
Globally, millions of people are at risk for the detrimental effects of chronic arsenic exposure with drinking water levels far exceeding the WHO guideline [1]. Prenatal arsenic exposure in human populations has been associated with pronounced long-term health consequences [4]. Here, we address the impact of maternal arsenic exposure on fetal gene expression in a human population. Our goals were 2-fold: first, to establish the extent to which chronic arsenic exposure in mothers impacts newborn gene expression, and second, to identify genes that could be used as potential biomarkers of prenatal arsenic exposure and targets for remedial therapy.
Differential expression testing of training populations of newborns whose mothers had varied exposures to arsenic identified three arsenic-associated gene expression signatures comprised of 170, 38, and 11 genes. Analysis of the predictive capacity of each of these gene sets using the Support Vector Machine two-class prediction algorithm showed that each of these gene sets is highly predictive of arsenic exposure in a test population. Notably, even the smallest gene set comprised of 11 genes was powerful, with 83% accuracy in predicting prenatal arsenic exposure in the test population. The 11 potential biomarkers of prenatal arsenic exposure include CXL1, DUSP1, EGR-1, IER2, JUNB, MIRN21, OSM, PTGS2, RNF149, SFRS5, and SOC3. The set of 11 genes show a striking dose response to prenatal arsenic exposure. Stress response and cell cycle regulation are associated molecular functions of the potential biomarker set. Arsenic exposure is known to activate stress-related transcripts in yeast, animal models and human subjects [24–26]. Here, we find that stress-response genes are differentially expressed among a population of newborns whose mothers were exposed to varying levels of arsenic.
To assess the genome-wide impact of prenatal arsenic exposure on newborn gene expression, we identified all transcripts that showed differential expression between two populations; the 21 newborns whose mothers had been exposed to arsenic versus the 11 newborns whose mothers were unexposed. These analyses identified ∼450 genes differentially expressed between the two populations, of which 90% had expression levels that were increased (rather than decreased) by arsenic exposure. Clearly, there is a robust genome-wide response to prenatal arsenic exposure with ∼3% of the expressed genes significantly altered in the newborn. Gene ontology and GSEA highlight the activation of stress-related transcripts in the cord blood of infants exposed prenatally to arsenic.
Furthermore, integration of the gene products of the ∼450 transcripts with known molecular interactions identified the existence of a large arsenic-modulated interacting network of 105 proteins. Embedded within this large interacting network are three sub-networks that highlight that prenatal arsenic exposure activates inflammation-related molecules. Specifically, the first of the sub-networks centers around NF-κB and IL1-β. NF-κB regulates a large number of genes critical for apoptosis, as well as inflammation-related molecules such as cytokines (interleukins). IL1-β belongs to the class of acute phase proteins known to be increased in response to inflammation. Links between prenatal arsenic exposure and the activation of a stress response are also evident in the second and third sub-networks. Prenatal arsenic exposure resulted in the induction of the stress-related transcription factors STAT1 and HIF-1α, both of which are known to be activated by arsenic in model systems [19]. Here, we identify STAT1 and HIF-1α activation in newborn cord blood upon prenatal arsenic exposure. The activation of stress-response proteins such as interleukin-8 (IL-8) in response to prenatal arsenic exposure is also evident in sub-network three. The gene expression signatures identified here as modulated by prenatal arsenic exposure were compared to arsenic-induced gene expression changes in the mouse model and also with a separate human population. These comparisons highlight the common pattern of activation of stress-related transcripts in response to arsenic exposure.
Additionally, eight of the 11 biomarker gene products were found to have significant interactions with the proinflammatory cytokine TNF-α. Several studies in animal models have shown that arsenic exposure results in TNF-α stimulation [27–29]. In this study, TNF-α activation is identified in newborn cord blood upon prenatal arsenic exposure. Taken together, the network findings underscore that a mother's arsenic exposure results in a robust response in the fetus, indicative of a systemic inflammatory response along with the modulation of numerous other biological processes including apoptosis, signal transduction, cell adhesion, and transcription.
We further show that the extensive genome-wide newborn response to prenatal arsenic exposure may be regulated by at least three transcription factors. Analysis of the promoter regions of the arsenic-modulated genes showed enrichment for NF-κB and SRF in all four arsenic-modulated gene sets. SRF transcriptionally activates the expression of immediate early response genes, including C-FOS and EGR-1 [30], two members of the potential arsenic biomarker set. Moreover, binding sites for the metal-responsive transcription factor-1 (MTF1) are enriched in three of the four gene sets (sets 1, 3, and 4). MTF1 was shown to be activated upon arsenic exposure in animal models [23,24]. That gene targets for a known arsenic-inducible transcription factor are found among the transcripts modulated in the cord blood of infants born to arsenic exposed mothers supports our conclusions that the transcriptional changes reported here are likely due to prenatal arsenic exposure.
Our findings clearly demonstrate the robust impact of a mother's arsenic consumption on gene expression in utero as evidenced by transcript levels in the newborn's cord blood. More specifically, our data suggest that prenatal arsenic exposure acts as an inflammatory stimulus that activates the NF-κB signaling cascade. NF-κB activation plays a critical role in inflammation-driven tumor progression [31], and thus key players in tumor progression are modulated in the blood of newborns exposed to arsenic. To determine the extent to which these exposures and the resultant expression changes are associated with susceptibility to disease in later life, the health status of these children is currently being followed.
In summary, class prediction algorithms identified gene expression signatures that predict arsenic exposure in a test population with about 80% accuracy. Notably, by integrating training populations with varied exposures, a highly predictive potential biomarker gene set composed of just 11 genes was identified. These genes are promising as genetic biomarkers for prenatal arsenic exposure. Currently, we cannot eliminate the possibility that the gene expression signatures identified here are not absolutely specific for arsenic; they may also be predictive of other environmental exposures, e.g., exposure to other heavy metals. Nevertheless, this study underscores that there is a robust prenatal response that correlates with arsenic-exposure levels that could modulate numerous biological pathways including apoptosis, cell signaling, the inflammatory response, and other stress responses, and ultimately affect health status. Arsenic contamination of the drinking water in the Ron Pibul area of Thailand is representative of that seen in many other areas of South East Asia, most notably Bangladesh [9], suggesting that prenatal exposures are likely to be endemic in these areas. Moreover, arsenic contamination of the Ron Pibul drinking water is roughly the same as that known to be present in many of the western United States [2,9], suggesting that prenatal arsenic exposure may also be a problem in the United States. These data contribute to our understanding of biological responses upon arsenic exposure, and show that prenatal exposure in humans results in measurable phenotypic responses in the newborn.
The study was conducted in Bangkok and the Ron Pibul District of the Nakhon Sri Thammarat Province located in the southern peninsula of Thailand (Figure S1). Five villages in the Ron Pibul district were selected for the study location as they had been classified as high level arsenic contaminated areas, and arsenicosis had been reported there [8]. Arsenicosis has not been reported in Central Thailand, specifically Bangkok, where arsenic concentrations in water and soil have been determined to be very low [8]. The study subjects consisted of 32 pregnant women (20–40 y old). All subjects were healthy, pregnant volunteers undergoing vaginal childbirth without birth stimulation or anesthesia. Twenty-three pregnant women living in the Ron Pibul District and nine women living in Bangkok for at least 1 y were recruited for the study. Women from both sites were age, educational level, and socioeconomically matched. Questionnaires were administered to all participants to obtain personal information regarding residential history, health history and potential confounding factors, birth and pregnancy information (number of births, abortions or complications), use of community drinking water and well water, plus water and food consumption habits. Cord blood samples were collected from January 2004 to December 2005 in the Ron Pibul Hospital (Ron Pibul District) and the Rajvithi Hospital (Bangkok). This study was conducted according to the recommendations of the Declaration of Helsinki (World Medical Association 1989) for international health research. All subjects gave written informed consent to participate in this study.
Pregnant participants were asked to provide toenail samples during pregnancy for analysis of total arsenic concentration, which was determined by Inductively Coupled Plasma-Mass Spectrometry (ICP-MS) (Agilent 7500c). After delivery, 2.5 ml of newborn cord blood was collected into a PAXgene Blood RNA (Qiagen) tube for study of gene expression. All cord blood samples were kept at −70 °C until analysis.
Total RNA was isolated from 32 cord blood samples according to the PAX gene protocol and Qiagen RNA extraction kit. RNA was labeled using a globin reduction protocol (Affymetrix) and hybridized to HGU133 Plus 2.0 full genome human arrays in technical duplicate for a total of 64 arrays. Data were first normalized using Robust Multi-Chip Average (RMA) [32] and filtered for expressed transcripts across all arrays (+2 standard deviations above mean background) resulting in reduction of the probesets from the original 54,675 to 15,265. A mean absolute expression value was calculated from technical duplicates of the arrays for all expressed transcripts. Differential gene expression and association with increasing arsenic concentration was calculated as follows. The samples comprising the training sets were separated into two groups based on arsenic exposure level. The two groups were unexposed (maternal toenail <0.5 μg/g) or exposed (maternal toenail ≥0.5 μg/g). The two-class exposure designation is based on the WHO standards for exposure to arsenic of 10 μg/l arsenic. A mean toenail arsenic concentration of 0.5 μg/g corresponding to chronic consumption of drinking water at 10 μg/l arsenic was derived from two studies associating arsenic toenail concentration and drinking water in a population from Bangladesh [12] and the United States [11]. Differential expression was determined as a significant difference in the expression of a gene (exposed versus unexposed) where the average fold change was greater than +/−1.5 and p < 0.05 (t-test). Additionally, significant association of gene expression and increasing arsenic level was determined by correlation measurements (r2 ≥ +0.6, r2 ≤ −0.6; p < 0.01) calculated using the linear regression model in S-PLUS 7.0 (http://www.insightful.com). The two-class prediction model used for assessing arsenic exposure in test populations was Support Vector Machine, carried out in Gene Pattern Software (version 2.0.1) (http://www.broad.mit.edu). Multivariate analysis was performed as follows: the expression values (Y) for each gene were modeled using Y = β1 + β2 ars (arsenic) + β3 loc (geographic location), where toenail arsenic concentration is a continuous variable and location is binary. Statistical significance was determined by subjecting β2 and β3 to t-statistics. A χ2 test for dependence (association) of the two factors (e.g., arsenic and geographic location) was performed for the set of arsenic biomarkers. A Fisher's exact test was employed to determine overrepresentation of the biomarkers within the genes significantly associated with either geographic source or arsenic exposure (p < 0.01). Network analyses were performed using the Ingenuity software (http://www.ingenuity.com). Gene ontology enrichment analysis was performed using GO Miner [33]. GSEA [34] was performed using the GSEA desktop software [35], with a false discovery rate correction (Benjamini-Hochberg) employed. Microarray data have been deposited to the Gene Expression Omnibus repository.
Transcription factor binding site analysis was performed using Expander software [36] and Genomatix software (http://www.genomatix.de). For both analyses, Affymetrix probesets were linked to sequence data for regions 1,000 base pairs upstream and 200 base pairs downstream of the transcription start sites, and these were analyzed for significant enrichment of transcription factor binding sites. Significance (p ≤ 0.05) was calculated where significance is the probability of obtaining an equal or greater number of sequences with a model match in a randomly drawn sample of the same size as the input sequence set.
Microarray data have been deposited to the National Center for Biotechnology Information (NCBI) Gene Expression Omnibus repository under Series Record GSE7967 (http://www.ncbi.nlm.nih.gov/geo/). |
10.1371/journal.pgen.1004186 | Within-Host Spatiotemporal Dynamics of Plant Virus Infection at the Cellular Level | A multicellular organism is not a monolayer of cells in a flask; it is a complex, spatially structured environment, offering both challenges and opportunities for viruses to thrive. Whereas virus infection dynamics at the host and within-cell levels have been documented, the intermediate between-cell level remains poorly understood. Here, we used flow cytometry to measure the infection status of thousands of individual cells in virus-infected plants. This approach allowed us to determine accurately the number of cells infected by two virus variants in the same host, over space and time as the virus colonizes the host. We found a low overall frequency of cellular infection (<0.3), and few cells were coinfected by both virus variants (<0.1). We then estimated the cellular contagion rate (R), the number of secondary infections per infected cell per day. R ranged from 2.43 to values not significantly different from zero, and generally decreased over time. Estimates of the cellular multiplicity of infection (MOI), the number of virions infecting a cell, were low (<1.5). Variance of virus-genotype frequencies increased strongly from leaf to cell levels, in agreement with a low MOI. Finally, there were leaf-dependent differences in the ease with which a leaf could be colonized, and the number of virions effectively colonizing a leaf. The modeling of infection patterns suggests that the aggregation of virus-infected cells plays a key role in limiting spread; matching the observation that cell-to-cell movement of plant viruses can result in patches of infection. Our results show that virus expansion at the between-cell level is restricted, probably due to the host environment and virus infection itself.
| A great deal is understood about how a virus infects an individual cell and manages to replicate. Patterns of disease progression in plant and animal hosts, such as virus titers and the appearance of symptoms, have also been described in great detail. On other hand, very little is known about what is happening at the intermediate levels during virus infection. Here, we use flow cytometry, a technique to rapidly measure large numbers of individual cells, to quantify the number of cells infected by a plant virus, in different leaves and at different times. We found that few cells become infected, and only one or two virus particles typically initiated cellular infection. Moreover, viruses from an infected cell will infect only one or two other cells. Therefore, although viruses replicate at astronomical rates within a cell, their rate of spread between individual cells can be much slower.
| For obligate intra-cellular micro-parasites such as viruses, the cell is the fundamental and minimal unit of infection. Important macro-scale phenomena in viral infection – immunity, virulence, transmission, and evolution – all depend on the infection outcome in individual cells. The biochemical and molecular bases of virus infection have received much scrutiny, and in the past decades there also have been major advances in understanding the dynamics at the host and host-population levels. The next great challenge is a unified picture of virus infection dynamics and evolution that integrates different spatiotemporal scales [1], [2]. However, integration across different spatiotemporal scales effectively has not occurred across the between-cell level due to practical and methodological considerations.
At present, there simply is not a coherent picture of infection dynamics at the between-cell level. A number of key issues have not been addressed adequately to date. First, virus replication in an individual cell can be extremely rapid [3], [4], as can the advance of infection and long-range movement [5]. However, little is known about the rate at which infection spreads at the cellular level [6]. What will be the number of newly infected cells per infected cell per day, a value we refer to as the cellular contagion rate (R)? Whereas a reproduction ratio estimates the number of cells directly infected by one cell [6], the contagion rate estimates the total number of newly infected cells occurring per infected cell over a given time period. For Tobacco mosaic virus (TMV) infection of Nicotiana benthamiana plants, a low R was estimated (0.5–0.6 cells/cell/d), although why this R value was so low was not discussed [7]. Given the rapid replication and spread of viruses, this result is unexpected and it is not at all clear whether other viruses will adhere to similar patterns. Furthermore, a constant R value was assumed in the analysis described in ref. [7], whereas a time-varying rate may provide more insights into the underlying dynamics [6]. Another important issue is that individual cells can be observed readily in cell culture systems, whereas gross infection patterns in multi-cellular hosts can be observed by means of virus-induced symptoms, molecular methods [8] or by monitoring infection of tagged viruses [5]. However, these methods do not render information on how the number of infected cells in different tissues changes over time. Finally, variation in genotype frequencies has been described only at higher levels of host organization [9]–[11]. By variation in genotype frequencies, we refer to the differences in the abundance of different virus variants, after a cohort of hosts is initially inoculated with a virus population containing two or more variants. How will this variation change from the population to the individual to the organ, and finally, to the cell? This variation is pivotal to studying the infection dynamics and evolution of viruses. Within-cell interactions between virus genotypes, such as recombination and the complementation of defective virus genotypes, will require that the presence of two genotypes within a host also carry over to the organ and individual cell levels. Whether genotypes carry over will depend on the genetic bottlenecks a virus population passes through when colonizing organs or infecting a cell, respectively.
Plant viruses are ideal model systems for studying virus infection at the between-cell level, and therefore infection dynamics at this level are probably best understood in these systems. The targets of primary infection by mechanical inoculation – epidermal cells – can be readily observed in situ [5], [12]–[14], allowing for the tracking of cell-to-cell movement [13]. Moreover, two approaches have been developed to determine whether protoplasts – intact cells extracted after degradation of the cell wall – are infected by different plant virus variants, based on fluorophores [7] or nested PCR [9]. Finally, there is an enviable characteristic of plants: their leaves are natural, biologically relevant compartments that can be removed cleanly (e.g. [15]) for further study.
The development of plant viruses as model systems to study between-cell infection dynamics has led to important insights and the estimation of some key infection parameters. First, as discussed above, a low R has been estimated for TMV [7]. Second, estimates of the cellular multiplicity of infection (MOI) have been made for three plant viruses. For TMV, MOI was found to be low (MOI<2) [7], [16]. Moreover, in this particular case a substantial proportion of cells (>0.1) remain uninfected [7]. However, a model-selection-based analysis of the TMV data suggests MOI might in fact be higher, whilst the number of coinfected cells is low due to spatial segregation of the two virus variants [17]. For Cauliflower mosaic virus (CaMV), MOI was reported to vary from 2 to 13 over time, and most cells were infected [9]. Furthermore, for CaMV virion concentrations in vascular tissue are correlated to MOI [18]. For Soil-borne wheat mosaic virus, MOI was estimated during the first rounds of cellular infection in the inoculated leaf, rendering an estimated of 5–6 [12]. Additionally, low level of potyvirus cellular coinfections suggest a low MOI for potyviruses [19]. Finally, for our model system, Tobacco etch virus (TEV; genus Potyvirus, family Potyviridae), the number of infected cells in systemic tissues early in infection depends on the number of primary infection foci, and the number of infected cells does not increase to a frequency greater than 0.5 [15].
Important omissions in our understanding of infection dynamics at the between-cell level remain, however. In particular, a comprehensive view of the between-cell level of infection is missing and the tracking of cell-level infection in multiple host organs or compartments has not been reported. We therefore opted to study these dynamics in TEV and devised an experimental setup in which we could measure infection at the cellular level, which was both sensitive and high-throughput. We opted to analyze the presence of viral variants in individual cells using a flow-cytometry-based method [15], [20]. This approach allows for quantitative measurements of the number of cellular infections for two virus variants in a large number of mesophyll cells, allowing for an analysis of infection dynamics in different host compartments and at different times. This large dataset allowed us to describe the dynamic pattern of the number of infected cells over time, estimate MOI, quantify R, and consider the variation in genotype frequencies at different levels of host organization as a consequence of bottlenecks.
We generated two TEV variants, TEV-BFP and TEV-Venus, which express blue or yellow fluorescent proteins, respectively. Fluorescent markers inserted in the TEV genome can be stable over multiple short rounds of infection [14], [21], and we confirmed the integrity of the marker sequences throughout the experiment (see Materials and Methods). Furthermore, the insertion of eGFP – a variant of the fluorescent protein from which BFP and Venus are derived – in the TEV genome has no effect on virus accumulation after 7 days post-inoculation (dpi) (see Materials and Methods). Therefore, these marked viruses have biological properties similar to the wild-type virus from which they are derived. We rub-inoculated the third true leaf of Nicotiana tabacum L. cv. Xanthi plants with a 1∶1 mixture of infectious saps (ground tissue in inoculation buffer) of the two variants. We then isolated protoplasts [15], [20] from the third, fifth, sixth, and seventh true leaves at 3, 5, 7, and 10 dpi, with five replicate plants for each time point. We did not analyze the fourth true leaf because under the current experimental conditions this leaf does not show any infection. Flow cytometry was used to determine which cells were uninfected, infected by one or by both virus variants. Using this approach we could quantitatively measure the distribution of cellular infection over space and time, for the two virus variants.
The frequency of virus-infected cells was low (mean ± 1 SD: 0.072±0.099), with the highest level of infection observed in any one sample being 0.424 (Leaf 7 at 10 dpi) (Figure 1A–D). The frequency of cells infected by both virus variants was also low (mean ± 1 SD: 0.012±0.023), with the highest level of coinfection observed in any sample being 0.112 (Leaf 6 at 7 dpi) (Figure 1A–D). These low levels of coinfection are in agreement with previous studies on plant RNA viruses [7], [13], [19], and suggest that MOI is low. Few cells were infected in any leaf at 3 dpi, with the greatest number of infections being found in Leaves 3 and 6. This surprising observation can be explained by the occurrence of limited, relatively slow TEV expansion at the macroscopic level in the inoculated leaf [8], combined with fast egress (<2 dpi) from Leaf 3 to Leaf 6 at high viral doses [15]. Both infection and coinfection appear to increase over time in the different leaves, although Leaf 5 shows very low levels of infection. Infection progresses slower in Leaf 3 than in Leaves 6 and 7. Leaf 6 becomes infected before Leaf 7, but the dynamics in these two leaves are otherwise very similar. The frequency of TEV-Venus infected cells was significantly higher than expected for a 1∶1 inoculum (one-sample t-test: t79 = 4.141, P<0.001), although the magnitude of the deviation was small (mean Laplace point estimator for the frequency of TEV-Venus infected cells ± 1 SD = 0.591±0.196). This deviation could occur because of a small discrepancy in the inoculum ratio, or a small difference in infectivity or in within-host competitive fitness of the two variants. To confirm that infection levels in Leaf 7 had saturated at 10 dpi, in a separate experiment we also analyzed infection in Leaf 7 at 13 dpi. The observed frequency of virus-infected cells was slightly lower than at 10 dpi, although the difference was not statistically significant (two-sample t-test: t8 = 1.251, P = 0.246). The data therefore suggest that infection levels had saturated in all analyzed leaves by 10 dpi.
To visually illustrate patterns of infection, we infected plants with TEV-eGFP and TEV-mCherry [14] under identical conditions. These viruses were used here, instead of TEV-BFP and TEV-Venus, because their fluorescent proteins are more suitable for microscopy. Even when infection appears to have saturated at both the cell and visible fluorescence level, we could see heterogeneities in the distribution of virus infection over the leaf at different spatial levels (Figure 1E–G).
We estimated the time-varying cellular contagion rate (R) from the data using a simple maximum likelihood method. This analysis was carried out on the total number of infected cells, regardless the virus variants present. For R>0 the number of infected cells increases, whereas for R<0 it decreases. Our estimates of R for individual leaves (Figure 2A–D) ranged from 2.43 cells/cell/d (95% CI: 1.80–3.39) (Leaf 6, 3 dpi;) to values not significantly different from zero (e.g., −0.327 cells/cell/d (95% CI: −0.539–0.271) for Leaf 5, 7 dpi). We do not expect R<0 in this system, since infection is not cleared and the number of infected cells can therefore not decrease. Our approach might slightly overestimate R in individual leaves because of between-leaf transmission, and we therefore also estimated R for pooled data from different leaves (Figure 2E). One disadvantage of this approach is that tissues with high infection levels will most strongly affect R estimates. These estimates of R (mean [95% CI]) ranged from 1.342 cells/cell/d [0.247–1.371], 3 dpi, to 0.196 cells/cell/d [0.041–0.244], 7 dpi, and were always significantly greater than zero. Overall, values of R appear to be surprisingly low given estimates of the rapid rate of cell-to-cell movement for TEV during initial infection, whilst they are similar to estimates of R for TMV (0.5–0.6 cells/cell/d) [7]. Low R values may therefore be commonplace in plant RNA viruses, although data from more pathosystems will be needed to confirm this idea.
Dolja et al. [5] observed that a primary infection focus starts with a single infected cell and grows to formation with a diameter of ten infected cells within 24 h, and hence cells/cell/d. This calculation is conservative and underestimates R because infection in the first infected cell cannot be observed at t = 0, and because it only takes into account infection in the epidermal cells. Note that such a high value – which probably far exceeds the number of other cells to which each cell is plasmodesmally connected [7] – is possible because of multiple rounds of replication can occur within a single day [5]. The R values we have measured are therefore extremely low compared to R values found in the inoculated leaf during early infection.
We wanted to test whether our understanding of the process that is likely to govern cell-level infection patterns was congruent with our empirical data. Specifically, we wanted to test whether there were leaf-dependent differences in key infection parameters, and whether there was evidence for aggregation of virus-infected cells limiting infection spread. We therefore developed a simple susceptible-infectious (SI) model of within-host infection dynamics. Each leaf in a plant represents a physically separated compartment - with its own physiological state - that a virus must colonize [22]. We therefore developed a simple meta-population dynamics model with between-leaf transmission from lower leaves to upper leaves. For the kth leaf, the rate of change of the fraction of infected cells (Ik) is:(1)where β is the within-leaf transmission coefficient (from cell to cell), χ is the between-leaf transmission coefficient and S is the fraction of susceptible cells. Between-leaf transmission depends on the total fraction of infected cells in the leaves below the kth leaf, given that systemic-movement for phloem-transported viruses is towards the apical sink leaves [5], [22]. Potyvirus infection appears to be marked by the aggregation of infected cells [19], and given that plant cells will largely retain their respective positions in developed leaves, the perfect mixing assumptions of the SI model will not be met. We therefore included a spatial aggregation factor of infectious units (i.e., infected cells) ψk in the model, such thatBy spatial aggregation of infected cells, we mean that infected cells are likely to be found together and are therefore not randomly distributed in the leaf. The mechanism resulting in the spatial aggregation of infected cells is probably the dependence of plant viruses on cell-to-cell movement for local infection to spread: the spread of virions, or in some cases unencapsidated genomes, from an infected cell to its direct neighbors [5], [13]. When ψk = 1 there is perfect mixing, whereas when ψk approaches 0 there is maximum aggregation of infected cells [23], [24]. The model was fitted using maximum likelihood methods, and model selection was performed to ensure the data supported the inclusion of all model parameters (see Materials and Methods). As with the estimates of R, this analysis was carried out on the total number of infected cells and does not distinguish between the two virus variants.
The SI meta-population model could describe the data well, clearly capturing the main trends in the data (Figure 2F). Spatial aggregation of infected cells (ψk) was indispensable to the model (Table S2), and parameter estimates varied over leaves; ψk was most pronounced in Leaves 3 and 5, and much lower in Leaves 6 and 7 (Figure 2G). The between-leaf transmission coefficients (χk) for Leaves 5 and 6 were similar, although infection never reaches even moderate levels in Leaf 5. χ7 was significantly lower than χ6 (non-overlapping 95% CIs of parameter estimates), although the number of infected cells in both leaves reached moderate levels eventually. Parameter estimates therefore suggest that infection dynamics vary for each leaf, even though the overall pattern (Figure 1A–B) is similar for Leaves 6 and 7.
The cellular MOI can be estimated from our data, as has been previously done for two plant viruses with a similar experimental setup [7], [9]. However, estimates of MOI can be influenced by the estimation method [17]. Model selection was therefore performed on a set of nine MOI-predicting models (see Materials and Methods), by testing which Poisson-based model best predicted the relationship between the fractions of uninfected and coinfected cells (i.e., those cells infected by both virus variants) [17]. The models incorporated spatial segregation of virus genotypes, spatial aggregation of infected cells, superinfection exclusion at the cellular level and combinations of all these effects. We could thereby identify the best model to generate MOI estimates (Tables S3 and S4). The best-supported model incorporated only a leaf-dependent aggregation factor ψ (Table S4). The MOI and SI model selection results are therefore in good agreement with each other, although estimated ψ values were higher than those obtained from the SI model (Figure 2G), indicating less aggregation (Figure 3A). These two separate model selection procedures therefore confirm the importance of the spatial aggregation of infected cells for understanding TEV infection dynamics at the between-cell level, as might be expected for a virus that spreads by cell-to-cell movement. On the other hand, in a similar model-selection-based analysis for TMV and CaMV MOI, two viruses that also move by cell-to-cell movement, spatial aggregation only marginally improved model fit for both datasets [17]. These two different model-selection results suggest that whether cell-to-cell movement really has an impact on MOI estimation will depend not only on the mechanism of movement. Other factors, such as the number and distribution of initially infected cells, and the frequency of infected cells, also may play an important role.
We then derived predictions of MOI using the best-supported model (Figure 3B). As could be expected from the low frequencies of cellular infection and coinfection (Figure 1A–D), the predicted MOIs were low, ranging from 1.001 (Leaf 5, 3 dpi) to 1.432 (Leaf 6, 7 dpi). Note that we report the estimated MOI value in infected cells only (i.e., mI in Materials and Methods), which has a minimum value of 1. The corresponding range of MOI values calculated over the whole population of infected and uninfected cells (mT) is 0.002 (Leaf 5, 3 dpi) to 0.735 (Leaf 6, 7 dpi). Although these estimates may seem low intuitively, MOI is assumed to follow a Poisson distribution over cells and some cells can still be infected by two or more virions, even when the mean of the distribution is low (Figure 3C–E). Our estimates of MOI are similar to the low estimates for TMV [7], [16], although model-selection-based estimates for the TMV data result in MOI values ranging to somewhat higher values (up to 2.1), due to the predicted occurrence of spatial segregation of virus genotypes [17]. For CaMV much higher MOI values were observed later in infection [9], but in our system infection levels remain low even then.
The experimental data also allow us to consider variation in the frequencies of viral genotypes at different levels of the host: leaf (Figure 4A–D), cells coinfected by both virus variants (Figure 4E–H), all infected cells (Figure 4I–L), but also at the level of the host-plant population (Figure 4M). Variance of TEV-Venus frequencies appears to increase strongly from the plant and leaf levels to the individual cell level (Figure 4A–M). The log-transformed genotype ratios (TEV-Venus∶TEV-BFP) in individual cells appear to be independent of the frequency of TEV-Venus in the leaf (Figure 5A), indicating a decoupling of processes occurring at the leaf and coinfected-cell levels. Low estimates of MOI (Figure 3B) imply that the virus population entering each cell is subject to a narrow genetic bottleneck. A decoupling of the infection processes at the leaf and cell levels is predicted to occur because very few cells are infected by more than 2 virions (Figure 3E). Hence, for the vast majority of coinfected cells the frequency of virus variants, as represented by the infecting virions, is limited to 1/3, 1/2 and 2/3. If our MOI estimates are correct, than stochasticity in the replication process within the cell accounts for high levels of variation. In line with these expectations, we observed high levels of variation in virus variants at the cellular level (Figure 4E–H) and a distribution of variants in coinfected cells that is independent of the frequency of virus variants in the leaf (Figure 5A). Note that there are couplings between the leaf and cell-level dynamics (i.e., MOI depends on the overall level of infection for the best-supported MOI models; see Materials and Methods), but our observations show that not all leaf-level characteristics of the virus population carry over to individual cells.
Finally, we estimated the effective population size, Ne, for individual leaves and the whole plant [25] (see Materials and Methods). For the inoculated leaf we obtained a Ne estimate of approximately 100 (Figure 5B), corresponding well to the approximate number of primary infection foci observed. For Leaf 6, Ne was also estimated to be approximately 100, although the confidence interval extends to ∞ and there is no evidence for a genetic bottleneck in this leaf. For Leaves 5 and 7, much lower estimates of Ne were obtained, suggesting that fewer virions infect these leaves and that it is more difficult for the virus to invade these compartments. A wide range of within-host effective population sizes at the leaf level has been reported for different viruses [10], [11], [18], [26]. Here we show a similar range of effective population sizes can occur with a single virus-host combination, probably due to the combined effects of host physiology, anatomy and immunity.
To link infection dynamics at the cell and host levels, we have measured the number of cells infected by two virus variants within individual plants over time and space. We have estimated R (the cellular contagion rate, expressed as newly infected cells per infected cell per day) over time for systemic virus infection. A conservative estimate of the maximum value for R is 1.4 cells/cell/d on day 3 (Figure 2E), and it falls to just under 0.2 cells/cell/d by day seven. These values are comparable to estimates of R for TMV infection of N. benthamiana of 0.5–0.6 cells/cell/day, although in this instance a constant R was estimated [7]. We can therefore conclude that for our model system, and perhaps more generally for plant RNA viruses, R is very low during systemic infection, suggesting that most cells will transmit virus to one or possibly even zero other cells during infection.
Here we have estimated the cellular contagion rate over a period of one day. Given that TEV infection has been reported to expand at a rate of one row of cells every 2 h [5], it is entirely possible that multiple rounds of infection will occur during one day. Therefore, the reproduction ratio at the cellular level (i.e., the number of cells to which one infected cell spreads infection over its lifetime) is probably similar to, or even lower than our estimates for the cellular contagion rate. These estimates are in principle the aggregated effect of local cell-to-cell movement and long-range systemic movement. What then accounts for these surprisingly low values of R, and are they reconcilable with high replication rates at the molecular level [3], [4] and fast virus expansion throughout the plant [5], [15]?
Decreases in cellular replication because the carrying capacity for infection has been reached do not explain these observations: low R values were estimated when infection levels were very low (e.g., compare Figures 2C and 2F). However, contagion rates at the cellular level can be much higher than those we have observed here: based on other results [5] we also estimate that during expansion in primary infection foci R≈78 cells/cell/d. We have observed early infection in systemically infected leaves that eventually reached high levels (i.e., Leaves 6 and 7), and especially in the case of Leaf 7 these infections appears to be initiated by a small number of virions. Hence, ceteris paribus we would have expected high R levels in these leaves as well, and moreover in Leaf 7 R does not reach the same levels as Leaf 6. These observations implicate two processes in slowing the observed rate of virus expansion at the between-cell level. First, host immune responses, particularly RNA silencing [27], is very likely to play a role. Moreover, since a specific RNA silencing signal progresses systemically to sink leaves [27], [28], we speculate that this may explain why there appear to be lower R levels in Leaf 7 than in Leaf 6. Second, our experimental approach limits us to analyzing the cells in a leaf as a whole, whereas the analysis of cells in the infection front would result in higher R values.
We found striking differences in infection dynamics in different leaves (Figure 1A–D). These differences were also reflected in estimates of parameters for the different models fitted to the data (Figures 2B, 3A and 5B). What can account for the infection dynamics in different leaves? First, sink-source transitions will play a major role in determining if and to what extent leaves can be colonized, because phloem-transported viruses cannot cross the sink-source boundary in any leaf [22]. This functional boundary separates the basal part of a developing leaf, which is importing photo assimilates, from the distal part that is already exporting them. Furthermore, sink-source transitions may further impact the spatial aggregation of infected cells on a smaller spatial level: sink-source transitions will determine from which classes of phloem the virus can unload, with much less restriction in smaller veins prior to the transition [22]. Hence the distribution of initially infected cells is likely to be more homogeneous – also on small spatial scale – in sink leaves, leading to less spatial aggregation of infected cells. We saw infection only in the basipetal region of Leaf 5, whereas about half of the surface of Leaf 6 became infected (Figure 1E). Therefore, we think that Leaf 4 has probably completed the sink-source transition, and is almost exclusively exporting photo assimilates, whereas it has not affected much of Leaf 7. These assertions on the physiological state of these different leafs are strongly supported by measurements of polyamine levels [29], which are molecular markers of proliferating source tissues. Putrescine and spermidine levels show that for N. tabacum cv. Xanthi of the same development stage as our plants at inoculation, the sink-source transition is virtually complete in Leaf 4, almost complete in Leaf 5, and has not yet commenced in Leaf 6. Note that whereas sink-source transitions probably account for virus aggregation on a large and intermediate scale (Figure 1E–1F), RNA silencing probably impedes infection at all scales, also resulting in aggregation of infected cells on the smallest scales (Figure 1G) [27]. Second, crossing from leaves at one side of the plant to the opposite can be hindered by the phloem connections between leaves [22]. Third, as aggregation of infected cells is increased, the rate of virus spread decreases [23] and the plant will have more time to mount an effective response [27]. Consequently, we hypothesize that large effective population sizes can only be achieved if (i) the virus can be readily transmitted between two particular leaves, and (ii) the subsequent aggregation of infected cells is moderate to low (e.g., the virus is not restricted to the basal part of the leaf by the sink-source transition), allowing infection to expand beyond the initial point of entry.
Based on these other studies, we therefore speculate on what processes can account for the leaf-dependent differences we have observed. Infection progresses relatively slow in Leaf 3, probably because under the conditions used the virus only expands locally and egresses from this source leaf [8]. Leaf 4 probably never shows any infection in our setup because it has completed the sink-source transition, and is moreover located opposite Leaf 3 (for leaf positions see Figure 1). Leaf 5 has a relatively high between-leaf transmission, strong aggregation, and a small bottleneck size (Figures 2, 3 and 5). Its position directly above the inoculated leaf explains high between-leaves transmission, whilst the nearly complete sink-source transition results in high aggregation, low levels of infection and therefore a de facto genetic bottleneck. In line with this explanation, the highest levels of aggregation were observed in Leaf 5, suggesting the virus expansion is very constrained in this leaf. Leaf 6 has a high between-leaf transmission due to its position above the inoculated leaf. Moreover, because the sink-source transition is far from complete there are high levels of infection, moderate aggregation and no genetic bottleneck. Finally, Leaf 7 is positioned on the far side of the plant, with respect to the inoculated leaf, and the increasing intensity of host immune responses results in low between-leaf transmission, and hence a genetic bottleneck occurs. However, since the sink-source transition is far from complete, those viruses that do enter the leaf can expand prolifically, resulting in lower estimated levels of aggregation and high infection levels. In summary, we think that plant anatomy and physiology may largely explain the leaf-dependent differences in infection patterns we have observed, although our explanation will require further testing.
Our analyses of infection spread and MOI support the idea that aggregation of virus-infected cells is also important for understanding dynamical patterns and therefore low R values. If there is aggregation of virus-infected cells, which is concurrent with potyviruses achieving local spread by cell-to-cell movement, only those cells on the edge of an aggregate can contribute to virus expansion, and even fortuitously situated cells may not actually infect those susceptible cells they are in contact with before neighboring cells do. The limitations on virus spread from an individual cell to its neighboring cells due to the overall rapid spatial spread of the virus is an effect we refer to as “self-shading”. The importance of self-shading in limiting between-hosts spread [23], [30], and its implications for virulence evolution [31], [32], have been recognized on larger spatial scales. Our results stress the importance of extending these concepts to within-host dynamics, although we anticipate that there will be differences in the between-host and within-host levels. For example, we hypothesize that a high cellular contagion rate may not incur a major cost in our model system; host cells are static and once a tissue has been infected there are no possibilities for further within-host spread, except for phloem loading in a minority of cells. Therefore, we speculate that aggregation and self-shading will, in this case, impose selection for fast viral replication and spread at the within-host level.
Our experimental approach consisted of the isolation of protoplasts, followed by measurements on individual cells by flow-cytometry. Advantages of this approach are its amenability to high-throughput, the high sensitivity of the flow-cytometer, and the fact that mesophyll cells – the primary targets of virus replication – can be analyzed. Disadvantages are the fact that sampling is destructive, and hence a time course cannot be analyzed, and the spatial information is lost during protoplast extraction. Compared to other techniques available for analyzing protoplasts [7], [9], the approach used here has a much higher throughput. Our approach may have a higher sensitivity than microscopy [7], although PCR-based methods are probably more sensitive [9]. Another alternative approach to analyze virus infection dynamics would have been microscopy on whole leaves, which renders spatial information and allows for longitudinal analyses [13]. Although this approach works very well in the inoculated leaf [13], it is not clear how well it would function in systemic leaves, and this is also a lower-throughput method. For a comprehensive analysis such as we have presented, the high-throughput nature of the assay is essential and dictated our choice of experimental approach.
For many other virus-host pathosystems, including those that result in disease in animals such as humans, important spatial characteristics of virus-plant pathosystems may be absent. Short-range virus infections can typically be achieved by diffusion of virions instead of cell-to-cell movement, and most host organs will not have the planar anatomy of leaves. However, there are general characteristics of virus-host interactions that suggest infection aggregation may be a very commonplace phenomenon. First, there are many physical barriers to virus expansion, structuring the host environment and naturally favoring aggregation of infected cells. Second, many viruses replicate in a limited number of cell types or tissues, thus leading to spatial aggregation. Third, epithelia are often targets of viral entry and one of the sites of replication, and consist of highly planar structures. Finally, even for free virions, diffusion and virion removal rates will determine at what distance infection tends to spread. Based on our results and these general considerations, we therefore speculate that aggregation of virus-infected cells and self-shading are likely to be key ingredients for cell-level infection dynamics in a broad range of intra-cellular pathogens infecting complex, multi-cellular hosts.
|
10.1371/journal.ppat.1002999 | Glycoprotein N of Human Cytomegalovirus Protects the Virus from Neutralizing Antibodies | Herpes viruses persist in the infected host and are transmitted between hosts in the presence of a fully functional humoral immune response, suggesting that they can evade neutralization by antiviral antibodies. Human cytomegalovirus (HCMV) encodes a number of polymorphic highly glycosylated virion glycoproteins (g), including the essential envelope glycoprotein, gN. We have tested the hypothesis that glycosylation of gN contributes to resistance of the virus to neutralizing antibodies. Recombinant viruses carrying deletions in serine/threonine rich sequences within the glycosylated surface domain of gN were constructed in the genetic background of HCMV strain AD169. The deletions had no influence on the formation of the gM/gN complex and in vitro replication of the respective viruses compared to the parent virus. The gN-truncated viruses were significantly more susceptible to neutralization by a gN-specific monoclonal antibody and in addition by a number of gB- and gH-specific monoclonal antibodies. Sera from individuals previously infected with HCMV also more efficiently neutralized gN-truncated viruses. Immunization of mice with viruses that expressed the truncated forms of gN resulted in significantly higher serum neutralizing antibody titers against the homologous strain that was accompanied by increased antibody titers against known neutralizing epitopes on gB and gH. Importantly, neutralization activity of sera from animals immunized with gN-truncated virus did not exhibit enhanced neutralizing activity against the parental wild type virus carrying the fully glycosylated wild type gN. Our results indicate that the extensive glycosylation of gN could represent a potentially important mechanism by which HCMV neutralization by a number of different antibody reactivities can be inhibited.
| Herpes viruses are transmitted between individuals in cell free form and successful spread benefits from mechanisms that limit the loss of infectivity by the activity of virus neutralizing antibodies. Human cytomegalovirus (HCMV) is an important pathogen and understanding how the virus can evade antiviral antibodies may be clinically relevant. HCMV particles contain a number of highly polymorphic, extensively glycosylated envelope proteins, one of which is glycoprotein N (gN). This protein is essential for replication of HCMV. We have hypothesized that the extensive glycosylation of gN may serve as a tool to evade neutralization by antiviral antibodies. Recombinant viruses were generated expressing gN proteins with reduced glycan modification. The loss of glycan modification had no detectable influence on the in vitro replication of the respective viruses. However, the recombinant viruses containing under-glycosylated forms of gN were significantly more susceptible to neutralization by a diverse array of antibody reactivities. Immunization of mice with viruses carrying fewer glycan modification induced significantly higher antibody titers against the homologous virus; however, the neutralization titers against the fully glycosylated virions, were not enhanced. Our results indicate that glycosylation of gN of HCMV represents a potentially important mechanism for evasion of antibody-mediated neutralization by a number of different antibody specificities.
| Cytomegaloviruses (CMV) have co-evolved with their respective hosts. During this long and continuing co-evolution these viruses have adapted to the host defense systems and vice versa to allow the life-long persistence of these viruses. As a result, infections in immunocompetent hosts are generally asymptomatic and a life-long persistent/latent infection is readily established. Development of symptoms or disease is prevented by a multilayered, in large parts redundant, innate as well as adaptive immune response [1]. Persistence and transmission between hosts eventually requires the evasion of immune control. Multiple mechanisms that permit evasion of immune control by the innate and adaptive cellular immune responses have been extensively documented [1]–[3]. In contrast, very little is known about mechanisms by which CMV can evade humoral immune responses that presumably consist of antiviral antibodies that potentially neutralize free virus or destroy infected cells via antibody mediated cytotoxicity. Since viral transmission between hosts in a community setting is thought to occur via cell free virus in most cases that have been studied, evading virus neutralizing antibodies is essential for successful spread and persistence of CMVs in the population.
On the population level, the extensive strain polymorphism that has been documented in different human and animal CMVs could serve as an immune evasion strategy [4]–[6]. CMV strains exhibiting antigenic and genetic variability are capable of super-infecting immune hosts and can be readily transmitted between immune individuals [7], [8]. Transmission of super-infecting strains has been well documented during pregnancy or following organ transplantation [9], [10]. Strain-specific virus neutralization is potentially a contributing factor to this phenomenon and strain-specific neutralization has been observed in a number of studies [11], [12]. Thus, virus strain-polymorphism can be considered as a mechanism that permits successful maintenance of CMV within the host population.
To evade neutralizing antibodies on the level of an individual viral strain within a single host requires an evasion strategy other than virus strain-polymorphism. The development of virus mutants during virus persistence in an individual host that resist virus neutralizing antibodies is, based on existing data, only a theoretical possibility. Herpes viruses in general are believed to be genetically stable secondary to the proofreading activity of their DNA polymerase. Moreover, the neutralization of CMV in vivo almost certainly involves antibodies directed at a multitude of different viral antigens. Thus, multiple mutations would be required for evasion of virus neutralizing activities that are likely to be present in a single individual host. Induction of neutralizing and non-neutralizing antibodies which compete for binding to the same antigenic determinant has been described as one possible mechanism for evasion of neutralization but this has been shown only for a single antigenic site on glycoprotein (g) B [13], [14]. Another possibility for viral evasion of neutralizing antibodies is the addition of carbohydrate to virion envelope glycoproteins that can alter antibody binding, a mechanism that has been extensively documented for viruses such as HIV and influenza, among others [15].
Human CMV (HCMV) is a structurally complex virus which contains a large number of envelope glycoproteins, several of which are predicted to be extensively glycosylated. One such extensively glycosylated envelope proteins is gN, a type I glycoprotein that is particularly interesting for several reasons. It is component of the gM/gN complex which is among the few envelope proteins that are conserved between the herpes viruses indicating an important function of this glycoprotein complex in the biology of herpes viruses [16]. For HCMV, gN is essential for virus replication whereas for alpha-herpes viruses it has been classified as non-essential [17]–[20]. In contrast to alpha-herpesviruses, the gM/gN complex is the most abundant protein complex in the HCMV virion envelope [21]. gN is an extremely polymorphic protein at the amino acid (aa) level. So far four major genotypes have been identified [22]. Differences between the gN genotypes are exclusively located within the surface domain of the protein, reaching amino acid differences of up to 50% between genotypes [22]. The protein is extensively modified by O-linked sugars contributing over 40 kDa of mass to the 15 kDa polypeptide backbone [23]. Despite the enormous amino acid variation in the surface domain of gN, the total number of serine (ser) and threonine (thr) residues remains constant at approximately 50. The conservation of the number of potential glycosylation sites in the face of significant primary sequence variation of non-ser/thr residues in the surface domain of this molecule suggests a strong selective pressure to maintain this precise level of glycan density. In contrast to gN from HCMV, which has a primary sequence of 138 aas, the gN proteins from alpha- and gamma-herpes viruses without exception are smaller molecules of approximately 100 aa that are not predicted to be extensively glycosylated. Experimental data have confirmed the prediction in those cases where the proteins have been studied [18], [24]–[26]. Even within the cytomegalovirus family, gN homologous proteins of the most species are predicted to represent small proteins with limited modifications [27], [28]. The extensive glycosylation of CMV gNs seems to be restricted to viruses derived from the great apes and humans since gN from chimpanzee CMV is predicted to contain a comparable number of O-glycosylation sites to HCMV while gN from rhesus and cynomolgus CMV are short, largely unmodified proteins [29]–[31].
The function(s) of the carbohydrate moieties of gN is largely unknown. The gCII complex, which has been shown to consist of gM and gN, has previously been proposed to be involved in the initial interaction process between the target cells and the virus since it was reported to bind to heparin [32]. Cell-surface heparan sulfate proteoglycans are thought to represent the initial molecules used by the virus to adhere to target cells [33]. With respect to the humoral immune response, gN has been identified as a target of neutralizing antibodies [34]. In fact, in human serum the capacity to neutralize infectious virus in vitro is comparable between anti-gN and anti-gB antibodies [34]. Moreover, exchanging the gN-genotypes in a single, genetically homogeneous HCMV strain resulted in strain-specific neutralization by human convalescent sera further emphasizing the importance of gN for the humoral immune response [12].
We hypothesized that the extensive glycosylation of HCMV gN could provide the virus with a mechanism to evade neutralization by antibodies. To test this hypothesis we generated gN-recombinant viruses with reduced carbohydrate modification. Our results indicated that under-glycosylation of gN increased the susceptibility to the neutralizing activity of antibodies directed at gN. Unexpectedly, we also demonstrated that recombinant viruses with under-glycosylated gN were more susceptible to antibodies directed against a number of different virion envelope proteins of HCMV that have been shown to be major targets of the neutralizing antibody response. Together these findings suggest that one function of the extensive glycosylation of gN could be to limit the activity of virus neutralizing antibodies directed at different envelope glycoproteins, a function similar to that of carbohydrates that serve as a glycan shield to limit antibody neutralization of RNA viruses.
More than 50% of the amino acids within the surface domain of gN are ser or thr residues that can serve as substrate for the addition of O-linked sugars. The finding that the viral protein migrates in SDS-PAGE as a 50–60 kDa diffuse species while the theoretical molecular mass of gN is 15 kDa indicating that a significant number of the potential glycosylation sites are utilized [23]. The consequences of this extensive glycosylation to the function of gN and ultimately in the biology of HCMV are unknown. Because site specific mutagenesis of ser and thr residues (total of 36 in HCMV strain AD169) individually and in combination in the surface domain of gN would represent an experiment of considerable complexity, we estimated the impact of reduced glycosylation of gN to the formation of the essential gM/gN complex using mutants constructed by deletion of stretches of ser or thr rich areas of the surface domain of gN. Expression plasmids were constructed that upon transfection into mammalian cells would give rise to truncated gN proteins that could be studied following transient expression. Deletion of aa 24–40 were made to yield plasmid gN-41sig, aa 24–60 and aa 24–89 to yield gN-61sig and gN-90sig, respectively (Figure 1). All plasmids were constructed as to maintain the authentic gN signal sequence which is predicted to be located between aa 1–21. To facilitate protein detection, all proteins were expressed with a myc-epitope at the carboxyl terminus.
The plasmids were individually co-transfected with a full length wild type gM encoding plasmid into Cos7 cells and complex formation was analyzed using the gN-specific monoclonal antibody (mab) 14-16A. This antibody has previously been shown to be specific for gN that is complexed with gM [23]. This mab does not react with gN that is not complexed with gM, thus providing an assay for the maintenance of sufficient structure of gN to allow complex formation with gM [23]. Reactivity with mab 14-16A was seen in cells transfected with gM combined with gN and gN-41sig as well as the localization of the protein in the TGN (trans-Golgi network), findings that indicated that a complex was formed between these two proteins and that the trafficking of the complex within the cell was similar to the wild type gM/gN complex (Figure 2). No reactivity with mab 14-16A was detected following co-transfection of gM plus gN-61sig, perhaps secondary to a loss of the epitope recognized by mab 14-16A. However, when the cells were stained with an anti-myc antibody we observed reactivity that co-localized the myc-tagged gN with markers for the TGN, indicating formation and correct transport of the gM/gN-61sig complex. Isolated expression of gN results in compact intracellular aggregation of the protein in structures containing endoplasmic reticulum markers and defective transport to the TGN [23]. A similar intracellular traffic defect was observed for the gN-truncated proteins and gN-41sig is shown as example (Figure 2). Transfection using gM combined with the gN-90sig plasmid did not result in a protein that could be detected by immunofluorescence with either antibody suggesting that this deletion resulted in loss of protein structure required for complex formation with gM (data not shown). Together these data demonstrated that deletion of stretches of ser/thr rich sequences of gN could be accomplished without loss of structure required for complex formation with gM and trafficking of this complex to the TGN in transfected cells.
To generate recombinant viruses expressing truncated gN proteins, we followed our previous strategy, i.e. construction of HCMV bacterial artificial chromosomes (BAC) carrying the truncated versions of the gN-gene in place of the full length gene, followed by reconstitution of infectious virus in human cells [19]. All viruses were constructed in the genetic background of HCMV strain AD169 [35]. Replicating virus was recovered for gN versions starting at aa 41 and aa 61 giving rise to RVgN-41sig and RVgN-61sig, respectively. In several attempts no replicating virus could be recovered from BACs carrying the gN-90sig mutation, indicating that this large deletion in gN is lethal, a finding that confirmed our previous results of the essential role of gN for replication of HCMV [36]. RVgN-41sig and RVgN-61sig recombinant viruses replicated with similar efficiency when compared to the parental virus RVAD169, a finding that was consistent with the capacity of these two gN mutants to form a complex with gM and traffic normally in transfected cells. (Figure 3A). In accordance with the similar efficiency of replication of the gN-truncated viruses, we observed no delay in expression of immediate early proteins as determined by indirect fluorescence arguing that the early infection events are similar for the three viruses (data not shown).
To determine if gM/gN complex formation occurred in cells infected with the respective viruses, we carried out indirect immunofluorescence analysis 5 days after infection (Figure 3B). In cells infected with the wild type RVAD169 virus, the gN signal colocalized with gB in a region in close proximity to the nucleus which has been termed the assembly compartment (AC) [37]. In cells infected with RVgN-41sig a similar staining pattern was observed, indicating that the truncated form of gN did not influence the trafficking of the gM/gN complex to the AC. As was demonstrated in transfected cells, expression of the gN-61sig protein could not be detected using the mab 14-16A. However, co-localization of gM and gB in cells infected with RVgN-61sig could be detected and because gM complex formation with gN has been shown to be required for transport of gM from the ER, this finding indicated that the gM/gN-61sig complex had been transported properly to the AC [23]. (Figure 3B). Co-localization of gN-61sig and gB was also observed when gN-61sig was detected via the myc epitope present on gN-61sig (Figure 3B).
To analyze incorporation of the different gM/gN complexes into extracellular virus particles, we gradient purified the respective viruses using ultracentrifugation through glycerol-tartrate gradients and analyzed the viral lysates for the presence of the gM/gN complex by western blot. When virion lysates were analyzed under reducing conditions we did not detect differences in the ratio of the major capsid protein (MCP) and gB between the viruses (Figure 3C).
The gM/gN complex was analyzed under non-reducing conditions since gN migrates as a smear under reducing conditions preventing an accurate estimation of the amount of protein [23]. To detect gN complexes present in all three recombinant viruses we used a gM/gN-specific polyclonal human serum, that was affinity purified from a HCMV hyperimmune globulin preparation [34]. It was shown to be monospecific for gM/gN [34]. The amount of protein that was applied to the analysis was adjusted to give a comparable gB-specific signal for all three viruses (Figure 3C). The results demonstrated that the RVAD169 and RVgN-41sig contained similar amounts of the respective gM/gN complex. For RVgN-61sig, the amount of gM/gN complex was more difficult to estimate due to the diffuse migration of the complex but appeared similar to the other two viruses. The explanation for the diffuse migration of the gM/gN-61sig complex is unknown but a plausible explanation is increased structural heterogeneity of the remaining carbohydrate modifications secondary to loss of a significant number of potential O-linked glycosylation sites. The presence of a similar gM/gN to gB ratio was also confirmed by western blot analyses for RVAD169 and RVgN-41sig using mab 14-16A (data not shown). To obtain more quantitative data on the different proteins in the respective virion particles we performed an ELISA using lysates from gradient purified virions as coating antigen. The ratio of MCP to the envelope proteins gB and gH was comparable for the three different recombinant viruses (Figure 3D). Note that for this analysis an anti-gH mab was used that is dependent on the native conformation of the antigen, indicating that the lysis procedure left the proteins largely intact [38]. Virions from RVgN-61sig gave a reduced signal with mab 14-16A compared to RVAD169 and RVgN-41sig, confirming the results of the indirect fluorescence analysis. Together these data argue that deletion of stretches of ser/thr rich sequences within the surface of domain did not alter the function of the gM/gN complex required for the production of replication competent viruses. Furthermore, the stoichiometry of the three major glycoprotein components of the virion envelope appeared to be unaltered in virions produced by these recombinant viruses suggesting that the deletion of these sequences in the surface domain of gN did not alter the incorporation of the gM/gN complex (or gB, gH) into the envelope of the virus.
The gM/gN complex of HCMV was originally designated gCII complex and it was reported that components of the gCII complex have heparin-binding capacity [32]. Heparin binding is most likely secondary to the carbohydrate modifications of gN since gM is minimally glycosylated and largely buried in the viral envelope. We therefore tested heparin for blocking infection of the gN-truncated viruses. In accordance with previous reports, we observed almost complete inhibition of infection in the presence of 2 µg/ml heparin [39]. Importantly, there was no difference between the three viruses in terms of their capacity to be inhibited by the addition of heparin (Figure 3D). In summary, when combined these data indicated that the behavior of gN-truncated viruses in these in vitro assays were phenotypically very similar if not identical to the parental RVAD169.
Several possible functions have been suggested as explanations for the extensive carbohydrate modifications of gN including a potential role in cell binding, possibly as a result of its interactions with cell surface glycosaminoglycans and/or serving to limit accessibility of anti-gN antibodies that could neutralize infectious virus. We examined this latter possibility by using the gN truncation mutant viruses in antibody mediated virus neutralization assays. To specifically compare the impact of the loss of carbohydrate modifications and the loss of amino acid sequence on susceptibility of the mutant viruses to neutralization by antibodies, we included antibodies reactive with gN as well as other envelope glycoproteins, gB and gH in these assays. With the exception of 14-16A, an IgM antibody that neutralizes HCMV only in the presence of complement, all antibodies neutralize HCMV in the absence of complement. The gN-specific mab 14-16A showed increased capacity to neutralize RVgN-41sig. RVgN-61sig was not neutralized, which was consistent with the findings that gN-61sig was not recognized by this antibody (Figure 4). These results argued that the loss of carbohydrate and not the antibody recognition site in the RVgN-41sig virus encoded gN was responsible for increased susceptibility to neutralization by this antibody. Moreover, it could also be argued that that carbohydrate modifications on wild type gN functioned to limit the virus neutralizing function of mabs directed against gN. Unexpectedly and perhaps more importantly, both of the gN-truncated viruses were significantly more susceptible to neutralization by other, non-gN specific neutralizing mabs utilized in these assays. The effect was most pronounced for the gH-specific murine mab 14-4b (gH1) and the human anti-gB mab ITC88 (gB-AD2), where differences in 50% neutralization titer of approximately 10-fold were detected. The human anti-gH mab MSL-109 (gH2) and the human anti-gB mabs SM5-1 (gB-AD4), 1G2 (gB-AD5) and C23 (gB-AD2) showed less drastic differences in virus neutralizing activity between the parental and the gN-truncated viruses (Figure 4). Note that the gH-specific mabs and two of the gB-specific mabs (SM5-1, 1G2) depend on native antigen conformation for binding. When non-neutralizing mabs against gB or gH were tested, we found no increase in neutralization activities of these antibodies. The anti-gB mab 27–156 (gB-AD1) is shown as an example. Because the biochemical composition of the envelope as measured by the amounts of three major glycoproteins in the virion envelope was unchanged in the mutant viruses that lacked wild type levels of glycosylation on gN, these results suggested that changes in carbohydrate content of a single envelope glycoprotein were responsible for the increased susceptibility of mutant virions to virus neutralizing antibodies directed at unrelated envelope proteins. These findings raised the possibility that the extensive carbohydrate modifications of gN could be functioning in a similar fashion as the glycan shield that has been proposed for other viruses, including HIV [40], [41].
Sera from HCMV infected individuals contain virus neutralizing antibodies directed against a number of different envelope glycoproteins. The majority of antibodies has been suggested to be directed against gB, gH and gN when such sera are analyzed with laboratory strains of HCMV such as strain AD169, which was used in this study [34], [42], [43]. To determine whether the effect observed when mabs were used to neutralize the gN-truncated viruses would be reflected in differences in virus neutralization by polyvalent human sera, we carried out neutralization assays with randomly selected sera from HCMV seropositive donors. A total of 11 specimens were tested and representative results are shown in Figure 5. As could be expected, the polyvalent sera showed less marked differences in neutralization titer between parental virus and the gN-truncated versions than was observed when assays were carried out with single antigen/epitope specific mabs. Two sera showed a significant difference (represented by serum 57 and 97), and the remaining sera differences between 1,3 and 2,2 fold, which, however, did not reach statistical significance (represented by serum ER). In addition, a commercial immunoglobulin preparation (Sandoglobin), presumably derived from a large number of donors, also showed a higher neutralization titer against the viruses expressing the truncated forms of gN, although this difference did not reach significance.
We have previously shown that neutralizing anti-gB antibodies do not prevent attachment of fully glycosylated virions to target cells making inhibition of attachment an unlikely mechanism responsible for the increased neutralizing sensitivity of the gN-truncated viruses to anti-gB antibodies [44]. However, the possibility existed that the removal of carbohydrates results in an altered steric orientation of the antibodies bound to the virion surface and thereby provide a new functional property that could alter virion attachment. Therefore, we tested attachment of the different viruses to fibroblast target cells in the presence of antibody. Virus/antibody mixtures were added to target cells at 4°C and the number of HCMV DNA copies attached to the cells was determined by quantitative real time PCR. Attachment of virions to fibroblasts was similar for RVAD169 and the gN-truncated viruses and was not influenced by addition of antibody (Figure 6A). The slight increase of bound virus in the presence of the gB-AD2 specific antibody C23 as compared to control was repeatedly observed and might reflect deposition of antibody/virus aggregates on the surface of cells. Neutralization of HCMV via AD2 specific antibodies requires both arms of the IgG Fab fragment and thus crosslinking of different viruses by C23 is a possibility [45]. We next determined the activity of the mabs towards virus that was adsorbed to cells by pre-adsorbing virus to cells for 1 h at 4°C before the gH-specific mab 14-4b was added. As shown in Figure 6B mab 14-4b was capable of neutralizing HCMV at a postadsorption step. In contrast to our findings from assays of antibody inhibition of attachment, the gN-truncated viruses were more susceptible to neutralization than the parental wild type virus. The higher antibody concentration that was required to completely neutralize adsorbed virus was presumably secondary to the requirement of blocking fusion of an already attached virion.
Finally, to determine whether antibodies that neutralized gN-truncated viruses more efficiently than the parent virus had altered kinetics of virion binding, we performed experiments that allowed an estimate of the rate of antibody-mediated virus neutralization. Virus and antibody were mixed at 4°C and either used immediately to infect cells or warmed to 37°C for 15 min or 30 min before adding to target cells and the percent neutralization was determined 24 h later. As can be seen in Figure 6C, the anti-gH antibody showed increased neutralization capacity towards the gN-truncated viruses at every time point, reflecting the observations that were made in our standard neutralization assays. Neutralization by the gB-AD2 specific human mab C23 was less affected by truncation of gN but clearly detectable (Figure 6C). Again the finding was consistent with the results from the standard neutralization assay as shown in Figure 4 and because the kinetics were similar, argued that the increased susceptibility to virus neutralizing antibody that was seen with the gN truncated viruses was not secondary to alterations in the kinetics of antibody binding to the virion.
In summary, these results indicated that the presence of antibody did not influence the attachment of the gN-truncated viruses to target cells. Moreover, increased neutralization of the gN-truncated viruses was observed very early after combining antibody and virions and even after virions had attached to fibroblasts, suggesting an improved accessibility of target epitopes for the gB- and gH-specific mabs.
The data presented thus far indicated that viruses with truncated gNs differed in the accessibility of epitopes on envelope proteins. Whether these differences would also alter the antigenicity of the viruses was tested in the next series of experiments. Groups of 3 mice each were immunized with equal amounts of gradient purified RVAD169, RVgN-41sig and RVgN-61sig, respectively. The resulting sera from each group were pooled and tested in an ELISA for production of HCMV specific antibodies using RVAD169 as antigen. The individual pools had comparable ELISA titers against whole HCMV antigen as well as against gB alone (Figure 7A). Control mice that had been injected with PBS did not develop anti-HCMV antibodies (Figure 7A). We then tested the serum pools for neutralization of RVAD169 and the gN-truncated viruses (Figure 7B). The individual viruses were neutralized by the different serum pools with similar titers. However, the 50% neutralization titers of the serum pools for RVAD169, RVgN-41sig and RVgN-61sig were different. Whereas, the three serum pools showed 50% neutralization at a dilution of approximately 1∶800 when RVAD169 was used, the titer increased to approximately 1∶3200 when RVgN-61sig was used as target (Figure 7B). 50% neutralization titers for RVgN-41sig by the three serum pools was observed in the range of 1∶1200 (Figure 7B).
The increased susceptibility of RVgN-61sig to virus neutralizing antibodies raised the question whether this phenomenon was based on an overall increase in neutralizing antibody titer or a specific increase in a selected set of antibodies. To test this, we performed ELISA assays using antigens known for binding of neutralizing antibodies. These included three well characterized antigenic domains on gB, namely AD1, AD2 and AD4 [44], [46], [47]. In addition, the AD86 epitope on gH as well as an antigenic region on the tegument protein pp150 was analyzed [48], [49]. For these analyses, the epitope detected by the anti-gH antibody 14-14b could not be assayed secondary to its conformational dependence. The serum pool derived from RVAD169 immunized mice showed comparable reactivity against epitopes located on gB, gH and pp150 (Figure 8). Sera from mice immunized with the gN-truncated viruses showed a different pattern of reactivity. Whereas reactivity was comparable between all three serum pools for the AD1 epitope on gB and the AD86 epitope on gH, sera from the gN-truncated immunized animals showed drastically enhanced antibody titers against gB AD2, gB AD4 and pp150. Testing the sera individually gave similar reaction pattern indicating that pooling the sera did not bias the result (data not shown). These data suggested that the increase in neutralization capacity in sera from gN-truncated viruses could be based on a selective increase of antibodies binding to neutralization relevant epitopes on different envelope glycoproteins that were less accessible in the wild type AD169 virus.
The remarkable variation in the primary sequence of gN that has been documented in different viral isolates and the extensive oligosaccharide modifications of the gN of HCMV are two characteristics of the HCMV gN that demonstrate its uniqueness among the gNs of human herpes viruses. The observation that the predicted number of potential sites for O-linked carbohydrate modifications in gN remains constant regardless of the variability of the primary sequence argues for a critical role of this modification for the biology of HCMV. Our results are consistent with the possibility that the carbohydrate modifications of this envelope protein limits the activity of virus neutralizing antibodies and suggests at least one functional role for this post-translational modification of this structural protein.
The results of our studies indicated that deletion of ser/thr rich primary sequence of the gN ectodomain and the associated carbohydrate modifications associated with these ser/thr residues had no measurable impact on the formation of the gM/gN complex, localization of the complex in sites of virus assembly, and virus replication. Furthermore, the early events of infection such as adsorption, penetration and expression of immediate-early proteins were similar for both gN-truncated viruses and the wild type parent virus. Likewise, for all viruses in this study attachment was inhibited to a similar extent by the addition of heparin, a finding that argued that either the carbohydrate modifications on gN were unimportant for cell surface proteoglycan binding or that residual carbohydrates on the truncated gNs viruses were sufficient for binding to these cellular molecules. Consistent with the former possibility is the finding that other envelope glycoproteins such as gB can also efficiently bind the heparan sulfate [39]. Overall, our results strongly argued that the deletions introduced into the surface domain of gN did not alter the structure of the molecule so as to limit its functional interactions with gM and its essential role in virus replication. In contrast to these results, deletion of oligosaccharides from gN had a significant impact on the susceptibility to virus neutralization by antibodies suggesting that the oligosaccharides played a role in limiting the activity of virus neutralizing antibodies. The effect was seen for antibodies whose activity depended on binding in cis, i.e. gN-specific antibodies. Importantly, deletion of 17 amino acids from the ser/thr sequence in the amino-terminus of gN resulted in a replication competent virus that had increased susceptibility to the neutralizing activity of a mab that reacted with gN from both the wild type and gN truncated virus. Because this antibody recognizes a non-conformation dependent binding site, based on its reactivity with denatured protein, the increased neutralizing activity of this mab for the RVgN-41sig virus is unlikely to be secondary to a conformational change in gN. Together these finding argued that alteration in the carbohydrate modifications of gN in the virus expressing the truncated form of gN likely accounted for the increased virus neutralizing activity of the anti-gN mab when assayed with the RVgN-41sig virus as compared to wild type virus.
Of perhaps greater interest was the finding that deletion of ser/thr rich sequences in the surface domain of gN resulted in viruses that were more susceptible to virus neutralizing activities of antibodies whose binding was in trans to gB and gH. This finding was of significance for several reasons. Perhaps the most obvious is that it argued strongly that the mutations introduced into gN that altered the carbohydrate content of gN also had an effect on the recognition of two other abundant envelope glycoproteins by neutralizing antibodies. The mutations in gN did not alter the biochemical content of the envelope as measured by the amount of gN/gM, gB, and gH in the envelope nor the conformation of gB or gH as revealed by their continued recognition by conformation dependent mabs. However, these mutations did alter the susceptibility of these mutant viruses to the neutralizing activity of these mabs. Removal of 17 or 37 residues from the extraviral part of gN resulted in a phenotype that was similar with regard to susceptibility to neutralization. This result indicated that the glycan modification in toto and not site specific carbohydrate modification may be required for the inhibition of virus neutralizing antibody activity that is observed in vitro. Whether similar requirements are operative in vivo is unknown but the maintenance of the number of O-linked glycosylation sites in the coding sequence of gN from a large number of clinical isolates would argue that for optimal evasion of antiviral antibodies, usage of all of the potential carbohydrate modification sites would be required.
Our findings argue for a role for the carbohydrate modifications present on gN in antibody recognition of envelope glycoproteins by virus neutralizing antibodies. The structural relationships between the gM/gN complex, gB, and gH/gL in the AD169 laboratory strain of HCMV are unknown but the complexity of protein composition of the virion envelope would suggest that the structure of the envelope would be complex. Because of this complexity we cannot definitively exclude the possibility that a deletion such as the 17 aa deletion in the amino terminus of the gN41-sig mutant could result in major structural changes in the envelope of the virion once the gM/gN41-sig complex was incorporated. Such structural changes resulting from this mutation could be proposed but not tested with available methodologies. Probing the envelope of the mutant viruses with mabs does argue that the overall structure of the envelope is likely intact and that similar stoichiometric relationships between the three major glycoproteins are maintained, findings that are also consistent with the similar in vitro replication of mutant and wild type viruses. However, we cannot formally exclude the possibility that the deletion of residues in the surface domain of gN lead to structural rearrangements within the envelope of infectious virions that increased the neutralizing activity of antibodies directed against three different envelope glycoproteins.
Glycan shields that protect viruses from antibody-mediated neutralization are a well described phenomenon and have been extensively investigated in a number of RNA viruses. The most well studied examples are HIV and influenza virus [40], [41]. For these viruses it is sufficient for the shield to work in cis meaning that the protein that carries the sugar is protected by carbohydrate modification, which could be expected since the viruses carry a single multifunctional protein that works in attachment, receptor binding and fusion. Thus, blocking antibody access of a limited set of protein domains that are crucial for the functional activity of neutralizing antibody is sufficient. The envelopes of herpes viruses are considerably more complex carrying several neutralization-relevant targets that exhibit redundant functions in the early events of virus infection. In case of HCMV these include at least gB and different gH complexes [50], [51]. Thus, evasion from neutralizing antibodies would require a potentially very different strategy and one that is likely much more complex because simply protecting a single protein would likely not be sufficient. The induction of neutralizing and non-neutralizing antibodies which compete for binding to the same antigenic domain is a mechanism that has been described for the gB neutralizing site AD1 [13]. Whether such a mechanism is operative for additional antibody binding sites on the HCMV envelope is unknown. More recent data show that there are domains on gB which are bound exclusively by neutralizing antibodies indicating that competition of neutralizing and non-neutralizing antibodies is probably not a general evasion mechanism [44]. Data from the present study raises the possibility that carbohydrate shielding of several glycoproteins by the heavily glycosylaetd gN envelope protein could be a second major mechanism that limits the neutralization of virus infectivity by antibodies. Deletion of a fraction of the predicted oligosaccharide addition sites from gN resulted in increased neutralization activity of a number of mabs that have been shown to be directed at different envelope proteins, although the effect was not equivalent for all antibodies. Whereas one anti-gH antibody and an anti-gB AD2 antibody required 5–10 times higher concentration to achieve 50% neutralization titers towards the gN-truncated viruses, antibodies against other epitopes on gB or gH were less affected by the gN truncations. This finding argues the functional efficiency of the proposed glycan shielding may be heterogeneous depending on the binding activity and target epitope of different mabs. However, with the possible exception of the anti-gB antibody C23, we did not see similar neutralization capacity of any antibody for the parent and the gN-truncated viruses.
We can only speculate on the mechanism(s) that are involved in the glycan mediated shielding of HCMV from the activity of virus neutralizing antibodies. A significant change in the protein composition of the envelope secondary to the gN-truncation could facilitate binding of neutralizing antibodies by drastically altering the structure of the virion envelope. However, this seems unlikely since we did not detect changes in the gB/gN ratio in purified virions nor significant changes in virion binding or internalization. In addition, the observed epitope-specific effect on virus neutralizing antibody activity together with the unchanged stoichiometry of the envelope glycoproteins gB, gH, and gM/gN in recombinant viruses with deletions in the ser/thr rich sequences of gN would argue against a global change in the viral envelope. In support of these arguments, deletion of individual envelope proteins in HSV-1 did not change the composition of others [52]. Thus, it is unlikely that deletion of 17 aa, as in the case of RVgN-41sig, resulted in significant changes in glycoprotein composition or structure of the viral envelope.
An obvious explanation could be the shielding of a fraction of epitopes on gB and gH by the glycans of gN thereby impeding antibody access to selected epitopes. Whether this mechanism is possible given the differences in protein size between gB (approx. 700 aa surface domain), gH (approx. 700 aa surface domain) and gN (approx. 100 aa surface domain) remains to be determined experimentally and will require structural information about the epitopes recognized by neutralizing antibodies that are exposed on the virion surface. Alternatively, removal of sugars from gN could alter the architecture of the outer surface of the envelope without affecting copy numbers of the individual protein components of the envelope. Mass spectrometry of HCMV has identified gM as the most abundant glycoprotein in the viral envelope. It can be assumed that gN is also highly abundant since it is covalently linked to gM via a disulfide bond [21]. The linkage is between cysteine at position 90 in gN and cysteine at position 44 in gM, thus it is not affected by the truncations that were introduced in gN [36] and the western blot analyses of the gN-truncated virions under non-reducing conditions supported this assumption. By occupying space, the bulky carbohydrate head of fully modified gN could result in closer packing of the surface domains of gB and gH, thereby impeding access of some gB- and gH-specific antibodies. If sugar is removed, it may result in a decreased density of the spacing of the respective proteins and permit easier access of antibodies to epitopes on gB and gH. Alternatively, structural changes may be induced resulting in altered binding avidity of immunoglobulin which could affect neutralization sensitivity [53]. The fact that we observed increased susceptibility for only a subset of neutralization relevant epitopes would be compatible with any of these mechanisms.
It is interesting to note, that HCMV carries a number of additional glycoproteins which have been shown to contain significant carbohydrate modifications, such as gB and gO [54]–[56]. If they were to have similar effects, the overall protection from the activity of virus neutralizing antibodies would be very significant. For gB of murid herpes virus 4, a significant effect of glycosylation on the evasion from virus neutralizing antibodies has been demonstrated [57]. There is indirect evidence that gO of HCMV may also protect the virus from neutralizing antibodies. Jiang et al. [58] have reported that the cell-to-cell spread of a recombinant virus lacking gO is more sensitive to neutralization by polyclonal sera than the parental virus. Thus, the protection from antibody-mediated neutralization by glycan modification on envelope glycoproteins could represent a more general phenomenon for herpes viruses than has been previously considered.
The immunization experiments indicated that gN-truncated viruses induced an antibody response that was not different from animals immunized with the parent virus in terms of the overall ELISA titer and gB ELISA titer. However, when the sera were tested in virus neutralization assays, a clear difference was seen. The RVgN-41sig virus and the RVgN-61sig virus were neutralized more efficiently by any serum pool, a result that was most apparent when the RVgN-61sig virus was used in these assays. These findings again emphasize that the gN-truncated viruses were more readily neutralized when compared to the wild type parental virus. Interestingly, on the epitope level, the gN-truncated viruses induced an antibody response that was different from the response against the parent virus. Epitopes which represented minor targets and induced only low levels of antibody in the wild type RVAD169 virus immunized animals became markedly more immunogenic in the gN-truncated virus immunized animals. This was most pronounced for the AD2 and AD4 epitopes on gB but also seen for the epitopes on gH and pp150. Interestingly, this was again an epitope specific effect because this response was not as apparent for the AD1 epitope on gB, possibly because of the dominance of this epitope in the antibody response to HCMV in mice and humans. Thus, the trans-effect that was seen in neutralization of the gN-truncated viruses with the different mabs was also reflected after immunization with viruses deficient in the glycan modifications of gN. The underlying mechanisms for the induction of a different set of antibodies by the fully glycosylated virus and the gN-truncated viruses are unclear at present but may involve activation of a different set of naïve B cells to produce antibodies after antigenic stimulation, a mechanism that has recently been suggested for the induction of neutralizing HIV-specific antibodies [59]. Defining the underlying mechanism for the enhanced immunogenicity of selected epitopes on the HCMV envelope proteins is a relevant question for the development of immunogenic components for vaccine development and is actively under investigation in our laboratories. Finally, a recent study of bovine herpes virus 4 that detailed the effect of O-glycosylation of gp180 on antibody evasion also found increased sensitivity of gp180-deficient virions to antibody-mediated neutralization but no difference between the immunogenicity of viruses with or without expression of gp180 [60].
It was surprising that we did not observe a difference in antibody binding ELISA assays when we used whole virus or purified gB as antigenic substrate. The most likely explanation is that the overwhelming composition of the antibody response against either antigen was directed at antibody binding sites irrelevant to the activity of virus neutralizing antibodies. Our analysis of the human antibody repertoire against gB has revealed that >95% of antibodies directed against gB are non-neutralizing [44]. The situation may be similar after immunization of mice with HCMV particles. Increased antibody titers against neutralization relevant epitopes, however, did not translate into increased neutralization capacity of the respective serum pool against the virus carrying full length gN suggesting that cryptic epitopes were not exposed in viruses lacking the full complement of carbohydrate modifications on gN. Thus, the resistance of viruses containing full length fully glyosylated gN can be most readily explained by the shielding by carbohydrates of the respective epitopes recognized by virus neutralizing antibodies. Lastly, what could be the relevance of our findings for the activities of virus neutralizing antibodies in vivo? The results from the immunization experiments in mice suggest that individual differences in the antibody repertoire induced in infected individuals in an outbred population such as humans has only limited consequences for the neutralization of the virus in vivo since neutralization relevant epitopes might be protected by the glycan shield of the virus.
This study was performed in strict accordance with German law (Tierschutzgesetz). The protocol was approved by the Committee on the Ethics of Animal Experiments at the Bavarian Government (Regierung von Mittelfranken, permit 54-2531.31-8/04). All efforts were made to minimize animal suffering.
The human kidney cell line 293T and human lung fibroblasts (MRC5) were maintained in Dulbecco modified Eagle medium supplemented with 10% fetal calf serum (FCS), glutamine (100 mg/L) and gentamycin (350 mg/L). Human foreskin fibroblasts (HFF) were kept in minimum essential medium supplemented with 5% FCS, glutamine and gentamycine. All viruses used were propagated on MRC5 cells and viral titers were determined in HFF using an indirect immunofluorescent assay with the mab p63-27, directed against the HCMV immediate-early 1 (IE1) protein [61]. Virions were isolated by glycerol-tartrate gradient centrifugation as described [62]. For growth curves, HFF, plated in 24-well dishes, were infected at a multiplicity of infection (m.o.i.) of approximately 0.1. After adsorption of virus (2 hours), the inoculum was removed and replaced by fresh medium. Supernatants were harvested at the indicated time points and stored at −80°C until use. Virus titers were determined by an indirect immunfluorescence assay using a mab against the IE-1 protein of HCMV as described [61].
The mabs used in this study have been described. Murine mab: gM-specific IMP91-3/1 [23], gB-specific 27–287 [63], gN-specific 14-16A [23], gH-specific 14-4b [38]; Human mab: gB-specific C23 (kindly provided by Teijin Pharma Limited, Japan) [64], ITC88 [14], SM5-1 and 1G2 [44]. For specificity of the mabs see Figure S1. Secondary antibodies were purchased from Dianova or DAKO. Sera from HCMV- positive and negative individuals were randomly selected from our diagnostic department.
Serial mab or serum dilutions were incubated with virus preparations for 1 h at 37°C. Viral titers were adjusted to give 100 to 150 infected cells counted on a fluorescence microscope using a 200× magnification, equivalent to 2000 infected cells/15000 cells. The virus antibody mixture was added to fibroblasts which were seeded at 1,5×104 per well in 96-well plates the day before. The medium was replaced 4 h later and the number of infected cells was counted 16 h later using indirect immunofluorescence with mab p63-27. Percent neutralization was calculated as reciprocal of infectivity with maximum infectivity being determined by incubation of virus without antibody. The number of infected cells without addition of antibody also served as reference for the determination of infectious units (IU). For the kinetic experiments, all reagent (virus, antibody, tissue culture medium) were cooled to 4°C before mixing. Mixtures were transferred to 37°C for the indicated time periods and added to fibroblasts at 37°C. The medium was replaced 4 h later and the number of infected cells determined as described above.
Balb/c mice were obtained from Charles River Laboratories, Inc.. Groups of 3 mice each were immunized with the respective virus. 200 µl containing 5 µg of gradient purified virus and 100 µl aluminum hydroxide adjuvant were administered intraperitoneally. Booster immunizations were given at weeks 5 and 8. Animals were sacrificed and sera of mice immunized with the same virus were pooled.
Plasmids expressing truncated forms of gN were constructed on the basis of pcDNA3.1myc/his (Invitrogen). First, the gN coding sequences for residues 1–23, which includes the predicted signal sequence, was inserted using the EcoRI and BamHI restriction sites of pcDNAmyc/his. In a second cloning step, gN sequences coding for aa 41–138, 61–138 and 90–138 were inserted into the gN-signal sequence containing plasmid via the BamHI and HindIII sites. All DNA fragments were generated by PCR using appropriate primers and integrity of the gN-expressing sequences of the resulting plasmids (gN-41sig, gN-61sig, gN-90sig) was confirmed by nucleotide sequencing.
Mutagenesis of HB5 [35] was performed using linear DNA fragments for homologous recombination. To generate BACmids containing the respective UL73 mutant sequences we used the plasmid pCPoΔUL73 [19]. This plasmid is a pCPo15 [65] derivative that contains the entire orf UL72 (nt 104560–105730, nomenclature according to Genbank Accession number X17403) at the 5′end and the entire orf UL74 (nt106095–107587) at the 3′end of the kanamycin resistance gene in plasmid pCP-o-15-Link2. Fragments containing the respective UL73 sequences were PCR-amplified from the plasmids described above and the amplimers were inserted into the 5′-flanking region of the kanamycin resistance gene in pCPoΔUL73. From these plasmids PCR fragments were generated encompassing the respective UL73-Kan-UL74 segment. Primers that were used included UL72up5 (nt 105682–105696) and UL73rec3 (nt 106271–106251). Recombination in pHB5 was done as described previously [19]. In brief, the DNA fragment was electroporated into E. coli DH10B carrying the BAC pHB5 and the plasmid pBAD for recE/T mediated recombination [66]. Bacterial colonies were selected on agar plates containing kanamycin (30 µg/ml) and chloramphenicol (30 µg/ml). To confirm the integrity of the recombined BAC, digestion of DNA with the appropriate restriction enzymes was carried out and analyzed via agarose gel electrophoresis in comparison to the parental pHB5. To confirm recombination at the predicted site Southern Blot analysis, PCR analysis as well as DNA sequence analysis of the UL72–UL75 region was performed. To remove the kanamycin resistance gene after successful recombination, plasmid pBT340 encoding the flp-recombinase was used as described [65].
To reconstitute infectious virus, MRC-5 cells (300.000 cells per well) were seeded into 6-well dishes. 48 h later 5 µg of BAC DNA together with 1 µg of pcDNApp71tag DNA (kindly provided by B. Plachter, University of Mainz, Mainz, Germany) were transfected with Superfect reagent (Qiagen) according to the manufacturer's instructions. 24 hours later culture medium was replaced by fresh medium and cells were cultivated for 7 days. Cells were then transferred to 25 cm2 flasks and cultured until a cytopathic effect was observed. The recombinant viruses were designated: RVAD169 (reconstituted from HB5), RVgN-41sig and RVgN-61sig, respectively.
Glycerol-tartrate gradient purified virus was subjected to urea-polyacrylamide-gelelectrophoresis as described previously [23]. Transfer of samples to nitrocellulose membranes was carried out by standard procedure. For visualization of antigens, gB-, gN- and gM-specific mabs were applied and detected with peroxidase-conjugated anti-mouse-IgG and anti-mouse-IgM, respectively, and the ECL detection system (Pharmacia Biotech).
The recombinant antigens used for the ELISA have been described [67]. Briefly, the following antigens were used: soluble gB (kindly provided by Sanofi-Pasteur, France); gB-AD1, containing aa 484–650 of gB; gB-AD2, containing aa 68–80 of gB; gB-AD4 containing aa 121–132 and 344–438; gH-AD86 containing aa 1–142 of gH strain AD169 and pp150 containing aa 555–705 of the tegument protein pp150 strain AD169. Proteins were diluted between 25 ng and 200 ng (depending on antigen) in 0.5 M sodium carbonate buffer, pH 9.6, or in 6 M urea (AD1) and 50 µl was used to coat microtiter plates overnight at 4°C. For the ELISA using viral lysates, wells were coated with virus lysates of 500 ng/well in 50 µl in 0.5 M sodium carbonate buffer, pH 9.6. Original virus lysates were prepared at 100 ng/µl in PBS/1% NP40. All subsequent steps were carried out at room temperature. Reaction wells were rinsed with PBS supplemented with 0.1% Tween 20 and blocked for 2 h with PBS containing 2% FCS. Plates were again rinsed with PBS supplemented with 0.1% Tween 20 and incubated with mabs or serum (50 µl/well) for 2 h. Unbound antibody was removed by washing and peroxidase-conjugated anti-human or anti-mouse IgG (Dako, Germany) was added at an appropriate dilution for 1 h. The plate was washed and 100 µl TMB (tetramethylbenzidine) peroxidase substrate, diluted 1∶1 in peroxidase substrate solution B (KPL, USA), added for 5 min. The reaction was stopped by the addition of 100 µl 1 M H3PO4 and the OD450 was determined using an Emax microplate reader (Eurofins MWG Operon, Germany). Dilution of all antibodies was done in PBS with 2% FCS. In all assays involving bacterially derived fusion proteins, the respective prokaryotic fusion partner was assayed in parallel and the optical density was subtracted from values obtained with the fusion protein.
Fibroblasts were seeded at 3×104 cells per well in 96-well plates. Virus was preincubated with individual mabs for 1 h at 37°C at concentrations ensuring complete neutralization. Cells and the virus/mab mixture were cooled to 4°C and the virus/mab mixture was added to the cells at a multiplicity of infection (m.o.i.) of 0.5. Following incubation for 1 h at 4°C, cells were washed three times with ice-cold PBS and cell lysates were prepared by freezing/thawing. DNA was extracted from the lysates using a MagNA Pure LC (Roche, Germany) instrument and quantitative real-time PCR was performed on an ABI PRISM 7500. To control for recovery of cells, copy numbers of albumine DNA was determined in parallel to HCMV and HCMV copies were calculated per 1000 copies albumine. Primers: CMV 5′:GAGCAGACTCTCAGAGGATCGG; CMV 3′: AAGCGGCCTCTGATAACCAAG; Albumine 5′: GTGAACAGGCGACCATGCT; Albumine 3′: GCATGGAAGGTGAATGTTTCAG.
Cos7 cells grown on glass coverslips in 24-well plates were transfected with 0.8 µg of plasmid DNA using Lipofectamin (Invitrogen). Fibroblasts, also grown on glass coverslips in 24-well plates, were infected with the respective viruses at a m.o.i. of 0.4. At the indicated times, the coverslips were washed and fixed in 3.0% paraformaldehyde in PBS. The fixed cells were permeabilized with PBS, 0.1% Triton X-100 for 4 min and then blocked using PBS 1% BSA for 15 min at room temperature. Primary antibodies were then added for 30 min at 37°C. Following washing, antibody binding was detected with the appropriate secondary antibody conjugated with either FITC or TRITC (Dianova). Images were collected using a Zeiss Axioplan 2 fluorescence microscope fitted with a Visitron Systems CCD camera (Puchheim, Germany). Images were processed using MetaView software and Adobe Photoshop.
|
10.1371/journal.pgen.1002649 | Transcriptional Regulation of Rod Photoreceptor Homeostasis Revealed by In Vivo NRL Targetome Analysis | A stringent control of homeostasis is critical for functional maintenance and survival of neurons. In the mammalian retina, the basic motif leucine zipper transcription factor NRL determines rod versus cone photoreceptor cell fate and activates the expression of many rod-specific genes. Here, we report an integrated analysis of NRL-centered gene regulatory network by coupling chromatin immunoprecipitation followed by high-throughput sequencing (ChIP–Seq) data from Illumina and ABI platforms with global expression profiling and in vivo knockdown studies. We identified approximately 300 direct NRL target genes. Of these, 22 NRL targets are associated with human retinal dystrophies, whereas 95 mapped to regions of as yet uncloned retinal disease loci. In silico analysis of NRL ChIP–Seq peak sequences revealed an enrichment of distinct sets of transcription factor binding sites. Specifically, we discovered that genes involved in photoreceptor function include binding sites for both NRL and homeodomain protein CRX. Evaluation of 26 ChIP–Seq regions validated their enhancer functions in reporter assays. In vivo knockdown of 16 NRL target genes resulted in death or abnormal morphology of rod photoreceptors, suggesting their importance in maintaining retinal function. We also identified histone demethylase Kdm5b as a novel secondary node in NRL transcriptional hierarchy. Exon array analysis of flow-sorted photoreceptors in which Kdm5b was knocked down by shRNA indicated its role in regulating rod-expressed genes. Our studies identify candidate genes for retinal dystrophies, define cis-regulatory module(s) for photoreceptor-expressed genes and provide a framework for decoding transcriptional regulatory networks that dictate rod homeostasis.
| The rod and cone photoreceptors in the retina are highly specialized neurons that capture photons under dim and bright light, respectively. Loss of rod photoreceptors is an early clinical manifestation in most retinal neurodegenerative diseases that eventually result in cone cell death and blindness. The transcription factor NRL is a key regulator of rod photoreceptor cell fate and gene expression. Here, we report an integrated analysis of the global transcriptional targets of NRL. We have discovered that both NRL and CRX binding sites are present in genes involved in photoreceptor function, implying their close synergistic relationship. In vivo loss-of-function analysis of 16 NRL target genes in the mouse retina resulted in death or abnormal morphology of photoreceptor cells. Furthermore, we identified histone demethylase Kdm5b as a secondary node in the NRL-centered gene regulatory network. Our studies identify NRL target genes as excellent candidates for mutation screening of patients with retinal degenerative diseases, and they provide the foundation for elucidating regulation of rod homeostasis and targets for therapeutic intervention in diseases involving photoreceptor dysfunction.
| Molecular mechanisms underlying neuronal differentiation and generation of complex sensory and behavioral circuits in the mammalian central nervous system are still poorly elucidated. Gene regulatory networks (GRNs) integrate key control elements that guide the development of distinct cell types [1], [2], [3] and contribute to precise maintenance of diverse cellular functions. As perturbations in homeostatic mechanisms (e.g., during aging and disease) can cause dysfunction or death of neurons [4], [5], a better understanding of GRNs that control neuronal homeostasis would augment the design of therapies for neurodegenerative diseases.
The rod and cone photoreceptors in mammalian retina are highly specialized neurons that transduce visual signals under dim and bright light conditions, respectively [6]. Daily renewal of almost 10% of outer segment membrane discs creates high metabolic demands, making the photoreceptors vulnerable to genetic and environmental insults [7]. Rods constitute over 95% of all photoreceptors in most mammals, including mice and humans; however, cones mediate high acuity and color vision [8]. Notably, functional impairment or loss of rod photoreceptors is an early clinical manifestation in most retinal neurodegenerative diseases that eventually results in cone cell death and blindness [9], [10], [11]. The GRNs that dictate homeostatic responses in mature rod photoreceptors have not been elucidated.
During development, rod and cone photoreceptors are produced from common pools of retinal progenitors under the control of multiple transcription factors and regulatory signaling pathways [11], [12], [13]. Furthermore, the basic motif-leucine zipper protein NRL is the dominant transcription factor that determines rod photoreceptor cell fate. In Nrl−/− mice, all post-mitotic cells originally fated to become rods instead generate a cone-only photoreceptor layer [14], whereas ectopic Nrl expression in photoreceptor precursors produces a rod-only retina [15]. Interestingly, knock-in mice where Nrl is replaced by thyroid hormone receptor β2 (Trb2) have an M-cone dominant retina, but the presence of both NRL and TRb2 yields a normal contingent of rods [16]. A key transcriptional target of NRL is the orphan nuclear receptor NR2E3 that primarily represses cone genes to establish rod identity [17], . The cone-rod homeobox CRX is another essential transcriptional activator of photoreceptor-specific genes as rods and cones in Crx−/− mice do not develop outer segments and eventually die [20], [21], [22]. NRL and CRX continue to be expressed at high levels in mature retina and in rod photoreceptors ([23]; Gotoh, Swaroop et al. unpublished data). Protein interaction and transcriptional activation assays, combined with expression profiling of knockout mice, demonstrate that NRL and CRX are the two major regulators of rod photoreceptor gene expression [24], [25], [26], [27].
We hypothesize that detailed mapping of a rod-specific GRN would lead to the development of better therapeutic interventions in blinding diseases involving photoreceptor degeneration. Here we report the genomewide NRL in vivo occupancy in adult mouse retina by chromatin immunoprecipitation followed by high-throughput sequencing (ChIP–Seq) using Illumina and ABI sequencing platforms. We perform an integrated analysis by coupling the NRL ChIP–Seq data with published photoreceptor-specific transcriptional profiles and CRX ChIP–Seq results. We use in vivo knockdown assays to examine the physiological relevance of NRL target genes and identify secondary regulatory nodes downstream of NRL in rod transcriptional hierarchy. Our studies establish NRL and CRX as the key regulatory nodes for rod-expressed genes, identify NRL targets as candidate genes for retinal diseases, and provide a framework for GRN that controls homeostasis in rod photoreceptors.
We performed chromatin immunoprecipitation experiments using anti-NRL antibody (with normal IgG as a control) to pull down the genomic fragments bound by NRL in vivo in adult mouse retina. The ChIP DNA was subjected to direct high-throughput sequencing using either Illumina 1G genome analyzer or ABI/SOLiD system (ABI). The workflow for the analysis of two datasets is shown in Figure 1A. (see www.nei.nih.gov/intramural/nnrldataresource.asp for raw sequence reads). Illumina and ABI datasets contained a total of 8 million 25-bp reads and 18.0 million 35-bp reads, respectively. Of these, respectively 5.3 million (66.3%) and 6.3 million (35%) reads were uniquely mapped to the mouse genome (NCBI Build 37, UCSC mm9), with overlapping sequence reads forming the NRL ChIP–Seq peaks (Table 1). We used NGS Analyzer (Genomatix) and MACS [28], in parallel, to determine NRL ChIP–Seq peaks with ChIP–Seq counts from negative control (normal IgG) libraries as thresholds. The peaks identified by both algorithms (intersected peaks) were kept for further analyses (Figure 1A, Table 2). Illumina and ABI platforms revealed 2790 and 5625 NRL ChIP–Seq peaks, respectively (Table 2). The number of peaks did not correlate with chromosome size (data not shown), indicating the interaction of NRL with specific genomic regions. The average peak widths were 398 bp (Illumina) and 408 bp (ABI), with average peak heights being 58.0 (Illumina) and 79.9 (ABI) and median peak heights of 30 (Illumina) and 47 (ABI) (Table 2). Illumina and ABI ChIP–Seq peak centers showed a strong correlation (Figure 1B), with almost 90% of Illumina peaks overlapping with ABI peaks (Figure 1B).
A large number of NRL ChIP–Seq peaks were mapped within 1 kb of the transcription start sites (Figure 1C, Table S1). Furthermore, over 70% of NRL ChIP–Seq peaks were present within 10 kb of 7–10% of mouse gene promoters (see Venn diagram in Figure 1C). The NRL ChIP–Seq peaks from both Illumina and ABI platforms are highly enriched in promoter regions, given that promoters only account for approximately 2% of the mouse genome (Figure 1D).
In order to identify physiologically relevant NRL target genes, we examined Illumina and ABI ChIP–Seq data in combination with global expression profiles of flow-sorted photoreceptors from wild type (WT) and Nrl−/− mouse retina [25]. Of 2143 genes associated with Illumina ChIP–Seq peaks, 216 exhibited at least 1.5 fold less expression and 80 genes showed higher expression in Nrl−/− photoreceptors (Figure 2A). Of 4085 genes associated with ABI ChIP–Seq data, we identified 291 genes with lower and 131 genes with higher expression in the Nrl−/− photoreceptors (Figure 2A). A combined analysis of Illumina and ABI ChIP–Seq datasets yielded 281 genes showing altered expression in Nrl−/− photoreceptors. A high correlation was detected between NRL ChIP–Seq peaks (from both Illumina and ABI datasets) and promoters of genes that are differentially expressed in rod photoreceptors of WT versus Nrl−/− retina (Figure 2B). For convenience, we will refer genes associated with NRL ChIP–Seq peaks and altered in Nrl−/− retina as direct transcriptional targets of NRL.
As transcription factor interactions determine the specificity of gene expression patterns [29], [30], we performed motif enrichment analysis (Genomatix RegionMiner, “Over-represented transcription factor binding sites” based on MatInspector [31], [32]) of sequences under the NRL ChIP–Seq peaks. As predicted, we noticed a significant enrichment of the binding sites for NRL and other AP1 related factors (AP1R) in peaks associated with genes that are up- or down-regulated in the absence of Nrl (Table S2).
An unbiased motif enrichment analysis of ChIP–Seq peak regions for NRL targets revealed binding sites for transcription factor families that include key photoreceptor regulatory proteins – CRX (BCDF family) [20], [24], [33], NR2E3 (NR2F family) [18], [19], [34], [35], RORβ (RORA family) [36], [37], ESRRβ (EREF family) [38] and MEF2C (MEF2 family) [39] (Table S2). Motifs for these transcription factors were significantly enriched within total NRL ChIP–Seq peaks and within the peaks associated with genes that are differentially expressed in Nrl−/− photoreceptors (except for MEF2C in ABI data) (Figure 2C). The motifs for AP1R (NRL), BCDF (CRX), RORA (RORβ) and EREF (ESRRβ) families were located close to the peak center whereas motifs for NR2E3 and MEF2C were not (Figure 2C). The composition and enrichment ranking of enriched transcription factor motifs were different between NRL target genes whose expression was down- or up-regulated in Nrl−/− photoreceptors (Table S2), suggesting that NRL cooperated with different proteins to activate or repress gene expression. However, the genomic distribution of NRL peaks is similar among the various groups (Figure S1).
As CRX is an established transcriptional activator of photoreceptor genes [21] and is shown to interact with NRL [24], we integrated NRL ChIP–Seq peaks with the previously published CRX ChIP–Seq data [22]. Interestingly, 65% of NRL ChIP–Seq peaks obtained from Illumina and 48% of those from ABI overlapped with the CRX peaks (Figure 2D), consistent with a previous finding that 51% of the down-regulated genes in Nrl−/− mice exhibit reduced expression in Crx−/− retina as well [40]. Motif enrichment analysis of NRL-CRX-overlapping peak regions and of non CRX-overlapping peaks revealed AP1R binding site (NRL) as the only common motif (Table S3). The motifs for other photoreceptor transcription factors ESRRβ, RORβ, NR2E3 were only enriched in NRL-CRX-overlapping peaks, and the most enriched motif for non CRX-overlapping NRL peaks was for SP1 family proteins (Table S3).
Ontology analysis revealed distinct biological functions for genes that were associated with NRL-CRX-overlapping ChIP–Seq peaks (involved in photoreceptor function) versus genes associated with non CRX-overlapping NRL peaks (basic cellular processes) (Table S4).
We first checked Illumina and ABI ChIP–Seq data for a few established NRL target genes that are involved in rod development or function (Figure 3). In addition to the reported NRL-binding sequences (at −75 bp for Rho and −3.5 kb for Nr2e3) [41], [42], [43], ChIP–Seq data further identified binding sites for NRL in Rho at −3 kb and −1.5 kb and in Nr2e3 at −1 kb and −100 bp. We also detected NRL binding in rod-specific genes (such as Pde6a, Gnat1) and Esrrb, an important regulator of rod gene expression [38]; the expression of these genes is decreased significantly in Nrl−/− mice. In Esrrb, we identified strong NRL binding to the second intron. Interestingly, a strong NRL ChIP–Seq peak was observed within an intron of the Nrl gene in addition to a peak in the promoter region. Kdm5b and Wisp1 are among additional genes that are regulated by NRL and play a role in rod homeostasis (see later). NRL also binds to cone-specific genes and may contribute to their down-regulation to maintain a rod phenotype, as proposed previously [11], [41].
In general, ABI ChIP–Seq peaks were higher than Illumina peaks although uniquely mapped reads in the two libraries were comparable (5.3 million vs 6.3 million) (Figure 3). Even though ABI data produced more peaks (e.g., Kdm5b and Nrl), Illumina data detected unique peaks that were not present in ABI (e.g., Esrrb) (Figure 3). We then plotted CRX ChIP–Seq peaks [22] relative to NRL peaks.
We then performed ChIP-qPCR validations for a number of known and novel NRL targets. To strictly control the ChIP-qPCR analyses, we used two sets of controls: normal IgG as an antibody control and retina from Nrl −/− mice as a tissue control. We compared ChIP-qPCR signals between anti-NRL antibody and normal IgG using WT mouse retina, and performed additional NRL ChIP analysis using WT and Nrl−/− mouse retina (Figure 4). The two sets of experiments were highly concordant and validated the ChIP–Seq findings for all 26 sites (with various ChIP–Seq peak heights) that were tested. ChIP-qPCR analysis did not detect the association of NRL with 5 genomic regions that did not include ChIP–Seq peaks (Figure 4).
To further test the functional relevance of NRL genome occupancy detected by ChIP–Seq, we generated enhancer-reporter constructs by cloning 26 randomly chosen ChIP–Seq peak regions (with a linear range of peak tags) and 5 non-peak genomic fragments of comparable sizes upstream of an SV40 basal promoter and a luciferase reporter gene. Of 26 NRL ChIP–Seq regions, at least 19 included CRX ChIP–Seq peaks. Five non-peak genomic fragments (3′Rho, 3′Pde6b, Gapdh, Hprt, Oct4) were negative for CRX peaks. Co-transfection of mouse NRL expression plasmid in HEK293T cells increased the luciferase reporter expression from all 26 enhancer constructs containing NRL ChIP–Seq peaks, but not from the 5 constructs containing non-peak fragments (Figure 5). Our data suggest that the genomic fragments spanning NRL ChIP–Seq peaks can function as enhancer elements and mediate NRL-driven transcriptional activation of target genes.
We also cloned and tested NRL peak regions associated with four cone genes (Gnat2, m-Opsin, Gngt2 and Pik3ap1) using the same reporter assay (2). Co-transfection of NRL expression plasmid increased the luciferase reporter expression from these enhancer constructs as well (Figure S2), validating the primary function of NRL as a transcriptional activator. However, we can not exclude the function of NRL in directly repressing cone genes in vivo as it may require interaction with native promoters and cis-elements, recruitment of appropriate cofactors, and/or native chromatin context, which are not provided in HEK293 cells.
We hypothesized that NRL target genes would contribute to rod photoreceptor homeostasis, and their abnormal regulation could lead to photoreceptor dysfunction and/or degeneration. We therefore integrated the chromosomal location of the human orthologs of NRL target genes with mapping information for human genetic loci for retinal diseases (RetNet http://www.sph.uth.tmc.edu/retnet/). We identified 21 NRL target genes that are known to be associated with retinal diseases involving photoreceptor degeneration (Table S5). Furthermore, almost 100 human NRL target genes map within the critical region of 29 as yet uncloned retinal disease loci (Table S5).
To directly examine the physiological function of 16 NRL target genes, we knocked down the expression of target genes by transfecting shRNA plasmids in vivo into the P0 mouse retina [44], [45]. For each target gene, three shRNA expression constructs were first evaluated for knockdown efficiency using a sensor construct in HEK293T cells (Figure S3). The most efficient shRNA was then used for in vivo knockdown experiments in the mouse retina, which were examined seven or twenty days (at P7 or P20) after electroporation (Figure 6, Figure 7, and Figures S4, S5). A GFP-expression plasmid (Ub-GFP) was co-transfected to mark the transfected retinal cells. Based on putative function and/or involvement in retinal disease (see Table S5), we selected 16 genes – Bach2, Cdr2, Dusp12, Esrrb, Gpsm2, Haus1, Kdm5b, Lman1, Lrp11, Lrrc2, Ncoa2, Plekha2, Ppargc1b, Trim36, Wisp1 and Zdhhc14. Eight of the genes have overlapping CRX ChIP–Seq peaks.
We consistently observed, in multiple biological replicates, smaller numbers of GFP+ cells in P20 retina that was transfected with shRNA against NRL target genes compared to the retina expressing control Gapdh shRNA (Figure 6, Figure 7, and Figure S5). The reduction in the number of GFP+ cells was more pronounced at P20 than at P7, and was most severe in retina transfected with Wisp1 shRNA, which led to a near total and consistent loss of GFP+ cells at P20. Thus, the function of a majority of NRL targets appears to be required for functional maintenance of photoreceptors.
In addition to the reduced number of GFP+ cells, the knockdown of Kdm5b, Lman1, or Wisp1 resulted in an abnormal morphology of the transfected photoreceptors at P20, including the abnormal location of their cell bodies (Figure 6A, 6B, and 6D) and short outer segments (Figure 6A, 6B, and 6E). The cell bodies of the GFP+ cells were positioned in the outer portion of the outer nuclear layer (ONL), reminiscent of cone nuclei [46], instead of spanning across the ONL (Figure 6).
To validate the specificity of knockdown data and rule out the possibility of general toxic effects of shRNA, we produced degenerate cDNA (dcDNA) constructs for two of the target genes (Lman1 and Wisp1) containing silent mutations that conferred resistance to shRNA mediated mRNA degradation. Co-transfection of Gapdh shRNA with dcDNA for Lman1 did not manifest a retinal phenotype, and more importantly, Lman1 dcDNA co-transfection rescued all of the Lman1 shRNA phenotypes in the retina (including the reduced number of GFP+ cells, cell body location and OS length) (Figure 7). Co-transfection with Wisp1 dcDNA also corrected the reduction of GFP+ cells; however, its overexpression led to a decrease in GFP+ cells (Figure 7), indicating that endogenous WISP1 levels are carefully controlled.
We were particularly intrigued by one of the NRL targets – Kdm5b (see Figure 6), which encodes lysine (K)-specific demethylase 5b, an enzyme that catalyzes the demethylation of active histone marks at methylated H3K4; thus, Kdm5b is involved in chromatin remodeling and functions as a transcriptional repressor [47], [48], [49]. To investigate its potential role as a second order node in photoreceptor GRN downstream of NRL, we dissociated the retina 20 days after knocking down Kdm5b or Gapdh expression by shRNA electroporation at P0, flow-sorted the electroporated cells, prepared total RNA, and performed global expression profiling using Affymetrix exon arrays (Figure 8A). Kdm5b knockdown resulted in up-regulation of 311 genes and down-regulation of 619 genes when compared to Gapdh knockdown. We detected 57 genes that are down-regulated and 20 that are up-regulated in both Kdm5b knockdown and Nrl−/− retina (Figure 8B), suggesting that some of the effects of loss of NRL (in Nrl−/− retina) are mediated through decreased Kdm5b expression. Some of the genes (e.g., Pde6a, Pde6b, Guca1b, Pde6c, Cngb3, Opn1sw) altered by Kdm5b knockdown are associated with the visual transduction, while a few others (Gadd45a, H2afz and Suv39h2) are associated with chromatin organization [50], [51], [52].
Visual impairment in a vast majority of retinal and macular degenerative diseases can be attributed to dysfunction or death of photoreceptors [7], [10], [11]. Despite the central role of cones in transduction of vision in humans, rods constitute 95% of all photoreceptors and are generally the first to die in retinal neurodegeneration. A relatively late onset of clinical manifestations in these diseases underscores the importance of stringently maintaining the function of highly metabolically active photoreceptors. The control of homeostasis must be exerted at multiple levels as quantitatively precise expression of phototransduction proteins and their transport to the modified sensory cilia (outer segments) are critical for photoreceptor survival. In addition to its essential role in photoreceptor differentiation, NRL has been implicated in the regulation of rod phototransduction genes, such as rhodopsin and cGMP phosphodiesterase α and β subunits [24], [27], [53], [54]. Here we identify global transcriptional targets of NRL and integrate our data with reported targets of CRX, another key regulator of photoreceptor genes. Our results show that NRL and CRX together control the expression of most, if not all, genes involved in rod phototransduction through a cis-regulatory module, which also includes the binding sites for NR2E3, ESRRβ, RORβ and in some cases MEF2C. Equally important is the finding that non-CRX containing NRL cis-regulatory modules fine-tune the expression of additional photoreceptor-expressed genes, which may contribute to high metabolic demand in rod photoreceptors.
ChIP–Seq has emerged as a cost effective, high-throughput technology for high-resolution genome-wide mapping of in vivo locations for chromatin modifications and transcription factor binding [55], [56], [57], [58]. Despite the fundamental difference in sequencing chemistry and nucleotide base calling software between the Illumina and ABI/SOLiD sequencing platforms [59], [60], our ChIP–Seq data from the two are remarkably comparable, further validating the in vivo NRL binding events reported here. In addition to enrichment in promoter regions, a number of NRL ChIP–Seq peaks are detected in intronic regions of annotated genes; some of these might reflect alternative promoter usage in photoreceptors as reported recently for Mef2c [39].
We previously proposed that photoreceptor precursors have a “default” S-cone fate and a “tug-of-war” among a selected few transcription factors specifies rod versus cone cell type [11]. NRL and TRβ2 respectively initiate the rod and M-cone pathways [16], with NRL being the dominant activator of rod genes and a suppressor of cone genes together with its target NR2E3 [15], [41]. Enrichment of a distinct set of transcription factor binding sites in NRL ChIP–Seq peaks in genes that are down- or up-regulated in Nrl−/− retina suggests specific and discrete cis-regulatory modules for rod versus cone photoreceptor expressed genes. CRX strongly activates the expression of both rod and cone genes [21], [22], [61]. An overlap of CRX peaks in over 50% of NRL ChIP–Seq peaks is consistent with their synergistic function in activating rod-expressed genes. Indeed, all rod phototransduction genes were included in this group. Notably, CRX ChIP–Seq peaks are much smaller than NRL peaks at the same loci and loss of NRL leads to more significant decrease in gene expression than in Crx−/− retina, suggesting a fundamental role of NRL in regulating rod genes. CRX likely enhances rod gene expression by altering the chromatin conformation via recruitment of histone acetylases [62]. In cone genes (up-regulated in Nrl−/− retina), binding of both CRX and NRL is consistent with the common photoreceptor precursor hypothesis [11], [16]. Additional studies (e.g., histone modifications) are needed to clarify differential regulation of specific genes by NRL and CRX in rod versus cone photoreceptors.
Like many key transcription factor nodes in GRNs [63], [64], [65], NRL likely auto-regulates its own expression as suggested by strong NRL ChIP–Seq peaks in Nrl promoter and intronic regions. While the key role of NR2E3 as a secondary node downstream of NRL is to repress cone-specific genes [17], [18], [41], two newly reported NRL targets – ESRRβ and MEF2C – function as transcriptional regulators for activation and/or maintenance of rod gene expression [38], [39]. A new secondary node in rod GRN that our studies identified is KDM5B (also called Jarid1b), a Jumonji-domain containing histone demethylase, which is associated with chromatin remodeling and transcriptional repression [47], [48]. KDM5B reportedly activates the expression of self-renewal-associated genes by suppressing cryptic initiation and maintaining proper H3K4me3 gradient for productive transcriptional elongation [66]. We observe a significant overlap between the genes altered by loss of NRL and KDM5B, indicating a broader role of KDM5B in regulating rod homeostasis downstream of NRL. We hypothesize that differential expression of KDM5B may contribute to chromatin organization and metabolic differences between rod and cone photoreceptors [8], [46], [67], [68].
Retinal and macular diseases are genetically heterogeneous with over 200 mapped loci; of these, almost 150 genes have been identified (http://www.sph.uth.tmc.edu/Retnet/). A catalog of genome-wide NRL targets with overlapping CRX binding sites, reported here, provides excellent candidate genes for mutation screening in patients with inherited retinal neurodegenerative diseases. We have listed almost 100 genes (see Table S5) that map to retinal disease loci. Interestingly, knockdown of 16 target genes, reported in this study, resulted in photoreceptor cell death or abnormal morphology, highlighting the importance of NRL targets in maintaining normal physiology and the association of perturbed target gene expression with retinal diseases.
A key aspect of photoreceptor homeostasis is the daily renewal of almost 10% of outer segment membrane discs, which requires a stringent control of the synthesis of specific phototransduction proteins and lipid molecules. Therefore, the target gene, Lman1, attracted our attention as its knockdown led to shorter photoreceptor outer segments and abnormal location of cell bodies (close to the sclera), which is characteristic of cone photoreceptors or late-born rods, whereas the early-born rods locate towards the vitreous side. LMAN1 participates in transport between the endoplasmic reticulum and Golgi [69]. Our data suggests that LMAN1 performs critical roles in photoreceptor homeostasis by controlling lipid homeostasis and/or biogenesis of membrane discs. Abnormal location of nuclei to scleral side in photoreceptors after its knockdown by in vivo electroporation could be due to rod to cone transformation in the absence of NRL, or delayed rod birth as a result of abnormal signaling for rod fate determination.
Wisp1, another interesting target of NRL, encodes the Wnt1-inducible signaling pathway protein 1 that exerts cytoprotective and/or growth promoting effects [70], [71] by repressing p53 and activation of Akt kinase [72]. WISP1 could therefore act as a survival or maintenance factor for photoreceptors. Further investigations on WISP1 may yield new targets for neuroprotective strategies in retinal degeneration.
Gene regulatory networks (GRN) control multiple pathways during development and homeostasis and provide conceptual framework for elucidating disease mechanisms [2], [73]. Transcription factors reside near the top of GRNs; their abnormal expression and/or activity can cause widespread changes in target genes [74], [75]. Our studies demonstrate a pivotal role of NRL in controlling rod homeostasis by modulating the expression of numerous target genes, which in turn maintain distinct aspects of cell function and survival. Elucidation of combinatorial regulation of genes by NRL and its co-regulators (specifically CRX) and identification of distinct downstream nodes (such as KDM5B) provide a framework to construct GRN for functional maintenance in mammalian rod photoreceptors (Figure 9).
All animal work must have been conducted according to relevant national and international guidelines. Animal Care and Use Committee of the National Eye Institute approved all mouse protocols. See also Text S1.
Retina from postnatal day (P) 28 C57Bl/6J mice was used for ChIP experiments with NRL antibody or normal IgG, as previously described [41]. Fifteen or 25 ng of ChIP DNA from parallel experiments was used for library preparation and sequencing on Illumina 1G Genome Analyzer or ABI SOLiD V2 system, respectively.
Raw sequencing reads from Illumina or ABI platforms were mapped to the mouse genome (NCBI build 37) using Genomatix Mining Station (GMS).
ChIP–Seq peaks were called using MACS [28] and NGS-Analyzer (Genomatix). The union of overlapping peak regions from both methods was used in subsequent analyses.
The NRL ChIP–Seq peaks were compared to the CRX ChIP–Seq regions [22] using GenomeInspector (Genomatix; [31]).
Affymetrix microarray data from flow-sorted photoreceptors of WT and Nrl−/− retina [25] was analyzed using ChipInsepector (Genomatix).
Transcription factor binding site enrichment analyses of sequences in ChIP–Seq peak regions were performed using RegionMiner and MatInspector (Genomatix; [31], [32]). The ChIP–Seq peaks were extended to −500 bp and +500 bp from the peak center. The sequences were scanned for TFBS matrices (Genomatix MatBase version 8.2) using MatInspector (Genomatix). Positional bias (P) was calculated [76], and the –log(P) was plotted against the scan windows' mid-positions. The over-represented TFBS positions for a TF family appear as peaks in these plots.
ChIP DNA was tested in triplicates by qPCR using SYBR Green [77]. We randomly tested 26 regions with peaks covering the majority range of the peak heights. Five regions without ChIP–Seq signals served as negative controls. Normal IgG served as the negative antibody control, and Nrl−/− retina was used as a negative tissue control. The complete ChIP-qPCR procedure was performed twice.
HEK293T cells were cultured in DMEM and transfected with Fugene 6 (Roche).
To generate shRNA-resistant dcDNA constructs, silent mutations that confer resistance to shRNA were introduced into Lman1 cDNA and Wisp1 cDNA using Quikchange kit (Stratagene).
To generate enhancer constructs, ChIP–Seq peak regions were amplified and cloned into pGL3-promoter vector (Promega). HEK293T cells were transfected with these enhancer constructs, a transfection control plasmid expressing Renilla luciferase (Promega), and NRL expression plasmid or empty vector. The luciferase activities were measured 48 hr after transfection. The experiments were performed three times.
To generate shRNA-sensor constructs for efficiency test, we cloned the shRNA target sequences into the 3′UTR of a GFP vector. The shRNA-sensor construct and CAG-HcRed (transfection control) were co-transfected with either shRNA against target or Gapdh shRNA, For each target gene, three shRNA constructs were evaluated for efficacy indicated by a decrease in GFP. The most efficient one was chosen for in vivo knockdown experiments.
shRNA alone or together with shRNA-resistant dcDNA was introduced in the retina of CD-1 P0 mouse pups by sub-retinal injection followed by in vivo electroporation, as previously described [44], [45]. The retinas were harvested at P7 or P20 for histology or immunohistochemistry.
Mouse retina was electroporated at P0 with Ub-GFP and Gapdh shRNA or Kdm5b shRNA and dissected at P20. GFP+ retinal cells were isolated from dissociated retina by FACS (FACSAria; BD Biosciences). RNA was extracted and cDNA was synthesized followed by sense transcript cDNA (ST-cDNA) generation using WT-Ovation Exon module (NuGEN Technologies). The ST-cDNA was fragmented and labeled with Encore Biotin Module (NuGen) and used for hybridization with GeneChip Mouse Exon 1.0 ST array (Affymetrix). The microarray data has been deposited in the Gene Expression Omnibus Database (accession #: will be available soon).
|
10.1371/journal.ppat.1003861 | STING-Dependent Type I IFN Production Inhibits Cell-Mediated Immunity to Listeria monocytogenes | Infection with Listeria monocytogenes strains that enter the host cell cytosol leads to a robust cytotoxic T cell response resulting in long-lived cell-mediated immunity (CMI). Upon entry into the cytosol, L. monocytogenes secretes cyclic diadenosine monophosphate (c-di-AMP) which activates the innate immune sensor STING leading to the expression of IFN-β and co-regulated genes. In this study, we examined the role of STING in the development of protective CMI to L. monocytogenes. Mice deficient for STING or its downstream effector IRF3 restricted a secondary lethal challenge with L. monocytogenes and exhibited enhanced immunity that was MyD88-independent. Conversely, enhancing STING activation during immunization by co-administration of c-di-AMP or by infection with a L. monocytogenes mutant that secretes elevated levels of c-di-AMP resulted in decreased protective immunity that was largely dependent on the type I interferon receptor. These data suggest that L. monocytogenes activation of STING downregulates CMI by induction of type I interferon.
| Current vaccines are successful at generating neutralizing antibodies, however there is a pressing medical need to find adjuvants that yield long-lived memory T cells. Immunization with the bacterium Listeria monocytogenes induces a robust protective immune response mediated by cytotoxic lymphocytes that are efficient at killing infected cells upon reinfection. When L. monocytogenes enters a cell, it secretes the small molecule cyclic diadenosine monophosphate (c-di-AMP), which activates the host protein STING leading to a type I interferon response. In this study, we tested whether STING activation plays a role in the generation of cytotoxic lymphocytes and protective immunity using a mouse immunization model. We found that in the absence of STING signaling mice restricted bacterial growth and maintained higher numbers of cytotoxic lymphocytes upon reinfection, whereas mice immunized in the presence of elevated levels of c-di-AMP were less protected. These results suggest that the inflammation induced by a bacterial pathogen can be detrimental to the development of adaptive immunity, which could provide new insights into vaccine development.
| Cell-mediated immunity (CMI) is a critical component for protection against intracellular pathogens. Upon infection, the innate immune response provides resistance and initiates the development of antigen-specific lymphocytes including cytotoxic CD8+ T cells, which ultimately kill host cells harboring pathogens [1]. The Gram-positive bacterium Listeria monocytogenes has been used for decades as a model organism to investigate the generation of CMI, as infection induces a robust effector and memory CD8+ T cell response that restricts bacterial growth following a lethal secondary challenge, resulting in long-lived sterilizing immunity [2]. Although it is generally agreed that activation of the innate immune system is critical for the initiation of adaptive immunity [3], the specific signaling pathways necessary to elicit a robust protective immune response to L. monocytogenes remain poorly understood.
L. monocytogenes is detected by multiple innate immune signaling pathways during infection [4]. Following engulfment by macrophages and dendritic cells, the bacteria reside within phagosomes where they are detected by Toll-Like Receptors (TLRs), resulting in the activation of MyD88-dependent response genes [5]. By secreting a pore-forming cytolysin, listerolysin O (LLO), L. monocytogenes escapes into the cytosol where it replicates and polymerizes actin to facilitate cell-to-cell spread [6]. L. monocytogenes is detected by several cytosolic innate immune pathways leading to a cytokine profile distinct from that of LLO-deficient bacteria, which are restricted to the phagosome [5], [7].
The primary cytosolic sensor of L. monocytogenes is STING (stimulator of interferon (IFN) genes, also known as MPYS, MITA and ERIS), an ER-localized transmembrane protein [8]. STING is activated by cyclic dinucleotides (CDNs) that are either produced by a pathogen or by an endogenous cyclic GMP-AMP synthase that is activated by DNA [9], [10]. Direct binding of CDNs to STING activates a downstream signaling cascade involving TBK1 and IRF3 [11], [12], [13]. In the case of L. monocytogenes, cyclic diadenosine monophosphate (c-di-AMP) is secreted through bacterial multi-drug efflux pumps, leading to STING activation and transcription of IFN-β and co-regulated genes [14], [15]. STING-deficient macrophages or mice are unable to produce IFN-β in response to L. monocytogenes infection indicating that STING is required for the type I IFN response to L. monocytogenes [16], [17].
Purified CDNs are immunostimulatory in vitro and in vivo. Murine and human dendritic cells exposed to cyclic diguanosine monophosphate (c-di-GMP) or c-di-AMP exhibit enhanced surface expression of costimulatory markers and T cell proliferation. Mice mount a significant antibody response following co-administration of protein antigens with c-di-GMP or c-di-AMP [13], [18], [19], [20]. CDNs also stimulate cellular immune responses. Antigen-stimulated splenocytes from mice immunized with β-ααgalactosidase in the presence of c-di-GMP or c-di-AMP proliferate and secrete cytokines [18], [19]. These data indicate that CDNs are sufficient to elicit a cell-mediated adaptive immune response.
Entry of L. monocytogenes into the host cytosol is necessary to generate secondary protective immunity, as phagosome-restricted heat-killed or LLO-deficient bacteria do not elicit functional cytotoxic T cells and long-term memory responses [21], [22], [23]. The attenuated ActA-deficient mutant strain, which escapes the phagosome but fails to polymerize actin and spread to neighboring cells, is fully immunogenic to mice [24]. Furthermore, MyD88-deficient mice, while highly susceptible to acute infection with virulent L. monocytogenes, are fully protected following secondary lethal challenge when immunized with the ActA-deficient mutant [25], [26], [27], [28]. These findings suggest that phagosomal detection of L. monocytogenes during immunization is not sufficient for the development of protective immunity.
STING activation induces an array of IRF3-dependent genes [5] as well as NF-κB and STAT6-dependent genes [29], [30]. Since LLO-deficient bacteria fail to enter the cytosol and induce STING-related genes [5], [7], we hypothesized that the detection of L. monocytogenes by STING is required for CMI. In this study, we tested whether STING signaling plays an important role in the generation of protective immunity to L. monocytogenes.
In the model of protective immunity used in these studies, mice were immunized with an attenuated yet immunogenic strain of L. monocytogenes that lacks the actA and inlB genes (ActA−Lm) and challenged 30–38 days later with 2LD50 (2×105 colony forming units (CFU)) of wild-type L. monocytogenes (WT Lm). Previous studies typically immunize mice with 0.1LD50 of L. monocytogenes (1×107 CFU ActA−Lm for C57BL/6 mice) [21]. At this high immunization dose, bacterial burdens in subsequently challenged mice are below the limit of detection. In contrast, a lower immunization dose of 103 CFU (∼0.00001LD50 for C57BL/6 mice) still generated significant immunity as compared to naïve mice, but did not induce saturating immunity and thus revealed differences that might be missed using higher doses (Fig. S1A).
To determine whether STING signaling is required for the generation of protective immunity to L. monocytogenes, mice lacking STING (Goldenticket, Gt) were immunized with 103 CFU of ActA−Lm expressing ovalbumin (ActA−Lm-OVA) and challenged 30–38 days later with 2LD50s of WT Lm-OVA. Surprisingly, whereas naïve STING-deficient mice had similar bacterial burdens as naïve C57BL/6 (B6) mice, immunized STING-deficient mice had approximately 1.5–2 logs fewer bacteria in spleens and livers compared to immunized B6 mice (Fig. 1A). At higher immunization doses (104 and 105 CFU), the majority of B6 and STING-deficient mice had bacterial numbers below the limit of detection in the spleen and thus no significant differences could be observed (Fig. S1B). Since cytotoxic CD8+ T cells are the major mediator of L. monocytogenes clearance following secondary challenge [2], [31], the number of OVA-specific CD8+ T cells were measured by staining splenocytes with a MHC class I restricted OVA tetramer (Kb/OVA257–264). STING-deficient mice had significantly higher total numbers of OVA-specific CD8+ T cells compared to B6 mice (Fig. 1B). These data suggested that STING-deficient mice exhibited enhanced immunity.
STING stimulation leads to the activation of the transcription factor, IRF3 [12]. In addition, IRF7 contributes to IFN-α production in response to L. monocytogenes in vivo [32]. To determine the role of these downstream effectors of STING signaling, mice lacking IRF3 (IRF3−/−) or IRF3 and IRF7 (IRF3/7−/−) were examined for protective immunity. Compared to B6 mice, IRF3-deficient mice had less bacteria in the livers and IRF3/7-deficient mice had less bacteria in the spleens and livers (Fig. 1A). Both groups had significantly higher numbers of OVA-specific CD8+ T cells than B6 mice (Fig. 1B), suggesting that these transcription factors contribute to the STING-mediated decrease in immunity.
To evaluate whether T cells from mice lacking STING signaling possessed effector functions, splenocytes from ActA−Lm-OVA-immunized mice were stimulated with either the MHC class I-restricted peptide OVA257–264, or the MHC class II-restricted peptide LLO190–201 from L. monocytogenes and measured for IFN-γ, TNF-α and IL-2 production by intracellular cytokine staining and flow cytometry. STING-deficient mice had significantly higher numbers of polyfunctional IFN-γ-, TNF-α- and IL-2- producing CD8+ and CD4+ T cells compared to B6 mice (Fig. 1C). IRF3/7-deficient but not IRF3-deficient mice also had higher numbers of IFN-γ- and TNF-α-producing CD8+ T cells. These data suggested that in the absence of STING or its downstream transcription factors IRF3 and IRF7, CD8+ T cell expansion, cytokine production and bacterial clearance was enhanced.
L. monocytogenes activates TLRs which signal via the adaptor MyD88 [4]. It is possible that both MyD88- and STING-dependent response genes play redundant roles in generating protective immunity to L. monocytogenes. To test this hypothesis, we bred mice lacking both MyD88 and STING (MyD88−/−Gt). Bone marrow-derived macrophages (BMMs) from MyD88/STING-deficient mice infected with WT Lm had low or non-detectable expression of the cytokines IFN-β, IL-12 p40, TNF-α and IL-6 (Fig. S2A).
To examine whether the loss of both the MyD88 and STING signaling pathways affect bacterial clearance during an acute infection, B6, MyD88- and MyD88/STING-deficient mice were infected with WT Lm. MyD88/STING-deficient mice had similar bacterial burdens as MyD88-deficient mice early after infection, however by day 3, had significantly higher CFU in the spleen and the liver (Fig. S2B). Although mice lacking MyD88 cannot survive infection with WT Lm, MyD88/STING-deficient mice died earlier than MyD88-deficient mice (data not shown). These data indicated that in the absence of MyD88, STING contributes to L. monocytogenes clearance during an acute response.
To determine whether the loss of MyD88 and STING affects initiation of the adaptive response to L. monocytogenes, the upregulation of costimulatory molecules on splenic dendritic cells from immunized B6, STING-, MyD88- and MyD88/STING-deficient mice was measured. Surface expression of CD86 and CD40 was decreased in MyD88- and STING-deficient mice and further reduced in the MyD88/STING-deficient mice, suggesting an additive effect of these two pathways (Fig. 2A). Upregulation of the activation marker CD69 on CD8+ and CD4+ T cells was also decreased in both MyD88- and STING-deficient mice and ablated in MyD88/STING-deficient mice (Fig. 2B). Furthermore, IL-6, TNF-α and MCP-1 in the serum were mostly dependent on MyD88 but further reduced in the MyD88/STING-deficient mice. IL-12 p70 was non-detectable in either MyD88- or MyD88/STING-deficient mice (Fig. 2C). These data indicated that mice lacking MyD88 and STING signaling had significantly reduced dendritic and T cell activation and cytokine production in vivo in response to L. monocytogenes immunization.
Mice lacking both MyD88 and STING were tested for the ability to develop protective immunity. Similar to B6 and MyD88-deficient mice, MyD88/STING-deficient mice were protected following secondary lethal challenge and showed no signs of disease unlike naïve mice which were either moribund or dead at the time of sacrifice. MyD88/STING-deficient mice had a small increase in the number of bacteria in the livers compared to MyD88-deficient mice (Fig. 3A). However, MyD88/STING-deficient mice had significantly higher number of OVA-specific CD8+ T cells compared to MyD88-deficient mice (Fig. 3B), suggesting that the higher bacterial loads in the liver was likely due to loss of innate rather than adaptive immune responses. These data indicated that regardless of the presence or absence of MyD88, protective immunity to L. monocytogenes is enhanced in the absence of STING.
Mice immunized with LLO-deficient L. monocytogenes fail to develop protective immunity [21], [22]. Since LLO-deficient bacteria do not activate STING, we hypothesized that c-di-AMP-mediated STING activation might be sufficient to restore immunity in LLO−Lm-immunized mice. In support of previous reports, we found that BMDCs stimulated with c-di-AMP secreted cytokines and upregulated costimulatory markers in vitro (Fig. S3). Furthermore, mice administered c-di-AMP intravenously exhibited increased surface expression of CD86 and CD40 on splenic dendritic cells and of the activation marker CD69 on splenic CD8+ and CD4+ T cells in a STING-dependent manner (Fig. 4A and 4B), indicating that c-di-AMP can induce an inflammatory response in our model.
To test our hypothesis, B6 mice were immunized with LLO−Lm-OVA in the presence of c-di-AMP. While ActA−Lm-OVA-immunized mice restricted bacterial growth following challenge, the co-administration of c-di-AMP with LLO−Lm-OVA did not rescue protective immunity (Fig. 4C). Instead, the presence of c-di-AMP significantly reduced immunity in ActA−Lm-OVA-immunized mice suggesting that STING signaling inhibits protective immunity to L. monocytogenes infection.
To further examine the effect of c-di-AMP on immunity to L. monocytogenes, B6, STING- and IRF3/7-deficient mice were immunized with ActA−Lm-OVA in the presence or absence of c-di-AMP. Following challenge, B6 mice immunized in the presence of c-di-AMP had significantly higher bacterial numbers and had fewer numbers of total and OVA-specific CD8+ T cells in the spleen than mice immunized with ActA−Lm-OVA alone (Fig. 4D, 4E and S4A). STING-deficient mice were protected and had a robust CD8+ T cell expansion confirming that c-di-AMP-mediated inhibition of immunity was STING-dependent. IRF3/7-deficient mice were significantly protected as compared to naïve mice and had a population of OVA-specific CD8+ T cells, suggesting that c-di-AMP-mediated inhibition is partially due to IRF3 and IRF7 (Fig. 4D and 4E). Interestingly, IRF3/7-deficient mice were less protected compared to mice immunized with ActA−Lm-OVA alone, indicating that STING-dependent, IRF3/7-independent signaling also plays a role in loss of protective immunity.
Next, we evaluated a L. monocytogenes mutant, tetR::Tn917, that secretes 20-fold more c-di-AMP than WT L. monocytogenes [14], [15]. B6, STING- and IRF3/7-deficient mice were immunized with either ActA−Lm-OVA or the tetR::Tn917 mutant in the ActA−Lm-OVA background (tetRActA−Lm-OVA). B6 mice immunized with tetRActA−Lm-OVA had significantly higher bacterial numbers in the spleen compared to ActA−Lm-OVA-immunized mice whereas STING- and IRF3/7-deficient mice exhibited no significant difference (Fig. 5A). Furthermore, B6 mice immunized with tetRActA−Lm-OVA had a smaller population of total and OVA-specific CD8+ T cells compared to ActA−Lm-OVA-immunized mice (Fig. 5B and S4B). These data supported our finding that enhanced STING signaling lead to a reduction in protective immunity.
To determine whether enhanced STING signaling reduces T cell priming, OVA-specific CD8+ T cells were measured at the peak of the primary response from mice immunized with ActA−Lm-OVA in the presence or absence of c-di-AMP or tetRActA−Lm-OVA. At 7 days post immunization, mice immunized in the presence of c-di-AMP had significantly fewer OVA-specific CD8+ T cells compared to mice immunized with ActA−Lm-OVA alone. Mice immunized with the tetRActA−Lm-OVA mutant also had a small decrease in antigen-specific cells (Fig. 6A). Furthermore, the small population of OVA-specific CD8+ T cells that were present in mice immunized with elevated STING activation had significantly higher surface expression of the naïve T cell maker CD62L, suggesting that enhanced STING signaling elicited fewer antigen-specific effector T cells (Fig. 6B). In addition, peptide-stimulated CD8+ and CD4+ splenocytes from ActA−Lm-OVA-immunized mice in the presence of c-di-AMP or tetRActA−Lm-OVA-immunized mice produced fewer cytokines than those from ActA−Lm-OVA-immunized mice (Fig. 6C). These data indicate that mice immunized in the presence of enhanced STING signaling exhibited reduced T cell priming.
To determine the role of type I IFNs in STING-mediated CMI, IFN-αβR-deficient mice were tested for protective immunity. IFN-αβR-deficient mice restricted bacterial growth better than B6 mice (Fig. 7A and 7B), indicating that like STING- and IRF3/7-deficient mice, IFN-αβR-deficient mice also hyper-immunize. Although IFN-αβR-deficient mice immunized in the presence of c-di-AMP had higher bacterial numbers compared to mice immunized with ActA−Lm-OVA alone, these mice were significantly more protected than naïve mice, indicating a role for both type I IFN-dependent and independent mechanisms of suppression (Fig. 7A). IFN-αβR-deficient mice immunized with tetRActA−Lm-OVA were completely protected (Fig. 7B). These data indicated that the c-di-AMP-mediated inhibition of protective immunity is largely dependent on type I IFNs. Interestingly, we found that mice immunized in the presence of the synthetic double-stranded RNA, polyinosinic∶polycytidylic acid (poly(I∶C)), a STING-independent agonist of TLR3 and IRF3, also lost the ability to restrict bacterial growth following challenge (Fig S5), suggesting that type I IFN-mediated inhibition of immunity is unlikely STING specific.
We next determined whether inhibition of T cell priming by STING-dependent type I IFNs acted directly on lymphocytes. CD8+ T cells lacking the IFN-αβR undergo clonal expansion in response to primary L. monocytogenes infection so an adoptive transfer model could be used [33], [34]. B6, STING-, or IFN-αβR-deficient mice were injected with WT and IFN-αβR-deficient OT-I splenocytes and subsequently immunized with ActA−Lm-OVA in the presence or absence of c-di-AMP. At 7 days, WT and IFN-αβR-deficient OT-I cells expanded in ActA−Lm-OVA-immunized B6 mice, whereas in the presence of c-di-AMP, both WT and IFN-αβR-deficient OT-I cells had significantly reduced populations indicating that type I IFNs were not directly blocking T cell priming (Fig. 8A and 8B). Inhibition of T cell expansion by c-di-AMP was rescued in IFN-αβR- and STING-deficient mice further indicating that type I IFN-mediated suppression of immunity is not T cell intrinsic.
The results of this study show that the STING signaling pathway is not required to elicit CMI to L. monocytogenes. In fact, the absence of STING or IRF3 and IRF7 led to a higher number of antigen-specific CD8+ T cells and increased levels of protective immunity. Mice lacking both MyD88 and STING were also protected upon secondary challenge, indicating that not only is STING dispensable for the generation of a protective response to L. monocytogenes, it does not act in a redundant fashion with the TLR-MyD88 signaling pathway. Conversely, when STING activity was enhanced either by administering c-di-AMP during immunization or using a bacterial mutant that secretes elevated levels of c-di-AMP, mice failed to immunize and had decreased numbers of antigen-specific T cells following reinfection. This suppressive effect was largely due to the induction of type I IFNs since IFN-αβR-deficient mice immunized with either ActA−Lm-OVA in the presence of c-di-AMP or the tetR mutant were protected. Collectively, these findings suggest that L. monocytogenes-induced STING activation reduces the host adaptive immune response by induction of type I IFN.
The mechanism of type I IFN-mediated inhibition of T cell priming remains unclear. We found that both WT and IFN-αβR-deficient CD8+ T cells had significantly reduced expansion in the presence of c-di-AMP, suggesting that c-di-AMP-mediated suppression of lymphocyte priming is T cell extrinsic. One possibility is that the uptake of type I IFN-induced apoptotic cells by macrophages results in the release of the immunosuppressive cytokine IL-10 [35]. However, mice administered c-di-AMP did not have detectable levels of IL-10 in the serum. Furthermore, similar to B6 mice, IL-10-deficient mice immunized in the presence of c-di-AMP were unable to elicit OVA-specific CD8+ T cells following challenge (data not shown). Thus, IL-10 is not the downstream effector of STING/type I IFN-mediated suppression of adaptive immunity in our model of protective immunity. Other mechanisms including indoleamine 2,3-dioxygenase (IDO)-mediated T cell suppression, which is upregulated by type I IFN [36], may play a role in c-di-AMP-mediated inhibition of immunity. Indeed, Huang et al. found that splenic DCs from mice treated with DNA nanoparticles suppress ex vivo T cell proliferation in a STING- and IDO-dependent manner [37].
While spleens from immunized STING-deficient mice contained fewer bacteria than B6 mice, bacterial clearance by immunized IRF3/7- and IFN-αβR-deficient mice was even higher. One possibility is that there are STING-independent sources of type I IFNs in response to L. monocytogenes. Supporting this hypothesis, we observed that MyD88/STING-deficient macrophages produced low levels of IFN-β eight hours post infection. Recent studies found that RIG-I-deficient cells had reduced IFN-β secretion in response to L. monocytogenes infection suggesting that RNA from L. monocytogenes can be detected by the host [38], [39]. Supporting this hypothesis, we found that immunized mice lacking MAVS, the signaling adaptor for RIG-I, were more protected than control mice upon reinfection (data not shown). Thus, RIG-I-dependent, STING-independent induction of type I IFNs via IRF3 and IRF7 may also contribute to the inhibition of protective immunity. Another possibility is that since naïve IRF3/7- and IFN-αβR-deficient mice exhibit heightened bacterial clearance compared to both naïve B6 and STING-deficient mice, innate immune factors are also likely restricting reinfection. In fact, IFN-αβR-deficient mice have higher neutrophil recruitment in response to L. monocytogenes infection [40]. Thus, careful examination of each mouse strain will be necessary to determine the extent of the contribution of innate versus adaptive immune mechanisms in protective immunity.
Type I IFN-mediated inhibition of immunity is unlikely specific to STING activation. Mice immunized in the presence of the poly(I∶C), which induces type I IFNs independently of STING, also lost the ability to restrict bacterial growth following challenge. Several studies have shown that administering poly(I∶C) prior to a protein antigen inhibits clonal expansion of antigen-specific CD8+ T cells, supporting our findings that systemic type I IFN-induced inflammation reduces T cell priming [41], [42].
The importance of type I IFNs during bacterial infection is less understood than for viruses [43]. For example, IFN-β is the highest upregulated gene following L. monocytogenes cytosolic invasion, yet mice lacking the IFN-αβR are more resistant to acute infection, suggesting that type I IFNs may promote pathogenesis [35], [44], [45]. However, strains of L. monocytogenes that secrete elevated levels of c-di-AMP and induce higher levels of type I IFNs, are not hypervirulent [46], but induce considerably less T cell immunity as shown in this study. Thus it is possible that secretion of c-di-AMP and consequent type I IFN production may play a role in L. monocytogenes pathogenesis by suppressing the development of adaptive immunity. Although L. monocytogenes generally causes acute infections, recent studies have found that type I IFNs promote chronic infections with LCMV by suppressing cell-mediated mechanisms of viral control [47], [48]. In the case of Mycobacterium tuberculosis, mice lacking IRF3 are more resistant to infection with M. tuberculosis suggesting that IRF3 activation is detrimental to host clearance [49]. In humans, IFN-α treatment leads to higher incidences of TB reactivation [50]. Furthermore, active TB patients exhibit an increase in type I IFN-inducible transcripts in the blood, which correlated with disease severity [51]. Therefore, type I IFNs may exacerbate or maintain secondary or long-term chronic infections. Interestingly, human STING often contains polymorphisms that makes it resistant to bacterial but not host derived CDNs [52].
STING-mediated suppression of protective immunity was not solely due to type I IFNs. Although protective immunity in tetRActA−Lm-OVA-immunized mice was rescued in the absence of IRF3/7 and the IFN-αβR, mice immunized in the presence of c-di-AMP exhibited type I IFN-independent suppression. Since STING activates NF-κB as well as IRF3, it is possible that NF-κB-dependent inflammation also plays a role in restricting immunity. We believe that administering c-di-AMP activates STING more robustly compared to infection with the tetRActA−Lm strain and thus inhibition of immunity by type I IFN-independent inflammation would become more apparent.
The results of this and other studies suggest an inverse relationship between the extent of inflammation and the development of adaptive immunity. For example, IL-12-deficient mice immunized with L. monocytogenes develop higher numbers of CD8+ memory T cells and are more resistant to reinfection [53]. In previous studies, we found that a L. monocytogenes strain engineered to activate the inflammasome, and consequently induce high levels of IL-1β secretion, was a poor inducer of adaptive immunity [54]. Furthermore, co-administration of heat-killed or LLO-deficient L. monocytogenes blocked immunity to WT bacteria in a MyD88-dependent manner [55]. Thus, activation of three distinct signaling pathways, STING, MyD88, and caspase-1 all resulted in the inhibition of the development of adaptive immunity. Therefore, there appears to be a dichotomy between innate immune pathways that are necessary for survival (for example, MyD88 for L. monocytogenes and type I IFN for viruses), and those that lead to adaptive immunity. In fact, our data and work from others suggest that lack of inflammation represents an ideal environment for the generation of memory T cells [31], [56]. Indeed, mice deficient for both MyD88 and STING are fully immunized by L. monocytogenes even though there was a significant reduction of dendritic and T cell activation and cytokine production following immunization.
Considering that the innate immune response is believed to be required for the initiation of adaptive immunity, we were surprised that MyD88/STING-deficient mice immunized with L. monocytogenes were protected after reinfection. This raises the question, which innate immune signaling pathways contribute to the initiation of T cell priming to L. monocytogenes? Previous work from our group found that NOD2 detects cytosolic L. monocytogenes [5]. However, immunized MyD88/NOD1/2-deficient mice clear bacteria upon secondary lethal challenge (data not shown), suggesting that protective immunity is not due to redundancy between MyD88 and NOD-like signaling pathways. Future studies to identify which innate immune detection pathways are required for L. monocytogenes-mediated CMI would provide a greater understanding of how pathogens and adjuvants elicit protective immunity, knowledge that can be used for the development and improvement of vaccines.
This study was carried out in strict accordance with the recommendations in the Guide for the Care and Use of Laboratory Animals of the National Institutes of Health. All protocols were reviewed and approved by the Animal Care and Use Committee at the University of California, Berkeley (MAUP# R235-0813B).
C57BL/6 mice were purchased from The Jackson Laboratory. Goldenticket (Gt) mice were generated from an ENU mutagenesis screen. Gt mice contain a single nucleotide mutation in STING resulting in the absence of the STING protein [17]. All mice were in the C57BL/6 genetic background. Gt, MyD88−/−, MyD88−/−Gt, IRF3−/−, IRF3/7−/− and IFNAR1−/− mice were bred in our facilities. GFP+/+ OT-I+/+ RAG2−/− and IFNAR1−/− OT-I+/+ Ly5.1+/+ RAG2−/− mice were generously provided by Ellen Robey.
All L. monocytogenes strains were in the 10403S background. ActA−Lm-OVA (ΔactAΔinlB) (DP-L6014) [57], WT Lm-OVA (DP-L6018) and LLO−Lm-OVA (Δhly) (DP-L6017) [21] L. monocytogenes were previously described. For tetRActA−Lm-OVA (DP-L6015), the tetR::Tn917 transposon [14] was transduced into ActA−Lm-OVA. L. monocytogenes were grown in brain heart infusion (BHI) media at 30°C overnight without shaking to stationary phase. For in vitro infections, L. monocytogenes was washed 3× in PBS. For in vivo infections, L. monocytogenes was diluted in BHI at 37°C shaking for ∼2 hours until they reached an OD600 0.4–0.6.
Eight to 12 week old sex-matched mice were infected intravenously with 103 CFU (unless otherwise indicated) of L. monocytogenes diluted in phosphate buffered saline (PBS) in a total volume of 200 µl. For acute infections and primary immunization studies, mice were sacrificed at 1, 2 and 3 or 7 days post infection, respectively. For challenge studies, mice immunized 30–38 days prior were infected with 2×105 CFU of WT Lm-OVA. Where indicated, mice were administered either 50 µg or 100 µg of c-di-AMP or 50 µg of poly(I∶C) (InvivoGen) with the bacterial inoculum. Three days later, spleens and livers were homogenized in 0.1% IGEPAL CA-630 (Sigma) and plated on LB-strep plates to enumerate CFU. For analysis of CD8+ T cell responses, spleens were divided and weighed. For splenic dendritic and T cell activation, mice were immunized with 105 CFU of ActA−Lm-OVA. A higher dose was used to allow for the easy detection of activated cells.
For OT-I cells, splenocytes from GFP+/+ OT-I+/+ RAG2−/− and IFNAR1−/− OT-I+/+ Ly5.1+/+ RAG2−/− mice were isolated and washed 3× with PBS. Percent of CD8+ OT-I cells was determined by flow cytometry. 2×104 of each cell type was injected intravenously into each mouse (4×104 total cells/mouse) 1 day prior to immunization with L. monocytogenes.
C-di-AMP was generated by Josh Woodward as previously described [58]. Purified c-di-AMP was resuspended in tissue culture grade PBS. LPS was removed from the prepared c-di-AMP using Detoxi-gel endotoxin removing gel (Pierce) according to the manufacturers instructions. Endotoxin content was measured using the Toxinsensor Chromogenic LAL Endotoxin assay kit (Genescript). The nucleotide solution was passed through the Detoxi-gel until endotoxin levels were <0.0125 EU/ml. Nucleotide was then diluted to 500 µg/ml.
For analysis of T cell responses, spleen halves were dissociated and filtered through a 70 µm cell strainer. Red blood cells were lysed with Red Blood Cell Lysing Buffer (Sigma). To determine OVA-specific cells, splenocytes were stained with anti-mouse CD8, CD44, CD62L and a Kb/OVA257–264 tetramer. Representative FACs plots are gated on CD8+ cells and values show the median percentage of Kb/OVA257–264 tetramer+ CD44+ within the CD8+ cell population ± SEM. For peptide stimulation assays, splenocytes were stimulated for 5 hours with 2 µM OVA257–264 or LLO190–201 peptide in the presence of GolgiPlug (BD Biosciences). Cells were surface stained with anti-mouse CD8 and CD4, fixed and permeabilized using Cytofix/Cytoperm (BD Biosciences), and stained for intracellular anti-mouse IFN-γ, TNF-α and IL-2. For splenic dendritic cells and T cells, splenocytes were stained with anti-mouse CD11b, CD11c, CD86 and CD40 or anti-mouse CD8, CD4 and CD69, respectively. Flurophore-conjugated antibodies were purchased from eBioscience. Samples were acquired using an LSRII flow cytometer (BD Biosciences) and analyzed using FlowJo software (Tree Star).
BMMs were generated as previously described [59]. In a 6-well plate, 2×106 BMMs were either infected with WT Lm at a multiplicity of infection (MOI) of 2 bacteria per cell or stimulated with 10 µM c-di-AMP. At 30 minutes post infection, gentamicin was added for a final concentration of 50 µg/ml. At 4 and 8 hours post infection, cells were harvested and RNA was purified using the RNAqueous kit (Ambion). RNA was then DNase treated, processed and analyzed as previously described [5].
Bone marrow from femurs was plated in media containing 20 ng/ml recombinant murine GMCSF (ProSpec) at a density of 5×105 cells/ml in a 24-well plate. At days 2, 4 and 5 media was replaced with fresh media containing 20 ng/ml GMCSF and cells were harvested on day 6. BMDCs were plated at 3×105 cells/well in 48-well plates. Cells were incubated with either 10 µM c-di-AMP, 100 ng/ml lipopolysaccharide (LPS), or 20 µg/ml poly(I∶C) (InvivoGen). After 24 hours, supernatant was assayed for IFN-β using ISRE-L929 cells as previously described [14], or for MCP-1, IL-12p40 (BD OptEIA kit, BD Biosciences) and IL-6 (eBioscience) by ELISA.
Serum cytokines were measured using the CBA Mouse Inflammation Kit (BD Biosciences) and analyzed on the LSRII flow cytometer.
A two-tailed, Mann-Whitney U test was used to analyze the significance of differences in the means between groups. Significance is indicated as * p<0.05, ** p<0.005, *** p<0.0005 or ns = not significant.
|
10.1371/journal.ppat.1005594 | Crystal Structure of Human Herpesvirus 6B Tegument Protein U14 | The tegument protein U14 of human herpesvirus 6B (HHV-6B) constitutes the viral virion structure and is essential for viral growth. To define the characteristics and functions of U14, we determined the crystal structure of the N-terminal domain of HHV-6B U14 (U14-NTD) at 1.85 Å resolution. U14-NTD forms an elongated helix-rich fold with a protruding β hairpin. U14-NTD exists as a dimer exhibiting broad electrostatic interactions and a network of hydrogen bonds. This is first report of the crystal structure and dimerization of HHV-6B U14. The surface of the U14-NTD dimer reveals multiple clusters of negatively- and positively-charged residues that coincide with potential functional sites of U14. Three successive residues, L424, E425 and V426, which relate to viral growth, reside on the β hairpin close to the dimer's two-fold axis. The hydrophobic side-chains of L424 and V426 that constitute a part of a hydrophobic patch are solvent-exposed, indicating the possibility that the β hairpin region is a key functional site of HHV-6 U14. Structure-based sequence comparison suggests that U14-NTD corresponds to the core fold conserved among U14 homologs, human herpesvirus 7 U14, and human cytomegalovirus UL25 and UL35, although dimerization appears to be a specific feature of the U14 group.
| Human herpesvirus 6B (HHV-6B), a causative agent of exanthema subitum for children and immunocompromised adults, encodes numerous tegument proteins that constitute the viral matrix. HHV-6B U14 is a tegument protein essential for viral propagation, and additionally it interacts with host factors such as tumor suppressor p53 and cellular protein EDD, thereby regulating host cell responses. Here, we report the molecular structure of HHV-6B U14 at an atomic resolution. The N-terminal domain of U14 (U14-NTD) adopts an elongated, helix-rich fold without any significant overall similarity to known structures. U14-NTD forms a 100 kDa homodimer through electrostatic interactions and a wide hydrogen bond network. The U14-NTD homodimer displays four clusters of electrostatic potential with deep grooves, implying multiple binding sites for other viral or host proteins. U14-NTD corresponds to the core fold shared by homologous proteins of human herpesvirus 7 (HHV-7) and of human cytomegalovirus, although dimerization seems to be specific to HHV-6 and HHV-7. The U14-NTD structure provides clues to promote further analysis on the role and behavior of U14 in the pathogenesis of HHV-6. It also leads to a comprehensive understanding of the U14 homologs in beta herpesviruses, and furthermore contributes to the overall knowledge about tegument proteins in herpesviruses.
| Human herpesvirus 6B (HHV-6B) and the closely-related virus HHV-6A are classified as Roseolovirus genus of beta herpesvirus subfamily [1] [2] [3] [4], which also includes human herpesvirus 7 (HHV-7) and human cytomegalovirus (HCMV). HHV-6B is a causative agent of exanthema subitum for children [5] [6] by primary infection and for immunocompromised adults by reactivation of the latent virus. Diseases induced by HHV-6 primary or reactivated infection are sometimes severe, causing encephalitis [7] [8].
Herpesviruses share a common architecture of the virion that is enveloped and contains the double-stranded DNA genome in a protein shell known as capsid. The space between the envelope and the capsid is filled with a pool of tegument proteins [9] [10]. The composition of tegument proteins differs among herpesviruses, and numerous tegument proteins have been identified for HHV-6B [11]. Tegument proteins are versatile proteins suggested to have additional functions other than acting as structural components of the viral tegument [12–14]. Thus, the characteristics and function of each tegument protein remain to be defined and could have a role in understanding herpesvirus pathogenesis.
HHV-6B U14 is a tegument protein that is 604 amino acid residues in length. U14 belongs to the herpes pp85 superfamily that is shared among beta herpesviruses, and has no homologs in alpha and gamma herpesvirus [11] [15] [16] [17]. HHV-6A, HHV-6B, and HHV-7 have U14 with relatively high sequence homology. Other members of the beta herpesvirus subfamily, including HCMV, have two tegument proteins belonging to the pp85 superfamily, UL25 and UL35 [18], although their sequence identities with U14 is less than 20%. Recently, HHV-6A U14 was revealed as an essential factor in the viral life cycle because a three amino-acid deletion in the U14 sequence resulted in a defect in viral growth [19]. In addition, U14 of HHV-6A and HHV-6B associate with the tumor suppressor protein p53 in the nucleus and cytoplasm, finally being incorporated into virions with p53 [20]. Furthermore, we found that HHV-6A U14 induces cell cycle arrest in G2/M phase by associating with a cellular protein, EDD during early phase of infection [21]. These results indicate that HHV-6 U14 functions not only as a virion tegument protein, but also in viral DNA replication cycles, suggesting that it is a multi-functional protein.
Structure determination of tegument proteins is an effective approach, providing information about their structural characteristics as well as a basis for mapping the results of biochemical experiments. In this study, we solved the crystal structure of the N-terminal region of U14 protein derived from HHV-6B. The structure represents a characteristic dimer form with potential functional sites. Through sequence comparison with HHV-6A U14, HHV-7 U14, and HCMV UL25 and UL35, shared and specific features among these homologs are discussed.
Full-length HHV-6B U14 (603 amino acids) was expressed in E. coli with MBP at the N-terminus (U14-MBP; Fig 1A). During purification, a fraction of U14-MBP was degraded to a smaller size, indicating that the C-terminal region of U14-MBP is unstable in E. coli (S1 Fig). Thus, a new construct was designed to express the U14-NTD corresponding to the N-terminal region (residues 2–458) in the form of an N-terminal MBP fusion (Fig 1A). MBP-U14-NTD was not degraded significantly during purification (S1 Fig), supporting the assumption that C-terminal region of MBP-U14 was degraded. Actually, the size of MBP-U14-NTD was similar to the degradation product of U14-MBP (S1 Fig). In the size-exclusion column chromatography experiment, the retention time of U14-NTD was shorter than expected from its size (50 kDa), indicating that U14-NTD forms a multimer in solution (Fig 1B). The size of U14-NTD estimated from the calibration curve was 118 kDa, which is slightly higher than the calculated size 100 kDa for a U14-NTD dimer.
Because there was no available structural information of any protein with high sequence homology to HHV-6B U14, a SeMet-derivative of U14-NTD was prepared to solve the phase problem by anomalous dispersion method. The structures of native U14-NTD and the SeMet-derivative were determined at 1.85 Å and 2.3 å resolutions, respectively (Table 1). There were two almost identical U14-NTD molecules in the asymmetric unit. Their RMS deviation for main-chain atoms and heavy atoms were 0.51 å and 1.09 å, respectively. Almost all of the U14-NTD residues were assigned to the electron density, with the exception of N-terminal residues 1–7 and C-terminal residues 456–458, indicating that the designed U14-NTD represents an actual structural domain of U14.
U14-NTD has an elongated helix-rich structure composed of sixteen α helices, four 310 helices and two β strands. To facilitate structure description, the U14-NTD structure was divided into four subdomains (SDs) based on secondary structure topology and spatial arrangement (Fig 2). The four-helix bundle SD2, which is composed of the N-terminal half of α4, the C-terminal half of α10, α11 and α12, forms a central part of U14-NTD (Fig 2, cyan). At the preceding N-terminal region, SD1 forms a compact fold including helices η1, α1, α2 and α3 (Fig 2, magenta), and is associated with SD2. SD3 is located at one side of the elongated long axis of U14-NTD (Fig 2, green). The C-terminal half of α4 and the N-terminal half of α10 are surrounded by five α helices (α5, α6, α7, α8, and α9) and form a compact fold of SD3. At the opposite side of the long axis, the C-terminal region of U14-NTD folds as SD4 composed of α13, α14, η2, η3, η4, η5, β1, β2, α15 and α16 (Fig 2, yellow). The β1 and β2 form a recognizable β hairpin that protrudes from the core overall fold.
Analysis by the DALI program [22] with the latest set of Protein Data Bank (PDB, www.rcsb.org, [23]) entries revealed that SD2 is similar to a variety of proteins characterized by four-helix bundles (Table 2). In addition, SD3 showed marginal similarity to Unc-51-like kinase 3 and other proteins (Table 2). For SD1 and SD4, no significant homology to known proteins was detected.
In the crystal structure, a dimer is formed along the long axis of U14-NTD in an antiparallel orientation (Fig 3). The dimer can be regarded as two right hands shaking one another with the protruding β hairpins forming the “thumbs”. A two-fold axis is located by the side of the β hairpin, resulting in an arrangement of crossed hairpins. All of the SDs are involved in the dimer interface. The calculated buried surface area per monomer is 4146 Å2 (Fig 3B, red), which is a relatively large value compared with those of known homodimer structures of similar size, at approximately 2800 Å2 [24].
One noticeable characteristic of the U14-NTD dimer is an internal cavity within the dimer interface (Fig 3C). The two-fold axis of the dimer penetrates the cavity. The volume of the cavity is approximately 1200 Å3 and corresponds to 2.1% of the volume of the monomer (57400 Å3). The internal cavity is enclosed with α4 and α10 of SD2 and α13 of SD4 from each monomer. The β hairpins of SD4s also face this cavity, forming a lid that separates it from the outer solvent.
Broadly spanning electrostatic interactions contribute to the dimerization (Fig 4A). At the dimer interface, a monomeric U14-NTD shows a negatively-charged surface between SD2 and SD4. On the other hand, a positively-charged surface is found between SD2 and SD3 in the same monomer. In the dimer form, the negatively-charged area of each monomer faces the positively-charged area of the opposite monomer. There are a lot of hydrogen bonds within and around the electrostatically attracting areas. Four clusters were found and named as interaction sites a, b, c, and d (Fig 4B). A total of 40 hydrogen bonds were formed in these areas, indicating tight and specific dimerization. Their distribution is summarized in S1 Table and the detailed interaction modes are shown in S2 Fig.
The U14-NTD dimer shows characteristic multiple clusters of positive and negative electrostatic potential on the surface (Fig 5A). At the β hairpin side, an extended negatively-charged area is formed across the two-fold symmetry axis (front side, Fig 5A, left). The dimer surface of this side is composed primarily of SD4. The β hairpin of each monomer contains six negatively-charged residues (S3A Fig). The region 342–378 of SD4, which corresponds to the outermost part of the long axis of U14-NTD, includes 11 negatively-charged residues (S3B Fig). On the opposite side (back side, Fig 5A right), the area around the two-fold axis is surrounded by seven positively-charged residues from each monomer (S3D Fig). At this same back side, a negatively-charged cluster consisting of ten negatively-charged residues is observed at the peripheral area distant from the two-fold axis (S3C Fig).
The three amino acids, L424, E425, and V426, of which deletion or substitutions to alanines caused a defect in viral growth [19], were mapped to the β hairpin (Fig 5B). The side-chains of L424 and V426 face the solvent side and constitute a continuous hydrophobic patch with the side-chain of I414 on the opposite β strand and the side-chain of L297 on SD2 of the partner monomer (Fig 5B).
In contrast to the flat surface of the front side, the back side has deep grooves along the dimer interface due to the staggered arrangement of monomers (Fig 5C). One side of the groove is exclusively composed of SD3, with SD4 and SD2 of the partner monomer forming the opposite wall. The length, depth, and width (distance between monomers) were roughly estimated to be ~30 Å, ~20 Å, and ~20 Å, respectively (S4 Fig).
To address issues of similarity and difference between HHV-6B U14 and its homologs, multiple sequence alignment was performed for HHV-6B U14, HHV-6A U14, HHV-7 U14, and HCMV UL35 (Fig 6). The alignment combined with the structural information of U14-NTD showed that U14-NTD is a core part conserved among all members. In the core region, HHV-6A U14 and HHV-7 U14 are well aligned with HHV-6B U14 across all SDs. HCMV UL35 was also aligned in the core part, except for the SD4, where short gaps are required for the alignment (Fig 6). For HCMV UL25, another pp85 family protein, similar alignment was obtained, although the pattern is different from that of HCMV UL35 (S5 Fig). These alignments suggest that these U14 homologs have a similar helix-rich fold at the core region.
Next, we examined the conservation of the residues involved in the hydrogen bond network in the HHV-6B U14-NTD dimer. Most of the HHV-6B U14-NTD dimer interaction sites are occupied by identical residues in HHV-6A U14 and HHV-7 U14, indicating that these homologs also dimerize in a similar manner. Of the 19 residues whose side-chains involved in the interaction, 17 and 16 residues are identical for HHV-6A U14 and HHV-7 U14, respectively (Fig 6). By contrast, the residues participating in the dimer interface are different from HHV-6B U14 in HCMV UL25 and UL35. For HCMV UL25 and UL35, only 4 and 3 residues are identical to HHV-6B, respectively. It suggests that HCMV UL25 and UL35 takes a different form to that of the HHV-6B U14-NTD dimer.
The HHV-6B U14-NTD structure, comprised of residues 2–458, reveals a helix-rich fold forming a compact homodimer. The broad and intricate interactions between each monomer (Figs 3 and 4), as well as the retention time of the size-exclusion column chromatography (Fig 1B), support the suggestion that a dimer is the natural form for U14-NTD. Multimerization of viral proteins has been frequently reported, particularly for structural proteins constituting the capsid and associated proteins. A number of tegument proteins have also been shown to form self-associated multimers, such as HSV-1 UL36 [26] and VP22 [27], HCMV pp65 [28] and pp28 [29], and murine gammaherpesvirus 68 ORF52 [30]. Compared with these, the ~50 kDa U14-NTD is relatively large as a dimerization domain with a broad interface in which all four SDs are included. Although the viral matrix is considered to be an amorphous/disordered protein pool in general, multimerization of its constituents would impose local order to some extent as a corollary to the symmetries of their own and of their interaction sites for other partners. Such local order in the viral matrix has been suggested for matrix proteins of RNA virus; multimerization of matrix proteins relates to the formation of a protein lattice in the matrix and contributes to the membrane deformation required for the budding process [31], [32]. Thus, the dimerization of U14 revealed in this research implies a role for this protein as a scaffold in the viral matrix. Analyzing the expression amount of U14 protein in virions would be required. As far as we know, the expression amount of HHV-6 U14 has not been investigated, hence it should be addressed in a future research. In the case of HHV-7, U14 is known as a major antigen pp85 [33], and U14 is thought to be relatively expressed abundantly. On the other hand, one of predominant major antigens of HHV-6 has been shown to be U11 [34], [35], which has been revealed to interact with U14 [36]. It may be noteworthy to mention that the HCMV UL25 was expressed abundantly especially in the dense body [37].
The HHV-6B U14 structure suggests multiple structural features as potential function sites, such as the negatively- and positively-charged clusters (Fig 5A), the β hairpin flanked by the hydrophobic patch (Fig 5B), and the grooves formed along the dimer interface (Fig 5C). These distinct structural features would be consistent with the multiple functions of U14. U14 is observed in at least three different locations during the protein's life cycle, namely in the nucleus of a host cell at an early phase of infection, in the cytoplasm at a late phase, and in the virion [20]. At each location, U14 could have a different role via interaction with different host/viral factors. Thus far, at least two associated host proteins are reported for U14: the tumor suppressor p53 [20] and EDD [21]. Experiments using deletion mutants of U14 revealed that three amino acids on the β hairpin and the C-terminal region outside U14-NTD are implicated in the interaction with p53 and EDD, respectively [21]. Among the potential function sites, the area around the β hairpin is of importance because substitution or deletion of three amino acids (L424, E425, and V426) on the β hairpin results in a defect in viral multiplication [19]. It is expected that the deletion of the three amino acids strongly affects the β hairpin structure due to the imbalanced length of the two β strands. The β hairpin contributes to the dimer interaction (Fig 4 and S2D Fig); such deletion could change either the fold of U14-NTD or its dimerization and function. On the other hand, substitutions probably maintain the β hairpin structure because the original side chains are not involved in folding and easily simulated to be substituted without any necessity to change its structure. To further assess the importance of the β hairpin, we performed immunoprecipitation assay with HHV-6A U14 mutants in which residues on the β hairpin were substituted (S6 Fig). The p53 interaction was abolished by the deletion of three amino acids corresponding to L424, E425 and V426 as reported previously [21]. In contrast, a single alanine substitution at the corresponding position to I414 (Fig 5B) did not affect the interaction. Therefore, p53 is suggested to be sensitive to the change in β hairpin structure due to the deletion. Another possibility is that the β hairpin is involved in binding to other viral proteins, such as tegument proteins to form the tegument structure or capsid or envelope proteins to form the virion structure. Recently, we identified a major tegument U11 as the binding partner of U14 [36], then the effect of the mutation/deletion on the β hairpin was also analyzed by the immunoprecipitation assay (S6 Fig). The interaction between U14 and U11 was abolished by the deletion of the corresponding residues of the L424, E425 and V426. Moreover, in contrast to p53, a single substitution at the corresponding residue of I414 (I414A) caused impaired interaction with U11. Thus, we suggest that the β hairpin is the binding site for U11 and the exposed hydrophobic sidechain of I414 observed in the U14-NTD structure is likely to be recognized by U11. Because U11 is an abundant and essential tegument protein of HHV-6 [36], further research focused on the interaction via the β hairpin will provide more information about the functionality of U14.
SD3 of U14-NTD showed structural similarity with the MIT (microtubule interacting and trafficking) domain of Unc-51-like kinase 3 (ULK3, PDB ID: 4wzx, [38]) by DALI analysis (Table 2 and Fig 7A). Recent research revealed that the ULK3-MIT domain interacts with the MIM2 motif of ESCRT-III and phosphorylates the site, resulting in inhibition of cytokinesis during the cell division process [38]. The MIM2 binding site of the ULK3-MIT domain was superposed to a part in the deep groove observed around U14-NTD SD3, and partially opened to the solvent (Fig 7B). Thus, it is tempting to speculate that SD3 and the nearby groove of U14-NTD dimer serve as the binding site for ESCRT-III or related proteins, thereby contributing to their transporting function. Considering that ESCRT-III is involved in the viral maturation/budding step [39] [40], further experiments are required to examine the relationship between U14 and the ESCRT system, thus further elucidating U14 function.
Structure-based sequence analysis revealed that the HHV-6B U14-NTD corresponds to the core part conserved between U14 and UL25/UL35 (Fig 6 and S5 Fig). The similarity across this region indicates that HHV-6A U14, HHV-7 U14, and HCMV UL25 and UL35 adopt the same elongated helix-rich fold. However, the absence of homology in HCMV UL25 and UL35 at the dimer interface region of HHV-6B U14 suggests that the dimer form is specific to U14 proteins. Because most of the structural features that likely constitute the functional sites of U14 depend on the dimer form, this information is not applicable to HCMV UL25 and UL35. The C-terminal region outside U14-NTD contains a large proportion of hydrophilic and glycine residues (S2 Table). This indicates that U14 consists of the core part with an intrinsically disordered tail [41]. The C-terminal region following the core fold differs in length among U14 and U25/U35 proteins (HHV-6B U14: 147 residues, HHV-6A U14: 146 residues, HHV-7 U14: 190 residues, HCMV UL25: 13 residues, and HCMV U35: 169 residues), posing difficulty in obtaining valid sequence alignments. The variation in the C-terminal region has been indicated from the alignment between HHV-6A and HHV-7 U14 [33]. HCMV UL25 has a long extension of 180 residues that precede the core part instead of the C-terminal region observed in other homologs. The amino acid compositions of these extensions share a common propensity. Similar to the HHV-6B U14 C-terminal region, those of other homologs are dominantly composed of hydrophilic and glycine residues (S2 Table). This indicates that these regions are intrinsically disordered without a stable structure in solution [41], and thus the construction, a core fold followed by an unstructured tail(s), is common for these proteins of the pp85 family. Interestingly, the proportion of serine residues is unusually high, around 20%, in these tail regions (S2 Table), which suggests that the tail could be the site of post-translational modifications such as phosphorylation and glycosylation. The importance of the C-terminal region has been established for U14 and UL35. The C-terminal region of HHV-6 U14 is required for interaction with EDD and, subsequently, cell cycle arrest [21]. HCMV UL35 has an isoform, UL35A, corresponding to the C-terminal 193 amino acids of UL35. UL35A functions to modulate expression of immediate early genes [18]. Alignment between HHV-6B U14 and HCMV UL35 showed that UL35A includes only a part of α15 and α16 in SD4, suggesting that UL35A is unlikely to form a stable fold on its own. The structural information derived from U14-NTD provides the basis for further structure-based analyses necessary for addressing the roles of these similar, but significantly different, tegument proteins.
The coding sequences for the U14 N-Terminal Domain (U14-NTD) was amplified by PCR from optimized viral DNA (optimized by GeneOptimizer) of HHV-6B strain HST using the U14 forward primer, with the HRV 3C protease site underlined, (5’- ACAGGATCCCTGGAGGTGCTGTTCCAGGGCCCCGAAGGCAGCAAGACCTTC-3’) and the U14 reverse primer (3’-ACAGTCGACTTACTCGTTCTGGTTCAGC-5’). The PCR product was subcloned into pMAL-C2 using BamHI and SalI restriction sites. The cloned DNA fragment was sequenced with a 3130 Genetic Analyzer (Applied Biosystems).
For native U14-NTD, freshly transformed Escherichia coli strain BL21 cells were incubated at 37°C overnight in 10 ml lysogeny broth (LB) starter culture supplemented with 50 μg ml-1 carbenicillin. The starter culture was diluted into 1 liter LB medium supplemented with 50 μg ml-1 carbenicillin and grown at 37°C until an OD600 of 0.6–0.7. Then the temperature was shifted to 20°C and the cells were induced with 0.3 mM isopropyl-β-D-thiogalactopyranoside (IPTG). The expression was induced for 24 h.
To prepare a selenomethionine (SeMet) derivative of U14-NTD for phase determination, Escherichia coli strain B834 was used as a host. Cells grown overnight in 10 ml LB medium were then diluted into 400 ml LB medium supplemented with 50 μg ml-1 carbenicillin and grown at 37°C until an OD600 of 0.9–1.0. Cells were harvested by centrifugation and suspended in SeMet M9 medium supplemented with 50 μg ml-1 carbenicillin. The final volume of medium was 1 liter when the main culture was started. Cells were grown at 37°C until the OD600 reached 0.6–0.7 before being induced with 0.3 mM IPTG. The expression was induced for 16–20 h.
Cells containing native U14-NTD or SeMet U14-NTD were harvested by centrifugation at 8000 ×g for 12 min at 4°C and suspended in column buffer (20 mM TrisHCl pH 7.4, 200 mM NaCl, and 0.1 mM DTT) with 0.5% v/v TritonX-100. The lysate was stored at -80°C for 12–14 h, and then disrupted by sonication. Insoluble proteins were removed by centrifugation at 11000 ×g for 15 min at 4°C. As the pMAL-C2-encoded U14-NTD contained an N-terminal maltose-binding protein (MBP) tag, Amylose Resin (NEW ENGLAND BioLabs) was added to the supernatant and gently rocked at 4°C for 10–12 h. The Amylose Resin was spun down by centrifugation at 500 ×g for 5 min at 4°C and washed with cold column buffer five times before being applied to a 20 ml column (BioRad). The column was washed with five column volumes of column buffer. The U14-NTD was eluted with column buffer containing 10 mM maltose. The MBP tag was removed by adding PreScission Protease (GE Healthcare; 15 U mg-1 U14-NTD with 0.4 mM DTT) using the HRV 3C protease site as described above. Further purification was carried out by size-exclusion chromatography using a Superdex 200pg column (GE Healthcare). The column was calibrated by HWM Calibration Kit (GE Healthcare). The protein was concentrated to 2.0–2.5 mg ml-1 using an Amicon Centrifugal Filter (molecular weight cut-off 30 KDa, Millipore) and the purity was assessed by SDS-polyacrylamide gel electrophoresis and Western blot using an antibody against MBP.
Purified U14-NTD was passed through a 0.22 μm Ultrafree Centrifugal Filter (Millipore) to remove aggregate. The concentration of the protein was estimated based on an A280 of 0.75 for 1 mg ml-1, calculated from the amino acid composition. Initial crystallization screening of U14-NTD was executed in 96-well plates at 4°C by the sitting-drop vapor-diffusion technique using the screening kit Index HTTM (Hampton Research). Each drop was prepared by mixing 0.5 μl of protein solution (both 2.5 mg and 1.25 mg ml-1 U14-NTD, 20 mM TrisHCl pH 8.0, 100 mM NaCl, and 0.1 mM DTT) with 0.5 μl reservoir solution, and was then equilibrated against 60 μl reservoir solution. Crystallization conditions were optimized by varying the pH, salt and precipitant concentration in 24-well plates. Finally, crystals suitable for X-ray analysis were obtained from drops prepared by mixing 1.0 μl protein solution (1.25 mg ml-1) with 1.0 μl reservoir solution consisting of 0.2 M Potassium sodium tartrate tetrahydrate and 16–18% w/v Polyethylene glycol 3,350 at 4°C. The crystals described here formed in 3–5 days and were harvested 30–40 days later to reach a size suitable for data collection.
X-ray diffraction data were collected on beamline BL26B1 and BL26B2 at SPring-8, Harima, Japan [42]. For data collection, crystals were transferred into a solution consisting of the reservoir solution supplemented with 25% glycerol prior to being flash frozen in liquid nitrogen. The data were processed using XDS [43] and indexed in space group P21212. Dataset of SeMet-U14-NTD was collected at the peak wavelength of Se K-edge and used for the experimental phasing by Phenix.autoSol [44,45]. Dataset of native U14-NTD was solved by molecular replacement method with Phenix.phaser [46]. The SeMet-U14-NTD model was used as the search model. Structural refinement was performed with Phenix.refine [44] [47] and Coot [48]. Structural analysis and image depiction were performed using MolMol [49] and UCSF Chimera [50]. The synchrotron radiation experiments were performed at BL26b1 and BL26b2 in SPring-8 with the approval of RIKEN (Proposal No. 2014B1234, 2015A1070, and 2015A1101). The coordinates and structure factors for the U14-NTD structure has been deposited in the Protein Data Bank under the accession number 5B1Q.
|
10.1371/journal.pntd.0006776 | Implementing active community-based surveillance-response system for Buruli ulcer early case detection and management in Ghana | Buruli Ulcer (BU) is one of the most neglected debilitating tropical diseases caused by Mycobacterium ulcerans, which causes considerable morbidity and disability. Building on earlier findings that community-based interventions could enhance case detection and reduce treatment dropout and defaulter rates, we established an active surveillance-response system in an endemic sub-district in the Ga West municipality of Ghana to enhance early case detection, diagnosis and treatment to reduce or eliminate severe ulcers and its related disabilities.
We established surveillance response system, implemented in collaboration with the sub-district disease control officers, selected clinical staff and trained community-based volunteers. The active community-based surveillance- response system was implemented for 12 months. Also, pre and post intervention surveys were conducted to document any change in perceptions on BU in the study population over the period. The baseline and endline surveys were conducted in August 2016 and August 2017 respectively.
On average, each person was seen 11 times in 12 months. In all 75 skin lesions were detected during surveillance rounds, out of which 17 were suspected to be BU and 12 out of the 17 were confirmed as BU using Polymerase chain reaction (PCR). Out of the 12, five, three and four were categories I, II and III lesions respectively. Physical examination was done on 94% of the people seen during the surveillance rounds. Knowledge on BU has also increased in the communities at the end of the study.
The findings from this study have demonstrated that it is possible to establish surveillance-response system for BU and by extension, other neglected tropical diseases to enhance control and elimination efforts through the use of community-based volunteers.
| The study revealed that it is feasible to train periphery health workers and community-based volunteers to implement a community-based active surveillance–response system for early buruli ulcer (BU) case detection, diagnosis and treatment at outpatient clinics. At the end of 12 months follow-up, there were 36084 surveillance person contacts made with physical bodily examination done on 93.77% of them. The average surveillance contact per person was 11.3 (91.7%). We believe that this was achieved largely because of the use of community-based volunteers for the surveillance visits. Given the high number of non-BU skin lesions detected during the surveillance period, it is recommended that any BU surveillance-response system must be an integrated one to aid the detection, diagnosis and treatment of other skin conditions to make it more cost effective, this has become even more imperative because the number of BU cases have been declining in most endemic communities in Ghana, since the introduction of antibiotics treatment.
| Buruli ulcer (BU) is one of the most neglected debilitating tropical diseases caused by Mycobacterium ulcerans. Though, BU-related fatality is very rare, it causes considerable morbidity and disability [1–3], which lead to stigma and anguish among infected individuals and affected families. The disease occurs mainly in remote areas, deprived of basic social infrastructure like health care facilities, potable water and accessible roads [1]. Most endemic countries lack efficient reporting systems, making the assessment of the disease burden unclear [2]. Buruli ulcer has recently received some attention focusing mainly on diagnosis, transmission and clinical management [4–6]. However, there is no active community-based surveillance-response system to generate the essential data necessary to capture and treat cases early enough to prevent complications that contribute so much to morbidity and disability, and cost of treatment to both the health system and affected families. In order to institute an effective public health response to the problem of BU, a strong surveillance system is needed to systematically collect, analyse and interpret data on the disease [7–9].
The epidemiology of BU in endemic countries is not entirely known, due to several factors including the focal distribution of cases, late reporting of cases and lack of health facilities among others in endemic countries of Africa [10]. In Ghana, the first passive surveillance system reported about 1,200 BU cases between 1993 and 1998 and more than 9,000 BU cases were also reported between 2004 and 2014 [9, 11]. A nation-wide active case search that was conducted in 1999 found BU in all the 10 administrative regions of Ghana with an overall prevalence of 20.7 per 100,000 of the population [9]. Currently, BU control in Ghana is mainly through early case detection [9, 12] and clinical diagnosis at peripheral health facilities designated by the National Buruli Ulcer Control Program (NBUCP) followed by laboratory confirmation and subsequent antimycobacterial therapy.
The mode of transmission of the pathogen is still elusive and therefore control relies mainly on case detection and treatment with Streptomycin and Rifampicin for eight weeks, followed by surgery to speed healing and correct deformities [11]. The success of this treatment modality depends very much on detecting suspected cases early for diagnosis and treatment at health facilities [12–14]. However due to socio-cultural and economic factors most cases are detected very late with large wounds with massive cell destruction by the cytopathic toxin, mycolactone [11, 15].
It is known that BU patients in West Africa do not typically seek care from health facilities or do so late with severe ulcers coupled with disabilities leading to under reporting of the disease burden at health facilities [5, 14–17]. However, it has been reported that community-based interventions could enhance case detection and reduced treatment dropout and defaulter rates [17, 18].
Huge resources are being invested to develop new diagnostic tools for BU and to improve on its clinical management. However the discovery of new diagnostic tools and improved clinical management techniques could only be useful when suspected patients seek health care for their conditions, especially at the early stages of the disease, before debilitating complications set in. When identified early, BU can be treated with a high degree of success with rifampicin and streptomycin. This antibiotic combination treatment is noted to have a positive impact on treatment outcomes as it has the potential to cure small lesions and limit surgery for larger lesions [19–21].
The overall aim of this study was to test the feasibility of training periphery health workers and community-based volunteers to implement community-based active surveillance–response system for early BU case detection, diagnosis and treatment on outpatient basis in an endemic sub-district in Ghana.
The study was reviewed and approved by the Institutional Review Board (IRB) of the Noguchi Memorial Institute for Medical Research (NMIMR) with Federal Wide Assurance Registration FWA 00001824. The protocol was assigned certified protocol number (CPN) 075/14-15. Written informed consent was taken from all participants (Informed consent from all adults aged 18 years and above, Parental consent for all children under 18 years old and Child assent from all children aged between 12 and 17 years.)
This was an epidemiologic study designed to test and evaluate a community surveillance-response system in the study area for 12 months. It was a longitudinal study with baseline and endline surveys to compare pre- and post- implementation perceptions among the population.
The Ga West Municipality with Amasaman as its capital has an estimated population size of about 215,824 with a growth rate of 3.4 percent [22]. The major economic activity in the municipality is farming, employing about 70% of the people. Other economic activities include fishing, stone quarrying and petty trading.
The district has the highest number of reported BU cases in the Greater Accra Region and is the fifth most BU endemic in the country in 2002. About 1000 cases of BU are reported in Ghana yearly, giving a nationwide prevalence of 20.7/100,000, in 1998. However, the district level prevalence of the Ga South district, where this study took place, was 87.7/100,000 population [23, 24].
The District Hospital at Amasaman has a specialized unit for BU treatment but due to poor road networks and socio-cultural factors, most cases report at the health facility late. The Amasaman hospital continues to be the main health services provider in the district with a few health posts, private clinics, and family planning and maternity homes doted within the municipality. Health care to the rural communities is mostly provided by the Ghana Health Service through monthly outreach services. Among the top five common diseases prevalent in the District are malaria, skin diseases, diarrhoea, HIV/AIDS (Human immunodeficiency virus/ Acquired immunodeficiency syndrome) and BU.
Ten communities namely; Kojo-Ashong, Onyansana, Otuaplam, Yahoman, Okushibiade, Adeyman, Kramo, Domsampaman, Kwashikuma, Odumtia/Akwakyere were identified to participate in the study. These communities were selected with the support of the district health management team and the National BU control programme. The Ga West municipality was selected because it has not only the highest number of reported BU cases in the Greater Accra Region and the fifth most BU endemic in the country but also continue to report the worst forms of the disease, category three ulcers in Ghana. The 10 communities selected were identified by the municipal health directorate as the most endemic communities in the district. Census was conducted in each of the selected communities to register everybody, which then formed the target populations for the establishment of the community-based surveillance -response system.
The current BU control in Ghana is based on the WHO recommended first line treatment for BU using oral rifampicin (10 mg/kg) plus intramuscular streptomycin injection (15 mg/kg), both given daily for 8 weeks under supervision coupled with passive surveillance. Case finding is based on passive surveillance coupled with occasional case search in communities with no on-going active surveillance in endemic communities for early case identification, though early treatment is being promoted.
At baseline, with an estimated population of 4,000 in the 10 selected communities and assuming that 50% of the people will be willing to participate in the interviews and 5% confidence limits and the design effect of 1.5, we arrived at a sample size of 570 (57/cluster) However, we sampled 60 respondents from each of the 10 communities (600 participants). At the endline survey, a known population size of 3255 was used for the calculation giving a sample size of 52/cluster, thus a total of 520 respondents in the 10 communities, however 526 respondents were interviewed. Participants were randomly selected from the adult population, using a community register, proportionally to represent both sexes.
Data entry and analyses were done using EpiInfo version 7. Findings were presented in descriptive statistics, especially frequencies and percentages. We compare proportions (percentages) from baseline and endline to determine any difference between the two data points. A difference of less than 0.05 (P<0.05) was considered statistically significant.
A total population of 3070 in 837 households in 10 communities was registered during the census. However, by the end of the 12 month, the population has increased by 185 (6.0%) to 3255. Beside the natural population growth, we believe that the increased in population was due mainly to people moving to settle in the area, which is not far from Accra the capital city of Ghana. Majority of the study participants were females (52.8%). At the end of 12 months, there were 36084 surveillance person contacts made with physical bodily examination done on 33835 (93.8%). The average surveillance contact per person was 11.1 (92.3%) out of the maximum number of 12 expected (Table 1).
Bodily examination was not done on 2249 (2249/36084 x 100 = 6.0%), mainly because participant were; not at home or travelled (70.7%), moved out of the community permanently (19.4%), refused to be examined 136 (6.0%) and others like infants or sick persons made up of only 3.9%. During the 12 months follow up period, a total of 75 skin lesions were encountered, out of which 17 (22.7%) were suspected to be BU by the volunteers. Out of the 17 suspected cases, 12 (70.6%) were confirmed as BU using PCR, making the clinical judgment of the volunteers very good. Out of the 12 confirmed cases, five, three and four were categories I, II and III respectively (Table 2). Thus, two out of three cases were picked from the community early and these could be easily managed at the peripheral health facilities at a lower cost to both the individuals and the health system. All but one (three out of four) of the category III lesions were invited by their relatives who were participating in the study to migrate into the study area in 2017 so that they could benefit from the project activities, especially facilitating laboratory diagnosis and treatment at the nearest health facility. The remaining category III case was hidden in the community for a long time. She reportedly went to the clinic once, many years ago, but had decided against biomedical treatment, and instead resorted to herbal or traditional treatments at home, however, through the surveillance response system, she was rediscovered, diagnosed and helped to receive treatment on an outpatient basis, and her wound has healed. The left leg (75.0%) was the dominant site of confirmed BU lesion (9 out of 12) with one each on the right leg, left and right arms respectively.
Majority (62.5% at baseline and 64.5% at endline) of the respondents had primary level education. As expected, majority (71.3% at baseline and 61.6% at endline) of the respondents belong to the Ga speaking ethnic group. The Christian religion was the most professed religion reported among respondents; 81.2% at baseline and 84.6 at endline (Table 3). In line with the dominant ethnic groups in the area, majority of the respondents (72.8 at baseline and 66.2% at endline) referred to the disease in the Ga/Agangbe local language as Odonti hela/ Aboagbonyo. Others include Detsifudor/detsifubi in Ewe language (12.4% at baseline and 14.3% at endline) and Kisi kuro/Asawa kuro in Akan language (2.5% at baseline and 3.2% at endline). Also, 11.2% at baseline and 15.0% at endline referred to it as Buruli. It is worth reporting that all the dominant local names could be translated literally to mean cotton wool wound or bad wound.
Majority of the respondents (89.8% at baseline and 92.6% at endline) said they know about Buruli ulcer as a disease that affects people in the community and have local names for it. The dominant local names reported—adonti hela (Ga), detsifu dor (Ewe) and asawa kro (Twi), could be translated literally to mean cotton wool wound and this could be linked to the cotton wool-like necrotic tissues that are usually found around the edges of BU wounds. The BU related knowledge was acquired from varied sources. The most commonly reported source of knowledge was to know someone with the infection, 71.2% at baseline and 53.6% at endline (Table 4).
Majority (69. 8% at baseline and 73.4% at endline) of the respondents said they could recognize BU at an early stage before it becomes a wound/sore or ulcerated. Among those who said they could recognize early BU, Nodule (87.1% at baseline and 96.9% at endline) was the most commonly reported sign. This was followed by Itching (10.3% at baseline and 9.8% at endline). Others were; Papule (1.9% at baseline and 1.3% at endline), Plaque (1.7% at baseline and 2.9% at endline) and Oedema (1.2% at baseline and 1.0% at endline).
Various treatment options were reported; chiefly among them was Hospital/clinic treatment (78.3% at baseline and 72.4% at endline). Others were; Traditionalists/Herbalists (34.4 at baseline and 22.2% at endline), Home prepared herbal medicine (17.5% at baseline and 5.7% at endline, self-medication with biomedicine (3.0% at baseline and 1.7% at endline with faith healers reported by two people only at baseline.
Majority (86.2% at baseline and 74.1% at endline) mentioned that BU can be prevented. Prominent prevention methods mentioned included; Cleaning the Environment (30.6% at baseline and 50.8% at end-line), Avoid bathing in the river (32.5% at baseline and 36.9% at end-line) and Regular use of biomedicine; 38.3% at baseline and 27.7% (Table 5).
At baseline, majority (62.8%) of the respondents said they were not aware of any active BU control activity in their communities. However, this had changed at the endline, where majority (55.1%) reported spontaneously that they were aware of a control activity going on in the study communities. When respondents were asked to explain the control activities in their own words, the dominant statement is represented in the following narrative “I know someone who moves from house to house to examine people for the disease and then refer those with the disease to the clinic for treatment.” This statement indicates that people were aware of the surveillance-response system that was implemented in these communities.
As expected, malaria was reported by the majority (88.5% at baseline and 86.88% at endline) as the single most common disease in all the study communities. This was not surprising as the area is noted for high malaria prevalence in the Greater Accra region, where malaria is generally low due to its urban nature. Though, only few respondents mentioned Buruli ulcer as a common disease (6.8% at baseline and 3.8 at end-line), it is important to know that it was considered as an important health problem in the study area. Other health problems mentioned at both baseline and endline were waist pains, stomachache, headache, eye problem and hypertension.
All respondents at endline testified that volunteers come to their homes on monthly basis. However, 92.01% reported that the volunteers did talk to them about Buruli ulcer whenever they visited. Also, 91.82% of the respondents mentioned that they were examined by the volunteers whenever they visited. Majority (91.0%) of the respondents said that they were happy with the work of the volunteers. Interestingly, 92.6% of the respondents indicated that they would like the visit to continue with the majority (67.6%) opting for once in a month visit, just as was done during the project implementation.
This study has demonstrated that it was possible to establish surveillance response system to conduct physical examination on participants on monthly basis. In the context of BU, this may enhance early case detection, diagnosis and treatment. The surveillance-response system implemented was not only able to pick cases at early stages to eliminate or, at least, reduce severe and debilitating ulcers associated with late reporting at health facilities, which causes morbidity and disabilities, but also rediscovered an old ulcerated case that was hidden from the view of the health system. This confirmed the report that patients are not only reporting late at health facilities but many more maybe hidden in the communities [16] and this may require active community-based surveillance to discover and support their diagnosis and treatment.
The surveillance system was able to pick all categories of BU lesions based on WHO (2008) categorization; category I—a single lesion <5 cm in diameter; category II—a single lesion between 5 and 15 cm in diameter; category III—a single lesion >15 cm in diameter, multiple lesions, lesion(s) at critical sites (eye, breast, genitalia) and osteomyelitis [23].
Proper community engagements and training of community-based volunteers, nominated by community members themselves to implement the surveillance-responses system has contributed to the acceptance and participation of the people in the project activities, where on average, each person received at least 11 surveillance contacts with physical examination done on over 90% of them. It was possible to achieve this because efforts were made to follow all community protocols and encouraged the people to see the project as their own. It was reported from a malaria study that, community participation is vital for the success of community based interventions and to achieve this may require full engagement of community members in the process from the start of the project, making them to claim ownership of it by observing existing community protocols and respecting established hierarchy of power within the study community [25].
With reduction in BU cases in endemic communities, other skin conditions must be integrated into BU surveillance-response system that is established to make it more cost effective. In our study, the other skin conditions were referred to the health facilities for management and most of them were treated without any laboratory diagnosis. Though the patients were treated, it will have been better if they were taken through diagnosis to know exactly what conditions were being treated. This may be useful to determine the kinds of lesions circulating in the study communities, to aid the design of any control and prevention strategies.
The socio-demographic characteristics of the study area has remained virtually the same as was reported in earlier studies [12,15,17,18,26], where the Ga ethnic groups were dominant with majority of them identifying with the Christian religion with primary level education. The knowledge of the disease was very high among respondents, gained mostly through the experiences of knowing someone with the infections. It was interesting to note that the socio-demographic characteristics did not affect the level of BU knowledge in the study area, implying that BU-related knowledge is evenly spread among the population. It was encouraging to know that majority of respondents could recognized early suspected lesions and this must be promoted to encourage early reporting at health facilities to promote early diagnosis and treatment as this may help reduce the disease burden on the individual, the affected family and the health system, since the current treatment is very effective when the infection is treated early [21].
Community-based volunteers have been used in Ghana in varied ways to support community-based public health service delivery. However, in the context of BU control, they have been involved in passive case finding in their communities and the added value for using them in this study was their involvement in active case finding through surveillance home visits, which makes it difficult for them to miss any case in the community. The volunteers were motivated with a token, which is about $20 every month to cover their communication and any other incidental cost that they might have incurred during surveillance rounds. In addition, bicycles were given to them to aid mobility during monthly rounds and this has been attested to by respondents as very helpful. The motivation was not solely in the token given to them but also in the fact that they saw themselves as contributing to bringing good health to their people and it was encouraging to hear some respondents asking for the continuation of the project, so that they could be visited at home monthly. This is an indication that with good community entry, respect for local authorities and rapports, it is possible to conduct community-based surveillance with over 90% bodily examinations in BU endemic communities. Since efforts in determining the transmission rout remains elusive, we could take advantage of the acceptance of the community-based surveillance-response system to not only pick cases early but also discover old but hidden ulcers for diagnosis and treatment.
A major limitation of the study was that, it was designed only to confirm BU lesions but not to test for any other cutaneous NTDs and this must be addressed in subsequent studies, given the current downward trend of BU cases in endemic communities including our study area. Secondly, the study period of 12 months was rather short for a longitudinal study of this nature.
The study was setup to determine the feasibility of training periphery health workers and community-based volunteers to implement a community-based active surveillance–response system for early buruli ulcer case detection, diagnosis and treatment at outpatient clinics. At the end of the study, it can be concluded that, it is feasible to train periphery health workers and community-based volunteers to carry out community-based surveillance-response system for early BU case detection, diagnosis and treatment. Given the high number of non-BU skin lesions detected during the 12 month surveillance period, it is recommended that any BU surveillance response system must be an integrated one to aid the detection, diagnosis and treatment of other skin conditions. This will make such an intervention more cost effective as the number of BU cases continues to decline in most endemic communities in Ghana after the introduction of antibiotics treatment.
|
10.1371/journal.pcbi.1004130 | Neuroblastoma Tyrosine Kinase Signaling Networks Involve FYN and LYN in Endosomes and Lipid Rafts | Protein phosphorylation plays a central role in creating a highly dynamic network of interacting proteins that reads and responds to signals from growth factors in the cellular microenvironment. Cells of the neural crest employ multiple signaling mechanisms to control migration and differentiation during development. It is known that defects in these mechanisms cause neuroblastoma, but how multiple signaling pathways interact to govern cell behavior is unknown. In a phosphoproteomic study of neuroblastoma cell lines and cell fractions, including endosomes and detergent-resistant membranes, 1622 phosphorylated proteins were detected, including more than half of the receptor tyrosine kinases in the human genome. Data were analyzed using a combination of graph theory and pattern recognition techniques that resolve data structure into networks that incorporate statistical relationships and protein-protein interaction data. Clusters of proteins in these networks are indicative of functional signaling pathways. The analysis indicates that receptor tyrosine kinases are functionally compartmentalized into distinct collaborative groups distinguished by activation and intracellular localization of SRC-family kinases, especially FYN and LYN. Changes in intracellular localization of activated FYN and LYN were observed in response to stimulation of the receptor tyrosine kinases, ALK and KIT. The results suggest a mechanism to distinguish signaling responses to activation of different receptors, or combinations of receptors, that govern the behavior of the neural crest, which gives rise to neuroblastoma.
| Neuroblastoma is a childhood cancer for which therapeutic progress has been slow. We analyzed a large number phosphorylated proteins in neuroblastoma cells to discern patterns that indicate functional signal transduction pathways. To analyze the data, we developed novel techniques that resolve data structure and visualize that structure as networks that represent both protein interactions and statistical relationships. We also fractionated neuroblastoma cells to examine the location of signaling proteins in different membrane fractions and organelles. The analysis revealed that signaling pathways are functionally and physically compartmentalized into distinct collaborative groups distinguished by phosphorylation patterns and intracellular localization. We found that two related proteins (FYN and LYN) act like central hubs in the tyrosine kinase signaling network that change intracellular localization and activity in response to activation of different receptors.
| Neuroblastoma arises from cells of the neural crest, a population of multipotent, migrating cells that differentiate into neurons in the peripheral nervous system, melanocytes, and structural cells [1]. Neuroblastoma represents 7–10% of childhood cancers and about half of all infant cancers. Positive prognosis ranges from 95% to 10% depending on age, markers expressed in tumor cells, and stage of progression. 70% of neuroblastomas are already metastatic at diagnosis. There is compelling evidence that stalled or incomplete cell differentiation is the primary defect that gives rise to this cancer [2–6]. Neural crest cells appear to restrict their range of cell fate choices in sequential steps [7,8], and the profound heterogeneity in neuroblastoma is caused by a failure to differentiate at different stages. Neuroblastoma tumors and cell lines thus represent a snapshot of failed differentiation at different stages in the neural crest sympathoadrenal lineage [2,4,7,8]. Anaplastic lymphoma kinase (ALK), a receptor tyrosine kinase (RTK), is frequently mutated and activated in both familial and spontaneous neuroblastomas, suggesting that this receptor can prevent a key differentiation step in neural crest cells [9–15]. Incompletely differentiated cells may give rise to a proliferating population when mutations occur that allow checkpoints in the cell division cycle and mechanisms of programmed cell death to be bypassed. The tragic outcome is too often a metastatic cancer with poor prognosis.
To address this clinically challenging problem, a greater understanding of the signaling mechanisms that are active in neural crest and neuroblastoma is required. Tyrosine kinase signaling networks play a major role in governing cell differentiation, including in neuroblastoma [16]. There are 90 tyrosine kinases in the human genome; 58 of these are receptor tyrosine kinases [17,18], many of which have unknown functions. Src Homology 2 (SH2) domains (and one-fifth of phosphotyrosine-binding or PTB domains) mediate selective protein–protein interactions with proteins phosphorylated on tyrosine residues, and thus mediate assembly of phosphotyrosine signaling networks [19]. The metazoan evolution of multicellular organisms coincided with expansion of tyrosine kinases, protein tyrosine phosphatases, and SH2 domains, which suggests that tyrosine kinase signaling mechanisms play a major role in cell differentiation [20–22]. Unfortunately, the system isn’t foolproof, and cancer results when the dynamic assembly of signaling complexes goes awry [23]. Thus, the complexity of kinase-substrate and other protein-protein interactions in tyrosine kinase signaling pathways is important to understand because these pathways govern the choice between differentiation and cancer.
Tyrosine kinase signaling mechanisms are intimately intertwined with mechanisms that govern protein interactions in endocytosis. Src Homology 3 (SH3) domains are among the most abundant protein domain modules encoded by eukaryotic genomes; over 300 SH3 domains are found in 213 human proteins [24,25]. SH3 domain-containing proteins, which typically bind to proline-rich motifs [26], are functionally linked to both endocytosis and tyrosine kinase signaling pathways [24]. SH3-domain-containing proteins play a role in endocytosis that is conserved in yeast, worms, and humans [26,27]. SH3 proteins may also contain other domains (e.g., kinase, phosphatase, GTP exchange, GTPase activating) to perform conserved functions in endocytosis and cytoskeletal dynamics, and, in metazoans, RTK signaling [28,29]. 36 human proteins contain one SH2 domain and one or more SH3 domain(s) (SH2-SH3 proteins) [25]. Most SH2-SH3 proteins are phosphorylated on multiple sites on tyrosine as well as serine and/or threonine residues. Half of them also have tyrosine kinase domains, e.g., the SRC-family kinases (SFKs). Interactions between proteins that contain SH2 and SH3 domains indicate that tyrosine kinase signaling and endocytosis are linked, and there is good evidence that endocytosis and signal transduction in general are integrated [30,31].
To identify patterns in tyrosine phosphorylation in neuroblastoma, we acquired phosphoproteomic data from 21 neuroblastoma cell lines and cell fractions including endosomes and detergent-resistant lipid rafts as previously characterized [32,33]. New approaches were devised to analyze these data. We previously experimented with different dimensionality reduction and clustering techniques and validated methods that effectively resolve clusters from lung cancer phosphoproteomic data [34]. An important first step is to represent missing values as “data not available” instead of zero in spectrometry data. By combining pattern recognition techniques with gene ontology (GO) and protein-protein interaction (PPI) data, we learned that clusters that contain interacting proteins are likely to indicate functional signaling pathways [34–40]. Here, we extend methods that employ graph theory and pattern recognition algorithms to introduce techniques to visualize data structure, namely a cluster-filtered network (CFN) and co-cluster correlation network (CCCN). We focussed primarily on proteins containing tyrosine kinase, tyrosine phosphatase, SH2 and SH3 domains, which collectively we call phosphotyrosine network control proteins (PNCPs).
To identify patterns in tyrosine phosphorylation in neuroblastoma, we analyzed tyrosine phosphoproteomic data acquired from 21 neuroblastoma cell lines using immunoprecipitation of tyrosine phosphorylated peptides as previously described [41,42]. Four cell lines [SH-SY5Y, LAN-6, SMS-KCN, and SK-N-BE(2)] were selected for further studies because of their different point mutations in ALK, p53 status, RTK expression, morphology, and growth patterns. These cells were fractionated to isolate endosomes and detergent-resistant lipid rafts [32,33], and analyzed under different conditions that changed the state of their signaling pathways. Quantification of immunoprecipitated phosphopeptides was obtained from the peak intensity of each peptide (from the MS1 spectrum of the intact peptide before fragmentation for MS/MS analysis) [41,42].
We experimented with different ways to analyze the mass spectrometry data (described in detail in Materials and Methods). For the first analysis described below, phosphopeptide amounts were summed for each protein in each sample, with the exception of the SRC-family kinases (SFKs), where the C-terminal inhibitory phosphorylation was summed separately and given the names SRC_i; LYN_i; FYN_i; and YES1_i. This provided an overview of which proteins were present and phosphorylated together in the same samples. For the second analysis, phosphopeptides were summed into individual phosphorylation sites, which were then clustered. Clustering data were obtained by treating all samples mathematically as different states in the neuroblastoma system. We describe analysis of the whole dataset first, then subsets of the data, focusing on signaling proteins in endosomes and detergent-resistant membranes.
1622 phosphorylated proteins were identified in all neuroblastoma samples (S1 Fig; S3 Dataset). 1203 of these were tyrosine phosphorylated, identified from peptides immunoprecipitated using an anti-phosphotyrosine antibody. 557 proteins were identified from phospho-AKT substrate immunoprecipitation; of these 419 were unique, and 138 were dually phosphorylated proteins also found in the phosphotyrosine data. Due to limits in mass spectrometric detection of peptides [43–47], these data were not an exhaustive determination of all phosphorylated proteins in all samples. To ask whether these data were complete enough for analysis of signaling pathways, we employed graph theory, which describes the properties of networks [35,38]. S1 Fig shows a network constructed using proteins identified in neuroblastoma phosphoproteomic data as nodes, and protein-protein interaction (PPI) edges merged as described [34]. We found that the entire neuroblastoma phosphoproteomic network of 1622 proteins and 18728 interactions is dense enough to have the structure and properties expected of biological networks, including clusters that can be usefully interpreted (S2 Fig). PPI databases are biased towards proteins best studied in the scientific literature [36–38], and not all protein-protein interactions in PPI databases may occur in neuroblastoma cells. Nevertheless, PPI network analysis indicates that the phosphoproteomic data are complete enough to examine further to gain insight into signal transduction pathways that are active in neuroblastoma (S2 Fig).
We hypothesize that proteins containing tyrosine kinase, tyrosine phosphatase, SH2 and SH3 domains (PNCPs) will collectively initiate and control phosphotyrosine signaling pathways [19,24]. In neuroblastoma phosphoproteomic data, we detected 31 phosphorylated RTKs out of 58 in the human genome (S3 Fig); 41 of 110 SH2-domain-containing proteins; 12 out of 38 (or 107 possible, based on open reading frames in the human genome) proteins containing the tyrosine phosphatase (PTPc) domain; and 61 out of the 216 human SH3-domain containing proteins. There are 36 proteins in the human genome that contain both SH2 and SH3 domains and 17 of these were detected in neuroblastoma phosphoproteomic data.
These data indicate that neuroblastoma cell lines express and phosphorylate a large fraction of the PNCPs in the human genome. This remarkable diversity in phosphotyrosine signaling pathways likely represents a snapshot of signaling pathways activated in the sympathoadrenal lineage of neural crest that gives rise to neuroblastoma at different stages of development [2–6]. The robust expression of RTK pathways that are known to function in neural crest differentiation suggests the hypothesis that neuroblastoma cells might be multipotent despite being selected for proliferation in culture. To test this hypothesis we transplanted neuroblastoma cells in to the developing neural tube of live chick embryos and indeed found that they were capable of both migration and terminal differentiation (S4 Fig). Notably, four different transplanted human neuroblastoma cell lines [LAN6, SK-N-BE(2), SMS-KCN, and SH-SY5Y] migrated to neural crest target sites, incorporated into the developing ganglia, and expressed neuronal markers specific to mature afferents (S4 Fig). The potential to migrate along the stereotypical neural crest migration pathways, and differentiate into most neural-crest-derived cell types, suggests that many of the RTK signaling pathways that control differentiation and migration were generally functional in these neuroblastoma cell lines. Thus, our phosphoproteomic data has relevance to pathways active in neural crest from which neuroblastoma is derived, and warrants detailed analysis.
We developed new methods to analyze proteomic data based on the hypothesis that data structure can be described using a combination of graph theory and pattern recognition techniques. The first key step was to recognize that missing data, which are common in mass spectrometry data due to stochastic variation in phosphopeptide detection, should not have a value of zero [34]. The next key step was to represent different statistical relationships by proximity on two- or three-dimensional graphs using an effective dimension reduction, or embedding, technique, t-distributed stochastic neighbor embedding (t-SNE) [48,49]. Clusters were identified by proximity on resulting three-dimensional data structures (embeddings) using a minimum spanning tree, single linkage method [34,50]. 75–80 clusters were identified from each embedding based on dissimilarity calculated in different ways (S1 Movie; S1 Dataset). Clusters were evaluated internally, based on the primary data, and externally, using PPI and gene ontology (GO) databases (S5 Fig). These evaluations confirm that these methods effectively resolve meaningful clusters as previously described [34].
We experimented with different approaches to use these clusters to define signaling pathways active in neuroblastoma. One approach was to apply a “hard” filter, or exclusive approach to identify groups of proteins that co-cluster from two or more dissimilarity representations. This exclusive approach separates groups of proteins that are most likely to define core units of signaling pathways [34]. Alternatively, an inclusive approach treats clusters derived from different embeddings as equally valid and therefore allows overlap between cluster membership. This inclusive approach recognizes that signaling pathways use common effectors. We show results from each of these approaches in turn.
For the first, exclusive cluster analysis, we focused on PNCPs and proteins whose phosphorylation pattern was statistically most similar determined by both Euclidean distance and Spearman correlation (Figs 1 and S6). Heat maps (Fig 1 and S6, right) indicate that the phosphorylation patterns in the primary data are reasonably consistent within each cluster. The RTK, ALK, clustered with two other RTKs (FGFR1, PDGFRA), activated FYN, and LYN phosphorylated on the C-terminal inhibitory site (LYN_i; Fig 1A). The tyrosine kinase, FAK (PTK2), and the adaptor molecules BCAR1, SHC1 and CBLB were included in this group of PNCPs. We also noted other clusters that suggest interactions among phosphorylated tyrosine kinases: IGF1R with LYN, FER, the phosphatase PTPN11/SHP-2, and the tyrosine kinase TNK2, whose interactions with other proteins in this group have not been previously characterized (Fig 1B). In addition, we found that EGFR and EPHB3 clustered with inhibited FYN and SRC as well as the SH3, SH2 containing tyrosine kinase, ABL1, and MPP5, a protein with PDZ, SH3, and guanylate kinase domains whose interactions are not characterized (Fig 1C). Examples of other clusters identified using this hard filter are shown in S6 Fig. These clusters define phosphorylated proteins most commonly phosphorylated together in the same samples in this data set, which suggests possible interactions among signaling proteins that were previously unknown. Assignment of proteins to one cluster should not be viewed as evidence for excluding it from participating in a signaling pathway identified in another cluster, however [34].
An alternative, inclusive approach is to recognize possible relationships defined by different measures of statistical similarity. Clusters derived from t-SNE applied to Spearman, Euclidean, and hybrid Spearman-Euclidean (SED) embeddings were typically overlapping but not identical, yet reasonably close in their ability to resolve meaningful clusters as determined by external and internal evaluations (S5 Fig; [34]). This suggests that statistical relationships independently defined by Euclidean distance or Spearman correlation are equally valid. Using this inclusive method that recognizes clusters derived from different embeddings had the advantage that it allows overlap between cluster membership, which makes sense biologically for these data because signaling pathways overlap and converge.
We employed the inclusive clustering strategy to filter protein interaction edges to obtain the cluster-filtered network (CFN) shown in Fig 2. In this graph, only edges among proteins that co-clustered based on Spearman, Euclidean, or hybrid Spearman-Euclidean (SED) dissimilarity are shown. This CFN data structure is useful because graph layouts that treat edges like springs (edge-weighted, spring embedded; force-directed) aggregate proteins that share a statistical relationship and interact with one another, so nearest neighbors are likely to represent functional groups (regions highlighted in Fig 2).
An alternative visualization of data structure is a co-cluster correlation network (CCCN; S7 Fig). In this graph, edges represent positive (yellow) or negative (blue) correlation, filtered to show only edges among proteins that clustered together and have a Spearman correlation coefficient greater than the absolute value of 0.5. The networks in Figs 2 and S7 are complementary because they apply a different filter to clustering results. Proteins that interact with one another may not tightly correlate, and co-clustered proteins that do tightly correlate may not have been studied previously for evidence of interactions. These filtered networks thus prune cluster members that have no evidence for interaction and do not tightly correlate with others in the group, yet allow potential interactions among pathways to be studied because overlapping cluster membership is defined by different embeddings.
Exploration of these networks reveals potential functional interactions among signaling proteins defined by the structure of neuroblastoma phosphoproteomic data. We noted two groups of highly phosphorylated RTKs that clustered together (Fig 3). Networks in Fig 3 show only positive correlation (yellow) and PPI (grey) edges between RTKs and co-clustered effector proteins, with proteins that link to three or more receptors grouped in the center of the graphs (Fig 3). The similarity in phosphorylation patterns for proteins in these groups can be seen in heat maps of the primary data (S8 Fig). Co-clustering of ALK with PDGFRA, FGFR1, and IGF1R (through co-clustering with FGFR1) is indicative of a collaborative relationship (Fig 3A). Similarly, EGFR co-clusters with PDGFRB, EPHA2, EPHB3, and DDR2 (Fig 3B), indicating that these RTKs form a separate collaborative group. While different RTKs within these collaborative groups share a number of co-clustering downstream proteins in common, the only effector proteins in common between these two collaborative groups are PIK3R2, FYN, and the SFK scaffold protein, PAG1 [51].
The following general conclusions can be made from these analyses so far. Clusters that contain proteins that interact with one another, identified using statistical relationships from phosphoproteomic data, likely indicate functional signaling pathways. New potential interactions are suggested when strong clustering is observed among proteins whose physical interactions have not been previously characterized (e.g., TNK2 and MPP5 in Fig 1). Common patterns of phosphorylation in neuroblastoma samples suggests collaboration among RTKs within functional groups (Fig 3). Since activation of different RTKs was associated with different states of activation and inhibition of different SFKs, particularly FYN and LYN (Figs 1 and 3), we next examined how stimulation or inhibition of RTKs affected phosphorylation of other tyrosine kinases.
RTK activation affects other RTKs, SFKs, and other tyrosine kinases. To examine the effects of RTK stimulation on other tyrosine kinases, we compared phosphoproteomic data from cells treated to influence RTK activity, or not treated, in the same experiment. Fig 4A shows tyrosine kinases whose total phosphorylation changed more than two-fold under experimental conditions where RTKs were stimulated by ligand or ALK was inhibited. For example, NGF treatment caused a more than twofold increase in total phosphorylation of DDR2, and more than fivefold decrease in phosphorylation of PDGFRA in both LAN-6 and SH-SY5Y cells. EGF treatment of SK-N-BE(2) cells activated EGFR and stimulated EPHA3 phosphorylation about 3-fold (Fig 4A). These data indicate that stimulation of one RTK affects the phosphorylation state of other RTKs in neuroblastoma cell lines.
Changes in inhibitory phosphorylation of LYN and SRC were also observed (Fig 4A, LYN_i; SRC_i), so individual phosphorylation sites on SFKs and other kinases were examined further. Phosphopeptides were assigned to phosphorylation sites based on peptide sequence homology (see Materials and Methods). The data revealed that both activating (SFK Y411-426) and inhibitory (SFK Y508-531) phosphorylation sites on the SFKs LYN, FYN, YES1, and SRC were significantly affected in different ways by treatments that influence RTK activity (Fig 4B). For example, the LYN inhibitory phosphorylation (LYN 508) was reduced by NGF treatment and increased by EGF treatment. In contrast, FYN inhibitory phosphorylation (FYN 531) was increased by NGF in two cell lines (Fig 4B). These data suggest the hypothesis that activation and inhibition of LYN and FYN distinguishes responses to different RTKs (Figs 1 and 3).
Tyrosine phosphorylation of RTKs is generally thought to be a measure of activation, but differences in different RTK phosphorylation sites were seen in these experiments. For example, NGF treatment both increased and decreased phosphorylation on different sites on EGFR, RET, IGF1R, ALK, and other RTKs in LAN-6 and SH-SY5Y cells (S9A Fig). Some variations in individual phosphorylation site responses to treatments were also observed for other tyrosine kinases (S9B Fig), but they were not as dramatic as those of SFKs (Fig 4B).
These data indicate that different RTKs initiate signaling mechanisms to cause distinct phosphorylation patterns on other tyrosine kinases, including RTKs and SFKs. Combined with the clustering patterns shown in Figs 1 and 3, the data suggest the hypothesis that SFKs, particularly FYN and LYN, discern and integrate signals from different RTKs. We hypothesized that functional interactions among these signaling proteins may occur in specific intracellular locations, namely endosomes and lipid rafts, and therefore we performed phosphoproteomic analyses on these fractions.
We asked whether particular signaling proteins were enriched in endosomes and detergent-resistant membranes (DRMs). RTKs are present in endosomes that can be distinguished from other types of receptors by size and density (S10 Fig) [32]. Phosphoproteomic analysis was also performed on detergent-resistant and-sensitive fractions distinguished by extraction with non-ionic detergent (S10 Fig) [33,52].
Endosomes from three neuroblastoma cell lines were characterized by phosphoproteomic analysis. In all endosome fractions from three cell lines (LAN-6, SMS-KCN, SK-N-BE(2)), 908 proteins were detected, including 22 RTKs, 10 tyrosine phosphatases; 30 SH2- and 44 SH3-domain-containing proteins. The most highly phosphorylated RTKs in neuroblastoma were those identified in Fig 5A by large yellow nodes that indicates large amounts detected in endosome fractions (e.g., DDR2, ALK, KIT, RET, EGFR, PDGFA, FGFR1). FYN and LYN containing both activating and inhibiting phosphorylations were also prominent in endosomes, along with PAG1, inhibited SRC (SRC_i), the SH3 adaptor protein BCAR1, several other adaptor proteins, two tyrosine phosphatases (PTPN11/SHP-2 and PTPRN), and PLCG1/PLCγ1, which was found previously in endosomes in PC12 cells [53]. Notably, 26 out of the 55 SH3-domain-containing proteins in the human genome that were predicted to have a function in endocytosis based on orthologous interactions in C. elegans were found in neuroblastoma endosome fractions, and 2 of the 55 were detected in lysosome fractions [24].
We asked whether particular phosphorylated proteins were enriched in endosomes and DRMs by calculating the ratio between amounts in those fractions compared to proteins in all other samples from the same cell line. ALK, FGFR1, RET, PDGFRA, DDR2, EGFR, and IGF1R were enriched in endosomes from two or more neuroblastoma cell lines, but there were profound differences among cell lines (Fig 5B). In Fig 6, enrichment was graphed in PPI networks as big yellow nodes for positive enrichment and small blue nodes for de-enrichment (defined as lower amounts in that fraction compared to elsewhere). In LAN-6 cells, most RTKs were enriched in endosomes, except EPHA2 and ROR1, which were enriched in DRMs (Fig 6A and 6B). In SK-N-BE(2) cells made to over-express NTRK1/TrkA, this receptor was enriched in endosomes and de-enriched in DRMs, whereas its related receptor, NTRK2/TrkB, had the opposite pattern, being enriched in DRMs and de-enriched in endosomes (Fig 6C and 6D). The SFKs, FYN and LYN were localized differently, with LYN (and LYN_i) being enriched in DRMs in LAN-6 and SK-N-BE(2) cells, and FYN (and FYN_i) being enriched in endosomes in LAN-6 cells, but not in SK-N-BE(2) cells (Fig 6). PAG1 was enriched in endosomes in LAN-6 cells (Fig 6A) and, in contrast, in DRMs in SK-N-BE(2) cells (Fig 6D).
We noted differences in distribution of SFK and PAG1 phosphorylation on individual phosphorylation sites between the two cell lines (Fig 5C). For example, PAG1 81 was consistently phosphorylated in endosomes, and PAG1 317 was consistently phosphorylated in DRMs in both cell lines, yet PAG 359 and other sites were highly phosphorylated in LAN-6, but not SK-N-BE(2) endosomes (Fig 5C). These data suggest a relationship between SFK and PAG1 phosphorylation on specific sites and intracellular localization.
These data suggest the hypothesis that stimulation of different RTKs should affect the activity and intracellular localization of FYN and LYN. We used a cell fractionation approach to assay intracellular localization after stimulation of ALK with PTN and KIT with SCF (Fig 7). Amounts of FYN and LYN increased with PTN and SCF treatment in organelles whose migration on velocity sedimentation gradients overlaps with Rab7 and acid phosphatase [32], markers for late endosomes and lysosomes (Fig 7A–7D, fractions 4–7). SCF also induced increases mainly in FYN localization to fractions 8–11 (Fig 7B and 7D), which contain endosomes marked by Rab4 and Rab5 [32]. LYN and FYN also increased in fractions 16–22 in response to both ligands (Fig 7A–7D). These fractions contain soluble, cytoplasmic proteins, and signaling particles, which were previously resolved on gradients centrifuged with greater force [52]. FYN and LYN were robustly associated with membranes that floated to the density of endosomes on floatation equilibrium gradients, and amounts increased in organelles of higher sedimentation velocity (E1) after PTN treatment (Fig 7E). Both FYN and LYN were predominantly phosphorylated on their activating sites in these membranes (Fig 7F). Differences between FYN and LYN localization to detergent-resistant and-soluble fractions were also observed. FYN’s response to PTN (enhanced DRM and diminished P1M association; Fig 7G) was different from that to SCF (reduced DRM, enhanced P1M association). In contrast, LYN’s response to both ligands was similar (reduced DRM, increased P1M association; Fig 7G). The magnitude of ligand-induced changes in FYN and LYN in organelle fractions were distinct in response to PTN and SCF (Fig 7H). Increased FYN and LYN in faster sedimenting organelles (lys and E1 fractions) likely reflects migration to multivesicular bodies, late endosomes, and possibly lysosomes [32]. These data are consistent with the hypothesis that RTK activation regulates FYN and LYN localization and activity in neuroblastoma cells in a manner that distinguishes responses to individual RTKs.
These data motivated further higher resolution interrogation of the relationships between individual protein phosphorylation events. We investigated the relationships among phosphorylation sites by clustering phosphorylation sites (summed from homologous phosphopeptides) and visualizing data structure as a co-cluster correlation network (CCCN). The edge-weighted, spring-embedded layout of this network showed several distinct groups of sites with statistical relationships to other groups (S11 Fig). The data were interrogated with a focus on the most highly phosphorylated sites on RTKs, SFKs, and PAG1 to ask if phosphorylation sites cluster together. Two distinct clusters are shown in Fig 8. ALK was detected in 22 distinct phosphopeptides in neuroblastoma samples, which could be collapsed into 13 distinct phosphorylation sites based on sequence homology. Fig 8A shows that the ALK phosphorylation site, ALK 1507, which was most frequently seen in neuroblastoma samples, was associated with inhibited LYN (LYN 508), and activated FYN (FYN 420; LCK 394; SRC 419; YES1 426; this site was assigned to FYN in total phosphorylation calculations because other FYN phosphopeptides were detected in the same samples; see Materials and Methods). Co-clustered phosphorylation sites on several other proteins in this cluster resemble the cluster in Fig 1A. Fig 8B shows that other ALK phosphorylation sites (ALK 1096 and 1604) clustered with the most prominently detected phosphorylation site on DDR2 (DDR2 481), along with activated LYN (LYN 411), and inhibited FYN and SRC (FYN 531; YES1 537 and SRC 530). Also co-clustered with the group in Fig 8B were phosphorylation sites from other RTKs represented in the cluster in Fig 1B.
Strikingly, a number of phosphorylation sites on PAG1 were detected, but none were statistically clustered with activated FYN and inhibited LYN (Fig 8A), while the most prominent PAG1 phosphorylation sites clustered with activated LYN and inhibited FYN and SRC (Fig 8B). The data suggest a mutually antagonistic relationship between different SFKs, particularly LYN and FYN, so that when one is activated, the other is inhibited. Phosphorylation of PAG1, which recruits SFKs and their inhibitory kinase, CSK, to bind to it [51], appears to be associated with the state where LYN is activated, and FYN and other SFKs are inhibited (Fig 8B).
The data suggest that RTK phosphorylation does not occur on all sites at once under all conditions, resulting in different phosphorylation sites on ALK and other RTKs clustering separately from one another. RTKs phosphorylated on different sites also fractionated to endosomes and DRMs selectively (S12 Fig). For example, some ALK and KIT phosphorylation sites were enriched in endosomes, while others were enriched in DRMs, with differences between the two cell lines examined (S12 Fig). In contrast, all EGFR and RET phosphopeptides were consistently enriched in endosomes. Phosphorylation on selected sites would be consistent with RTKs acting as effectors as well as initiators of signal transduction. Phosphorylation by other tyrosine kinases, such as other RTKs or SFKs, may favor particular sites, and thus influence intracellular location, providing different contexts for signaling pathways to influence cell responses.
There is considerable interest in tyrosine kinase signaling mechanisms because of their roles in tumor initiation and metastasis. Tyrosine kinase signaling mechanisms arose during evolution when multicellular organisms evolved [19,54], and many RTKs are known to be involved in governing cell behaviors such as cell division, cell death, differentiation, and migration. Acquisition of phosphoproteomic data from a migratory, multipotent tumor cell type was motivated by these considerations. The complexity of the data forced us to develop new approaches to understand signaling mechanisms that involve tyrosine phosphorylation. Indeed, modeling dynamic complex systems and their interacting macromolecules remains a general challenge that lags far behind large-scale acquisition of biological data [35,55]. To make progress, we found it useful to apply techniques from the fields of pattern recognition and graph (network) theory and combine them with external PPI and GO data [34], thus extending the concept of using a variety of statistical techniques for exploratory data analysis [56]. Exploratory data analysis is inherently descriptive in its initial stages, but allows generation of hypotheses which then motivates more directed data interrogation and subsequent experiments. In this way, initial phosphoproteomic analysis of neuroblastoma cell lines motivated further experiments where cells were treated to perturb signaling pathways and subjected to organelle fractionation.
Several technical hurdles had to be overcome when analyzing these data. 1) Phosphoproteomic data, like any mass spectrometry data, has missing values because many peptides are not analyzed by the detector, and using a “data not available” marker (NA) instead of zero facilitated calculation of statistical relationships based only on observed data [34]. 2) Employing an effective embedding technique [48,49] prior to clustering allowed resolution of patterns that were difficult to discern otherwise [34]. 3) The analysis treated all samples mathematically as different states in the total neuroblastoma signaling system. Embeddings were first performed on data from cell lysates from 21 neuroblastoma cell lines grown in culture without treatments. Preliminary analysis led to emerging trends in clustering that resembled the more robustly defined clusters derived when all samples were included, including different cell lines, cells treated in ways to perturb signaling pathways, and cell fractions. In fact, we noted different phosphoproteomic results from the same cell lines cultured under nominally similar conditions when grown by different investigators or at different times. This heterogeneity could be due to differences in serum batches, selection pressure by passaging, or other factors. Mathematically, any heterogeneity is useful for statistical analyses because different phosphorylation patterns help distinguish signaling pathways. 4) Visualizing data as networks was informative in several ways. Initially, networks that included all PPI edges allowed us to determine that the dataset was complete enough for further analysis (S1 and S2 Figs). However the multitude of known interactions among signaling proteins was too complex to be informative, and not specific to neuroblastoma.
Importantly, to visualize data structure in a new and informative way, we developed an innovation that may be generally useful, namely to filter edges to show interactions only between co-clustered components (CFN, Fig 3; CCCN, S7 and S11 Figs). Including clusters from different (equally valid) embeddings recognizes that tyrosine kinase signaling pathways are highly interconnected by conveniently allowing overlap in cluster membership. Applying this approach to individual phosphorylation sites (S11 Fig) elucidated phosphorylation patterns and relationships among signaling pathways with high resolution. These graphs allowed exploration of data structure using network analysis in a visually accessible graph. We focussed here on PNCPs (proteins with tyrosine kinase, tyrosine phosphatase, SH2 and SH3 domains). Cytoscape-accessible files of these graphs are provided online for investigators interested in exploring the data further S2 Dataset.
The observation that neuroblastoma cell lines expressed more than half of the RTKs in the human genome (S3 Fig), and responded to signals from growth factors in the embryonic microenvironment to migrate and differentiate into a number of neural crest target sites (S4 Fig), suggests that neuroblastoma, and neural crest from which it is derived, takes full advantage of RTK signaling mechanisms to govern cell fate decisions. We found functional compartmentalization of tyrosine kinase signaling pathways in neuroblastoma cells from different tumor origins, with different sets of RTKs forming collaborative groups that interact with each other and common downstream effectors (Fig 3). There was also physical compartmentalization of signaling components within neuroblastoma cells. By combining cell fractionation with phosphoproteomics, we found that there was non-uniform distribution of signaling components, and moreover non-uniform distribution of phosphorylated residues on individual proteins (Figs 5, 6 and S12 Fig).
Compartmentalization of signal-initiating receptors and downstream effectors may be employed to distinguish extracellular instructions that determine cell fate. Many receptors signal from endosomes to amplify signals, activate different effectors than those activated at the plasma membrane, or convey signals to different intracellular locations [30,57–61]. In fact, there is evidence that endosomal signaling from a number of different receptors affects cell fate decisions during development [62–66].
Different RTKs elicit different cellular responses, yet all appear to activate the canonical RAS/ERK, PLC-γ, SFK, and PI3K/AKT pathways. Differential responses may be obtained by affecting the duration of downstream effector activation [67], or by modulating the relative strength of downstream pathway signaling, as has been elegantly shown for the ratio of activation of AKT and ERK pathways that distinguishes the proliferation and neurite-outgrowth (differentiation) response in PC12 cells [68]. Our data suggest that SFKs, especially FYN and LYN, function as signal integrating devices—central hubs in the tyrosine kinase signaling network—to distinguish RTK signal transduction pathways, in part by activating distinct mechanisms specifically in endosomes and lipid rafts. FYN and LYN were highly phosphorylated in endosomes and detergent resistant membranes, and their activity and localization was affected by cell type (Figs 5 and 6) and changed in different ways in response to receptor activation (Fig 7). FYN and LYN appear to have a partially antagonistic relationship because when one is activated, the other is frequently phosphorylated on its C-terminal inhibitory site (Figs 1 and 8). How localized activation of FYN and LYN may in turn affect the relative strength and duration of effector pathways, or the ratio of activation of AKT and ERK, remains to be determined.
Previous work supports the hypothesis that SFKs function to affect signal integration and protein localization. SFK family members are differentially palmitoylated, which affects their localization on endosomes and the plasma membrane [69,70]. SFKs have been implicated in the regulation of endocytosis by a variety of mechanisms. These include phosphorylation of clathrin [71], modification of Rho proteins and actin assembly [69,72], and regulation of the Cbl family of ubiquitin ligases [51,73,74], which control RTK sorting in endosomes [75,76]. The localization of SFKs to lipid rafts is thought to be important for their signaling function [77]. For example, it has been shown that FYN plays a role in localizing NTRK2/TrkB to lipid rafts [78], and LYN, which is enriched in lipid rafts (Fig 6) is a key effector of NTRK1/TrkA for terminal differentiation [79].
The transmembrane SFK scaffold protein PAG1 (Cbp/PAG) has been previously described to associate with lipid rafts [80]. Consistent with this, we found PAG1 in DRMs (Figs 5 and 6). PAG1 was also one of the most highly phosphorylated proteins in endosomes (Figs 5 and 6). PAG1 binds several different SFK family members, and can bind to more than one at a time, as well as to the kinase that phosphorylates the inhibitory site on them, CSK [51]. In fact, PAG1 can form a complex with a number of SFK regulatory proteins in addition to CSK: the phosphatase, PEP, PTPN22 and SOCS1, which catalyses SFK ubiqutination [51]. PAG1 also binds PLCG1/PLCγ1 and PI 3-kinase; and PLCG1 and PIK3R1/2 were detected in endosome fractions in this study. The phosphatase, PTPN11/SHP-2, which was also prominently detected in neuroblastoma endosomes (Fig 5A) may also be part of this complex [81]. Different patterns of PAG1 and PTPN11 phosphorylation in leukemia and prostate cancer are associated with different activation states of SFKs and other signaling effectors [82,83]. This array of proteins bound to the PAG1 scaffold may either positively or negatively regulate SFK activity as well as other effectors, depending on context. Interestingly, we found phosphorylated PAG1 to be clustered with activated LYN and inhibited FYN (and SRC), but not activated FYN and inhibited LYN (Fig 8).
The collaborative groups that emerged from these data (Fig 3) suggest the hypothesis that receptors within these groups might be likely to cause transactivation of other RTKs within the same group. IGFR1 1161 phosphorylation was decreased by the ALK inhibitor on a similar scale to ALK 1507 and 1509 (S9 Fig), which is consistent with the hypothesis that ALK and IGFR1 activities are linked (Fig 3A). The data show substantial variability on different RTK phosphorylation sites, however. When we performed clustering on individual phosphorylation sites (S1 Fig), different phosphorylation sites on ALK and other RTKs clustered separately from one another. For example, ALK 1507 was associated with the group of sites shown in Fig 8A, while ALK 1096 and 1604 was associated with the group in Fig 8B. These phosphorylation patterns may be due to selective phosphorylation or dephosphorylation. Phosphorylation on selected sites would be consistent with RTKs acting as effectors as well as initiators of signal transduction; other tyrosine kinases, such as other RTKs or SFKs, may favor phosphorylation on particular sites. One mechanism of RTK transactivation could involve heterodimerization of different RTKs or multiprotein receptor clusters. Heterodimers have been inferred from co-immunoprecipitation between MET, EGFR, and ERBB3/Her3 [84]; PDGFR and EGFR [85]; AXL and EGFR [86]; and among similar EGFR and FGFR family members [87].
SFKs may also play a role in RTK transactivation [87]. SFK SH2 domains bind to phosphorylated tyrosine residues on RTKs, and can phosphorylate RTKs directly, in some cases mimicking those sites phosphorylated during ligand-induced receptor activation [74]. SFKs associate with RTKs in protein complexes and play a direct role in transducing their signals [74,88]. It has been shown that transactivation between PDGFR and EGFR depends on SFKs [85], and SRC is recruited to PDGFRB and the GPCR, MBTPS1/S1P1, which form a complex that is endocytosed as a unit [89]. In addition, phosphatases may favor particular sites. For example, the phosphatase, PTPN6/SHP-1, acts on NTRK1/TrkA, mainly at Y674 and Y675 [90], and association of PTPN6/SHP-1 with lipid rafts suggests localized dephosphorylation of NTRK1/TrkA [33].
Point mutations in the RTK, ALK, are the primary cause of familial neuroblastoma and account for 8–12% of sporadic neuroblastomas [15]. ALK is expressed earlier than Trks (NTRK1-3) in neural crest development [91], highly expressed in paravertebral sympathetic ganglia [92], and co-expressed with NTRK1/TrkA and RET in a subtype of dorsal root ganglia neurons during development [93]. Overexpression of full-length ALK in PC12 cells causes increased phosphorylation of PTPN11/SHP-2 and STAT3 [94]. We found PTPN11 clustered with ALK and IGF1R (Figs 1B and 3A), and localized in endosomes (Fig 5A). STAT3 co-clustered with ALK, IGF1R and PDGFRA as part of the same collaborative group (Fig 3A). That phosphorylated ALK was present in many neuroblastoma cell lines is consistent with its role as an important marker, or driver, of neuroblastoma.
Like ALK, KIT is an emerging marker for aggressive neuroblastoma that leads to poor prognosis [95]. We found KIT in a subset of neuroblastoma cell lines, and enriched in endosomes in SK-N-BE(2) cells (Figs 6 and S6C). Activation of either ALK or KIT caused increased association of FYN and LYN with endosomes (Fig 7). Both ALK and KIT are expressed early in neural crest, giving rise to the hypothesis that cells derived from an earlier stage of the neural crest sympathoadrenal lineage are more likely to give rise to more aggressive tumors and poor clinical outcome. Sox10+/Kit+, but not Sox10+/Kit- cells, remain multipotent even after reaching their final target tissue [96,97]. Neuroblastoma cells that express high levels of KIT can induce tumors ninefold more efficiently than those with low KIT expression [95]. Interestingly, KIT clustered with ROR1 (S6C Fig), which is also expressed early in development and is a marker for cell migration and invasiveness in neuroblastoma and other cancers [98]. The data suggest that both KIT and ALK may be active early in neural crest development and their activity signifies, or causes, incomplete differentiation.
Neurotrophin receptors are of interest in neuroblastoma because they are markers for clinical prognosis. NTRK1/TrkA is a marker for neuroblastoma tumors that spontaneously undergo apoptosis and regression, while NTRK2/TrkB is often expressed with its ligand, (BDNF), forming an autocrine loop that predicts poor prognosis [99–101]. The pan-neurotrophin receptor, p75NTR enhances sensitivity to low neurotrophin levels, which affects response and outcome in NTRK1/2-expressing cells [102]. Overexpression of NTRK1/TrkA in LAN-6 cells caused apoptosis, but was tolerated in SK-N-BE(2) neuroblastoma cells that express non-functional p53, in agreement with previous work [103]. The differential localization of NTRK1/TrkA, which preferred endosomes, and NTRK2/TrkB, which was enriched in DRMs (Fig 6C and 6D) may provide a clue as to how these two similar receptors have such profoundly different effects in neuroblastoma. Neurotrophin receptors signaling from lipid rafts vs. endosomes may account for the selectivity of their transduced signals and the resulting effects on cell behavior [33,104,105].
Neuroblastoma cell lines offer insight into neural crest signaling pathways that is difficult to obtain directly from migrating immature neural crest cells. While signaling pathways activated by oncogenic mechanisms and cell culture conditions no doubt contribute to the phosphorylation patterns we identified here, the fact that these cells retained the capacity to migrate and differentiate (S4 Fig) indicates that neuroblastoma cell lines retain signaling pathways activated in immature, multipotent neural crest [2,4,7,8]. That neuroblastoma cells express so many RTKs suggests that mechanisms to discern and integrate different receptors’ signals must play a role in cell fate decisions in neural crest and neuroblastoma [106–108]. SFKs, which contain a tyrosine kinase domain, a SH2 domain that recognizes phosphorylated tyrosine, and a SH3 domain that plays a conserved role in endocytosis (and other) mechanisms, appear to be constructed for signal integration. The activation and dynamic intracellular location of LYN and FYN, and a scaffold protein (PAG1) that binds to them, suggest that these SFKs function to discern and integrate signals from different RTKs.
Discovery of new pathways activated in neuroblastoma cells provides potentially new therapeutic approaches [109]. Co-activation of two or more RTKs, which is not uncommon in cancer, leads to therapeutic challenges that compel consideration of treatment with multiple inhibitors [110–113]. The data and analysis presented here suggest, for example, that ALK-driven tumors might also present activated IGF1R, FGF1R, and/or PDGFRA. When challenged by ALK inhibitor therapy, these receptors could take over as drivers to activate similar signaling pathways (Fig 3A). The data also suggest that there are different routes to cell proliferation in neuroblastoma, such as the distinct mechanisms activated by the EGFR group (Fig 3B), or KIT (Figs S6C and 7). In the future, it will be important to compare our results to pathways activated in neuroblastoma primary tumors in different microenvironments. This study, and other large-scale gene expression or proteomic studies that include network and pathway analysis [82,83,114] and gene ontology [115], will help determine likely control points for cell growth, migration, and differentiation in individual tumors.
21 neuroblastoma cell lines were obtained by Cell Signaling Technology (Danvers, MA) from American Type Culture Collection (ATCC; Manassas, VA); Leibniz Institute DSMZ-German Collection of Microorganisms and Cell Cultures (DSMZ; Braunschweig Germany); Coriell Institute for Medical Research (Camden, NJ); and Interlab Cell Line Collection (ICLC; Genova, Italy). SMS-KCN, LAN6, SK-N-BE(2), and SH-SY5Y, were obtained by M.G. from Children’s Hospital of Los Angeles, CA, except for SH-SY5Y cells, which were provided by Dr. Mark Israel (University of California, San Francisco, CA). Neuroblastoma cells were grown in RPMI 1640 medium (Thermo Scientific HyClone, U.S.) supplemented with NaHCO3 (Sigma, U.S.) and 10% fetal bovine serum (Thermo Scientific HyClone, U.S.).
Tyrosine phosphoproteomic data for 21 neuroblastoma cell lines were initially acquired at Cell Signaling Technology using techniques described previously [41,42]. Cells were incubated overnight in media without serum prior to harvesting for mass spectrometry. Four cell lines [SH-SY5Y, LAN-6, SMS-KCN, and SK-N-BE(2)] were selected for further studies because of different point mutations in ALK, p53 status, RTK expression, morphology, and growth characteristics. A sub-line of adherent SMS-KCN cells, named SMS-KCN-A, was selected by culturing SMS-KCN cells on collagen coated plates and removing floating cell spheres. SMS-KCN-A cells required trypsin for passage and retained their adherent phenotype after passaging. SK-N-BE(2) cells were made to overexpress Rat TrkA with CFP insert at amino acid 587 (in the cytoplasmic tail) using a γ-retroviral expression vector (a gift from Mary Beth Eiden, NIH [116]). The construct was made using transposon-mediated insertion [117], and shown to be functional as assayed by NGF-induced tyrosine phosphorylation and neurite outgrowth in PC12nnr5 cells (in which endogenous TrkA is non-functional). The γ-retroviral genomic vector plasmid (pRT43.2TrkCFP), helper plasmids (pIK6.1.gagpol+ATG and pLP-VSVG) were transfected into HEK293T cells using calcium phosphate. Cell culture media (Dulbecco’s Modified Eagle Medium/10% FBS) was changed approximately 16 hours post-transfection. Supernatant containing viral particles was harvested at 48 and 72 hours post-transfection and pooled together.
Cell lines were treated (or left untreated, control) with ligands or the ALK inhibitor TAE684 as indicated in Table 1.
For organelle fractionation phosphoproteomics, cell lines were treated with ligands (LAN-6 and TrkA-CFP-expressing SK-N-BE(2): NGF, SMS-KCN:BDNF) for 10 min at 37°C. For cell fractionation experiments after ALK and KIT stimulation (Fig 7), LAN-6 cells were serum-starved for 2hr, then treated with 50 nM PTN or 5 nM SCF (R & D Systems). Ligands were bound to cells at 4°C for 1 hr, then cells were warmed to 37°C for 10 min or 1 hr. Organelles were isolated from mechanically permeabilized cells using velocity sedimentation only (Fig 7A–7D) or velocity sedimentation followed by flotation equilibrium centrifugation as described [32]. Phosphoproteomic analysis was performed on two endosome (E1, E2) and lysosome (Lys) fractions as shown in S10 Fig mass-density plots. In addition, E3 and cytosol (cyt) fractions collected from velocity gradients as indicated in Fig 7C were methanol/chloroform precipitated for gel electrophoresis and western blot analysis. Detergent-resistant (DRM) and-soluble (P1M) fractions were prepared as described [33] except that flotation of detergent-resistant membranes was not performed for mass spectrometry experiments or gel electrophoresis and western blotting.
Antibodies used in Fig 7 were from Cell Signaling Technologies (Danvers, MA): anti-LYN (#2796);-FYN(#4023);-pSRC (Y416; #2101);-non-pSRC (Y416; #2102). HRP-linked secondary antibodies were from GE Healthcare UK Limted: anti-Rabbit HRP (# NA934V); anti-Mouse HRP (#NA931V). Ligands were from R & D Systems: PTN (#252-PL); SCF (#255-SC).
Quantification of immunoprecipitated phosphopeptides was obtained from the peak intensity of each peptide (from the MS1 spectrum of the intact peptide before fragmentation for MS/MS analysis) [41,42]. Data were processed using R [118,119]. Gene names were mapped and converted to unique gene identifier names (according to genenames.org). In cases where conserved peptide sequences identified multiple proteins, if a protein was identified by a different peptide in the sample, the peptide was assigned to that protein, otherwise the first name was used (this is referred to as exclusively summed). Where phosphorylation sites were known to have inhibitory effect on protein activity (Regulatory_sites.gz), peak intensity values were converted to negative values (this allows graphing network nodes as blue, as in Fig 1). Peak intensity was summed for each protein in each sample (i.e., cell line) using functions written in R [34], except in the case of SRC-family kinases (SFKs), where peptides phosphorylated on C-terminal inhibitory sites were tracked separately (denoted FYN_i, LYN_i, SRC_i, YES1_i, FRK_i). Due to limits in mass spectrometry detection, data were not expected to be complete; for example SMS-KCN cells express NTRK2 (TrkB), but NTRK2 peptides were masked; and NTRK1 was not always detected in cell lines known to express it. Therefore, missing values were treated as NA (data not available) for statistical calculations [34].
In cases where duplicate mass spectrometry analyses were conducted on the same cell line, under the same conditions within a short time frame (e.g., duplicate runs of the same experiment), data were merged to include the average of the two runs, ignoring missing values. Otherwise, each experiment was treated as an independent sample for data analysis.
Summarized data are available online as Supplemental Data (S3 Dataset). Primary phosphoproteomics data are available from PhosphoSitePlus (http://www.phosphosite.org/browseDiseaseResultAction.do?id=66&type=true) using curation set (CS) numbers 1119, 1121, 1148, 1154, 1157, 1206, 1448, 1613, 1762, 1763, 1764, 1886, 1887, 1888, 2042, 2043, 2044, 3492, 3493, 3495, 4010, 4011, 5180, 5181, 5182, 5269, 5270, 5271, 5272, 6121, 6122, 6123, 6124, 6125, 6151, 6152, 6153, 6154, 6155, 9206, 9208, 9942, 9943, 9944, 9946, 10553, 10554, 10555, and 10557. Supplementary data, and links to spectra and other important experimental parameters from this submission will be also made available via the ‘Reprints, References, Supplemental Tables’ page http://www.phosphosite.org/staticSupp.do.
Duplicate MS runs on the same samples in the same experiment were LAN6.Control (CS 5269, 5270); LAN6.NGF (CS 5271, 5272); SMSKCN.Control (CS 3492, 3493); SMSKCN.NGF (CS 3494, 3495); SMSKCN.Control (CS 3549, 3550); SMSKCN.BDNF (CS 3551, 3552).
1203 tyrosine phosphorylated and 557 AKT-substrate (using RxRxxS/T consensus sequence antibodies) proteins were identified in all samples; 138 were in common between phosphotyrosine and phospho-AKT substrate data.
For analysis of proteins, the total amount of phosphorylation of each protein was determined by summing peak intensity signals from all peptides for each protein in each sample. In cases where conserved sequences did not allow unambiguous assignment to a particular protein, peptides were assigned to proteins that were detected by other phosphopeptides in the same sample or the first name was used. We thus obtained a data matrix in which each row corresponds to a protein and each column corresponds to a neuroblastoma cell line or organelle fraction (i.e. a sample; Fig 1). The elements of the data matrix contain the total peak intensity signals. All samples were treated as different states in the neuroblastoma system. To ensure that all samples were weighted equally in statistical calculations, data were normalized by scaling by sample standard deviations without centering. The statistical similarity of any two proteins was determined by the extent to which they were detected in similar amounts in each sample. This relationship was represented in different ways. First, the Euclidean distance between the row vectors corresponding to the two proteins was stored in a distance matrix. A dissimilarity matrix (also called dissimilarity representation or feature vector) is similar to a distance matrix except the values do not necessarily specify Euclidean distance [120,121]. For the second method, dissimilarity was represented by one minus the absolute value of the Spearman correlation of each protein with every other protein. A third method employed combining equally scaled Euclidean distance and Spearman dissimilarity as a dissimilarity matrix, referred to as Spearman-Euclidean dissimilarity, or SED [34,122].
The dimension-reduction (embedding) technique, t-distributed stochastic neighbor embedding (t-SNE) was employed to visualize the proteins in a scatter plot based on the distance or dissimilarity matrices [34,48,123]. This machine learning technique aims to represent each protein by a two-(or three-) dimensional point, arranging the points in such a way that nearby points in the scatter plot correspond to proteins with statistical similarity and distant points to dissimilar proteins. Proteins close to one another in this data structure were identified as clusters by the minimum spanning tree, single linkage method [50]. Three dimensional embeddings of data structure were visualized with PyMOL and Cytoscape, the latter using RCytoscape and three dimensional manipulation functions from the R package, rgl (S1 Movie). Filters were applied to focus on proteins containing tyrosine kinase, tyrosine phosphatase, SH2, and SH3 domains (PNCPs), or to focus on proteins that clustered together using both Spearman and Euclidean dissimilarity embeddings [34].
Clusters were evaluated by several quantitative measures. For comparison, 70 non-overlapping random clusters were generated containing gene names from the data set; the number of members was also randomized to mimic the number of genes identified in clusters defined by t-SNE embedding and minimum spanning tree methods. Evaluations based on examining the primary data (internal evaluations) were performed as described [34]. A quantitative index was used to evaluate the density of data (percent NA or missing values) and the conformity to the pattern in the group, weighted by the total phosphopeptide signal (S5A and S5B Fig). External evaluations with data from PPI (S5C and S5D Fig) and GO (S5E and S5F Fig) databases were also compared to 20 random clusters [34].
PPI edges from String (string.embl.de/) [124], GeneMANIA (genemania.org/) [125], and the kinase-substrate interactions from PhosphoSitePlus (phosphosite.org) [126] were merged as described [34]. Network modules or highly interconnected regions of the neuroblastoma phosphoproteomic network (S2B Fig) were determined using Cytoscape plugins MCODE and NeMo.
Gene Ontology (GO) was determined as described [34]. Enriched gene function annotations, or GO terms for gene groups determined by clustering methods, and for the randomly selected genes as described above, were retrieved using Bioconductor libraries “GO.db,” “GOstats,” and “org.Hs.eg.db” ([127] bioconductor.org/) using a p-value <0.01. If there was enrichment, at least two genes in the cluster should have the same GO term, so terms with single genes were discarded. The enriched GO terms per gene was compared to the average background for randomly selected genes from the dataset; this background was about one enriched GO term for every three genes [34]. When the number of enriched GO terms is more than five fold over background, this is strong evidence for enrichment [34].
For phosphorylation site analyses, peptide peak intensity values were summed based on sequence homology and phosphorylation site, independent of the presence or absence of oxidized methionine. In cases where conserved sequences did not allow unambiguous assignment to a particular protein, the peptide name either retained multiple names, for example “FYN 420; LCK 394; SRC 419; YES1 426,” or were merged into all possible larger peptides, for example MAPKs and C-terminal inhibitory phosphorylations on SRC, FYN, and YES1 (referred to as inclusively summed).
Four neuroblastoma cell lines (LAN6, SK-N-BE(2), SMS-KCN, and SY5Y) were cultured in 25 μL hanging drops containing approximately 5,000 neuroblastoma cells. Cells were transplanted into the neural crest of developing chick embryos to determine if these cells could survive transplantation and subsequently integrate into the migration pathways of the chick neural crest cells. Of the 14 embryos that were injected, 10 survived the transplantation process: three of these were injected with LAN6 cells, two with SK-N-BE(2) cells, three with SMS-KCN cells, and two with SY5Y cells. Fluorescent imaging of embryo sections showed that all four cell lines were successfully transplanted and could be located within various areas of the embryo with the use of GFP infection with adeno-associated virus that expresses GFP (AAV-GFP, a gift from Dr. D. Poulsen, University of Montana), anti-GFP (Rockland, Gilbertsville, PA), anti-ERGIC (Alexis Biochemicals, U.S.), and fluorescent secondary antibodies (Alexa Fluor 514 goat anti-mouse and Alexa Fluor 488 chicken anti-rabbit from Invitrogen Molecular Probes, U.S.). Further details of embro transplantation methods are provided in S1 Text.
|
10.1371/journal.pntd.0000894 | Viral and Epidemiological Determinants of the Invasion Dynamics of Novel Dengue Genotypes | Dengue has become a major concern for international public health. Frequent epidemic outbreaks are believed to be driven by a complex interplay of immunological interactions between its four co-circulating serotypes and large fluctuations in mosquito densities. Viral lineage replacement events, caused for example by different levels of cross-protection or differences in viral fitness, have also been linked to a temporary change in dengue epidemiology. A major replacement event was recently described for South-East Asia where the Asian-1 genotype of dengue serotype 2 replaced the resident Asian/American type. Although this was proposed to be due to increased viral fitness in terms of enhanced human-to-mosquito transmission, no major change in dengue epidemiology could be observed.
Here we investigate the invasion dynamics of a novel, advantageous dengue genotype within a model system and determine the factors influencing the success and rate of fixation as well as their epidemiological consequences. We find that while viral fitness overall correlates with invasion success and competitive exclusion of the resident genotype, the epidemiological landscape plays a more significant role for successful emergence. Novel genotypes can thus face high risks of stochastic extinction despite their fitness advantage if they get introduced during episodes of high dengue prevalence, especially with respect to that particular serotype.
The rarity of markers for positive selection has often been explained by strong purifying selection whereby the constraints imposed by dengue's two-host cycle are expected to result in a high rate of deleterious mutations. Our results demonstrate that even highly beneficial mutants are under severe threat of extinction, which would suggest that apart from purifying selection, stochastic effects and genetic drift beyond seasonal bottlenecks are equally important in shaping dengue's viral ecology and evolution.
| Dengue fever and the more severe dengue haemorrhagic fever and dengue shock syndrome are mosquito borne viral infections that have seen a major increase in terms of global distribution and total case numbers over the last few decades. There are currently four antigenically distinct and potentially co-circulating dengue serotypes and each serotype shows substantial genetic diversity, organised into phylogenetically distinct genotypes or lineages. While there is some evidence for positive selection, the evolutionary dynamics of dengue virus (DENV) is supposed to be mostly dominated by purifying selection due to the constraints imposed by its two-host life-cycle. Motivated by a recent genotype replacement event whereby the resident American/Asian lineage of dengue virus serotype 2 (DENV2) had been displaced by the fitter Asian-1 lineage we investigated some of the epidemiological factors that might determine the success and invasion dynamics of a novel, advantageous dengue genotype. Our results show that although small differences in viral fitness can explain the rapid expansion and fixation of novel genotypes, their fate is ultimately determined by the epidemiological landscape in which they arise.
| Dengue virus (DENV) is the most wide-spread arbovirus affecting human populations. During the last decades it has increasingly become a major public health problem with significant economic and social impact [1]–[3]. It is transmitted between humans in urban and peri-urban settings predominantly by the Aedes aegypti and Aedes albopictus mosquitoes vector [4]. Ae. aegypti is extremely well adapted to urban environments where it efficiently breeds in artificial water containers, such as flower pots, plastic bags or discarded car tires, near human habitations. Both vectors have undergone rapid expansion worldwide in the last couple of decades leading to DENV endemicity in more than 100 countries [5].
There are four closely related and potentially co-circulating serotypes of DENV (DENV1-DENV4) [6], [7] and recovery from infection is believed to provide life-long immunity to the infecting serotype but only a brief period of heterologous protection to all other serotypes [8]. Most primary infections are self-limited and clinically silent but can occasionally result in a short-lived febrile illness which is commonly known as dengue fever (DF). In some cases this may progress to more severe and life-threatening illness such as dengue haemorrhagic fever (DHF) or dengue shock syndrome (DSS) [9]. While several risk factors for developing DHF/DSS have been described, including host genetic background, viral genotype, order of infecting serotype, time between infections or age of infection [1], [9], the most widely cited explanation is that of Antibody Dependent Enhancement (ADE) (e.g. [10]–[13]) whereby subneutralizing antibodies from primary infection can mediate viral entry into host cells leading to increased replication and disease manifestations [14]–[18].
The temporal epidemiological pattern of dengue is characterized by semi-periodic outbreaks whilst the inter-epidemic cycles in DF/DHF incidence highly correlate with the seasonal variations in vector population size (see e.g. [19]). Furthermore, individual serotype prevalences show cyclical replacements in dominance (Figure 1A) which are believed to be induced by the immune profile of the human population [20], [21].
Phylogenetic studies based on complete sequences of structural genes of all 4 serotypes have demonstrated the existence of multiple lineages in which different genotypes can be clustered [6], [7]. Despite a general bias in the literature towards studies based on single-gene approaches, spatio-temporal patterns of genotype replacement in endemic regions have been widely recovered from data [6], [7], [22]–[24]. With the extrinsic pressures on DENV, such as seasonal or human-forced reductions in vector population size or abundance and mobility of susceptible hosts, it has been proposed that genetic drift plays a major role in the observed phylodynamics [22], [25]. Furthermore, most studies have reported that DENV recent molecular evolution is marked by strong purifying selection, possibly due to the requirement of its two-host life cycle, and few reports have been able to show convincing evidence for positive selection either by the existence of non-synonymous mutations or in measures of fitness advantage in viral traits [6], [7], [23], [24], [26].
Following earlier reports of inter-serotypic difference in virulence (see e.g. [27]) one of the first convincing evidences for genetic determinants in disease outcome came from epidemiological studies suggesting that the DENV2 Asian genotype was associated with higher frequencies in DHF compared to the American genotype [28]. In vitro studies have since shown that the replication rate in both human monocyte-derived macrophages and dendritic cells as well as the vector's susceptibility were higher for the Asian genotype [29], [30]. It was also found that the Asian genotype of DENV2 had a slightly higher replication rate within the mosquito and a shorter extrinsic incubation period [31]. These results provided a rational explanation for the replacement patterns observed in the Americas, where displacement of the American genotype by the Asian genotype has taken place in several countries in recent years [28], [29], [32]. A similar lineage replacement event has also occurred in SE Asia, with Asian-1 lineage viruses having displaced Asian/American viruses from Viet Nam (Figure 1B), Cambodia and Thailand. This displacement was proposed to be due to difference in in vivo fitness, with higher viraemia levels observed in Asian-1 infected patients that could lead to an enhanced probability of human-to-mosquito transmission [33].
The study by Hang et al. [33] demonstrated some other intriguing aspects about the invasion dynamics of Asian-1. A phylogenetic analysis suggested that the Asian genotype was introduced into the population years before it had been detected, and once it was detected it reached fixation within a relatively short period of time. The rate at which this genotype replaced the Asian/American type would suggest a significant fitness advantage not only over the resident genotype but possibly also over the other circulating serotypes; however, there was no discernible difference in the overall epidemiological dynamics in the period before or after fixation. Although these results suggested that a fitness advantage in a specific viral trait played a decisive role, the emergence of advantageous genotypes are as likely to be driven by the level of transmission and the underlying immune status of the human population.
Here we have constructed an epidemiological model of dengue to qualitatively address the impact of immunity and transmission on the invasion and replacement patterns of a novel advantageous dengue genotype. Our results suggest that the observed replacement events can be explained by competition between genotypes of relatively small fitness differences which, although sufficient for displacement, do not interfere with the overall serotype dynamics. Furthermore, we show that invasion success and total time required for fixation are strongly influenced by inter- and intra-serotype competition at the time of introduction.
The model is an extension of the 4-serotype mathematical framework analysed by Recker et al. [34] and includes a mosquito vector component, temporary cross-immunity after primary infection and seasonal forcing in mosquito biting. In summary, we disregard the effect of maternal antibodies and instead assume that human individuals are born susceptible to all 4 serotypes. After recovery from primary infection they acquire life-long immunity to the infecting serotype and cross-immunity to any other serotype for a short period of time. As temporary immunity wanes, individuals become susceptible to secondary heterologous infection. For simplicity and because of the relative rarity of reported third and fourth infections we assume that after recovery from secondary infections individuals remain fully protected against further challenges [4], [35]. The system can then be given by the following set of differential equations describing the rate of change in humans either susceptible, infected with, temporarily immune or recovered from dengue serotypes , DENV1, DENV2, DENV2′, DENV3 or DENV4:(1)(2)(3)(4)(5)(6)with the force of infection of serotype affecting the human population, , given as(7)We denote as the mosquito biting rate and as the vector-to-human transmission probability; and are the respective durations of infection and cross-immunity. Given the short period of infection we do not account for the possibility of co-infections by two or more serotypes. We assume a constant human population size and further assume that infection has a negligible effect on the average death rate, . To account for seasonal variation we assume a periodically forced biting rate, that is we set(8)where is a positive integer influencing the ‘seasonality’ where results in shorter and more pronounced seasons.
The dynamics of the mosquito population is given as follows:(9)(10)with the force of infection from humans to mosquitoes given as(11)
In accordance with our previous model [34] we assume that antibody-dependent enhancement acts to increase both susceptibility to and transmissibility of secondary heterologous infection by factors and , respectively, with values described in Table 1.
To investigate the invasion patterns of a novel and fitter dengue genotype we assume that DENV2 is represented by two genotypes which differ in relative fitness but are antigenically equivalent. That is, individuals previously infected by DENV2 are immune to type DENV2′ and vice versa. We consider four different fitness traits which we can vary independently: (i) transmissibility from human to mosquito, e.g. through increased viral load, , (ii) longer life-expectancy of mosquitoes infected with DENV2′ to emulate a shorter extrinsic incubation period (EIP), , (iii) longer infectious period in humans, , and (iv) an increased level of enhancement of secondary infections, . These can simply be given using:(12)(13)(14)(15)That is, , can be considered as the degree of the fitness advantage. In line with the suggestion by Hang et al. [33], most of our analysis is concentrated on the fitness advantage due to increased viral load and thus transmissibility from the infected human individual to the mosquito vector, . In fact, we found that the results presented here are invariant to the actual viral trait that is enhanced; results obtained under changes to other viral traits can be found in the supporting material.
To address certain aspects of the invasion process of a more probabilistic nature, such as invasion success rates and fixation events, we also implemented the above model as a stochastic framework using a tau-leap Gillespie algorithm [36]. Stochastic simulations were initialized with equilibrium population status derived from the deterministic framework with parameter values the same as given in Table 1 (see Figure S7 and S8 for general model behaviour).
We used a simple epidemiological model of dengue to investigate the effect of host population immunity structures and transmission settings on the invasion pattern of a novel DENV2 genotype, hereby denoted as DENV2′. The model is based on a previously introduced deterministic, multi-serotype framework (e.g. [34], [37], [38]) but extended to include the mosquito vector population, with seasonal fluctuations in biting frequencies, and a period of temporary cross-immunity; full model details are given in the Methods section. We verified our model predictions within a stochastic framework which allowed us to more adequately address and further explore certain aspects of the invasion and replacement dynamics and their determinants [39].
The general dynamics generated by our model under parameter values given in Table 1 and prior to the introduction of a novel DENV2 genotype are characterised by semi-regular epidemic outbreaks and asynchronous cyclical behaviour in serotype prevalence (Figure 2). In accordance with previous studies (e.g. [34], [37], [40]) a wide range of incidence and serotype dynamics with different inter-epidemic periods can also be found under changes to key parameters values, especially those relating to the degree of enhancement of secondary infection or the period of temporary cross-immunity (Figures S1 and S2). For the remainder of this work, however, we kept most parameter values constant to allow for better comparisons between invasion patterns and their epidemiological determinants.
We examined the dynamics of a novel genotype introduced into a dengue endemic population by either an infected human individual or via an infected mosquito. The novel genotype is here denoted as DENV2′, to represent the Asian-1 genotype of serotype 2, whereas the resident type is denoted as DENV2 to represent the Asian/American type. Figure 3 shows the result of an invasion scenario where the invading genotype has a small fitness advantage over the resident type (, corresponding to a fitness advantage of ). In this case, higher viral fitness was realised through enhanced transmissibility from infected human individuals to the mosquito vectors, i.e. . In agreement with the data, two important features of the invasion dynamics can be observed and are highlighted in Figure 3B. Despite the eventual fast rate at which the advantageous genotype replaces the resident type, there is a significant lag between the point of introduction and the time when DENV2′ genotype would reach a detectable level of prevalence within the population; we refer to this level of prevalence as detection threshold. Furthermore, despite the expected temporary rise in dengue incidence, compared to the situation without invasion, the overall dynamics in both disease incidence and serotype prevalence remain largely invariant (Figure 3A). This suggests that both the time lag between introduction and first detection and also the rapid exclusion of the resident genotype, such as reported by Hang et al. [33], can be explained by a relatively small fitness advantage of the invading genotype.
The same qualitative behaviour can be also found when changing other viral traits which could determine the fitness advantage. That is, shortening the extrinsic incubation period, , increasing the duration of infection, , or the level of enhancement of secondary infection, , have the same effect as increasing the transmission probability from infected humans to mosquitoes, . Notably, though, when considering low advantages, smaller differences in terms of viral fitness are required to achieve the same rate of fixation if the fitness advantage manifests itself in longer infectious periods compared to an increase in transmissibility (Figure S3). Interestingly, while similar levels of fitness advantages in either EIP or transmissibility result in the same fixation times (Figure S4), the disturbance on the epidemiological pattern of dengue is less severe when the fitness advantage is expressed in the mosquito (Figure S5). From now on, we concentrate only on a fitness advantage through the proposed increase in human-to-vector transmission.
As shown in Figure 3, a small increase in transmissibility from human to mosquito seems sufficient for a novel genotype to displace a resident type within a short period of time. The actual rate of competitive exclusion and overall time from introduction of the advantageous genotype to its fixation in the population is likely to depend on various factors including fitness advantage, rate of transmission and immune profile within the human population. As shown in Figure 4A, increasing viral fitness accelerates the rate at which the invading genotype drives the resident type, DENV2, to extinction, resulting in a shorter period between introduction and fixation. For example, increasing the fitness advantage from to reduces the time to fixation from years down to years. However, this increase in viral fitness has a major effect on dengue incidence patterns and the dynamics of the other serotypes. In this case it leads to a significantly bigger epidemic outbreak at the time of replacement followed by a long period of low transmission and low prevalence of serotype 2 which could endanger its continuous persistence; this is highlighted in Figure 4B (compare to Figure 3A).
We next addressed the effect of the time of introduction on the invasion dynamics. This was simply motivated by the fact that serotype competition is not constant over time but is strongly affected by the level of transmission which itself is dependent on host immunity level and seasonal variation in mosquito densities. Not surprisingly, we found that the time of introduction can significantly alter the time taken for a novel genotype to reach fixation. Figure 5A shows the decrease in the frequency of DENV2, relative to the fitter genotype DENV2′, for two different time points of introduction. However, while the overall duration from invasion to fixation is dependent on the time when DENV2′ gets introduced, the actual rate of replacement remains constant. In other words, the time taken from DENV2′ passing a detection threshold, relative to DENV2, to reaching fixation is independent of the time of introduction (Figure 5B) and therefore independent of the overall epidemiological dynamics. This, on the other hand, suggests that the time lag between introduction and the point when it has spread sufficiently for detection, or waiting time, is strongly influenced by the epidemiological profile at that time.
To investigate further the determinants for fixation time we simulated a number of invasion events at various time points over a four year period and recorded the total time to fixation for each event with respect to (i) the number of naive individuals, (ii) serotype 2 susceptible individuals, (iii) disease prevalence and (iv) mosquito biting frequency. While we could not find a clear correlation between any of these population profiles and fixation time, we observed a trend for longer fixation times during the time window where the relative prevalence of serotype 2 was increasing (Figure S6).
The results from our deterministic model suggest that novel genotypes can face long periods at very low prevalence before breaching a detection threshold and going to fixation. Within a more realistic setting these periods signify an enhanced risk of stochastic extinction of the novel type despite its fitness advantage over the resident type. To better address the invasion success of DENV2′ we used a stochastic formulation of our model (see Methods) and simulated a number of invasion events over a period of four years and recorded the success rate of invasion, here defined as the successful introduction into a population followed by competitive exclusion of the resident type. As demonstrated in Figure 6A we observed that invasion success shows an oscillatory behaviour whose phase seems negatively correlated to total dengue prevalence at time of introduction. This suggests that the invasion of a newly advantageous genotype can be hampered by serotype competition during epidemics and favoured during off-season periods. Moreover, the amplitude of oscillation, i.e. the maximum success rate, is dependent on and again negatively correlated to serotype 2 prevalence. Figure 6B shows the increase in relative prevalence of DENV2 over the 4-year period which clearly correlates with a decline in the success rate of DENV2′.
Since the time taken from passing a detection threshold to reaching fixation was shown to be independent of the time of introduction (Figure 5B), we focused on the relationship between serotype 2 prevalence and the time to emergence, i.e. the period between introduction and reaching a prevalence threshold. Figure 7 clearly illustrates that a novel and advantageous genotype entering the population during periods of high DENV2 prevalence will face significantly longer emergence times than those introduced during periods of low prevalence. Together our results indicate that the fate of a novel genotype is strongly determined by both inter- and intra-serotype competition at the time of introduction.
We analysed the invasion pattern of a novel dengue genotype into an endemic population with 4 co-circulating serotypes. Within our framework we assumed that the invading genotype, representing the Asian-1 genotype of dengue virus serotype 2, possesses a fitness advantage over the resident type, the Asian/American genotype, through enhanced transmissibility from infected human individuals to the mosquito vectors. This assumption was based on the findings by Hang et al. [33] which showed increased plasma viraemia levels in patients infected by Asian-1 DENV2 viruses. In contrast to other studies [30], [41], Hang and colleagues did not find increased infectivity of Asian-1 viruses to Ae. aegypti mosquitoes per se; however, it is easy to envisage how higher viral titers could enhance the ‘per bite’ probability of human-to-vector transmission. By thus focusing on the hypothesis of a small increase in transmissibility during primary and secondary infections, and in agreement with the data, we observed that the total time for genotype replacement is composed of a period during which the invading type can circulate at very low prevalence levels for several years, followed by a rapid shift in dominance and competitive exclusion after the invading genotype had emerged; here we defined ‘emergence’ as a threshold level of prevalence where widespread detection would be highly likely.
Of particular interest is the time lag between introduction and emergence, or waiting time, when the detection of the new dengue genotype might be difficult by surveillance systems based on low viral sampling numbers and/or infrequent genotyping. Not surprisingly, we found that this period is strongly and positively affected by the difference in viral fitness between the resident and novel genotype. In the case of small fitness advantages several years could pass before the invading type has spread sufficiently to outcompete the resident type on a population-wide level. Furthermore, as the epidemiological pattern would remain largely invariant, passive surveillance systems based simply on case numbers could also easily fail to detect this intra-serotype replacement event. These results therefore support the findings of Hang et al. [33] who hypothesised that a small enhancement of human-to-mosquito transmission through increased viral load is sufficient to explain the observed invasion pattern in Southern Viet Nam where Asian-1 was first detected in 2003 despite the phylogenetic analyses dating the introductory event sometime during the late 1990's.
Apart from increased transmission from infected humans to the mosquito vectors we also considered other viral traits that could be enhanced in the Asian-1 genotype, such as longer infectious periods or shorter extrinsic incubation periods (EIP). The latter is of particular interest as it can potentially lead to a significantly increase in vectorial capacity [31]. While the actual viral trait which is enhanced does not alter the overall invasion pattern or results presented in this work (Figures S3, S4, S5, S9, S10, and S11), we found that viral fitness traits have an additive effect (Figure S4). This means that even smaller individual enhancements are sufficient to explain the observed invasion dynamics of the Asian-1 genotype, especially under the assumption that this replacement event did not have a major effect on the sero-epidemiological pattern of dengue. Interestingly, though, our results suggest that dengue incidence and serotype dynamics are less disturbed when the fitness advantage is manifested through shorter EIP than increased infectivity or transmissibility (Figure S5).
In addition to viral fitness, the time point at which a novel genotype enters a population is crucially important in determining its invasion dynamics and ultimately success. Whereas the relative fitness advantage affects the overall time between introduction and fixation, the epidemiological profile more strongly determines the period of low level prevalence before the advantageous genotype emerges. We tested various epidemiological factors for their influence on the waiting time but to our surprise only found the relative prevalence of DENV2 to have a strong effect. That is, whereas population susceptibility to either dengue in general or serotype 2 in particular had no immediate influence on the time between introduction and wide-spread detection, we found that the relative prevalence of DENV2 at the time of introduction positively correlates with extended periods during which the novel genotype circulates below a detection threshold. Therefore, while transmission intensities strongly affect the success of an invasion event, the dominance level of serotype 2 within the population determines both the invasion success rate and, independently, the period before the invading genotype would reach a sufficient level of prevalence to be widely detecable. Our results thus confirm that serotype interactions and the resulting epidemiological landscape can have a big influence on intra-serotype dynamics and thus viral evolution, as previously noted by Zhang and colleagues [23].
There is considerable interest in determining the evolutionary processes that underlie the observed structures and genetic variation of dengue virus populations (both inter- and intra-serotypic). Overall, low estimates of selection pressure, in terms of average values, and the fact that dengue has a two-host life-cycle are commonly used to place purifying selection as the strongest selective force acting on dengue evolution [23], [26], [42]. However, it is also clear that dengue viruses exhibit strong spatio-temporal variations. Various phylogenetic studies have identified frequent DENV lineage turnover events which have resulted in the characteristic, ladder-like tree (e.g. [24], [42]) and which are commonly ascribed to positive selection [24], [32], [43]. In addition, genetic drift has also been proposed to play a major part in dengue evolution such that the replacement of viral lineages or clades could be explained through stochastic processes alone. For example, repeated bottlenecks due to large seasonal fluctuations in mosquito densities imply that the emergence of novel and possibly advantageous genotypes could be a recurrent phenomenon followed by a strong probability for extinction in the subsequent circulating seasons which could explain the weak signature for positive selection in the data (compared to purifying selection). This in turn would also suggest that the success of a genotype does not always reflect its viral fitness [7]. In fact, we have shown that novel genotypes, especially those that arise during large epidemic outbreaks, can face high risks of extinction despite possessing a fitness advantage. Furthermore, even successful genotypes, i.e. those that eventually reach fixation, potentially undergo prolonged periods of low frequency which can span for several transmission seasons independently of the epidemics therein. Therefore, low measures of adaptive selection in this case would not necessarily imply strong purifying selection but could equally be explained by other epidemiological factors. This, however, needs to be confirmed within a more rigorous framework.
Dengue's two-host life-cycle implies a significant evolutionary constraint whereby the majority of newly arising variants are likely to be deleterious and selectively removed from the population. We have shown that even novel and advantageous DENV genotypes can undergo periods of several years prior reaching sufficiently large population sizes to escape the risk of extinction. Our results thus indicate that in addition to purifying selection, the epidemiological landscape and stochastic effects might be equally important determinants in shaping the viral evolutionary ecology.
|
10.1371/journal.pntd.0003803 | Antiviral Biologic Produced in DNA Vaccine/Goose Platform Protects Hamsters Against Hantavirus Pulmonary Syndrome When Administered Post-exposure | Andes virus (ANDV) and ANDV-like viruses are responsible for most hantavirus pulmonary syndrome (HPS) cases in South America. Recent studies in Chile indicate that passive transfer of convalescent human plasma shows promise as a possible treatment for HPS. Unfortunately, availability of convalescent plasma from survivors of this lethal disease is very limited. We are interested in exploring the concept of using DNA vaccine technology to produce antiviral biologics, including polyclonal neutralizing antibodies for use in humans. Geese produce IgY and an alternatively spliced form, IgYΔFc, that can be purified at high concentrations from egg yolks. IgY lacks the properties of mammalian Fc that make antibodies produced in horses, sheep, and rabbits reactogenic in humans. Geese were vaccinated with an ANDV DNA vaccine encoding the virus envelope glycoproteins. All geese developed high-titer neutralizing antibodies after the second vaccination, and maintained high-levels of neutralizing antibodies as measured by a pseudovirion neutralization assay (PsVNA) for over 1 year. A booster vaccination resulted in extraordinarily high levels of neutralizing antibodies (i.e., PsVNA80 titers >100,000). Analysis of IgY and IgYΔFc by epitope mapping show these antibodies to be highly reactive to specific amino acid sequences of ANDV envelope glycoproteins. We examined the protective efficacy of the goose-derived antibody in the hamster model of lethal HPS. α-ANDV immune sera, or IgY/IgYΔFc purified from eggs, were passively transferred to hamsters subcutaneously starting 5 days after an IM challenge with ANDV (25 LD50). Both immune sera, and egg-derived purified IgY/IgYΔFc, protected 8 of 8 and 7 of 8 hamsters, respectively. In contrast, all hamsters receiving IgY/IgYΔFc purified from normal geese (n=8), or no-treatment (n=8), developed lethal HPS. These findings demonstrate that the DNA vaccine/goose platform can be used to produce a candidate antiviral biological product capable of preventing a lethal disease when administered post-exposure.
| Our studies show the utility of combining DNA vaccination with the goose platform for the development of polyclonal avian antibodies for use as candidate medical countermeasures. We demonstrate that these antibodies have potent anti-viral neutralizing activity in cell culture and are efficacious in preventing hantavirus pulmonary syndrome in Syrian hamsters when administered as a post-exposure prophylactic. The polyclonal anti-Andes virus antibodies were not effective if administered late in the disease course indicating that the effective use of an avian polyclonal antibody-based approach to preventing hantavirus disease will require rapid diagnosis and treatment of persons presenting signs of hantavirus disease.
| Andes virus (ANDV) is a New World hantavirus from the genus Hantavirus within the family Bunyaviridae, an etiological agent of hantavirus pulmonary syndrome (HPS). Hantaviruses are enveloped viruses with trisegmented single-stranded, negative-sense RNA genomes. The three genome segments S, M, and L encode for three structural proteins: the nucleocapsid (N) protein, two glycoproteins Gn and Gc, and an RNA-dependent RNA-polymerase (RdRp), respectively [1]. ANDV was first reported and identified in southwestern Argentina in the mid-1990s [2,3], and since then outbreaks of HPS have occurred throughout South and Central America including Brazil, Chile, and Uruguay [4,5]. Most of these HPS cases are caused by ANDV, or ANDV-like viruses. Hantaviruses persist within rodents; whereas humans most likely become infected by inhalation or ingestion of virus containing urine or feces or by exposure to saliva through a bite from an infected rodent. ANDV is the only hantavirus known to be transmitted person-to-person [6,7]. Clinical HPS is characterized by a progression from flu-like symptoms and fever to non-cardiogenic pulmonary edema caused by vascular leakage. In fatal cases it is common for cardiogenic shock to develop [8]. The case fatality rate for HPS is 35–40% [4]. Despite the high mortality rate and the potential for person-to-person transmission, there are presently no approved vaccines, post-exposure prophylactics, or therapeutic treatments for HPS.
Studies emphasize the importance of the humoral immune response in hantavirus disease outcome support the use of antibodies as a potential treatment option for ANDV infection. In HPS cases, higher neutralizing antibody titers in patient’s serum have been shown to correlate with mild disease outcome [9]. Also, higher hantavirus specific IgG levels early in disease have been associated with survival [10]. In other hantavirus infections, hantavirus neutralizing activity has been related to antibodies directed to the surface glycoproteins, since monoclonal antibodies to Gn and Gc but not to N, have been shown to neutralize viral infection in vitro [11]. Specific to ANDV, a DNA vaccine expressing the M genome segment of the virus has been developed [12]. When either rhesus macaques or rabbits are vaccinated with this DNA vaccine, high-titer neutralizing antibodies are produced. Serum from these vaccinated animals, when passively transferred to ANDV-infected Syrian hamsters, protected the hamsters from lethal disease when given either before or after ANDV challenge [12,13]. It has also been shown that fresh frozen plasma (FFP) from convalescent HPS patients infected with ANDV was able to protect in the ANDV/Syrian hamster model when administered before or post-exposure [14]. In the post-exposure studies, it was necessary to administer the antibodies before high levels of serum viremia were detectable.
As an alternative to using human convalescent sera or mammalian antibodies as a source of neutralizing antibodies against ANDV, which are in short supply or have risks of reactogenicity in humans, antibodies against ANDV were purified from the eggs of ducks vaccinated with the ANDV DNA vaccine [14]. The α-ANDV IgY/IgYΔFc antibodies purified from the duck eggs were able to protect Syrian hamsters when administered after ANDV challenge [14]. An advantage to using IgY over human serum or other mammalian antibodies is the potential to prevent adverse reactions that the mammalian antibodies can have in the human body. Birds produce three different antibody isotypes: IgM, IgA, and IgY. IgY is the primary serum immunoglobulin of birds. It is the functional equivalent of mammalian IgG, but the two antibodies differ structurally. The advantage of IgY when used as a treatment include the inability to activate mammalian complement [15–20], interact with mammalian Fc receptors, or other receptors that are Fc binding, that could mediate an inflammatory reaction [21–26]. In addition to full length IgY, an alternatively spliced form of IgY, known as IgYΔFc, can be found in anseriformes birds e.g. geese and ducks, lacking the last two constant domains of the heavy chain [22,27]. In ducks, when immunized, the ratio of IgY to IgYΔFc shifts from predominantly IgY to predominantly IgYΔFc [28]. And indeed, a combination of IgY and IgYΔFc pass from mother to offspring via yolk-sac transmission [29]. Biologic and immunochemical studies have shown that IgYΔFc retains the ability to neutralize [28,30]. IgYΔFc from the eggs of vaccinated ducks or geese provide a unique source of neutralizing antibodies that is unobtainable from chicken eggs (express only full-length and membrane receptor forms of IgY) [28,31].
The transition from duck/egg to goose/egg was made based on preliminary studies showing increased frequency of high-titer responses in DNA-vaccinated geese versus ducks. Other advantages included a well-defined lineage of birds, 30+ generation breeding stock, expertise in breeding/hatching/goose husbandry, and the ability to collect more mg of antibody per yolk.
Here, we used geese vaccinated with the ANDV DNA vaccine to determine the post-exposure potential of goose-derived α-ANDV antibodies. The neutralizing capabilities of α-ANDV IgY/IgYΔFc purified from egg yolks of DNA vaccinated geese were determined after initial and long-term booster vaccinations. In addition, we identified epitopes within the ANDV glycoproteins recognized by IgY/IgYΔFc purified from egg yolks of DNA vaccinated geese. We show here, that neutralizing α-ANDV titers were maintained during the time period between initial and booster vaccinations, and that neutralizing titers were further increased after long-range booster vaccination. Finally, we demonstrate that polyclonal neutralizing antibodies produced using the DNA vaccine/goose platform administered post-exposure are capable of preventing disease in the ANDV/Syrian hamster model of lethal HPS.
Geese were vaccinated with an ANDV DNA vaccine, either pWRG/AND-M(opt) for Group A or pWRG/AND-M(1.1) for Group B, both containing the full-length M genome segment of ANDV strain Chile- 9717869. The difference between the plasmids used has been described previously [32]. The initial vaccination series included intramuscular vaccination at two-week intervals up to10 weeks. The first generation PharmaJet DSJI (blue) device was used for all vaccinations. One year later, the immune geese were boosted with the original or codon-optimized version of the AND-M vaccine, denoted either pWRG/AND-M(opt) or pWRG/AND-M(opt2) at week 52, 54, 56, 60, 61, and 62 (Fig 1A). Sera were collected from the vaccinated geese throughout the initial vaccination series and the long-range booster vaccination series. Sera neutralizing antibody titers were determined using the pseudovirion neutralization assay (PsVNA) [32]. Individual geese titers are shown for the initial vaccination period (Fig 1C) and the long-range boost (Fig 1D). The mean of Group A (receiving the pWRG/AND-M(opt) boost) and Group B (receiving the pWRG/AND-M(opt2) boost) indicate an order of magnitude increase in titer after the long-range boost. PsVNA neutralization 80% PsVNA80) titers as high as 100,000 were produced in the serum of vaccinated geese. Similarly, plaque reduction neutralization (PRNT) 50% titers of 20,480 were produced (S1 Fig). The PRNT assay is similar to the PsVNA (to obtain neutralizing antibody titers); however, ANDV is used rather than an ANDV pseudovirion. There was no significant difference between the AND-M DNA vaccines used during the initial vaccination series, or the long-range boost (Fig 1B).
Goose eggs were collected after the initial vaccination and immediately after the long-range boost (following week 54 vaccination) (Fig 1A). Total IgY was purified from eggs yolks and evaluated for α-ANDV neutralizing activity by PsVNA. Similar to serum collected from the vaccinated geese, eggs collected after long-range boost showed an order of magnitude increase in neutralization titer (p = 0.0007) (Fig 2A). A pool was made containing purified IgY with the highest neutralization titer for animal experiments. The PRNT80 titer of the egg-derived IgY/ IgYΔFc was 2,560 and the PsVNA80 was 60,231, with a final protein concentration of 12.5 mg/ml. The yield of purified IgY is between 50–160 mg/yolk. Purified IgY from eggs (98% purity) collected at time points throughout the long-range boost were visualized by Coomassie staining (Fig 2B). Both IgY and IgYΔFc are detectable in eggs collected from Group A and Group B-vaccinated geese. Under nonreducing conditions, IgY is detected at approximately 180 kDa and IgYΔFc is detected at approximately 120 kDa (S2A Fig). Lower molecular weight bands were tested by Western blot using antibodies specific for IgY heavy and light chain subunits and found to correspond to the IgY light chain and heavy chains from IgY and IgYΔFc (S2B and S2C Fig). Total IgY isolated from eggs collected from geese from Group A show a shift in the ratio of IgY to IgYΔFc towards a preponderance of full-length IgY later in the long-range boost vaccination series.
To determine the specificity of the IgY/IgYΔFc antibodies purified from the egg yolks of geese vaccinated with the ANDV M gene-based DNA vaccine, linear epitope mapping was completed using microarray slides. The M genome segment is the primary viral component of the ANDV DNA vaccine; two glycoproteins Gn and Gc are synthesized from the M segment and are present as oligomers on the outside surface of mature ANDV virions. To identify potential linear epitopes of ANDV glycoproteins Gn and Gc recognized by α-ANDV goose IgY and IgYΔFc, microarray slides were covalently linked with 13-mer peptides with a 10 amino acid overlap, and 3 amino acid offset, for a total of 376 peptides, spanning the entire sequence of the glycoproteins from ANDV strain Chile-9717869 (GenBank accession number AF291703). IgY, IgYΔFc, or IgY/IgYΔFc purified from the eggs of geese vaccinated at both initial and booster vaccination time points, as described in the Methods section, were incubated with microarray slides. Reactivity was compared to negative control features on the slide that did not contain any protein; the average of the negative controls was taken and used for comparison.
Eleven IgY/IgYΔFc reactive epitopes were identified across both glycoproteins Gn and Gc (Fig 3) outside of regions recognized by normal goose IgY/IgYΔFc. Six of the epitopes were specific for the Gn glycoprotein, peptides starting with aa 34–40, 82–85, 154–163, 259, 283–286, and 574 and the remaining five epitopes were within the Gc glycoprotein, reactive peptides started with aa 685–697, 853–856, 940–952, 1060–1066, and 1117–1123. When comparing the epitopes recognized by IgY/IgYΔFc, to the separated IgY or IgYΔFc, all goose antibody preparations recognized the same epitopes with similar levels of reactivity. There was no reactivity detected with any negative control slide features.
Because geese receiving the long-range booster vaccination showed increased IgY neutralizing antibody titers compared to the initial vaccination, we wanted to see if there was any change to the epitopes recognized by IgY/IgYΔFc from the geese receiving the long-range boost. Fig 3 shows that IgY/IgYΔFc from the egg yolks of the long-range boosted geese reacted with all of the same epitopes as the IgY/IgYΔFc from the initial vaccination and reactivity was either at the same level or increased. Areas of special interest because of increased activity after booster vaccination include areas starting with aa 34–40, 82–85, 259, 940–952, and 1060–1066 (Fig 3).
Sera from rabbits that were previously vaccinated with pWRG/AND-M [33] were also incubated with microarray slides to determine potential epitopes recognized by mammalian antibodies generated after vaccination. Most of the rabbit sera IgG antibodies recognized the same epitopes as IgY/IgYΔFc. However, in the majority of cases there was a noticeable difference in the level of binding of the rabbit IgG when compared to the IgY/IgYΔFc. There was increased binding to the peptide starting with aa 313 in Gn. There was decreased binding to peptides starting with aa 685–697 (Fig 3). When comparing the rabbit IgG to the IgY/IgYΔFc from the eggs of booster vaccinated geese, there was one obvious region of decreased reactivity in Gn corresponding to the peptide starting with aa 34 (Fig 3). There were two unique epitopes recognized strongly by the rabbit IgG compared to IgY/IgYΔFc from the eggs of vaccinated geese comprised of peptides starting with aa 223–229 in Gn along with peptide 760 in Gc (Fig 3).
Important protective epitopes are most likely the five areas with increased IgY/IgYΔFc binding after booster vaccination. These five regions were also recognized by the rabbit sera, but to a lesser reactivity level than IgY/IgYΔFc from booster vaccinated geese and all IgY treatments (IgY, IgYΔFc, and IgY/IgYΔFc) for some regions. All five regions with increased reactivity are outside of the regions identified as reactive by human sera from ANDV patients, and sera from naturally infected rodents [34]. Epitopes in the study by Tischler et al. characterized as having strong reactivity for serum from humans ANDV HPS patients across both Gn and Gc were aa 14–26, 691–703, and 955–967. The strongly reactive epitopes for serum from naturally infected rodents were aa 599–611 within Gn and aa 691–703, 918–30, and 955–967 within Gc [34]. In addition, the highly reactive IgY epitopes do not overlap with regions recognized by monoclonal antibodies that neutralize other hantaviruses [35–37]. Taken together these results show that IgY isolated from eggs yolks of vaccinated geese was ANDV specific and recognized unique epitopes compared to human and rabbit serum.
Having produced potent α-ANDV neutralizing antibodies using the goose platform, we were interested in testing the efficacy of this antiviral biologic in an animal model of hantavirus disease. Before performing a protection study in the ANDV/Syrian hamster model of lethal HPS, a bioavailability experiment was performed testing the goose-generated IgY/IgYΔFc in hamsters (Fig 4). Groups of three hamsters were injected by the subcutaneous route with 64,000 NAU/kg (neutralizing antibody units/kilogram) or 12,000 NAU/kg α-ANDV IgY/IgYΔFc. Serum samples were collected on days 1, 3, 6, 9, 15, and 21 after antibody injection, and titers were assessed by PsVNA. α-ANDV neutralizing antibodies were detected on days 1 and 3 for the group of hamsters receiving 64,000 NAU/kg α-ANDV IgY/IgYΔFc, then dropped below the level of detection for the assay. On day 3, only a single hamster from the group receiving 12,000 NAU/kg α-ANDV IgY/IgYΔFc had a positive neutralizing antibody titer, which was reduced compared to the high dosage group. These findings were used to determine the dose and schedule of antibody injections in protection experiments described below.
We next determined the protective efficacy of goose-generated α-ANDV IgY/IgYΔFc in the hamster model of lethal HPS. On day 0, all hamsters were challenged with 200 PFU of ANDV by the i.m. route (Fig 5A). On day 5, groups of 8 hamsters were administered 12,000 NAU/kg α-ANDV rabbit sera as a positive control, 20,000 NAU/kg α-ANDV goose sera, 20,000 NAU/kg α-ANDV IgY/IgYΔFc, an equivalent dose of normal goose sera, or an equivalent protein concentration of normal goose IgY/IgYΔFc as negative controls. All antibody injections were by the s.c. route. On day 8, these same groups were administered a second treatment of 20,000 NAU/kg α-ANDV goose sera, 20,000 NAU/kg α-ANDV IgY/IgYΔFc, an equivalent dose of normal goose sera, or an equivalent dose of normal goose IgY/IgYΔFc by the s.c. route. All hamsters receiving normal goose sera, normal goose IgY/IgYΔFc, or no antibody treatment succumbed to HPS between days 10 and 17 postinfection displaying clinical signs of HPS (i.e. staggered gait, tachypnea). Of the α-ANDV treatment groups, 7/8 hamsters receiving positive control rabbit sera survived to day 28 (p<0.0001), 8/8 hamsters receiving α-ANDV goose sera survived to day 28 (p<0.0001), and 7/8 hamsters receiving α-ANDV goose-egg derived purified IgY/IgYΔFc survived to day 28 (p = 0.0003). None of the surviving hamsters displayed clinical signs of HPS. Sera from all surviving hamsters was collected on day 28 and subjected to an N-ELISA (Fig 5B). Results from the ELISA indicated that all hamsters were productively infected with ANDV, yet did not succumb to lethal HPS.
Having evaluated α-ANDV IgY/IgYΔFc administered as a post-exposure prophylactic (prior to the onset of viremia in the ANDV hamster model), we next determined the efficacy of α-ANDV IgY/IgYΔFc when administered as a potential therapeutic (after the onset of viremia). For a 200 PFU i.m. ANDV challenge, the mean day-to-death is 11 days post-infection [14,33]. Therefore, treatment beginning on day 8 post-infection would be 2 days after the onset of viremia (day 6) and 3 days before the mean day-to-death. In this experiment, the dosage of α-ANDV neutralizing antibodies was doubled from 20,000 to 40,000 NAU/kg. In addition to α-ANDV IgY/IgYΔFc, α-ANDV FFP from a Chilean HPS survivor [38] was evaluated for its efficacy as a therapeutic. On day 0, all hamsters were challenged with 200 PFU of ANDV by the i.m. route (Fig 6A). On days 8 and 10, groups of hamsters were passively transferred with 40,000 NAU/kg α-ANDV IgY/IgYΔFc, 40,000 NAU/kg α-ANDV FFP, or equivalent protein concentration of normal IgY or equivalent dose normal FFP. A group of 8 hamsters were untreated to serve as an infection control. All hamsters from the α-ANDV IgY/IgYΔFc, α-ANDV FFP, normal IgY, and normal FFP succumbed to lethal disease between days 9–12 post-infection. Two hamsters from the ANDV infection only control group survived to day 28 (end of study). Sera from these hamsters were subjected to an N-ELISA demonstrating a productive infection (S3 Fig).
RNA isolated from the lungs of a subset of ANDV-infected hamsters on day 10 was evaluated for ANDV viral genome by RT-PCR (Fig 6B). These results show no statistically significant difference between any of the treatment groups and control groups. These findings indicate that a late treatment with a dose of α-ANDV antibodies, regardless of goose or human source, is insufficient to lower the amount of virus accumulating in the lung, and is unable to reverse the disease course in the hamster model.
ANDV has been associated with a majority of HPS cases [3] and continues to be the only hantavirus capable of human-to-human transmission. In spite of the continuing number of HPS cases and the staggering mortality rate of 35–40% there are no available treatments or preventative vaccines. The potential for the use of passive treatments to protect against HPS has already been established by previous studies utilizing immune serum from infected patients and isolated antibodies from vaccinated animals in the ANDV/hamster model [12–14,39]. In addition, evaluation of HPS patients has highlighted the importance of neutralizing antibody production in recovery from infection [9,10]. An obvious source of antibodies for passive treatment would be immune plasma from convalescent patients, but immune plasma is in short supply and can pose problems of reactogenicity when given to other patients if not blood group-typed appropriately. Protective monoclonal antibodies are another option for treatment and have been used with other viral infections [40], but thus far, ANDV neutralizing monoclonal antibodies have not been described. Polyclonal antibodies have been generated against toxins and venoms in vaccinated sheep and horses, but this method has yet to be successful in generating any antiviral treatments. Avian antibodies are yet another logical alternative source of passive therapeutics with the potential to overcome the shortcomings of current therapeutics antibodies; be it source shortages, reactivity to Fc portions of mammalian antibodies, or lack of reactivity with neutralizing epitopes. IgY is the primary serum antibody of birds and is transferred to the egg yolk via receptors on the surface of the yolk membrane that is specific for IgY translocation, causing the yolk to have high IgY concentrations [27,41–44]. IgY can then be purified from egg yolks and in large quantities for use as therapeutics. Current and previous research using polyclonal avian IgY has already established a baseline for its therapeutic potential against infectious agents e.g. Pseudomonas aeruginosa [22,45–47] and Candida albicans [48]. IgY antibodies have also been developed against different venoms and antitoxins [49–55]. Most important for this research, DNA vaccinated birds have also been used to produce virus-specific IgY [14,56,57].
DNA vaccination of ducks with pWRG/AND-M had already been shown to result in the production of ANDV neutralizing antibodies [14]. Here, geese were vaccinated and initial total IgY PsVNA80 titers from eggs were 2,567 (geometric mean titer, GMT from Fig 2). Long-range booster vaccination resulted in an increase in total egg IgY neutralizing antibody PsVNA80, titer GMT = 62,195. The increase in titers correlated with an increase in ANDV-specific neutralization PsVNA80 titers for IgY/IgYΔFc in sera. PsVNA80 titers for sera increased from >1,000 just 4 weeks into the initial vaccination series to >10,000 3 weeks into booster vaccination (Fig 1). We demonstrated that neutralizing antibody titers were maintained during the year between initial vaccination and booster vaccination. While unexpected, these results point out the advantage of long-range boost leading to an order of magnitude enhanced antibody response.
It is interesting to note the shift in IgY to IgYΔFc ratio to predominantly full-length IgY during the long-range boost vaccination series (Fig 2B). In an early publication detailing duck immunoglobulins, the author describes a shift from predominantly IgY to IgYΔFc in serum from hyper-vaccinated ducks [28]. We speculate that this difference could potentially be attributed to the use of a protein vaccine as opposed to a DNA vaccine. However, this difference could also be attributed to the species used. Following DNA vaccination, total IgY isolated from duck eggs contained 75% IgYΔFc [14]. Regardless, the shift towards IgY or IgYΔFc does not appear to have a differential effect in efficacy studies. Future experiments isolating α-ANDV full-length IgY from α-ANDV IgYΔFc will determine the contribution of Fc in protection studies.
Epitope mapping using IgY purified from egg yolks of vaccinated geese identified several regions of reactivity on the ANDV surface glycoproteins Gn and Gc. It is likely that many epitopes recognized by ANDV neutralizing antibodies are conformational; however, it is possible that there are linear neutralizing epitopes as well. In total, following the initial vaccination there were 11 epitopes recognized. After the long-range booster vaccination there were five epitopes with increased reactivity. The increase in neutralizing capabilities of both sera from DNA vaccinated geese along with IgY/IgYΔFc isolated from egg yolks of vaccinated geese supports the protective potential of these five epitopes making these epitopes of high interest. Four of the five highly reactive regions in glycoproteins Gn and Gc did not overlap with regions previously identified as reactive by serum from human ANDV HPS patients or serum from naturally infected rodents [34]. The only overlap with human and rodent epitopes was at the epitope made up of aa 685–697. This region overlapped the strongly reactive rodent and human serum epitope made up of amino acids 691–703 [58]. This is a potentially immunodominant domain since it is being recognized by antibodies from multiple species.
Looking at the predicted structure of the ANDV glycoproteins as part of a mature virion, all five epitopes with increased reactivity after long-range boost are within regions potentially accessible from the surface of the virion. AA 34–40, 82–85, and 259 are part of the predicted Gn ectodomain prior to WAASA the cleavage site, found at aa 647–651, where the glycoprotein precursor is cleaved into the two glycoproteins Gn and Gc [58]. Peptides starting with aa 34–40 are near the n terminus of Gn. The secondary structure of this region is predicted to be a mixture of α-helices and β strands. For the Gc glycoprotein highly reactive regions aa 940–952 and 1060–1066 are within the predicted Gc ectodomain of the protein with the secondary structure primarily being made of β-sheets and random coils. In addition all five of these regions are outside of predicted hydrophobic regions, or glycosylation sites of the glycoproteins [58]. These epitopes of high interest recognized by the IgY/IgYΔFc from the egg yolks of geese vaccinated with at long-range booster represent novel potentially neutralizing epitopes on the ANDV glycoproteins.
We have shown previously that neutralizing antibodies administered on day 5 following a 200 PFU ANDV i.m. challenge can protect hamsters from lethal HPS [14]. In that study, 75% of hamsters receiving 12,000 NAU/kg α-ANDV duck IgY/IgYΔFc survived ANDV infection. In order to bolster the survival percentage, 20,000 NAU/kg α-ANDV goose IgY/IgYΔFc was administered to hamsters on days 5 and 8 post-infection (based on bioavailability results). This resulted in 88% survival (7/8 hamsters). In a clinical setting, it is likely that increasing the neutralizing antibody per dose, and increased frequency of dosing, would increase the efficacy of any antibody-based treatment for HPS.
In order to evaluate the use of IgY/IgYΔFc as a therapeutic, an experiment was conducted to determine the efficacy of administering a high concentration of α-ANDV neutralizing antibody after the onset of viremia, which in the 200 PFU i.m. challenge starts on day 6 [59]. By doubling the concentration of antibodies delivered (from 20,000 NAU/kg to 40,000 NAU/kg) and starting treatment on day 8 post-infection, we were unable to protect ANDV-infected hamsters from lethal HPS. It is unknown if a higher concentration of neutralizing antibody alone, changing the route of administration, and/or a treatment starting on day 7, which would still be post-viremia, would be sufficient to achieve a survival outcome. Ribavirin, favipiravir (T-705), and neutralizing antibodies have all been proven efficacious prior to the onset of viremia [12,14,60,61]. To date, there is no treatment option for HPS that has demonstrated efficacy after the onset of viremia. This highlights the need for a treatment option that expands the current therapeutic window.
A compassionate open trial is underway in Chile using α-ANDV human FFP to treat ANDV infections. ABO-compatible immune plasma at a dosage of 5,000 NAU administered to confirmed hantavirus cases resulted in a reduction in the rate of lethality from 32% to 14% [62]. This is a borderline statistically significant reduction in lethality; however, it is suggestive of the potential clinical benefit of this type of therapy. We speculate that the timing of α-ANDV plasma therapy plays a critical role in the ability of the antibodies to neutralize virus and prevent morbidity and mortality. A greatly increased concentration of neutralizing antibodies (40,000 NAU/kg) was insufficient to protect hamsters from lethal HPS if administered 3 days prior to the mean day-to-death, further supporting the argument for early intervention in suspected HPS cases.
In this study we used the DNA vaccine/goose platform to produce a biologic consisting of purified IgY antibodies targeting both ANDV Gn and GC envelope glycoproteins. This candidate product has potent anti-viral neutralizing activity and is effective at preventing disease when administered as a post-exposure prophylactic in the Syrian hamster model. Translating an egg-derived polyclonal antibody product to the clinic is a daunting challenge; however, the ongoing phase III clinical trial test the efficacy of avian polyclonal IgY antibodies against Pseudomonas aeruginosa for the treatment of cystic fibrosis under the auspices of the European Medicines Agency indicates that novel avian antibody-based approaches to the development of medicines to prevent and treat infectious disease has merit.
A twice plaque-purified ANDV strain Chile-9717869 passaged in Vero E6 cells (Vero C1008, ATCC CRL 1586) was described previously [63]. Cells were maintained in Eagle’s minimum essential medium with Earle’s salts (EMEM) supplemented with 10% fetal bovine serum, 10nM HEPES (pH 7.4), 200 U/ml penicillin, 200 μg/ml streptomycin, 1X nonessential amino acids, 1.5 μg/ml amphotericin B, and 50 μg/ml gentamicin sulfate (cEMEM) at 37°C in a 5% CO2 incubator.
pWRG/AND-M (1.1) has been described previously [13]. Details of codon-optimized DNA vaccines pWRG/AND-M(opt) and pWRG/AND-M(opt2) are contained in a separate manuscript [32]. Briefly, pWRG/AND-M(1.1) was generated using viral RNA from ANDV-infected Vero E6 cells. pWRG/AND-M(opt) is identical to pWRG/AND-M(1.1) with the ORF codon-optimized for Homo sapiens but missing a stop codon, resulting in the addition of 24 amino acids to the Gc C-terminus. pWRG/AND-M(opt2) is the corrected DNA vaccine.
Two groups containing four geese each (Anser domesticus, 25 months old) were immunized with the indicated AND-M DNA vaccine intramuscularly using the v1.0 PharmaJet injector. Five 1mg DNA inoculations were delivered to the breast at indicated times. The same delivery device and site was used for the long-range boost immunized at approximately 1 year after the initial immunization to match the start of the next goose laying season.
Yolks were isolated, rinsed with water, and punctured to drain the contents. These contents were diluted 1:10 with cold, deionized water, stirred, and acidified to pH 5.0. The diluted yolk was centrifuged at 10,000 x g for 30 min and supernatant was filtered. An equal volume of 100% saturated ammonium sulfate was added to the filtered supernatant to give a 50% saturation, stirred for 30 min, and centrifuged at 10,000 x g for 15 min. The pellet was suspended in 50 mM TrisCl pH 8.0. Further purification was achieved via hydrophobic charge induction chromatography on 4-Mercapto-Ethyl-Pyridine-lined (MEP) HyperCel sorbent (Pall Biosciences) followed by concentration using a Tangential flow filtration and diafilitration with 1x PBS buffer. This method results in a combination of full-length IgY and truncated IgYΔFc.
The PsVNA was run as previously described [32]. Briefly, a 1:10 dilution of heat-inactivated sera was made followed by five-fold serial dilutions that were mixed with equal volume of cEMEM containing 4,000 FFU ANDV PsV with 10% guinea pig complement. The mixture was incubated overnight at 4°C. Following this incubation, 50μl was inoculated onto Vero cell monolayers in a clear bottom black-walled 96-well plate (Corning) in triplicate. Plates were incubated at 37°C for 18–24 hrs. The media was discarded, and cells were lysed according to the luciferase kit protocol (Promega #E2820). A Tecan M200 Pro was used to acquire raw luciferase data. The values were graphed using GraphPad Prism software (version 6) to calculate the 80% neutralization and then interpolate to obtain PsVNA80 titers.
Duplicate IgY/IgYΔFc samples were separated by 4–15% gradient SDS-PAGE run under nonreducing conditions. The gel was incubated in Bio-Safe Coomassie G-250 stain (Bio-Rad Laboratories) for 30 minutes and subsequently destained (deionized water) over 2.0 hrs before visualized. Signals were captured using AlphaView software and AlphaImager HP Imaging System (Alpha Innotech).
Linear IgY epitopes were identified using JPT PepStar microarrays. The entire glycoprotein precursor sequence of the ANDV strain Chile-9717869 was synthesized into 13 amino acid peptides. The 376 resulting 13-mer peptides were covalently attached to a microarray slide with a 10 amino acid overlap. Peptides were analyzed for their reactivity with IgY antibodies isolated from geese eggs, following protocols recommended by JPT. Briefly, slides were incubated with 30μg/mL primary antibody at 4°C overnight in a moist environment. The slide was washed and incubated with a fluorescently labeled secondary antibody, goat anti-chicken IgY conjugated to Cy5 (1ug/mL) (Abcam, for 1 hour at 30°C. After washing and drying the slide, bound antibodies were detected using a microarray reader (Genepix 4000). Fluorescence was measured at a 10um pixel size and mean values with the background corrected were calculated and used for analysis. The reactivity was classified based on a spectrum ranging from no activity in white, mild reactivity in gray, to strong reactivity in red.
Female Syrian hamsters aged 6–8 weeks (Harlan) were anesthetized by inhalation of vaporized isoflurane using an IMPAC 6 veterinary anesthesia machine. Once anesthetized, hamsters were injected with 200 PFU of virus diluted in PBS. Intramuscular (i.m.) (caudal thigh) injections consisted of 0.2ml delivered with a 1ml syringe with a 25-gauge, 5/8in needle.
Hamsters were anesthetized by inhalation of vaporized isoflurane using an IMPAC 6 veterinary anesthesia machine. Once anesthetized, hamsters were injection with antibody diluted in PBS. Subcutaneous injections consisted of 1–2ml delivered with a 3ml syringe with a 23-gauge, 1in needle. Positive control rabbit sera was collected from rabbits vaccinated with pWRG/AND-M four times by muscle electroporation [33].
The enzyme-linked immunosorbent assay (ELISA) used to detect N-specific antibodies (N-ELISA) was described previously [64,65]. The endpoint titer was determined as the highest dilution that had an optical density (OD) greater than the mean OD for serum samples from negative-control wells plus 3 standard deviations. The PUUV N antigen was used to detect ANDV N-specific antibodies as previously reported [63].
Approximately 250 mg of lung tissue was homogenized in 1.0 ml TRIzol reagent using gentleMACS M tubes and a gentleMACS dissociator on the RNA setting. RNA was extracted from TRIzol samples as recommended by the manufacturer. The concentration of the extracted RNA was determined using a NanoDrop 8000 instrument and standardized to a final concentration of 100 ng/ul. Real-time PCR was conducted on a BioRad CFX thermal cycler using an Invitrogen Power SYBR Green RNA-to-Ct one-step kit according to the manufacturer’s protocols. Primer sequences are ANDV S 41F 5’-GAA TGA GCA CCC TCC AAG AAT TG-3’ and ANDV S 107R 5’-CGA GCA GTC ACG AGC TGT TG-3’ [66]. Cycling conditions were 30 min at 48°C, 10 min at 95°C, followed by 35 cycles of 15 sec at 95°C and 1 min at 60°C. Data acquisition occurs following the annealing step.
Comparison of egg IgY titers was done using Student’s t-test (two-tailed). P values of less than 0.05 were considered significant. Survival analyses were compared using Kaplan-Meier survival analysis with log rank tests and P-values adjusted by simulation or by Dunnett’s test to account for multiple comparisons. Analyses were conducted using GraphPad Prism (version 6).
The goose work was approved by the University of North Dakota Institutional Animal Care and Use Committee: Protocol Number 1403–1. The hamster work was approved by the USAMRIID Institutional Animal Care and Use Committee.
Research was conducted under an IACUC approved protocol in compliance with the Animal Welfare Act, PHS Policy, and other Federal statutes and regulations relating to animals and experiments involving animals. The USAMRIID is accredited by the Association for Assessment and Accreditation of Laboratory Animal Care, International and adheres to principles stated in the Guide for the Care and Use of Laboratory Animals, National Research Council, 2011. Opinions, interpretations, conclusions, and recommendations are ours and not necessarily endorsed by the U.S. Army or the Department of Defense.
|
10.1371/journal.pbio.2006387 | Classes and continua of hippocampal CA1 inhibitory neurons revealed by single-cell transcriptomics | Understanding any brain circuit will require a categorization of its constituent neurons. In hippocampal area CA1, at least 23 classes of GABAergic neuron have been proposed to date. However, this list may be incomplete; additionally, it is unclear whether discrete classes are sufficient to describe the diversity of cortical inhibitory neurons or whether continuous modes of variability are also required. We studied the transcriptomes of 3,663 CA1 inhibitory cells, revealing 10 major GABAergic groups that divided into 49 fine-scale clusters. All previously described and several novel cell classes were identified, with three previously described classes unexpectedly found to be identical. A division into discrete classes, however, was not sufficient to describe the diversity of these cells, as continuous variation also occurred between and within classes. Latent factor analysis revealed that a single continuous variable could predict the expression levels of several genes, which correlated similarly with it across multiple cell types. Analysis of the genes correlating with this variable suggested it reflects a range from metabolically highly active faster-spiking cells that proximally target pyramidal cells to slower-spiking cells targeting distal dendrites or interneurons. These results elucidate the complexity of inhibitory neurons in one of the simplest cortical structures and show that characterizing these cells requires continuous modes of variation as well as discrete cell classes.
| Single-cell RNA sequencing allows scientists to count the number of copies of each gene expressed in multiple individually isolated cells. Because different cell types express genes in different amounts, “clusters” of cells with similar expression patterns are likely to correspond to different cell types. As well as discrete classes, however, cells also show continuous variation in gene expression. To study the relationship between cell classes and continua in a well-understood brain system, we applied new analysis methods to a dataset of inhibitory interneurons from area CA1 of the mouse hippocampus. Thanks to decades of intensive work, at least 23 classes of CA1 interneurons have been previously defined. We were able to identify them all with our transcriptomic clusters but unexpectedly found three to be identical. Because the connectivity of these cells has already been established, we were also able to identify the primary mode of continuous variation in these cells, which related to their axon target location. This in-depth understanding of the relatively simple cortical circuit of CA1 not only clarifies the cellular composition of this important brain structure but also will form a solid foundation for understanding more complex structures, such as the isocortex.
| Cortical circuits are composed of highly diverse neurons, and a clear definition of cortical cell types is essential for the explanation of their contribution to network activity patterns and behavior. Cortical neuronal diversity is strongest amongst GABAergic neurons. In hippocampal area CA1—one of the architecturally simplest cortical structures—GABAergic neurons have been divided so far into at least 23 classes of distinct connectivity, firing patterns, and molecular content [1–6]. A complete categorization of CA1 inhibitory neurons would provide not only essential information to understand the computational mechanisms of the hippocampus but also a canonical example to inform studies of more complex structures, such as six-layered isocortex.
CA1 GABAergic neurons have been divided into six major groups based on connectivity and expression patterns of currently used molecular markers. Parvalbumin (PVALB)-positive neurons (including basket, bistratified, and axo-axonic cells) target pyramidal cells’ somata, proximal dendrites, or axon initial segments, firing fast spikes that lead to strong and rapid suppression of activity [7,8]. Somatostatin (SST)-positive oriens/lacunosum-moleculare (O-LM) cells target pyramidal cell distal dendrites and exhibit slower firing patterns [9]. GABAergic long-range projection cells send information to distal targets and comprise many subtypes, including SST-positive hippocamposeptal cells; NOS1-positive backprojection cells targeting dentate gyrus and CA3; and several classes of hippocamposubicular cells, including trilaminar, radiatum-retrohippocampal, and PENK-positive neurons [10–14]. Cholecystokinin (CCK)-positive interneurons are a diverse class characterized by asynchronous neurotransmitter release [15,16] that have been divided into at least five subtypes targeting different points along the somadendritic axis of pyramidal cells [17–21]. Neurogliaform and Ivy cells release GABA diffusely from dense local axons and can mediate volume transmission as well as conventional synapses [22,23]. Interneuron-selective (I-S) interneurons comprise at least three subtypes specifically targeting other inhibitory neurons and expressing one or both of Vasoactive intestinal polypeptite (VIP) and calretinin (CALB2) [2,24–26]. Finally, additional rare types, such as large SST/NOS1 cells [27], have been described at a molecular level, but their axonal targets and relationship to other subtypes is unclear.
This already complex picture likely underestimates the intricacy of CA1 inhibitory neurons. Currently defined classes likely divide into several further subtypes, and additional neuronal classes likely remain to be found (e.g., [28]). Furthermore, it is unclear whether a categorization into discrete classes is even sufficient to describe the diversity of cortical inhibitory neurons [29,30]. For example, several CCK interneuron classes have been described, targeting pyramidal cells at multiple locations ranging from their somata to distal dendrites, and the molecular profile and spiking phenotype of these cells correlates with their synaptic target location, with fast-spiking cells more likely to target proximal segments of pyramidal neurons [18,19,21]. Do such cells represent discrete classes with sharp interclass boundaries, or do they represent points along a continuum? Finally, while a cell’s large-scale axonal and dendritic structure likely remains fixed throughout life, both gene expression and electrophysiological properties can be modified by factors such as neuronal activity [31–34]. To what extent is the observed molecular diversity of interneurons consistent with activity-dependent modulation of gene expression?
Single-cell RNA sequencing (scRNA-seq)—which can read out the expression levels of all genes in large numbers of individual cells—provides a powerful opportunity to address these questions. This method has successfully identified the major cell classes in several brain regions [35–46]. Nevertheless, identifying fine cortical cell classes has not been straightforward because of both incomplete prior information on the underlying cell types and complicating factors, such as potential continuous variability within these classes. The large body of prior work on CA1 interneurons provides a valuable opportunity to identify transcriptomic clusters with known cell types in an important cortical circuit, enabling confident identification of known and novel classes and investigation of questions such as continuous variability.
Here, we describe a transcriptomic analysis of 3,663 inhibitory neurons from mouse CA1. This analysis revealed 49 clusters, of which we could identify 41 with previously described cell types, with the remaining 8 representing putative novel cell types. All previously described CA1 GABAergic classes could be identified in our database, but our results unexpectedly suggest that three of them are identical. The larger number of clusters occurring in our transcriptomic analysis reflected several previously unappreciated subtypes of existing classes and tiling of continua by multiple clusters. Our data suggest a common genetic continuum exists between and within classes, from faster-firing cells targeting principal cell somata and proximal dendrites, to slower-firing cells targeting distal dendrites or interneurons. Several classes previously described as discrete represent ranges along this continuum of gene expression.
We collected cells from six Slc32a1-Cre;R26R-tdTomato mice, three of age p60 and three of age p27. Cells were procured using enzymatic digestion and manual dissociation [46], and data were analyzed using the 10X Genomics “cellranger” pipeline, which uses unique molecular identifiers (UMIs) to produce an absolute integer quantification of each gene in each cell. The great majority of cells (4,572/6,971 cells total; 3,283/3,663 high-quality interneurons) came from the older animals. Because we observed no major difference in interneuron classes between ages, data were pooled between them (S1 Fig). Fluorescence-activated cell sorting (FACS) yielded an enriched but not completely pure population of GABAergic neurons. A first-round clustering (using the method described below) was therefore run on the 5,940 cells passing quality control, identifying 3,663 GABAergic neurons (as judged by the expression of genes Gad1 and Slc32a1).
We analyzed the data using a suite of four novel algorithms derived from a probabilistic model of RNA distributions. All four methods were based on the observation that RNA counts within a homogeneous population can be approximated using a negative binomial distribution (see Methods, [47,48]). The negative binomial distribution accurately models the high variance of transcriptomic read counts (S2A and S2B Fig). As a consequence, algorithms based on this distribution weight the presence or absence of a gene more than its numerical expression level—for example, this distribution treats read counts of 0 and 10 as more dissimilar than read counts of 500 and 1,000 (S2C Fig).
The algorithm we used for clustering was termed ProMMT (Probabilistic Mixture Modeling for Transcriptomics). This algorithm fits gene expression in each cluster k by a multivariate negative binomial distribution with cluster-specific mean μk. The mean expression levels of only a small subset of genes are allowed to vary between clusters (150 for the current analysis; S3 Fig); these genes are selected automatically by the algorithm by maximum likelihood methods. The use of such “sparse” methods is essential for probabilistic classification of high-dimensional data [49], and the genes selected represent those most informative for cluster differentiation. The number of clusters was chosen automatically using the Bayesian Information Criterion (BIC) [50]. The ProMMT algorithm also provides a natural measure of the distinctness of each cluster, which we term the isolation metric (see Methods).
The ProMMT algorithm divided CA1 interneurons into 49 clusters (Fig 1). We named the clusters using a multilevel scheme after genes that are strongly expressed at different hierarchical levels; for example, the cluster Calb2.Vip.Nos1 belongs to a first-level group characterized by strong expression of Calb2 (indicating I-S interneurons); a second-level group Calb2.Vip; and a third-level group distinguished from other Calb2.Vip cells by stronger expression of Nos1. This naming scheme was based on the results of hierarchical cluster analysis of cluster means, using a distance metric based on the negative binomial model (Methods; Fig 1).
To visualize cell classes in two dimensions, we modified the t-stochastic neighbor embedding (tSNE) algorithm [51] for data with negative binomial variability, terming this approach nbtSNE (negative binomial tSNE). In conventional tSNE, the similarity between data points is defined by their Euclidean distance, which corresponds to conditional probabilities under a Gaussian distribution. We obtained greater separation of clusters and a closer correspondence to known cell types by replacing the Gaussian distribution with the same negative binomial distribution used in our clustering algorithm (see Methods; S4 Fig).
The nbtSNE maps revealed that cells were arranged in 10 major “continents” (Fig 2). The way expression of a single gene differed between classes could be conveniently visualized on these maps by adjusting the symbol size for each cell according to that gene’s expression level. Consistent with previous transcriptomic analyses, we found that classes were rarely, if ever, identified by single genes but rather by combinatorial expression patterns. Thanks to the extensive literature on CA1 interneurons, 25 genes together sufficed to identify the main continents with known cell classes (Fig 3), and it was also possible to identify nearly all the finer subclasses using additional genes specific to each class (S1 Text).
Previous work has extensively characterized the connectivity, physiology, and firing patterns of CA1 inhibitory neurons, and these cellular properties have been related to the expression of large numbers of marker genes. We next sought to identify our transcriptomic clusters with previously defined cell types, taking advantage of the “Rosetta stone” provided by this extensive prior research. Explaining how the identifications were made requires an extensive discussion of the previous literature, which is presented in full as S1 Text, online. Here, we briefly summarize the major subtypes identified (summarized in Fig 4).
Continent 1 was identified with the Sst-positive hippocamposeptal and O-LM cells of stratum oriens (so). These cells all expressed Sst and Grm1, and were further divided into two Npy+/Ngf+ clusters identified as hippocamposeptal neurons [52] and three Pnoc+/Reln+/Npy− clusters identified with O-LM cells [9]. In addition, continent 1 contains a previously undescribed subclass positive for Sst, Npy, and Reln.
Continent 2 was identified as basket and bistratified cells. These were all positive for Tac1 (the precursor to the neuropeptide Substance P), as well as Satb1 and Erbb4, but were negative for Grm1. They were divided into two Pvalb+/Sst− clusters identified with basket cells, two Pvalb+/Sst+/Npy+ clusters identified with bistratified cells (Klausberger et al., 2004), and three Pvalb− clusters identified with Oriens-Bistratified (O-Bi) cells [53].
Continent 3 was identified as axo-axonic cells because of their expression of Pvalb but not Satb1 [54]. This continent’s three clusters were Tac1 negative but positive for other markers, including Snca, Pthlh, and C1ql1, which have also been associated with axo-axonic cells in isocortex [43,44]. We note that this dichotomy of Pvalb interneurons into Tac1-positive and -negative subclasses is likely homologous to previous observations in isocortex [55].
Continent 4 was identified as Ivy cells and medial ganglionic eminence (MGE)-derived neurogliaform cells. These cells expressed Cacna2d1, which we propose as a unique identifier of hippocampal neurogliaform/ivy cells, as well as Lhx6 and Nos1 [56]. They were divided into a Reln+ cluster identified with MGE-derived neurogliaform cells and a Reln−/Vwa5a+ cluster identified with Ivy cells [23]. This continent is homologous to the isocortical Igtp class defined by Tasic and colleagues [44], which we hypothesize may represent isocortical neurogliaform cells of MGE origin; this hypothesis could be confirmed using fate mapping.
Continent 5 was identified as caudal ganglionic eminence (CGE)-derived neurogliaform cells. Its three clusters contained Cacna2d1 and many other genes in common with those of continent 4, but lacked Lhx6 and Nos1 [56]. Similar to isocortical putative neurogliaform cells, this continent expressed Ndnf and contained a distinct subtype positive for Cxcl14 [44]. As with continent 4, continent 5 mainly expressed Reln but also contained a small Reln-negative cluster, which we suggest forms a rare and novel class of CGE-derived ivy cell.
Continent 6 was identified with Sst-negative long-range projection interneurons. It divided into two distinct clusters, both of which were strongly positive for Ntng1. The first strongly expressed Chrm2 but lacked Sst and Pvalb, identifying them as trilaminar cells [12,57]. The second subgroup lacked most classical molecular markers; this fact, together with their inferred laminar location at the stratum radiatum / stratum lacunosum-moleculare (sr/slm) border, identified them as putative radiatum-retrohippocampal neurons that project to the retrosplenial cortex [12,58].
Continents 7 and 8 were identified as what are traditionally called Cck interneurons. This term is somewhat unfortunate: while these cells indeed strongly express Cck, many other inhibitory classes express Cck at lower levels, including even Pvalb+ basket cells [59]. Continents 7 and 8 cells comprised 13 highly diverse clusters but shared strong expression of Cnr1, Sncg, Trp53i11, and several other novel genes. Continent 8 is distinguished by expression of Cxcl14, which localizes these cells to the sr/slm border. This continent comprised a continuum ranging from soma-targeting basket cells, identified by their Slc17a8 (vGlut3) expression, to dendrite-targeting cells, identified by expression of Calb1 or Reln [19,21]. Continent 7, lacking Cxcl14, was identified as Cck cells of other layers and contained multiple subtypes characterized by the familiar markers Calb1, Vip, and Slc17a8 [21] as well novel markers such as Sema5a and Calca. Associated with continent 8 were several apparently novel subtypes: a rare and distinct group positive for both Scl17a8 and Calb1 and marked nearly exclusively by Lypd1; a Ntng1+/Ndnf+ subgroup related to cells of continent 6; and a group strongly expressing both Vip and Cxcl14, which therefore likely corresponds to a novel Vip+/Cck+ interneuron at the sr/slm border.
Continent 9 was identified as I-S interneurons. Its eight clusters fell into three groups: Calb2+/Vip− neurons identified as IS-1 cells, Calb2−/Vip+ neurons identified as IS-2 cells, and Calb2+/Vip+ neurons identified as IS-3 cells [2,24,26,60]. All expressed Penk [61]. These cells contained at least two novel subgroups: an IS-3 subtype positive for Nos1 and Myl1, homologous to the Vip Mybpc2 class defined in isocortex [44], and a rare subclass of IS-1 cells positive for Igfbp6.
Continent 10 contained a single highly distinct cluster located in an “island” off continent 1. It contained cells strongly positive for Sst and Nos1 [27], whose expression pattern is consistent with that of both backprojection cells [13] and PENK-positive projection cells [10], suggesting that these three previously identified classes reflect a single cell type.
Our finding of 49 clusters in a sample of 3,663 CA1 cells contrasts with a previous study of isocortical area V1 (primary visual cortex), which found 23 clusters from a sample of 761 inhibitory neurons [44]. One can imagine three reasons for the greater number of clusters found in the present study: the larger sample size used here may have resulted in our resolving more clusters, the use of a different clustering algorithm may have allowed the current study to reveal finer cell types, or area CA1 might genuinely contain more diverse inhibitory neurons than isocortex. To address these questions, we performed two analyses. First, we applied our clustering algorithm to the data of Tasic and colleagues (2016), and second we reanalyzed subsamples of the data of both the current study and of Tasic and colleagues (2016) to see how the number of clusters found varies with cell count and with sequencing depth.
Applying the ProMMT algorithm to the Tasic dataset yielded 30 clusters (S5A and S5B Fig). The cluster assignments almost completely overlapped as far as top-level groupings but showed some more subtle distinctions in finer-level clusters (S5C Fig). We examined three of these novel classes’ differences in more depth, to ask whether the finer distinctions found by the ProMMT algorithm could correspond to genuine biological cell classes. The most notable of these was cluster 11, which contained neurons that had previously been assigned to the neurogliaform clusters Ndnf Cxcl14, Ndnf Car4, but lacked common neurogliaform markers such as Lamp5 and Gabrd. Instead, cells in these clusters expressed Calb2 and Penk but not Vip, suggesting cells homologous to hippocampal IS-1 cells and potentially matching the Vip-negative I-S layer 1 “single-bouquet cells” (SBCs) described by Jiang and colleagues [62,63]. To test whether cluster 11 indeed corresponds to SBCs, we took advantage of a Patch-seq study [64] that contrasted gene expression in anatomically identified layer 1 SBCs and neurogliaform cells (S5D Fig). We found that the genes that Cadwell and colleagues had reported as distinguishing SBCs from neurogliaform cells indeed occurred in almost entirely nonoverlapping populations of cells; furthermore, these populations closely matched the ProMMT clusters identified with SBCs and neurogliaform cells. Examination of two further subdivisions found by the ProMMT algorithm again revealed genes uniquely expressed in nonoverlapping subpopulations of the Sst Cbln4 and Vip Parm1 clusters (S6 Fig). We conclude that the larger number of clusters identified by the ProMMT algorithm at least in part results from its ability to distinguish subtle variations in gene expression between related cell types.
To ask whether the greater number of clusters found in the current study might in part arise from its larger sample size, we reran the cluster analysis on randomly selected subsets of cells from our dataset. We found a strong linear increase in the number of clusters found with increasing sample size (S7A Fig). To investigate what effect sequencing depth may have had, we resampled our dataset to simulate lower read counts for the same cells, and again found an approximately linear increase in the number of identified clusters with read count (S7B and S7C Fig). We performed similar analyses on Tasic and colleagues’ data and obtained similar results (S7D and S7E Fig).
We therefore conclude that the larger number of clusters found by the current study is more likely to reflect a combination of larger sample size and more sensitive clustering algorithms than a greater number of biological cell types in CA1 than in V1. Furthermore, we expect that an even larger sample size or greater sequencing depth would have revealed yet more, finely distinguished cell types.
Although the major continents of the expression map were clearly separated, clusters within these continents often appeared to blend into each other continuously. This suggests continuous gradation in gene expression patterns: while our probabilistic mixture model will group cells from a common negative binomial distribution into a single cluster, it will tile cells of continuously graded mean expression into multiple clusters.
Although visualization methods such as nbtSNE can suggest whether classes are discrete or continuously separated, they are not sufficient to confirm the suggestion. Such methods exhibit local optima, raising the possibility that apparent continuity only occurs for particular initialization conditions. Furthermore, as nbtSNE is based on a subset of genes, it is conceivable that discrete/continuous patterns occur only for this subset.
To confirm the apparent continuity or discreteness of these groups, we therefore employed a novel method of negative binomial discriminant analysis that is independent of nbtSNE and considers all genes. Given a pair of cell classes, this method compares how close each cell’s whole-genome expression pattern is to each class, using a cross-validated likelihood ratio statistic. For two classes identified as basket and axo-axonic cells, the histogram of likelihood ratios was clearly bimodal (Fig 5A, top), indicating that every cell exhibited a much stronger fit to its own class than to the other, and confirming the discrete separation of these classes. A second example of clusters identified with Ivy and MGE-neurogliaform cells, however, showed different behavior (Fig 5A, middle): a unimodal likelihood ratio histogram indicated that the two clusters ran smoothly into each other, tiling a continuum of gene expression patterns. The bimodality of the likelihood ratio can be captured by a d’ statistic, which for these two examples was 7.2 and 1.5, respectively. Perhaps ironically, the degree to which two neighboring classes are discrete or continuous was itself a continuous variable. For example, Slc17a8-expressing Cxcl14/Cck neurons showed largely continuous overlap with their neighboring Cck/Cxcl14 cells, but with some small indication of bimodality, characterized by a d’ of 3.1 (Fig 5A, bottom). We conclude that while truly discrete cluster separations do exist, the dataset is not fully described as a set of discrete classes, and that many clusters tile continuous dimensions (Fig 5B).
The existence of continuous variation in gene expression suggests that cluster analysis is not giving a complete picture of neuronal gene expression patterns. To further study the biological significance of continuously varying gene expression, we therefore applied a complementary method, latent factor analysis. Cluster analysis can be viewed as an attempt to summarize the expression of all genes using only a single discrete label per cell (the cell’s cluster identity); the value this label takes for each cell is not directly observed but is “latent” and inferred from the data. Latent factor analysis also attempts to predict the expression of all genes using only a single variable (the “latent factor”), but now with a continuous rather than discrete distribution. As with cluster analysis, the latent factor is not directly observed but is inferred for each cell. Latent factor analysis operates without knowledge of cluster identity and therefore requires that the same rules be used to predict gene expression from the latent factor for cells of all types. Clearly, one should expect neither method to precisely predict the expression of all genes from a single variable, but the rules of cellular organization they reveal may provide important biological information.
As expected, latent factor analysis produced a complementary view to cluster analysis (Fig 6A). Knowing a cell’s cluster identity did not suffice to predict its latent factor value, and vice versa. For example, the ranges of latent factor values for cells in the clusters identified with Cck and Pvalb basket cells overlapped. Nevertheless, the range of possible latent factor values was not identical between clusters, and the mean latent factor value of each cluster differed in a manner that had a clear biological interpretation.
The mean latent factor value of each cluster correlated with the axon target location of the corresponding cell type (Fig 6A). The clusters showing largest mean latent factor values were identified with soma-targeting basket cells (both Pvalb and Cck expressing) and with axo-axonic cells. Lower values of the latent factor were found in clusters identified with dendrite-targeting Cck cells and with bistratified, Ivy, and hippocamposeptal cells, which target pyramidal cells’ proximal dendrites [14,23]. Still lower values of the latent factor were found in clusters identified with neurogliaform and O-LM cells, which target pyramidal distal dendrites. The lowest values of all were found in clusters identified with cells synapsing on inhibitory interneurons: the IS cells of continent 9 and the Sst/Penk/Nos1 cells of continent 10, whose local targets are Pvalb cells [10].
While mean values of the latent factor differed between continents, there was also substantial variability within cells of a single continent. For example, a gradient of latent factor values was seen within continent 8 (identified with Cck-positive neurons at the sr/slm border), with larger values in the west smoothly transitioning to smaller values in the east (Fig 6B). Comparison of gene expression patterns in continent 8 to previous work again suggested that this gradient in latent factor values correlates with axon target location. Indeed, immunohistochemistry has demonstrated that CCK-positive cells expressing SLC17A8 (expressed in western continent 8) project to the pyramidal layer [21], while those expressing CALB1 (expressed in the east) target pyramidal cell dendrites [3,18,65]. The cannabinoid receptor Cnr1, which is more strongly expressed in soma-targeting neurons [66,67], was also more strongly expressed in western cells with larger latent factor values.
As expected, the expression levels of many individual genes correlated with the latent factor; furthermore, the directions of these correlations were consistent, even within distantly related cell types. We investigated the relationships of genes to the latent factor by focusing initially on the Pvalb- and Cck-expressing cells of continents 2 and 8 (Fig 6C). Most genes correlated similarly with the latent factor in both classes. For example, the Na+/K+ pump Atp1b1 and the GABA synthesis enzyme Gad1 correlated positively with the latent factor for multiple cell types, while 6330403K07Rik, a gene of unknown function, correlated negatively. Some genes’ expression levels depended on both cell type and latent factor value. For example, the ion channel Kcnc1 (which enables rapid action potential repolarization in fast-spiking cells) correlated positively with the latent factor in both Pvalb and Cck cells, but its expression was stronger in Pvalb cells, even for the same latent factor value. Other genes showed correlations with the latent factor, but only within the specific classes that expressed them. For example, expression of Pvalb correlated with the latent factor within cells of continent 2, but the gene was essentially absent from cells of continent 8; conversely, Cnr1 expression correlated with the latent factor in continent 8 but was essentially absent in cells of continent 2. Thus, the latent factor value is not alone sufficient to predict a cell’s gene expression pattern but provides a summary of continuous gradation in the expression of multiple genes in multiple cell types.
The relationship of genes to latent factor values was statistically similar across cell types. To demonstrate this, we computed the Spearman correlation of each gene’s expression level with the latent factor, separately, within cells of each continent (S1 Table). As expected from the scatterplots (Fig 6C), the correlation coefficients for Atp1b1, Gad1, and 6330403K07Rik were similar between continents 2 and 8 (Fig 6D). Also as expected, Pvalb and Cnr1 showed strong positive correlations with the latent factor within the continent where these genes were expressed, but correlations close to zero within the continent where they were barely expressed. In general, the correlation coefficients of genes with the latent factor were preserved between continents 2 and 8 (Fig 6D; Spearman rank correlation ρ = 0.58, p < 10−100). A similar relationship was found across all continents (Fig 6D, inset; p < 10−100 in each case), although cells of continents 9 and 10 showed less similarity than continents 1–8. Furthermore, similar results were obtained when analyzing isocortical data, most notably in isocortical Pvalb cells (S8 Fig).
In summary, the expression of many genes correlates with a single continuous variable, the latent factor value assigned to each cell. While this latent factor does not provide a complete summary of a cell’s gene expression pattern, the direction and strength of the correlation of individual genes to the latent factor is largely preserved across cell types. Furthermore, while a cell’s latent factor value was not simply a function of its cell class, mean latent factor values differed between clusters, being largest for clusters identified with cell types whose axons target pyramidal somata or axon initial segments and smallest for clusters identified with cell types targeting pyramidal distal dendrites or interneurons.
The above results suggest that the expression of a large set of genes is modulated in a largely consistent way across multiple cell types in a manner that correlate with their axonal targets. What biological functions might these genes serve? While one might certainly expect structural genes be differentially expressed between soma- and dendrite-targeting interneurons, these cells also differ in their physiology. Indeed, Pvalb-expressing basket cells are known for their fast-spiking phenotype, which produces rapid, powerful perisomatic inhibition and is mediated by a set of rapidly acting ion channels and synaptic proteins, including Kcnc1, Kcna1, Scn1a, Scn8a, and Syt2 [8]. Although most other interneurons show regular-spiking phenotype, CCK-expressing basket cells with a fast-spiking phenotype have also been reported [18,20]. We therefore hypothesized that genes responsible for the fast-spiking phenotype might be positively correlated with the latent factor, because of increased expression in soma-targeting cells of all classes.
Consistent with this hypothesis, genes associated with fast-spiking phenotype (Kcnc1, Kcna1, Scn1a, Scn8a, Syt2) were amongst the genes most positively correlated with the latent factor in both Pvalb and Cck basket cells (Fig 6D). However, this positive correlation was not restricted to these cell types: in an ordering of the correlations of all genes with the latent factor (taking into account cells of all types), these genes ranked in the 99.9th, 98.3rd, 99.5th, 98.9th, and 95th percentiles, respectively (S1 Table).
Other gene families positively correlated with the latent factor included genes associated with mitochondria (e.g., mt-Cytb), ion exchange and metabolism (e.g., Atp1b1; Slc24a2), GABA synthesis and transport (e.g., Gad1, Slc6a1), vesicular release (e.g., Syp, Sv2a, Cplx2, Vamp1), and fast ionotropic glutamate and GABA receptors (e.g., Gria1, Gabra1) as well as GABAB receptors (e.g., Gabbr1, Gabbr2, Kcnj3, Kctd12). The genes correlating negatively with the latent factor were less familiar but included Atp1b2, a second isoform of the Na+/K+ pump; Fxyd6, which modulates its activity; Nrsn1, whose translation is suppressed after learning [68], as well as many neuropeptides (e.g., Sst, Vip, Cartpt, Tac2, Penk, Crh; exceptional neuropeptides such as Cck showed positive correlation). Genes associated with neurofilaments and intermediate filaments (e.g., Nefh, Nefl, Krt73) tended to show positive weights, while genes associated with actin processing (e.g., Gap43, Stmn1, Tmsb10) tended to show negative weights. Many other genes of as yet unknown function correlated positively and negatively with the latent factor (for example, 6330403K07Rik). Relating the latent factor correlations of each gene to their gene ontology (GO) annotations (which are not granular enough to list annotations such as fast-spiking physiology) suggested that negatively correlated genes tended to be associated with translation and ribosomes, while positively correlated genes were associated with diverse functions, including transcription, signal transduction, ion transport, and vesicular function and with cellular compartments, including mitochondria, axons, and dendrites (S2 Table).
We therefore suggest that cells with large values of the latent factor not only target more proximal components of pyramidal cells but also express genes enabling a faster spiking firing pattern, more synaptic vesicles, and larger amounts of GABA release; receipt of stronger excitatory and inhibitory inputs; and faster metabolism. These are all characteristics of Pvalb-expressing fast-spiking interneurons [8]; however, a similar continuum was observed within all cell types, suggesting that these genes are commonly regulated in all CA1 interneurons.
The fact that the latent factor differs systematically between cells with different axonal targets suggests that this property is in good measure fixed, as it seems unlikely that neurons would make major changes to their axonal targets in adulthood. Nevertheless, interneuronal gene expression can be modulated by activity, and some of the genes that were most strongly correlated with the latent factor (Pvalb, Kcna1) are amongst those with activity-dependent modulation [31–34,69].
To investigate whether the genes correlated with the latent factor might also be partially modulated by neuronal activity, we correlated each gene’s latent factor score with that gene’s modulation by in vivo light exposure after dark housing, using data from three classes of visual cortical interneurons (made available by Mardinly and colleagues [33]). We observed a moderate relationship of latent factor weighting to activity modulation in Sst neurons (r = 0.26; p < 10−12; S9 Fig), suggesting that activity-dependent modulation of Sst cells may cause them to move along the continuum of latent factor values. A weaker but still significant correlation was observed for Pvalb neurons (r = 0.11; p < 0.002), whereas no significant relationship was found for Vip neurons (p = 0.17). These data therefore suggest that a portion of the continuous variability of gene expression observed in CA1 interneurons may arise from activity-dependent modulation but that such modulation is unlikely to be a full explanation for the genetic continua revealed by latent factor analysis.
The transcriptomic classification we derived makes a large number of predictions for the combinatorial expression patterns of familiar and novel molecular markers in distinct CA1 interneuron types. To verify our transcriptomic classification, we set out to test some of these predictions using traditional methods of molecular histology.
Our first tests regarded the very distinct Sst.Nos1 cluster of continent 10. This cluster’s expression pattern matched three previously reported rare hippocampal inhibitory cell types: large SST-immunopositive cells that are intensely immunoreactive for NOS1 throughout the cytoplasm, revealing their full dendrites [27]; PENK-positive projection cells [10]; and strongly NADPH diaphorase-labeled (i.e., NOS1-positive) backprojection cells [13]. We therefore hypothesized that these cell types, previously regarded as separate, may in fact be identical. To test this hypothesis, we performed a series of triple and quadruple immunoreactions, focusing on the intensely NOS1-positive neurons (n = 3 mice, n = 70 cells: 39% in so/alveus; 10% in stratum pyramidale (sp); 27% in sr; 24% at the sr/slm border). Similar to previously reported PENK-projection, backprojection, and SST/NOS1 cells [10,13,27]—but unlike SST-positive O-LM cells [9]—these neurons all showed spiny or sparsely spiny dendrites. As expected from the Sst.Nos1 cluster, we found that they were all SST/NPY double positive (n = 20/20) and were virtually all weakly positive for CHRM2 (n = 36/38) and GRM1 (n = 17/17) in the somatodendritic plasma membrane, strongly positive for PCP4 (n = 19/21) in the cytoplasm and nucleus, and strongly positive for PENK (n = 35/42) in the Golgi apparatus and granules in the soma and proximal dendrites (Fig 7). By contrast, the more numerous moderately NOS1-positive cells (which include many interneuron types such as ivy, MGE-neurogliaform, and a subset of IS-3 neurons) were mostly immunonegative for CHRM2, PCP4, and PENK, although some were positive for GRM1. Our results are therefore consistent with the hypothesis that all three previously reported classes correspond to the Sst1.Nos1 cluster.
A second prediction of our classification was the expression of Npy in multiple subclasses of Cck cell, most notably the Slc17a8- and Calb1-expressing clusters of continent 8. This was unexpected, as NPY (at least at the protein level) has instead been traditionally associated with SST-expressing neurons and ivy/neurogliaform cells (Fuentealba et al., 2008a; Katona et al., 2014). Nevertheless, no studies to our knowledge have yet examined immunohistochemically whether the neuropeptides NPY and CCK can be colocalized in the same interneurons. We therefore tested this by double immunohistochemistry in sr and slm (Fig 8A, n = 3 mice). Consistent with our predictions, 119 out of 162 (74% ± 6%) of the cells immunopositive for pro-CCK were also positive for NPY (an additional 73 cells were positive for NPY only, which, according to our identifications, should represent neurogliaform and radiatum-retrohippocampal cells). A subset (176 cells) of NPY and/or pro-CCK immunopositive neurons were further tested for CALB1 in triple immunoreactions. As expected, nearly all CALB1-positive neurons were pro-CCK positive (89% ± 2%), and CALB1 immunoreactivity was seen in a subset of the cells containing both pro-CCK and NPY (27% ± 3%). Additional triple immunohistochemistry for NPY, pro-CCK, and SLC17A8 (VGLUT3) revealed triple positive cells in sr and particularly at the sr/slm border, as predicted by the class Cck.Cxcl14.Slc17a8 (Fig 8B). Because of the low level of somatic immunoreactivity for SLC17A8 (which, as a vesicular transporter, is primarily trafficked to axon terminals), we could not count these cells reliably; however, of the cells that were unambiguously immunopositive for SLC17A8, in a majority we detected NPY. Additional analysis combining double in situ hybridization for Slc17a8 and Npy with immunohistochemistry for pro-CCK (Fig 8C, n = 3 mice) confirmed that the great majority of Slc17a8-expressing cells were also positive for Npy and pro-CCK (84% ± 3%). As predicted by our identifications, the converse was not true: a substantial population of Npy/pro-CCK double-positive cells (57% ± 7% of the total) did not show detectable Slc17a8, which we identify with dendrite-targeting neurons in the east of continent 8.
Several cell types in our classification expressed Cxcl14, a gene whose expression pattern in the Allen Brain Atlas shows localization largely at the sr/slm border. The Cxcl14-positive population includes all clusters of continent 8, which express Cck and contain subclusters expressing Npy, Calb1, Reln, and Vip; a subtype of CGE-derived neurogliaform cell that expresses Reln and Npy but lacks Nos1 and expresses Kit at most weakly; as well as IS-1, IS-2, and radiatum-retrohippocampal cells. However, as all Cxcl14-positive clusters lacked Lhx6, we conclude they should be distinct from all MGE-derived neurons, including MGE-derived neurogliaform cells.
To test these predictions, we performed in situ hybridization for Cxcl14 simultaneously with in situ hybridization or immunohistochemistry to detect Reln, Npy, CALB1, CCK, PVALB, Sst, Nos1, and Kit (n = 3 mice; Fig 9). In addition, we combined fluorescent in situ hybridization for Cxcl14 with immunohistochemistry for yellow fluorescent protein (YFP) in Lhx6-Cre/R26R-YFP mice, which allows identification of developmental origin by marking MGE-derived interneurons (Fogarty et al., 2007). The results of these experiments were consistent with our hypotheses. We found that within CA1, Cxcl14-expressing cells were primarily located at the sr/slm border (71% ± 3%), although a subpopulation of cells were also found in other layers. We found no overlap of Cxcl14 with YFP in the Lhx6-Cre/R26R-YFP mouse, confirming the CGE origin of Cxcl14-expressing neurons (Fig 9A). The majority of Cxcl14-positive cells expressed Reln (72% ± 4%), accounting for 42% ± 9% of Reln-expressing neurons (substantial populations of Reln+/Cxcl14− cells located in so and slm likely represent O-LM and MGE-neurogliaform cells, respectively (Fig 9B). Indeed, although less than half of Reln cells were located at the R-LM border (44% ± 1%), the great majority of Reln+/Cxcl14+ cells were found there (88% ± 6%). Consistent with the expected properties of continent 8 cells, a large fraction of the Cxcl14 population were immunoreactive for pro-CCK (62% ± 6%; Fig 9C), while substantial minorities were positive for CALB1 (29% ± 2%; Fig 9D) or Npy (25% ± 5%; Fig 9E). However, as expected from the lack of Cxcl14 in MGE-derived neurogliaform and IS-3 cells, we observed no overlap of Cxcl14 with Nos1 (0 out of 209 cells; Fig 9F) and very weak overlap with Kit, which is primarily expressed in clusters Cacna2d1.Ndnf.Npy and Cacna2d1.Ndnf.Rgs10, associated with the Cxcl14-negative CGE-neurogliaform population (1 of 264 cells, respectively, from all mice; Fig 9G).
The cluster Cck.Cxcl14.Vip presented a puzzle, because Cxcl14 is located primarily at the sr/slm border, whereas immunohistochemistry in rat has localized CCK/VIP basket cells to sp [24]. Because Cxcl14 expression can sometimes also be found in sp, we tested whether this cluster reflects sp cells by combining in situ hybridization for Cxcl14 with immunohistochemistry against VIP in mouse CA1 (Fig 10). This revealed frequent co-expression at the sr/slm border (8% ± 1% Cxcl14 cells positive for Vip; 23% ± 1% Vip cells positive for Cxcl14) but very few Cxcl14 cells in sp, and essentially no double labeling (1 of 147 Vip cells in sp was weakly labeled for Cxcl14). We therefore conclude that this cluster indeed represents a novel cell type located at the sr/slm border, expressing Cck, Vip, and Cxcl14.
The molecular architecture of CA1 interneurons has been intensively studied over the last decades, leading to the identification of 23 inhibitory classes. Our transcriptomic data showed a remarkable correspondence to this previous work, with all previously described classes identified in our database. Our analysis also revealed a continuous mode of variability common across multiple cell types, eight hypothesized novel classes, as well as additional molecular subdivisions of previously described cell types.
Surprisingly, these data suggest that three previously described CA1 cell groups in fact represent a single cell class, a fact previously overlooked because of the limited combinations of molecules tested in prior work. The Sst.Nos1 class is strongly positive for Nos1 and also expresses Sst, Npy, Chrm2, Pcp4, and Penk, but unlike Penk-positive I-S cells of continent 9, it lacks Vip. This class is homologous to the “Int1” and “Sst Chodl” classes defined in isocortex, which have been identified with long-range projecting sleep-active neurons [44,46,70,71]. The three previously described classes identified with Sst.Nos1 are PENK-immunopositive neurons with projections to subiculum, which were shown to be VIP negative, but not tested for SST or NOS1 (Fuentealba et al., 2008b); the NADPH diaphorase-labeled (i.e., strongly NOS1-positive) axons reported by Sik and colleagues (1994) as projecting to CA3 and dentate, but not tested for SST or PENK; and the SST/NOS1 cells identified by Jinno and Kosaka (2004) in mouse, which were not tested for long-range projections or for PENK. While it remains possible that a larger transcriptomic sample of these rare neurons would reveal subclasses, our present data suggest that Sst.Nos1 cells are a homogeneous population: the nbtSNE algorithm, BIC criterion, and further manual exploration failed to reveal any finer distinctions. We therefore suggest that they constitute a class of inhibitory neurons with diverse long-range projection targets. Interestingly, the targets of PENK-positive projection cells are most commonly PVALB-positive interneurons, unlike conventional IS cells, which preferentially target SST cells [10]. As these cells are identified as sleep active, this fact may provide an important clue to the mechanisms underlying sleep in cortical circuits.
The match between our transcriptomic analysis and previous immunohistochemical work (primarily in rat) is so close that it is simpler to describe the few areas of disagreement than the many areas of agreement. First, ACTN2 has been used as a neurogliaform marker in rat [72] but was almost completely absent from any cell type of our database. We suggest this reflects a species difference, as previous attempts with multiple ACTN2 antibodies have been unsuccessful in mouse (J. H.-L., unpublished observations), and Actn2 labeling is not detectable in the Allen atlas [73]. Second, we observed Calb2 in a subset of putative O-LM cells; these Calb2-expressing neurons typically also expressed Calb1. Such O-LM cells have not been described in rat [9], but CALB2/SST neurons have been observed in mouse isocortex [44,74]. A third inconsistency regards NCALD, which in rat was reported not to overlap with PVALB, SST, or NPY [75], but did so in our data. Finally, it has previously been reported that a subset of O-LM cells show Htr3a expression [76]. In our data, we observed at most weak expression of Htr3a in Sst cells, and the cells showing it belonged to clusters identified as hippocamposeptal rather than O-LM cells.
Our analysis revealed several rare and presumably novel cell groups, although we cannot exclude that some of these were inadvertently included from neighboring areas such as subiculum (S10 Fig). Sst.Npy.Serpine2 and Sst.Npy.Mgat4c, which simultaneously expressed Sst, Npy, and Reln, fit the expected expression pattern of neither O-LM nor hippocamposeptal cells; Sst.Erbb4.Rgs10 is a distinct group related to Pvalb basket and bistratified cells; Cck.Lypd1 formed a rare and highly distinct class expressing Cck, Slc17a8, and Calb1; Ntng1.Synpr showed an expression pattern with features of both sr/slm Cck neurons and projection cells; and Cck.Cxcl14.Vip represents a cell class strongly positive for both Cck and Vip located at the sr/slm border that appears to be a pyramidal- rather than interneuron-targeting class. The analysis also revealed subdivisions of known types, such as the division of IS-3 cells into Nos1-positive and -negative groups, and the division of CGE-NGF cells into Car4- and Cxcl14-expressing subtypes. Finally, our data suggested that with more cells or deeper sequencing, even rarer types are likely to be found, as subsetting analysis showed a linear increase in the number of clusters with cell count and read depth, with little sign of saturation as yet. The data appeared to contain several novel cell types not containing enough cells to overcome the algorithm’s parsimony penalty, such as a small group of cells with features of both basket and axo-axonic cells located off the coast of continent 3; such cells have indeed been rarely encountered by quantitative electron microscopic analysis of synaptic targets in the rat (P. S., unpublished observations).
Latent factor analysis revealed a common continuum of gene expression across the database, suggesting a large “module” of genes that are coregulated in multiple types of hippocampal interneuron. The latent factor differed between clusters, and clusters with larger latent factor values were identified with interneuron types targeting pyramidal cell somas or proximal dendrites (such as Pvalb or Cck/Slc17a8 expressing basket cells), while those with low mean values were identified with interneurons targeting pyramidal distal dendrites (such as Sst or Cck/Calb1 expressing dendrite-targeting cells) or targeting other interneurons. Subtler differences in latent factor were found within clusters, suggesting that a similar continuum exists within cells of a single type. Genes positively correlated with the latent factor are associated with fast-spiking phenotype, presynaptic function, GABA release, and metabolism. Consistent with this expression pattern, perisomatic inhibitory cells show fast-spiking phenotypes and deliver powerful, accurately timed inhibition [8], but interneurons targeting distal dendrites show slower-spiking patterns; presumably because distal inputs are subject to passive dendritic filtering, their presynaptic vesicle release does not need to be so accurately timed. I-S cells had the lowest mean values of the latent factor, consistent with their small axonal trees and metabolic machinery [77]. The stronger expression of many neuropeptides in cells of low latent factor suggests that these slower, distal-targeting interneurons may also rely more heavily on neuropeptide signaling, for which slow firing rates support outputs transduced by slower G-protein–coupled receptors. Interestingly, a study conducted independently of the present work identified enriched expression of a gene module similar to our latent factor in isocortical Pvalb neurons [43] and suggested it is controlled by the transcription factor PGC-1α [78,79]. Our results suggest that Cck-expressing basket cells have a similar expression pattern and that, more generally, expression of this module correlates with a neuron’s axonal target location.
Several novel genes correlating with the factor appear to be interesting candidates for future research, such as Trp53i11, Yjefn3, and Rgs10, associated with faster-spiking Cck cells; Zcchc12 and 6330403K07Rik, both associated with slower-firing cells of all classes; and Fxyd6, associated with slow spiking, which may modulate ion exchange. Intriguingly, genes for neurofilaments and other intermediate filaments were positively correlated with the latent factor, while genes involved in actin processing were negatively correlated; we hypothesize that this might reflect a different cytoskeletal organization required for somatic- and dendritic-targeting neurons.
The question of how many cell classes a given neural circuit contains is often asked of transcriptomic analyses, but we argue this question will not have a clearly defined answer. For example, our data indicate no sharp dividing line between ivy cells and MGE-derived neurogliaform cells. Yet cells at the two ends of the continuum are clearly different: not only do their gene expression patterns differ substantially, but their different axonal targets indicate different roles in circuit function [23]. In statistics, multiple criteria can be used to define how many clusters should be assigned to a dataset; a common approach (which is used by the ProMMT algorithm) is to consider a cluster indivisible if within-cluster fluctuations cannot be distinguished from random noise. Using this criterion, the number of clusters of CA1 interneurons increased with the number of cells and read depth analyzed, showing no sign of saturation in the current dataset. Furthermore, we observed several apparent rare classes that were too small to be assigned their own clusters at present, together with further subtle gradations within currently assigned clusters. The fact that we observed more clusters in CA1 than the 23 previously identified in isocortex [44] should therefore not be taken as implying that CA1 is a more complex circuit but simply that our larger sample size and different clustering algorithm were able to detect finer distinctions. Indeed, our data suggest that while the divisions between the 10 major “continents” are unambiguous, the organization of gene expression within these continents is complex and subtle, and likely far more detailed than characterized by our present 49 clusters. An understanding of this multiscale variability in gene expression in CA1 interneurons will be a key tool to understanding the function of this circuit.
Six adult (20 wk old) male C57BL/6J mice (Charles River, Oxford, UK) were perfusion fixed following anesthesia and tissue preparation for immunofluorescence (Katona et al., 2014) and analysis using wide-field epifluorescence microscopy [21] was performed as described. The following primary antibodies were used: anti-calbindin (goat, Fronteir Inst, Af104); anti-pro-CCK (rabbit, 1:2,000, Somogyi et al., 2004); anti-metabotropic glutamate receptor 1a (GRM1, rabbit, 1:1,000; guinea pig, 1:500; gifts from Prof. M. Watanabe, Frontier Institute); anti-muscarinic acetylcholine receptor 2 (CHRM2, rat, 1:400, EMD Millipore Corporation, MAB367); anti-NOS1 (rabbit, 1:1,000, EMD Millipore Corporation, AB5380; mouse, 1:1,000, Sigma-Aldrich, N2280); anti-NPY (mouse, 1:5,000, Abcam, #ab112473); anti-Purkinje cell protein 4 (PCP4, rabbit, 1:500, Santa Cruz Biotechnology, sc-74816); anti-pre-pro-enkephalin (PENK, guinea pig, 1:1,000, gift from Takahiro Furuta, Kyoto University, Japan; rabbit, 1:5,000, LifeSpan Biosciences, LS-C23084); anti-SST (sheep, 1:500, Fitzgerald Industries International, CR2056SP); anti-VGLUT3 (guinea pig, Somogyi et al 2004). Secondary antibodies were raised in donkey against immunoglobulin G of the species of origin of the primary antibodies and conjugated to Violet 421 (1:250); DyLight405 (1:250); Alexa 488 (1:1,000); cyanine 3 (1:400); Alexa 647 (1:250); cyanine 5 (Cy5, 1:250). With the exception of donkey‐antimouse‐Alexa488 purchased from Invitrogen, all secondary antibodies were purchased from Stratech.
For cell counting, image stacks (212 × 212 μm area; 512 × 512 pixels; z stack height on average 12 μm) were acquired using LSM 710/AxioImager.Z1 (Carl Zeiss) laser scanning confocal microscope equipped with Plan-Apochromat 40×/1.3 Oil DIC M27 objective and controlled using ZEN (2008 v5.0 Black, Carl Zeiss). In a second set of sections, images were taken using Leitz DM RB (Leica) epifluorescence microscope equipped with PL Fluotar 40×/0.5 objective. Counting was performed either using ImageJ (v1.50b, Cell Counter plugin) on the confocal image stacks or OPENLAB software for the epifluorescence documentation. For the CCK counts, numbers were pooled from two separate reactions testing for a given combination of primary antibodies (n = 3 mice each reaction, 2–3 sections each mouse) and reported as average values ± standard deviation. For the testing of intensely nNOS-positive neurons, cells were selected using Leitz DM RB (Leica) epifluorescence microscope equipped with PL Fluotar 40×/0.5 objective. Cells were pooled from three separate reactions testing for a given combination of primary antibodies (n = 3 mice each reaction, 2 sections each mouse) and reported as pooled data. Image processing was performed using ZEN (2012 Blue, Carl Zeiss), ImageJ (v1.51m, open source), Inkscape (0.92, open source), and Photoshop (CS5, Adobe).
Wild-type (C57BL/6/CBA) male and female adult (p30) mice and Lhx6-CreTg transgenic mice were perfusion-fixed, as previously described (Rubin et al., 2010), followed by immersion fixation overnight in 4% paraformaldehyde. Fixed samples were cryoprotected by overnight immersion in 20% sucrose, embedded in optimal cutting temperature (OCT) compound (Tissue Tek, Raymond Lamb Ltd Medical Supplies, Eastbourne, UK), and frozen on dry ice. 30 μm cryosections were collected in DEPC-treated PBS and double in situ hybridization was carried out as described (Rubin et al., 2010). Probes used included either a Cxcl14-(digoxgenin)DIG RNA probe in combination with Reln-(fluorescein)FITC; Npy-FITC, Sst-FITC, or Vip-FITC probes; or a Cxcl14-FITC probe with Nos1-DIG, Kit-DIG, Scl17a8-DIG, or Pvalb-DIG probes. DIG-labeled probes were detected with an anti-DIG-alkaline phosphatase (AP)-conjugated antibody followed by application of a Fast Red (Sigma) substrate. The first reaction was stopped by washing 3 × 10 min in PBS, and the sections were incubated with an anti-FITC-Peroxidase (POD)-conjugated antibody (1:1,500—Roche) overnight. The POD signal was developed by incubating the sections with Tyramide-FITC:amplification buffer (1:100, TSA-Plus, Perkin Elmer) for 10 min at room temperature. For immunohistochemistry after in situ hybridization, the following antibodies were used: anti-Calbindin (rabbit, 1:1,000, Swant, Bellinzona, Switzerland); anti-pro-CCK (rabbit, 1:2,000, Somogyi et al., 2004); anti-GFP (chicken, 1:500, Aves Labs). All sections were counterstained with Hoechst 33258 dye (Sigma, 1,000-fold dilution) and mounted with Dako Fluorescence Mounting Medium (DAKO).
For cell counts, images (at least two sections per mouse) were acquired on an epifluorescence microscope (Zeiss) with a 10× objective. Several images spanning the entire hippocampal CA1 were stitched using Microsoft Image Composite Editor. Cells were counted manually in the CA1 area, including sr and slm, and in a subregion spanning 100 μm across the border between sr and slm, where most Cxcl14-positive cells are located. Confocal images (z stack height on average 25 μm, 2 μm spacing) were taken on a Leica confocal microscope under a 10× objective and processed for contrast and brightness enhancement with Photoshop (CS5, Adobe). A final composite was generated in Adobe Illustrator (CS5, Adobe).
Code for cluster analysis and all other algorithms can be found at https://github.com/cortex-lab/Transcriptomics.
To measure how well separated each cluster is from its neighbors, we define an isolation metric equal to the deletion loss (described in the previous section), divided by Nklog(2), where Nk is the number of cells assigned to cluster k. This has an information-theoretic interpretation, as the number of additional bits that would be required to communicate the gene expression pattern of a cell in cluster k, using a code defined by the probability model if cluster k were deleted.
Each cluster produced by the EM algorithm is specified by a mean expression vector. To understand the relationship between these cluster means, we applied a clustering method to the clusters themselves. This was achieved using Ward’s method, with a distance matrix given by the K-L divergence between cluster means, weighted by the number of cells per cluster.
To visualize the locations of the cells, we derived a variant of the tSNE algorithm [51] appropriate for data following a negative binomial distribution.
Stochastic neighbor embedding algorithms such as tSNE start by converting Euclidean distances between pairs of high-dimensional vectors xi into conditional probabilities according to a Gaussian distribution: pj|i=N(xj;xi,σi2)/∑k≠iN(xk;xi,σi2). The tSNE algorithm then adjusts the locations of low-dimensional representation yi in order to minimize the K-L divergence of a symmetrized pj|i, with a t-distribution on the yi.
The Gaussian distribution, however, is not the most appropriate choice for transcriptomic data. We found that we obtained better results using the same negative binomial distribution as in the ProMMT algorithm:
pj|i=NB(xj;xi,r)/∑k≠iNB(xk;xi,r)
where
NB(xj;xi,r)=exp[∑g∈Sxgjlog(xgixgi+r)+rlog(rxgi+r)]
excluding a binomial coefficient that cancels when computing pj|i. The sum runs over the set of genes g that were chosen by the ProMMT algorithm.
In the original tSNE algorithm, variations in distance between the points xi are overcome by adjusting the variance σi2 for each point i to achieve constant perplexity of the symmetrized conditional distributions. We took the same approach, finding a scale factor λi for each cell i to ensure that the scaled symmetrized distribution
pji=expλi(log(pj|i)+log(pi|j))
had a fixed perplexity of 15. This computation and the implementation K-L minimization was achieved using Laurens van der Maaten’s drtoolbox (https://lvdmaaten.github.io/drtoolbox/). The algorithm was initialized by placing all points on a unit circle, with angular position determined by their parent cluster, linearly ordered by the hierarchical cluster clustering.
For comparison, we ran four other methods of tSNE analysis (S4 Fig) using either all genes or the 150 genes found by ProMMT, and either a Euclidean metric or a Euclidean metric after log(1+x) transformation. Perplexity of 15 was again used and initialization was the same as before. Using all genes gave results that were difficult to interpret, particularly for log-transformed data, which we ascribe the noise arising from the large number of weakly expressed genes in the database. Using the gene subset provided more interpretable results, and combining the gene subset with log(1+x) transformation yielded results similar to nbtSNE, while a Euclidean metric yielded a less clear distinction of isolated classes such as Cck.Lypd1 and Sst.Nos1. We conclude that the alignment of nbtSNE to the probability distribution of RNA counts allows the algorithm to take into account differences between weakly expressed genes, and that a log(1+x) transformation approximates this probability distribution. We also conclude that gene subsetting prevents noise from the large number of genes that do not differ between classes from swamping the signal, and that this is particularly important with algorithms sensitive to changes in weakly expressed genes. We suggest that nbtSNE provides a principled probabilistic method for choosing the transformation and gene subset required for informative tSNE analysis.
To investigate whether a pair of clusters were discretely separated or tiled a continuum we developed a method of cross-validated negative binomial discriminant analysis. This analysis assesses the separation of two clusters k1 and k2 by computing the log likelihood ratio for each cell to belong to the two clusters. It is simple to show that this ratio for a cell c is given by
Δc=∑g∈Gxc,g(logpg,k1-logpg,k2)+r(log(1-pg,k1)-log(1-pg,k2))
The sum runs over all genes g in the database, not just the set S found by the ProMMT algorithm.
The degree to which clusters k1 and k2 are discrete is visible by the bimodality of the histogram of Δc, which can be quantified using a d’ statistic, μ1-μ2(σ12+σ22)/2, where μi and σi represent the mean and standard deviation of Δc for cells arising in cluster i. In this analysis, it is essential that the ratios Δc are computed on a separate “test set” of cells to the “training set” used to estimate pg,k, otherwise even a random division of a single homogeneous cluster would give an apparently bimodal histogram because of overfitting.
To model continuous variation between cells, we used a negative binomial latent factor model. The model is parametrized by two matrices, W and F of size Ngenes × Nfactors and Nfactors × Ncells. The distribution of each cell follows a negative binomial distribution with mean μgc = r exp(∑f Wgf Ffc):
Pr(xgc;W,F)=NB(xgc;rexp(∑fWgfFfc),r)
This corresponds to the natural parameterization of the negative binomial, p = 1 / (1 + exp(∑f Wgf Ffc)). As usual, we take a fixed value of r = 2. For the analysis described in this study, we use only a single latent factor, but add a second column to F of all ones to allow the mean expression level to vary between genes.
Given a dataset xgc, we fit the matrices W and F by maximum likelihood. As the negative binomial distribution with fixed r belongs to the exponential family, we can use the simple alternating method of Collins and colleagues [83]. Note that we do not require a sparse algorithm because (unlike in clustering), the number of parameters is fixed. However, to avoid instability, only genes that have reasonable expression levels in the database are kept (genes are included if at least 10 cells express at least 5 copies of the RNA), and a quadratic regularization penalty −50[|W|2 + |F|2] is added to the log likelihood.
To relate the correlations of each gene with the latent factor to their GO categories (S2 Table), we used the MGI mouse GO database (downloaded 2 April 2018), accessed via MATLAB’s bioinformatics toolbox. An enrichment score was computed for each GO term by summing the Spearman rank correlations of gene expression with the latent factor, over all genes annotated with that term.
|
10.1371/journal.pntd.0001618 | Untargeted Metabolomics Reveals a Lack Of Synergy between Nifurtimox and Eflornithine against Trypanosoma brucei | A non-targeted metabolomics-based approach is presented that enables the study of pathways in response to drug action with the aim of defining the mode of action of trypanocides. Eflornithine, a polyamine pathway inhibitor, and nifurtimox, whose mode of action involves its metabolic activation, are currently used in combination as first line treatment against stage 2, CNS-involved, human African trypanosomiasis (HAT). Drug action was assessed using an LC-MS based non-targeted metabolomics approach. Eflornithine revealed the expected changes to the polyamine pathway as well as several unexpected changes that point to pathways and metabolites not previously described in bloodstream form trypanosomes, including a lack of arginase activity and N-acetylated ornithine and putrescine. Nifurtimox was shown to be converted to a trinitrile metabolite indicative of metabolic activation, as well as inducing changes in levels of metabolites involved in carbohydrate and nucleotide metabolism. However, eflornithine and nifurtimox failed to synergise anti-trypanosomal activity in vitro, and the metabolomic changes associated with the combination are the sum of those found in each monotherapy with no indication of additional effects. The study reveals how untargeted metabolomics can yield rapid information on drug targets that could be adapted to any pharmacological situation.
| Understanding drug mode of action is of fundamental importance. Of the five drugs in use against human African trypanosomiasis (HAT), convincing evidence on a specific mode of action has been proposed only for the polyamine pathway inhibitor eflornithine. Eflornithine is currently used with nifurtimox as first line treatment of stage 2 CNS-involved HAT. Here, we present a new way of determining the mode of action of a drug by measuring how the levels of small molecules comprising the cellular metabolome are perturbed when exposed to drugs. We show that eflornithine causes the changes in polyamine metabolism previously known to underlie its mode of action. Furthermore, we show that nifurtimox, is rapidly metabolised and significantly alters metabolism. Nifurtimox and eflornithine do not show the predicted synergy with regard to trypanocidal activity and this is reflected in metabolomic analysis where perturbations to the metabolome are additive with no additional changes observed in the combination.
| Human African trypanosomiasis (HAT) is a parasitic infection in sub-Saharan Africa transmitted by tsetse flies. Its causative agent is the flagellated protozoan Trypanosoma brucei, with two sub-species, T. b. gambiense and T. b. rhodesiense responsible for human disease [1], [2].
There are five drugs in use against HAT. Of these five, only eflornithine has a confirmed mode of action (MOA), namely, inhibition of ornithine decarboxylase (ODC) [3], [4] with concomitant perturbation of the polyamine pathway. In addition to the four licensed drugs, nifurtimox has been recommended by the World Health Organisation for use against late-stage disease in combination with eflornithine [5]. The MOA for nifurtimox has, however, yet to be fully elucidated. For many years it was presumed to exert its action through the generation of oxidative stress associated with reduction of the nitro group with subsequent reduction of oxygen to toxic reactive oxygen species [6], [7]. In trypanosomes polyamines serve an unusual role in combining with glutathione to create the metabolite trypanothione [8], which carries out many of the cellular roles usually attributed to glutathione in other cell types, including protection against oxidative stress. This indicated that eflornithine, which inhibits polyamine biosynthesis [9], [10] and subsequently trypanothione biosynthesis, would synergise with nifurtimox as result of a reduced ability of cells to deal with oxidative stress. However, the data that lead to the conclusion that nifurtimox causes oxidative stress is inconclusive [7] and recent evidence shows that nifurtimox is activated upon metabolism to an open chain nitrile [11] and that this nitrile is as toxic as the parent drug. In mice there was no indication that either drug enhanced uptake of the other into brain [12], indeed eflornithine diminished brain penetration of nifurtimox in short term uptake assays. Moreover, isobologram analysis indicated that the two drugs were not synergistic in vitro [13].
It is very rare for a new chemotherapeutic agent to be licensed without prior knowledge of its MOA. In 2009, 19 drugs were approved by the FDA's centre for drug evaluation and research in the US, only one of which had a wholly unknown MOA [14]. A knowledge of the MOA reduces the risk of unexpected toxicity and allows synergism and resistance mechanisms to be predicted. Currently, the MOA of a drug is predicted using expensive and time-consuming enzyme-based assays, followed by targeted analyses of whether cellular death is associated with changes consistent with loss of the predicted target.
Metabolomics is a relatively new technology that enables the simultaneous identification of hundreds of metabolites within a given system. In principle, if an enzyme is inhibited by a drug then the concentration of substrate should rise within a system and the concentration of product fall. We have recently introduced metabolomics approaches to investigate metabolism in trypanosomes [15]–[19]. Here we use our metabolomics platform to test the mode of action of eflornithine (an ornithine decarboxylase suicide inhibitor that has had its MOA validated) and nifurtimox (a drug for which the MOA is incompletely understood). The combination therapy was also tested to determine any synergy between the drugs at the level of metabolism. Broad changes to cellular metabolomes in response to drug have been determined before [20]–[23], and an analysis of the effect of eflornithine along with an inhibitor of s-adenosyl methionine decarboxylase were studied using a targeted multi reaction monitoring (MRM) approach in Plasmodium parasites, responsible for malaria [24] is one of few studies have focussed on individual changes that can be mapped to specific targets in the metabolic network. Here we reveal that an untargeted LC-MS based metabolomics approach identifies specific changes in the metabolome of trypanosomes that can be related directly to effects induced by these drugs.
Bloodstream form trypanosomes were grown in HMI-9 (Biosera) [25] supplemented with 10% foetal calf serum (Biosera), incubated at 37°C, 5% CO2. Cells for metabolomics assays were grown in 500 cm3 Corning vented culture flasks to a maximum volume of 175 ml per flask. The Alamar blue assay developed by Raz et al. [26] for bloodstream form trypanosomes was used to determine activity of drugs against T. brucei strain 427. For isobologram analyses, alamar blue assays were conducted for one drug in the presence of three different concentrations (IC50, 2× IC50 and 0.5× IC50) of another drug.
A commercial kit (QuantiChrom, BioAssay Systems) was used to measure the arginase activity in cell extracts spectrophometrically (Dynex, wavelength 450 nM) by the amount of urea produced following manufacturers' specifications.
A rapid oil/stop spin protocol, previously described by Carter & Fairlamb [27], was used to determine uptake of radiolabelled ornithine (4,5-3H-ornithine, American Radiochemicals). Briefly, 100 µl of oil (1-Bromodo-decane, density: 1.066 g/cm3) (Aldrich) and 100 µl, 20 µM, 1% (v/v) radiolabelled ornithine in CBSS buffer were set up in a tube and 100 µl of cell suspension (108 per mL) for varying lengths of time before stopping the reaction by centrifugation. Alternatively, radiation was used at 1% (v/v) and cold ornithine levels were varied, while keeping the incubation time constant at one minute.
The resulting cell pellet was flash frozen in liquid nitrogen and the base of the tube containing the pellet was cut into 200 µl of 2% SDS in scintillation vials and left for 30 minutes. Three ml of scintillation fluid was added to each vial and these were left overnight at room temperature.
Counts per minute were read on a 1450 microbeta liquid scintillation counter (Perkin Elmer) and normalised between samples for the cell density. Michaelis-Menten kinetic analyses were performed using Graphpad Prism 5 software.
Metabolite extraction methods were adapted from Leishmania spp extraction techniques developed previously [13], [28], [29]. Cultures were kept in log phase growth (below 1×106/ml) in the presence of drug. At the time of harvest, 4×107 cells were rapidly cooled to 4°C to quench metabolism by submersion of the flask in a dry ice-ethanol bath, and kept on ice for all subsequent steps. The cold cell culture was centrifuged at 1,250 RCF for 10 minutes and the supernatant completely removed. Cell lysis and protein denaturation was achieved by addition of 200 µL of 4°C chloroform∶methanol∶water (ratio 1∶3∶1) plus internal standards (theophylline, 5-fluorouridine, Cl-phenyl cAMP, N-methyl glucamine, canavanine and piperazine, all at 1 µM), followed by vigorous shaking for 1 hour at 4°C. Extract mixtures were centrifuged for two minutes at 16, 000 RCF, 4°C. The supernatant was collected, frozen and stored at −80°C under argon until further analysis.
For heavy metabolite tracking analyses, log phase cells were centrifuged 1,250 RCF for 10 minutes and resuspended in CBSS (20 mM HEPES, 120 mM NaCl, 5.4 mM KCl, 0.55 mM CaCl2, 0.4 mM MgSO4, 5.6 mM Na2HPO4, 11.1 mM glucose) or HMI-9 as outlined in the Results section. Heavy atom labelled amino acids were obtained with 15N incorporation from Cambridge Isotope Laboratories (L-threonine (98% enrichment, one incorporation, alpha-N, cat:NLM-742-0), L-glutamine (98% enrichment, one incorporation, alpha-N, cat: NLM-1016-0), L-arginine (98% enrichment, four incorporations, allo-N, cat: NLM-396-0), L-ornithine (98% enrichment, two incorporations, allo-N, cat: NLM-3610-0), L-lysine (95–99% enrichment, one incorporation, alpha-N, cat: NLM-143-0)) or Sigma Aldrich (L-proline (98% enrichment, one incorporation, alpha-N, cat: 608998), L-glutamate (98% enrichment, one incorporation, alpha-N, cat: 332143)). Quenching of metabolism was achieved through rapid cooling and metabolite extraction was conducted as above.
Samples were analysed on an Exactive Orbitrap mass spectrometer (Thermo Fisher) in both positive and negative modes (rapid switching), coupled to a U3000 RSLC HPLC (Dionex) with a ZIC-HILIC column (Sequant) as has previously been described [13]. All samples from an individual experiment were analysed in the same analytical batch and the quality of chromatography and signal reproducibility were checked by analysis of quality control samples, internal standards and total ion chromatograms. The few samples that displayed unacceptable analytical variation (retention time drift) were removed from further analysis. A standard mix containing approximately 200 metabolites (including members of the polyamine pathway) was run at the start of every analysis batch to aid metabolite identification.
Untargeted metabolite analysis was conducted with the freely available software packages mzMatch [30] and Ideom (http://mzmatch.sourceforge.net/ideom.html). Raw LCMS data was converted to mzXML format and peak detection was performed with XCMS [31] and saved in peakML format. MzMatch was used for peak filtering (based on reproducibility, peak shape and an intensity threshold of 3000), gap filling and annotation of related peaks. Ideom was used to remove contaminants and LCMS artefact peaks and to perform metabolite identification. Metabolite identities were confirmed by exact mass (after correction for loss or gain of a proton in negative mode or positive mode ESI respectively) and retention time for metabolites where authentic standards were available for analysis, and putative identification of all other metabolites was made on the basis of exact mass and predicted retention time of all metabolites from the KEGG, MetaCyc and Lipidmaps databases [17]. Additional manual curation was performed on all datasets to confirm the identification of metabolites that changed significantly in response to drug treatment, and to remove false-identifications based on the LCMS meta-data recorded in Ideom. In cases where identification was putative, the most likely metabolite was chosen based on available chemical and biological knowledge, however accurate isomer identification is inherently difficult with LCMS data and lists of alternative identifications and meta-data for each identified formula are accessible in the macro-enabled Ideom files (Figure S1, Figure S2, Figure S3 and Figure S4; help documentation available at mzmatch.sourceforge.net/ideom.html). Quantification is based on raw peak heights, and expressed relative to the average peak height observed in untreated cells from the same experiment.
Unidentified peaks in the LCMS data were also investigated for drug-induced changes, however, after removal of LCMS artefacts and known contaminants, the only reproducible change (<3-fold) amongst the unidentified peaks was the appearance of C10H15N3O3S (mass = 257.0834, RT = 13.5) in the nifurtimox-treated cells. This mass is in agreement with the saturated open chain nitrile metabolite of nifurtimox.
The IC50 of eflornithine on bloodstream form cells in vitro was 35 µM using a standard alamar blue assay [23]. The IC50 of nifurtimox was 4 µM (Table 1). The drugs were widely believed to be synergistic given the fact that eflornithine ultimately diminishes polyamine production and in turn production of trypanothione, the trypanosome's principal anti-oxidant, whilst nifurtimox had been proposed to generate oxidative stress [6], [7]. However, we showed in isobologram analyses that the action of nifurtimox and eflornithine did not synergise when nifurtimox action was assayed in the presence of several fixed concentrations of eflornithine [13] and Fig. 1A. Indeed, an antagonistic effect was seen with a fractional inhibitory concentration of 1.61.
To determine the levels at which eflornithine is cytostatic and cytotoxic, time course assays were conducted with drug at various concentrations (Fig. 1B). Eflornithine was found to be cytostatic (cells remained at the same density even at 500 µM until around 55 hours in drug, when they died). There was no overt sign of differentiation to stumpy forms, but as the 427 strain is monomorphic, and thus incapable of the morphological changes associated with differentiation in field isolates, this would not be expected. Nifurtimox, on the other hand, had lysed all trypanosomes by 8 hours in 60 µM drug. It is possible that eflornithine's antagonistic effect could relate to its cytostatic potential, if, for example, nifurtimox activity depends on cellular proliferation.
The purine analogues, NA42 and NA134 are also known to be cytostatic [32] and these compounds were tested in combination with nifurtimox and also found to be antagonistic with FICs (fractional inhibitory concentrations) of 1.40 and 1.56 for NA42 and NA134 respectively. DB75, a known potent trypanocidal agent [33], on the other hand, was shown to be additive in its activity with nifurtimox (FIC: 1.09).
In order to detect molecular targets of eflornithine, a first experiment using sub-lethal levels (20 µM) of drug was used, with the cellular metabolome measured at 0, 1, 24, 48 and 72 hours following exposure to drug. Eflornithine was added to the 427 bloodstream form wild type cell line in the same growth medium in which IC50 values had been determined, so that cells were metabolising as normal apart from the perturbation by the drug.
The stringent filtering systems in the mzMatch and IDEOM software reduced the number of peaks in the spectra from several hundred thousand to a few hundred robust signals with putative metabolite identities (Fig. S1). Most metabolite levels were unaltered over the time points taken, indicating a high level of robustness within the trypanosome metabolome. Ornithine (mass: 132.0899, RT: 27.9 minutes), the substrate of eflornithine's known target, ornithine decarboxylase (ODC), was the most significantly modulated metabolite over the time course (7.5 fold increased at 48 hours). Putrescine (mass: 88.1001, RT: 36.91 minutes), the product of the ODC reaction was the only known metabolite in the T. brucei metabolite database at KEGG, to significantly decrease (by 66% at 48 hours) over time. Acetylated ornithine and putrescine were also detected, and these correlated highly with their non-acetylated counterparts. N-acetyl ornithine (mass: 174.1004, RT: 15.3 minutes) showed the most striking correlation. N-acetyl-putrescine (mass: 130.1106, RT: 15.5 minutes) was seen in early samples, but levels rapidly fell below the level of detection (1,000) from an average intensity of 41,000 (peak height) before drug addition, correlating with the decrease in putrescine.
Cells were also treated with 500 µM eflornithine, a lethal dose of the drug. At this dose bloodstream form trypanosomes exhibit division arrest over 48 hours in drug before dying between 48 and 55 hours (Fig. 1B). This was reflected by many more changes to the metabolome (Figure S2). Changes to polyamine pathway metabolites were again consistent with inhibition of ODC, with significant increases in ornithine and N-acetyl ornithine, and decreases in putrescine and N-acetyl putrescine, observed within 5 hours and maintained for the duration of treatment. Spermidine was significantly decreased by 24 hours, confirming the downstream effect of ODC inhibition on polyamine levels (Fig. 2). Additional metabolites that significantly increased within 24 hours were putatively identified as N-acetyl spermidine, N-acetyl lysine and N5-(L-1-Carboxyethyl)-L-ornithine (a known bacterial metabolite formed from ornithine and pyruvate, although we are not in a position to rule out its generation as a non-enzymatic liaison between these chemicals during sample preparation). These metabolites, along with N-acetyl ornithine, demonstrate metabolic derivitisation of ornithine and other polyamine metabolites, which may be an upregulated process in response to the elevated ornithine levels.
Aside from the polyamines, most major decreases in metabolite levels over 24 hours were observed among the phospholipids. Polyamines have previously been shown to be key mediators of membrane stability [34]–[36], and the lipid degradation observed here is consistent with cell membranes being compromised by polyamine depletion. Furthermore, the majority of metabolites in the cell decrease at the 48 hour time point, indicating a possibility that the cell membrane has been compromised and metabolites may be leaking from the cell during incubation and/or sample preparation. The processing of the cells involves cooling them to 0°C in a dry ice–ethanol bath and two centrifugation steps. These weakened cells are therefore potentially more leaky than cells that have not been compromised by prolonged exposure to eflornithine.
Several methionine-related metabolites (cystathionine, S-adenosyl methionine, methylthioadenosine and methyl-methionine) increased over the first five hours in drug, which was not reported in previous studies. S-adenosyl methionine is the aminopropyl donor involved in spermidine synthesis, and it is possible that this pathway has been upregulated in response to the declining polyamine levels. Methionine levels do not increase over this time course, however, this may be due to the high concentration of methionine in the growth medium (200 µM) and robust transport [37] masking any changes within the cells.
Despite the significant decrease observed for spermidine, levels of trypanothione disulphide were not affected during the first 24 hours of treatment. A significant decrease was observed at 48 hours. The analytical platform used here is not capable of reporting the oxidation state of trypanothione or other thiols.
The other significant changes observed during the first 24 hours of eflornithine treatment were not expected. Sedoheptulose (mass: 210.0738, RT: 14.9 minutes) and sedoheptulose phosphate (mass: 290.0400, RT: 25.4 minutes) were increased, as well as a metabolite with the chemical formula C7H12O5 (mass: 176.0683, RT: 7.52 minutes), putatively identified as propylmalate, but possibly diacetylglycerol or another isomer.
Our metabolomics analysis above reveals ODC to be the primary target of eflornithine, as was already clear based on previous work and the design of the compound as a specific inhibitor of the enzyme. Surprisingly, however, we could find no previous work that has focused on the cellular source of ornithine in T. brucei. In many eukaryotes, ornithine is produced from arginine via the enzyme arginase. In Leishmania parasites, which belong to the same taxonomic group as T. brucei, for example, an arginase enzyme has been characterised in some detail [38]–[41]. T. brucei, however, lacks a gene that is syntenic with the known Leishmania arginase. A second gene related to arginase is present in Leishmania and an orthologue is present in T. brucei (Tb927.8.2020). This latter predicted enzyme, however, lacks key arginase residues and is currently annotated as a putative agmatinase (although this also seems unlikely given the lack of conservation of key active site residues). We measured arginase activity in Leishmania mexicana extracts and compared this to T. brucei extracts where we show that the trypanosome contains little or no classical arginase activity when compared to Leishmania (Fig. 3A). The absence of a classical arginase raises questions about other potential sources of ornithine in T. brucei. Our experiments did reveal the presence of N-acetyl ornithine in T. brucei, the abundance of which was closely correlated to ornithine. Differences in the retention times between ornithine (RT = 27.9 minutes) and acetylornithine (RT = 15.3 minutes) confirm that the two metabolites are not mass spectrometry artefacts. In a variety of bacteria ornithine is produced from glutamate in a pathway that involves N-acetyl ornithine as an intermediate [42], [43].
We used heavy-nitrogen labelled metabolites to trace whether a similar pathway exists in T. brucei. However, cells incubated with isotopically-labelled extracellular 15N-glutamate failed to accumulate this amino acid to a detectable level. We therefore provided 15N labelled glutamine, which was converted to glutamate (albeit at a relatively low level of 5% of the non-labelled metabolite) after two hours and 15N-proline which was converted to glutamate at levels of 3.1% of the unlabelled glutamate generated within these cells. However, the heavy atom labelled glutamate was not further converted to ornithine, N-acetyl ornithine or N-acetyl glutamate semialdehyde (another metabolite of the glutamate to ornithine pathway). Furthermore, no orthologues, other than N-acetyl ornithine deacetylase (Tb927.8.1910), encoding enzymes of the bacterial pathway could be identified in the trypanosome genome indicating that this pathway is not operative in trypanosomes.
Although ornithine is not a component of HMI-9 medium, metabolomics analysis of our medium indicated that the commercial supply we used did contain ornithine and we were able to measure its concentration at 77 µM, using a calibration curve with isotopically labelled ornithine. We therefore measured the ability of 3H-ornithine to enter trypanosomes. This indicated a possible external source of ornithine and we tested the ability of this nutrient to enter trypanosomes. At 10 µM, ornithine was shown to enter bloodstream form T. brucei at a rate of approximately 10 pmol/107 cells/min (Fig. 3B). Kinetic analysis of ornithine transport indicated a Km of 310 µM and Vmax of 15.9 pmol/107 cells/min (Fig. 3C). Given that ornithine is present in serum and cerebrospinal fluid (at concentrations of 54–100 µM in plasma and 5 µM in CSF (according to the human metabolome database, http://www.hmdb.ca/)), this would indicate that T. brucei is capable of fulfilling its ornithine requirements directly by transport from the bodily fluids in which it resides. When we used 15N-ornithine externally to trace its metabolism we showed that N-acetyl ornithine, spermidine and trypanothione disulphide when added to cells growing in HMI-9. 15N-labelled arginine was converted to ornithine when administered in CBSS (Carter's balanced saline solution), but not when administered in HMI-9 growth medium. This suggested that when exogenous ornithine is present, uptake of ornithine is sufficient for polyamine synthesis, but when absent, synthesis from arginine is possible. This was confirmed by the addition of exogenous ornithine in addition to heavy arginine in CBSS, where synthesis of heavy ornithine from arginine was no longer detected (heavy ornithine being present at 40% of unlabelled ornithine levels when exogenous ornithine was not added under the same conditions). The enzymatic route by which arginine is converted to ornithine in the absence of canonical arginase is not known.
At the sub-lethal dose of 1.5 µM nifurtimox, no significant changes to the metabolome were recorded (data not shown). However, at a lethal dose of 60 µM changes to the metabolome at 0, 1, 2 and 5 hours following exposure to drug, were seen (Figure S3). Nifurtimox (mass: 287.0577, RT: 5.25 minutes) was observed in all treated samples, in addition to a mass (mass: 257.0834, RT: 13.5 minutes) consistent with the saturated open chain nitrile metabolite [11] (Fig. 4A) which was recently shown to be the end product of the multi-step 2-electron reduction of nifurtimox by type-1 nitroreductase. Previous work in a cell-free system showed the saturated nitrile only after 24 hours of drug exposure to the nitroreductase [11]. Our metabolomics platform allows identification of this metabolite within the cell, and shows the process to be rapid with significant levels detectable at the first, 1 hour time point. The implicit intermediates from this reductive activation cascade, including the unsaturated open chain nitrile proposed to mediate trypanocidal activity, were not observed, indicating either that the reduction is rapid and intermediates in the pathway do not persist at detectable concentrations, or that the reactive intermediates indeed react rapidly with intracellular macromolecules. An exhaustive search of all known metabolites in our database revealed no detectable masses that correspond to a hypothetical adduct between the unsaturated open chain nitrile and any known metabolite. Our metabolomics platform, by definition, was unable to detect the potential formation of adducts between nifurtimox metabolites and macromolecules (proteins or nucleic acids).
A number of cellular metabolites were shown to change in abundance over the nifurtimox exposure time course (Table 2), although 95% of putatively identified metabolites were stable. There was an increase in concentrations of nucleotides and nucleobases (adenine, deoxyadenosine, AMP, GMP, uracil and UMP) during the time course, which may result from degradation of RNA and DNA consistent with the hypothesis [11] that the nifurtimox active metabolite (the open chain nitrile) binds to macromolecules including nucleic acids, by the ability of the unsaturated nitrile intermediate to act as a Michael acceptor [11].
Glycolysis appeared to be downregulated, with significant decreases in hexose 6-phosphates, and similar trends for glyceraldehyde 3-phosphate and 3-phosphoglycerate. The metabolite that decreased most following nifurtimox treatment was deoxyribose, which may indicate reduced synthesis from the glycolytic intermediates, or could be related to nucleic acid homeostasis. Lipid metabolism was largely unaffected with the exception of decreased levels of monounsaturated ether-linked lysophosphatidylcholines (14∶1, 15∶1 and 16∶1) and ethanolamine phosphate.
Metabolites of the polyamine pathway were not significantly altered over the nifurtimox time course, although decreased thiol levels (trypanothione disulphide and glutathionyl-cysteine disulphide) were observed, suggesting that oxidative stress may be induced on exposure to nifurtimox in agreement with previous studies [6], [7], [44], although the role of this stress in ultimate trypanocidal effect is uncertain. It is noted that this untargeted metabolomics approach is not suited for assessment of redox balance (as reduced thiols are oxidised during sample preparation and analysis), however and it is assumed that the observed disulphide levels are indicative of total thiol levels. The presence of oxidative stress may also explain the observed inhibition of glycolysis [45], and the decreased levels of arginine phosphate [46].
We also investigated changes to the metabolome associated with exposure to eflornithine and nifurtimox simultaneously (Figure S4). The metabolome of NECT treated cells was measured using drug levels that were toxic in the monotherapies (500 µM for eflornithine and 60 µM for nifurtimox) and the time points used in the nifurtimox toxicity assay (0, 1, 2 and 5 hours), after which cells died without remaining viable for as long as studied in the eflornithine monotherapy study. The rapid reduction of Nifurtimox (within 1 hour) to the saturated open chain nitrile was still observed (Fig. 4B). This indicates that the nitroreductase activity known to be responsible for metabolic activation of nifurtimox [11] is not diminished in a short term response to eflornithine.
The combination therapy showed qualitatively most of the same changes that were present in each of the monotherapies alone (Table 2 and Fig. 5). This indicates that both of the drugs are able to exert their individual effects and no additional effects were apparent using the combination. The eflornithine-induced changes to polyamine pathway metabolites were observed in the combination (Fig. 2), although the later effects of eflornithine could not be measured as cells died from nifurtimox toxicity before these were apparent. The nifurtimox-induced changes to nucleotides, glycolysis intermediates, deoxyribose and thiols were all observed to a similar extent in the combination treatment.
Understanding how small chemicals interfere with cellular metabolism is a critical part of modern drug development. Here we show how a relatively simple LC-MS based metabolomics platform can be used to identify drug modes of action in the causative agent of human African trypanosomiasis, Trypanosoma brucei. Using each of the drugs currently used in combination as a first line treatment against stage two HAT we reveal how modes of action of drugs can be rapidly ascertained.
At low levels of drug (sub IC50) specific changes to the metabolome can be detected as was evidenced with eflornithine. The data reveal very localised changes to the metabolome with little indication of broadly disseminated affects consistent with the theory that metabolic networks are generally robust to perturbations [47].
This study reveals the power of metabolomics for predicting the MOA of compounds with a metabolic (enzyme inhibition) mode of action. As ornithine accumulation and putrescine loss were the most significant changes between treated and untreated cells, ornithine decarboxylase emerges as the most likely target for this drug. In this case, the outcome was already known hence the follow up experiments e.g. showing that ODC is essential using gene knockout [48] and that addition of polyamines to the medium can bypass eflornithine toxicity [48] have already been performed. With unknown drugs, of course, these validation experiments are still required once the hypothesis has been set using metabolomics. The presence of the open chain trinitrile in nifurtimox-treated cells confirms the trypanosome-mediated metabolic activation of this drug, as was recently demonstrated following substantial targeted analysis of nifurtimox [11]. It will be of interest to extend studies to other current trypanocides and also to systematically include metabolomics in any test of action for compounds emerging from screens.
Eflornithine inhibited ODC relatively quickly with levels of ornithine and putrescine demonstrably altered after just five hours in toxic doses of drug. Trypanothione is a glutathione-spermidine adduct and its overall levels are diminished by around 73% prior to death in the eflornithine study, which is similar to the 66% reduction determined after eflornithine exposure in vivo [10], but it should be noted that many unrelated metabolites were also diminished at 48 hours. An advantage of the non-targeted metabolomics platform used here over a strictly targeted approach to report on individual metabolites is thus clear. Loss of putrescine and spermidine appears to contribute to cellular toxicity independently of their role in trypanothione biosynthesis as rescue experiments where spermidine is given exogenously to ODC knock down cells were unsuccessful [49]. Our studies indicate that eflornithine is trypanostatic for 48 hours, before killing the parasites after apparently compromising the membrane of the cell, as judged by a general loss of metabolites from the cell and particularly changes in lipid content. Since polyamines have been proposed to help stabilise membrane phospholipids [34], [35] this could indicate the actual cause of death following reduction in polyamine biosynthesis. In vivo, changes to membrane integrity would also expose new ligands to the immune systems, possibly explaining the need of an active immune system for optimal eflornithine activity [47]. Nifurtimox did not show the same depletion in membrane integrity prior to cell death.
The untargeted metabolomics approach was particularly useful for the identification of unexpected metabolites. Acetylated ornithine and putrescine have not been previously described in trypanosomes, and would likely not have been assessed with a classical targeted approach, but these results clearly show the presence of acetylated polyamines, and their dynamic relationships with polyamine levels, with N-acetyl ornithine correlating particularly well with ornithine. This metabolite has an unknown function within trypanosomes but appears to be created directly from ornithine transported into the cell. We have also shown that trypanosomes do not use classical arginase activity comparable to that found in related Leishmania spp. parasites to create ornithine from arginine but do have the ability to transport ornithine which is present in plasma and CSF, indicating that they probably fulfil ornithine needs by acquiring it directly from the host. Interestingly they can, nevertheless, convert arginine to ornithine, but apparently only when exogenous ornithine is not available.
An increase in sedoheptulose and sedoheptulose phosphate in eflornithine-treated cells was also of interest. Sedoheptulose phosphate is a seven carbon sugar of the pentose phosphate pathway, formed, along with glyceraldehyde 3-phosphate, from ribose 5-phosphate and xylulose 5-phosphate through transketolase (Tb927.8.6170) action. Transketolase activity, however, is absent in bloodstream-form trypanosomes [50], [51], although it is induced in parasites transforming between bloodstream and procyclic forms. It is possible, therefore, that the increase in sedoheptulose 7-phosphate could relate to transketolase being switched on in relation to the proposed induction of differentiation between slender bloodstream form and stumpy form organisms. Although the s427 strain used here does not differentiate to stumpy forms, other biochemical events such as the induction of enzymes usually repressed in the non-dividing stumpy stage may occur. Nifurtimox treatment did not induce any changes to sedoheptulose or its phosphate's levels.
Toxic doses of nifurtimox revealed alterations to levels of various nucleotide, carbohydrate and lipid metabolites. More work is required to ascertain why these metabolites' levels are altered with nifurtimox treatment and how these changes relate to death. However, our data reveal that this metabolomics approach can confirm previous findings that relate oxidative stress to nifurtimox treatment, and demonstrate the production of an open chain trinitrile metabolite in agreement with the proposed mechanism for the drug's selective activity against trypanosomes [11]. We show also that the appearance of this metabolite is relatively fast, being detectable within an hour of exposure.
The nifurtimox-eflornithine combination therapy, which was previously assumed to be synergistic, was shown to be mildly antagonistic in vitro. The theory behind synergy was based on the assumption that eflornithine would decease cellular trypanothione levels thus decreasing the ability of these cells to defend against oxidative stress. Since nifurtimox was generally believed to exert an effect through generation of reactive oxygen species [6], [7] it followed that eflornithine treated cells would show enhanced susceptibility to nifurtimox. However, the metabolic perturbations observed in this study suggest that oxidative stress is not the primary MOA for either drug (despite some indication of oxidative stress observed with nifurtimox), and if nifurtimox actually acts through production of the reactive open chain trinitrile and its ability to covalently modify macromolecules, then the proposed synergy would not exist. It should be noted too, however, that our studies in vitro need not reflect the situation in vivo where pharmacokinetic factors lead to very different exposure of parasites to drug and where other host related factors, not least the immune response, contribute to effects of the drugs, although in mice at least neither drug facilitates entry of the other into the brain.
A potential reason why the drug combination is mildly antagonistic in vitro could relate to the activation of nifurtimox and its target based effects depending upon growth status of the cell. There was no evidence that activation of nifurtimox was reduced in the eflornithine co-treated cells. Instead, therefore, it is possible that cells entering a state of reduced growth are less affected by the impact of nifurtimox on energy and nucleic acid metabolism. This hypothesis was supported by the antagonism to nifurtimox seen with the trypanostatic purine analogues NA42 and NA134 [36].
The examples we provide here demonstrate how a relatively simple metabolomics platform can elucidate the mode of action of a drug in a relatively short time frame. This study shows that our metabolomics platform yields hypothesis-free data that confirm the known MOA of eflornithine and create testable hypotheses for the nifurtimox MOA as well as confirming a lack of synergy of NECT. The approach we provide here can be readily adapted for other drugs and cellular systems.
|
10.1371/journal.ppat.1003611 | Dual Analysis of the Murine Cytomegalovirus and Host Cell Transcriptomes Reveal New Aspects of the Virus-Host Cell Interface | Major gaps in our knowledge of pathogen genes and how these gene products interact with host gene products to cause disease represent a major obstacle to progress in vaccine and antiviral drug development for the herpesviruses. To begin to bridge these gaps, we conducted a dual analysis of Murine Cytomegalovirus (MCMV) and host cell transcriptomes during lytic infection. We analyzed the MCMV transcriptome during lytic infection using both classical cDNA cloning and sequencing of viral transcripts and next generation sequencing of transcripts (RNA-Seq). We also investigated the host transcriptome using RNA-Seq combined with differential gene expression analysis, biological pathway analysis, and gene ontology analysis.
We identify numerous novel spliced and unspliced transcripts of MCMV. Unexpectedly, the most abundantly transcribed viral genes are of unknown function. We found that the most abundant viral transcript, recently identified as a noncoding RNA regulating cellular microRNAs, also codes for a novel protein. To our knowledge, this is the first viral transcript that functions both as a noncoding RNA and an mRNA. We also report that lytic infection elicits a profound cellular response in fibroblasts. Highly upregulated and induced host genes included those involved in inflammation and immunity, but also many unexpected transcription factors and host genes related to development and differentiation. Many top downregulated and repressed genes are associated with functions whose roles in infection are obscure, including host long intergenic noncoding RNAs, antisense RNAs or small nucleolar RNAs. Correspondingly, many differentially expressed genes cluster in biological pathways that may shed new light on cytomegalovirus pathogenesis. Together, these findings provide new insights into the molecular warfare at the virus-host interface and suggest new areas of research to advance the understanding and treatment of cytomegalovirus-associated diseases.
| We have conducted a comprehensive analysis of the murine cytomegalovirus and host cell transcriptomes during lytic infection. We identify numerous novel spliced and unspliced transcripts of MCMV. Unexpectedly, the most abundantly transcribed viral genes are of unknown function. We found that the most abundant viral transcript, recently identified as a noncoding RNA regulating cellular microRNAs, also codes for a novel protein. To our knowledge, this is the first viral transcript that functions both as a noncoding RNA and an mRNA. Infection alters expression of many unexpected host genes, including many noncoding RNA genes. Correspondingly, many cluster in unexpected biological pathways that may shed new light on cytomegalovirus pathogenesis. Together, these findings provide new insights into the molecular warfare at the virus-host interface and suggest new areas of research to advance the understanding and treatment of cytomegalovirus-associated diseases.
| The cytomegaloviruses, classified within the Betherpesvirinae subfamily, are a group of species-specific herpes viruses that establish life-long infection of their hosts. Human cytomegalovirus (HCMV) can cause devastating disease and death in congenitally-infected infants, and long-term neurological complications in survivors. In adults, HCMV can cause a spectrum of diseases in immune compromised patients involving multiple organs and tissues and is a primary cause of graft loss in transplant patients [1], [2]. In recent years, HCMV has been linked to lung injury in trauma patients [3] and is also postulated to act as a cofactor in atherosclerosis and some cancers [4], [5]. For these reasons, there is an urgent need for an effective vaccine and new antiviral intervention strategies that mitigate the toxicity and drug resistance shortcomings of current antiviral drugs [1], [6].
There exist a number of challenges to our understanding of CMV pathogenesis as well as progress in vaccine and antiviral drug development. Two outstanding challenges are the gaps in our knowledge of viral genes and how these gene products interact with host cellular gene products to cause disease. Despite the publication of the first sequence of the HCMV genome in 1990 [7], [8], and the first sequence of the murine cytomegalovirus (MCMV) genome in 1996 [9], there are still important questions regarding the nature and number of genes for these viruses.
MCMV is the most widely used model to study HCMV diseases and recapitulates many of clinical and pathological findings found in human diseases. Our understanding of MCMV viral genes and genomes has evolved with the technology used to study them. A major milestone in understanding MCMV came with decoding the first MCMV complete genome sequence by Rawlinson and colleagues [9]. The authors identified a 230 kb genome predicted to encode 170 genes.
Subsequent refinements in the annotation of the MCMV were introduced by classical molecular and biochemical studies that are reflected in the current NCBI reference sequence. The application of new technologies to study the MCMV genome emerged in the last decade and include proteomic [10], in silico [11], and gene array [12], [13] approaches that have led to major revisions in gene annotation. More recently Cheng and colleagues [14] proposed additional changes after sequencing isolates to measure genome stability after in vitro and in vivo passage. Also, Lacaze and colleagues [13] extended the microarray approach to include probes specific to both strands of the genome, leading to the discovery of noncoding and bi-directional transcription at late stages of MCMV infection. Finally, a recent transcriptomic analysis of newly synthesized RNA in MCMV infected fibroblasts [15] applied RNA-Seq technology to study regulation of viral gene expression and identified a very early peak of viral gene transcriptional activity at 1–2 hours post infection followed by rapid cellular countermeasures but did not attempt to re-annotate MCMV genome.
Altogether, these new technologies have refined and advanced our knowledge of viral genes and the MCMV genome. Nevertheless, we still lack definitive annotation for the standard lab strains of MCMV and specific knowledge of how many of these genes function during natural infection and disease. Currently, two annotations of MCMV genomes are used – the original Rawlinson's annotation with minor modifications (GenBank accession no. GU305914.1) where 170 open reading frames (ORFs) are identified and the NCBI reference sequence annotation (GenBank accession no: NC_004065.1) with 160 ORFs. We previously used a transcriptomic approach to analyze gene products of HCMV [16]. This was the first report to characterize abundant antisense and noncoding transcription in the HCMV genome showing that there is greater complexity of herpesvirus genomes than previously appreciated. Using RNA-Seq technology, Gatherer et al. [17] showed that most protein coding genes are also transcribed in antisense but are generally expressed at lower levels than their sense counterparts. A more recent analysis of translational products of HCMV [18] by ribosomal footprinting indentified 751 translated ORFs, further underscoring the complexity of herpes virus genomes.
We describe MCMV transcriptional products that differ from predicted ORFs, novel spliced transcripts, and novel transcripts derived from intergenic regions of the genome. Additionally, we found that the most abundant viral transcript (MAT) is a spliced transcript recently identified as a noncoding RNA that limits accumulation of cellular miRNAs [19], [20]. Here we report that this transcript also specifies a novel protein and to our knowledge, this is the first viral transcript that functions both as a noncoding RNA and mRNA. Analysis of the host transcriptional response to infection revealed many unexpected host genes that are regulated by virus infection, including many noncoding RNA genes. Correspondingly, many host genes regulated by virus infection cluster in unexpected biological pathways that may shed new light on the pathogenesis of cytomegalovirus-associated diseases. Together, these findings suggest important revisions are required for MCMV genome annotation and emphasize numerous aspects of MCMV biology and the host response to this infection that are unknown and require further study.
In this study, we set out to complete a transcriptomic analysis of MCMV infection. We analyzed viral transcripts through classical cDNA cloning and sequencing and through next generation sequencing of cDNA generated from total cellular RNA (RNA-Seq). Analysis of cDNA libraries is a well-proven approach to analyze viral transcripts based on isolation of long transcripts, molecular cloning of the transcripts, and traditional Sanger-based sequencing of the cDNA clones. Traditional cloning has many advantages, including isolation of novel transcripts that may not be identified by probe-based technologies, as well as precise analysis of transcript splice sites and transcript 3′ ends. The introduction of massively parallel sequencing techniques represents a major new technology to study gene expression. Basically, RNA (total or fractionated) is converted to a library of smaller cDNA fragments. Adaptors are added to the fragments, and the shorter fragments are sequenced in a high-throughput manner using next generation sequencing technology. This RNA-sequencing (RNA-Seq) approach is free of selection biases associated with traditional cloning or probe-based methods and allows for the entire transcriptome to be analyzed in a quantitative manner (reviewed in [21]).
First, cDNA libraries representing the major temporal classes of viral gene expression were generated by collecting RNA from infected mouse embryonic fibroblasts (MEFs) at 9 time points after infection. For RNA-Seq analysis, RNAs collected at the same 9 time points were pooled, converted to cDNA, and sequenced on the Illumina Genome Analyzer IIx. Of the 33,995,400 reads that passed the filter from infected cells, 11% aligned to MCMV genome indicating a 585-fold coverage of the viral genome.
A total of 448 cDNA clones were included in the final analyses [84 from the immediate early (IE) library, 163 from the early (E) library, and 201 from the late (L) library]. Generally, temporal assignment of cDNA clones in this study agrees with previous studies and a detailed comparison, including discrepancies to earlier studies is provided in Dataset S1.
As shown in Figures 1 and 2, transcriptomic data generated using these two experimental approaches were compared to currently available genome annotation (the NCBI reference sequence, GenBank accession. no. NC_004065.1, and a more recent sequence analysis of the Smith strain, GenBank accession no. GU305914.1). Using this schematic overview, current annotations (red and blue arrows) largely agree. The MCMV transcripts identified through our classical cDNA cloning and sequencing (green arrows) and the RNA-Seq expression profiles (gray histograms), showed complementary results to each other but diverged dramatically from current annotations. A summary of the cDNA clones relative to genes annotated in the NCBI reference sequence (NC_004065.1) is shown in Table S1, and a complete list of the 448 cDNA clones isolated in this study and their characteristics are presented in Table S2. A detailed comparison of transcripts cloned in this study compared to current annotations (GU305914.1 and NC_004065) is presented in Table S3, and includes estimated relative expression measured as reads per kilobase per million mapped reads (RPKM) derived from RNA-Seq analysis.
Analysis of the cDNA clones revealed many novel transcripts with the following characteristics:
Approximately 31% of the viral cDNA clones and 41% of all viral reads from the RNA-Seq analysis mapped to the novel spliced transcript at the right end of the genome. The structure of this transcript relative to current gene annotation is shown in Figure 3A and the spliced nature of this transcript is also apparent in the RNA-Seq profile (gray histogram). The longest predicted ORF extends into the second exon and predicts a protein of 147 AA, of which the first 127 residues matches the predicted m169 protein sequence (Figure 3B). To determine if this transcript is translated, an antibody was prepared to the protein sequence predicted for m169. This antibody was used in immunoblot analysis of cell lysates prepared from mock-infected cells and cells infected with wild-type (WT) BAC-derived Smith strain virus, or various multi-gene and single-gene mutant virus strains. This antibody reacted with a protein of approximately the expected size (17 kDa) (Figure 3D) in cells infected with WT virus and mutant viruses that express most or all or the MAT transcript as determined by northern blot analysis (Figure 3C).
While the MAT protein is first detected 24 hrs after infection, maximal amounts are observed at 60 and 72 hrs after infection (Figure 3E). MAT protein was also detected in fibroblasts exposed to five different strains of MCMV isolated from wild mice, indicating that the coding region of the gene is conserved in laboratory and wild strains of MCMV (Figure 3G). Most remarkably, these findings demonstrate that the MAT gene generates a single transcript with both noncoding [19], [20] and protein coding functions.
We also investigated the possibility that MAT protein accumulation is directly related to control of the MAT transcript levels by cellular miR-27 [20]. Marcinowski and colleagues have shown that when the binding site for miR-27 is mutated (m169-mut virus), MAT transcript levels were increased twofold in comparison to levels obtained in cells infected with WT MCMV at 24 hours after infection due to loss of transcript regulation by the microRNA. The difference in MAT transcript abundance between WT and the m169-mut virus was lost by 48 hours after infection. However, we observed that MAT protein levels were similar in cells infected with the WT and m169-mut viruses at 24 and 48 hrs after infection (Figure 3H). We conclude that the noncoding function of the MAT transcript (regulation of cellular miR-27) is unrelated to MAT protein accumulation.
In addition to providing valuable insight into transcript structure, RNA-Seq analysis revealed several new facets of the viral transcriptome. First, accumulation of individual viral transcripts varies by several orders of magnitude. Figure 4A depicts the number of RNA-Seq reads mapped against the MCMV genome in which rows 1, 2 and 3 visualize the data with the maximum number reads set to 50,000, 5000, and 500, respectively. Validating the classical cloning study, the most abundant transcript identified by RNA-Seq is the MAT transcript. Enumeration (RPKM) of the most abundant transcripts is presented in Figures 4B and 4C, and shows that after the MAT transcript, the most highly expressed genes are m119, M116, and m48, all genes without assigned functions. Also highly expressed are the immune evasion genes m04 and m138, M55 (glycoprotein B), and additional genes of unknown functions (M73 and m15). Second, as shown by comparing Figure 4B to C, both the overall magnitude of expression and the ranking of the abundance of different transcripts vary according to annotation used for the RPKM analysis. Third, analysis of reads mapped at the highest resolution (Figure 4A, row 3) indicates that most of the viral genome is transcribed to some degree. Remarkably, 30 or 35% of the reads mapped to intergenic regions, depending on annotation (Figure 4D). This percentage is reduced to 14% when annotation is modified to reflect the correct MAT gene structure identified in this study. RNA-Seq also detected significant transcription in m74-M75 (Figure S4), M85-M87 and M88-m90 (Figure 4E) intergenic regions. In contrast, the transcriptional profile of the annotated M87 shows less transcriptional activity than the adjacent intergenic regions. Similarly, RNA-Seq identified transcription from genes that were not isolated in the classical cDNA library or in previous studies using microarray technology [12], [13]. A detailed analysis of the sensitivity of this RNA-Seq study to previous studies is provided in Supplemental Datasets S1A–C. We also compared our RNA-Seq data to a recent RNA-Seq analysis of the MCMV transcriptome using BAC-derived WT virus on NIH-3T3 fibroblasts [15]. As shown in Supplemental Figure S1, the profiles obtained from these two different RNA-Seq experiments are remarkably similar despite using different sequencing platforms and library generation approaches. Also, either seven or eight of the 10 most abundant genes were identical in both datasets (Supplemental Dataset S1C). Minor differences in abundance of some transcripts can be attributed to differences in the time points analyzed in these two studies as well as the fact that our analysis achieved an order of magnitude greater sequencing depth (compare reads analyzed for each histogram set in Figure S1).
Together these findings demonstrate that RNA-Seq analysis is a highly sensitive method for detection of viral gene expression during infection. Moreover, these findings highlight numerous incongruencies with current annotation for the MCMV genome. Finally, RNA-Seq analysis revealed that many of the most abundantly expressed viral genes are of unknown function.
Because cDNA cloning and RNA-Seq identified significant differences between the MCMV transcriptome and current annotations, we performed an in depth analysis of several genomic regions by northern analyses (Figure 5, Figures S2, S3, S4, S5) using our cDNA clones to generate strand specific riboprobes (Table 1).
To investigate genomic regions where transcripts overlapping more than one gene were detected, we analyzed transcription in m15–16 and m19–20 regions. In both regions multiple transcripts were detected with different temporal expression patterns. Smaller transcripts tended to accumulate at later time points, a feature previously reported for certain transcripts in both HCMV and MCMV [22]–[24]. In the m15–m16 region 5 transcripts were cloned, all of which overlapped the predicted m15 and m16 genes, and one transcript was spliced (Figure 5A). The RNA-Seq profile (Figure 5A, gray histogram) also strongly indicated transcription that spans both predicted genes. Consistent with our cDNA library, no antisense transcripts were detected while the sense probe detected 5 transcripts (Figure 5A, bands 1–5). The 3′ end of all cDNA clones end at or close to nucleotide position 15700 (Supplemental Table S2) and RNA-Seq data alignment to MCMV genome indicates a sudden drop in reads around this nucleotide position. Assuming that transcripts in this region are co-terminal, band sizes predict transcript initiation sites in m11, m12, m13 or m14, and m15 (Figure 5A, bands 1–4). Similar results were found in this region in cells infected with wild isolates of MCMV (Alec Redwood, personal communication). While the smallest band observed by northern analysis (Figure 5A, band 5) corresponds in size to the novel spliced transcript, E253 (566 bp), we could not confirm splicing by PCR using intron-flanking primers (data not shown). Therefore, it is likely this spliced transcript represents an aberrant transcript or a result of intramolecular template switching during reverse transcription [25].
The m20 gene region also diverged substantially from current annotation. Similar to m15–16 region, in the m20 region 5 transcripts with differential temporal expression patterns were detected by northern analyses using clone IE205 as a probe (Figure 5B). No transcript was detected using an AS probe derived from clone IE205 or L57, indicating that m19 is not transcribed in the sense orientation (Figure 5B and Figure S2). We therefore propose that m19 should be removed from MCMV genome annotation. This is consistent with our cloning study where 4 transcripts have been isolated in this region and none overlap m19 in the sense orientation (Supplemental Table S2). Evidence from the cDNA library and RNA-Seq alignment (Table S2 and Figure 5B) indicates that 5 bands detected in northern are co-terminal, with the 3′ end located close to nucleotide position 20430. The largest band at 4 kb (Figure 5B, band 1) is detectable at 30 and 60 hours PI and we predict that this transcript initiates in M23. We failed to detect transcripts between the m20 and m25 genes consistent with previous studies [12] and northern analyses using mutant viruses lacking genes in this region (data not shown). The lack of cDNA clones can be explained by low abundance and size of this transcript, as well as by propensity of cDNA libraries to enrich 3′ ends. The band slightly smaller than 3 kb (Figure 5B, band 2) shows a peak accumulation at 60 hours PI and is consistent with a transcript overlapping m19-m21 (approx locations 20430–23220, 2.79 kb). The band of approximately 2 kb (Figure 5B, band 3) shows peak accumulation 10 hours PI. Based on the RNA-Seq profile this band could represent transcripts that initiate at nucleotide position 22060. Finally, the late time points are dominated by smaller transcripts of approximately 1 kb (Figure 5B, band 5; predicted start site at 21440) which correspond in size to the cDNA clones detected in our study. In short, northern analyses support the conclusion that transcripts overlapping multiple genes in the m15-m16 region and m19-m20 region accumulate in infected cells and indicate that additional, larger transcripts are transcribed in these regions which have yet to be characterized.
Next we analyzed gene regions in which we detected novel spliced transcripts. The M116 region was chosen as an example of a highly abundant spliced transcript of unknown function in addition to m168–169 transcript. Current annotations predict an ORF of 1.9 kb whereas RNA-Seq profiles and cDNA study both detected a slightly shorter (1.6 kb) transcript with an 81 bp intron. Northern blot analysis (Figure 5C) identified a strong band of appropriate size (1.6 kb) that starts to accumulate at IE times and peaks at E and L times after infection. Due to the small intron and high abundance of this transcript, unspliced transcripts could not be definitively resolved by northern analysis but were confirmed by PCR using primers flanking the intron (Figure S3). Additionally, northern analysis detected another less abundant band of approximately 3 kb. Leatham and colleagues [26] have detected a band of similar size in the homologous region in HCMV of 3.2 kb that encompasses UL119-115 genes. While we failed to isolate cDNA clones overlapping m117 region, we predict the larger, less abundant 3 kb band observed initiates in m117, though additional northern or 5′RACE studies are needed to confirm the start site of the larger transcript.
The m72-m74 region was previously shown to have a very complex transcriptional profile [23], [27]. cDNA library data, RNA-Seq data and results of northern analyses with L42 as a probe all are in agreement with the findings of Scalzo et al. [23] of multiple spliced transcripts that share exon 2. Bands 5–7 (Figure 5D) correspond to previously reported m60, m73 and M73.5 spliced transcripts. In our cloning study four isolated clones correspond to M73.5 transcripts (represented by the longest clone, L443) and one to M73 (L33) (Supplemental Table S4). Transcripts corresponding to m60 were not isolated in the cloning study, however the RNA-Seq profile in the region corresponding to m60 exon1 shows active transcription (Figure S4D). We have also detected a 1.1 kb band (Figure 5D, band 6) that corresponds to longer M73 and M75 transcripts, and bands corresponding to unspliced transcripts of M73 and M73.5 (approx. 2 kb, Figure 5D, band 4). In addition to these previously published transcripts, we have also cloned one novel spliced transcript from this region, E180. In accordance with work in the m60 region [23], we propose M71S as a designation for this novel gene. Like other spliced transcripts from this region, E180 shares exon 2 with other transcripts while its splice donor site is located at 102830. The spliced nature of this transcript has been confirmed by PCR (Figure S4C), however more analyses are needed to determine its exact 5′ start site. Northern analysis revealed a band of 0.5 kb (Figure 5D, band 7) that corresponds in size to the E180 spliced transcript whereas the unspliced version is detected around 3 kb (Figure 5D, band 3). In addition, a band of similar size (3.5 kb) transcribed from the plus genomic strand detected by the L147 probe (Figure Figure S4B) could correspond to the unspliced version of E180. All probes used in this region detected bands transcribed from negative genomic strand that correspond to those previously reported by Rapp et al. [27]. Based on additional northern blots using cDNA clones L69 (AS to m72) and L147 (AS to m74) (Figure S4A and B) as well as previous reports [23], [27] we conclude that the 5 kb transcript starts in m75 and ends in m72 and corresponds to transcript encoding gH while the 3 kb transcript starts in m74 and ends in m72 and corresponds to transcript encoding dUTPase (Supplemental Figure S4). Additional very large transcripts transcribed from plus or minus genomic strands detected by the L42 and L147 probes, respectively, have yet to be characterized but underscore the complex transcriptional patterns in this region.
Last, we set out to confirm novel antisense transcripts detected in the cDNA library. Analysis of transcription in M100-M103 region confirmed previously published findings. We have detected single M102 transcript from plus genomic strand as described by Scalzo [28] (Figure S5A). A probe derived from M100 detected a single transcript from negative genomic strand corresponding to M100 [28], and 2 from the positive genomic strand that correspond to those described by Cranmer et al. [29] and are in line with our cDNA and RNA-Seq analysis (Figure S5B). The presence of sense and antisense transcripts in this gene region corresponds to findings for HCMV [16]. Finally, in the M103 gene region we detect 2 transcripts from plus genomic strand that correspond to those described by Lyons et al [30] (Figure S5C). Temporal expression of transcripts detected by northern in this region is in line with our cDNA analysis and previously published data [13], [15].
Based on northern analyses of 5 regions, we conclude that our cDNA and RNA-Seq analyses faithfully represents the MCMV transcriptome in infected primary fibroblasts and confirms the presence of novel transcripts. Moreover, the distribution of clones in the IE, E and L cDNA libraries accurately reflected the accumulation of transcripts detected by northern analyses.
RNA-Seq analysis also enabled us to investigate changes in the host transcriptome. Differentially expressed (DE) murine genes in MCMV-infected cells compared to mock-infected cells were identified by calculating RPKM. This analysis identified 10,748 statistically significant (p<0.05) genes altered by infection (Table S5). The top induced, upregulated, repressed and downregulated genes are presented in Tables 2–5, (genes associated with characterized biological pathways are in bold). Interferon β (Ifnb1) and the interferon-inducible pyhin1 (a.k.a. ifi-209, ifix) were among the top induced genes, consistent with the expected host response to virus infection. Also congruous with expected host responses to infection were two highly induced genes associated with apoptosis induction (Hrk and Tnfsf10 [a.k.a. TRAIL]). Interestingly, transcription factors (Foxa1, En2, Insm1, Tbx21, [a.k.a T-bet], and Tp73) were among the most strongly induced by MCMV infection.
Chemokine ligands dominated the group of the top upregulated genes. Genes encoding proteins with roles in intrinsic cellular defense were also highly upregulated, including OAS1, Mx1, Gpb5 and Rsad2 (a.k.a. viperin). There were also a surprising number of genes involved in development, differentiation, and stem cell renewal strongly induced or upregulated by infection, including FoxA1, Spint1, Lin 28B, En2, Gabrq, Esx1, Trim71, Trp73, Cpne5, Cdh7, Cited 1, Pou4f1, and Jag2. The relevance of these genes, as well as others including Art3, Ugt8, and Trank1 to infection is not clear. We analyzed protein levels of several induced and upregulated transcripts whose relevance to MCMV infection is unknown (Figure 6) including the notch ligand Jagged 2, the homeobox-containing transcriptional factor Engrailed 2, and the E3 ubiquitin-protein ligase Trim 71. Protein levels of all proteins tested correlated with their transcript levels in infected BALB/c fibroblasts.
Interestingly, the top repressed and downregulated genes are primarily of unknown relevance to infection, though many are receptor or cell surface molecules (Npy6R, Rxfp, Mc2r, Cd200r3, Antxrl, Scara5, Il1r2, Agtr2, GPR165, the olfactory receptor genes, Olfr1314 and Olfr78, and the lectin or lectin-like genes Clec 3b and Reg3A). MCMV infection also caused repression or downregulation of noncoding (nc)RNAs including the small nucleolar RNA gene, Snord15A and genes of unknown function including 3 long intergenic noncoding RNAs (lincRNAs), the miscellanous RNA, 4930412B13Rik, and 2 antisense transcripts (Gm12963, Gm15883). To summarize, while many of the top upregulated genes are associated with host responses to infection, the function of many of the top downregulated and repressed genes during infection are obscure.
Genes and their products do not work in isolation but rather form pathways and networks. Even small perturbations in gene expression in a pathway can exert profound influences on eventual processes or functions. Therefore, we analyzed gene lists for shared common pathways. As expected, the top scoring gene networks for all differentially expressed (DE) genes included (i) infectious disease, antimicrobial response, inflammatory response (28 focus molecules); (ii) inflammatory response, cellular development, cell-mediated immune response (27 focus molecules) and (iii) cell morphology, hematological system development and function, inflammatory response (19 focus molecules) (Table S6A). These were also top networks when the subset of up- and down-regulated genes were evaluated (Table S6B). Also identified were networks associated with cell morphology and hematological system development and function. When this analysis was conducted with only induced and repressed genes, the top networks included cellular development, cell-mediated immune response, cellular function and maintenance, gene expression and embryonic development (Table S6C). The relationships among the molecules in top networks for differentially regulated and induced/repressed genes are shown in Figures Figure S6 and Figure S7. Thus, an unexpected outcome of this analysis is that MCMV infection influences a subset of networks controlling development..
The biological functions and/or diseases that were most significant to the molecules in the MCMV-regulated networks are shown in Figure 7A. Immunological disease, cardiovascular disease, genetic disorders, and skeletal and muscular disorders ranked as the top bio-functions connected with genes altered by MCMV infection. Among molecular and cellular functions, cell growth and proliferation were the top ranked perturbed functions, consistent with known effects of lytic MCMV infection of cells. Nervous system development and function is at the top of the list of physiological and developmental biofunctions, followed by organismal and tissue development and, surprisingly, behavior with 92 associated genes. DE genes were also evaluated for canonical pathways in the Ingenuity library (Figure 7B). The pathways most affected by MCMV included G-protein coupled receptor signaling followed by pathogenesis of multiple sclerosis and GABA receptor signaling. Together, these analyses point to known and expected consequences of infection at the cellular level (i.e., cell growth and proliferation, G-protein coupled receptor signaling) and physiological level (i.e. nervous system development) but also highlight unexpected cell and molecular functions, as well as physiological systems and disorders that may advance the understanding of CMV pathogenesis.
Gene ontology (GO) enrichment using GOrilla ranked lists analysis [31], [32] was also used to analyze DE genes. The full list of enriched GO terms long with associated genes is shown in Table S7. GOrilla analysis highlighted processes associated with upregulated genes including cell differentiation, neuron differentiation, regulation of ion transport, and the G-protein coupled receptor signaling pathway. Genes downregulated during MCMV infection were associated with many processes, including regulation of cell shape, adhesion, motility, and the extracellular matrix. Altogether, GOrilla analyses support results of the Ingenuity pathway analysis and suggest novel processes regulated in infected cells, notably suggesting that infection leads to a restructuring of the extracellular environment of the infected cells.
We report a comprehensive analysis of the MCMV transcriptome during lytic infection derived from cloning and sequencing of viral transcripts and next generation sequencing (RNA-Seq). By combining the approaches of RNA-Seq and traditional cDNA cloning as well as northern and RT-PCR analyses in certain complex regions, we were able to construct a comprehensive profile of viral and host transcription during lytic infection. We also investigated the host transcriptome using RNA-Seq combined with differential gene expression analysis, pathway analysis, and gene ontology analysis.
The major findings are as follows: 1) The MCMV transcriptome diverges substantially from that predicted by current annotation; 2) the identification of a novel viral protein specified by the MAT transcript indicates that this transcript functions as an mRNA and a non-coding RNA; 3) the majority of the most abundantly transcribed viral genes are of unknown function; and 4) the host response to infection includes regulation of many host genes and gene networks of unknown relevance to infection.
There are four major findings from the analysis of the MCMV transcriptome. First, we demonstrate novel transcripts of MCMV including novel splice variants, transcripts that map to noncoding regions, and transcripts overlapping multiple genes. Earlier, we reported similar novel transcripts of HCMV through analysis of a classical cDNA library [16]. This study revealed a dramatic increase in the complexity of viral gene products compared to currently available predictions and its findings were later on confirmed by RNA-Seq analysis [17]. A more recent analysis of HCMV translational products [18] by ribosomal footprinting identified over 700 translated ORFs – a strikingly high number compared to annotated genes. This discrepancy is, at least in part, a consequence of the polycistronic nature of HCMV transcripts which appear to code for many more ORFs than previously predicted (internal in frame or out-of-frame ORFs, uORFs) as well as ORFs coming from antisense or dedicated short transcripts. Our analysis demonstrated that the MCMV transcriptome is similarly complex: we identified several regions where multiple 3′ co-terminal transcripts expressed in different temporal phases are being transcribed. Such transcripts have the potential to code for truncated protein forms or even completely new proteins as described for HCMV, suggesting that the size and complexity of the MCMV proteome, like the MCMV transcriptome, is currently underestimated. Accumulation of ncRNAs is also a prominent feature of the cytomegalovirus transcriptomes. Our RNA-Seq analysis shows intense transcription in previously described stable MCMV introns and in intergenic regions, consistent with abundant ncRNAs reported for HCMV and MCMV [16], [17], [33]. These findings have a profound implication for understanding studies of CMV genes functions and underscore the need for transcriptomic maps in addition to genomic maps depicting only ORFs. The functions of many MCMV genes have been elucidated by using deletion mutants [34]. However in a transcriptionally complex region of the genome any deletion will likely impact multiple transcripts and possibly multiple proteins resulting in complex phenotypes.
In line with previous studies [13], we identified novel AS transcripts of MCMV. Interestingly, preliminary estimates in our cloning study indicate that AS transcripts occur at much lower frequency than reported for HCMV [16]. There are likely to be additional AS transcripts of MCMV. Because we did not capture every known sense transcript of MCMV, we may presume that the cDNA cloning study did not capture all AS transcripts. In addition, the RNA-Seq analysis performed in this study was limited by the fact that the methods employed did not provide strand-specific information and could not identify novel AS transcripts. AS transcripts, even those expressed at low levels, may possess noncoding RNA functions and contribute to complexity of the proteome as has described for HCMV [20]. Therefore, further studies are needed to determine the number and nature of AS transcripts derived from MCMV and will be critical to generating definitive transcriptome and proteome maps of this virus. The cDNA library analysis does suggest that the extent of MCMV AS transcription is lower than that described for other herpesviruses, including HCMV. These results are consistent with a strand-specific RNA-Seq experiment performed by Dölken group [15] that also show poor AS transcription in comparison to sense counterparts. Very little antisense transcription was also noted for the anguillid herpesvirus 1 (AngHV1) infecting eels [35], though extensive antisense transcription was reported for other herpesviruses, including KSHV and MHVγ68 [36], [37]. We conclude that different members of the Herpesviridae family differ in the extent of antisense transcription during lytic infection.
Second, we observed similar inconsistencies between transcriptomic data and gene annotation for MCMV as previously reported for HCMV [16]. These discrepancies can profoundly impact future studies related to the quantitative analyses of gene expression, interpretation of microarray studies, comparisons to newly sequenced virus strains, and studies using deletion mutant virus strains. The results presented here represent an important first step in re-annotation of the MCMV genome and underscore the utility of transcriptome studies in validating and refining genome annotation for microbial pathogens.
Third, analysis of the MCMV transcriptome revealed the striking abundance of the spliced MAT transcript. This gene is also largely conserved in wild isolates of MCMV (Alec Redwood, personal communication and [38]) and the protein is expressed by wild isolates tested in this study. MAT abundance may reflect its multiple functions. The 3′ untranslated region (UTR) of this transcript facilitates degradation of murine miR-27, establishing that this transcript functions as a noncoding RNA molecule [19], [20]. Members of the alpha, beta, and gamma herpes virus subfamilies all encode for abundant, largely enigmatic noncoding RNAs including the latency associated transcript (LAT) of herpes simplex virus (HSV), EBNA RNAs of Esptein-Barr virus (EBV), the β2.7 transcript of HCMV, the PAN RNA of Kaposi's sarcoma herpes virus (KSHV) and the HSUR transcripts of herpesvirus Samiri (HVS) which also downregulates the cellular miR27 (reviewed in [39]). In addition to the noncoding function of the MAT, we demonstrate that this transcript also encodes for a novel small protein of approximately 17 kDa. To our knowledge, this is the first herpes virus transcript we know of that functions both as a noncoding RNA, and an mRNA that specifies a novel viral protein.
Fourth, a somewhat startling finding from the quantitative RNA-Seq analysis was that after MAT, the most abundant viral transcripts in infected cells are derived from genes without known functions, including M116. We report that M116 is a novel spliced transcript predicted to specify a much smaller protein compared to current annotation. These results highlight fundamental gaps in our understanding of basic MCMV biology.
We found that the cDNA library and RNA-Seq approaches yielded remarkably complementary data including identification of novel transcripts and new insights into transcript abundance, despite different biases in each of these methods. For example, while there may be selection bias for isolating transcripts with long tracts of adenosines during cDNA library construction [16], GC content, bias in the sites of fragmentation, primer affinity and transcript-end effects may influence RNA-Seq results [40]. Future RNA-Seq studies may also facilitate novel gene identification as RNA-Seq has now been applied to ab initio reconstruction of gene structure [41] using only RNA-Seq data and the genome sequence. However, currently available algorithms are unable to cope with highly dense genomes, such as MCMV and other viral genomes. Until such tools are developed for very dense genomes, RNA-Seq data relies upon comparison to existing gene annotation and other experimental methods for gene structure prediction. In this study, we compared RNA-Seq to currently used annotations but also to the cDNA library study, northern analysis, and RT-PCR studies to identify and validate numerous novel transcripts.
We also report that lytic infection elicits a profound cellular response in fibroblasts. This study identified 10,748 differentially regulated genes. As the number of mouse genes is estimated to be 33,207 [42] we estimate that over 31% of mouse genes are altered in response to infection. Many of the top upregulated and induced genes and gene networks were associated with immune responses to infection, including interferon and interferon-inducible genes such as phyin1, a potential activator of p53 [43], the inflammasone regulator Gpb5 [44] and Rsad2 (a.k.a. viperin), also known to be induced by HCMV [45].
Inflammatory chemokine ligand genes are also highly upregulated during infection. MCMV encodes virally-derived chemokine homologs specified by the m131/m129 genes [46], [47] and at least one chemokine receptor homolog, M33 [48]. Numerous host chemokine receptors are also upregulated by infection, suggesting a remarkably complex interplay between MCMV-derived and host derived chemokine signaling during infection. Induction of inflammatory gene networks by MCMV also lends credence to the hypothesis that inflammatory responses link CMV infection to chronic diseases, such as chronic allograft rejection, cardiovascular disease, and cancer [2], [4], [5].
Numerous transcription factors are also induced or upregulated by infection including insulinoma-associated 1 (Insm1). Recently, Insm1 was found to be strongly upregulated by HSV-1 infection and shown to promote HSV gene expression, probably by binding the HSV infected cell protein (ICP)0 promoter. [49]. This raises the intriguing possibility that INSM1 plays a similar role in promoting virus gene expression during MCMV infection. Another induced transcription factor induced at the transcript and protein level is engrailed-2 (EN2). This transcription factor is key to patterning cerebellar foliation during development [50]. We previously described a profound dysregulation of cerebellar development in brains of neonatal mice infected with MCMV [51], suggesting a possible physiological link to regulation of this gene. Another top induced gene was the GABA receptor, Gabrq. Glutamate receptor signaling was also identified as significantly impacted canonical pathways in our dataset. In the developing brain GABA and glutamate receptors influence neuronal proliferation, migration, differentiation or survival processes [52]. Whether and how these observations relate to our previous findings that MCMV infection of neonates results in decreased granular neuron proliferation and migration [51] are important areas for future study and may impact our understanding of neurological damage and sequelae associated with HCMV in congenitally-infected infants.
Perhaps most importantly, many top regulated genes, especially downregulated and repressed genes, are associated with functions whose roles in infection are obscure, including many genes of unknown function. Many downregulated or repressed genes are cell surface molecules, host lincRNAs, antisense RNAs, or small nucleolar RNAs. Regulation of lincRNAs was recently observed during infection with severe acute respiratory syndrome coronavirus (SARS-CoV) and influenza virus, and have been suggested to impact host defenses and innate immunity [53]. Further studies to identify the functions of these downregulated and repressed genes and noncoding RNAs during MCMV infection may well provide novel insights into the virus-host molecular interface as well as possible therapeutic targets.
This analysis also revealed immunological disease, cardiovascular disease, genetic disorders and skeletal and muscular disorders as top bio-functions connected with genes altered by MCMV infection. While MCMV involvement in cardiovascular disease is a subject of intensive research, potential involvement in skeletal and muscular disorders is not well documented but may be relevant to the novel observation that MCMV infection of mice with a heterozygous Trp53 mutation develop rhabdomyosarcomas at high frequency [54].
A primary caveat of RNA-Seq analysis is determining whether changes in gene transcript levels are also reflected at the protein level. This is particularly important as herpesviruses can control protein accumulation at the post-transcriptional, translational, and post-translational levels [55]–[57]. We confirmed that the notch ligand, Jagged 2, is highly upregulated by infection at both the transcript and protein level. Notch signaling is a highly conserved signaling pathway that plays important roles in development, including neurogenesis and differentiation of immune cell subsets [58]. Jagged 2 is also upregulated by the alphaherpesviruses, HSV-1 and Psuedorabies viruses [59]. KSHV and EBV also exploit the notch signaling pathway to facilitate aspects of their life cycle [60] and notch signaling is proposed to influence HSV-2-induced interferon responses [61]. We show for the first time that the betaherpesvirus, MCMV, also influences notch signaling. Dysregulation of Jagged2 as a consequence of MCMV infection is highly interesting since it plays a role in important processes affected by CMV including inner ear development [62], [63], generation of motor neurons [64] and differentiation of immune cell subsets [65], [66].
To summarize, this study has refined the understanding of MCMV gene expression and identified new areas of research to advance our understanding of the host response to these ancient viruses. We describe what is to our knowledge, the first herpes virus transcript that functions as both a noncoding RNA that limits accumulation of cellular miRNAs, and an mRNA that specifies a protein. This study also revealed novel features of the host response to infection. Perhaps most importantly, this study identified many virus and host genes of unknown function that are regulated during infection. It is highly likely that further study of these genes may lead to breakthroughs in the understanding and treatment of cytomegalovirus-related diseases.
Primary mouse embryonic fibroblasts (MEFs) from BALB/c or Balb.K mice were prepared and maintained as described [67] and used between passages 3–8. Immortalized murine BALB.K MEFs, (MEF.K) [68] and SVEC4-10 (ATCC CRL-2181) were used for immunoblot studies. MCMV Smith strain (ATCC VR-1399) was propagated and titrated on primary Balb/c MEF by standard plaque assay as described in detail in [69]. Wild type MCMV isolates K181 (GenBank acc no: AM886412.1), C4D, K6 and WP15B (GenBank acc no: EU579860.1) [38] were a kind gift from A. Redwood (University of Western Australia, Australia). Construction of the Δ7S3, Δm167, Δm168, Δm169, Δm170, and m169-mut mutant viruses were previously described and were generated by ET-cloning [70] using the full-length MCMV BAC pSM3fr [71]. The double deletion mutants (Δm168Δm169 and Δm169Δm170) were constructed exactly as described previously [20]. Primers for construction of the double deletion mutants are also as described [20] using the forward primer for the first gene and reverse primer for the second gene. All infections were conducted by exposing cells to 0.3 PFU/cell followed by centrifugal enhancement for 30 minutes at 800 g, as described in [69].
Smith MCMV infected cBalb/c MEFs were harvested 72 h post infection and viral DNA was isolated as described [16].
RNA was extracted from Smith MCMV infected Balb/c MEF at 4, 8 and 12 hrs after infection (IE library); 16, 24 and 32 hrs after infection (E library); and 40, 60 and 80 hrs after infection (L library). No drug was used to select for different temporal classes of transcripts and equal amounts of RNA from each time point were pooled prior to library construction. cDNA libraries were generated as described previously for HCMV [16] by following the instruction manual for the SuperScript Plasmid System with Gateway Technology for cDNA Synthesis and Cloning (Invitrogen) with some minor modifications. Briefly, total RNA was isolated using the TRIZOL Reagent (Invitrogen, CA, USA). A poly(T)-tailed PacI primer-adapter was used for first-strand cDNA synthesis (5′-GCGGCCGCTTAATTAACC(T)15-3′). After second-strand synthesis, an EcoRI-PmeI adapter was added to the 5′ end and cDNAs were cleaved with EcoRI and PacI. The EcoRI-PmeI adapter was generated by annealing following oligonucleotides: 5′-AATTCCCGCGGGTTTAAACG-3′ and 5′-Pho-CGTTTAAACCCGCGGG-3′. cDNA fragments were inserted into a modified pcDNA3.1(+) previously digested with EcoRI and PacI and transformed into XL1-Blue Supercompetent E. coli cells (Stratagene, CA, USA).
Positive selection of viral cDNA clones was performed as described previously [16]. Mse I-digested genomic MCMV DNA was labeled using a DIG High Prime DNA Labeling Detection Starter Kit II (Roche Applied Science) and used to identify virally-derived cDNA clones. Plasmids harboring cDNA clones that reacted with probe were isolated and sequenced from the 5′ end using T7 primer for pcDNA3.1(+) or the 3′ ends using primer (5′GCACCTTCCAGGGTCAAGGAAG) or standard poly (T) primers at the OSU Plant-Microbe Genomics Facility. Sequences were compared to the MCMV Smith strain genome [GenBank accession no. NC_004065] using mega BLAST.
Total RNA was extracted from Balb/c MEF cells cultured in 100 mm2 petri dishes and exposed to 0.3 PFU/cell of the MW 97.01 strain of murine cytomegalovirus or mock-infected. At 4, 8, 12, 16, 24, 32, 40, 60 and 80 hours after infection, RNA was isolated using TRIZOL Reagent. RNA integrity was assessed on Agilent Bioanalyzer and only samples with RNA index values of at least 9 were used. Equal amounts of RNA from each time point were pooled (0.3 µg of RNA per time point) and treated with DNaseI. Libraries were prepared with Illumina TruSeq RNA kit according to manufacturer's instructions and sequenced on Illumina Genome Analyzer IIx as single-end 36 bp reads. The Illumina TruSeq RNA kit employed does not allow for strand-specific information to be derived from the sequence data. Datasets are available at the National Center for Biotechnology Information (NCBI) Sequence Read Archive (SRA) accession no. SRR953479 (sequence reads from MCMV-infected MEFs) and accession no. SRR953859 (sequence reads from mock-infected MEFs).
Reads were aligned to mouse (NCBI37/mm9 assembly) and MCMV genome (GenBank acc.no. NC_004065.1) using ELAND aligner or Bowtie aligner (for comparison with data provided by Lars Dölken). It is important to note that both ELAND and Bowtie aligners do not map splice junctions and thus give concordant results. Alignments were visualized using Integrative Genomics Viewer (http://www.broadinstitute.org/igv/) [72]. Differential gene expression was assessed by calculating RPKM (reads per kilobase of million mapped reads (RPKM) using SAMMate 2.6.1. release with EdgeR (http://sammate.sourceforge.net/) [73]. Gene ontology (GO) enrichment analysis was performed on filtered lists of differentially expressed genes (p<0.05) using GOrilla ranked lists analysis [31], [32]. Ingenuity Core Analysis (Ingenuity Systems, www.ingenuity.com) was used for gene interaction network and canonical pathway analysis. Gene lists were filtered for statistically significant differentially expressed genes (p<0.05) and a fold change cutoff of 2 was set to identify molecules whose expression was significantly differentially regulated. For network generation, these molecules (Network Eligible molecules), were overlaid onto a global molecular network developed from information in the Ingenuity Knowledge Base based on their connectivity. The Functional Analysis of a network identified the biological functions and/or diseases that were most significant to the molecules in the network. Right-tailed Fisher's exact test was used to confirm that biological functions and/or disease assigned to data sets were not due to chance. The nature of individual DE genes was also investigated using the Mouse Genome Informatics databases (http://www.informatics.jax.org/) [74] and Entrez Gene (http://www.ncbi.nlm.nih.gov/gene) [75].
RNA was isolated using Trizol reagent from mock or MCMV-infected Balb/c MEF at 24 hours (MAT) or 10, 30 and 60 hrs after infection. RNA (1 µg/lane or 10 µg/lane (MAT)) was separated by formaldehyde agarose gel electrophoresis and transferred to positively charged nylon membrane and crosslinked by UV irradiation. Membranes were reacted to DIG-labeled probes overnight at 67°C. For MAT detection, a DIG labeled double-stranded DNA probe was made using fragments corresponding to the MAT gene sequences derived from cDNA library clones E1, E125 and E134 using Roche's DIG-High Prime DNA Labeling and Detection Starter Kit I. For all other northern blots, single-stranded DIG-labeled RNA probes were used generated using Roche's DIG Northern Starter Kit. Antisense probes were generated by in vitro transcription from T7 promoter present in pcDNA3.1 plasmids containing cDNA clones that harbor the desired gene fragments (Table 1). Therefore antisense probes are identical to transcripts cloned in cDNA library and can detect transcripts antisense to cDNA clones. To generate sense probes, T3 promoter was added to 5′ end of complimentary strand of the gene fragments used for antisense probes by PCR (Table 1). The PCR fragments were then in vitro transcribed and DIG labeled using T3 RNA polymerase. Care was taken to generate sense probes of length comparable to corresponding antisense probes.
The m169 gene sequence was amplified by PCR using viral DNA isolated from MCMV BAC pSM3fr using following primers: F: 5′-TTTTTGGATCCATGAGCAACGCGGTCCCGTTC-3′ and R: 5′- TTTTTCTGCAGTCATCACGGGGGGCACCTACC-3′, reacted with BamHI and HindIII (New England Biolabs), inserted into pQE30 expression vector and introduced to E.coli Bl21 pREP4 strain (Qiagen). The protein was induced according to manufacturers' instructions and purified on a His-tag column. Purified protein was used for immunization of Balb/c mice and antibody titer in blood serum was measured by ELISA. When antibody titer in serum reached adequate levels, animals were sacrificed, their spleens isolated and fused with SP2/O cells. Supernatants from motherwells were tested by immunoblot blot on purified MAT protein and positive wells were rescreened by immunoblot using lysates from MEFs infected with WT, Δ7S3, Δ168-169, Δ169-170, Δm168, Δm169 and Δm170 mutants as described below.
RNA from Mock- or MCMV-infected cells isolated for northern blots was also reverse transcribed using oligo-dT primers (ProtoScript M-MuLV Taq RT-PCR Kit, New England BioLabs). No reverse transcriptase (-RT) controls were run in parallel. Splicing was then verified by PCR amplification using primers that flank putative introns (M116; F: CTTCATCGGATTCGGAGGC; R: TGTTGTTGTCGACGTCTGATGTG; m71–m75; F: ATCTCCTCTGCCTCCGACCTC, R: CGATGTCATCTTGGAATCCGACGA; m72–m75; F: CCGGATACGACCGTCAGC, R: CGATGTCATCTTGGAATCCGACGA) using Phusion high fidelity polymerase (New England BioLabs).
Mock-infected or MCMV-infected primary MEFs, or murine cell lines (MEF.K, SVEC4-10) were lysed in RIPA buffer. Protein lysates were separated by SDS-PAGE and transferred to PVDF. MAT protein was detected with anti-m169 antibody described above, Jag2 with antibody N-19 (Santa Cruz), Engrailed 2 with En2 PA5-14363 antibody (Thermo Scientific), Trim71 with PA5-19282 (Thermo Scientific), and actin with antibody C4 (Millipore) followed by peroxidase-labeled secondary antibodies (Jackson ImmunoResearch or Abcam). Proteins were visualized using Amersham ECL Prime Western blotting reagent (GE Healthcare) and quantified using ImageJ software (http://rsbweb.nih.gov/ij/).
|
10.1371/journal.ppat.1005979 | Pheromone Recognition and Selectivity by ComR Proteins among Streptococcus Species | Natural transformation, or competence, is an ability inherent to bacteria for the uptake of extracellular DNA. This process is central to bacterial evolution and allows for the rapid acquirement of new traits, such as antibiotic resistance in pathogenic microorganisms. For the Gram-positive bacteria genus Streptococcus, genes required for competence are under the regulation of quorum sensing (QS) mediated by peptide pheromones. One such system, ComRS, consists of a peptide (ComS) that is processed (XIP), secreted, and later imported into the cytoplasm, where it binds and activates the transcription factor ComR. ComR then engages in a positive feedback loop for the expression of ComS and the alternative sigma-factor SigX. Although ComRS are present in the majority of Streptococcus species, the sequence of both ComS/XIP and ComR diverge significantly, suggesting a mechanism for species-specific communication. To study possible cross-talk between streptococcal species in the regulation of competence, and to explore in detail the molecular interaction between ComR and XIP we undertook an interdisciplinary approach. We developed a ‘test-bed’ assay to measure the activity of different ComR proteins in response to cognate and heterologous XIP peptides in vivo, revealing distinct ComR classes of strict, intermediate, and promiscuous specificity among species. We then solved an X-ray crystal structure of ComR from S. suis to further understand the interaction with XIP and to search for structural features in ComR proteins that may explain XIP recognition. Using the structure as a guide, we probed the apo conformation of the XIP-binding pocket by site-directed mutagenesis, both in test-bed cultures and biochemically in vitro. In alignments with ComR proteins from other species, we find that the pocket is lined by a variable and a conserved face, where residues of the conserved face contribute to ligand binding and the variable face discriminate among XIP peptides. Together, our results not only provide a model for XIP recognition and specificity, but also allow for the prediction of novel XIP peptides that induce ComR activity.
| Bacteria transmit chemical signals to each other in a process known as quorum sensing. This adaptation is central to pathogenesis as it allows bacteria to coordinate a group response, such as when to form a biofilm, secrete virulence factors, or uptake new DNA. Quorum sensing in many species of Gram-positive bacteria utilizes secreted short peptide signals that are sensed by protein receptors that alter gene expression. The ComR-ComS receptor-signal pair is a regulator of competence for genetic transformation that is broadly conserved in a majority of Streptococcus species, including those pathogenic in humans and animals. Despite this conservation, we observe sequence diversity among ComRS orthologs that raises questions relating to sensory specificity and the molecular mechanism of peptide recognition. To address this, we directly tested the possibility for signaling cross-talk and identified three general categories of ComR receptors displaying strict, intermediate, and promiscuous responses to heterologous peptides. To elucidate the molecular mechanism of receptor specificity we determined an X-ray crystal structure of ComR from S. suis. We observe a conserved face of the ligand pocket important for peptide binding and a variable face that functions in peptide specificity. Finally, basic criteria necessary for peptide responses were used to redesign active peptides from inactive templates.
| Competence, or the ability of bacteria to import extracellular DNA, is a trait widely conserved across both Gram-negative and Gram-positive bacteria. This natural transformation process allows bacteria to acquire new genes that increase genetic diversity and fitness, such as the gain of antibiotic-resistance determinants [1]. Common among bacteria that are able to undergo natural transformation is a minimal set of genes enabling transport and integration of DNA by homologous recombination. Broad conservation of these genes, even among bacteria for which natural transformation has not been demonstrated in laboratory settings [2] provides a mechanistic explanation for the widespread evidence of horizontal gene transfer in bacteria, and strongly suggests their evolutionary importance.
Within the genus Streptococcus, incorporating both pathogenic and commensal species to humans and animals, and those useful in industrial applications [3], the core competence genes are tightly regulated and their expression is coordinated by quorum sensing (QS). Generally, QS is utilized by both Gram-positive and Gram-negative bacteria to organize a community-wide response to environmental stimuli. For example, pathogens use QS to time the expression and secretion of virulence factors during pathogenesis [4, 5]. For most QS signaling, a small molecule such as an N-acyl homoserine lactone [6] or a peptide [7] is used as the secreted chemical messenger, although inter-bacterial communication can also be accomplished though the interaction of secreted proteins [8]. Competence in Streptococcus spp, is regulated either by the ComCDE system [9] or by the subject of our study, the ComRS pathway [10].
The ComCDE pathway is utilized by members of the Mitis and Anginosus Streptococcus groups and its peptide pheromone is sensed by the bacteria through a two-component signal transduction pathway [9, 11]. In contrast, all other groups of Streptococcus (Pyogenic, Mutans, Bovis, and Salivarius) employ the ComRS system, in which the pheromone is imported into the cytoplasm to bind the transcriptional regulator directly. Although the details of these QS pathways differ, both induce the expression of sigX, the Streptococcus alternative sigma-factor and its regulon [7]. It is SigX that subsequently initiates transcription of the competence genes required for the incorporation of the newly-acquired DNA [12].
The gene products of comRS comprise the pheromone precursor and the receptor of the pheromone. In Streptococcus mutans, ComS is a 17 amino acid pre-peptide that undergoes proteolytic maturation and secretion from the cell by an uncharacterized pathway to produce the active 7 aa pheromone XIP (sigX-inducing peptide) [13, 14]. To be sensed by bacteria, XIP must reach the cytoplasm, and does so by way of generalized peptide transporters of the oligopeptide permease (Opp) family [10, 15]. Once in the cytoplasm, XIP directly binds to and activates ComR [10].
ComR proteins are members of the RNPP (Rap/NprR/PrgX/PlcR) family of QS regulators, including Rgg, which are prevalent throughout the Lactobacillales order [16, 17]. Members of this family share several biochemical and structural similarities, most importantly the ability to bind peptide pheromones to govern protein activity. However, the exact mechanisms by which they respond to bound pheromone can differ significantly. For example, NprR undergoes oligomerization from a dimer to a tetramer in response to peptide binding [18], whereas Rgg proteins are thought to remain as dimers in both apo- and ligand-bound forms [19]. ComR is predicted to contain an N-terminal helix-turn-helix (HTH) DNA binding domain (DBD) and a C-terminal tetratricopeptide repeat (TPR) domain, which contains the site of pheromone binding [19, 20]. Once activated, the ComR/XIP complex binds at the promoter regions of comS and sigX, thus producing a pheromone-dependent positive-feedback regulatory loop and activating the SigX regulon [10, 21].
There are three general classes of ComRS pathways, type-I (Salivarius), type-II (Bovis, Mutans, Pyogenic), and type-III (Suis) [10, 22]-[23]. The classification is based upon observed differences in both the comS/sigX promoters and XIP sequences, which in turn correspond with phylogenetic groupings (Fig 1). For example, the type-I and type-II comS/sigX promoter sequences differ within inverted repeats, at which ComR/XIP binds [10]. Additionally, the type-II comS genes encode XIP peptides containing a C-terminal WW-motif that is required for ComS and SigX activation in S. mutans UA159 [10], whereas this WW-motif is not found in type-I XIPs [22]. The type-III ComS contains two tryptophans similar to the type-II peptides, but differs in that the WW-motif is split by two residues [23]. As type-II ComRs represent the larger sample group and include pathogenic species, we chose to focus our study in these systems. Despite this classification and the broad conservation of type-II ComRS genes, there are significant sequence variations among both the XIP and ComR sequences. This observation and recent work suggesting communication between Streptococcus species [24] led us to investigate roles of these variations in intercellular communication, in particular XIP recognition.
To address the effect of these sequence varieties, we developed a ‘test-bed’ assay using S. mutans to measure the response of different type-II ComR proteins to various XIPs in bacterial culture. Not only did this reveal distinct degrees of specificity for XIP, but it also allowed demonstration of active ComRS systems that were only previously predicted. We also expanded this analysis to include the type-III representative, S. suis, to broaden our exploration of signaling peptide recognition. These experiments provided an initial survey of streptococcal species that are amenable to inter-species induction of competence. Additionally, we obtained an X-ray crystal structure of the ComR from S. suis and used this as a guide to probe the structural features of the apo conformation of the XIP binding-site by site-directed mutagenesis, both in vivo by the test bed assay and directly in vitro biochemically. The combined biochemical and structural data provide a model for how ComR discriminates between cognate and foreign pheromones for activation. Finally, we use our combined functional and structural data to predict and create new active peptide-pheromones.
The entire known SigX regulon, as well as the proximal transcriptional regulator of sigX, ComR, is conserved throughout the Mutans, Bovis, and Pyogenic groups (Fig 1, top) [10]. Additionally, within these groups, the mature form of the peptide pheromone ComS contains a conserved double-tryptophan motif (‘WW’) (Fig 1, bottom) [10]. This led us to ask whether the putative ComRs encoded by members of these groups could respond to cognate and heterologous XIPs to activate comS and sigX transcription. To test this possibility, we developed a strain that would allow for control of the ComR and XIP variants present within the system. Previous findings showed that an S. mutans UA159 ΔcomS strain robustly activated luciferase genes expressed from the sigX promoter (PsigX::luxAB) upon addition of synthetic cognate XIP. Additionally, when a multi-copy plasmid carrying wild-type comR was provided to a ΔcomR strain, the ability to naturally transform exogenous genomic DNA was rescued [10]. With these observations in mind, we created an S. mutans UA159 ‘test bed’ strain, ES1, in which the endogenous comRS locus was replaced with a spectinomycin-resistance cassette (Fig 2A). Because this strain contains neither comS nor comR, it is unable to produce or respond to XIP pheromones. A plasmid was used as a vehicle to provide different comR variants, under control of the native promoter, with adjoining comR-comS intergenic regions containing ComR binding sites and the comS promoter. The plasmid also contains the bacterial luciferase genes, luxAB, under transcriptional control of the cloned comS promoter. Clones harboring each comR-variant plasmid were constructed, grown, and treated with titrations of synthetic XIP peptides. Since productive interaction between ComR and XIP results in transcriptional activation of the comS promoter, we could test the functionality of ComR–XIP interactions by assessing the specific bioluminescent activity of each culture. To ensure that the test bed recapitulated observed results seen in wild-type S. mutans UA159, we assessed the UA159 ComR with serial dilutions of the cognate XIP-C7 (the C-terminal seven amino acids of ComS) and measured bioluminescence levels (Fig 2B). The observed EC50 value (effective concentration of XIP eliciting 50% of the log maximum luminescence) was 0.2 nM, with a maximal induction of 140-fold at 2 nM XIP (Fig 2B). These results confirmed that episomal expression of comR and exogenous provision of XIP in the test bed recapitulated reported observations in wild-type S. mutans [10, 25, 26].
To determine if other conserved ComRS pathways selected from the Bovis and Pyogenic groups also demonstrated ComR and XIP-dependent activity similar to S. mutans UA159, we developed a series of plasmids, each containing the analogous segment of DNA. These reporter plasmids were transferred into the ES1 background, allowing us to test whether the encoded ComR could activate the cognate PcomS in response to added synthetic cognate XIP. As XIP peptide length may vary between ComRS systems, we synthesized and assessed two or three length variants for most species’ putative XIPs to determine the optimal peptide size needed to activate a given ComR. Each ComR tested exhibited PcomS reporter activity with at least one length-variant of the putative XIP (Fig 2D and S1 Fig).
Among all strains tested, the two Bovis representatives, strains SboB and 83, provided the most similar ComR alleles (82.6% identity) and XIP sequences (differing only in Ala or Ser at position 5 from the C-terminus, C-5). A roughly equal response to both S. bovis XIP-C7 variants was observed for both, and each displayed the lowest EC50 values recorded in our dataset (<<1 nM) (Fig 2C and 2D).
Among Pyogenic species, all five of the ComR proteins tested activated the cognate comS promoter, but to different maximal levels of activity and with variable sensitivities to XIP. Two Group B Streptococcus (GBS) ComRs (Streptococcus agalactiae 2603 v/r and A909) exhibited the strongest induction in reporter activity among the Pyogenics, with maximal activity observed with the C7 length of XIP. The ComRS alleles from Group A Streptococcus (GAS, S. pyogenes), MGAS5005 (a serotype M1 strain) and MGAS315 (serotype M3), showed slightly different activation profiles. The MGAS5005 ComR allele activated PcomS::luxAB transcription at an EC50 value of approximately 250 nM with the C8 cognate XIP. However, upon treating the MGAS5005 reporter strain with 50 μM of the MGAS5005 C8 XIP, PcomS::luxAB transcription increased approximately 25-fold while the MGAS315 C8 XIP elicited only 25% of the maximum transcription level at the same concentration (S1D Fig). One representative of the Lancefield G antigen group, S. dysgalactiae subsp. equisimilis strain ATCC35666, was included in our studies and responded best to the C7 cognate XIP, with an EC50 value of 15 μM and displayed ~10-fold increase in reporter activity at 50 μM XIP.
Additionally, we included S. suis in our analysis to expand the test-set to include more distantly related ComRS alleles. Unlike the type-I and type-II classifications, the S. suis XIP has a split aromatic motif with an acidic C-terminal side chain, in stark contrast to an aliphatic residue or glycine (Fig 1). However, ComR S. suis behaved similarly to the type-II ComRs in our transcriptional assay as it responded with the highest activity to the cognate C7 peptide with a EC50 of 93 nM (Fig 2C). Taken together, these results show that heterologous expression of comR alleles provides the capacity to implement pheromone responses, and they suggest that each comRS system provides quorum sensing functionality, albeit with potentially varying sensitivity, in their respective species.
Although each XIP contains conserved aromatic residues, there is significant sequence variation of signaling peptides between species (Fig 1). Given this, we suspected that there might be structural aspects of the pheromones that denote specificity for their cognate ComR. To search for a possible pattern of substrate specificity, each ComR in our study was stimulated with non-cognate XIP peptides using the test-bed assay. Nine comR alleles were expressed in the S. mutans test bed strain, treated with titrations of each length-variant of XIP, and luciferase values were determined (S1 Fig). From these data, distinct activity patterns using heterologous XIPs were observed, leading to categorization of ComR proteins as “strict”, “intermediate”, and “promiscuous” (Fig 3).
The strict, or high-specificity group includes ComR alleles from S. mutans UA159 and S. suis 05ZYH33, both of which only exhibited detectable PcomS reporter activity upon addition of cognate XIP. The intermediate-specificity group included S. pyogenes MGAS315 and S. agalactiae strains 2603 and A909. These alleles displayed detectable reporter activities from one to three non-cognate XIP types. The promiscuous, or low-specificity group included ComR alleles capable of activation by a wide range of XIP variants, and included S. pyogenes MGAS5005, S. dysgalactiae subsp. equisimilis, and both S. bovis alleles. Each of these responded to at least five different XIP types, and S. bovis was found to have measurable responses to all eight other XIP peptides tested including the most disparate peptide derived from S. suis (Fig 3 and S1 Fig).
To better understand the protein-pheromone interaction, we also pursued the structure of ComR. Although a structure for a type-II ComR proved elusive, ComR S. suis was of interest as it is part of the strict response group and recognizes a unique XIP. Furthermore, as ComR proteins are homologous, S. suis will provide a template for the accurate modeling of the other orthologs in our study. The native crystal structure of the apo-form of ComR S. suis reveals a dimerized α-helical protein in the asymmetric unit that resembles related Rgg-proteins [19] (Fig 4A, Table 1). However, the construct used in this study behaves mostly as a monomer in solution when assayed by size-exclusion chromatography (S2 Fig). As the active form of Rgg2 shows a domain swap of the DBD domains (Fig 4A) [19], the observed conformation of the ComR apo structure appears incompatible with DNA binding and would require reorientation of the DBD domains. This is reinforced by the observation that the DBD of ComR packs tightly against the TPR using α-helices 3 and 4 (DBD) and α-helices 8 and 9 (TPR) to form a significant interface region (Fig 4B). The observed dimer interface between the TPR domains appears to help promote crystal-packing in this structure and the two monomers exhibit an identical conformation (RMSD = 0.192 Å2). The monomer is shown in Fig 4B, revealing that ComR S. suis is comprised of 16 α-helices divided into two domains. Helices 1–5 (residues 7–72) constitute the highly conserved DNA-binding-domain (DBD) and helices 6–16 (residues 79–304) comprise the tetratricopeptide repeat (TPR) domain that is the predicted site of interaction with XIP. These two domains are separated by a 6 aa linker region. It is also important to note that the space group of the selenomethionine derivative compared to the native crystal resulted in a different packing arrangement with only one monomer per asymmetric unit versus two monomers for the native crystal. Due to these packing differences, clear electron density for α-helices 15 and 16 of the TPR domain is only visible in the native data set (S3 Fig). Furthermore, the derivative had other areas of discontinuous density leading to an incomplete model and a very poor overall refinement. As several areas of weak density leaves doubts about the model and the conformations of α-helices 15 and 16, the derivative structure has been omitted despite the slightly better resolution of the dataset.
Although this general architecture of the monomer is observed in a related Rgg protein [19] and other RNPP family members such as PlcR [17, 19, 27], ComR S. suis differs from them in several ways, foremost in that the apo-form of the protein in solution is a monomer, not a dimer (Fig 4 and S2 Fig). ComR S. suis also lacks the disulfide bond observed in the DBD of Rgg2 that is thought to help facilitate its stable dimeric assembly and orient these domains for interaction with DNA [19]. Furthermore, a comparison of the ComR TPR to the Rgg2 TPR reveals several differences (Fig 4C), indicating that the activation mechanism likely differs significantly from Rgg proteins. Overall, the size and shape of the ligand binding pocket is altered due to a re-arrangement of several of the helices. Notably, α7 extends into the putative peptide binding site and helices α6–8 are rotated by 45° relative to Rgg2. This and the additional length of α7 in ComR allows for an ~11Å extension to contact helices α14 and α16. Finally, the C-terminal capping helix (CAP) [19] of ComR S. suis is translated outward by ~7Å relative to Rgg2. These differences may determine substrate specificity. For example, the configuration of ComR would block the binding of Cyclosporin A, an inhibitor of Rgg2 [19, 28].
To understand how ComR S. suis recognizes its peptide and how ComR proteins generally may discriminate or accommodate heterologous XIP substrates, the molecular surface of the apo conformation of ComR was examined in detail (Fig 5). ComR from strains used in this study were aligned with Clustal Omega [29] and submitted to the Consurf server [30, 31] to plot areas of sequence conservation. As shown in Fig 5A and 5D, although the DBD is highly conserved, the TPRs diverge significantly so that only one side of the putative XIP binding-site pocket exhibits sequence conservation among the ComR species in this study. This creates a “conserved face” and “variable face” on opposing surfaces of the cavity for potential interaction with peptide pheromones. Residues that show strict conservation line the outer edge of one side of the pocket, whereas the inside of the pocket shows some minor variation. The variable face likewise lines one-half of the binding surface and includes a deep hydrophobic groove in the unbound state of the pocket, at least in S. suis (Fig 5A and 5B). Based on this observation we hypothesize that the conserved face of the XIP interface contains elements required for peptide binding and that the variable face acts in substrate specificity. However, these activities are likely not mutually exclusive, as suggested by the electrostatic potential of ComR S. suis generated by the Adaptive Poisson-Boltzman Solver software (APBS) [32, 33] (Fig 5C). The homologous residues of the inner pocket are electropositive (R103, R170, and K250) which is likely required for optimal interaction with the C-terminal glutamic acid moiety of the ComR S. suis XIP (Fig 1).
When considering general residue conservation, the DBD-TPR domain interface is also a site of interest (Fig 5A inlay). All residues at the domain interface are conserved or homologous, polar, and create an extensive network of hydrogen bonds and salt-bridges that stabilize the observed conformation, showing that this is not a crystal-packing artifact. Indeed, relative to Rgg2 the DBD of S. suis appears to be held in an inactive conformation to occlude helices α3 and α4 from interacting with DNA and possibly inhibiting dimerization.
Given the specificity of various ComRs to heterologous XIPs (S1 Fig), we also performed a molecular surface comparison between strict and promiscuous ComR proteins to look for common or differing structural features that may correlate to these activities. All proteins used in this study were modeled using the program Modeller with apo-ComR S. suis as a template [34] and their surface properties explored, with four representative structures selected to represent the analysis (S. suis, S. mutans, MGAS5005, and S. bovis 83) (S4 Fig). We observe that strict ComR proteins appear to have a significantly more electropositive pocket than ComRs that can recognize multiple pheromones (S4A Fig), but that the overall hydrophobicity of the pocket does not seem to differ between recognition types (S4B Fig). We examined the size and shape of the apo-pocket using the 3V webserver [35], which suggests that the promiscuous ComR binding pockets maybe smaller and shallower, with slightly more aliphatic character (S4C Fig). However, it is important to note that the difference in pocket size is not extremely pronounced and thus likely falls within the range of the error of the homology model.
As we observed possible trends in the overall molecular features of the pheromone binding site, we examined the variable face for specific features that may be common to strict or promiscuous ComR proteins. As shown in Fig 6A, there is a residue consensus for the strict ComR proteins that occurs at the lower end of the pocket. Positions 226, 228, 264 and 265 are identical between S. suis and S. mutans, but variable in the promiscuous examples. This is in contrast to residues in the middle of the variable face, which are unconserved (Fig 5A). However, residues on the conserved face do show some variability, although they are homologous and preserve the general side chain characteristics. For example, both strict ComR proteins have lysine at position 250 as opposed to arginine for the promiscuous examples (Fig 6A), but it is unlikely that this subtle difference affects peptide binding.
Recent work comparing RNPP family structures has shown that the presence of an Asn residue in the XIP binding pocket appears to be a common feature [36], however it is not clear if this is true for ComR proteins. No residue on the conserved face is an asparagine (Fig 5), and although ComR S. thermophilus uses N208 to contact XIP (see Talagas. et. al. [37]), that same position in ComR S. suis is D213 (S5A Fig). Given that an aspartic acid likely fills a homologous role, it is tempting to speculate that variation of the position of an Asn residue may be related to peptide specificity. However, it is not readily apparent if the position an Asn relates to our observed categories of responses to XIP, for example (Fig 3 and S5B and S5C Fig). Furthermore, an alignment of the ComR X-ray structures and models used in this study reveals that MGAS5005 may not have an Asn or Asp in a position for contact with XIP (Fig 5 and S5C Fig).
Although the features highlighted in Fig 6A show a trend, in the type-I ComR-XIP S. thermophilus structure (Talagas, et al. [37]) the lower end of the pocket does not interact with the bound peptide, including residue K260 on the conserved face that is strictly conserved in all ComR strains used in our study (Fig 5, Fig 6B, S5A and S5B Fig). This could suggest a different mode of interaction with XIP, possibly in a more extended conformation, or that K260 performs an accessory role such as help to orient the CAP helix given that it forms a salt-bridge with D302 (Fig 6). However, our crystal structure contains a crystal-packing artifact that not only hints at how ComR S. suis recognizes its unique XIP, but provides clues for initial peptide recognition or selection. As shown in Fig 6B, the N-terminal residues (1 to 7 plus 4 residues of the cloning linker, RGSH-MEVWFMN) from a symmetry-related ComR bind in the apo-conformation of the XIP pocket. Although bound in the reverse orientation to S. thermophilus (S5 Fig), several hydrogen-bonds and van der Waals contacts are made to both the conserved and variable face with an interaction surface area of 1269 Å2 for the entire contact, as determined by PISA (http://www.ebi.ac.uk/pdbe/prot_int/pistart.html) [38]. Notably, W4 slots into the deep hydrophobic pocket of the variable face (Fig 5 and Fig 6B) while packing against K260, and E2 forms a salt-bridge with K250 at the back of the pocket, possibly mimicking the side-chain interactions of the S. suis C7 peptide. Additionally, residues that are common to strict ComR proteins (Fig 6A) hydrogen-bond to the N7 side-chain (residues K226 and Y228) while the variable face residue N220 is involved in hydrogen-bonding to the peptide backbone of the artifact. It is also important to note that this peptide fails to make contacts with residues T93 and R103 on α7 (T90 and K100 in S. thermophilus), which appear crucial to inducing the conformational change of the TPR necessary for activation (S5A and S5B Fig). Given this, we hypothesize that this may mimic in part an initial peptide interaction event that occurs when the XIP binding pocket is in the apo conformation. If the interacting XIP is accepted it can then trigger the conformational change of the TPR domain to open the DBD-TPR interface to allow dimer formation and activation as proposed by Talagas, et al. (Fig 7 and S5 Fig) [37].
To further explore the molecular features of the XIP binding pocket, we also compared the type-III ComR structure directly to the type-I ComR apo and activated structures reported in Talagas, et. al. (Fig 7 and S5 Fig) [37]. As shown in Fig 7A, the conformation of the TPR domain of S. suis and apo S. thermophilus structure are similar, although there are noticeable differences. Helices α6, α9, and α10 adopt clearly distinct conformations, and S. suis α7 extends further over the XIP binding pocket to contact the CAP helix α16. This arrangement is perhaps in part responsible for the observed conformational difference of the DBD domain relative to the TPR. Furthermore, the loop region between α8 and α9 (residues 128–144 S. suis and 131–139 S. thermophilus) is longer in S. suis, and the loop connecting α10-α11 (residues 178–183 S. suis and 173–181 S. thermophilus) adopts a different conformation between the two structures. In S. thermophilus this loop folds facing into the XIP pocket, but in S. suis this loop faces out towards the solvent (Fig 7A and S5A Fig).
Interestingly, if we compare these structural features to the activated ternary complex (Fig 7B) we observe that the XIP binding pocket of apo-ComR S. thermophilus shares more in common with the activated complex than the ComR S. suis structure (Fig 7C). Helices α10, α12, and α16 of S. suis differ from the S. thermophilus structures and the general position of the α10-α11 loop that makes contacts with XIP is preserved in both ComR structures of S. thermophilus regardless of whether the peptide is bound (Fig 7B and S5A Fig). In the activated state, the TPR as a whole does undergo significant changes most notably at α9 and α10 (including the loop region) and the repositioning of α6-α8 through contact with XIP. When activated, α6-α8 appear to rotate by ~30° relative to the apo-conformation (Fig 7B). However, we see that α6 and α7 in S. suis exhibit a different conformation from both type-I structures as they appear to be rotated counter clockwise (Fig 7C). Given that the type-I and type-III ComR proteins studied here recognize significantly different peptides (Fig 1) and have insertions/deletions of their primary sequence, these differences could potentially correlate to how XIP is recognized. However, as the DBD is inherently dynamic and the observed crystal packing interactions include some of the dimer contacts (Fig 4A), we cannot rule out influence from the crystallization process on the observed conformations of the S. suis or S. thermophilus structures.
As we have hypothesized that specific contributions of the conserved and variable faces participate in binding the peptide pheromone, and that ComR discriminates between, or filters, XIPs with the apo conformation of the pocket including residues in the lower end of the putative interface, we tested the interaction of ComR S. suis with peptide by site-directed mutagenesis. We created several protein variants of ComR S. suis to probe the assorted molecular features, including the solvent-exposed surface of the DBD (R19A, Q28A, and T42A), the DBD-TPR interface (Q40A and R43A), a residue on the conserved surface and lower end of the pocket (K260A), the electropositive surface of the back of the pocket (R103A, K100 in S. thermophilus see Talagas et. al. [37]), and the variable face (N220A) (Fig 5, S5B Fig and Fig 8A). These variants were placed into the ES1 test-bed background, and luciferase activity was recorded after stimulation with the S. suis XIP-C7 as outlined in Fig 2A. Additionally as a control, several variants were also introduced into the expression plasmid, purified as the wild-type, and assayed for structural integrity by circular dichroism (CD) spectroscopy (Fig 8B). The spectrum of each variant tested shows a folded α-helical protein similar to wild-type ComR S. suis, indicating that the selected mutations did not disrupt the protein fold.
The DBD variants Q28A and T42A retained significant activity, with Q28A responding to XIP similarly to wild-type. ComR S. suis R19A showed no activity and likely contributes indirectly through contacts made to residues on α-helix 3 that make close contacts with DNA (Fig 8A) [19]. Both residues at the DBD-TPR interface (Q40A and R43A) also exhibited significantly reduced activity, although Q40A could be partially stimulated with high levels of XIP. As Q40A and R43 are on α-helix 3, a conformational change would be required for interaction with DNA (Fig 5A, Fig 7, and Talagas et. al. [37]). The loss in activity could result from either disruption of the regulator mechanism or reduced affinity for DNA. In the case of the conserved residue K260 of the lower pocket, when it is mutated to alanine only negligible activity was observed, indicating that it is required for the activation of ComR. The same effect is observed for R103, implying that proper electrostatic interactions with the S. suis XIP are necessary. Interestingly, we also observed that mutation of residue N220 severely reduced activation, demonstrating that the variable face of ComR S. suis participates directly in the activation of ComR.
We next sought to examine further the specific mechanism of how ComR recognizes XIP directly in vitro through isothermal titration calorimetry (ITC) using three of the purified variants in Fig 8B. The titration curves of each experiment are shown in Fig 9 and the results summarized in Table 2. As shown in Fig 9B, S. suis has a high affinity for its cognate XIP-C7 with a Kd similar to that observed for SHPs and Rggs (Table 2) [39]. Additionally, the binding mechanism is driven by hydrogen bonding and likely electrostatic interactions as indicated by the large enthalpy, which again emphasizes the importance of the electropositive pocket (Fig 5C). When K260A was titrated with peptide (Fig 9A) we observed only background heats, indicating that K260A is absolutely required for binding of XIP. Mutation of N220 in the variable face of the pocket reduced the affinity for peptide by more than an order of magnitude (0.86 μM to 13.7 μM), although the overall mechanism appears to be similar (Table 2, the sign for ΔH and ΔS is the same as wild-type). Oddly, titration of XIP with the DBD-TPR interface variant Q40A showed rapid saturation suggesting an altered interaction with XIP or aggregation of the protein (Fig 9D). As Q40 may be part of the regulation mechanism, the protein variants that could still bind XIP were also examined by dynamic light scattering (DLS) (S6 Fig). The purified wild-type ComR S. suis showed one species in solution and then conversion to a larger species in the presence of XIP (S6A and S6B Fig). Likewise, N220A had the same behavior in solution as wild-type in the absence of XIP (S6C Fig). However, Q40A on its own appeared to have already dimerized and was in equilibrium with larger aggregates at a ratio of 4 to 1 (S6D Fig). Given that the DBD-TPR interface is likely compromised in this variant, the loss of activity of Q40A appears to be due to dis-regulation of dimer formation, which results in a significant amount of non-specific unproductive oligomerization. In this context, due to the observed aggregation by DLS, the interaction of XIP with Q40A remains unclear.
As we were interested in the binding of heterologous XIPs by ComR (S1 Fig and Fig 3), ComR S. suis was also titrated with S. mutans XIP C7 (Fig 9E). Although ComR S. suis cannot be activated by S. mutans XIP (Fig 3 and S1 Fig), it is still able to bind the pheromone directly as observed by ITC. However, the affinity is reduced by 7-fold and the mechanism of binding has changed, most notably a positive contribution from entropy (Table 2). Additionally, we titrated N220A with S. mutans XIP and observed the same Kd and mechanism as wild-type ComR S. suis (Fig 9F, Table 2). Together these data support our hypothesis that the conserved face residues provide the required contacts for pheromone binding while the variable face provides XIP-selectivity.
Observed pheromone specificity patterns (S1 Fig and Fig 3) and clues revealed from the ComR structure (Fig 5) suggested it might be possible to activate a selective ComR variant using a non-inducing XIP peptide redesigned to satisfy criteria for activation. For these experiments, the intermediately-restrictive S. agalactiae 2603 ComR was chosen since two disparate XIP peptides displayed activity upon this ComR, offering a basis for activation criteria. By comparing common traits of active peptides to those of inactive peptides, we identified two primary characteristics we hypothesized would account for activity. First, all peptides active with ComR S. agalactiae 2603 have three residues after the WW-motif, whereas all inactive peptides have one or two residues. Second, active peptides typically contain a terminal glycine or hydrophobic residue. A chemically-compatible terminal residue appears to be critical as the structure of the S. thermophilus ComR-XIP complex shows that it is inserted into the back of the XIP pocket to induce dimer formation (Talagas, et al. [37]). This is also observed in S. suis where the positive charge of the pocket and negative charge of the XIP play a role in binding (Fig 5 and Fig 7). Using these observations as a basis for ComR S. agalactiae 2603 activity requirements, we chose three XIP variants from S. mutans, S. suis and different S. agalactiae strain (NEM316, XIP TMGWWGL) that were unable to activate ComR S. agalactiae 2603 and predicted sequence rearrangements of the peptides that would lead to ComR activation. Each peptide was rearranged to fit the guidelines for ComR S. agalactiae 2603 to include three residues after the adjacent WW motif and contain a C-terminal glycine, leading to EBS1 (LDWWSLG), EBS2 (TGWWMLG), and EBS3 (ETEWWNVG). When tested by titrating the synthetic peptides to the ES1 reporter strain carrying ComR S. agalactiae 2603, EBS1 and EBS2 exhibited robust reporter induction (Fig 10) indicating that the guidelines successfully generated novel XIP ligands in two of three instances. However, the EBS3 XIP, which contains eight residues (rather than seven in EBS1 and ESB2) did not elicit strong reporter activity, indicating that pheromone length is an important determinant of activation, as reported in ComR S. thermophilus.
Through efforts to understand in detail the species specificity and molecular mechanism of ComRS quorum sensing, we developed an in vivo technology to test and monitor Streptococcus spp. communication in the regulation of competence. In prior studies with S. mutans and S. thermophilus, the exogenous addition of synthetic XIP was sufficient to activate ComR and induce transformation when cultures were grown in a chemically defined medium (CDM [10, 22, 40]). These discoveries enhanced the ability to genetically manipulate naturally transformable species under laboratory conditions [41, 42]. Nonetheless, for many notable pathogens, including Streptococcus pyogenes, exogenous addition of XIP leads to SigX induction, but transformation under laboratory conditions remains elusive [14]. The use of the S. mutans ‘test bed’ not only opens up the study of new ComRS systems, but has shown that ComR proteins can be strict or promiscuous in their ability to recognize the QS messengers of other species. We expanded these results by determining an X-ray crystal structure of ComR S. suis to help explain observations of pheromone specificity and develop a model for initial peptide recognition and activation of ComR. The biochemical data show that the XIP binding pocket consists of two faces, a conserved face that is required for peptide binding, and a variable face that we propose functions in pheromone specificity. Furthermore, the structural and biochemical observations together strongly suggest that the observed apo conformation of the XIP binding pocket allows the binding of varied XIP sequences. However, only those that satisfy the correct molecular interactions with both the conserved and variable face with sufficient affinity can induce the conformational change for dimerization that is required to bind DNA.
Genetic manipulation of many important pathogenic streptococcal species has been notoriously tedious, making the study of several species prohibitively difficult [43]. Here we demonstrate the expression of heterologous genes in an S. mutans test bed as a way to bypass slow and arduous genetic manipulation of native species to study genetic pathways. Through exchange of coding and regulatory regions of comRS loci from various species, we determined the functionality of ComR/XIP pairings within an otherwise intact quorum-sensing system. Our results indicate that nine studied systems exhibited productive pairings between pheromone and receptor, providing further evidence that the ComRS QS pathway is functionally maintained for the induction of the SigX regulon across Suis, Bovis, Mutans and Pyogenic species of Streptococcus [10, 14, 41, 42]. These findings demonstrate that previously unstudied ComRs from S. dysgalactiae and S. agalactiae bind the cognate XIP in vivo to activate transcription. Furthermore, the use of S. mutans as a host system could be easily translated to the study of other RNPP proteins and encoded peptide ligands found among the order Lactobacillales, or potentially be adapted to study the activity of other gene pairs. Taking into account the significant level of sequence variability of ComR and ComS peptides, we used this test-bed method to assess the ability of different ComR proteins to recognize foreign XIPs. We selected species from the Pyogenic (GAS and GBS), Bovis, and Mutans groups, and included the peculiar example of S. suis due to its distinctive split-tryptophan motif in ComS, which allowed us to discover unique response profiles that ranged from highly-specific to highly-promiscuous.
The two most promiscuous ComRs included in our study, SboB and Sbo 83, both recognized peptides encoded by species of each group within the type-II ComRS family. Surprisingly, both ComRs also recognized the atypical XIP encoded by S. suis. Together, these results suggest that the Bovis peptide-binding pocket’s intrinsic properties allow for more permissive ligand binding. Recently, Morrison et al. demonstrated that natural transformation in S. infantarius is activated under laboratory conditions upon treatment with synthetic cognate XIP [41]. Our results suggest, assuming successful import via Opp, that all type-II and type-III S. suis derived XIPs can activate competence in this Bovis species.
The results presented here agree with previous findings that ComR and putative XIPs up-regulate the competence regulon in S. infantarius, S. suis, and S. pyogenes, [14, 41, 42]; however, modest differences with published findings were seen with the test-bed system. One observed difference from published ComR/XIP behavior was for the two S. pyogenes strains used in our study, MGAS5005 (serotype M1) and MGAS315 (serotype M3). In the ES1 background, both the MGAS5005 and MGAS315 ComRs exhibited specificity with respect to the M1 and M3 versions of XIP, contrary to the robust cross-strain activation of PsigX::luxAB between the GAS M1 and M3 ComRS systems observed by Mashburn-Warren [14]. Though we are unable to explain these discrepancies, it is possible that M1 and M3 XIPs display differential stability within S. mutans extracellular or intracellular environments, possibly due to expressed peptidases that degrade peptide signals seen in other systems [44, 45]. Another observed difference is in comparison to past work showing that S. suis S10 (serotype 2) had the highest transformability upon addition of the XIP-C9 length variant (GNWGTWVEE), and that the C8 variant (NWGTWVEE) activated higher levels of transformation than that of the C7 variant [42]. Our findings indicate that an S. suis ComR of a different strain (05ZYH33) but same serotype expressed in the S. mutans ES1 background more strongly activated PcomS::luxAB upon treatment with the C7 length (WGTWVEE) than with C8. As both Suis strains encode identical ComR and ComS genes, this difference in activity might be explained by a length preference of the S. mutans peptide permease Opp. Substrate selectivity of peptide transporters has been reported, as in studies with Lactococcus lactis, where Detmers et al. observed a significant decrease in affinity between OppA and peptide substrates when the peptides contained 8 or 10 amino acids instead of the preferred 9, regardless of amino acid identities [46].
In comparing the biochemical features of ComR with other Rgg-like proteins, we found that unlike Rgg2, apo-ComR is a mostly a monomer in solution (S2 Fig) [19]. Previous work demonstrated that the active ComR-XIP complex is a dimer [21], and our DLS results demonstrate dimerization in vitro upon the addition of XIP (S6 Fig). However, the asymmetric unit of the S. suis ComR crystal structure resembles that of dimeric Rgg2, save for the lack the DBD domain swap (Fig 4A). The DBD is instead packed against the TPR by a substantial hydrogen-bonding network, which includes several residues implicated in DNA binding (Fig 4B, Fig 5A). Interestingly, when the residues at the domain interface are mutated, activity is abolished as determined by transcriptional reporter (Q40A, R43A Fig 8). Moreover, if the hydrogen-bonding stabilization of the DBD-TPR interface is disrupted, apo-ComR forms dimers and higher order oligomers, at least in vitro, that likely interfere with proper activity (Fig 5 and S6 Fig). From this we conclude that the apo-ComR structure holds the DBD against the TPR to not only time the DBD domain swap with XIP binding, but to assure proper oligomerization for the interaction with DNA. These observations are in complete agreement with Talagas et. al. [37] suggesting that the proposed dimerization mechanism is persevered in type-I, type-II, and type-III ComR proteins.
Due to the activation mechanism of ComR, there are several additional significant structural differences from other Rgg proteins. Not only does ComR lack the Rgg2 DBD disulfide to force oligomerization [19], but also the pheromone binding pocket itself differs drastically. As shown in Fig 4C, the arrangement of the ComR TPR produces a significantly different pocket. The most important feature is the extended helix α7 that makes contacts with the ComR CAP helix. We see a large shift of the ComR CAP outwards which not only produces a deeper apo-pocket compared to Rgg2 but the CAP-α7 interaction is likely broken by XIP binding, which in turn would allow for a shifting of the α6-7-8 helix bundle to begin the release of the DBD for dimerization. This suggests a completely different mode of peptide binding and activation of ComR compared to Rgg proteins and other members of the RNPP family.
ComR-XIP interaction results from the test bed experiments led us to classify ComR proteins as strict, intermediate, and promiscuous (S1 Fig and Fig 3). Regardless, a general comparison of the molecular features of the XIP binding pocket between these types did not reveal outstanding molecular differences to account for specificity (S4 Fig). Although some qualitative differences appear to exist between strict ComRs and promiscuous ComR TPR domains, these features did not correlate directly to complementary XIP charge or XIP side-chain size. Instead we observe that the ability to be activated by XIP depends on key interactions with the pocket side-chains themselves, likely in a species-specific manner. Supporting this hypothesis, we have observed that the apo-ComR XIP binding pocket consists of a conserved face and a variable face (Fig 5).
The amino acid conservation map of the apo-ComR surface in Fig 5 demonstrates one half of the pocket to be conserved and the other variable. As expected, if the conserved residues are mutated transcription is abolished (R103A, K260A Fig 8A) due to the inability of XIP to bind ComR (K260A Fig 9A). Moreover, K260 does not contact the activated peptide conformation as shown in Talagas et. al. [37] and S5 Fig. This suggests either a different conformation for the S. suis XIP, or that the lower end of the pocket makes contacts to the pheromone before it adopts a final conformation. However, we cannot discount that K260 may also, or instead, serve an alternative role such as stabilization of the CAP helix. In the case of the variable face, if that surface is mutated significantly reduced levels of transcription are observed (N220A Fig 8A) and the corresponding affinity of ComR for XIP is reduced by more than an order of magnitude (N220A Fig 9; Table 2). Moreover, wild-type S. suis ComR stimulated with S. mutans XIP did not activate in vivo transcription in the test-bed system (Fig 3 and S1 Fig) although we did observe binding in vitro. Interestingly, the in vitro interaction occurred with the same affinity and mechanism as the variable face mutant (N220A) titrated with the Suis XIP (Fig 9, Table 2). This shows that the conserved face provides contacts with XIP that are required for binding and the variable face provides contacts that provide further refinement of XIP selection. This is somewhat reminiscent of how the Rap proteins recognize their cognate peptides [47]. However, unlike the Rap family the ComR pheromone binding pocket positions the conserved and variable residues on opposite sides of the pheromone interface, and at least for the strict ComR proteins, a single residue substitution on the variable face does not appear sufficient to allow activation by heterologous XIPs (Fig 5, Fig 8 and Fig 9).
So how does ComR initially recognize and discriminate between possible peptide pheromones? A clue to this model comes directly from the crystal packing artifact observed in the crystal structure (Fig 6B). The N-terminus of ComR shares some features with its own XIP (specifically, tryptophan and glutamate side chains) and shows how a non-productive peptide might first fit into the apo-ComR pocket. We observe the W4 side-chain fitting into one of the likely aromatic-binding pockets with packing against K260. Although it makes some contacts with the variable face, it fails to make enough contacts with the conserved residues required for activity, especially in the back of the pocket (S5A and S5B Fig). We see that the residues N-terminal to M1 are swung outwards, making few contacts with ComR and fail to push into the back of the pocket to begin the conformational change for dimerization. One can hypothesize that the S. mutans peptide may fit into ComR S. suis in this manner, and without N220 to anchor the S. suis XIP it does not bind tight enough to lock in to the back of the pocket and induce activation. This model would allow for the initial binding of many different peptides, even those that are quite foreign like the promiscuous ComRs, yet allow for selectivity as demonstrated by the strict ComRs. Considering the biochemical data, pheromone recognition is ultimately accomplished by fine tuning the residues on the variable face. Furthermore, this artifact hints at how the unorthodox S. suis XIP might bind. The W4 position is likely the same as the N-terminal W of the C7 XIP, which would allow for the two C-terminal glutamate residues to fit into the electropositive back of the pocket to disrupt the interactions stabilizing the observed apo conformation and allow for the necessary conformational change.
In line with the structural and biochemical observations on XIP specificity, it is interesting to observe that in the work by Talagas et. al [37] the mechanism and affinity of type-I ComR interaction with pheromone differs in ITC experiments. For the type-III interaction studied here the affinity is ~800 nM and exothermic compared to ~40 nM for type-I and endothermic (Table 2). This contrast likely arises from the differences in the peptide sequence, namely that the S. suis XIP is charged and the S. thermophilus XIP is more hydrophobic. This suggests that the S. suis interaction is driven more by hydrogen bonding and electrostatics, whereas S. thermophilus is driven largely by the hydrophobic effect and entropy. We speculate that this most likely contributes to filtering of potential XIP peptides, but does not impact the overall mechanism of ComR activation. Additionally, given that the binding of XIP induces dimerization (S6 Fig, Talagas et. al. [37]), it was interesting to observe that a one-site model of the ITC data produced the best fit and expected ComR-XIP stoichiometry. This suggests that perhaps the binding of XIP to ComR is directly coupled and that the ComR-XIP monomer is an extremely short lived species
Taking the sum of our data into account, we could identify patterns associated with individual ComRs that displayed an intermediate selectivity. Using GBS 2603 as an example, we identified patterns within active XIPs to design agonists. As predicted, all three designed XIPs activated the GBS 2603 ComR, with two of the three peptides strongly inducing light activity. In this way, we were able to design ligand agonists without using receptor structural data. Whether these peptides function in the native strain to activate PcomS remains to be tested. Due to the finite number of active XIP variants tested of which we could compare to inactive variants, we were limited in determining specific residues that would improve an already productive interaction.
Here we have cataloged responses of several type-II ComR proteins to non-cognate XIPs and developed a model for how ComR selects, or filters for the proper XIP. Although we observed clear variations in response to heterologous pheromones among species, specificity did not appear to correlate with phylogenic relationship (diversity was apparent within species groups, e.g. S. pyogenes). The difficult question at hand is what does the XIP selectivity we observed mean for bacterial interactions occurring in nature? For example, could the promiscuous Bovis species activate its competence pathway upon detecting foreign XIP in its local environment? Likewise, is the control of competence in S. mutans so strictly regulated such that only Mutans-derived XIP will activate this trait? Both species are naturally found in complex microbial populations and the evolution of comRS to become either strict or promiscuous may provide an unseen advantage under some circumstances or in some communities. In particular, aside from the core competence regulon, SigX controls many other accessory genes [12] whose governance may impact fitness or interactions with other microorganisms. The limited size of our study made it difficult to correlate ComR specificity to virulence potential of individual strain isolates. For instance, while S. pyogenes strains MGAS5005 and MGAS315 are each documented as invasive strains and are of the most commonly isolated serotypes [48, 49], their observed responses to heterologous XIPs indicated distinct patterns, even though both alleles were categorized as intermediate. Larger sample sets are likely to be required to identify any associations between ComR specificity and phenotype, and with such analyses a more informative study, perhaps, would be to investigate correlations between ComR specificity and frequencies of horizontal gene transfer events, and might help to explain adaptability of strains or emergence of serotypes [50–52].
Although work remains to fully understand to what extent the ComRS system plays in species-cross communication in the regulation of natural competence, we have provided the ground work for studies aimed at revealing the role of ComRS in a myriad Streptococcus species, and have developed a detailed structural and biochemical model that describes how the quorum sensing receptor ComR recognizes and discriminates between pheromone molecules.
S. mutans strains were incubated at 37° C in anaerobic conditions (5% CO2) in Todd Hewitt Broth (THB; Difco) or in Chemically Defined Media (CDM; [53]) supplied with 1% (m/v) glucose and 20% glycerol for storage at -80° C. E. coli strains were incubated at 30° C or 37° C, as indicated, under aerobic conditions, in Luria Broth (LB; Difco). To select for strains, 0.75% agar was added to media with selective antibiotics and cultures were plated and incubated at either 30° C or 37° and under aerobic or anaerobic conditions, as indicated. Antibiotics used for selection of S. mutans strains were spectinomycin (200 μg mL-1) and erythromycin (1 μg mL-1). Antibiotics used for selection of E. coli strains were erythromycin (500 μg mL-1) and ampicillin (100 μg mL-1). Oligonucleotides were synthesized by IDT (Coralville, Iowa) and are listed in S1 Table.
To construct ΔcomRS::specR linear DNA, 1029 bps directly upstream of the comR (designated US-ComRS) encoding region as well as the 1097 bps directly downstream of the comRS inter-genic (IR) region (starting with the ATG site of comS; designated DS-ComR-IR) were amplified using the primer pairs LW4-26/LW6-46 and LW4-40/LW4-41, respectively. The LW6-46 primer contains a SalI tail while the LW4-40 contains a PstI tail. A 1009 bp product containing the spectinomycin resistance cassette (referred to as specR) encoded within pLZ12 spec [54] was amplified using the primers LW2-33, containing a SalI cut site, and LW2-37, containing a PstI cut site. In a 1:1:1 molar ration, the specR segment was incubated with T4 DNA ligase (NEB) in-between the US-ComRS and the DS-ComR-IR segments at the SalI and PstI sites, respectively. To naturally transform S. mutans strain UA159 with the completed ligation reaction, we used the method described by Desai et al. [40] with modifications. Briefly, overnight cultures of S. mutans strain UA159 were diluted to an OD600 of 0.05 in 0.5 mL CDM + 1% (m/v) glucose and grown under anaerobic conditions at 37° C for 1 hour before adding 1 uM UA159 C7 XIP. Cultures were incubated an additional 1 hour at 37° C before addition of the ligation reaction. After 1 hour of incubation with DNA, cultures were plated on selective media and plates were incubated under anaerobic conditions at 37° C. Electro-competent cells from the resulting S. mutans ΔcomRS::spec strain (ES1) were made using a modified S. pyogenes preparation procedure [55]. Briefly, overnight cultures of ES1 were grown in THY + 1 mM glycine. The overnight cultures were diluted 1:20 in 50 mL THY + 1 mM glycine and grown until an OD600 of ~0.3 was reached. At that time, cells were immediately spun down to pellets before undergoing an additional spin after re-suspension in sterile deionized water. An additional wash step was done with 15% glycerol before storage of aliquots at -80°C. Electro-competent ES1 aliquots were electroporated with ~200 ng of purified ComR reporter plasmid and recovered in 1 mL THY at 37°C for 90 minutes. Cultures were plated on selective media and incubated at 37° under anaerobic conditions until colonies formed.
Reporter plasmid DNA was derived from genomic DNA isolated from S. mutans strain UA159, S. pyogenes strain MGAS5005, S. pyogenes strain MGAS315, S. agalactiae strain A909, S. suis clinical strain 01–18929, S. agalactiae strain 2603 v/r, S. bovis strain SboB, S. bovis strain Sbo 83, and S. dysgalactiae subsp. equisimilis strain ATCC35666. (S2 Table). Reporter plasmids were constructed either by restriction digest followed by ligation or by Gibson assembly as previously described [56] and as indicated in S3 Table. The plasmid insertions for pMRW101, pMRW103, pMRW104, pMRW105, pMRW107, pERB1, pERB2, pERB3, pERB4, pERB5, pERB6, and pERB7 were each amplified with primer pairs containing either a 5’-PstI tail and a 3’-SalI or -KpnI tail as indicated in S4 Table. The plasmid insertions for pERB20, pERB21, pERB22, pERB23, pERB24, pERB36, pERB37, pERB38, pERB39, pERB40, pERB41, and pERB42 were amplified using two segments of DNA amplified from pERB5 which overlapped the bp(s) mutated to alanine as listed in S4 Table. The S. suis mutant reporter plasmid backbone, pERB5, was linearized with PstI and KpnI and underwent Gibson Assembly with the two segments of DNA containing the overlap region harboring the bp mutation, as described by Gibson et al. [56]. S. suis wild type and mutant ComR expression plasmids (pERB15, pERB30, pERB31, pERB32, pERB33, pERB34, pERB43, pERB44, pERB45, pERB46, pERB47, pERB48, pERB51) were synthesized using Gibson assembly after amplification of the wild type or mutated allele from the respective reporter plasmid (S4 Table) using primers EB147/EB148 containing 5’ and 3’ overlapping regions allowing recombination with the expression vector pET-15b (Novagen, Billerica, MA). Electro-competent cells were prepared from E. coli strain BH10C [57] and BL21(DE3) [58]. De-salted ligation reactions and 1:3 dilutions of Gibson assembly reactions were transferred into electro-competent cells. After electroporation, cells were recovered in LB, at 30° C for 90 minutes before plating and incubation at 30° C. All constructs were confirmed by DNA sequencing.
Predicted XIP peptide sequences were determined by identifying open reading frames downstream from the putative ComR gene by using NCBI BLAST to identify UA159 ComR homologues. One to three different lengths of peptide from each species were ordered from NeoScientific (Woburn, MA). Peptide crude extracts (purity 45–75%) were dissolved in DMSO (Fisher Scientific, Hampton, NH) at a concentration of 1 mM based upon the specific purity and stored at -20° C.
Starter cultures, from cultures stored in 20% glycerol at -80° C, of each ComR reporter strain were diluted to an OD600 of 0.05 in CDM + erm and incubated at 37° C until an OD600 of ~0.1. 190 uL aliquots of reporter strain cell culture were added to wells within a 96-well (Greiner Bio-One; Monroe, NC) lidded, clear, flat-bottom plate, each containing 10 μL of DMSO or a 5x-serial dilution of the peptide of interest in DMSO (final concentration of 5% DMSO/well). Concentrations tested for each peptide were 1.3 nM, 6.4 nM, 38.6 nM, 231.5 nM, 1388.9 nM, 83.3 μM, and 50 μM. 50 μL of a 1% decanal solution in mineral oil was added between each well in the plate to provide substrate for the luciferase reporter. Prepared plates were lidded and sealed with Parafilm (Bemis, Oshkosh, WI) before incubation at 37° C, with continuous shaking, in a Synergy 2 plate reader (Biotek). OD600 and luminescence readings were measured every 15 min. over a 4-hr period. The maximum RLU measurement per well, between 60–240 minutes was recorded using Gen5 data analysis software. For each well, baseline-corrected induction curves, EC50 values, and significance values between 50 μM and 0 μM peptide treatment (p <0.05) were calculated to determine the activity profile of each ComR (GraphPad, La Jolla, CA).
ComR S. suis 05ZYH33 (residues 1–304) was cloned as an N-terminal 6-his tag in the vector pET15b. BL21(DE3) cells were transformed with the expression vector and grown in LB media at 37°C until an OD of > 0.6 at 600 nm with protein expression induced by the addition of 1 mM isopropyl β-D-1-thiogalactopyranoside (IPTG). The temperature was reduced to 20°C and cultures were allowed to grow overnight. Cells were harvested by centrifugation followed by lysis with an Emulsiflex-C5 (Avestin). The lysate was cleared by centrifugation at 16,000 rpm for 30 minutes and then passed over a nickel NTA gravity column (Pierce) followed by a wash with 50 column volumes of chilled buffer (50 mM Tris pH 7.5, 500 mM NaCl, 25 mM imidazole). The protein was eluted with 5 column volumes elution buffer (50 mM Tris pH 7.5, 500 mM NaCl, 500 mM imidazole) and further purified using an SD75 16/60 superdex gel filtration column (GE Healthcare) via AKTA (GE Healthcare) at 4°C in a final buffer of 20 mM Tris pH 7.5, 100 mM NaCl, 1 mM beta-mercaptoethanol (β-ME). Selenomethionine labeled protein was produced using metabolic inhibition [59]. Briefly, cells were grown in M9 media at 37°C followed by reducing the temperature to 20°C and adding 0.05 g/L seleomethionine, leucine, valine, proline, and 0.1 g/L lysine, threonine, phenylalanine. The cells were allowed to grow for an additional 30 minutes before induction with 1 mM IPTG. All ComR protein variants were created by quick-change mutagenesis (Strategene) and expressed and purified similar to wild-type.
ComR S. suis 05ZYH33 was concentrated to 25 mg/mL for initial screening in commercially available conditions with a Tecan Freedom Evo 200 robot at the University of Illinois at Chicago Research Resources Center High Through-put facility. The crystallization conditions were 10 to 25 mg/mL ComR with a 1:1 mixture of 15% PEG 3350, 0.2 M sodium citrate. Crystals were grown by sitting drop vapor diffusion at 4°C with micro-seeding using SeedBead (Hampton Research).
Diffraction data was collected at the Advanced Photon Source at Argonne National Laboratories as part of the LS-CAT, Sector 21. Protein crystals were prepared by overnight dehydration in mother liquor substituted with 15% PEG8000 against 0.2 M sodium citrate and 35% PEG8000 followed by flash freezing in liquid nitrogen. Data was processed using XDS [60] and phases determined by single-wavelength anomalous dispersion (SAD) on data collected near the selenium peak using both the Phenix package [61] and Sharp/Autosharp [62, 63]. The initial model was further built and refined using Coot [64], Refmac5 [65] from the CCP4 suite of programs [66], Phenix [61], and TLS refinement [67]. PDB_REDO was used for final validation and refinement [68]. The final model has an R/Rfree of 22.0/25.5% with 99.7% of residues in the allowed region of the Ramachandran plot. The coordinates and structure factors (code 5FD4) have been deposited to the Protein Data Bank, Research Collaboratory for Structural bioinformatics, Rutgers University, New Brunswick, NJ (www.rcsb.org and www.pdb.org).
Purified ComR S. suis and variant proteins were diluted into 20 mM KHPO4 at pH 7.5 in a 0.2 cm path-length cuvette. Spectra were collected on a Jasco-815 spectropolarimeter from a wavelength of 260 nm to 190 nm. Background spectra of the buffer diluted into 20 mM KHPO4 was subtracted from the data. To plot the data as mean residue ellipticity, a mean residue weight of 118.1 was used with protein concentrations of 0.05 mg/mL for wild-type, 0.01 mg/mL for K260A, 0.03 mg/mL for N220A, R103A, and Q40A.
Purified ComR S. suis and variant proteins (K260A, N220A, and Q40A) were dialyzed overnight at 4°C into the same buffer stock of 20 mM Tris pH 7.5 100 mM NaCl 1 mM β-ME. HPLC purified synthetic peptides for both S. suis 05ZYH33 (WGTWVEE) and S. mutans U159 (GLDWWSL) were obtained from NeoScientific (Woburn, MA). Each XIP was reconstituted in the experimental buffer followed by centrifugation at 12,000 rpm to clear undissolved or precipitating material. XIP was used at 300 μM concentration and injected into 20 μM ComR. All experiments were performed using a VP-ITC calorimeter (Malvern) at 25°C. Controls included titration of XIP into buffer alone or the titration of buffer into ComR. The final heats of binding were analyzed using Origin Software (Malvern) using a one-site model.
Purified ComR S. suis and protein variants were injected into a DynaPro-801 (Protein Solutions) using a syringe with a 0.1 μm filter in 20 mM Tris pH 7.5 100 mM NaCl 1 mM β-ME after centrifugation at 6000 rpm for 15 minutes to remove any larger precipitates. The concentration of ComR was ~ 20 μM except for Q40A which was diluted to ~ 5 μM due to increased signal from aggregates. XIP was added to wild-type ComR at a concentration of 50 μM. Data was analyzed with the provided software as mono-modal or bi-modal using an aqueous buffer model.
|
10.1371/journal.pntd.0005749 | Geographic strain differentiation of Schistosoma japonicum in the Philippines using microsatellite markers | Microsatellites have been found to be useful in determining genetic diversities of various medically-important parasites which can be used as basis for an effective disease management and control program. In Asia and Africa, the identification of different geographical strains of Schistosoma japonicum, S. haematobium and S. mansoni as determined through microsatellites could pave the way for a better understanding of the transmission epidemiology of the parasite. Thus, the present study aims to apply microsatellite markers in analyzing the populations of S. japonicum from different endemic areas in the Philippines for possible strain differentiation.
Experimental mice were infected using the cercariae of S. japonicum collected from infected Oncomelania hupensis quadrasi snails in seven endemic municipalities. Adult worms were harvested from infected mice after 45 days of infection and their DNA analyzed against ten previously characterized microsatellite loci. High genetic diversity was observed in areas with high endemicity. The degree of genetic differentiation of the parasite population between endemic areas varies. Geographical separation was considered as one of the factors accounting for the observed difference between populations. Two subgroups have been observed in one of the study sites, suggesting that co-infection with several genotypes of the parasite might be present in the population. Clustering analysis showed no particular spatial structuring between parasite populations from different endemic areas. This result could possibly suggest varying degrees of effects of the ongoing control programs and the existing gene flow in the populations, which might be attributed to migration and active movement of infected hosts from one endemic area to another.
Based on the results of the study, it is reasonable to conclude that genetic diversity could be one possible criterion to assess the infection status in highly endemic areas. Genetic surveillance using microsatellites is therefore important to predict the ongoing gene flow and degree of genetic diversity, which indirectly reflects the success of the control program in schistosomiasis-endemic areas.
| Schistosomiasis is one of the important neglected tropical diseases endemic in 78 countries throughout the world. The disease is caused by the parasitic worms known as schistosomes. In the Philippines, S. japonicum is the causative agent of the disease. The prevalence of the disease varies in endemic areas, suggesting that the parasite populations might differ in their genetic composition and infectivity to the human host. In this study, DNA samples of adult worms from seven endemic municipalities were analyzed. Characterization of S. japonicum samples in different endemic sites with varying prevalence provides information on the genetic diversity of the parasite. Results of this study showed that samples in high prevalence endemic areas like Irosin, Catarman and Socorro were genetically diverse as compared to other areas. Information on parasite genetic diversity is therefore important in planning disease control strategies. The results suggest ongoing parasite transmission across geographic endemic areas which should be monitored and used as reference for genetic diversity of the schistosomes attributed to geographic areas, thus a safeguarding precaution should be implemented to ensure localized elimination of the disease.
| Schistosomiasis is one of the most important neglected tropical diseases affecting almost 240 million people throughout the world with more than 700 million considered to be at risk of infection [1]. Five species of schistosomes are known to cause human infection: Schistosoma haematobium, S. mansoni, S. mekongi, S. intercalatum, and S. japonicum. Among these species, S. japonicum is considered the most virulent because of the larger number of eggs it can produce as compared to other species, causing severe disease pathology. In addition, the zoonotic nature of S. japonicum contributes to increased disease transmission, making schistosomiasis control more difficult [2, 3].
Microsatellite markers have recently been used in determining S. japonicum genetic diversity and in estimating the levels of gene flow in the population. Previous studies have recommended the use of microsatellite markers to determine schistosome genetic diversity because of their codominant expression and their ability to serve as neutral markers [4, 5]. The easy observation of heterozygosity and the reasonable number of alleles per polymorphic locus in the samples make microsatellite analysis a more powerful tool in genetic studies than the use of rapid amplified polymorphic DNA (RAPD) and mitochondrial DNA [6]. Studying the genetic variation of S. japonicum populations provides an opportunity to link some genotypes associated with disease prevalence, which can then be used in formulating effective control measures. Previous studies using microsatellite markers suggested that the prevalence of S. haematobium and S. mansoni infections could be closely related with the parasite’s genetic variation. High prevalence of infection was observed in those areas with high genetic diversity and low prevalence in those areas with low genetic diversity [7, 8]. This information makes DNA microsatellites useful in population genetics studies.
In addition, microsatellite markers had been used in the identification of different geographical strains of S. japonicum in China [5]. Previous studies suggested that these markers could provide useful information for assessing the efficacy of mass drug administration (MDA) [9]. In the Philippines, schistosomiasis has shown considerable variations in the intensity and prevalence of the disease [10]. We hypothesized that the population genetic structure of schistosome parasites might partly contribute to these differences. The parasite population in each endemic area was therefore characterized for their genetic backgrounds using the microsatellite markers. The information found in this study can therefore provide basic information on the population genetic structure of S. japonicum in the Philippines that can be used as basis to evaluate and modify the widely used current control strategies for human schistosomiasis. For example, cognizant of the genetic types of the parasites, interventions can be modified so as to address differences in disease prevalence.
Animal experiments done in this study were conducted according to ethical guidelines for the use of animal samples permitted by the University of the Philippines Manila Institutional Animal Care and Use Committee (2011–009), as well as by Obihiro University of Agriculture and Veterinary Medicine (Permit No. 28–30). Infected mice were anesthesized using isoflurane before they were sacrificed. Perfusion method with normal saline solution was done to collect the adult S. japonicum.
The snail intermediate host, O. h. quadrasi, were collected from seven municipalities (in seven provinces) in 2013 to 2015 where the disease is endemic, namely Catarman (in Northern Samar), Gonzaga (in Cagayan Province), New Corella (in Davao del Norte), Irosin (in Sorsogon), Talibon (in Bohol), Alang-Alang (in Leyte) and Socorro (in Oriental Mindoro) (Fig 1). The snails were crushed between glass slides to examine the presence of cercariae under the microscope. The cercariae were then pooled separately for each endemic municipality and used for mice infection. Ten BALB/c mice were infected percutaneously with 50 cercariae from each municipality. The infected mice were sacrificed six weeks after the infection, and the adult worms were collected from their mesenteric veins and washed with saline for DNA extraction.
Genomic DNA was extracted from individual male and female adult worms using the DNeasy Blood & Tissue Kit (QIAGEN, Japan) following the manufacturers’ protocol. Table 1 showed the total number of DNA samples tested in each endemic municipality.
PCR amplifications were performed using Veriti 96-well Thermal Cycler (Applied Biosystems, Carlsbad, CA). Amplifications were performed in 10 μl reactions containing 1 μl of 10X PCR buffer, 0.4 μl of 1.5 mM MgCl2, 0.2 μl of 2.5 mM dNTP, 0.2 μl each of 10 pmol/μl primer, 0.1 μl of 5 U/μl Taq DNA polymerase (Takara, Otsu, Japan), and 1 μl of template DNA. The conditions for thermal cycling were as follows: 5 minutes at 94°C, followed by 30 cycles of 1 minute at 94°C, 1 minute at locus-specific temperature, 1 minute at 72°C, with a final extension at 72°C for 10 minutes [11].
The DNA of each individual S. japonicum worm was genotyped using the previously characterized microsatellite loci RRPS, M5A, TS2, MPA, 2AAA, J5, SJP1, SJP5, SJP6, and SJP9 [11, 12]. In our study, we have screened twenty microsatellite markers but among which only ten worked well. The 5’ end of the forward primer for each locus was fluorescently labeled with 6-FAM, VIC and Ned dyes. Different dyes were used for those loci with overlapping fragment size. Two μl of the PCR product with LIZ 600 labeled size standard (Applied Biosystems) was subjected to the 3500 ABI Prism Genetic Analyzer for fragment analysis assay. The allele sizes were determined using the Gene Mapper software version 4.0 (Applied Biosystems). In each run, S. japonicum sample from Gonzaga, Cagayan, which has good DNA volume and concentration, served as the reference genotype for which the microsatellite sizes for the 10 loci had been determined by sequencing. S. japonicum Yamanashi strain (Japanese isolate) was also genotyped as a control group to confirm that the microsatellite markers could differentiate between samples from different origin. A total of 201 DNA samples were tested; however, only 186 were successfully genotyped due to poor DNA quality.
For each population, the genetic diversity was examined by calculating the number of alleles using rarefaction analysis. Expected heterozygosity (gene diversity) (He) and observed heterozygosity (Ho) were determined using the GenAlEx 6.5 software [13]. Rarefaction analysis was performed to make the alleles comparable in the population. Genetic differentiation was determined using Wright’s F-statistics (Fst) in Arlequin, and the significance of the Fst values was tested at p value <0.05 [14]. The following qualitative guidelines were used for the interpretation of Fst genetic differentiation: 0–0.05 (little), 0.05–0.15 (moderate), 0.15–0.25 (great), and >0.25 indicate (very great genetic differentiation) [5]. The Analysis of Molecular Variance (AMOVA) was used to partition the genetic variation within and among populations using the software Arlequin version 3.5. The inbreeding coefficient (FIS), which measures the extent of nonrandom mating, was computed in the study. Nonrandom mating occurs when there is inbreeding.
Principal Coordinate Analysis (PCoA) was done to determine the clustering pattern of S. japonicum population based on their genetic distance using GenAlEx 6.5. Cercariae derived from a snail infected with only a single miracidium is assumed to be genetically identical. Hence, duplicate multi locus genotypes in a population are a consequence of clonal replication within snails [12, 15]. The presence of duplicate multilocus genotypes in adult worms was identified as one possible source of bias [15]. Duplicate multilocus genotype (MLG) was therefore removed, leaving a single representative of each in the dataset. Recent studies revealed that removal of clones in the dataset improved the assignment and clustering pattern of S. japonicum population [15]. The GENECLASS software 2.0 was used to identify migrant individuals [16].
To visualize relationships among populations, a Neighbor-joining tree was constructed based on FST genetic distance using 100 bootstrap replications in POPTREE2 [17]. The FST is one of the well-known parameters used in measuring genetic differentiation between populations using microsatellite data [18]. The S. japonicum Yamanashi strain was used as an outgroup.
Sequences of microsatellite loci reported here have been deposited in GenBank with the following accession numbers, RRPS (U22167), M5A (AF244896), TS2 (AF244896), MPA (U11895), 2AAA (M32280), J5 (M26212), SJP1 (EU262604), SJP5 (EU262608), SJP6 (EU262609) and SJP9 (EU262612).
A total of 186 individual S. japonicum worms collected from seven endemic municipalities were analyzed. Highest gene diversity indices (He) was observed in Catarman (0.727) followed by Irosin (0.694), Socorro (0.677), and Gonzaga (0.605) while the lowest was in Alang-Alang (0.495). Those from New Corella (0.587) and Talibon (0.566) were comparable. Similarly, allelic richness after sample size correction was highest in Irosin (4.630) followed by Catarman (4.500) and Socorro (4.280) and lowest in Alang-Alang (2.570) followed by Talibon (2.920) (Table 1).
Population-specific inbreeding coefficient was determined in this study to measure the extent of nonrandom mating. Highest inbreeding coefficient values (FIS) was observed in Irosin (0.239) while the lowest was in Catarman (0.012) (Table 1). The lowest inbreeding coefficient values in Catarman may be related to the increased heterozygosity in this area. Inbreeding increases the homozygosity of the alleles. The pairwise FST values ranged from 0.019 to 0.0188, indicating varied levels of pairwise population genetic differentiation (Table 2). Great genetic differentiation was observed in the New Corella samples. The AMOVA showed that greater genetic variation in the samples occurred within the population (91.95%) rather than among populations (8.05%) (Table 3).
The PCoA showed no particular geographical structuring among the S. japonicum populations (Fig 2). The neighbor-joining tree method showed clustering of the samples into two groups. Populations from Catarman, New Corella, Gonzaga and Talibon grouped together (86% NJ bootstraps), whereas populations from Alang-Alang Socorro and Irosin belong to a separate cluster (61% NJ bootstraps) (S1 Fig). However, there was no correlation between the clustering of populations and their geographic distribution as shown in the neighbor joining tree (S1 Fig). These findings further support the results of the PCoA (Fig 2), suggesting the existing gene flow in the population. Two subgroups were observed in Catarman (Northern Samar) using the PCoA analysis (Fig 2). The presence of two subgroups in the Catarman samples may account for the high genetic variation within population (Table 1). Six samples, namely 3 from Gonzaga, 2 from Irosin and 1 from New Corella, were identified by GENECLASS as migrants (Table 4). These individuals showed a probability below 0.05.
In this study, based on the hypothesis that the population genetic structure of S. japonicum might explain the variations in the intensity and prevalence of schistosomiasis in the Philippines, the genetic polymorphism of the parasite population from different endemic areas was examined. A large number of different alleles were observed in the samples examined, especially in Irosin, Catarman and Socorro where high prevalence of infections was reported [10] (S1 Table). There is a greater potential for these populations to possess the alleles responsible for the parasite infectivity, causing high infection [7, 19]. These findings were in agreement with that of previous studies where the prevalence of infection was directly proportional to the number of alleles [7, 8, 22]. This situation somehow follows a general pattern in our current study where high prevalence of infection either in humans and snail hosts was observed in those areas with high allelic richness while low or zero prevalence in those areas with low allele numbers (S1 Table). However, this is in contrast to our results in Alang-Alang where low number of different alleles has been observed where high prevalence of the infection was also reported. This could be due to the prolonged utilization of praziquantel from the annual MDA since Leyte has been one of the oldest endemic foci in the Philippines. To confirm such findings the effect of drug selective pressure brought by praziquantel on the parasite genetic diversity should be analysed. It has been known in other parasitic infections such as malaria that selective drug pressure brought by extensive drug use can lead to a reduction in genetic diversity of the parasite [20, 21]. Currently, there are no microsatellite markers that can be linked with parasite infectivity. The alleles that might contribute to the high infection rate might be present in those area with high prevalence, however further studies should be done to confirm this.
In this study, it should be noted that the prevalence data presented in S1 Table was collected from 2013 to 2015 while, our samples obtained from Talibon (Bohol) were collected prior to this period. Bohol is considered as a near-elimination area based on the absence of human cases for many years now. However, the presence of infection in water buffaloes continues to indicate an ongoing transmission even if there are no more human cases (S1 Table). Hence, the possibility of human infection is always present.
Among the seven endemic municipalities analyzed, the Catarman samples showed the highest gene diversity indices (Table 1). Catarman has been reported with high prevalence of infection both in humans and snail intermediate hosts (S1 Table). In addition, this study also revealed that water buffaloes and dogs in this municipality had high prevalence of infection (S1 Table). High infection rate in humans and animal hosts will then increase the probability of snail infection. A study by Rudge et al. (2008) in the Philippines using S. japonicum larval stages, found high levels of parasite gene flow between humans and dogs suggesting strongly the frequent transmission of S. japonicum infection across host species and between villages [23]. Furthermore, the role of animals in disease transmission was further supported by a population genetics study in China, where S. japonicum from cattle showed high genetic diversity in the marshland areas, whereas parasites from humans and dogs were more diverse in the hilly region [24, 25]. These previous studies have therefore demonstrated the contribution of animal host species in the genetic diversity, and gene flow pattern of the parasite. Thus, the zoonotic nature of S. japonicum infecting animals should be seriously considered in the increased disease transmission [26, 27, 28]. Currently, we are now performing direct genotyping of stool-derived eggs collected from humans and animals particularly in water buffaloes and dogs using microsatellites. We will measure the infection intensities in humans and animals together with the parasite’s genetic variation in our ongoing study. Moreover, in this study we chose to use adult worms isolated from mice experimentally infected with snail-derived cercariae for genotyping. This is because adult worms can generally provide DNA with higher quality and quantity suitable for genotyping than snail-derived cercariae or eggs in stool samples. However, fitness of parasites to mice may serve as a bias to the genetically diverse population, leading to bottlenecking of genotypes [29, 30]. The S. japonicum cercariae shed from infected snails collected from endemic areas will also be analyzed in our future studies. Because there are several studies on parasite’s population genetics by using cercariae-derived samples [23, 24, 25], such the direct genotyping may be feasible and provide an advantage to skip in vivo passage.
The level of genetic differentiation differs between endemic areas. Great genetic differentiations were observed in the New Corella samples than those from other endemic sites (Table 2). The large geographical distance separating New Corella from other endemic sites could possibly limit the contact between the hosts, eventually resulting to high genetic differentiation in this municipality. New Corella is located in Southern Mindanao and is expected to be more genetically differentiated because of its geographic location (Fig 1). As seen in Fig 1, New Corella is the farthest of the endemic municipalities being separated by a wide distance from other endemic municipalities. Previous studies showed that the high genetic differentiation observed among peripheral populations such as those of New Corella can be explained by their strong spatial isolation [31].
Furthermore, the possibility of the snail hosts influencing the genetic variation of the parasite population should also be taken into consideration. Presently, there is no study using microsatellites on the genetic variation of the snail population in the sampling areas. However, previous studies on mitochondrial DNA had provided some insights into the genetic variation of the snail population of S. japonicum, and suggested that examination of naturally infected snails may exhibit co-evolutionary relationships with their parasites [32]. Thus, a snail population may reflect the population genetic parameters of their parasites [33, 34]. Nevertheless, it is worth mentioning that these previous studies on the genetic diversity of Oncomelania populations are based on mitochondrial markers. Hence, future studies using microsatellites on the snail populations from each endemic area is essential to obtain results which can be analyzed together with those of S. japonicum population.
The genetic variation observed using AMOVA was greater within each S. japonicum population (91.95%) than the variation among the populations (8.05%) (Table 3). This might be due to the snails being infected by genetically different cercariae having multiple genotypes within the endemic areas [7, 22]. Mixing of infected snails and of their parasites brought about by flooding may explain the higher genetic variation within the population [5, 9]. Also, the continuous rainfall and subsequent floods in these endemic areas might facilitate host-parasite contact, exposing people and animals to contaminated waters that result to higher infection. Thus, people and animals moving from one village to another to escape flooding, take advantage of employment opportunities and there is also animal trade where water buffaloes are exported to other areas could facilitate parasite transmission, contributing to high genetic variation within each endemic area. Another reason could be due to the snail sampling being done in three villages for each endemic municipality where a high village-level variation might exist. Genetic variance within population was accounted for most of the genetic diversity of S. japonicum population in endemic provinces in China [5].
S. japonicum samples obtained from different endemic areas did not form a particular spatial structuring. The lack of geographical structuring suggests that there is still an ongoing gene flow among the S. japonicum populations in all the study areas despite execution of control measures [22]. These findings might imply that there is a continuing transmission of S. japonicum across geographic areas, and therefore reflect the inadequate effect of MDA implementation. The current national control strategy for schistosomiasis in the Philippines is annual MDA using 40 mg/kg of praziquantel in all schistosomiasis-endemic villages including the sampling areas. However, the compliance rate was reported to be <50% [35, 36].
The ongoing gene flow in the populations might be attributed to migration and movement of infected hosts as also suggested otherwise by previous studies done on S. mansoni [22]. The infected hosts could therefore serve as means of allele dispersal in endemic sites. Therefore, the existence of gene flow among the schistosome populations might increase the opportunity for the spread of alleles conferring parasite traits such as infectivity, virulence and drug resistance [8, 22].
Two subgroups were observed in the Catarman samples using the PCoA analysis (Fig 2), indicating that co-infection with several genotypes of the parasite might be infecting the hosts in this endemic site. Catarman is surrounded by other endemic areas such as Leyte, Negros Occidental, Bohol in the Visayas, so the possibility of intermixing of the parasite is very high leading to high genetic variation within the area. A higher transmission and infection success is expected to occur more in mixed parasite genotypes than in single-genotype infection as reported in previous studies [37]. There might be a decrease in the effectiveness of the host immune system to cope with the infection due to the simultaneous attack of the parasite with different genotypes leading to a higher infection success [37]. Furthermore, the high genetic diversity in Catarman may be explained by the lowest inbreeding coefficient values. Inbreeding increases the similarity of the alleles in the parasite population. Previous studies suggested that co-infection by multiple genotypes decreases the possibility of inbreeding [37].
Some parasite populations in Gonzaga, New Corella and Irosin were identified as migrants using Gene Class 2.0 software (Table 4). Gonzaga has just been identified as a new endemic focus at the start of the 21st century [38], and it is presumed that some parasite populations are introduced in this area from Talibon and Catarman. Infected people or animals might have moved from these areas and started the disease transmission in Gonzaga. The theory proposed for the emergence of schistosomiasis in Gonzaga was based on the history of a big geothermal project by PNOC (Philippine National Oil Company) in Gonzaga that recruited workers from the Visayas and Mindanao (Fig 1). The movement of people from these endemic areas into Gonzaga brought in cases, and with the presence of snail hosts in the area, the emergence of the disease became just a matter of time [38]. Interestingly, migrants detected in Irosin were rooted from Gonzaga, further providing evidence for the continuous transmission flow of the parasite among the endemic municipalities in the Philippines (Table 4). Migrant detection in the study supports the clustering analysis using the neighbor joining method wherein the population, including those individuals that were considered to be migrants, clustered together with their source population; for instance, Gonzaga received migrants from Talibon, which clustered together in the NJ tree (S1 Fig).
Genetic diversity found in this study among the parasite population in each endemic site in the Philippines is vital in the parasites’ ability to survive the effect of selective pressures such as those brought by drug treatment. At the same time, selection pressures increase the frequency of favorable alleles across all populations [4]. In this sense, the present finding of high diversity among the parasite populations imply that the MDA with praziquantel has varying degree of impact in interrupting the parasite’s life cycle. Aside from looking at several factors that can contribute to possible treatment failures including low compliance and the quality of the drugs used, it is also important that the effects of MDA be monitored on schistosome populations for its genetic background. This can be done by using microsatellite markers to measure the genetic diversity parameters which include the allelic richness and the heterozygosity of the alleles in the parasite population before and after MDA implementation. Alleles contributing to the severity of the disease such as the ones responsible for the fecundity and survival of S. japonicum inside the host should be identified and need to be further studied.
In conclusion, the use of microsatellites in this study has shown that there is an ongoing gene flow among the S. japonicum population from different endemic areas, indicating the active movement of infected humans and animals from one endemic area to another. Aside from the control programs being implemented in each endemic area, an effective surveillance to monitoring these movements in humans and animals in each endemic site should be in place. Thus, a better cooperation between the medical and veterinary sectors would be highly recommended to ensure a strengthened control program for schistosomiasis. In addition, the diversity will indirectly explain the varying degree of the effects of the ongoing control programs done in these endemic areas. A regular MDA should be implemented and monitored regularly for its efficacy in endemic areas. Considering that only 10 microsatellite markers were analyzed in this study for determining genetic diversity and gene flow of the parasite, we therefore recommend the use of additional highly polymorphic microsatellite markers not only for S. japonicum, but for S. mansoni and S. haematobium to be used in future studies for more precise analysis.
|
10.1371/journal.pcbi.1004808 | Reconstruction of Tissue-Specific Metabolic Networks Using CORDA | Human metabolism involves thousands of reactions and metabolites. To interpret this complexity, computational modeling becomes an essential experimental tool. One of the most popular techniques to study human metabolism as a whole is genome scale modeling. A key challenge to applying genome scale modeling is identifying critical metabolic reactions across diverse human tissues. Here we introduce a novel algorithm called Cost Optimization Reaction Dependency Assessment (CORDA) to build genome scale models in a tissue-specific manner. CORDA performs more efficiently computationally, shows better agreement to experimental data, and displays better model functionality and capacity when compared to previous algorithms. CORDA also returns reaction associations that can greatly assist in any manual curation to be performed following the automated reconstruction process. Using CORDA, we developed a library of 76 healthy and 20 cancer tissue-specific reconstructions. These reconstructions identified which metabolic pathways are shared across diverse human tissues. Moreover, we identified changes in reactions and pathways that are differentially included and present different capacity profiles in cancer compared to healthy tissues, including up-regulation of folate metabolism, the down-regulation of thiamine metabolism, and tight regulation of oxidative phosphorylation.
| Cellular metabolism is defined by a large, intricate network of thousands of components, and plays a fundamental role in many diseases. To study this network in its entirety, metabolic models have been built which encompass all known biochemical reactions in the human metabolism. However, since not all metabolic reactions take place in any given tissue, these generalized models need to be tailored to study specific cell types. Algorithms developed to date to perform this tailoring process have focused on keeping tissue-specific models as concise as possible. This approach, however, can remove essential reactions from the model and hamper subsequent analysis. Here we present CORDA, a tissue-specific building algorithm that yields concise, but not minimalistic, tissue-specific models. CORDA has many advantages over previous methods, including better agreement with experimental data and better model functionality. Using CORDA, we developed a library of 76 healthy and 20 cancer-specific models of metabolism, which we used to identify similarities between healthy and cancerous tissues, as well as metabolic pathways that are unique to cancer. Results of this work provide a broadly applicable tool to model cell- and tissue-specific metabolism, while highlighting potential new pathway targets for cancer therapies.
| Genome-wide Metabolic Reconstructions (GEMs) computationally model the molecules and reactions responsible for metabolism in any given organism, and have been applied across a variety of fields including metabolic engineering and evolutionary analysis [1]. Computational methods developed to study GEMs [2] have generated novel hypotheses about the structure of metabolic networks in microorganisms, and helped elucidate gaps in our knowledge of metabolism [3, 4]. Since the publication of the comprehensive human metabolic reconstruction Recon1 [5], human GEMs have enabled the study of human metabolism at a genome level [6]. These studies include the prediction of novel metabolic functions [7], prediction of metabolic biomarkers for congenital genetic disorders [8, 9], context analysis of omics data [10–12], comparison between humans and other mammals through gene homolog mapping [13, 14], and prediction of suitable cancer drugs [15, 16] and drug targets [17–19].
A particularly prolific subfield of human GEMs is the development of tissue-specific reconstructions. Different groups of metabolic reactions occur in different cell types. Hence, numerous studies have been dedicated to generating tissue specific or cell specific models of metabolism [20, 21]. These tissue-specific reconstructions can be built by piecing together the model based on previously established biological evidence obtained by reviewing the literature [22–26], through the integration of omics data and computational methods in order to tailor generic, published human reconstructions [5, 9, 27–29] to the desired cell type [15, 16, 30–33], or through a combination of computational algorithms and manual curation [27, 28, 34–36].
Automated tissue-specific reconstruction algorithms developed to date can be broadly categorized into two groups [20]: “flux-dependent” and “pruning” methods. Flux dependent methods find an optimal flux distribution through the general reconstruction which contains the maximum number of high confidence reactions (i.e. reactions whose presence is supported by significant experimental data) [15, 31, 32, 37–39]. These algorithms have been successfully used to predict gene essentiality in cancer tissues [19, 33], cancer specific metabolic pathways [31], metabolic biomarkers for congenital genetic disorders [8, 9], and cancer specific anti-growth factors [15, 16]. One of the main advantages of flux-dependent methods is the fact that they predict a flux distribution along with the tissue-specific model [20]. While this characteristic can be desirable, it also renders flux-dependent reconstructions “snapshots” of the metabolic state defined by the data, as opposed to comprehensive, functional metabolic models [15, 20].
The second category of tissue-specific reconstruction methods are pruning algorithms, which include MBA [34], mCADRE [30] and fastCORE [40]. Models generated using these algorithms have been used to calculate metabolic flux values in hepatocytes [34], identify pathways specific to cancer [30], and predict cancer drug targets [17, 18]. These algorithms start with a core set of reactions, obtained through literature review or experimental data, and proceed by removing the remaining reactions in the generalized human reconstruction while maintaining functionality in the core set. In these algorithms, a tradeoff can be defined between maintaining the model as concise as possible and including all core reactions. That is, if a core reaction requires too many undesirable reactions to carry flux, the algorithm may remove this core reaction from the tissue model, a tradeoff referred to as flexible core.
There are two main advantages to defining a core set of reactions before performing the tissue-specific algorithm. The first advantage is the possible inclusion of multiple sources of data and biochemical information [20, 34]. The definition of the reactions core is left to the user’s discretion, allowing for both the combination of data sources and the manual inclusion of reactions. Secondly, reactions with overwhelming evidence are always included in the final tissue model, since a non-flexible set of high confidence reactions can be defined [20]. This pruning approach then allows for the construction of comprehensive tissue models, containing all reactions that may be in a tissue’s metabolism, as opposed to a snapshot of the metabolic state returned by the flux-dependent methods [15, 20].
Current pruning methods are also accompanied by two major limitations, however. First, the order in which reactions are removed from the model plays a major role in the final reconstruction. Second, similar to flux-dependent methods, current algorithms aim to keep the final tissue-reconstruction as concise as possible, an approach referred to as parsimonious. These algorithms aim to remove from the tissue-specific model all reactions for which experimental data is unsupportive or unavailable, such as reactions with low levels of gene expression or non-gene associated reactions. While a concise tissue-specific reconstruction is desirable, keeping the reconstruction as parsimonious as possible may lead to the removal of fundamental reactions and physiologically unlikely flux distributions. In Recon 1, for instance, oxygen and H2O exchange reactions can be removed from the reconstruction with no effect on model functionality (Fig 1A). During simulations, however, these would be replaced by the uptake of the toxic metabolites superoxide anion and hydrogen peroxide respectively, leading to the prediction of physiologically inaccurate flux distributions (Fig 1A). The oxygen exchange reaction is in fact not present in the MBA and mCADRE liver reconstructions, and the water exchange reaction is not present in the mCADRE liver reconstruction.
Hence, in order to ensure our algorithm did not rely on alternative, physiologically unlikely pathways, and that it was independent of any ordering assignments, we chose to take an approach which was not parsimonious. Here we introduce a novel tissue-specific reconstruction algorithm based on Cost Optimization Reaction Dependency Assessment (CORDA). CORDA returns a concise, functional tissue-specific reconstruction, and features a flexible reactions core. CORDA does not depend on Flux Variability Analysis [41] or Mixed Integer Linear Programming (MILP) problems, but only on Flux Balance Analysis [42] (FBA), which is dependent on Linear Programming (LP). This characteristic renders CORDA considerably faster than previous, similar methods. Finally, the CORDA algorithm returns reaction associations that assist in any manual curation to be performed following the automated reconstruction process.
In line with previous studies [43], we apply CORDA to generate a library of 76 healthy and 20 cancer-specific metabolic reconstructions. These reconstructions enabled us to identify metabolic similarities amongst healthy tissues as well as key differences between healthy and cancerous tissues. Furthermore, by sampling the feasible solution space in cancer and healthy models, this library can be used to predict the up- and down-regulation of cancer-specific pathways in cancer metabolism.
The CORDA algorithm is based on a novel approach to identify the dependency of desirable reactions (i.e. reactions with high experimental evidence) on undesirable reactions (i.e. reactions with no experimental evidence), a method referred to here as dependency assessment. In the dependency assessment approach, the metabolic network is modified in four ways (Fig 1B). First, reversible reactions are split into forward and backward components. Second, a pseudo-metabolite is added as a product for every reaction in the model. At this point, undesirable reactions will carry a higher stoichiometric coefficient for this added metabolite, assigning these reactions a higher “cost”. Third, a reaction consuming this pseudo-metabolite is added to the model. Finally, a positive lower bound is set for the reaction being tested in order to force that reaction to carry flux. After modifying the network, FBA (Materials and Methods) is performed while minimizing the flux through the cost-consuming reaction (Fig 1B). The flux distribution returned will then use high cost, undesirable reactions only as necessary for the reaction being tested to carry flux. Throughout the manuscript, we will refer to high cost reactions predicted to carry flux as associated with the reaction being tested. In order to identify pathways with the same cost (i.e. same number of undesirable reactions), multiple dependency assessment can be performed while adding a small amount of noise to the cost of each reaction.
Using this dependency assessment, we have developed the CORDA algorithm for the reconstruction of tissue-specific models (Fig 1C). CORDA takes as input the reactions in the generalized human reconstruction separated into high (HC), medium (MC), and negative (NC) confidence groups (see Materials and Methods section for a detailed description). All remaining reactions in the reconstruction (i.e. non gene associated reactions or reactions for which no data is available) are designated as others (OT). All HC reactions are included in the model, and the maximum number of MC reactions is included while minimizing the inclusion of NC reactions. While the definition of these four reaction groups are left to the user’s discretion, here we categorize them according to proteomics data from the Human Protein Atlas (HPA)[44, 45] and a methodology used in previous studies [30, 32, 37] (Materials and Methods). To begin the algorithm, all HC reactions are moved into the tissue reconstruction (RE). In a first step, MC and NC reactions associated with each RE reaction (which are the same as the HC group at this point) are identified using the dependency assessment and moved into the RE group. In a second step, NC reactions associated with a high number of MC reactions are identified and moved into the tissue model, and all remaining NC reactions are blocked (upper and lower bounds set to zero). Next, all MC reactions still able to carry flux are also moved to the RE group. Finally, in the final step of the algorithm, all OT reactions associated with any RE reaction are moved to the RE group for the final tissue-specific model. A detailed description of the CORDA method, including detailed steps, algorithm parameters, and categorization of model reactions is available in the Materials and Methods section.
Following the algorithm validation, we generated a library of 76 healthy and 20 cancer tissue-specific models using CORDA. In order to generate the most comprehensive models possible, we used the generalized human reconstruction Recon2 [9] in the calculation of this library. Recon2 is one of the most comprehensive human reconstructions performed to date, containing approximately twice the amount of reactions than Recon1, 1.7 times more unique metabolites, and 1.2 times more unique genes. Details of how the reconstructions were calculated can be found in the Materials and Methods section.
Here we introduced a novel tissue-specific algorithm based on Cost Optimization Reaction Dependency Assessment (CORDA). CORDA relies solely on FBA, rendering it more computationally efficient than previous methods. CORDA takes a non-parsimonious approach to the reconstruction process, based on the addition of valuable reactions to the reconstruction as opposed to the removal of non-essential reactions. We showed that the CORDA algorithm provides reconstructions that agree better with experimental data, and that demonstrate better metabolic functionality than prior methods like MBA and mCADRE. Furthermore, CORDA provides reaction associations that can greatly assist subsequent manual curation, while maintaining the reconstructions only slightly larger than previous parsimonious approaches. Monte-Carlo sampling analysis also demonstrates that the CORDA generated models provide better predictions of tissue-specific functionality.
In addition to the algorithm validation, we generated a library of 76 healthy and 20 cancer tissue-specific reconstructions, which show considerable agreement with our current knowledge of healthy tissue and cancer metabolism. First, as an initial validation of our cancer and healthy tissue models, we computationally predicted metabolites that are more frequently essential in cancer models than healthy tissues [15, 16, 54]. Two metabolites were implicated in this analysis: phosphatidylethanolamine (pe_hs) and triglyceride (tag_hs), both of which are part of metabolic pathways previously implicated as cancer specific [15, 16]. While future work is merited to identify more specific essential metabolites (e.g. through the inclusion of more comprehensive metabolic tasks in the tissue reconstruction process, and more metabolites in the essential metabolite identification algorithm), these results help validate the cancer and healthy tissue reconstructions presented here.
Following this analysis, we demonstrated that the tissue models calculated by CORDA cluster largely according to tissue type. Similar clustering patterns, based on gene expression and proteomics data, have been observed experimentally. In particular, based on the expression of over 30,000 genes across multiple individuals and tissues, one study found that brain, muscle, and liver tissues, as well as Epstein-Barr virus-transformed lymphocytes, form well defined groups, while skin, adipocytes, and nerve tissues cluster closely together [117]. A separate study used in the generation of the HPA, based on protein evidence from almost 17,000 protein-coding genes in 44 major tissues and organs, also showed that tonsils, spleen, appendix, and lymph node tissues cluster closely together, and that bone marrow clusters separately, but close to these lymphoid tissues [45].
Evidence supporting many of the apparent exceptions identified by our clustering analysis is also available. For instance, Uhlén et. al. found that brain and liver tissues, along with testis, cluster considerably separate from other tissues and closer to each other, which is what we observed by clustering the CORDA models. The same study found that prostate tissue clusters closely with salivary glands [45]. It is worth noting that good agreement with the data by Uhlén et. al. is expected, given that a subset of this data was used to generate the tissue-specific models. This agreement, however, suggests that the similarities between tissues shown by Uhlén et. al.[45] and Melé et. al.[117] at the gene expression and protein level are also present in the metabolic enzymes level.
Additionally, breast and salivary glands are known to share many morphological features, and studies have shown that both can give rise to tumors with similar morphology [118, 119] and myoepithelial differentiation [120]. These finding can explain why breast and salivary glands clustered with epithelial and myoepithelial cells, as opposed to glandular cells. Finally, skin cancer and non-Hodgkin’s lymphoma appear frequently as secondary cancers in immunosuppressed individuals [121, 122]. This could lead to cancers with significantly different metabolic profiles, supporting their separation from the remaining cancer models.
Clustering of tissue-specific models according to subsystems has also highlighted many differences between healthy and cancerous tissues at the pathway level (Fig 5). Evidence for many of these differences are also available in the literature, including:
Single reactions included most often in cancer or healthy tissue models were also analyzed, and again literature evidence has been found to support many of them (Table 3). Two surprising findings stemmed from this analysis. First is the predicted down-regulation of CoA synthesis reactions, implicated in both the subsystem and single reaction analyses. Upon further inspection, we traced this differential inclusion to the gene PPCS, the only gene related to this pathway included in the reconstruction process, which is significantly down-regulated in cancer cells [44, 45]. Second, the exclusion of ACOAO7p from most cancer models is also unexpected, since this reaction is part of the fatty-acid oxidation pathway, which has been shown to be up-regulated in cancer tissues [123, 124]. Protein evidence of this reaction’s associated gene, ACOX1, supports this exclusion from cancer models [44, 45], suggesting an alternate pathway for palmitoyl-CoA oxidation in cancer tissues.
Finally, Monte-Carlo sampling was also performed in all healthy and cancer tissue models. Sampling results demonstrate that cancer models show an increased capacity through pathways that are largely up-regulated in cancer metabolism, and a reduced capacity through pathways previously shown to be down-regulated. Interestingly, mitochondrial respiration showed a slightly reduced and tightly constrained capacity in cancer over healthy tissue models, despite the presence of a larger number of oxidative phosphorylation reactions in cancer models (Fig 5). For decades, the role of mitochondrial respiration was thought to be decreased in cancer tissues due to their high glycolytic capacity. In recent years, however, researchers have shown that this pathway actually plays an important role in cancer metabolism [125, 126]. Our results suggest that although a larger number of oxidative phosphorylation reactions are present in cancer models, the activity of this pathway is tightly regulated by cancer metabolism topology (Fig 6). On one hand, the low probability of cancer models reaching high cytochrome c oxidase flux values compared to healthy tissues is in line with cancer’s high glycolytic potential. At the other extreme, the low probability of cancer models reaching relatively low cytochrome c oxidase sampled fluxes is in line with the key role played by mitochondrial respiration in cancer metabolism uncovered in recent years.
We have also investigated the differences in glycine hydroxymethyltransferase capacity in cancer versus healthy tissue models (S1 Text). This reaction is dependent on two proteins, SHMT1 and SHMT2, which correspond the cytosolic and mitochondrial isozymes respectively. Both these proteins have been shown to be up-regulated in cancer over healthy tissue models [127], although SHMT2 has been so to a greater extent [71, 127]. The over expression of these proteins, however, has been shown to be heavily dependent on cancer type [127]. This claim is supported by the protein expression of SHMT2 in the HPA, where half the cancer types considered have samples with both high and not detected SHMT2 expression. This variability could explain why the distribution of reactions associated with these genes is similar between cancer and healthy tissue models (S1 Text). Some cancer types, however, show a considerable increase in SHMT2 expression when compared to their healthy counterparts, including breast, glioma, head and neck, lung, stomach, testicular, and thyroid cancer. In all but one of these models (glioma), the flux distribution of glycine hydroxymethyltransferase was shown to be considerably shifted towards higher values when compared to their healthy counterparts (S1 Text). These results demonstrate CORDA’s ability to predict cancer type specific functionality, and not only differences between all cancer and healthy tissues taken together.
The CORDA tissue-specific reconstruction algorithm, as well as the healthy and cancer tissue-specific reconstructions presented here, introduce a new approach for the development of comprehensive tissue-specific metabolic reconstructions. These reconstructions can generate novel insights into both healthy and diseased human metabolic behavior. Furthermore, the ability of CORDA to generate models based solely on experimental data, along with the computational efficiency of this algorithm, allows for continuous updates of this library of tissue-specific models, both as more experimental data is updated and made available, and as more comprehensive human metabolic reconstructions are developed.
While previous methods determined reaction dependencies using Flux Variability Analysis (FVA), the CORDA algorithm takes a different approach, referred here as dependency assessment. The novelty of this method lies not in the LP formulation itself, which is the same as the widely established Flux Balance Analysis (FBA), but in the model modifications performed prior to the application of FBA, as well as the interpretation of the flux distribution returned. Assuming we want to test whether a given reaction, x, is dependent on the presence of a group of reactions, Y, to carry flux, CORDA proceeds in five steps. The parameters required for the CORDA algorithm are summarized in Table 4:
It is worth noting that the high cost reactions implicated in step five are not necessarily essential for x to carry a flux ±ϵ, but are the set of reactions in Y that combined carry the minimal amount of flux. That is, no flux distribution through the metabolic network allows for the predefined flux through x with a lower combined flux through the reactions of Y. For instance, if one of the reactions in Y deemed associated with x were to be removed from the reconstruction, x could still be able to carry a flux ±ϵ, but the combined flux through the reactions in Y would be larger than before. This way, this dependency assessment does not minimize the number of undesirable reactions to allow x to carry flux, but instead the combined flux through them. Naturally, however, a lower number of reactions would more easily allow for a lower combined flux. It is also for this reason that throughout the manuscript we use the term associate instead of dependent. Throughout the literature, referring to one reaction as dependent on another means the removal of the later from the model negates the former’s ability to carry flux, which is not necessarily the case for the reaction associations defined here.
Another significant advantage of this dependency assessment over previous pruning algorithms is that it requires only the LP problem solved during FBA, rendering it much faster than previous methods. While MBA and mCADRE used a much faster variation of FVA, it is still considerably more computationally expensive than LP. Although mCADRE is up to three orders of magnitude faster than MBA [30], the mCADRE model used in this study took about 4 hours to be calculated in a 2.34 GHz CPU with 4G RAM using the IBM CPLEX solver [30]. The CORDA reconstruction, on the other hand, using the same data and general human reconstruction, took under 30 minutes in a 2.66 GHz CPU with 4G RAM using the Gurobi solver [128].
In order to obtain a tissue-specific metabolic reconstruction using this dependency assessment, we define the Cost Optimization Reaction Dependency Assessment (CORDA) algorithm. This algorithm takes as input the reactions in the generalized human reconstruction divided into four categories:
Here, we also allow for the inclusion of metabolic tasks in the HC group. That is, during the CORDA algorithm, sinks can be specified for given metabolites, and added to the model when tested to ensure the final tissue model can produce these metabolites. These reactions are added when being tested then immediately removed from the model, so that none of these metabolic task reactions are present when other reactions are being tested, and no two test reactions are present in the model at the same time. The 32 metabolic tasks included in all CORDA reconstructions in this manuscript are available in S1 Table.
While the definition of these reaction groups can be left to the user’s discretion, here we defined the four groups according to proteomics data from the HPA [44, 45], and boolean gene-reaction rules included in the generalized reconstructions Recon1 and Recon2. In the HPA, each protein is classified as being Not Detected, or present at Low, Medium or High levels in each tissue. The gene-reaction association rules are composed of gene names and “AND” and “OR” boolean associations. For instance, the reaction r0634 in Recon2 has the boolean rule “HADHB AND (ACAA2 OR ACAA1)”, and can therefore be considered active if the gene HADHB, as well as ACAA2 or ACAA1, are active.
Using this boolean mapping, gene IDs were first replaced by the numerical values -1, 1, 2, and 3, corresponding to Not detected, Low, Medium and High protein expression levels respectively. Genes not included in the dataset were assigned a numerical value of zero. Next, AND boolean associations were replaced by the function MIN; OR boolean associations were replaced by the function MAX; and the expression was evaluated. Reactions with a final score of 3 were assigned to the HC group; reactions with scores of 1 or 2 were assigned to the MC group; and reactions with a score of -1 were assigned to the NC group. Reaction scores of -1, 1, 2, and 3 also correspond to Not Detected, Low, Medium, and High expression levels expressed in Fig 2.
As an example, HADHB is expressed at low levels in cerebellum Purkinje cells; ACAA2 is not detected; and ACAA1 is expressed at high levels. The r0634 gene-reaction rule mentioned above was then be replaced by “MIN(1,MAX(-1,3))”, which evaluates to 1. During the Purkinje cells reconstruction, this reaction was then placed in the MC group. Similar approaches have been used by previous studies to assign reaction confidence scores [30, 32, 37].
Aside from the four reaction groups, the CORDA algorithm also requires 5 parameters to operate, which are summarized in Table 4. To begin the algorithm, all HC reactions are moved into the tissue-specific reconstruction (RE), since these are sure to be included in the final model. Given the remaining three reaction groups, the CORDA algorithm proceeds in three steps:
It is worth noting that one of the main advantages of CORDA over pruning algorithms is the fact that it is independent of how reactions are ordered. This is due to the fact that reaction associations are calculated for each step, and at the end of each step a decision is made as to which reactions are added to the tissue reconstruction. This way, the order in which reaction dependencies are calculated does not affect the final tissue reconstruction.
The CORDA reconstructions used for comparison to previous methods were generated using γ = 105, the highest cost value tested, κ = 10-2, the lowest noise value tested, ϵ = 1, a threshold similar to a previous study [32], n = 5, to allow for the inclusion of a larger number of OT reactions, and p = 2.
For a direct comparison to previous methods, the CORDA reconstructions used during the parameter sensitivity analysis, cross-validation, and comparison to previous methods were performed using the same data used for the mCADRE hepatocyte reconstruction. For the Monte-Carlo sampling analysis, a new reconstruction was generated using the most up-to-date data from the HPA. Both of these reconstructions are available in the supplemental material (S1 File). All calculations in this study were performed using the COBRA toolbox [129] and the Gurobi optimizer [128]. The MATLAB function file used for CORDA reconstructions is also available in the supplemental material (S2 File). Finally, an example of the CORDA algorithm, applied to small sample networks, is available in S2 Text.
While CORDA requires a number of different parameters, many of these values can be arbitrarily assigned. For instance, γ can be arbitrarily large, while ϵ and κ can be arbitrarily small. In order to demonstrate that the CORDA algorithm is robust to a wide range of parameters, we performed 108 hepatocyte specific reconstructions varying all parameters but p (which was set to be equal to two) to a wide range of values. A separate sensitivity analysis of p was performed and is included in S1 Text. The parameter p can be set in order to define a more or less flexible MC and NC core, and can be set to the user’s discretion.
These 108 reconstructions were based on the generalized human reconstruction Recon1 [5], using the same set of protein expression data (total of 560) and 32 of the metabolic tests used in the mCADRE hepatocyte specific reconstruction [30]. The data used in this step, as well as the metabolic tests and calculated reaction groups, are available in the supplemental information (S1 Table). Metabolic tests were included as single reactions in the reconstruction in order to assure the model was able to produce certain metabolites. Each metabolic test was added to the model when being tested then immediately removed, so that no two tests were present in the model at the same time, and no metabolic test reaction was included when other reactions were being assessed. Details of this analysis are available in S1 Text.
During the metabolic tasks validation analysis, the exchange rate of the basal inputs carbon dioxide (co2[e]), water (h2o[e]), protons (h[e]), oxygen (o2[e]), phosphate (pi[e]), hydrogen peroxide (h2o2[e]), superoxide anion (o2s[e]), bicarbonate (hco3[e]) and carbon monoxide (co[e]) were unconstrained. All other uptake reactions were blocked unless otherwise specified.
For each of the 20 amino-acid recycling tests, the uptake rate of the given amino acid and glucose were set to an arbitrary value, so that the amino-acid being tested was the only source or nitrogen. Next, the production of urea was set to a strictly positive value, and FBA was performed while optimizing the production of urea. The same test was also performed for ammonium. For each of the 21 glucogenic tests, the uptake rate of the given metabolite was set to an arbitrary value, and the production of glucose was optimized. For both the amino-acid and glucogenic tests, if the model returned a feasible flux distribution the test was considered passed, otherwise it was considered failed. If the exchange reaction of the given metabolite was not present in the model, the result was considered inconsistent. The generalized Recon1 reconstruction failed two of the glucogenic tests, so the results of the remaining 19 tests are reported in the main text.
For the eight nucleotide production tests, a sink consuming the given nucleotide was added to the cytosolic compartment. The model was allowed to uptake glucose and ammonium (as a source of nitrogen), and the flux through the sink was optimized. If the model was able to produce the given nucleotide, the test was considered passed.
Following the validation of the CORDA algorithm, we generated a library of 76 healthy and 20 cancer tissue-specific reconstructions using the generalized human reconstruction Recon2 [9] and the most recent proteomics data from the HPA [44, 45]. All reactions used to generate the tissue-specific models are available in S1 Table, and tissue-specific models are available in SBML and MATLAB format at [130]. The healthy tissue models were calculated using the same classification as described in the algorithm description section, since data for each protein was categorized as not detected, low, medium or highly expressed in each cell type. For cancer models, the same classification was available for any number of samples for each protein in each cancer type. In this case, values of -1, 1, 2 and 3 were assigned to each sample according to not detected, low, medium or high expression levels respectively, and these values were averaged for a final protein score in that particular cancer type. These protein values were then used in the gene-reaction boolean association as described in the algorithm description for a final reaction score. Reactions with a score equal to or greater than 2.5 were assigned to the HC group, less than 2.5 but greater than 1 to the MC group, and less than or equal to -0.5 to the NC group.
For instance, in renal cancer samples, protein HADHB has been analyzed in 12 different samples in the HPA, and was found to be expressed in high levels in 2 of them, medium levels in 8, and in low levels in 2. The protein score associated with HADHB in renal cancer is then calculated as ( 2 · 3 ) + ( 8 · 2 ) + ( 2 · 1 ) 12 = 2. Similarly, ACAA1 expression was calculated as medium in 5 samples, low in two samples, and not detected in four samples of renal cancer, yielding a score of ( 5 · 2 ) + ( 2 · 1 ) + ( 4 · ( - 1 ) ) 11 = 0 . 73. Finally, ACAA2 is present in high levels in one sample, medium level in 5 samples, low levels in one sample and not detected in 3 samples of renal cancer, giving this protein a score of ( 1 · 3 ) + ( 5 · 2 ) + ( 1 · 1 ) + ( 3 · ( - 1 ) ) 10 = 1 . 1. With that, the score for r0634 is calculated as “MIN(2,MAX(0.73,1.1))”, which is 1.1, putting this reaction in the MC group during the renal cancer reconstruction. Data and reaction distributions used during these calculations can be found in S1 Text.
Healthy and cancer specific models were clustered according to reactions present in each model. For that, 4,205 reactions present in at least one, but not all models were obtained. A binary vector was then calculated for each model indicating whether reactions were present (1) or not present (-1). These vectors were then clustered using hierarchical clustering with Hamming distance as the similarity metric, and average linkage. Leaf orders were also calculated in order to maximize the similarity between neighbors in the hierarchical binary cluster tree dendrogram. These results are summarized in Fig 4.
Next, in order to divide the clusters according to subsystem expression, a total of 4,751 reactions present in any of the models was obtained. These reactions were then divided by subsystem according to their classification in the Recon2 reconstruction. For each of the clusters of models calculated in the previous step, the average number of reactions from each subsystem included in the cluster’s models was then calculated. Finally, this number was divided by the total number of reactions in that subsystem which were included in any of the models for a final score between zero and one. These values were then clustered using hierarchical clustering with Euclidean distance as the similarity metric, and average linkage. Leaf orders were again organized to maximize similarity between neighbors to yield Fig 5.
Perhaps the most widely used method to analyze GEMs is Flux Balance Analysis (FBA) [42]. FBA predicts a flux distribution through the metabolic network which optimizes (maximizes or minimizes) a given objective function, defined as a single reaction or group of reactions in the network. This flux distribution is subject to upper- and lower-bound constraints, which include exchange reactions, and a steady state assumption for all model metabolites, so that no metabolite has a net production or consumption rate.
The mathematical formulation of GEMs are defined at the core by a stoichiometric matrix S, where each row defines a metabolite, each column defines a reactions, and each entry the stoichiometric coefficient of that metabolite in that particular reaction. Vectors defining lower (lb) and upper (ub) bounds for each reaction, as well as an objective vector (c) of the same length, are also defined. Given this model, FBA finds a flux vector v through all reactions in the GEM such that:
S ⋅ v = 0 l b i ≤ v i ≤ u b i o p t i m i z e : v ⋅ c T
During the dependency assessment described here, the stoichiometric matrix S is altered to reflect the changes described above. Given a reaction j being tested, a group of undesirable reactions Y, and a matrix S of size m by n, let κ ¨ denote a random number drawn uniformly between 0 and κ. The GEM is modified in the following ways:
With these constraints in place, FBA is performed as described above while minimizing the objective function. For each reaction in the reconstruction, if i ∈ Y and vi ≠ 0, the reaction i is deemed associated with j.
Monte-Carlo sampling was performed in a manner similarly to Bordbar et. al.[25] and Lewis et. al.[49]. This sampling method is a slight variation of the Artificially Centered Hit and Run (ACHR) algorithm developed by Kaufman and Smith [131]. In this algorithm, warmup points are initially generated at random corners of the solution space by solving an LP problem with objective vectors containing randomly generated ones and negative ones. The center point between all points is then computed. Next, for each point sampled, a random direction is selected as the difference between a randomly selected point and the center point. By selecting the direction this way, the direction is biased in the longer direction of the solution space, speeding up the rate of mixing while maintaining uniformity. After a direction is chosen, the limit of how far the current point can travel in that direction is calculated, and a new point is randomly chosen along that line. After several iterations, the set of generated points will be well mixed and approach a uniform sampling of the solution space.
The termination condition imposed on the ACHR algorithm here is the same imposed by Bordbar et. al.[25] and Lewis et. al.[49], introducing the concept of mixed fraction. For that, a partition is created over the set of points by drawing a line at the median value, with half the points on either side of the partition. The mixed fraction is the number of points that cross this line during mixing. Initially, the mixed fraction is one as all the points are on their original side of the line. As the sample solutions are mixed, the probability of each point crossing the median line approaches 0.5 asymptotically. The sampled points were initially mixed using the warmup points created as described above until the mixed fraction reached a particular threshold. Following that, the samples were mixed two more times, using the previous iteration’s final points as warmup points, until the same mixed fraction was reached. For the comparison between CORDA and other tissue-specific algorithms, a mixed fraction threshold of 0.52 was chosen as the termination condition. For the cancer and healthy tissue-specific models, a mixed fraction threshold of 0.6 was chosen to make the 96 sampling experiments computationally feasible.
Due to the heterogeneity between tissue-specific models, sampled flux values were evaluated between all cancer and healthy tissue models separately. That is, all sampled flux values for the given reaction were obtained from all cancer models that contain that reaction, and compared to all sampled values from healthy tissue models that contain the reaction. Results of this analysis are presented in Fig 6. In some cases, two or more reactions were combined: MTHFD2* combines reactions MTHFD2 and MTHFD2m, GHMT2r* combines reactions GHMT2r and GHMT2rm, and SPODM* combines reactions SPODM, SPODMe, SPODMm, SPODMn and SPODMx. These are the same reactions taking place in different cellular compartments. For these, flux values from each of these groups of reactions were added within each sampled flux distribution when plotting Fig 6.
|
10.1371/journal.pntd.0002030 | Naturally Occurring Incompatibilities between Different Culex pipiens pallens Populations as the Basis of Potential Mosquito Control Measures | Vector-borne diseases remain a threat to public health, especially in tropical countries. The incompatible insect technique has been explored as a potential control strategy for several important insect vectors. However, this strategy has not been tested in Culex pipiens pallens, the most prevalent mosquito species in China. Previous works used introgression to generate new strains that matched the genetic backgrounds of target populations while harboring a new Wolbachia endosymbiont, resulting in mating competitiveness and cytoplasmic incompatibility. The generation of these incompatible insects is often time-consuming, and the long-term stability of the newly created insect-Wolbachia symbiosis is uncertain. Considering the wide distribution of Cx. pipiens pallens and hence possible isolation of different populations, we sought to test for incompatibilities between natural populations and the possibility of exploiting these incompatibilities as a control strategy.
Three field populations were collected from three geographic locations in eastern China. Reciprocal cross results showed that bi-directional patterns of incompatibility existed between some populations. Mating competition experiments indicated that incompatible males could compete with cognate males in mating with females, leading to reduced overall fecundity. F1 offspring from incompatible crosses maintained their maternal crossing types. All three populations tested positive for Wolbachia. Removal of Wolbachia by tetracycline rendered matings between these populations fully compatible.
Our findings indicate that naturally occurring patterns of cytoplasmic incompatibility between Cx. pipiens pallens populations can be the basis of a control strategy for this important vector species. The observed incompatibilities are caused by Wolbachia. More tests including field trials are warranted to evaluate the feasibility of this strategy as a supplement to other control measures.
| Population suppression is an important component of mosquito control measures. The incompatible insect technique exploits the monogamous mating behavior of female mosquitoes to decrease the percentage of females inseminated by compatible males and hence reduce overall fecundity. Previous studies used genetically engineered Wolbachia-infected mosquitoes as the sources of incompatible males. The long-term stability of these mosquitoes is unknown. In this study, we examined naturally occurring incompatibilities between different field populations of Culex pipiens pallens, a vector of West Nile Virus and filarial worms widely distributed in China. We found that bi-directional patterns of incompatibility existed between some Cx. pipiens pallens populations. Incompatible males could compete against cognate males in mating with females and suppress reproduction. The level of suppression depended on the relative population sizes of incompatible and cognate males. We also found that these incompatibilities were preserved in the offspring from incompatible crosses, indicating that this control strategy is likely to be sustainable when applied repeatedly to successive mosquito generations. Our data also indicate that these incompatibilities were caused by endosymbiont Wolbachia, which is also the basis for cytoplasmic incompatibility in a variety of mosquito species. Our results may help to design simpler and more time-effective control strategies for a number of vector-borne diseases.
| Vector-borne diseases especially those transmitted by mosquitoes such as malaria, dengue fever, Japanese encephalitis, West Nile fever, Chikungunya and lymphatic filariasis are still important scourges responsible for millions of deaths each year, especially in tropical and developing countries. The lack of effective vaccines for major mosquito-borne diseases and the development of resistance to available chemotherapies both underscore the importance of reducing the numbers of major vector mosquitoes. Insecticides are the major weapons for mosquito control. Earlier efforts using insecticides to reduce malaria and other neglected tropical diseases had been met with success [1], [2]. But, long-term intensive use of insecticides has led to the development of insecticide resistance in important vector mosquito species, jeopardizing the effectiveness of insecticide-based vector control [3]. In addition, the adverse effects of insecticides on human health and the environment could not be ignored [4], [5].
With these concerns, biological approaches are called upon as alternatives to chemical control. One approach is to render the arthropods incapable of disease transmission. This strategy uses transgenesis or paratransgenesis to introduce foreign genes into arthropod populations. The expression of the transgenes makes these arthropods resistant to infections or unsuitable for parasite development to their infective stages [6], [7]. Introduction of endosymbionts such as Wolbachia, maternally inherited obligatory intracellular bacteria, into certain mosquito species can also decrease their vectorial capacity and have yielded novel strategies for population replacement [8]–[11]. Another approach resembles chemical control in that it aims at population reduction. For example, transgenes or biological agents such as Wolbachia can be introduced into mosquito populations to reduce fertility [12]–[15].
Because females of many insect species mate only once in their life-time, unproductive mating effectively precludes their reproduction or reduces their fecundity. Two strategies have been developed to take advantage of their monogamous behaviors, namely the sterile insect technique (SIT) and the incompatible insect technique (IIT). SIT uses radiologically or chemically sterilized males to suppress target populations. IIT uses incompatibilities especially cytoplasmic incompatibility (CI) between males and females to reduce female fecundity. CI is an incompatibility between a sperm and an egg that interferes with the normal development of a zygote. It is caused by endosymbionts in the cytoplasm such as Wolbachia [15]–[17]. Since their initial identification in the ovaries of Culex mosquitoes [18], Wolbachia have been found in the reproductive tissues of the majority of tested arthropod species [19]. Wolbachia cause a variety of reproductive abnormalities besides CI, such as feminization of genetic males, thelytokous parthenogenesis (female offspring being produced from unfertilized eggs), and male-killing [15], [20]. Among these abnormalities, CI has been explored as a potential tool for the development of novel and environmentally friendly strategies for insect controls [9], [21]. Because CI exists between sperm of Wolbachia-infected mosquitoes and eggs of an uninfected or an incompatible strain of Wolbachia-infected mosquitoes, CI-based IIT is applicable to both Wolbachia-negative and Wolbachia-positive target populations. CI prevents embryonic development, thus making the mating events unproductive. Consequently, those mated females are unable to reproduce or have their fecundity greatly reduced.
The application of CI to suppress insect populations predated the identification of Wolbachia as its causative agent. Laven used an introgressed Culex pipiens fatigans strain with maternal cytoplasm derived from a Paris strain to successfully suppress the incompatible local Cx. pipiens fatigans populations in Myanmar [22]. Subsequently, CI-based suppression was trialed for the European cherry fruit fly Rhagoletis cerasi [23] and almond warehouse moth Cadra (Ephestia) cautella [24]. To improve the safety of this suppression strategy, an irradiation step was included in several studies on Culex mosquitoes to minimize inadvertent release of fertile females and minority-type (compatible) males [25]–[27]. Medfly Ceratitis capitata transinfected with Wolbachia derived from R. cerasi have been generated for IIT approaches to suppress medfly populations [28]. CI has also been utilized to develop strategies to control Aedes mosquitoes. Through interspecific hybridization and introgression, an Aedes polynesiensis strain harboring Wolbachia derived from Aedes riversi was generated. This new strain is bi-directionally incompatible with the natural strain [29].
Wolbachia-infected mosquito strains that display novel patterns of CI for population suppression are usually created by introgression or transinfection [30]–[32]. The preparation of incompatible mosquitoes using such methods is often laborious and time-consuming. Because introgression involves putting Wolbachia into new host genetic backgrounds, the long-term stability of such novel mosquito-Wolbachia symbioses is unknown. In addition, although IIT has been applied to a variety of insects, it has not been tested in Culex pipiens pallens, one of the most prevalent mosquito species in China. In this study, using three Cx. pipiens pallens field populations collected in eastern China, we tested the applicability of exploiting incompatibilities between naturally occurring populations to suppress target populations. We report that incompatible males could compete with cognate males in mating with females, resulting in reduced fecundity. Though the incompatibilities were not absolute, they were preserved with hybrid offspring keeping the maternal crossing types. Our data indicate that this strategy may be a viable and sustainable approach over time to suppress populations of Cx. pipiens pallens.
WX (Wuxi, Jiangsu Province, 31°33′58.47″N, 120°18′9.88″E), NJ (Nanjing, Jiangsu Province, 32°3′30.11″N, 118°47′47.28″E), and TK (Tangkou, Shandong Province, 34°52′34.97″N, 117°22′53.69″E) populations of Cx. pipiens pallens were used in this study (Figure 1) [33]. WX and TK populations were collected in 2008 from July to August. NJ population was collected in July, 2006. For each population, several hundreds of larvae were collected with 350-ml dippers from over 50 larval habitats in public areas of the respective location. Fourth-instar larvae were identified to species by morphology, including the aspect ratio of the air tube and the number of pectens and tufts on it. The larvae were then brought back and reared in the insectary. Mosquitoes were kept at 28°C and 75% relative humidity with 14 h∶10 h light∶dark cycle. Adult mosquitoes were fed 10% (w/v) glucose solution prior to blood meals.
To separate virgin females and males, pupae from each population were put into 15-ml tubes with water for individual emergence. Afterward, male and female adults were raised in 30.5×30.5×30.5 cm cages. Females 2 days after emergence and males 3 days after emergence were used in mating experiments. Each set of crossings included combination groups of virgin males and females from two different Cx. pipiens pallens populations, with combinations of males and females from the same populations as positive controls. Females and males placed in the same cages were given 2 days to mate. Females were blood fed after mating, then the egg rafts were given 48 hours after oviposition to hatch in separate containers. The numbers of eggs and larvae were counted under stereoscope and the hatching rates (HR) for individual rafts were calculated. Unhatched egg rafts were all examined for fertilization through microscopic observation of embryonic development [34]. Each experiment was performed twice, shown here is a representative result.
Total DNA of individual Cx. pipiens pallens mosquito was extracted using a method previously described [35]. Polymerase chain reaction (PCR) was first carried out using generic primers to amplify wsp gene which encodes the major surface protein of Wolbachia to determine the infection status of the mosquito [36]. Studies have reported that the majority of insect Wolbachia belong to supergroups A and B, specific wsp primers were used accordingly for typing [37]. All primers were adopted from published studies. Primers wF and wR (see Table 1 for primer sequences) were used to amplify a 590–632 bp fragment from all Wolbachia strains; primers wAF and wR were used to amplify a 556 bp fragment characteristic of supergroup A; primers wBpipF and wR were used to amplify a 501 bp fragment characteristic of wPip strains of supergroup B; and primers wBcauBF and wR were used to amplify a 466 bp fragment characteristic of the wCauB strains of supergroup B. In addition, PCR was carried out to amplify the ankyrin domain-encoding gene ank2 to differentiate wPip strains into groups I-V based on 313–511 bp amplification fragments [38].
All PCR reactions were composed as follows: 1 unit of Pyrobest Taq DNA polymerase (Takara, Japan), 5 µl 10× Pyrobest Buffer II (Mg2+ Plus), 4 µl of 2.5 mM dNTPs, 5 ng total DNA, 2 µl of each 10 µM primer, and ddH2O was added to bring up the total volume to 50 µl. The amplification of wsp gene was performed as follows: 32 cycles of 94°C for 30 s, 55°C for 30 s and 72°C for 60 s, followed by a final extension at 72°C for 7 min. The amplification of ank2 gene was performed as follows: 35 cycles of 94°C for 30 s, 52°C for 30 s and 72°C for 60 s, followed by a final extension at 72°C for 7 min. PCR products were resolved by 1% agarose gel electrophoresis and stained with ethidium bromide. PCR products of ten mosquitoes from each population were used for sequencing of wsp and ank2 alleles by chain-termination method (Invitrogen). Sequence analysis was carried out using DNAMAN software. The unique DNA sequences were deposited into GenBank (accession numbers JX050182 - JX050187).
Tetracycline treatment to eliminate Wolbachia from Culex populations was carried out according to published methods [17]. Tetracycline (Amresco) at a concentration of 0.05 mg/ml was used for the treatment through both larval and pupal stages. Eggs were placed on tetracycline water solution to hatch. Surviving larvae were transferred to fresh tetracycline solution every 24 hours. A normal infusion was prepared in parallel and fed to larvae in tetracycline solution. The elimination of Wolbachia was checked by Hoechst 33342 (Sigma) staining [39]. Egg rafts within 1 hour of oviposition were placed in 5 ml fixation buffer (182 mM KCl, 46 mM NaCl, 3 mM CaCl2, 10 mM Tris pH 7.2, 3.7% formaldehyde) and overlaid with 5 ml n-heptane in a 50-ml tube. After incubation at room temperature for 15 min with constant shaking, fixation buffer was replaced with 10 ml methanol. The eggs were incubated at room temperature for 10 min with constant shaking. The n-heptane layer was replaced with 10 ml methanol. The eggs were washed for 2 times with methanol, and stained with Hoechst 33342 (1 µg/ml in PBS) for 15 min before visualization under microscope.
Statistical differences in hatching rate among crossing groups were examined using the Student's t-test. Linear regression analysis was conducted to determine the correlation coefficient between hatching rate and percentage of incompatible males. All statistical analyses were carried out using SPSS Statistics 17.0.
We first tested the compatibilities between available Cx. pipiens pallens populations collected from three geographic locations in China. Reciprocal crosses among these three field populations were performed as outlined in Table S1. In the mating combinations of NJ and WX populations, no significant incompatibility was detected in either NJ♀×WX♂ or WX♀×NJ♂ cross as compared to NJ♀×NJ♂ and WX♀×WX♂ control groups (Table S1 and Figure 2A). In the mating combinations of TK and WX populations, bi-directional incompatibility was observed, with both TK♀×WX♂ and WX♀×TK♂ crosses having significantly reduced average hatching rate as compared to TK♀×TK♂ and WX♀×WX♂ control groups (Table S1 and Figure 2B). In the WX♀×TK♂ group, 14 larvae hatched out of 21 egg rafts, making the average hatching rate 0.005±0.026. The incompatibility was more pronounced in the TK♀×WX♂ group, with 2 larvae hatched out of a total of 18 egg rafts, resulting in an average hatching rate of 0.001±0.001. In comparison, the hatching rates in the control groups TK♀×TK♂ and WX♀×WX♂ were 0.873±0.028 and 0.856±0.003, respectively.
Similarly, bi-directional incompatibility was detected in the mating combinations of NJ and TK populations (Table S1 and Figure 2C). The average hatching rates for TK♀×NJ♂ and NJ♀×TK♂ crosses were 0.004±0.003 and 0.004±0.002, respectively. In comparison, the average hatching rates were 0.883±0.028 and 0.922±0.015 for TK♀×TK♂ and NJ♀×NJ♂ control groups, respectively.
The observed bi-directional incompatibility between TK and WX populations and between TK and NJ populations could be a result of mating failure or some post-mating events. To distinguish between these two possibilities, insemination status and embryonic development in incompatible crosses were examined. The female genitalia of Culex mosquitoes include three oval-shaped spermathecal capsules for spermatozoa storage which are connected to vagina through spermathecal ducts. During coitus, some or all of these capsules are filled with seminal fluid from a male and become inflated. The females from incompatible crosses were examined and compared to virgin females. Because three spermathecal capsules from the same individual female had variable sizes before insemination as seen in virgin females, and it was likely that they also varied from mosquito to mosquito, inflation was not always a reliable indicator of successful insemination. Instead, the presence of spermatozoa inside the spermathecal capsules or ovarioles was checked. In all crossing experiments we conducted, the females from both TK♀×NJ♂ and TK♀×WX♂ crosses contained spermatozoa, indicating that they had mated. In addition, the egg rafts were examined 48 hours after oviposition. The eggs from incompatible crosses showed clear embryonic development (Figure S1). Inside the eggs from TK♀×WX♂ cross, there was segment formation along the axis. At the anterior end of the eggs, two red spots representing primitive eyes (stemmata) were visible. Consistent with previous reports, these embryos showed some abnormalities [34]. One evident difference between TK♀×WX♂ cross and compatible crosses was that the incompatible embryos had disorientated bristles. In the eggs from TK♀×NJ♂ cross, similar development was observed. 1–3 (mostly 2) pigmented stemmata were formed. Most stemmata were located in the head region, but some did not seem to have a specific localization. In some eggs, two stemmata were aligned anteroposteriorly instead of being side-by-side. In addition, there was segment formation along the axis in some eggs. Disoriented bristles were also observed in the eggs from TK♀×NJ♂ cross. These observations further confirmed that the females in the incompatible crosses had mated. For those females needed for subsequent mating experiments, only embryonic development inside the eggs was checked to confirm their insemination status.
Female monogamy in insects is common but often not absolute. Considering these populations were geographically isolated, it was possible that mating with incompatible males did not preclude subsequent mating of the inseminated females. To test if mating with incompatible males made the females refractory to re-mating, females were retrieved from the incompatible crosses and tested for their ability to mate with cognate males. In the aforementioned crossing experiment, the combination of TK and WX populations displayed higher level of incompatibility than that of TK and NJ populations, so TK and WX populations were chosen in subsequent experiments. TK♀ from TK♀×WX♂ cross and WX♀ from WX♀×TK♂ cross were separated from WX♂ and TK♂, respectively. Each group was equally divided into two subgroups, with one subgroup mixed with cognate males and the other kept alone (Table S2). If subsequent mating could happen, the subgroup mixed with cognate males would become inseminated with both compatible and incompatible spermatozoa and result in higher hatching rates than the subgroup kept separate from cognate males. As shown in Figure 3 and Table S2, the hatching rates in TK♀ subgroup mixed with TK♂ and in TK♀ subgroup kept alone without males were not significantly different (t = −1.013, df = 18, P = 0.324). Similarly, the hatching rates in WX♀ subgroup mixed with WX♂ and in WX♀ subgroup kept alone without males were not significantly different (t = −1.0, df = 6, P = 0.356). These results indicate that both TK♀ and WX♀ became refractory to subsequent mating after they mated with incompatible males.
Because these incompatible females and males can successfully mate in the absence of compatible males and produce reduced numbers of offspring, we then tested if incompatible mating would still occur when compatible males were available. We also tested if the reduction in fecundity depended on the number of incompatible males introduced. To that end, two sets of experiments were performed using TK and WX females (Table S3). In each set, an equal number of females were placed in different cages together with no male (blank), equal number of compatible males (positive control), equal number of compatible males plus equal number of incompatible males, or equal number of males plus 3× as many incompatible males. In addition, as in the previous experiment, 48 hours after the first oviposition, egg rafts were checked for embryonic development to ascertain that the females were inseminated. Subsequently, females were collected from each group and divided into two subgroups to be mixed with either no male or cognate males (Table S4). The hatching rates from the second oviposition were compared to determine if mating with mixed male populations made these females refractory to subsequent mating even when only cognate males were available.
The results (Table S3, Figure 4A) show that in the positive control group (TK♀×TK♂) the average hatching rate was 0.861±0.020. When an equal number of WX males were included in the cage [TK♀×(TK♂+WX♂)], the average hatching rate dropped to 0.211±0.071. This value was further decreased to 0.063±0.039 in the group that included 3× as many WX males [TK♀×(TK♂+3×WX♂)]. In comparison, the average hatching rate of TK♀ mixed with only WX♂ was 0.004±0.003 and TK♀ kept alone produced no larva. These data indicate TK♀ mated with TK♂ or WX♂ in the presence of mixed male population, i.e., even in the presence of TK♂, mating between TK♀ and WX♂ still occurred, which resulted in reduced overall fecundity of the groups. The extent of fecundity reduction correlated with the ratio of WX♂ to TK♂.
Similarly, in parallel experiment with WX♀ (Table S3, Figure 4B), the average hatching rate of WX♀ was reduced from 0.903±0.013 in the positive control group (WX♀×WX♂) to 0.607±0.066 in the group with an equal number of competing TK♂ included [WX♀×(WX♂+TK♂)] and further to 0.407±0.077 in the group with 3× competing TK♂ included [WX♀×(WX♂+3×TK♂)]. The average hatching rate of WX♀ mixed with only TK♂ (WX♀×TK♂) was 0.001±0.001. WX♀ kept alone produced no larva. These data also indicate that the frequency of WX♀×TK♂ mating in the presence of mixed male population can also be increased with an increasing TK♂∶WX♂ ratio.
As shown in Figure 4C, after retrieved TK♀ were mixed with TK♂ or no male, these two subgroups produced similar hatching rates in their second oviposition. Similarly, subsequent mixing of retrieved WX♀ with WX♂ did not significantly increase the hatching rates in the second oviposition compared to the subgroups of retrieved WX♀ kept alone (Figure 4D). These results indicate both TK♀ and WX♀ became refractory to subsequent mating after they mated with mixed male populations. This was similar to the scenario in which their first mating was with incompatible males only. Polyandry was not common or completely absent for these female populations. In addition, the average hatching rates in their subsequent ovipositions correlated with the hatching rates in their first oviposition. A higher number of incompatible males during the first mating event resulted in lower fecundity during the first and subsequent gonotrophic cycles.
The correlation between hatching rate and the proportion of incompatible males was plotted for both TK♀ and WX♀. As shown in Figure 5, based on the curves, the fecundity of TK♀ could be diminished when inundated by excess WX♂. 9× WX♂ can achieve nearly complete suppression of TK♀ fecundity. On the other hand, the maximal fecundity reduction of WX♀ by TK♂ is around 90% in one generation. The r2 value is 0.92 (P<0.05) for TK♀ curve (Figure 5A), and 0.9363 (P<0.05) for WX♀ curve (Figure 5B).
Since the incompatibility between TK and WX populations is not absolute, when incompatible males are released to suppress a target population, some hybrids will be generated. If these hybrids are compatible with the released males, then subsequent male release will help these hybrids to reproduce. This would pose a danger of gradually replacing the target population with a compatible hybrid population or creating a balance between the original target population and the hybrid population. In these cases, the populations are not effectively suppressed by male release. To test if the use of naturally incompatible populations as a control strategy is sustainable, the compatibility between the hybrids and TK and WX populations were measured.
In the case of releasing WX♂ to suppress TK population, both male and female F1 offspring will be generated from TK♀×WX♂ cross, designated as F1♀(TK♀×WX♂) and F1♂(TK♀×WX♂). These hybrids will encounter TK♀, TK♂ and WX♂, resulting in six possible mating combinations. Three extra mating combinations WX♀×F1♂(TK♀×WX♂), WX♀×WX♂ and WX♀×TK♂ that would not accompany this population suppression measure were also included in this set to provide more information about the crossing type of F1♂(TK♀×WX♂). These nine mating combinations were compared for fecundity. As shown in Table S5 and Figure 6A, the mating combinations F1♀(TK♀×WX♂)×F1♂(TK♀×WX♂) and F1♀(TK♀×WX♂)×TK♂ had comparably high hatching rates, while F1♀(TK♀×WX♂)×WX♂ produced no larva. These indicate F1♀(TK♀×WX♂) maintained the crossing type of TK♀. On the other hand, the average hatching rates of TK♀×F1♂(TK♀×WX♂) and TK♀×TK♂ were comparably high, while the hatching rates in WX♀×F1♂(TK♀×WX♂) and WX♀×TK♂ crosses were comparably low, indicating that F1♂(TK♀×WX♂) maintained the crossing type of TK♂. These results demonstrate that in TK♀×WX♂ cross, both male and female F1 offspring maintained their maternal crossing type.
Reciprocally, the strategy of using TK♂ to suppress WX population was tested for sustainability. To test the crossing type of F1♀(WX♀×TK♂) and F1♂(WX♀×TK♂) generated from the cross between WX♀ and TK♂, nine possible mating combinations were compared for hatching rate. As shown in Table S5 and Figure 6B, the average hatching rates of F1♀(WX♀×TK♂)×F1♂(WX♀×TK♂) and F1♀(WX♀×TK♂)×WX♂ were comparably high, while the hatching rate was low for F1♀(WX♀×TK♂)×TK♂ group. These indicate F1♀(WX♀×TK♂) maintained the crossing type of WX♀. On the other hand, the average hatching rates of WX♀×F1♂(WX♀×TK♂) and WX♀×WX♂ were comparably high, while the average hatching rates of TK♀×WX♂ and TK♀×F1♂(WX♀×TK♂) were comparably low. These indicate F1♂(WX♀×TK♂) maintained the crossing type of WX♂. In WX♀×TK♂ cross, both male and female F1 offspring maintained their maternal crossing type.
Taken together, when using incompatible males to suppress a target Cx. pipiens pallens population, hybrid offspring may be generated if the incompatibility is not absolute. However, these hybrids maintain their maternal crossing types. This phenomenon indicates that both original target population and inadvertently generated hybrids are subject to suppression by the released incompatible males, suggesting that this control strategy is sustainable.
The crossing types observed in these Cx. pipiens pallens populations were inherited maternally, suggesting the possibility that Wolbachia was the causal factor. To detect possible infections of Wolbachia, PCR was carried out using total DNA extracted from WX, NJ and TK populations. The primers were selected to amplify the wsp gene [36], [37] and the ank2 gene [38] according to published studies. For each population, 100% infection rate was detected (30 positive out of 30 tested mosquitoes for each population). The wsp sequence can distinguish supergroup A, wPip strains of supergroup B and wCauB strains of supergroup B. The ank2 sequence can further distinguish five phylogenetic groups (wPip-I to wPip-V) of wPip strains [38]. PCR results show that using the generic wF-wR primer pair a fragment around 600 bp was amplified from all three natural populations of Cx. pipiens pallens (Figure S2A). A fragment around 500 bp was amplified from all three populations using the wPip-specific wBpipF-wR primer pair. Neither wAF-wR primer pair (specific for supergroup A) nor wBcauBF-wR primer pair (specific for wCauB of supergroup B) generated any amplification product. These results are consistent with previous reports that most mosquito-infecting Wolbachia are wPip strains of supergroup B. This was confirmed by subsequent sequencing analysis, which also revealed that the wsp genes from these three populations are identical (GenBank accession numbers JX050185 - JX050187). PCR and sequencing analysis of ank2 gene revealed that NJ population was infected by Wolbachia of wPip-III group (GenBank accession number JX050182), while WX and TK populations were both infected by Wolbachia of wPip-IV group (GenBank accession numbers JX050183 and JX050184). The ank2 genes from WX and TK populations are identical (Figure S2B and Figure S3).
To confirm that the observed incompatibilities were caused by Wolbachia, the mosquitoes were treated with tetracycline [17]. The elimination of Wolbachia was checked by Hoechst 33342 staining [39]. As shown in Figure S4, eggs from untreated mosquitoes had strong fluorescence at both anterior and posterior ends, indicating Wolbachia were concentrated at these poles. In contrast, eggs from tetracycline-treated mosquitoes had even distribution of background fluorescence. No strong fluorescence was observed around the micropyle or at the posterior end. These results indicate that Wolbachia were removed from these mosquito populations. In addition, tetracycline-treated strains all tested negative for Wolbachia by PCR using wsp-specific primers wF and wR (data not shown). Crossing experiments were carried out using both Wolbachia-positive and Wolbachia-negative populations. WX females were crossed with Wolbachia-positive TK males (TK♂), Wolbachia-negative TK males (TKtet♂) and WX males. Similarly, TK females were crossed with Wolbachia-positive WX males (WX♂), Wolbachia-negative WX males (WXtet♂) and TK males. As shown in Figure 7, the hatching rate of WX♀×TKtet♂ cross was not significantly different from that of WX♀×WX♂ cross, but was significantly higher than that of WX♀×TK♂ cross. The hatching rate of TK♀×WXtet♂ cross was not significantly different from that of TK♀×TK♂ cross, but was significantly higher than that of TK♀×WX♂ cross. These results indicate that the bi-directional incompatibility between TK and WX populations is dependent on the presence of Wolbachia, i.e., it is Wolbachia-induced CI.
The GenBank accession numbers for sequences mentioned in the paper are ankyrin domain protein ank2 genes of Wolbachia in Nanjing, Tangkou and Wuxi populations of Cx. pipiens pallens (JX050182 - JX050184), and surface protein precursor wsp genes of Wolbachia in Nanjing, Tangkou and Wuxi populations of Cx. pipiens pallens (JX050185 - JX050187).
Biological control is a low-pollution component of integrated pest management. A variety of strategies have been developed to suppress populations of insect pests. Taking advantage of many insects' monogamous mating behavior, SIT uses sterile male insects to compete with normal males for their mates, which results in some females producing no or a reduced number of offspring. Two common methods to produce sterile male insects are irradiation and chemical treatments. SIT has been reported to be successful in a number of trials [13], [40]–[42]. One potential problem with SIT is the dosage of radiation or sterilizing agent used to treat male insects. Insufficient dosages might result in the release of fertile males while overtreatment might damage the males so much that they cannot compete with normal male insects. In addition, the residual chemicals in treated insects could potentially cause pollution to the environment and harm untargeted insect species [43], [44].
Alternatively, IIT exploits incompatibilities between insect populations and uses incompatible male insects to compete with compatible insects for their mates. Because IIT relies on the genetic traits of incompatible populations, it could be more reproducible than SIT. Most IITs are based on Wolbachia-induced cytoplasmic incompatibility. Infection with Wolbachia has been reported in many species, including some disease vectors such as Aedes albopictus and Culex quinquefasciatus ( = Cx. pipiens fatigans) [15], [21], [30]. For naturally uninfected insects, artificial Wolbachia infection would generate an incompatible strain that could be used to suppress target populations [28]. Introgression using Wolbachia-infected females and Wolbachia-negative target population males can also be used to generate an incompatible strain. For Wolbachia-positive target populations, a common approach is introgression with females from another strain infected with an incompatible Wolbachia. Sometimes, antibiotic treatment to eliminate Wolbachia from target strain is needed to facilitate the introgression process [29]. The advantages of this strategy is the generation of insect population with a desired host genetic background (often to match that of the target population) to increase the likelihood of survival in the target environment and successful competition with native males. However, both artificial infection and introgression can be technically difficult. Even with successful establishment of an incompatible strain, because of the placement of Wolbachia in a different host genetic background, the long-term stability of this new symbiosis needs to be tested.
In our study, we exploited incompatibilities between naturally existent populations from different geographic locations to suppress mosquito populations. This spares the time and effort to create incompatible populations. Another advantage is that the stability of Wolbachia can be assured since this symbiotic relationship has been naturally selected for a long time. IIT involves repeated release of incompatible males that can competitively inseminate females in a target population. Since these males are not meant to have offspring, any potential adverse effect of their genetic makeup on their fitness in the target environment would not be an issue so long as these males can live long enough to mate and are sexually competitive. We tested the mating behavior of three populations collected from different locations. When mixed with males of a different population, females became inseminated, as spermatozoa were present in the spermathecal capsules of these females. TK females mixed with NJ males or WX males laid comparable number of egg rafts, although these eggs had much lower hatching rates. NJ females or WX females also laid comparable numbers of egg rafts when mixed with TK males compared to when they were mixed with males of their own populations. Egg rafts with low hatching rate also showed clear embryonic development. All these data indicate that these males and females from different populations could successfully mate. Geographical and chronic isolation did not appear to create a mating barrier between these populations.
IIT works effectively only if the incompatible males are competitive enough. Although these mosquito populations could mate with each other, we also tested if the females had strong preference of their cognate males in the presence of both cognate and incompatible males using a mating competition experiment. Our results indicated significant mating preference was not observed. In Figure 5, a linear curve would indicate random mating, while a strong preference of cognate males would result in an arch curve with the median value greater than the average of the two extrema. Both curves are basically linear, supporting that these females chose their mates randomly. With an increasing ratio of incompatible to cognate males, the number of offspring decreased. Our results are consistent with previous reports from other groups that assortative mating usually does not occur in natural populations when numerous Wolbachia strains coexist within those populations [45]. These data suggest that although Wolbachia may cause CI in mosquitoes and skew survival in favor of those embryos that harbor the same or compatible Wolbachia, it does not affect the mate selection significantly. Instead, insects including mosquitoes select their mates based on their own genetic traits [46], [47]. The selection of suitable populations for mosquito control would depend on finding a population whose genome is compatible with that of target population, while also having a Wolbachia infection that induces cytoplasmic incompatibility in that context.
Another factor that influences IIT success is the monogamous mating behavior of females. After female mosquitoes are inseminated, they usually become refractory to re-mating. This phenomenon has been attributed to the effects of proteins in the seminal fluid received from male mosquitoes. But female monogamy is not always absolute. Polyandry has been reported in a number of mosquito species, such as Aedes aegypti and Culex tarsalis. The likelihood of female re-mating increases after these females go through gonotrophic cycles [48], [49], possibly due to the waning of seminal fluid proteins. One potential obstacle to the use of mosquitoes from different geographic locations to suppress target populations is the incompatibility between ligands in seminal fluid of one population and receptors in females of the other. In this scenario, insemination would not prevent the female mosquitoes from mating with reproductively compatible males. In our study, TK female mosquitoes became refractory to re-mating after being inseminated by WX or NJ male mosquitoes, even after these females had gone through a gonotrophic cycle, indicating seminal fluid proteins of WX or NJ population can act on receptors of TK population to cause post-coital behavioral changes in TK females. These results suggest that the seminal fluid proteins and their receptors are conserved enough between these populations, or their interactions are flexible enough to tolerate certain mutations. Further studies are needed to reveal the interactions between seminal fluid proteins and their receptors of different mosquito populations. Potentially, there are other factors causing female monogamy. If so, these factors seem to be functional across different populations.
Population suppression usually requires repeated effort. It is important for IIT to work for many generations. We tested if hybrids of incompatible populations became compatible with their parental populations. Our data indicate the incompatibility between different mosquito populations is hereditarily stable. This suggests that using males from incompatible natural populations to suppress a target population is sustainable. Although in our study, F1 offspring from incompatible crosses maintained the maternal crossing types, it should be noted that the preservation of maternal crossing type is not always the case. For example, F1♀ from the cross of bi-directionally incompatible Cx. quinquefasciatus strains Bei and Pel became fully compatible with Pel males [50]. This cautions us that when choosing incompatible populations, the crossing types of F1 hybrids should be carefully tested as well.
Although the climates at the sites where these mosquitoes originated from are different due to their latitudes, etc., these mosquitoes did not demonstrate significant behavioral differences in the same artificial environment of our insectary. More trials are needed to determine how they would behave and interact in semi-field and field trials. It remains to be tested whether incompatible male mosquitoes released into different natural environments will remain sexually competitive, and reduce the fecundity of local females.
Both maternal inheritance of crossing types and elimination of incompatibility by tetracycline treatment indicate that the observed incompatibility is caused by Wolbachia. To test if these mosquito populations are infected with different strains of Wolbachia, specific DNA fragments were PCR amplified from three mosquito populations. We sequenced wsp and ank2 genes, two specific genes commonly used to type Wolbachia. However, sequence analysis shows that wsp gene is identical in these three mosquito populations. The ank2 gene is identical between TK and WX populations, while ank2 gene of NJ population is different from TK or WX population. These Wolbachia are all wPip strains of supergroup B. It is currently unclear how they differ from each other. The definitive answer would require more sequencing or other typing tests which are beyond the scope of the current study. ANK genes (including ank2) were initially proposed to be involved in CI [50]. This hypothesis was challenged by subsequent analysis that failed to find any association of ANK polymorphism and CI [38]. Our data show NJ and WX populations are bi-directionally compatible, yet their ank2 genes are different. On the other hand, TK and WX populations are bi-directionally incompatible, yet their ank2 genes are identical (or at least the sequenced segments). These provide support to the view that homology between these typing markers does not correlate to the level of CI, suggesting molecular markers that are polymorphic and more closely associated with CI factors are yet to be found [38], [51]. Nuclear contribution to incompatibility has also been reported [50]. Being isolated from each other in the wild, these mosquitoes might have accumulated enough mutations in their genomes and/or even mitochondrial DNA to make some crosses incompatible. These differences in the host could directly cause incompatibility or more likely cause incompatibility by modulating Wolbachia gene expression. Consequently, although the incompatibilities between these populations are dependent on Wolbachia, contribution of host genomes cannot be ruled out at present. Our observation of embryonic development of incompatible eggs is in accordance with previously reported Wolbachia-mediated CI. It has been reported that incompatible eggs from two Wolbachia-positive populations have higher level of development than incompatible eggs from Wolbachia-negative females (which are usually undistinguishable from unfertilized eggs) [34]. We observed stemmata, bristles and segmentation in both TK♀×WX♂ and TK♀×NJ♂ crosses. This would be consistent with the multi-factorial mod resc model: even incompatible Wolbachia provides partial rescue in the eggs [15].
Although incompatibility between insect populations (including CI) is not fully understood, its applicability is promising. Our study proves that IIT is a feasible control strategy for Cx. pipiens pallens. The use of naturally occurring populations without genetic manipulation will save time and effort, and require less technical knowhow. This is an advantage for many less developed regions that deserves consideration. The wide distribution of mosquitoes in varied environments may be turned against them because it provides rich diversity in incompatibility; so that it is likely to find naturally incompatible and sexually competitive strains for many target populations. The conclusions from our study on Cx. pipiens pallens might offer reference to control measures of other mosquitoes, as well.
|
10.1371/journal.pcbi.1000938 | Drug Off-Target Effects Predicted Using Structural Analysis in the Context of a Metabolic Network Model | Recent advances in structural bioinformatics have enabled the prediction of protein-drug off-targets based on their ligand binding sites. Concurrent developments in systems biology allow for prediction of the functional effects of system perturbations using large-scale network models. Integration of these two capabilities provides a framework for evaluating metabolic drug response phenotypes in silico. This combined approach was applied to investigate the hypertensive side effect of the cholesteryl ester transfer protein inhibitor torcetrapib in the context of human renal function. A metabolic kidney model was generated in which to simulate drug treatment. Causal drug off-targets were predicted that have previously been observed to impact renal function in gene-deficient patients and may play a role in the adverse side effects observed in clinical trials. Genetic risk factors for drug treatment were also predicted that correspond to both characterized and unknown renal metabolic disorders as well as cryptic genetic deficiencies that are not expected to exhibit a renal disorder phenotype except under drug treatment. This study represents a novel integration of structural and systems biology and a first step towards computational systems medicine. The methodology introduced herein has important implications for drug development and personalized medicine.
| Pharmaceutical science is only beginning to scratch the surface on the exact mechanisms of drug action that lead to a drug's breadth of patient responses, both intended and side effects. Decades of clinical trials, molecular studies, and more recent computational analysis have sought to characterize the interactions between a drug and the cell's molecular machinery. We have devised an integrated computational approach to assess how a drug may affect a particular system, in our study the metabolism of the human kidney, and its capacity for filtration of the contents of the blood. We applied this approach to retrospectively investigate potential causal drug targets leading to increased blood pressure in participants of clinical trials for the drug torcetrapib in an effort to display how our approach could be directly useful in the drug development process. Our results suggest specific metabolic enzymes that may be directly responsible for the side effect. The drug screening framework we have developed could be used to link adverse side effects to particular drug targets, discover new uses for old drugs, identify biomarkers for metabolic disease and drug response, and suggest genetic or dietary risk factors to help guide personalized patient care.
| Despite the advantages gained from drug therapy in medicine, drug development has historically presented an expensive and frequently perplexing challenge for researchers. Identifying useful drug targets for treating disease and matching them to chemical compounds that can elicit the desired effect through drug-target interaction has been the paradigm for the drug development process in the era of molecular medicine. However, this approach has yielded many failed drug treatments and an incomplete understanding of the consequences of treatments for human health, even with drugs that have made it to market and been prescribed for decades. Two major contributing factors that confound individual molecular target-based drug discovery are drug off-target binding and the lack of systems-level understanding of drug response [1]. Adopting a new, systems-based approach to drug development is therefore a desirable goal in the era of systems medicine.
The growing wealth of omics data offers a valuable opportunity for novel approaches in systems medicine but also presents significant challenges for data integration [2]. Increasingly sophisticated computational approaches are being developed to analyze and manipulate omics data in order to gain a greater understanding of complex biological systems. An algorithm for identifying and comparing ligand binding sites on protein structures [3] was recently employed to predict drug off-target binding sites across the proteome [4]. Such a tool offers unique capabilities for drug development by providing a comprehensive survey of uncharacterized drug targets that may participate directly in drug response, which is likely to be important as polypharmacology interactions suggest that drug promiscuity is a predominant property of existing drugs [5].
Biological systems exhibit redundant pathways and synergistic effects conferring a robustness of phenotype when confronted with external stimuli. As a result, multi-target drugs are generally more clinically efficacious than single-target drugs. These facts highlight the critical importance of studying polypharmacology in a systems level context [6]. The increasing use of genome-scale metabolic network models for a variety of applications [7], [8] has established this research platform as a promising means for studying the emergent properties of complex systems. The published applications of metabolic models for drug development have thus far focused on identifying drug targets for antibacterial treatment in such pathogens as M. tuberculosis [9], [10], S. aureus [11], [10], H. pylori, and E. coli [10]. However, the human metabolic network reconstruction (Recon 1) [12] and developed context-specific metabolic modeling algorithms [13], [14] permit human-centered in silico drug studies. Integrating these structural bioinformatics and human system modeling techniques for application in drug development represents a first computational step into the era of systems medicine. As an example of this integrative approach, the results of protein off-target prediction for the drug torcetrapib [4], a cholesteryl ester transfer protein (CETP) inhibitor, were evaluated in the context of a model of renal metabolism.
CETP inhibitors are intended to treat patients at risk for atherosclerosis and other cardiovascular diseases by raising high-density lipoprotein cholesterol (HDL-C) and lowering low-density lipoprotein cholesterol (LDL-C) [15]. Torcetrapib was withdrawn from phase III clinical trials after a substantial investment of labor and capital due to its observed side effect of fatal hypertension in some patients [16]. It has since been of great interest to elucidate the cause of this side effect in order to avert such failures in the future and to better define the potential of CETP inhibitors for treatment [17]. Subsequent studies have provided evidence in favor of the hypothesis that the cause of this side effect was not due directly to the mechanism of HDL-C and LDL-C regulation via CETP inhibition [18]. Instead, it has been suggested that the hypertensive side effect may result from uncharacterized drug off-target effects [17]. Two other CETP inhibitors are now under clinical trial, anacetrapib [18] and JTT-705 [19]. Thus far, studies have not indicated the same risk of hypertension associated with the latter two drugs; however, these studies have been limited to relatively small patient groups lacking in diversity and over relatively short-term treatment. Even if these alternative CETP inhibitors do not carry the same adverse side effects, it is still of value to future drug development to determine the exact mechanism of torcetrapib's adverse action. It has been suggested that off-target effects of torcetrapib lead to increased activity of the renin-angiotensin-aldosterone-system (RAAS) and thereby hypertension [4], [20], but a recent review of the published CETP inhibitor clinical studies [21] concludes that the effect on blood pressure is most likely independent of the increase in aldosterone. Currently the exact cause of the hypertensive side effect of torcetrapib remains to be unambiguously identified.
The predicted torcetrapib off-targets include many metabolic enzymes and metabolite transport proteins. Although there are several mechanisms involved in regulating blood pressure that may be responsible for the hypertensive side effect, one possible mode is the renal regulation of blood pressure via metabolite reabsorption and secretion. The kidneys are the primary organs that filter the blood and therefore are strong contributors to maintaining a normotensive state even independent of RAAS function. Thus a model of renal metabolism was developed as the system context in which to analyze torcetrapib off-targets and predict drug response phenotypes. The two best-supported causal off-targets predicted in this study are prostaglandin I2 synthase (PTGIS), due to decreased capacity for renal prostaglandin I2 (PGI2) secretion, and acyl-CoA oxidase 1 (ACOX1), due to decreased capacity for renal citrate and amino acid reabsorption. Four other predicted off-targets are also predicted to impact amino acid, glucose, citrate, or bicarbonate reabsorption. As well, the model predicts no effect on renal reabsorption or secretion for a number of other predicted off-target metabolic proteins.
The goal of this study is not only to provide new insight into the torcetrapib problem but also to reveal the theoretical implications that this computational systems medicine platform has for drug development and personalized medicine. Characterizing the influence that genetic variation has in determining drug response phenotypes has been recognized as a crucial goal for the future of drug development [22]. To this end, the renal model was also used to analyze metabolic disorders resulting from genetic deficiencies and to identify those deficiencies that may pose additional risks for drug treatment in select individuals.
Although many of the predictions generated by this approach are supported by clinical and other experimental evidence that describe the impact of loss of function for predicted causal off-targets and genetic deficiencies, the full set of exact metabolic mechanisms of drug action predicted by our model remain to be completely validated. While this is seen as a limitation of this study, it also offers a number of opportunities to experimentally evaluate promising hypotheses that, if validated, will lead to significant advancements in developing CETP inhibitors for treatment and novel insight into certain renal disorders.
The approach for context-specific organ modeling proposed in this study (see Materials and Methods and Figure 1) yielded a renal metabolic model capturing functions of the kidney for reabsorption and secretion (Table 1). Many components of the renal objective function are factors known to be relevant determinants of blood pressure. However, there is currently incomplete knowledge about the exact role that some of these components play in blood pressure regulation. Calcium reabsorption, for example, leads to vasoconstriction in kidney glomeruli through the action of L-type and N-type calcium ion channels [23] suggesting a resulting increase in blood pressure if this mechanism applies across all vascular tissues. Calcium reabsorption also leads to an inhibition of renal sodium reabsorption in the proximal tubule [24] suggesting a blood pressure lowering effect consistent with the observation that increased dietary calcium also lowers blood pressure [25]. This highlights the complexity of the effect certain renal reasborptions have on blood pressure. Nevertheless, the many components accounted for in the renal objective function enabled explicit predictions about how system perturbations such as drug treatment and genetic deficiencies affect the kidney's ability to regulate the small molecule content of the blood.
The kidney model included 336 explicitly predicted active metabolic genes (Table S1) that met criteria for activity as summarized in Figure 2. The majority, 243 genes, satisfied the gene expression significance threshold (see Materials and Methods), although the activity of 58 genes was predicted despite expression values below the threshold. These genes were activated by the GIMME algorithm [13] to optimally achieve the renal objectives while remaining minimally inconsistent with gene expression data and may represent post-transcriptionally upregulated genes. The other 35 genes were predicted to be active without penalty since no corresponding probesets existed on the microarray upon which the transcriptomic data was obtained. Since many of these genes participated in optimal pathways for achieving renal objectives, it is projected that experimental measurement would confirm their expression if performed.
The active reactions in the model reflect both the possible pathways by which the kidney can achieve the specified renal objectives as well as other functions supported by the gene expression data. The model included 1587 active reactions (Table S2), excluding model-based reactions such as objective functions, exchanges, and demands. Of these active reactions, 333 comprised a single connected sub-model accounting for all pathways which could possibly support the specified renal objectives. We refer to this sub-model as the reduced kidney model (see Table S1 and Table S2 for the contents of the reduced model and Dataset S1 for the actual model in SBML format). It should be noted that because the reduced model included all reactions that can carry flux in support of the renal objectives, it had the exact same effective flux state solution space as the full renal model. The reduced kidney model reactions spanned a broad range of metabolic subsystems (Figure 3). The largest subsystem consisted of plasma membrane-spanning transport reactions, which is expected given that this model captured renal filtration and secretion functions. The second largest subsystem represented intracellular transport, signifying the importance of interaction among sub-cellular compartments in renal function including the cytosol, endoplasimic reticulum, Golgi apparatus, and mitochondria. A significant proportion of the other active subsystems in the reduced kidney model were involved directly in the metabolism of components of the renal objective function including carbohydrate, amino acid, vitamin, lipid, carboxylate, and glutathione metabolism as well as the urea cycle. These permitted the indirect reabsorption of metabolites as well as the required synthetic pathways for renal secretions.
The integrative framework adopted for predicting causal drug targets associated with response phenotypes employed both structural bioinformatics tools as well as modeling techniques of systems biology (see Materials and Methods and Figure 4). The workflow begins with screening of the entire human structural proteome, with each subsequent step in the process narrowing the list of proteins ultimately into a set of targets for which a response phenotype was predicted upon functional inhibition. The first step of this process identified putative off-target drug binding sites using a ligand-binding site structural alignment algorithm (see Materials and Methods). The 41 predicted metabolic protein off-targets were the focus of this study (see Table S3), 28 of which had predicted drug binding sites overlapping with their functional sites. Simulated inhibition of these targets in the reduced kidney model (see Materials and Methods) predicted response phenotypes for 6 of the off-target proteins with respect to renal function (Figure 5). The results of all analysis steps for these 6 off-targets are summarized in Table 2. The expression of all of these targets was determined to be the most limiting for their associated metabolic reactions included in the reduced kidney model (see Materials and Methods), providing additional evidence supporting that inhibition of these targets would be expected to have at least some deleterious impact on those reactions.
The renal response phenotypes for inhibition of two of the predicted drug off-targets were supported by existing scientific literature. Simulated PTGIS inhibition completely precluded PGI2 secretion. Based on the relation of renal PGI2 secretion to blood pressure (see Table 1), this inhibition would be expected to have a hypertensive effect. Experimental studies confirmed that PTGIS is associated with essential hypertension in humans [26] and that transgenic rats highly expressing human PTGIS exhibited decreased mean pulmonary arterial pressure despite treatment with monocrotaline to induce hypertension [27]. Inhibition of hydroxyacid oxidase 2 (HAO2) in the reduced kidney model led to reduced glutamate, glycine, and serine reabsorption suggesting a possible role for HAO2 in the hypertensive side effect following CETP inhibitor treatment based on the association of amino acid reabsorption with vasodilation and hypertension (see Table 1). HAO2 is highly expressed in human kidney [28] and was identified as a candidate quantitative trait locus for blood pressure in rat kidney in a study comparing normal to hypertensive rats [29].
Two predicted causal CETP inhibitor off-targets, PTGIS and ACOX1, exhibited notable binding affinity differences when comparing docking results for their endogenous substrates to those for the three CETP inhibitors (Figure 6). The mean predicted binding affinity of PTGIS for its endogenous substrate prostaglandin H2 was weaker than for all three CETP inhibitors (Figure 6A). Anacetrapib was predicted to have the strongest mean binding affinity of all four tested molecules for PTGIS and JTT-705 the weakest of the three drugs. The predicted mean binding affinity of ACOX1 for its endogenous substrate palmitoyl-CoA was weaker than for torcetrapib and anacetrapib but stronger than the affinity of the protein for JTT-705 (Figure 6B). These results supported potential competitive inhibition of PTGIS and ACOX1 by torcetrapib and anacetrapib, but the predictions suggested a lesser effect of JTT-705 on ACOX1.
Similar to the use of the model to test inhibitory effects on drug targets, the model was also used to predict genetic deficiencies that lead to renal disorders and drug off-targets that act synergistically with genetic deficiencies. Simulated gene knockouts predicted to impact renal objective functions are displayed in Figure S1, Figure S2 and Table S4. The 118 deficient genes predicted to cause disorders impacted a variety of renal secretions and absorptions to varying degrees. Thirteen of these deficiencies predicted total loss of at least one renal function (see Figure S2).
Some renal disorders were only predicted in the gene-deficient models in combination with drug treatment, not in the untreated gene-deficient models or in the normal drug-treated model, and are referred to in this study as cryptic genetic risk factors. Five such gene deficiencies were predicted (see Table S4). A deficiency in CYP27B1, which impacted vitamin D secretion alone, also exhibited defects in proline reabsorption when combined with drug treatment in simulation. Defects in three amino acid transport proteins (SLC7A10, SLC3A1, and SLC7A9) were predicted to decrease renal glycine reabsorption in combination with drug treatment along with the disorders predicted in the absence of drug treatment. The model deficient in the ATP-binding cassette sub-family C member 1 gene (ABCC1) was predicted to exhibit a cryptic deficiency in renal phosphate reabsorption under drug treatment. These predictions are of special importance because they suggest that these renal phenotypes would only surface in gene-deficient individuals under certain conditions, such as when treated with CETP inhibitors.
Multiple evaluations were performed to analyze and validate the content of the reduced kidney model. The reduced kidney model effectively predicted activity of significantly expressed metabolic genes. The ability of our modeling approach to correctly and robustly predict activity of highly expressing genes was evaluated by a five-fold cross validation (see Materials and Methods). Our approach showed significant recall of the 20% most highly expressed metabolic genes, p-value = 4.57×10−22. This observation is especially notable since the reduced kidney model was not a global model of kidney metabolism, and the result suggests the relative importance of the renal functions captured by our model within the context of total kidney gene activity.
We compared the metabolic gene activity predictions from the reduced kidney model to the set of significantly expressed genes as well as to a proteomic dataset derived from normal, healthy human kidney glomerulus tissue [30] (Figure 7A). A total of 164 genes active in the reduced kidney model, 72% of the predicted activities, were supported by either significantly expressed mRNA levels, high-confidence protein detection, or both (see Table S1 for a detailed list). The remaining 64 gene activities accounted for in the model include 23 genes with no corresponding microarray probesets, and therefore not experimentally measured mRNA, and 41 genes that were determined to express more marginally below the established significance threshold. Despite a strong overlap between the transcriptomic and proteomic datasets, there were also large proportions of both which are unique. This disagreement may be due to tissue samples being taken from different kidney sub-tissues in each experiment, absent probesets on the microarray, or the propensity of mass spectrometry proteomic experiments to produce false negatives. All of the counted activities in Figure 7A were included in the full human metabolic network, signifying that the reduced kidney model was not a global kidney model and that there is potential for expansion to account for more metabolic functions than those of concern in this study.
The literature-curated renal functions achievable by the kidney model were also compared to those achievable by a model derived from the predictions of Shlomi et al (Figure 7B). While the kidney model developed in this study was compatible with all 41 curated renal functions, the predictions of Shlomi et al were only compatible with 25 functions. This difference in functionality was due to false negative inactivity predictions made by Shlomi et al such as inactive urea transport, prostaglandin synthesis, and ATP synthesis. These results underscore the need to manually curate automatically generated metabolic network reconstructions and the advantage of integrating objective functions with context-specific modeling.
Next, the model was functionally validated by comparing the gene deficiencies predicted to cause renal disorder to disease phenotypes in the OMIM database collected from clinical studies. Twenty known gene deficiencies leading to specific disease phenotypes were accurately predicted using the model (see Table S4). Loss of function mutations in the gene encoding 25-hydroxyvitamin D3-1-alpha hydroxylase (CYP27B1) have been linked to vitamin D-dependent rickets type I in both human patients [31] and pigs [32] consistent with the predicted inability of the gene-deficient model to secrete calcitriol. Hypouricosuria, low urinary excretion of urate, is a symptom of xanthinuria that is caused by xanthine dehydrogenase (XDH) deficiency [33], which is consistent with the deficient model's inability to excrete urate. Similarly, hypouricemia, low blood serum urate, is a consequence of nucleoside phosphorylase (NP) deficiency [34] also predicted in the model. Deficiency of aromatic L-amino acid decarboxylase (DDC) leads to increased urinary excretion of 5-hydroxytryptophan [35], which is consistent with the decreased ability to reabsorb tryptophan and secrete tryptamine predicted through simulation. Mutations in the mitochondrial cytochrome c oxidase gene (COX6B1) lead to de Toni-Fanconi-Debre renal syndrome, whose symptoms include a deficiency in the renal reabsorption of glucose, amino acids, and bicarbonate [36], [37], all of which were predicted in the model. Deficiencies in seven NADH dehydrogenase genes all lead to hypoglycemia, confirmed in simulation, and a decreased ability to oxidize citrate and glutamate [38], reactions important for indirect renal reabsorption of citrate and glutamate in the model. Proline dehydrogenase (PRODH) deficiency causes an inability to oxidize proline in kidney and other tissues leading to hyperprolinemia that includes increased urinary excretion of proline as a symptom [39]–[41], which is also consistent with the predicted decrease in renal proline reabsorption. Deficiencies in two genes that take part in the ubiquinol-cytochrome c reductase complex III (UQCRQ and UQCRB) lead to proximal tubulopathy, including an inability to reabsorb amino acids [42]; the gene-deficient model exhibited reduced renal reabsorption of alanine, glutamate, and proline. Fumarate hydratase (FH) deficiency leads to defects in glutamate oxidation in kidney and other tissues [43], [44], which is also consistent with the decreased indirect renal reabsorption of glutamate predicted by the model. Renal glucosuria, recapitulated in the model, results from deficiency in a sodium-glucose transporter (SLC5A2) [45]. Dicarboxylicamino aciduria [46] exhibits impaired renal glutamate and aspartate reabsorption and hypoglycemia resulting from a deficient glutamate transporter (SLC1A1), all symptoms predicted by the model. Severe dehydration is one symptom resulting from another deficient transporter (SLC5A1) [47], confirmed through decreased reabsorption of water in the model. These results qualitatively describe the ability of our modeling approach to predict perturbed phenotypic states.
To more rigorously quantify the predictive ability of our model simulation approach, we performed area under receiver operating characteristic (AROC) analysis based on not only the abovementioned clinical validations of our gene-deficient phenotype predictions but based on the entire set of such known clinical phenotypes that could potentially have been investigated using our model (see Figure S3 and Materials and Methods). The sharp declines in rates with increasingly stringent classifier ratio thresholds (see Figure S3) reflect the likely low coverage of actual disorder phenotypes by existing clinical studies. Nevertheless, our approach performed very well based on this analysis, with an AROC of 0.7565. Permutation trials resulted in a mean AROC of 0.5112, in close agreement with the expected theoretical randomly achievable AROC of 0.5. Our approach achieved a significantly greater AROC than could be expected by chance, p-value = 8.71×10−70. Given the relatively low number of actual clinical negatives available (see Table S5), we also assessed the significance of our prediction results based purely on the true positive rates determined through the AROC analysis. The mean true positive rate of our results in this analysis was 0.2859, significantly greater than the 0.0215 mean true positive rate obtained randomly, p-value = 3.29×10−127. These analysis results illustrate that our approach for predicting perturbation phenotypes exhibits both favorable sensitivity and specificity based on actual clinical data and should hold not only for predicting genetic deficiency phenotypes but also enzyme inhibition by drugs, which exhibits a similarly deleterious phenotypic effect.
In order to assess the effects of some of the critical assumptions made in the model development and simulation procedures, we performed sensitivity analysis with respect to the predicted renal disorder phenotypes.
First, we compared the predictive capability of our reduced kidney model to that of the original, unconstrained human Recon1 metabolic network. The same approach to simulating renal disorder states was employed using both models (see Materials and Methods). We simulated all single gene knock outs in both models and assessed the renal disorder phenotypes with respect to each individual component of the renal objective function based on the ratio of maximum objective flux in the perturbed state to maximum objective flux in the unperturbed state. Comparing the results achieved by each model (Figure 8), it is apparent that although there are a few cases where both models predict an equal degree of renal disorder given the same genetic perturbation, the vast majority of disorder phenotypes are more apparent in the reduced kidney model than in Recon1 alone. In fact, 427 out of the 608 (71%) disorder phenotypes predicted by the reduced kidney model showed no degree of disorder relative to the unperturbed state in Recon1, including 36 of the most severe phenotypes for which a total loss of renal function was predicted by the reduced kidney model. These observations display the predictive ability gained through integration of the gene expression data via the GIMME algorithm, incorporating metabolomics data to set exchange constraints, and the addition of six key membrane transport reactions during the limited function-enabling manual curation of the model. These reactions involve the transport of prostaglandins I2 and H2, calcitriol, and carnosine. It should be noted that the 7 disorders for which Recon1 predicted a more severe phenotype than the kidney model result directly from the addition of these transporters in that these transporters have enabled additional pathways in the kidney model that are absent in Recon1. All but one of the predictions concerning CETP inhibitors showed a clearer phenotype in the kidney model as well; this off-target is PTGIS for which both models predict a complete loss of function when fully inhibited. Finally, 28 out of the 33 clinically validated phenotypes are predicted more noticeably by the kidney model, 17 of these showing no disorder phenotype in Recon1. Overall, this comparison establishes the relative contribution of context-specific modeling in studying disorder and drug response phenotypes.
Second, we investigated the sensitivity of our drug off-target response phenotype predictions to the variability of two important parameters used in our simulations, the system boundary flux constraint, set as equal fractions of the upper bound on renal objective fluxes (see Materials and Methods), and the degree of enzymatic activity inhibition assumed to result from drug treatment.
The system boundary flux constraint was imposed upon demand and exchange reactions other than those optimized during a given simulation. By default we set this constraint assuming that all allowed boundary fluxes can carry an equal fraction of the potential maximum renal objective flux. This assumption was made to allow all pathways that could possibly contribute to the objective to be used simultaneously in the optimal flux state, providing the most flexible state while maintaining maximum sensitivity of our model to additional system perturbations such as gene deficiencies or drug effects. This approach was unbiased in that it did not favor any possible pathway over another in achieving a set objective without imposing additional constraints, which may not always reflect biological reality but was the most conservative assumption in the absence of additional experimental data required to more precisely set these flux constraints. In our sensitivity analysis, we varied this parameter between 0 and 1000 flux units, the absolute lower and upper magnitudes possible in our model, and repeated the simulations of drug off-target effects. The result of this analysis (Figure S4) was captured in the normalized sensitivity coefficient computed for each simulation (see Materials and Methods). The coefficient can vary between negative and positive unity and displays the deviation from a base result, the primary predictions we have presented in this study. The base result is indicated by a black star in Figure S4, and the parameter value in this case equals 13.5 flux units.
It is clear from Figure S4 that PTGIS inhibition resulted in the same renal disorder phenotype regardless of the value of the system boundary flux constraint parameter. This was because there was only one pathway in the model by which prostaglandin I2 could be secreted. Most other disorder phenotype predictions begin to diverge from the base result around a parameter value of 200 flux units, a fairly permissive value, which shows that the predictions were fairly robust to variability of this parameter. The closer to 1000 flux units this parameter was set, the more completely alternative pathways could compensate for a loss of function in the simulations. If alternative pathways existed to achieve a renal function, it was guaranteed that the ability to predict a disorder phenotype with respect to that function would be completely lost at the maximum possible parameter value of 1000.
We similarly analyzed the sensitivity of our predictions to changes in the degree of enzyme inhibition assumed to follow from drug treatment (Figure S5). For the primary results presented in this study, we assumed complete inhibition of activity by the drug, corresponding to a fraction of maximum enzymatic reaction flux equal to 0 in Figure S5. Similar to the default setting of our system boundary flux constraint, this default of complete inhibition was chosen in order to maximize the sensitivity of our model in detecting disorder phenotypes. Most of the phenotypes were still detectable to varying degrees with as much as 25% of the maximum activity of drug targets. The predicted phenotypes associated with PTGIS, ACOX1, and AK3L1 were especially robust to variation in degree of inhibition, still exhibiting a phenotype near 50% of maximum activity. Decreased glucose and bicarbonate reabsorption under drug-induced MT-COI and UQCRC1 inhibition exhibited the most sensitivity to variability in this parameter, although none of the predicted phenotypes required complete inhibition of the drug target in order to be detected.
A novel approach for making functional predictions of drug response phenotypes has been introduced that integrates techniques of both structural bioinformatics and systems biology. Although the current study focused on a specific metabolic system, the general methodology excluding techniques particular to metabolic modeling are extensible to other systems such as signaling or transcriptional regulation. Non-metabolic protein drug off-targets are predictable using the same structural analysis tools, and many such off-targets have indeed been predicted as well for CETP inhibitors [4].
The context-specific organ metabolic modeling strategy employed in this study represents an improvement upon previous efforts in this realm. Model development algorithms such as GIMME [13] or that developed by Shlomi et al, when integrated with multiple omics datasets, can lead to more biologically realistic models. It is also of critical importance to include context-specific metabolic objective functions in the model development process in order to yield a fully functional and predictive model, as is evident from the functional comparisons of models performed in this study.
As an early effort at modeling such a context-specific metabolic system it is important to discuss the limitations of our model. Although the functional validations presented here are compelling, currently available clinical data only permits the assessment of a subset of the predictions possible in the model. Also, the functional portion of the model, the reduced kidney model, does not and is not intended to represent a global model of kidney metabolism but only the specific renal functions studied in this work. As such, our model does not fully resolve of complexity of the human kidney. The human kidney fulfills a number of functions not studied here and is a spatially distributed system across multiple distinct tissue types. Here we have summarily replaced the various kidney sub-tissues with a single, net system model. Because we integrated expression data with curated renal functions that operate across multiple kidney tissues, it is likely that our model approximates a superset of the metabolic pathways supporting these functions. Although we have made several simplifying assumptions in the model development process, even the current level of model validation suggests that the gene and reaction content of the model is fairly accurate and that simulations in this model indeed hold predictive capability.
The simulation approach taken, optimization of a linear objective function, does not fully capture the full physiological role of the kidney. The goal of these simulations was to determine drug-target effects that may limit the capacity of the kidney to move towards a homeostatic nominal state from a state of high blood pressure, thereby decreasing the capacity of the kidney to lower blood pressure. This strategy is appropriate for the goals of the current study but would not be appropriate to simulate all physiological states of interest in the kidney. On a related note, the choice to define a disorder state based on the ratio of perturbed to unperturbed maximum achievable renal objective flux demonstrates a difference in the capacity of the renal function and not necessarily a precise flux state. Therefore this strategy too is not appropriate for modeling all physiological states.
The predictions made for CETP inhibitors in this study serve as illustrative examples of many important implications that this approach has for drug development and personalized medicine. Predicted causal off-targets for renal metabolic disorders related to blood pressure may be responsible in part or full for the clinically observed hypertensive side effect of torcetrapib. The evidence resulting from this study suggests that PTGIS and ACOX1 are both potential causal torcetrapib off-targets, the inhibition of which may explain the side effect of hypertension. In addition, AK3L1, HAO2, MT-COI, and UQCRC1 may also play a role in this side effect as we have predicted, although our docking trials did not suggest that they are bound as strongly by torcetrapib. The specific predicted deficiencies in renal function associated with the drug off-targets can serve as biomarkers to be measured in patients participating in clinical trials. A positive correlation of these biomarkers with side effects would lend support to the predictions of this study and confirm these biomarkers as risk indicators in future patient treatment. It is important to note that although these predictions comprise the basis for a renal filtration and secretion-based hypothesis explaining the hypertensive side effect of torcetrapib, these results do not refute the hypothesis based on a RAAS-mediated mechanism. These two hypotheses are not mutually exclusive and could potentially contribute alternatively or synergistically to the clinically observed side effects. This possibility illustrates the major tenet for systems biology: studying a single protein or even a single pathway is not necessarily sufficient to explain complex biological phenomena.
Aside from the confirmation that some of our predicted off-targets are known to be involved in renal disorders, we do not currently present direct experimental verification that torcetrapib binds and inhibits the predicted targets and that this inhibition leads to the predicted response phenotypes. Although this would be the obvious next step, a retrospective validation is currently hampered by the availability of the drug and the nature of the phenotypes both predicted and known. Ideally, relevant physiological studies would be carried out during actual clinical trials, when a method such as ours would be most useful, in preclinical and clinical phases of drug development.
The extended structural analysis of causal drug off-targets to identify differential binding affinities for endogenous substrates and drug molecules suggests possible differences in drug response phenotypes across the CETP inhibitors tested. The results suggest that anacetrapib may potentially lead to a similar response phenotype to that of torcetrapib, while JTT-705 may not carry the same adverse effect, at least with respect to the off-targets detailed in this study. This particular type of analysis may aid in differentiating between likely response phenotypes expected for chemically and functionally similar drugs. Results of the computational pipeline for interaction prediction between proteins and CETP inhibitors employed in this study, SMAP and docking, have yet to be confirmed experimentally. Although we are currently unable to provide direct experimental evidence for the off-target interaction predictions for this class of drugs, multiple recent studies have shown experimental support for the general efficacy of this approach for interaction prediction [48], [49].
The predicted renal metabolic disorders with a genetic basis suggest classes of individuals in which treatment with CETP inhibitors may pose a higher risk for adverse side effects. These predictions suggest a likely relationship between participants in torcetrapib clinical trials exhibiting symptoms of these disorders and the observed adverse side effects. The concept of cryptic genetic risk factors for drug treatment introduced in this study suggests a novel approach to personalized medicine. Should polymorphisms within these genes be clinically linked to side effects of drug treatment, the result would comprise a basis for genetic screening to assess the risk of drug treatment for future patients. Given that these cryptic risk factors are not expected to elicit the predicted abnormal phenotypes in the absence of drug treatment, identification of causal polymorphisms through association studies could only occur during clinical phase when a sufficient number of patients could be observed to gain the statistical power needed to draw significant correlations.
As illustrated above, this approach for in silico drug testing could become an indispensible tool during the pre-clinical and clinical phases of new drug development for studying the nature of adverse side effects. In addition, this platform holds obvious potential for analyzing drug efficacy in general and identification of potential beneficent drug side effects that may be useful for drug repositioning and could also be easily adapted for studying combinatorial drug treatment. For a failed drug like torcetrapib, results from this approach could reinitiate the drug development process, providing new insight to help target patients who could benefit from the treatment without the risk of serious adverse side effects.
The binding site for CETP inhibitors on the CETP structure and the predicted off-target binding sites for this class of drug across the proteome were assumed to be as previously predicted using the SMAP program [4], which implements the Sequence Order Independent Profile-Profile Alignment (SOIPPA) algorithm to identify significant structural similarity to a given ligand-binding site [3]. The results contained proteins from all organisms represented in the PDB, not just human structures.
In order to integrate the result of drug off-target predictions with the metabolic network, it was necessary to first map all PDB structures (http://www.pdb.org) corresponding to human metabolic proteins included in Recon1, downloaded from the BiGG database, to their respective gene identifiers as represented in Recon1. The BiGG database requires registration and a password, which can be requested by visiting (http://bigg.ucsd.edu/bigg/home.pl). The UniProt ID mapping tool (http://www.uniprot.org/) was used to map PDB structures corresponding to human proteins to gene identifiers linked to metabolic reactions in Recon1 accounting for all predicted human metabolic protein drug off-targets. All non-human predicted metabolic protein drug off-targets were mapped to their human orthologs using the Basic Local Alignment Search Tool (BLAST) [50] to perform a bi-directional BLAST with a mutual best hit criterion. BLAST was also used to resolve inconsistencies in functional annotation between Recon1 gene-protein-reaction associations (GPRs) and gene annotations from the Entrez Gene database (http://www.ncbi.nlm.nih.gov/sites/entrez?db=gene) with respect to predicted drug targets, leading to the reannotation of three Recon 1 GPRs. The overall result of this mapping was that 97 metabolic reactions in Recon1 were linked to 41 predicted CETP inhibitor off-targets.
The metabolic enzymes predicted as CETP inhibitor off-targets using SMAP were evaluated to determine potential enzymatic inhibition by the drug. The predicted drug-binding sites of the putative off-targets were compared to endogenous ligand-binding sites from existing PDB protein-ligand complex structures (http://www.pdb.org) and catalytic sites from the Catalytic Site Atlas (http://www.ebi.ac.uk/thornton-srv/databases/CSA/). Ligand-binding sites were defined as amino acid residues lying within 4.5 Å from atoms of the ligand. Drug-binding sites were defined as residues aligned with the cholesteryl ester binding sites on the CETP structure using SMAP. Overlap between endogenous ligand-binding sites and drug-binding sites was defined by a sharing of any amino acid residues between the sites. The function of predicted drug targets present in Recon1 with at least a partial such overlap was considered to be competitively inhibitable by the drug.
Enzyme substrates were identified from Recon1 reaction formulas. Certain molecules (H+, H2O, O2, phosphate, ferricytochrome C, and ferrocytochrome C) were excluded from docking trials due to size or structural challenges prohibiting a useful docking result for the purposes of binding affinity predictions. All protein structures used in this study were downloaded from the PDB (http://www.pdb.org). Three-dimensional structures for endogenous enzyme substrates were downloaded directly from the PDB if available. If the PDB ligand structure did not exist or was non-functional for docking, the structure was searched for in PubChem (http://pubchem.ncbi.nlm.nih.gov/). The subsequently downloaded SDF file was converted to PDB format using the ChemAxon web applet available at the PDB website (http://www.rcsb.org/pdb/ligand/chemAdvSearch.do). If the three-dimensional ligand structure could not be found in PubChem, the two-dimensional structure was derived from the canonical SMILES [51] representation of the compound available in PubChem and then converted to a three-dimensional structure in PDB format using the Clean3D Fine Build tool available through the Marvin web applet (http://www.chemaxon.com/marvin/sketch/index.jsp). The three-dimensional structures for glycolipids were derived from their KEGG glycan structures (http://www.genome.jp/kegg/glycan/) using SWEET-II (http://www.glycosciences.de/spec/sweet2/doc/index.php).
Protein structures were pre-processed for docking using AutoDockTools (ADT) version 1.5.2 [52] by adding polar hydrogen atoms, removing all non-protein molecules from the PDB structure including water, detergents, and ligands, adding Kollman charges to the protein and converting it to PDBQT format. Ligand structures were also prepared using ADT, using the default method for preparing ligands for docking that adds hydrogens and charges. The default rotatable bonds were accepted as well, and the structure was converted to PDBQT format. The search space for docking was determined visually by centering the Grid Box in ADT central to the experimentally determined binding site of the ligand and expanding the dimensions of the cubic search space to just completely encompass the site.
Docking was performed using AutoDock Vina [53] with default parameter settings other than the search space specification described above, and the mean predicted binding affinity from the set of predicted binding poses was accepted as the true binding affinity for each docking run. The predicted binding affinities for endogenous substrates were compared to the affinity of the same site for the CETP inhibitor drugs in order to make predictions about differential responses with respect to each of the drugs.
As the preliminary step in modeling human renal function, the scientific literature was reviewed to compile a list of specific metabolic functions of the kidney, with a focus on functions implicated as determinants of blood pressure. This list includes a number of renal reabsorptions and secretions. Each function in this list was tested for compatibility with Recon1, downloaded from the BiGG database (http://bigg.ucsd.edu/bigg/home.pl), by performing flux balance analysis (FBA) on the fully unconstrained network optimizing for the given function. Those functions compatible with Recon1 were those that could achieve a positive flux and are summarized in Table 1. These metabolic functions were combined with a basic ATP maintenance function to form a single model reaction that represents the kidney's ability to filter the metabolic content of blood with preference for lowering blood pressure. This model reaction was used as the objective function in developing the metabolic kidney model and is referred to as the renal objective function in this study. All stoichiometric coefficients in this reaction were set equal to one, which is a safe assumption for the model development step as this only significantly impacts the magnitude of fluxes through pathways that support each individual renal objective and not generally whether or not certain fluxes will be active in the resulting model. For the full renal objective function reaction to be seen as useful in performing simulations, more careful balancing of these coefficients based on experimental evidence would be required. As such, the full renal objective function was not used in any subsequent simulations with the model, instead being substituted as an objective by the reactions representing individual reabsorptions or secretions.
Metabolite exchange and transport reactions needed to achieve some of the renal functions were also added to the network. It was observed that Recon1 as a base model could not achieve flux through certain key renal metabolite reabsorptions: sodium, calcium, chloride, potassium, and oxalate. These deficiencies were corrected for by simply adding demand fluxes for these metabolites in the cytosol model compartment. Demand fluxes were also added for the remaining kidney reabsorptions and secretions as well to enable an array of simulations involving individual components of the renal objective function to be tested. In the case of reabsorption, this allows for direct reabsorption of metabolites in addition to indirect reabsorption in which the absorbed metabolite is first metabolized into other compounds and then reabsorbed into the blood, as is the primary mechanism of reabsorption for some metabolites, such as reduced glutathione (GSH) [54].
A preliminary model was created by imposing kidney-specific exchange flux constraints representing the metabolic exchanges the kidney carries out with the blood and urine. The preliminary model was initialized by loading Recon1 into the COBRA Toolbox [55] and, by default, unbounding all reaction fluxes by setting them to the default maximum magnitude of 1000 flux units. Next, the renal objective function was added to the network as a single reaction. Exchange fluxes for kidney secretion objectives were constrained to preclude uptake of those metabolites to achieve the renal objective, forcing the model to synthesize those metabolites in order to secrete them. The resulting preliminary model included 407 exchange fluxes, only 49 of which were explicitly unconstrained based on literature-curated kidney functions and the most basic of metabolic precursor requirements. The basic metabolic exchanges assumed to take place include ions and other inorganic compounds.
The Human Metabolomics Database (HMDB) (http://www.hmdb.ca/) was queried to derive further evidence in support of allowable exchange fluxes for the kidney. All 407 exchange metabolites in the preliminary model were searched in HMDB for experimental detection in specific biofluids and tissues. Those metabolites detected both in the blood and kidney tissue were assumed to be freely exchangeable in the kidney model, leading to 78 more explicitly unconstrained exchanges beyond what was derived from basic and curated kidney-specific metabolic functions. This assumption is based on the kidney's physiological role of filtering the blood and the observation that if both the blood and kidney contain a metabolite, it must either be exchanged between the two or synthesized separately in both. In the former case, this data provides evidence of that exchange. In the later case, although the model might allow a biologically unrealistic exchange, because the metabolite exists in both blood and kidney, the impact on simulations using the resulting model should be merely quantitative in terms of the maximum renal objective fluxes achievable by the unperturbed model. The integration of gene expression data in the model development process described below should reduce the propensity for biologically unsound metabolic pathway activation that could follow from precursors introduced by any biologically unsound exchanges. Those metabolites detected both in the urine and kidney were assumed to be possible excretions, and exchange constraints were set accordingly. Excretions determined utilizing the urine metabolomics data mostly showed redundancy in determining exchange constraints with exchanges determined using blood data or literature curation with the exception of 4 additional metabolites. The remaining 276 exchange fluxes for which no evidence was found to support were tentatively constrained to 0 flux units.
The resulting preliminary model was again tested for the ability to achieve all kidney-specific metabolic functions. It was found that this model could not absorb and metabolize GSH, without also absorbing oxidized glutathione, the exchange of which was subsequently unconstrained. Also, L-threonine and L-methionine could not be absorbed and metabolized in this model without exchange of 2-hydroxybutyrate and 2-methylcitrate, the exchanges of which were similarly unconstrained as a corrective measure. The resulting preliminary model could still achieve all the same renal objectives as the fully unconstrained model. As a final preliminary constraining measure, all system effluxes were bound to equal fractions of the default upper bound on influxes of 1000 flux units; we term this parameter the system boundary flux constraint. This was done so that any available direct or indirect reabsorption pathways could possibly be used to achieve metabolite reabsorption without biasing the model towards use of any particular pathways without further evidence. This represents the state of the model just prior to final processing using the GIMME algorithm. The fitting of the allowable fluxes to the gene expression data by GIMME ultimately determined the usable reabsorption and secretion pathways in accordance with gene expression.
Two gene expression microarray dataset for normal, healthy kidney tissue [56] were obtained from the GEO database (http://www.ncbi.nlm.nih.gov/geo/), accession GSE803. The two background-subtracted datasets were first normalized using a global normalization factor equal to the sum of probe intensities from the first dataset divided by the sum of probe intensities from the second dataset to account for any systematic differences in procedure between the two experiments. The resulting data were then normalized using the Lowess method [57] to reduce random noise. The resulting normalized datasets were then weighted equally as replicates in determining the final data for integration with the human metabolic network by taking the mean of the two normalized datasets.
The gene-protein-reaction associations (GPRs) in Recon1 use Entrez Gene IDs to annotate reactions in the network. To map the data from the AffyHG-U95 chips to Recon1, all genes included in Recon1 were mapped to corresponding AffyHG-U95 probesets using Bioconductor [58] and the most recent chip annotations [59]. A single expression value was then assigned to each gene in Recon1 based on the maximum normalized data value associated with any of the probesets mapped to a given gene. Next, a single expression value was assigned to each reaction in Recon1 by evaluating the Boolean rules in the GPRs with respect to the normalized expression data. The minimum data point was chosen for genes linked by an AND operator in a GPR, and the maximum data point was chosen for genes linked by an OR operator in a GPR.
Finally, a significant expression threshold was established for subsequent use in the GIMME algorithm. This was done by fitting the normalized gene expression data to a Gaussian distribution, estimating the mean and standard deviation of this distribution, and calculating p-values associated with each data point by subtracting the cumulative distribution function from one. The normalized data value corresponding to the p-value closest to but not exceeding 0.05 was chosen as the significance threshold; this resulted in a threshold of 991.3698 for the normalized expression data.
To integrate the renal objective function and kidney gene expression data with the preliminary model to derive a functional kidney model, the GIMME algorithm [13] was implemented. The GIMME algorithm takes a metabolic network model, a gene expression dataset, and specified required metabolic functions as input and solves a linear programming optimization to yield the network flux activity state that maximizes the specified functions while remaining as consistent as possible with the gene expression data. The complete renal objective function, the combination of all functions presented in Table 1, was set as the metabolic objective with a minimum requirement of 90% of the maximum possible flux set as a parameter for GIMME in determining the final kidney model. The reaction expression threshold parameter was set as described above. GIMME was run with these parameters and the normalized expression data and preliminary model as inputs. The resulting reaction activity predictions were used to constrain metabolic reactions yielding the full kidney model. Subsequently, the connected sub-graph of this full kidney model, which includes all functioning reactions possible for achieving the renal objectives, was isolated and is this portion of the model we focused on for validation and simulation. We refer to this sub-model as the reduced kidney model (available in SBML format as Dataset S1).
Gene activity predictions made when deriving the metabolic kidney model were compared to the set of expressed genes with normalized expression values above the significance threshold described above. Activity predictions were also validated against a comprehensive proteomics dataset from normal human kidney glomerulus tissue [30] for overlap with network-associated proteins detected with high confidence, that is, identified through detection of two or more peptides.
To evaluate the modeling approach used in this study, a five-fold cross validation was performed in which the data corresponding to the most highly expressed 20% of network-associated genes were excluded before deriving the kidney model. The ability of each approach to correctly predict the activity of these most highly expressed 20% of genes was measured from the overlap of predictions with the highly expressed gene set assuming a hypergeometric distribution, and the resulting probability was Bonferroni-adjusted.
All predicted metabolic protein drug off-targets were tested in the kidney model to assess the drug response phenotype caused by inhibitory effects in this system. Inhibition of metabolic proteins by the drug was modeled by constraining corresponding reactions catalyzed by drug targets to 0 flux units. Simulations of the consequences of these drug effects were performed using FBA as implemented in the COBRA Toolbox [55] in the MATLAB programming environment. Each drug target was evaluated with respect to its impact on each individual renal function to determine if target inhibition by the drug leads to a renal deficiency relative to the untreated normal kidney model. This was done by optimizing single exchange or demand fluxes at a time, representing reabsorptions and secretions respectively, out of the full set listed in Table 1. The cumulative effect of all predicted drug targets being simultaneously inhibited was also tested against each individual renal function. Renal secretion fluxes were maximized in simulation. Renal reabsorption fluxes were set as unbounded and then maximized while the remainder of allowable uptakes were constrained to equal fractions of the default maximum magnitude of 1000 flux units. The additional constraints were imposed for reabsorption simulations in order to allow the resulting network flux state to include concurrently active alternative optimal direct and indirect reabsorption pathways rather than having to identify alternative optimal pathways by performing multiple simulations.
Single gene deficiencies were also simulated in the kidney model to assess their effects on renal function and their potential as risk factors for treatment with CETP inhibitors. Each of the genes annotated to reactions in Recon1 was knocked-out of the kidney model and simulations were run using the gene-deficient kidney model both with and without drug treatment to assess effects on each individual renal reabsorption and secretion.
Drug response and metabolic disorder phenotypes were assessed by taking the ratio of maximum gene-deficient, untreated renal function flux to maximum normal, untreated renal function flux. A ratio of less than unity indicates a deleterious phenotype. Predicted metabolic disorder phenotypes were validated against previous clinical studies as represented in the Online Mendelian Inheritance in Man (OMIM) database (http://www.ncbi.nlm.nih.gov/omim/).
Cryptic genetic risk factors for drug treatment were also predicted for which the maximum gene-deficient, untreated renal objective flux equals the maximum normal, untreated renal objective flux but the ratio of maximum gene-deficient, drug-treated renal objective flux to maximum normal, drug-treated renal objective flux is less than unity.
Sensitivity of our prediction approach to variability in parameters was performed through repeated simulation in which we varied the parameter value across the full range of possible values. We investigated sensitivity with respect to each parameter independently. A normalized sensitivity coefficient was calculated as the result of each of these simulations. This coefficient was calculated by first taking the percent difference in the predicted outcome relative to a base case, our primary results, and then dividing it by the maximum possible percent difference.
Benchmark data was collected from the OMIM database (http://www.ncbi.nlm.nih.gov/omim/) by searching for all metabolic disorders related to renal reabsorptions or secretions that are associated with deficiencies in genes included in the reduced kidney model. The resulting list of disorders was manually curated using literature references to identify precisely which metabolic renal reabsorptions and secretions were impacted. These included not only those renal functions captured in Table 1, but also other renal exchanges. All resulting reabsorptions and secretions that can have corresponding non-zero fluxes under unperturbed conditions in the reduced kidney model were included in our benchmark data set (see Table S5). Every phenotype in the benchmark data was investigated through our model as described for simulating drug target effects and renal metabolic disorders, taking the ratio of perturbed to unperturbed flux capacities as a measure of phenotype, where a ratio of one signifies no disorder phenotype and a ratio of less than one signifies some degree of disorder. Next, the ratio threshold for classifying normal versus disorder phenotype was iteratively set to assess the sensitivity and specificity of our approach for predicting true and false positives across the full range from zero to one. Note that a threshold of one was used by default for the main results presented in this study. The true positive rate was plotted against the false positive rate (see Figure S3), the ROC curve, and the AROC was computed using the trapezoidal rule for approximating definite integrals. The statistical significance of our result was determined by comparison to 100 permutation trials in which all reaction flux ratios, perturbed to unperturbed, were randomly shuffled for each simulated gene deficiency and AROC-analyzed. The permutation trials exhibited true positive and false negative rates expected for random disorder phenotype classification (see Figure S3), and thus comprised an appropriate assessment of the predictive ability of our model simulation approach relative to chance. One-sample left-tailed student t-tests were performed using an alpha value of 0.05 to assess the statistical significance of the AROC and mean true positive rate achieved by our model simulation approach relative to the permutation results.
|
10.1371/journal.pbio.0050292 | A sex-ratio Meiotic Drive System in Drosophila simulans. I: An Autosomal Suppressor | Sex ratio distortion (sex-ratio for short) has been reported in numerous species such as Drosophila, where distortion can readily be detected in experimental crosses, but the molecular mechanisms remain elusive. Here we characterize an autosomal sex-ratio suppressor from D. simulans that we designate as not much yang (nmy, polytene chromosome position 87F3). Nmy suppresses an X-linked sex-ratio distorter, contains a pair of near-perfect inverted repeats of 345 bp, and evidently originated through retrotransposition from the distorter itself. The suppression is likely mediated by sequence homology between the suppressor and distorter. The strength of sex-ratio is greatly enhanced by lower temperature. This temperature sensitivity was used to assign the sex-ratio etiology to the maturation process of the Y-bearing sperm, a hypothesis corroborated by both light microscope observations and ultrastructural studies. It has long been suggested that an X-linked sex-ratio distorter can evolve by exploiting loopholes in the meiotic machinery for its own transmission advantage, which may be offset by other changes in the genome that control the selfish distorter. Data obtained in this study help to understand this evolutionary mechanism in molecular detail and provide insight regarding its evolutionary impact on genomic architecture and speciation.
| Genetic conflicts among genes happen when their modes of transmission differ. Genes in the heterogametic (XY) sex can be grouped as X-linked, Y-linked, autosomal, or cytoplasmic. Sex ratio in the progeny greatly affects the transmission advantage of each of the four types of genes, with the optimal sex ratio for each type being respectively 100%, 0%, 50%, and 100% of females. Sex ratio can often be biased from the normal 50% by genes that distort the sex ratio toward their own optimal transmission. Here we report genetic and molecular characterization of an autosomal gene that functions as a suppressor of an X-linked sex-ratio distorter. Male mutants give rise to female-biased progeny. The cause of the distortion was assigned to a failure of the Y-bearing sperm to mature. The DNA sequence of the suppressor gives clues to understanding the sex-ratio meiotic drive at the molecular level. More generally, the genetic conflict over sex ratio may be important in determining the evolutionary dynamics and the architecture of eukaryotic genomes.
| As a rule, most dioecious species produce equal numbers of male and female progeny because the rarer sex has a mating advantage. R. A. Fisher in 1930, as well as Carl Düsing in 1884 before him, framed arguments based on parental expenditure that mandates an equilibrium sex ratio of 50% female [1,2]. This so-called Fisher's principle is based on the premise that genes have maximum representation in future generations only if they are equally transmitted through both sexes [3]. However, sex ratios in nature sometimes show marked departures from 50%, and there are many ecological and genetic situations where the assumptions of Fisher's principle do not hold [4,5]. One prominent example is related to sex-linked genes of the heterogametic sex. Because of their unisexual transmission, mutations that distort their meiotic transmission—thus the sex ratio—in their own favor can gain selective advantage [6]. This type of sex ratio meiotic drive, also called sex-ratio, conflicts with selection for equal transmission of autosomal genes and thus represents a typical intragenomic conflict among genes when their transmission optima are not congruent.
A sex-ratio distorter can invade a population as long as its deleterious effects on viability and fertility are offset by the biased segregation [7]. Invasion by a sex-ratio distorter creates a genetic context promoting strong selection for sex-linked or autosomal suppressors that ameliorate the segregation distortion and/or its deleterious pleiotropic effects [8]. This type of recurrent intragenomic conflict generates a dynamic of genetic “attack” and “defense,” even in the absence of any external abiotic or biotic challenges [4,9,10]. As a result, genetic conflicts over sex ratio may be a key factor in driving the evolution of gametogenesis in the heterogametic sex, thereby leading to the nearly ubiquitous pattern of Haldane's rule observed in speciation [11–13]. Because the evolution of genetic suppression would make sex-ratio transient on an evolutionary time scale, the uncovering of suppressed sex-ratio mutations requires sophisticated crossing schemes, often between individuals from different populations or incipient species. Perhaps it is because Drosophila geneticists have often studied matings of this type that most of the known examples of sex-ratio are found in Drosophila [6], and new cases continue to be reported regularly [14–16].
At least three independent sex-ratio meiotic drive systems have been uncovered in the species D. simulans [15,17–20]. To facilitate discussion, we will assign each sex-ratio system a different name according to the location where the original stocks were collected or the research was carried out. In the most thoroughly analyzed case, which we will refer to as the Paris sex-ratio, at least two X-linked distorters have been mapped, and they are currently being characterized at the molecular level [19,21,22]. The Paris sex-ratio (SR) appears to have originated relatively recently, and evolutionary signatures of a recent selective sweep are evident [23]. Multiple suppressors on the autosomes and the Y chromosome have been reported [19,21,24]. Both the distorters and the suppressors are polymorphic across the worldwide distribution of D. simulans [25–28]. Cytogenetic studies show that most (92%–96%) of the Y chromosomes in SR males do not undergo normal disjunction during meiosis II. As a consequence, the Y chromosome is often fragmented and lost. Sperm without a Y chromosome do participate in fertilization, and account for the 15%–30% of sterile F1 progeny males observed [29]; however most of the Y-bearing spermatids fail to mature in spermiogenesis [30]. The second sex-ratio system, which we call the Durham sex-ratio, was uncovered by introgressing regions of the third chromosome from D. mauritiana into D. simulans. The gene responsible for the suppression was named too much yin (tmy) [15]. An intriguing observation is that among about 20 hybrid male sterility loci on the third chromosome between these two species, tmy also had the strongest sterilizing effect on hybrid males. This observation provides a direct link between meiotic drive and interspecific hybrid male sterility [15]. The third sex-ratio system, which we call the Winters sex-ratio, was first revealed in interspecific hybrids between D. simulans and D. sechellia, using a D. simulans stock collected in Winters, California [20]. The gene responsible for revealing the sex-ratio is autosomal recessive. It was thought to have been introgressed from D. sechellia into a largely D. simulans genetic background and to have replaced the dominant suppressing D. simulans allele; however, two alternative explanations could not be ruled out. In the first alternative, a cryptic meiotic drive system from D. sechellia may have been introgressed into the D. simulans background and became reactivated. In the second alternative, a sex-ratio system may have been created de novo by an interaction between D. simulans and D. sechellia genes, even though these genes are irrelevant to sex ratio control in their native genomes [20].
Two additional sex-ratio cases have been reported in D. simulans [17,18]. For a case discovered in Brazil, an autosomal recessive mutation was implicated [18]. Similarly, in flies collected in California in 1959 [17], a recessive mutation (sxr) was mapped on the third chromosome. Unfortunately, the stocks of these two cases are no longer available and hence their relationship to the above three systems cannot be ascertained.
Given the frequent observations of sex-ratio and its potential biological significance, it is surprising that no single gene has as yet been identified for any sex-ratio system. Two difficulties prevent a molecular genetic analysis. First, many sex-ratio cases have been found in the genus Drosophila but not yet in natural populations of the model species D. melanogaster. Second, sex-ratio genes are often associated with inversions, making even conventional genetic mapping difficult. In our study we took advantage of the genomic and genetic resources developed in recent years for Drosophila to characterize the Winters sex-ratio system at the molecular level. Evidence for the existence of the other two independent sex-ratio meiotic drive systems in D. simulans will also be elaborated. The study of sex-ratio systems may help us gain new insights into mechanisms of speciation, the maintenance of Mendelian segregation, and the evolutionary principles of genome organization.
We first mapped a recessive gene responsible for uncovering the Winters sex-ratio to the vicinity of the locus pe (polytene chromosome position 85A6) on the third chromosome through various crosses by using stocks from a previous study, where the Winters sex-ratio was revealed [20] (Text S1; Figures S1–S5). We reasoned that the recessive allele is likely a loss-of-function mutation of a dominant sex-ratio suppressor, rather than corresponding to the distorter itself, because a sex-ratio distorter has no transmission advantage unless it is sex-linked (Figure 1A). A sex-ratio line SSR12-2-7 was constructed by extracting the third chromosome from the stock SSR12 into a pure D. simulans background (Figure S2). This stock was used throughout this study for genetic and phenotypic analyses.
Further mapping of the sex-ratio suppressor was made possible through a 2-P mapping scheme [15], where the P[w+]-tagged D. mauritiana introgression stocks were constructed for other purposes (mapping hybrid male sterility between D. mauritiana and D. simulans) [31]. The P inserts here conveniently functioned as semi-dominant markers (Figure S6). Their positions and the introgression lines with two P inserts (2-P lines) in the pe region are shown in Figure 1B. The results of the 2-P mapping clearly indicate that the target gene must be localized to a ∼2,700-kb interval between P40 and P38 (Figure S6). With allele-specific oligonucleotide (ASO) markers (Table S2), we genotyped 129 recombinants generated from the 2P-10 line to narrow down the target to an interval of 88 kb between CG10841 and CG31337 (Figure 1C). There are no obvious candidate genes for the sex-ratio suppressor in this region.
To pinpoint the target gene more precisely, we screened an additional 3,100 recombinants within the P40 – P38 interval with 40 P38 and 50 P40 recombinants falling within the 88-kb interval between CG10841 and CG31337 (Figure 1C). Analysis of 79 of these recombinants narrowed the target to the 7-kb interval between CG14369 and K44 region (Figure 1C). In such a small region, it is unlikely that more than one gene is responsible for uncovering the Winters sex-ratio. We named this gene not much yang (nmy). We use the symbol Nmy to denote the dominant suppressing allele in D. simulans and D. mauritiana. The recessive loss-of-function allele that allows the sex-ratio phenotype to be expressed is denoted nmy.
We sequenced the parental chromosomes from D. mauritiana mau12 (Nmy) and D. simulans SSR12-2-7 (nmy) across the CG14369–K44 interval (Figure 1D). From this region, sequences homologous to CG14370 of D. melanogaster were recognized. Insertions of 2,791 bp and 1,427 bp were found in the coding region of CG14370 in mau12 (Nmy) and SSR12-2-7 (nmy), respectively. Notably, a pair of inverted repeats of 380 bp was found in the 2.8-kb Nmy insert, and three large deletions were found in the 1.4-kb nmy insert relative to the 2.8-kb Nmy insert. Because of these deletions, only one repeat remains in the 1.4-kb nmy insert (Figure 1D). The target region was further narrowed down to the 2.3-kb interval of nmy529–nmy2827 by genotyping the last 11 recombinants with seven additional ASO probes that mark the region from nmy135 to nmy7123 (Figure 1D, Table S3). Sequencing showed that the only difference between P40.B13 Nmy and P40.L12 nmy is in the two deletions D307 and D388 (Figure 1D). We conclude that the wild-type function of Nmy as a sex-ratio suppressor will be lost if the pair of inverted repeats in the Nmy insert are not intact. Hereafter, we use brackets to denote the species and length of the insert for the Nmy allele. For example, Nmy[mau2791] and nmy[sim1427] are used to represent the alleles as well as the inserts found in mau12 and SSR12-2-7, respectively.
We compared the CG14370/Nmy region from several species of the D. melanogaster subgroup. The Nmy inserts are found in D. simulans (four strains) and D. mauritiana (one strain) but not in the other species (one strain each) (Figure 2A). Importantly, we found no Nmy insert in the D. sechellia strain (3588) but two inserts (Nmy[sim2041] and nmy[sim1427]) in the D. simulans strain (sim2). These two strains were used in constructing the D. simulans × D. sechellia recombinant inbred lines where the Winters sex-ratio was first observed [20]. Evidently, the allele nmy[sim1427] does not come from D. sechellia. Rather, it is still segregating in the sim2 strain with a frequency of 6.1% (n = 294). There is an intact inverted-repeat structure within the allele Nmy[sim2041] that has full function as a sex-ratio suppressor (unpublished data). In addition, a null allele of nmy (i.e., CG14370 without any insert) was also found in a local Massachusetts population of D. simulans (unpublished data). Taken together, the evidence indicates that Nmy is a gain-of-function mutation that suppresses the Winters sex-ratio distorter.
The three Nmy inserts are compared in Figure 2B. The length of a single inverted repeat in Nmy[mau2791] is 35 bp longer than that in Nmy[sim2041]. The sequences between the inverted repeats in these two Nmy alleles are not homologous, except for a 93-bp element found in reverse orientation. There is a 664-bp fragment (D664 in Figure 2B) in Nmy[mau2791] that is absent in the two D. simulans alleles. The allele nmy[sim1427] was derived from Nmy[sim2041] by the loss of one inverted repeat and the loss of most of the sequence located between the inverted repeats, except for the 93-bp element in reverse orientation. This observation again suggests that the inverted-repeat structure is essential for the sex-ratio suppression function. Nmy does not appear to be a conventional gene (Figure 2B). Both ends of the transcripts across the Nmy[sim2041] or nmy[sim1427] region were determined by 5′ and 3′ rapid amplification of cDNA ends (RACE) and they are identical to those of CG14370 found in D. melanogaster. The full length of the mRNA from nmy[sim1427] was determined, and two alternately spliced introns (I, II) were identified (Figure 2B). We failed, however, to obtain the internal sequences of mRNA from Nmy[sim2041], because the reverse transcription was evidently stalled, most likely by a stem-loop secondary structure formed between the inverted repeats (see Text S3). The coding potential is limited for transcripts from both nmy[sim1427] and Nmy[sim2041] because of the premature termination of the CG14370 open reading frame (ORF) (Figure 2B). Although we cannot rule out the existence of other functional ORFs in the Nmy[sim2041] transcripts, it seems more plausible that the stem-loop structure is functional and essential as a suppressor. We predict that small interfering RNAs (siRNAs) generated from this stem-loop structure would contain the information that is necessary for the specificity of suppression.
Further evidence for a possible siRNA mechanism comes from the mapping and cloning of a sex-ratio distorter on the X chromosome (Dox). Dox is a novel gene that is responsible for the Winters sex-ratio distortion, whereas Nmy suppresses Dox and results in a normal sex ratio [32]. Sequence comparisons strongly suggest that Nmy originated from part of Dox through retrotransposition (Figure 2C). The retrotransposition event is evident by the following observations. First, there is a tandem duplication of the dinucleotide (TA) at the insertion site within CG14370, and 11 base pairs (TTGTTTAATTT) that are proximal to the 5′ end of the Dox cDNA are also duplicated in the 5′ end of the Nmy insert. These two duplications are the telltale signs of a retrotransposition event through a target-primed reverse transcription mechanism, which are possibly catalyzed by some non–long terminal repeat type retrotransposon [33–35]. Second, the three introns of Dox, with lengths of 57, 63, and 63 bp, respectively, are not found in Nmy. On the other hand, it is unclear whether the precursor mRNA in the retrotransposition event was a bona fide transcript of Dox, because the 3′ end of the Nmy inserts matches to genomic region 100–200 bp upstream from Dox. There are several explanations for this discrepancy. First, the 5′ end of an alternative Dox transcript may have been missed in our 5′ RACE experiments. Second, the evolutionary precursor of the current Dox might have a longer transcript. Third, the Nmy insert may derive from a longer and aberrant transcript from Dox.
Sequence comparisons support further inferences about the molecular evolution of the Nmy gene. Other than the left-most inverted repeat (IR' in Figure 2C), sequences of Nmy and the cDNA from Dox are largely colinear, implying that IR' originated as a secondary duplication after the retrotransposition. Many nucleotide substitutions are shared by Nmy[sim2041] and Nmy[mau2791], although each also has its unique substitutes. This observation suggests that these two alleles shared a common ancestor before the split between D. simulans and D. mauritiana. Even though there are some gross sequence changes along the two Nmy lineages, their homologous sites have only minor differences (3/1,467bp) relative to their divergence from the Dox sequence (Figure 2D). Furthermore, the inverted repeats from all alleles are virtually identical, again suggesting the functional importance of the stem-loop secondary structure (Figure 2D).
Distorted sex ratio can be caused by a number of mechanisms, including male-specific lethality, phenotypic sex reversal, or a deficiency of functional Y-bearing sperm. We inferred abnormalities in spermatogenesis from a cross of sex-ratio males with C(1)RM y w females, where the Y-bearing sperm are transmitted to female progeny. Progeny from this cross show a male biased sex ratio, implying a deficiency of functional Y-bearing sperm from sex-ratio males (Table 1). In another cross of sex-ratio males to females carrying the X-linked marker forked (f), no f female progeny were ever observed, ruling out the feminization of XY males to phenotypic females as the cause of the female-biased sex ratio (Table 1). Additionally, survival rates from egg to adult in the progeny of sex-ratio males are not different from controls, confirming earlier suggestion that male-specific lethality is not the cause of sex-ratio (Table S4) [20].
A lack of functional Y-bearing sperm could be caused by failure in development, motility, or ability to fertilize. To distinguish among these possibilities, we exploited the temperature sensitivity of the Winters sex-ratio. In particular, at 18 °C, the sex ratio in progeny of nmy males can reach as high as 93%, whereas at 25 °C, the sex ratio decreases to about 60% (Figure 3A). The age of the males has a much smaller but also significant effect on the sex ratio. In experiments carried out in such a way that each functional sperm was maximized in its chance of fertilization, we estimated that a nmy male produces about as many sperm as a simB male does at room temperature or 25 °C, but produces less than half as many functional sperm at 18 °C (Figure 3A). To assess whether any reproductive process taking place in females contributes to the sex ratio distortion, we carried out an experiment in which Nmy and nmy males and females reared at two temperatures (18 °C and 25 °C) were mated, and the F1 progeny were reared separately at these two temperatures (Figure 3B). The resulting sex ratio depended only on the temperature at which the parental males were reared, suggesting that spermatogenesis, not sperm competition and/or fertilization, is the stage where the etiology of the sex-ratio happens.
An ultrastructural study of spermatogenesis from Nmy and nmy males further narrowed the etiology of sex-ratio to abnormal spermiogenesis. Normal spermiogenesis follows an elaborate program of nuclear elongation, microtubule assembly and depolymerization, chromatin condensation, elimination of excess nuclear envelope and neoplasm, and individualization at the final stage [36–38]. During this process, a vast majority (>99%) of the nuclear content is eliminated as waste, and the microtubules perform essential cytoskeletal and transport functions [38]. Abnormalities in various stages of spermiogenesis in nmy males are evident (Figure 4). In early spermiogenesis, the individual mitochondria aggregate and fuse to form two giant interleaved structures (the onion stage, or nebenkern). Normally, the nuclei appear to be filled with a homogeneously fine granulofibrillar material, punctuated with occasional coarse granules and a prominent protein body [39]. However, the mutant spermatid nuclei start to accumulate nucleoplasmic vacuoles at the late onion stage. This is the earliest morphological abnormality that might be associated with malfunctions in spermiogenesis (Figure 4A and 4B).
In the subsequent elongation period, the nuclear envelope has already been demarcated into two well-defined areas, fenestrated and nonfenestrated. In the early stage of elongation, the fenestrated portion of the nuclear envelope is apposed to rows of microtubules running parallel to the elongating axis (Figure 4C). In the following periods, the microtubules migrate from the fenestrated to the nonfenestrated areas, at the same time that chromatin condensation begins on the nonfenestrated side. Along with the continuing nuclear condensation, much of the nuclear envelope and excess nucleoplasm are eliminated. The end product is a highly packed lanceolate nucleus, which is only about 1/200 of the original size. It is believed that the microtubules play essential roles in the process of chromatin condensation by performing transport and support functions [40] (Figure 4E). In nmy genotypes, some spermatid nuclei develop nucleoplasmic vacuoles that become very large in the early elongation stage. Because the rows of microtubules appear to be normal (Figure 4D), the failed transport across the nuclear envelope is not likely caused by microtubule malfunctions. In other words, it is more likely a failed process within the nuclei (e.g., chromatin packaging) that underlies the formation of the nucleoplasmic vacuoles. Later in the post-elongation period, these vacuoles may cause rupture of the nuclear envelope when forces generated from the microtubule arrays as well as from chromatin condensation exert a strong pressure on it (Figure 4F).
As an apparent consequence of the failed chromatin condensation, the sperm individualization process in nmy males is also perturbed. Normally, the individualization complex (IC) is initiated at the head region of the spermatid bundle and traverses along the entire length of the bundle. Concomitant with the passing of the IC, excess nucleoplasm, nuclear envelope, syncytial bridges, and degenerated spermatids are stripped off and collected in the waste bag. Each individual spermatid becomes invested in its own membrane [36] (Figure 4G and 4I). In nmy males, the nucleoplasmic vacuoles are prominently present in the IC, likely blocking its passage through the tails (Figure 4H). These abnormal tails remain in the syncytium, while the rest of the tails are individualized and separated (Figure 4J).
Counts of abnormal and normal nuclei or tails in transmission electron microscopy (TEM) micrographs from testes developed at 16 °C or 26 °C are shown in Table 2. During sectioning, the grids were prepared from sections at least 10 μm apart to avoid sampling the same nuclei for imaging, whereas only one grid from each testis was used for the tail images. The average number of heads per bundle observed was much lower than the possible 64, in part because perfectly perpendicular sections through the heads of a spermatid bundle are rare. It is however noteworthy that the average number of heads (mean ± standard error of the mean [SEM] = 19.1 ± 1.0 per bundle, n = 80) from 16 °C nmy male is significantly fewer than that from control (27.4 ± 2.0, n = 51) (t-test, p < 0.001), suggesting the bundles in nmy males are less tightly packed in the head region because of the many big heads inflated by the large nucleoplasmic vacuoles (Figure 4D). For nmy males raised at 16 °C, neither the percentages of abnormal heads (25.2% ± 1.4%, n = 80) nor that of tails (30.6% ± 3.1%, n = 39) is sufficient to account for the biased sex ratio of 97.0%, because the maximum sex ratio would be 71.9% and 66.5%, respectively, assuming that all abnormal sperm are Y-bearing and the maximum sex ratio being calculated as 0.5 × (abnormal + normal)/normal. This discrepancy suggests that many Y-carrying spermatids with normal appearance in TEM were nonfunctional. For nmy males raised at 26 °C, there were very few abnormal heads and tails observed, consistent with the slight bias in sex ratio.
Under the fluorescence light microscope, the earliest sign of abnormal nuclear transformation can be detected in the elongation stage but not in the onion stage (Figure 5A–5C). After slight fixation and 4',6-diamidino-2-phenylindole (DAPI) staining, it is very easy to spread spermatid heads within a bundle. The dimorphism in nuclear transformation becomes very clear for nmy males developed at 16 °C but is attenuated with increasing temperature. Reared at 26 °C, nmy males have hardly any abnormal nuclei (Figure 5B–5E and Table 3). The correlation between predicted sex ratio and the sex ratio observed is highly significant (r2 = 0.931, p = 0.0078). The high correlation implies that most if not all of the abnormally developed spermatids are Y-bearing.
Are the abnormalities observed at the onion stage the primary developmental lesion of nmy? In Drosophila spermatogenesis, many mutant phenotypes appear to be similar even though the underlying mutations (primary lesions) have dramatic differences in their normal functions [41]. In this case we took advantage of the temperature-sensitivity to detect the primary lesion under the assumption that it is also sensitive to temperature. In temperature-shift experiments, where the chance of mixing early and late sperm is small and the stages of spermatogenesis can be cytologically determined (see Materials and Methods), the critical developmental stages in spermatogenesis can be inferred directly from the sex ratio data (Figures 6 and 7). Up to the early elongation stage, the abnormal nucleoplasmic vacuoles formed at 18 °C can be eliminated by a temperature shift to 25 °C (Figure 7A), whereas a temperature shift is no longer effective after full elongation has been reached (Figure 7B–7D). For the shifts to lower temperature, 18 °C can still cause sex ratio distortion even if the nuclei have reached the middle elongation stage (Figure E), but a switch to 18 °C has little effect once individualization has commenced (Figure 7F–7L). These observations unambiguously agree with the electron microscopy data that the earliest detectable developmental lesions are indeed the abnormal nucleoplasmic vacuoles and that they have a deleterious effect in disrupting nuclear condensation. We do not know what upstream biochemical processes are involved, but chromatin modulation and traffic across the nuclear envelope are among the primary suspects.
We have presented genetic, molecular, and cytological data characterizing the suppression of the Winters sex-ratio in D. simulans. We have shown that the Winters sex-ratio is polymorphic within natural populations of D. simulans, and that it has no relation to D. sechellia as previously thought. The etiology of the aberrant sex-ratio is attributed to the failure of nuclear condensation in the Y-bearing sperm. We mapped and cloned an autosomal suppressor that acts to suppress a sex-ratio distorter on the X chromosome. The nucleotide sequence and structure of the suppressor strongly suggests a mechanism involving siRNAs. Remarkably, the evolutionary origin of the suppressor is from a transcript of the sex-ratio distorter gene itself. Our work supports the theory that intragenomic conflicts are important evolutionary processes that may ultimately underlie several seemingly unrelated major biological phenomena including Mendelian segregation, genome organization and speciation.
The cases of sex-ratio recorded in D. simulans can be recognized as at least three independent systems, as highlighted below.
Intragenomic conflicts are struggles within a genome over hereditary transmission [49]. Meiotic drive is one type of intragenomic conflict in that the driving allele or haplotype has more than 50% representation in next generation. For nuclear genes that freely recombine the evolutionary stable strategy is exact Mendelian segregation [8]. Thus for most of the genome, there is strong evolutionary pressure for selecting modifiers that increase the fidelity of Mendelian segregation. This explains why segregation in meiosis is usually Mendelian, and it also underlies Fisher's argument about the sex ratio based on parental expenditure [8]. However, this logic does not apply to cases where free recombination is inhibited. For example, in the two classic cases of autosomal meiotic drive—Segregation Distortion (SD) in D. melanogaster [50] and the t- complex in Mus musculus [51]— the distorter(s) and insensitive responder(s) are locked together within complex inversions. Similarly, the evolution of sex-linked meiotic drive is facilitated by the lack of recombination between the X and the Y chromosomes. It has been reasoned that sex-ratio meiotic drive is a more potent evolutionary force than autosomal drive, based on two arguments [11,12]. First, sex-ratio drive can evolve more readily. When the sex chromosomes do not undergo recombination along most of their length, which includes most cases of heteromorphic sex chromosomes, many sex-linked genes can potentially mutate to sex-ratio distorter. This is usually not true for an autosomal distorter, because the precondition for its invasion is satisfied only in special circumstances, such as in the centromeric region with its reduced recombination or within inversions. Second, sex-ratio has a much greater effect on the rest of genome precisely because it affects the sex ratio, and thus, the transmission rates of different genomic compartments. In addition to favoring modifiers that reduce any fitness cost, sex-ratio favors modifiers that render the sex ratio more equal, as mandated by Fisher's principle [52]. The evolutionary cycle of distorter and suppressor could go on indefinitely as long as new sex-ratio mutations unaffected by existing suppressors can occur.
When a sex-ratio mutation invades a population, its fate can be fixation, stable polymorphism, or extinction, depending on the configuration of fitness components (particularly the fitness of sex-ratio males) [53]. Polymorphisms of distorters and suppressors have been reported in several Drosophila species, including D. simulans [25], D. quinaria [54,55], D. obscura [56], D. paramelanica [57], and D. mediopunctata [58–60]. Fixation of a suppressor is likely the case for the cryptic sex-ratio suppressor Tmy in D. simulans [15]. However, no sex-ratio suppressors have been found in D. pseudoobscura, and the cause may be related to high fitness cost of sex-ratio males [61,62]. It is remarkable in itself that modifiers that increase the fitness of sex-ratio males seem not to have been selected for the past 0.7–1.3 My [63]. The answer may lie in the molecular mechanism of this sex-ratio distorter that is still unknown. Molecular characterization of the Winters sex-ratio system makes possible future studies on the molecular population genetics and ecological dynamics of suppressors in natural populations.
We have shown here that an RNAi mechanism is likely involved in the Winters sex-ratio suppression. Remarkably, the suppressor Nmy was generated through the use of the sequence information of the Dox distorter itself by means of retrotransposition. An interesting observation is that the abnormal transformation of the Y-bearing sperm in Dox;nmy male is exacerbated at lower temperature. If a lack of small RNAs encoded by the nmy allele is indeed the cause for derepressing the expression of Dox, then the molecules encoded by Dox (RNA or protein) may be more harmful to the Y-bearing sperm at lower temperature. Further in-depth molecular analysis of this hypothesis is required. This Dox-Nmy system is reminiscent of a possible case of cryptic sex-ratio system Ste/Su(Ste) in D. melanogaster [64,65]. The distorter Ste and the suppressor Su(Ste) share common sequences, and an RNAi mechanism has been convincingly shown to be responsible for the suppression [66,67].
The cytological defects in nmy are also reminiscent of abnormal nuclear condensation in the classical meiotic drive system SD in D. melanogaster [50]. The gene Sd targets the Responder (Rsp) locus and causes degeneration of Rsp-bearing sperm [68]. The Sd-bearing sperm is thus transmitted to more than 95% of the progeny from Sd/Rsp males. Sd is a truncated duplication of RanGAP and is still enzymatically active [69,70]. It may cause problems in establishing a normal RanGTP–RanGDP gradient across the nuclear envelope [71], which is a critical condition for many cytological functions including transport of small RNAs [72]. The Rsp locus consists of a cluster of satellite repeats whose sensitivity is proportional to copy number [73]. One might speculate that an interruption of heterochromatin condensation in and around the Rsp cluster directly causes the degeneration of Rsp-bearing sperm, because normal heterochromatin regulation requires small RNAs homologous to the heterochromatic region [74], whereas small RNAs originating and/or targeting the Rsp cluster may be misregulated in Sd/Rsp males. Indeed, small RNAs originated from the Rsp locus have been recorded [75].
Segregation distortion in males may happen at either meiotic or post-meiotic stages. For example, the Y chromosome can be fragmented or lost during meiosis II [29,76], or the Y-carrying spermatids may not mature [77,78]. Post-meiotic failure in spermiogenesis can result from aberrations in meiosis. For example, failure of X-Y pairing during prophase of meiosis I has been shown to cause spermiogenic failure in D. melanogaster [77,79]. For the Winters sex-ratio, it is unclear whether there is any pairing problem between the sex chromosomes, and whether chromosomal behavior during meiosis is normal. Through light and electron microscopy, we have demonstrated that the primary lesion in nmy mutants is a defect in nuclear condensation during spermiogenesis. In any case, segregation distortion in males usually results from a failure to produce a class of sperm, or a failure of a class of sperm to function, and there is a concomitant reduction in fertility.
Whatever the actual molecular mechanisms or developmental stages involved, it is logical to make an evolutionary link between sex-ratio meiotic drive and speciation. A meiotic drive distorter first exploits the meiotic and post-meiotic mechanisms for the biased transmission of itself, then the genome regains balanced transmission through the evolution of suppressors that correct the meiotic or post-meiotic aberrations. During each episode of distortion and suppression, male fertility might be compromised and recovered. Because of these dynamics, the meiotic/post-meiotic genes have an accelerated rate of evolution, thus promoting the reproductive isolation among isolated populations or incipient species that do not share common distorters and suppressors [13]. Under this meiotic drive scenario, it is easy to explain two ubiquitously observed genetic patterns for postzygotic reproductive isolation: a much faster accumulation of hybrid male sterility among Drosophila species, and a “large X” effect for the distribution of the hybrid male sterility genes [13]. Two Drosophila cases where one gene expresses both hybrid male sterility and sex-ratio lends direct support to a role for meiotic drive in speciation [14,15]. A testable prediction is that genes that are responsible for the initial postzygotic reproductive isolation between species most likely function in meiotic and post-meiotic processes such as chromosome segregation and chromatin condensation.
Five NSR (normal sex ratio) lines (IG88, IG113, IG118, IG132, and IG143) and six SSR (skewed sex ratio) lines (IG12, IG33, IG54, IG151, IG73, and Q15.3) were used from a previous D. simulans × D. schellia hybridization in which the parental D. simulans stock, sim2, had been collected in Winters, California [20]. The D. mauritiana × D. simulans introgression lines have been described before [13,31]. The following lines were used in this study: homozygous introgression line P18.16 and P38.7 and heterozygous line P40.6. The heterozygous 2-P lines were made by recombining two P[w+] inserts in cis from two introgression lines: 2P-10 (P40.8 × P38.4), 2P-20 (P37.1 × P35.2), 2P-19 (P37.3 × P38.5), 2P-17 (P35.2 × P46.17), and 2P-15 (P38.1 × P46.17). The expression of the w+ allele affecting eye color in P[w+] is sensitive to its position and copy number. Essentially, these P[w+] inserts provide semi-dominant markers along the third chromosome in D. simulans.
A D. simulans stock with the multiply marked third chromosome jv st e pe has been described before [80]. These mutations and their genetic positions are javelin 3–19.2, scarlet 3–49.5, ebony 3–63.0, and peach 3–104.9 [81]. They are allelic to D. melanogaster mutations javelin 3–19.2, 65A5-E1 [82], scarlet 3–44.0, 73A3 [82], ebony 3–70.7, 93C7-D1 [47], and pink 3–48.0, 85A6 [82], respectively. Note that because there is a large inversion 84F-94F on 3R in D. simulans as compared with D. melanogaster, the map positions of e pe in D. simulans are reversed on the map of D. melanogaster. Other D. simulans stocks simB (w; nt; III) and w; e have been described before [31]; and C(1)RM y w/lzS was kindly provided by J. Coyne.
All flies were reared on cornmeal-molasses-agar medium sprinkled with yeast grains at room temperature (22 ± 1 °C) if not otherwise indicated. The sex-ratio phenotype of a male was scored by mating the male with three tester virgin females, usually of the stock w; e, for 7 d before clearing all adults. The progeny were sexed and counted three times until the 19th day. The sex ratio (k) was calculated as the proportion of females.
The use of ASO markers was described previously [31]. Some key techniques and the reagents/kits used are as follows: long PCR (Takara LA Taq); PCR product cloning (Topo TA and Topo XL PCR cloning kits, Invitrogen); sequencing of large DNA fragment (EZ-Tn5 Insertion Kit, Epicentre); phage genomic library of D. simulans simB (Lambda ZAP II vector predigested with EcoR I, stratagene); RNA isolation (TRIZOL Reagent, Invitrogen); RT-PCR (reverse transcription polymerase chain reaction) (3′ and 5′- RACE kits and SuperScirpt II Reverse Transcriptase, Invitrogen).
Testes or gonads from crawling larvae or young adult males were dissected into saline (0.7% NaCl). Spermatids or mature sperm were released from gonads/testes with a fine tungsten needle. Live specimens were observed directly under phase contrast. For a reliable count of abnormal spermatids, the specimens were fixed for 20 s to 1 min on a microscope slide with 10 μl 2% glutaraldehyde in phosphate-buffered saline (PBS) (2.7 mM KCl, 137 mM NaCl, 8.0 mM KH2PO4); and rinsed with PBS for 5 min before staining with DAPI (100 ng/ml in PBST–0.1% Triton X-100 in PBS) for 5 min. The specimens were spread with vigorous tapping on the cover slip before being observed under epifluorescence. To visualize the IC [83], an additional step was added in the above protocol: fixed specimens were rinsed for 5 min in PBSTB (PBS with 0.1% Triton X-100 and 1% BSA) and stained 10 min with Alexa Fluor 488 Phalloidin (10 μ/ml; Invitrogen). The specimens were rinsed again for 3 min with PBSTB before DAPI staining. For longer storage, specimens were mounted in SlowFade Gold (Invitrogen). Images were taken on an Axioskop2 or Leica DMRB with digital camera, and processed with Adobe Photoshop.
In a drop of chilled 0.067 M pH 7.4 phosphate buffer, testes and accessory glands were dissected from young males (1–3 d old) with a fine tungsten needle and were transferred immediately to 2% glutaraldehyde in 0.067 M phosphate buffer on ice. The specimens were fixed for 2 h at 4 °C in 1% paraformaldehyde and 2% glutaraldehyde in 0.067 M phosphate buffer, followed by a post-fixation of 1 h in 2% OsO4 at 4 °C. The specimens were treated with 1% uranyl acetate at room temperature for 1 h before processing through an ethanol series to dehydrate. The specimens were trimmed after ethanol dehydration so that only one of each pair of testes was used in embedding. The Epon resin was made by mixing Taab Epon (42.4 g), DDSA (19.8 g), and NMA (18.0 g) for 30 min before adding 1.9 g DMP-30. Before final embedding, each testis was cut into 4–5 segments and aligned on the bottom of the mold in a straight line with the apical tip facing out. Sections were cut on a Reichert ultracut-S microtome, followed by staining with uranyl acetate and lead citrate. The grids were observed with a Tecnai G2 Spirit BioTWIN electron microscope.
Vials with SSR12-2-7 (nmy) eggs collected at 25 °C were cultured at 18 °C or 25 °C and shifted to the other temperature at time points of 0.5 or 2 d apart (0.5 d for the more critical stages) until adult flies emerged. The first few males after eclosion from each vial were singly mated to ten w; e 2-d-old virgins, which were adequate to exhaust the sperm of a single male (unpublished data). Each male was aspirated to new vials containing 10 virgin females every day for 3–5 d. Offspring from each vial were counted, but only those from the first vials were used in data analysis once their total reached over 50. On average, 148 and 108 offspring per male were used for calculating the sex ratio for the 18 °C to 25 °C shift and the 25 °C to 18 °C shift, respectively. In this way the first batch of sperm developing from the first few sperm bundles were exhausted and assayed. Critical intervals that were sensitive to temperature were detected from the change in sex ratio. Furthermore, the spermatogenetic stages of the first most mature bundles at these sensitive intervals were determined cytologically, either through phase contrast optics as described in D. melanogaster [41], or using Alexa Fluor 488 Phalloidin for viewing the individualization complex [83].
All sequences have been deposited in the GenBank (http://www.ncbi.nlm.nih.gov/Genbank/index.html) database and have been assigned the accession numbers EF565211–EF565217.
|
10.1371/journal.pntd.0003162 | In Situ Immune Response in Human Chromoblastomycosis – A Possible Role for Regulatory and Th17 T Cells | Chromoblastomycosis is a chronic fungal infection that affects skin and subcutaneous tissue. Lesions can be classified in tumorous, verrucous, cicatricial and plaque type. The cellular immune response in the severe form of the disease seems to correlate with a Th2 pattern of cytokines. The humoral immune response also seems to play a role. We intended to explore the populations of regulatory T cells and the Th17 pattern.
Twenty-three biopsies of verrucous form were obtained from patients with clinical, culture and histopathological diagnostic of chromoblastomycosis, without treatment. It was performed an immunohistochemistry method to detect Foxp3, CD25, TGF-β, IL-6, IL-17 and IL-23.
IL-17 was the only cytokine with high expression in CBM when compared to normal skin. The expression of Treg cells, TGF- β, IL-6 and IL-23 were similar to normal skin.
The constitution of a local immune response with high expression of IL-17 and low expression of other cytokines could be at least in part, an attempt to help the immune system against fungal infection. On the other hand, high levels of local immune response mediated by Th17 profile could overcome the role of Treg cells. The inefficient immunomodulation as a consequence of the unbalance by Treg/Th17 cells seems to corroborate with the less effective immune response against fungi.
| Chromoblastomycosis (CBM) is a chronic infection that affects skin and subcutaneous tissue, caused by some fungi which have a brownish color due to the presence of melanin pigments. The most frequent lesions are of verrucous type. Here we describe the participation of regulatory T cells and cells with the Th17 pattern of cytokines. High levels of Th17 cells participate in chronic inflammatory conditions, once at least in part they could improve the immune response and act in concomitance to the Th1 and Th2 patterns. Our results indicate a predominance of the Th17 pattern over Treg cells in verrucous lesions. We speculate that the local immune imbalance in CBM lesions characterized by exacerbated Th17 response, probably by suppressing the Treg response, is not effective for total fungal elimination. Even after long-period treatments, most patients has no absolute cure and often there is recurrence of the lesions. We believe that our study could contribute to the understanding of the immunopathogenesis of CBM and in such a way, presents some aspects to new possible therapies.
| Chromoblastomycosis (CBM) is a chronic granulomatous fungal infection that affects skin and subcutaneous tissue, especially the lower limbs. It is a cosmopolitan disease, but classically it is found in tropical and subtropical regions [1].
The causative agents of CBM are dimorphic fungi [2], [3] which have a brownish color due to the presence of melanin pigments in their cell wall and the spherical morphology that allow easy histological and mycological identification [4]–[8].
The infection occurs following traumatic inoculation of conidia or mycelial fragments from dematiaceous fungus [9]. The most frequently isolated species are Fonsecae pedrosoi, Phialophora verrucosa, Cladophialophora carrionii and eventually Rhinocladiella aquaspersa [10].
Clinically, the lesions can be classified in two instances: one that takes into account the appearance (tumorous, verrucous, cicatricial and plaque type) and the other, considering the gravity (mild, moderate or severe, according to the number and size of lesions) [11], [12], [13].
In general, CBM begins with the eruption of papules or nodules that develop slowly into warts. Regularly lesions appear in the lower limbs, knees and hands. There are reports of involvement of the face, chest, buttocks and other areas; dissemination via the lymphatic system is possible but infrequent [12].
The mechanisms of host defense in CBM are not fully elucidated. It is known that the immune response against the CBM is primarily cellular, where the process is ordered by phagocytic macrophages and humoral response also plays a role. Langerhans cells and Factor XIIIa+ dermal dendrocytes appear to be involved to a lesser extent in phagocytosis and antigen presentation against F. pedrosoi [6], [14]–[16].
Some studies have investigated the polarity of CBM and demonstrated, both experimentally and in situ, that the severe form of the disease, or warty lesions, correlates with a Th2 pattern of immune response by presenting the production of IL-10, high fungal burden, and also TNF-α. The average form is related to the Th1 profile, with high production of IFN-γ, low levels of IL-10, scarce number of fungi and relates to the better granulomatous immune response, resulting in less severe injuries (plaque type lesion) [17], [18].
Several studies have been conducted to better understand the Th1/Th2 paradigm of immune response or, in some diseases, the presence of both patterns of cytokines in the host. It is known that there is a subset of T lymphocytes that can modulate the immune response against pathogens, self-antigens and allergens which is also a constituent of immunological tolerance, called regulatory T cells (Treg). They are characterized by the expression of high levels of CD25 (α chain of the IL-2) whose function depends directly on the transcription factor Foxp3 [18]–[20].
A cytokine of high value for the studies of the immune response is TGF-β, since it is involved in the healing process in order to minimize tissue damage [21] and suppresses CD8+ cells [22], transform CD4+CD25− cells into CD4+CD25+ [23], induce naive T cells to differentiate into Foxp3+ [24] and still participates in differentiation of CD4 T cells in Th17 cells [25].
High concentrations of TGF-β added to the absence of pro-inflammatory cytokines direct the immune response to the development of regulatory T cells. Similarly, low concentrations of TGF-β associated with pro-inflammatory cytokines such as IL-1 β, IL-6, IL-21 and IL-23 promote expression of the IL-23 receptor (IL-23R), factor that allows the differentiation of CD4 T cells in Th17 [26], [27].
The Th17 lineage is a subpopulation of CD4+ T cells characterized by the secretion of IL-17 that seems to reinforce the protection of the host when the immune profiles of Th1 and Th2 cells are not totally effective against intracellular pathogens. Target of scientific spotlight, these cells appear to be protagonists in chronic inflammatory conditions, such as psoriasis [28], [29].
The study of both cell lines was performed by Pagliari et al. [30] in specimens from patients with paracoccidioidomycosis skin and mucous membranes lesions. The analysis revealed the involvement of both cellular profiles, identifying both the presence of immunoregulatory mechanism as the strengthening of effector T cells expressing IL-17.
Taking into account that, although the Th17 cells and regulatory T have similar ontogeny but distinct roles in the generation and control of infections [31], exploring their relationship could improve the understanding of the immunopathology of chromoblastomycosis and contribute to the most effective therapeutic strategies against the disease.
Twenty-three biopsies from skin lesions were kindly provided by Dermatopathology Laboratory, Division of Clinical Dermatology, Hospital das Clinicas, Faculty of Medicine, University of São Paulo, obtained from patients with clinical, culture and histological diagnostic of chromoblastomycosis by F. pedrosoi (86.96% males, mean age 61 years old, SD 15.28). The control group was constituted by ten specimens of normal skin, free of infectious or inflammatory activity at the time of surgery.
In addition to the skin control group without inflammatory activity (n = 10), it was also used a group of twenty skin lesions of paracoccidioidomycosis (PCM). This disease is caused by the dimorphic fungus Paracoccidioides brasiliensis and the host immune response against this fungus shares some similarities with chromoblastomycosis. Moreover, some markers of immune response proposed in this work have been described and discussed in PCM.
The use of the material that constituted the casuistic was approved by the ethics committee of Hospital das Clínicas da Faculdade de Medicina da Universidade de São Paulo, under the number 0317/11.
It was performed a streptavidin-biotin peroxidase method. The specimens were deparaffinized and hydrated in ethanol, the antigens were retrieved in TRIS/EDTA buffer pH 9.0 for 20 minutes at 95°C. The primary antibodies anti-Foxp3 (clone 236A/E7), anti-CD25 (clone 4C9), anti-TGF-β (clone TGFB17), IL-6 (clone 10C12), IL-17 (clone IL17A) and IL-23 (clone HLT2736) were diluted in 1% bovine albumin solution and incubated over-night at 4°C. Following, it was applied the second antibody and Streptavidin-peroxidase complex. 3,3-diaminobenzidine tetrahydroxychloride was used as chromogen and the slides were counterstained with hematoxylin. All reactions were performed with positive and negative controls. The second ones were constituted by the use of isotype controls and the omission of the primary antibody.
Cells were quantified by counting the number of immunolabeled cells in nine randomized high-power fields for each specimen with an ×10 ocular lens with a square grid area of 0.0625 mm2. The number of positive cells was statistically analyzed with the Mann-Whitney test with the level of significance set at 95%.
The group of CBM specimens consisted of lesions with a verrucous aspect. The histopathological analysis evidenced epidermal changes as hyperkeratosis, irregular acanthosis and microabscesses. The dermis was constituted by suppurative granulomas, inflammatory infiltrate consisting of giant cells, epithelioid cells, macrophages, lymphocytes, plasma cells, and eosinophils. The semi-quantitative analysis of parasitism ranged from moderate to intense. Eventually it was possible to identify ulceration of the skin, pseudocarcinomatous hyperplasia and fibrosis (Fig. 1).
The immunohistochemical method allowed observing cells expressing Foxp3 and CD25, both in the inflammatory infiltrate and around granulomas. The expression of TGF-β was present in mononuclear cells of the inflammatory infiltrate. There was a discrete expression of IL-6. The expression of IL-17 was visualized in mononuclear and polymorphonuclear cells, mainly in granulomatous areas. Cells expressing IL-23 were present in the inflammatory infiltrates (Fig. 2 and 3).
The group of PCM was characterized by the presence of all markers, expressed in mononuclear cells in the inflammatory infiltrate similar to the CBM group.
The control group of normal skin also presented the markers. However, the expression was low or even absent in some cases.
Cells expressing IL-23 were detected in 66% of the normal skin specimens. Those cells were observed only in the dermis.
The statistical analysis of cells expressing Foxp3 evidenced a decreased number of such cells in CBM group when compared to PCM group (p<0.001) and similar number when compared to normal skin.
The expression of CD25 was similar between CBM and PCM groups, however CBM group presented a statistically significant higher number of positive cells when compared to normal skin (p<0.001).
The number of cells expressing TGF-β was similar among the three groups (p = 0.05), the same to IL-6 and IL-23 (p>0.05).
The three groups presented cells expressing IL-17. However, CBM group had an increased number when compared to the two other (p<0.001).
The quantitative analysis can be visualized in figure 3 and table S1.
The study of CBM skin lesions is important because these injuries can reflect the immune status of the patient. The cutaneous tissue is often responsible for the initiation of immune cascade and, in the case of chromoblastomycosis, it is the structure to which the fungus has tropism.
Studies concerning immune response in CBM show that verrucous lesions present the mycotic granuloma formation with a pattern of Th2 response consistent with worse response against fungus. On the contrary, lesions of plaque type are characterized by a Th1 pattern of cytokines and therefore a better immune response, according to Minotto (unpublished data) and D'ávila [17]. However, there are no reports studying the immune dynamics of the lesions or inferences about what it takes to develop such responses.
We evidenced the predominance of cells expressing IL-17 and the presence of TGF- β and IL-23, although in low number. According to the literature published, this pattern of cytokines characterizes the Th17 immune response. The concomitant presence of a Th17 pattern could represent at first, an attempt of the in situ immune system to restrain the fungal infection.
With respect to the presence of regulatory T cells, although to a lesser extent when compared to cells expressing IL-17, we were also able to verify considerable number of Foxp3+ cells (40% of cases) and CD25+ cells (90%).
According to Melo and Carvalho [32], regulatory T cells have a key role in maintaining tolerance and regulation of the immune response. In the same way, in a work of Weaver [28] it is discussed that Th17 cells appear to be the translation of adaptability of the immune system favoring the protection of the host, so that the elucidation of the characteristics of both groups of cells, i.e. Treg and Th17 cells, could not only identify the mechanisms of invading organisms, but also to assist in the development of more effective therapies for numerous diseases.
Treg cells are characterized by a CD4+CD25+ phenotype [33], however the CD25 molecule is expressed by others populations of cells, such as others T cells and B lymphocytes or activated macrophages. Studies demonstrated that the gene Foxp3 has a nuclear expression in Treg cells and therefore a reliable marker of this cell line [34].
As already seen, the cytokine TGF-β plays a dichotomy role and mediates the targeting of Treg and Th17 responses. This selection between both profiles seems to depend on the levels of cells and the cytokines present in the microenvironment. High levels of TGF- β and low expression of pro-inflammatory cytokines favor the regulatory T-profile. On the contrary, low levels of TGF-β associated to significant presence of pro-inflammatory cytokines promote the synthesis of IL-23, development and maintenance of Th17 lymphocytes [35]–[41].
The cytokine IL-17 has received a considerable attention in many contexts since its discovery in 1993 [42]. The cytokine IL-23 promotes the secretion of IL-17 produced mainly by CD4+ T cells [43], [44].
According to many studies, high levels of Th17 cells are protagonists in chronic inflammatory conditions, once this cell population, at least in part, could improve the immune response and act in concomitance to the Th1 and Th2 patterns of response [28]. The performance of this cell line is characterized by secretion of IL-6, IL-17, IL-22 and TNF-α [45].
In the same context of investigation we consider patients affected by paracoccidioidomycosis (PCM), a systemic mycosis similar in some aspects to CBM. The disease follows inhalation of conidia of Paracoccidioides brasiliensis, a dimorphic fungus [46], and the primary focus are the lungs. It is described the cutaneous involvement at about 30% of cases. Unlike CBM, the lympho hematogenous spread of fungi is more frequent [47].
The cellular immune response in PCM is mainly mediated by macrophages and CD4+T cells [48], [49]. The Th1 profile of cytokines is considered the most important and some studies have also demonstrated the role of Treg cells and the profile of Th17 cytokines [30].
Considering previous investigations on the role of Treg cells in PCM, our results could suggest that this cell population not only have the capacity to interfere in the efficient immune response against fungi in chromoblastomycosis, but also benefit the host, by being able to reduce the tissue damage that follows a local immune response [30], [50].
In a recent study, the authors discussed the interaction of cells producing IL-17 and Treg cells and the homeostasis of the intestinal mucosal tissue [51]. The unregulated interaction of pro-inflammatory activity of IL-17 with pathogens seems to change the balance between regulatory and effector response predisposing the individual to the chronicity of the disease. It is noteworthy that even being subjected to long-period treatments, most patients affected by CBM has no absolute cure and often there is recurrence of the lesions. Thus, the unbalance between the populations of Treg/Th17 cells seems to restrain the effective immune response against the fungus.
Finally, it was interesting the similar number of cells expressing TGF-β, IL-6 and IL-23 when we compared the groups of lesions and normal skin. We expected that both CBM and PCM specimens presented more cells expressing those markers. We speculate that, at least in part, the presence of such cytokines in normal skin could be produced by the components of skin immune system [52].
Interestingly, in a previous work, Esterre et al. (1994) observed an overexpression of TGF-beta mainly at the periphery of the granulomas in areas of fibrosis [53].
In this work, we could suggest that the low number of cells with TGF-beta in lesions of CBM, with no difference from normal skin, could be explained by the randomized counting of positive cells throughout the dermis and not specifically in areas of fibrosis where such cells were also observed.
We suggest that our study could contribute to the understanding of the immunopathogenesis of chromoblastomycosis and in such a way, presents some aspects that could assist in the possibility of new therapies to modulate the immune system to the most effective immune profile of patients.
|
10.1371/journal.pgen.1005870 | Systems Level Analyses Reveal Multiple Regulatory Activities of CodY Controlling Metabolism, Motility and Virulence in Listeria monocytogenes | Bacteria sense and respond to many environmental cues, rewiring their regulatory network to facilitate adaptation to new conditions/niches. Global transcription factors that co-regulate multiple pathways simultaneously are essential to this regulatory rewiring. CodY is one such global regulator, controlling expression of both metabolic and virulence genes in Gram-positive bacteria. Branch chained amino acids (BCAAs) serve as a ligand for CodY and modulate its activity. Classically, CodY was considered to function primarily as a repressor under rich growth conditions. However, our previous studies of the bacterial pathogen Listeria monocytogenes revealed that CodY is active also when the bacteria are starved for BCAAs. Under these conditions, CodY loses the ability to repress genes (e.g., metabolic genes) and functions as a direct activator of the master virulence regulator gene, prfA. This observation raised the possibility that CodY possesses multiple functions that allow it to coordinate gene expression across a wide spectrum of metabolic growth conditions, and thus better adapt bacteria to the mammalian niche. To gain a deeper understanding of CodY’s regulatory repertoire and identify direct target genes, we performed a genome wide analysis of the CodY regulon and DNA binding under both rich and minimal growth conditions, using RNA-Seq and ChIP-Seq techniques. We demonstrate here that CodY is indeed active (i.e., binds DNA) under both conditions, serving as a repressor and activator of different genes. Further, we identified new genes and pathways that are directly regulated by CodY (e.g., sigB, arg, his, actA, glpF, gadG, gdhA, poxB, glnR and fla genes), integrating metabolism, stress responses, motility and virulence in L. monocytogenes. This study establishes CodY as a multifaceted factor regulating L. monocytogenes physiology in a highly versatile manner.
| Bacterial pathogens sense multiple host-related metabolic signals that alert them of host localization and result in induction of virulence traits. The Gram-positive foodborne pathogen Listeria monocytogenes activates the transcription of its virulence genes in response to low levels of branch-chained amino acids (BCAAs). This phenomenon is dependent on the global transcription regulator CodY, which binds BCAAs as a ligand and this binding affects its regulatory functions. CodY is classically thought to function under rich growth conditions, when bound to its ligands, however we recently reported that CodY directly activates L. monocytogenes virulence when BCAAs are limited. Identifying this novel CodY activity prompt us to further investigate CodY functions under different growth conditions in a genome wide level. For this purpose, we set on analyzing CodY’s regulon in L. monocytogenes in both rich and minimal growth conditions using genome-wide sequencing techniques. Remarkably, we identified for the first time a global regulatory role for CodY when BCAAs are limited, that are similar to those within the mammalian niche. Furthermore, our data establish CodY as a central regulator that integrates metabolism, motility, stress responses and virulence in L. monocytogenes.
| Listeria monocytogenes is a Gram-positive facultative intracellular pathogen transmitted by ingesting contaminated foods. L. monocytogenes causes a disease termed listeriosis associated with a mortality rate of up to 30%. Listeriosis typically manifests as a mild gastroenteritis in healthy people, however it can lead to meningitis in elderly and immunocompromised people and cause stillbirth in pregnant women [1]. The replicative niche of L. monocytogenes inside the host is within the cell cytosol [2]. The bacteria invade non-phagocytic host cells by expressing specialized proteins termed internalins that induce active internalization; in the case of phagocytic cells the bacteria are simply phagocytosed [3, 4]. Subsequently, L. monocytogenes rapidly escapes from the endosome/phagosome vacuole using primarily the listeriolysin O toxin (LLO), two phospholipases (PlcA and PlcB), and components of the competence system [5–8]. Having gained entry to the host cell cytosol L. monocytogenes replicates rapidly (at a growth rate similar to that exhibited in rich laboratory medium), and spreads from cell to cell using actin based motility, which is mediated by the virulence factor, ActA [9, 10]. Remarkably, all the above-mentioned virulence factors (and other factors) are positively regulated by PrfA, a Crp/Fnr like transcription regulator that is considered the master virulence activator of L. monocytogenes [11, 12].
The transition from saprophyte to deadly pathogen relies largely on L. monocytogenes sensing multiple host-specific signals that are transduced to trigger PrfA expression and activity [13]. For example, sensing of temperature, certain carbon sources (e.g., glucose-1-phosphate), the availability of amino acids (e.g., isoleucine), iron and glutathione, were all shown to trigger PrfA via different mechanisms, and thus activate downstream virulence genes [14–21]. These examples highlight that multiple mechanisms have evolved to sense various host-specific cues and metabolic signals that conjointly inform L. monocytogenes of its intracellular location and the need to switch to the virulent state. To better understand the metabolic environment within the host cell and the signals that activate L. monocytogenes virulence genes during intracellular growth, we previously performed a genome scale integrative study that combined transcriptome analysis and metabolic modeling in silico [18]. Various bacterial metabolic pathways were identified to be highly active during L. monocytogenes infection and contribute to bacterial intracellular growth in macrophage cells. Notably, the biosynthesis of branch-chained amino acids (BCAAs) (i.e., isoleucine, leucine and valine (ILV)) was highly induced in intracellularly grown bacteria, a pathway encoded by the ilv operon, suggesting that BCAAs may be limiting in macrophage cells. In light of this finding, we reasoned that sensing of metabolite availability within the host cell might alert the bacteria of their intracellular location and the need to activate the virulence state. Accordingly, we found that growing L. monocytogenes in minimal defined medium with limiting amounts of BCAAs indeed leads to robust activation of the virulence genes [18]. Under low concentrations of BCAAs, in particular of isoleucine, prfA and some of its downstream-regulated genes were highly expressed concomitantly with the BCAA biosynthesis pathway. These observations identified BCAAs as an important metabolic signal for L. monocytogenes within the mammalian niche.
Subsequently, we demonstrated that CodY, a global regulator and sensor of BCAAs, is responsible for the upregulation of prfA and the virulence genes under low BCAAs concentrations [17, 18]. CodY was found to bind directly within the coding sequence of the prfA gene (15 nucleotides down-stream the ATG start codon) and activate transcription, in turn leading to upregulation of virulence genes. These findings were surprising since CodY was thought to function primarily as a repressor and to bind DNA primarily under rich media conditions (i.e. under high BCAA concentrations) [17].
CodY is a Gram-positive specific global regulator that was discovered two decades ago in Bacillus subtilis as a general repressor of stationary phase genes, though now it is known to regulate many cellular processes [22–24]. CodY responds to cellular levels of BCAAs by directly binding these amino acids, an interaction that influences its structural conformation and activity [25–27]. In some Gram-positive bacteria CodY also binds GTP, however the effect of this interaction is not completely understood [28]. Initial studies indicated that CodY acts as a general repressor (when bound to GTP or BCAAs) of many metabolic genes including the BCAAs biosynthesis pathway. However, later studies in B. subtilis demonstrated that CodY also functions as an activator, in the presence of its ligand (i.e., under high concentrations of BCAAs) [29, 30]. In L. monocytogenes, CodY was shown to repress genes involved in amino acid metabolism, nitrogen assimilation and sugar uptake under conditions of high BCAA concentrations [31]. A role for CodY in regulation of virulence was demonstrated in Gram-positive pathogens other than L. monocytogenes, including Clostridium perfringens, Bacillus anthracis and Streptococcus pyogenes where CodY was shown to activate indirectly the expression of certain virulence genes [32–35].
As mentioned above, before the present study, CodY activity was primarily documented under rich growth conditions, with the ability of CodY to bind DNA demonstrated in the presence of its regulatory ligand [25, 36]. However, our previous observation that CodY possesses a regulatory activity and DNA binding capacity also under minimal growth conditions, exhibiting limited amounts of BCAAs (as shown for the prfA gene), raised the possibility that CodY may bind DNA and function also when BCAAs are at low concentrations, maybe even in its unliganded form [17]. Support for this premise came from the observation that a CodY protein mutated within its BCAA-binding site (i.e., harboring a R61A substitution within its GAF domain) loses the ability to repress metabolic genes under high BCAAs concentrations but retains the ability to activate prfA under low BCAAs concentrations [17].
To further delineate if indeed CodY possesses diverse activities under rich and minimal growth conditions and thus regulates different target genes, we took a system-level approach employing L. monocytogenes bacteria. The CodY regulon, in particular its direct target genes, were analyzed in bacteria grown in rich and minimal media (the latter containing low concentrations of BCAAs) using RNA-Seq and ChIP-Seq techniques. Notably, the data reveal that CodY retains multiple regulatory activities under both conditions, orchestrating the expression of metabolic, stress and virulence genes in a highly versatile manner.
Before the present study, CodY’s regulon and direct target genes were assessed under rich growth conditions (or that are rich in BCAAs), and thus a role for CodY was documented in the presence of its ligand isoleucine [33, 37–42]. To better decipher CodY’s activity in relation to the availability of its regulatory ligand/s, we analyzed the CodY regulon under both rich and minimal growth conditions (the latter containing low levels of BCAAs) using the RNA-Seq technique. Wild-type (WT) and ΔcodY L. monocytogenes bacteria were grown in rich brain heart infusion (BHI, containing excess amounts of BCAAs > 800 μM) and low BCAAs minimal defined media (LBMM; containing ~80 μM of BCAAs). Total bacterial RNA was extracted at mid-logarithmic growth and subjected to deep sequence analysis using Illumina HiSeq 2500. Of note, these experimental conditions were chosen in concordance with our previous findings demonstrating that during growth in BHI, CodY effectively represses metabolic genes (e.g., the ilv operon, the hallmark of CodY regulation), whereas during growth in LBMM this repression is relieved concomitantly with CodY activation of the prfA gene [17, 18]. As reported previously, WT and ΔcodY bacteria exhibit similar growth in LBMM medium, whereas ΔcodY bacteria grown in BHI displayed a slightly reduced growth as compared to WT bacteria (S1 Fig) [17]. RNA-Seq analysis was performed in triplicate and the reproducibility of the biological repeats was high (a mean R2–0.956) (S2 Fig).
A total of 368 genes (~14% of L. monocytogenes genome) were found to be affected by CodY (directly and indirectly) under both conditions. Among these, 237 genes were upregulated and 131 genes were down regulated in the ΔcodY mutant in comparison to WT bacteria. In BHI medium, 334 genes were differentially regulated by CodY (directly and indirectly), in agreement with previous reports [37–43]. Under this condition, 111 genes were down regulated and 223 genes were upregulated in the ΔcodY mutant in comparison to WT bacteria (Fig 1A and 1C). Notably in LBMM, 181 genes were differentially regulated by CodY, among them 55 genes were down regulated and 126 were upregulated in the ΔcodY mutant, demonstrating for the first time a global regulatory role for CodY under low BCAAs conditions (Fig 1B and 1C). Notably, among all of the upregulated genes (237 genes), 112 (45%) were upregulated under both rich and minimal conditions, whereas among the down regulated genes (131 genes), 36 (~27%) were down regulated under both conditions (Fig 1B). Overall these findings indicate that CodY may serve as both a repressor and as an activator of genes under high and low BCAAs levels. Moreover, they demonstrate that another mode of CodY regulation might exist that is independent of BCAAs. The latter may be mediated by additional regulatory factors such as GTP. Applying manually curated criteria clustering and hierarchical clustering on all of the differentially regulated genes yielded 6 distinct gene clusters representing all modes of CodY regulation (Fig 1B and 1C and S1 Table). Specifically, 111 genes were identified to be repressed (cluster I) and 76 genes to be activated (cluster II) by CodY (directly and indirectly) exclusively under rich growth conditions, 14 genes were identified to be repressed (cluster III) and 19 genes to be activated (cluster IV) by CodY exclusively under minimal growth conditions, while 112 genes were identified to be repressed (cluster V) and 36 genes to be activated (cluster VI) by CodY under both rich and minimal growth conditions (clusters are listed in S1 Table). Taken together, the data highlight CodY’s plasticity and ability to regulate genes in diverse spectrum of growth conditions.
Pathways and responses were identified manually and by functional enrichment analysis using the MIPS server (Fig 2 and S2 Table) [44]. Genes repressed by CodY under rich growth conditions are involved in amino acid metabolism and transport (e.g., BCAAs and histidine biosynthesis pathways), peptide and sugar transport systems (e.g., the OppC and OppF permeases and the glycerol uptake protein) and stress responses (e.g., bile salt hydrolase (bsh), clpC, and heat shock proteins) (cluster I) (Fig 2 and S1 Table). Notably, this cluster includes several virulence-associated genes such as internalin A and B and the glycerol transporter and kinase, which are expressed during L. monocytogenes infection of mammalian cells [15, 18]. Several metabolic genes/pathways were activated by CodY under this condition, such as arginine biosynthesis, assimilation of ammonia, certain PTS systems and enzymes of the tricarboxylic acid (TCA) pathway (cluster II) (Fig 2 and S2 Table).
Under minimal growth conditions, we identified prfA and its downstream virulence genes (e.g., hly, actA, plcA, plcB and mpl) to be activated by CodY (cluster IV), which is in accordance with our previous results [17, 18] (S1 Table). Notably, all the other genes within this cluster (except for one; LMRG_01981) were previously found to be induced during L. monocytogenes infection of mammalian cells [18], implicating a potential role in L. monocytogenes virulence. Some of these genes mediate ammonium transport, nitrogen regulation, cell wall synthesis and certain PTS systems, while others represent conserved hypothetical genes. Of note, it is not known yet whether these genes are under the direct regulation of PrfA. Overall, these findings strengthen the premise that CodY plays an important role in the activation of L. monocytogenes virulence under low BCAAs conditions, which resemble the mammalian cytosolic niche. Under these conditions, we found CodY to repress genes involved in purine metabolism, and iron transport, as well as some genes that encode hypothetical proteins (Cluster III) (S1 and S2 Tables). Genes repressed by CodY under both rich and minimal conditions are involved in nitrogen metabolism, arginine deiminase, osmotic and salt stress responses, distinct PTS systems and encode amino acid transport proteins, which most likely reflects nutrient availability within the media and the growth conditions tested (cluster V). Genes/pathways activated by CodY under both conditions include various metabolic genes and transport systems, D-alanine dipeptide synthesis, as well flagella biosynthesis and chemotaxis (cluster VI) (S1 and S2 Tables). Notably, several transcription regulators and regulatory proteins were identified within the CodY regulon (distributed among the different clusters), such as sigma-B, GlnR, GntR and certain CRP/FNR transcription regulators, as well sigma-54 regulatory proteins, indicating a hierarchy in CodY gene regulation. Overall, the RNA–Seq analysis highlights the breadth of CodY regulation in L. monocytogenes and its potential regulatory functions under varying metabolic environments.
To delineate which genes in the CodY regulon are directly regulated by CodY, a genome wide chromatin immunoprecipitation was performed in combination with DNA sequence analysis (ChIP-Seq) using Illumina HiSeq 2500. An L. monocytogenes codY-6his strain, in which the codY gene was replaced with a 6-histidine tagged codY, was grown in both BHI and LBMM media to mid-exponential phase. Bacteria were then cross-linked using formaldehyde and subjected to ChIP as described in the Materials and Methods. Of note, we have established previously that the CodY-6His protein functions similarly to the native CodY [17]. The ChIP-Seq analysis revealed 302 DNA regions bound by CodY under both conditions (270 in LBMM and 131 in BHI), with ~30% overlap (S3 Table). Among the 302 binding regions identified, 61 were mapped to transcriptional units (genes and operons) that were shown to be differentially regulated by CodY in the RNA-Seq analysis (corresponding to a total of 127 genes), and thus may represent bona fide targets of CodY under the tested conditions. Notably, 48 DNA regions were associated with transcriptional units regulated in LBMM (activated and repressed) and 33 DNA regions were associated with transcriptional units regulated in BHI (activated and repressed), whereas 20 DNA regions overlapped the two conditions (Fig 3A and S4 Table). Importantly, the data revealed that CodY directly regulates 33% of the genes within its regulon, and that it directly binds more regulatory regions under minimal growth conditions harboring low levels of BCAAs than under rich growth conditions, which is surprising given what is known about this regulator (see Discussion). Among the 302 CodY-binding regions identified, 71 (~23%) contained a putative CodY-box/s similar to those found in Lactococcus lactis and B. subtilis [37, 45], whereas among the 61 DNA regions that coincided with CodY regulated genes, 24 (~40%) exhibited a CodY-box, as determined by the MAST algorithm that predicts the presence of CodY binding motifs (Fig 3B and S3 Table). This observation suggests that CodY may employ additional binding sites or mechanisms to directly regulate genes. As expected, we found that the ilvD promoter region was bound specifically under high BCAA concentrations, whereas the virulence genes region was bound specifically under low BCAA concentrations; these data points serve effectively as positive controls for the ChIP-Seq dataset (Fig 3B and S4 Table).
By combining the data of the RNA-Seq and ChIP-Seq analyses, we were able to determine CodY direct-regulated genes, which further fall into the six described clusters. Under rich growth conditions, CodY directly represses the transcription of BCAAs, histidine, methionine, purine and riboflavin biosynthesis genes as well as the transcription of sigma B, clpC, glycerol uptake and phosphorylation, and a few general metabolic genes (cluster I). Under these same conditions, CodY directly activates the transcription of genes that encode for arginine biosynthesis enzymes, peptidoglycan deacetylation enzymes (on N-acytelglucosamine), and several PTS systems (cluster II) (S4 Table). Under minimal growth conditions, CodY directly represses purine biosynthesis, iron transport, a gene involved in pyrimidine biosynthesis and an amino acids permease (cluster III). Under the same conditions, CodY directly activates a cysteine transporter and a specific PTS system, in addition to the virulence regulator, prfA (cluster IV). Surprisingly, within this latter cluster we identified a novel CodY binding region upstream to the actA gene, which is responsible for L. monocytogenes intracellular actin based motility, and that is itself under the regulation of PrfA (S4 Table). Of note, this regulatory relationship represents an additional direct role for CodY in L. monocytogenes virulence. Under both rich and minimal growth conditions, CodY directly represses amino acids transport, PTS systems, genes involved in nitrogen (e.g., glutamate dehydrogenase, gdhA gene), pyruvate and lipids metabolism, and directly activates motility and chemotaxis genes, the GlnR regulator, other PTS systems and additional metabolic genes (S4 Table).
Next, as real-time quantitative PCR (RT-qPCR) analysis is still considered the gold standard in gene transcription analysis, we employed it together with ChIP RT-qPCR analysis to validate that CodY directly regulates representative genes from each cluster (clusters I-VI). To this end, we chose genes/operons that contain a putative CodY-box in their regulatory region (S3 Table and S3 Fig). From cluster I, comprising genes repressed by CodY in BHI, we chose the BCAAs and the histidine biosynthesis pathways (genes tested: ilvD, ilvC, hisG, hisA and hisI), as well the sigma B regulator (sigB) and the glycerol uptake transporter (glpF). From cluster II, comprising genes activated by CodY in BHI, we chose the arginine biosynthesis pathway and the glutamate decarboxylase gene (genes tested: argH, argF and gadG). From cluster III, comprising genes repressed by CodY in LBMM, iron uptake genes were chosen (e.g., feoA). From cluster IV, comprising genes activated by CodY in LBMM, prfA and actA genes were chosen. From cluster V, comprising genes repressed by CodY under both conditions, the gdhA gene was chosen [39, 48] and poxB gene encoding a pyruvate oxidase. From cluster VI, comprising genes activated by CodY under both conditions, flagella and motility genes were chosen (e.g., motB, flhA and fliP) as well the glnR gene, encoding the nitrogen metabolism regulator GlnR. In general, we found the RT-qPCR transcription profiles of the tested genes to be similar to those observed using RNA-Seq analysis. Genes predicted to be repressed by CodY were up regulated in the ΔcodY mutant in comparison to WT bacteria, whereas genes predicted to be activated by CodY were down regulated in the ΔcodY mutant (Fig 4A). The only exception is the observation that the ilv genes exhibit a higher transcriptional level in WT bacteria versus ΔcodY mutant during growth in LBMM. These results suggest that CodY further activates these genes when BCAAs concentrations drop significantly, a phenotype that was previously observed at [18].
ChIP RT-qPCR experiments were performed to verify direct binding of CodY to the regulatory regions of the chosen genes/operons (in the case of an operon the first gene of the operon was tested). Similar to the ChIP-Seq experiments, a CodY-6His variant was used to precipitate DNA fragments during L. monocytogenes growth in BHI and LBMM. Amplification of CodY binding regions upstream to the different genes/operons indeed verified that CodY binds all the tested regulatory regions (Fig 4B). In most cases, the binding of CodY correlated with the corresponding conditions in which CodY’s regulatory activity was observed.
Next, to affirm that CodY regulation responds primarily to the availability of BCAAs, we repeated the RT-PCR and the ChIP RT-PCR experiments in bacteria grown in minimal defined medium containing high levels of BCAAs (HBMM, containing 800 μM of BCAAs) and compared it to BHI and LBMM conditions (Fig 4A and 4B). Interestingly, most of the representative genes were regulated in HBMM as in BHI (R2 = 0.81), and were different from LBMM (R2 = 0.005), indicating that indeed BCAAs represent the primary ligand of CodY under these conditions. An exception, were the genes of the histidine and the arginine biosynthesis pathways, which were regulated by CodY in BHI (repressed and activated, respectively), but not in HBMM, suggesting that in conjunction with CodY, additional factors (e.g., GTP) mediate the regulation of these pathways under rich conditions.
To further characterize the binding of CodY to regulatory regions of select genes/operons, an electrophoresis mobility shift analysis (EMSA) was performed using DNA probes comprising the upstream intergenic sequences plus a ~100 bp of the gene 5’-coding sequence. In accord with the ChIP data, we observed binding of CodY to all tested probes with varying affinities. Although EMSA reactions are not at equilibrium and should be considered qualitatively, apparent KD values were derived as following: 150 ± 34 nM for hisZ, 151.5 ± 0.5 nM for rbsV-sigB, 114 ± 13 nM for glpF, 70 ± 11 nM for argG, 25 ± 13 nM for gadC-gadG, 90 ± 48 nM for feoA, 208 ± 19 nM for actA, 57 ± 3 nM for gdhA, 55 ± 2 nM for poxB, 74 ± 32 nM for glnR and 26 ± 9 nM for fliN (Fig 5 and S4 Fig). Of note, binding of CodY to the regulatory regions of genes specifically regulated in LBMM (Clusters III and IV) was measured in the absence of BCAAs. Taken together, these experiments corroborate CodY’s ability to directly bind and regulate different genes in a versatile manner and identified hisZ, rbsV-sigB, glpF, argG, gadC-gadG, feoA, actA, gdhA, poxB, glnR and fliN as novel L. monocytogenes CodY direct target genes, representing metabolic and virulence genes.
Finally, intrigued by the observation that CodY directly activates genes involved in flagella biosynthesis, we examined the ability of the ΔcodY mutant to swarm on soft agar plates. WT and ΔcodY bacteria were subjected to a swarming assay on plates containing BHI and LBMM media. In accordance with our findings, the ΔcodY mutant was found to be severely impaired in motility in comparison to WT bacteria on both media tested (Fig 6). On BHI plates, swarming regions with mean diameters of 1.07 ± 0.07 cm and 0.625 ± 0.06 cm were measured for WT and ΔcodY bacteria, respectively (Fig 6A), while on LBMM plates, mean diameters of 1.3 cm ± 0.09 cm and 0.76 ± 0.04 cm were measured for WT and ΔcodY bacteria, respectively (Fig 6B). Introducing an ectopic copy of the codY gene to the ΔcodY mutant (ΔcodY+pLIV2-codY) restored bacterial motility on both media to WT levels (Fig 6A and 6B). Since the flagella plays a critical role in L. monocytogenes attachment to mammalian cells [49, 50], we further examined the ability of the ΔcodY mutant to attach to Caco2 epithelial cells. An attachment assay was performed with WT bacteria, ΔcodY, ΔcodY+pLIV2-codY and a ΔflaA mutant as a control. Indeed we observed a reduced attachment of ΔcodY bacteria to Caco2 cells, a defect that was rescued in a ΔcodY complemented strain (ΔcodY+pLIV2-codY) (Fig 6C). Overall, these findings support a more central role for CodY in L. monocytogenes pathogenesis, beyond activation of virulence genes per se, a role that encompasses regulation of metabolic and motility genes, functions important for successful mammalian infection.
In this study, we applied a genome-wide approach to identify CodY’s regulon, target genes and regulatory functions in L. monocytogenes. Unlike previous studies of CodY, we examined both rich and minimal growth conditions, in attempt to explore further the possibility that CodY retains activity also when BCAAs, its primary ligands, are in limiting amounts. In contrast to the current dogma that CodY functions only in the presence of its ligand BCAAs, our results clearly demonstrate that CodY functions under both conditions, rich and minimal, and that under each condition it can serve as a repressor and as an activator of genes, establishing for the first time a global regulatory role for CodY under low levels of BCAAs (Fig 7). Furthermore, this study reveals a broader role for CodY in L. monocytogenes physiology, and particularly in regulation of virulence, as novel CodY direct target genes were discovered that are known to contribute to L. monocytogenes infection of mammalian cells (discussed below). Notably, this study not only builds on previous knowledge of how CodY serves to monitor and fine tune bacterial metabolism, but also expands our understanding of CodY’s spectrum of activities and impact on other central bacterial processes. For the first time, CodY is established as an integrator of bacterial motility, stress related and virulence functions and metabolic adaptations.
The two systems level analyses that we employed to delineate the role of CodY in L. monocytogenes growth under rich and minimal conditions were RNA-Seq and ChIP-Seq. The RNA–Seq analysis revealed that CodY affects the expression of hundreds of genes, establishing this transcription factor as a central regulator of L. monocytogenes. The ChIP-Seq analysis showed that CodY directly regulates 33% of its regulon, under both rich and minimal growth conditions. In total, more genes were regulated by CodY (directly and indirectly) under rich nutrient conditions (i.e., in BHI), with the majority repressed, essentially validating CodY’s global role as a repressor of metabolic genes. Nevertheless, a third of the CodY genes regulated under this condition were found to be activated by CodY, many of them related to the TCA cycle (discussed below), demonstrating CodY’s ability to serve as an activator as well.
As previously reported in other bacteria, we found that CodY represses amino acid biosynthesis (mainly BCAAs and histidine), purine, riboflavin and certain carbon and nitrogen metabolism genes under rich nutrient conditions [37–42, 48]. However, we found that CodY activates critical enzymes of the TCA cycle, including glutamate/glutamine derivatives and the arginine biosynthesis pathway, in contrast to what was shown in B. subtilis and L. lactis, where CodY was found to repress these genes [37, 41, 51]. This intriguing discrepancy suggests L. monocytogenes may have evolved distinct metabolic network/fluxes to fit its unique lifestyle. Specifically, our data predict that under rich nutrient conditions CodY directs metabolic flux from pyruvate to the TCA cycle through pyruvate carboxylase (PycA), which generates oxaloacetate, while blocking pyruvate flux to the BCAAs biosynthesis pathway through direct repression of the pyruvate oxidase gene, poxB and the ilv operon (both shown to be directly repressed by CodY) (Fig 8). This model is based on our findings that CodY upregulates most of the downstream TCA cycle genes (encoding four consecutive enzymes converting oxaloacetate to 2-oxoglutarate). Since the TCA cycle of L. monocytogenes is missing the enzyme that converts 2-oxoglutarate to succinate (2-oxoglutarate dehydrogenase) [52, 53], this step may be bypassed by glutamate synthase converting 2-oxoglutarate together with glutamine to two molecules of glutamate, which are then further converted to GABA by glutamate decarboxylase (GadG), and to succinate by the GABA shunt [54]. A support for this metabolic bypass is provided by the observation that the gene encoding glutamate decarboxylase (gadG) was also found to be directly activated by CodY, while the enzyme that reverts glutamate to 2-oxoglutarate, namely glutamate dehydrogenase (encoded by gdhA), was found to be directly repressed by CodY under these conditions (Fig 8). Moreover, further down this pathway, we found that CodY activates expression of fumarate reductase, which converts succinate to fumarate, in agreement with the premise that CodY positively regulates the TCA cycle and the glutamine/glutamate-GABA bypass (Fig 8). Interestingly, since L. monocytogenes is also missing the enzyme malate dehydrogenase, which converts malate to oxaloacetate [52, 53], conversion of fumarate to malate is a dead end reaction. In this regard, our observation that CodY activates expression of arginine biosynthesis genes during growth in BHI may be explained as a way to consume fumarate through a reversed arginine pathway, generating carbamoyl phosphate that can further feed to nitrogen or pyrimidine metabolism (Fig 8). In general, this alternative metabolic flux may generate energy and essential precursors to other metabolic pathways to support rapid growth of L. monocytogenes in rich medium. In contrast, during growth in LBMM CodY does not up regulate the expression of the TCA cycle or the arginine metabolism genes and most likely diverts the flux of pyruvate to the generation of BCAAs (Fig 8). This model may explain the observation that the bacteria grow more slowly in LBMM than in rich medium (S1 Fig).
Recently, the second messenger molecule c-di-AMP was shown to be involved in regulation of the TCA cycle and was reported to be essential specifically during growth in rich medium conditions, but not during growth in minimal medium [55]. Interestingly c-di-AMP was found to negatively regulate the pyruvate carboxylase, PycA, (via an allosteric binding) thus controlling the conversion of pyruvate to oxaloacetate [56]. Therefore, under conditions where c-di-AMP is absent (simulated by a diadenylate cyclase mutant, ΔdacA) and it is expected that PycA activity is enhanced, it follows that the TCA cycle is accelerated, resulting in generation and accumulation of intermediates and byproducts that may be toxic to the bacteria. We propose that while c-di-AMP may play a role in regulation of the TCA cycle under rich conditions (i.e., by preventing a high flux of pyruvate to the TCA cycle), in minimal medium this function may be dispensable, since part of the pyruvate flux is directed by CodY to the BCAAs biosynthesis pathway. A corollary of this hypothesis is that inactivation of CodY, which in turn reduces flux through the TCA cycle, may alleviate the toxic phenotype of the ΔdacA mutant under rich nutrients conditions. Interestingly, a recent study reported an indirect but intriguing link between DacA and CodY, whereby a mutation in the codY gene rescued the virulence defect of a ΔrelA mutant, which in turn rescued the growth defect of ΔdacA [55]. While more research should be done to explore these phenotypes, it is clear that c-di-AMP and CodY play important roles in shaping L. monocytogenes core metabolism and by doing so affect virulence. Notably, the former is a known listerial immunostimulatory ligand, which is recognized by the innate immune system during infection, and thus exerts additional phenotypes within the host [57–59].
Another novel metabolic relationship identified in this study is the direct activation of GlnR expression by CodY. GlnR is a conserved transcription regulator of nitrogen metabolism genes (such as those involved in glutamine, glutamate and ammonium metabolism) [60]. Although both CodY and GlnR were previously reported to independently repress nitrogen related genes (e.g., gdhA) [39, 48], a direct relationship was not previously documented. The activation of glnR by CodY may underlie CodY’s observed robust regulation of nitrogen metabolism genes.
In addition to metabolic regulation, this study identified a role for CodY in regulation of stress responses. Specifically, under rich nutrient conditions we found that CodY represses, directly and indirectly, the stress induced protease HslUV, the chaperon GroEL-GroES, the osmoprotectant transport system OpuCA and the stress responsive alternative sigma factor, SigB (σB). The latter is shown here for the first time to be a direct target of CodY, indicating a hierarchical regulation of stress related genes down-stream to CodY. Notably, σB is a critical sigma factor of L. monocytogenes, playing a major role in regulation of stress related and virulence genes during mammalian infection [61, 62]. Sigma B itself is directly involved in PrfA regulation, as it binds the prfA promoter and facilitates transcription during infection [61, 63, 64]. Our novel finding that CodY regulates sigB raises the possibility that during mammalian infection, conditions in which BCAAs are limited, CodY may promote prfA transcription by at least two mechanisms: directly via binding to the prfA gene and indirectly by relieving sigB repression. In this regard, we identified also several virulence related genes to be indirectly repressed by CodY, e.g., inlA and inlB, which mediate bacterial internalization into mammalian cells [3, 4]. These genes were shown previously to be positively regulated by σB [65] and thus may be repressed in BHI as a result of sigB repression by CodY. Similarly, the bile salt hydrolase (bsh) and the osmoprotectant system opuCA are indirectly repressed by CodY during growth in BHI and are both known to be positively regulated by σB [61, 65]. More generally, these findings reveal tight regulatory cross-talk between three central factors, σB, PrfA and CodY that together coordinate L. monocytogenes adaptation to the mammalian niche with regards to stress, virulence and metabolism, respectively.
One of the big surprises of this study was the extent of CodY regulation under conditions when BCAAs are limiting. We found that CodY regulates (activates and represses) genes involved in metabolism, motility and virulence. Cluster IV is of particular interest, comprising genes activated by CodY in LBMM, as this cluster includes most of L. monocytogenes major virulence factors in addition to some metabolic genes. CodY regulation of Cluster IV strengthens the premise that CodY attunes virulence functions and metabolic requirements to better adapt the bacterium to the intracellular niche. Notably, three novel CodY direct targets were identified within this cluster in addition to the prfA gene: the actA gene; a cysteine transporter gene; and a PTS system operon (S3 Fig). We confirmed CodY binding to the actA regulatory region by ChIP-RT-qPCR and EMSA assays. Nevertheless, it remains somewhat intriguing why CodY directly regulates both actA and prfA genes, as actA is already under the direct regulation of PrfA. One possible explanation is that by regulating actA CodY is serving as a direct regulatory link between metabolism and motility (in this case intracellular motility), in addition to general regulation of virulence. A role for CodY in L. monocytogenes motility is further highlighted by our findings that CodY directly activates flagellar and chemotaxis genes (i.e., extracellular motility) under both growth conditions. Taken together, these observations suggest that CodY plays a major role in coordinating sensing of nutrients with bacterial movement under diverse conditions and niches. CodY regulation of motility was also documented in B. cereus, where it was shown that CodY positively regulates motility genes and that a ΔcodY strain is less motile [40]. Similarly, we showed in the present study that an L. monocytogenes codY mutant is impaired in motility and attachment to mammalian cells.
Above all, the present study establishes that CodY regulation is more complex than classically considered. Previous studies depicted CodY as a transcriptional regulator that represses gene expression either by binding to promoter regions to interfere with RNA polymerase binding or by binding to internal sites leading to transcriptional roadblocks [45, 66, 67]. In light of our new data, it appears that CodY could function in all possible states and forms, for example under high and low BCAAs levels, as a repressor and as an activator (under both conditions), with the ability to bind multiple binding sites with different affinities around the genome; such complexity will make future study of CodY highly interesting yet challenging. It is most likely that in vivo the different binding sites are subject to genome wide binding competition, which is dependent on the concentrations of the different CodY forms (for example bound or unbound to a specific ligand). The data also suggests that additional factors are involved in mediating CodY binding to the different sites, as many strong sites that are bound under rich conditions (e.g., upstream to the ilv operon) do not appear to be bound in minimal conditions, although BCAAs are still present, while other binding sites are bound under both conditions. A model of cooperative binding was suggested before for CodY to explain such phenomena [45], where under a given condition CodY may cooperatively bind DNA in high affinity, but if conditions are changed even a bit (e.g., a slight drop in BCAA levels), binding is completely lost. Under this scenario, other binding sites with lower affinities might now be accessible for CodY binding and thus better compete. This model may explain why we were able to detect many direct binding sites that are specific to LBMM. In addition, it is most likely that CodY does not work alone and that other transcription regulators and factors influence its activity, as was observed in the case of the histidine and the arginine operons. Such factors can affect CodY’s DNA accessibility, binding affinity or conformation and thus modulate gene expression [28, 68, 69]. Notably, GTP is another known ligand of CodY that was shown in other bacteria to affect CodY activity [25, 28]. It is possible that GTP modulates CodY activity also in L. monocytogenes, and that CodY responds to varying concentrations of GTP in a similar manner it responds to BCAAs. Under this scenario, the relative concentrations of the two ligands under changing environments may determine CodY activities and binding affinities, a hypothesis that awaits further investigation. Of note, we did not observe differences in the transcription of the three relA paralogs of L. monocytogenes (relA, relP and relQ) in the RNA-seq data, genes that their products are known to affect intracellular levels of GTP.
As in other genome wide binding studies, we observed many CodY binding regions that do not appear to be associated with transcriptional regulation. This phenomenon is generally explained by one of the following scenarios: non-specific protein binding that influences DNA topology; an active mechanism that serves to titer the regulator itself; or redundancy with other co-regulators, as was recently shown for CodY and ScoC interactive regulation, where CodY deletion alone did not result in changes in gene expression of the BCAAs permase, braB [69–75]. This notwithstanding, the observation that many of the identified binding regions do not contain a CodY putative binding box as identified in other bacteria [72, 76, 77], suggests that CodY binding sites in L. monocytogenes are more diverged than in B. subitlis and L. lactis and/or that additional CodY recognition sites might exist. Unfortunately, the resolution of our ChIP-Seq data did not allow us to identify such new sites, though new techniques may lead in the future to identification of such sites. Overall, this study expands our understanding of CodY functions in L. monocytogenes, and we expect these insights to impact the study of CodY in other pathogenic and non-pathogenic bacteria.
L. monocytogenes 10403S [78] was used as the wild type strain. Brain heart infusion (BHI, Merck) was used as a rich medium, while high BCAAs minimal medium (HBMM) was used as defined medium as in [79] with 100 μg ml-1 of BCAAs [isoleucine, leucine and valine] and low BCAAs minimal medium (LBMM) was made with 10 μg ml-1 for each BCAA, which is 76 μM for leucine and isoleucine and 85 μM for valine. Bacteria were grown at 37°C with agitation in BHI, HBMM or LBMM. Bacterial strains used in this study are listed in S5 Table.
Bacteria (WT and ΔcodY strains) were grown to mid-exponential phase in BHI, HBMM or LBMM at 37°C (OD600 = 0.35). RNA was extracted using the RNAsnap method [80]. For RNA-Seq samples, DNase I (Qiagen) treatment was performed on Qiagen RNAeasy columns. For RT-PCR analysis, DNase I (Fermentas) treatment was followed by phenol-chloroform extraction. For RNA-Seq samples, the RNA integrity number (RIN) was evaluated using a TapeStation instrument (Agilent Technologies) and then rRNA was depleted using the RiboZero kit (Epicentere). RNA-Seq libraries were prepared using the TruSeq RNA sample Prep kit (Illumina) and sequenced (50 nt per read) by HiSeq 2500 instrument (Illumina) at the Technion Genome Center (Haifa, Israel). For RT-qPCR analysis, 1 μg of total RNA was reverse transcribed using QScript reverse transcription kit (Quatna). 16 ng of cDNA were used for RT-qPCR analysis with FastStart Universal Green Master Mix (Roche) using a StepOnePlus instrument (Applied Biosystems). The transcription level of each gene of interest was normalized to that of the rpoD mRNA. For the ChIP-RT-PCR analysis, each target gene was first normalized by the levels of rpoD and bglA DNA in each sample (normalization to multiple control genes is recommended and done using StepOne software) and then the ChIP sample was normalized to its no-ChIP sample using the StepOne V2.3 software. First, ΔCt for each sample is calculated as ΔCt = Average Ct—Normalization Factor (NF), while Normalization Factor is the mean of the selected endogenous controls (single or multiple genes), which is used to normalize the Ct value of each sample. Next, Fold enrichment is calculated as RQ = 2 (–ΔCt (ChIP sample)) / 2(–ΔCt (no-ChIP sample)) RT-qPCR primers are described in S5 Table.
Bacterial ChIP analysis was performed as described in [17, 81]. Bacteria were grown in 50 ml of BHI, HBMM or LBMM at 37°C with shaking at 250 rpm to O.D. of ~0.35. 1% formaldehyde was added to the cultures, which were then incubated at room temperature with shaking at 100 rpm for 20 min. 0.5 M Glycine was added to quench excess formaldehyde by shaking for 5 min at room temperature at 100 rpm. Afterwards, the samples were centrifuged at 4000 rpm (2600 g) for 10 min at 4°C, washed twice with cold TBS (50 mM Tris-HCl pH 7.5, 150 mM NaCl) and kept at -80°C. Cross-linked samples were resuspended in 0.2 ml of lysis buffer (10 mM Tris pH 8, 20% sucrose, 50 mM NaCl, 10 mM EDTA and 10 mg ml-1 of lysozyme) and incubated for 30 min at 37°C, and then 0.8 ml of IP-buffer (50 mM HEPES-KOH pH 7.5, 150 mM NaCl, 1 mM EDTA, 1% Triton X100, 0.1% sodium deoxycholate and 0.1% SDS) supplemented with 1 mM PMSF was added. The samples were lysed using sonication (6 rounds of 30 s) and were centrifuged 10 min at 14000 rpm (18000 g) at 4°C and the supernatants were transferred to new 1.5ml tubes. Sheered DNA was analyzed by electrophoresis to observe bands between 200–500 bps. 0.8 ml of sonicated sample was used for immuno-precipitation by adding 20 μl of magnetic A/G beads (Millipore, Cat. 16–663) and 5 μl of anti-his tag antibody (Abcam, Cat. 18184). Samples were then incubated with rotation overnight at 4°C to allow immuno-binding. Beads were collected using magnetic stands and the supernatant was transferred to a new tube to be used as control DNA. Beads were washed twice with 0.5 ml of IP-buffer, once with 0.5 ml of IP-buffer supplemented with 500 mM NaCl, once with 1 ml of washing buffer (10 mM Tris pH 8, 250 mM LiCl, 1 mM EDTA, 0.5% IP-40 and 0.5% sodium deoxycholate) and once with TE buffer (50 mM Tris pH 7.5 and 10 mM EDTA). Finally, the samples were resuspended in 0.1 ml of elution buffer (50 mM Tris pH 7.5, 10 mM EDTA and 1% SDS) and incubated 10 min at 65°C. Then the beads were removed by magnetic stands and the supernatants were transferred to a new tube. 80 μl of TE buffer and 2.5 μl of RNase A (8 mg ml-1) were added and incubated for 1.5 h at 42°C, followed by 2 h incubation at 42°C with 20 μl of Proteinase K (Fermentas). Then, the samples were incubated overnight at 65°C and the next day the DNA was isolated using Qiagen DNA concentration and cleanup kit. For ChIP-Seq analysis, 50 ng of ChIP and control DNA were used to prepare ChIP DNA libraries using the TruSeq ChIP sample Prep kit and sequenced (50 nt per read) on HiSeq 2500 instrument (Illumina) at the Technion Genome Center (Haifa, Israel).
On average ~10 million reads were obtained per cDNA library in fastq file, providing a 168-fold genomic coverage (data was deposited in GEO, accession number—GSE76159). The quality of the reads was evaluated using FastQC (Babraham Bioinformatics) and if needed, trimming of reads was performed using the fastX tool kit (http://hannonlab.cshl.edu/fastx_toolkit/). Mapping of reads followed by upper quartile normalization by gene expression and differential expression analysis by the negative binomial distribution as the statistical model was performed using Rockhopper V1.3 with default parameters [82]. For curated clustering of differentially expressed genes 3 criteria were used: minimal value of 10 normalized counts in at least one of the samples, minimal fold change between the WT and ΔcodY samples of at least 1.8 in either the BHI or LBMM media, and a significant q-value (<0.05) of the negative binomial analysis performed by Rockhopper V1.3. Hierarchical clustering of the differentially expressed genes was performed by the average linkage method using Cluster V3.0 (http://bonsai.hgc.jp/~mdehoon/software/cluster/software.htm), and a heatmap was generated using TreeView V3 (http://jtreeview.sourceforge.net).
fastq files were obtained and processed as mentioned above (data was deposited in GEO, accession number—GSE76821). Reads were mapped to L. monocytogenes genome using Bowtie2 running with default parameters [83]. Peak calling for ChIP-Seq analysis was performed using MACS V1.4.2 with default parameters [84] and with SeqMonk. CodY motif search based on L. lactis and B. subtilis CodY motifs was performed using MAST [47].
Differentially expressed genes from each cluster of the RNA-Seq analysis were used as input for the MIPS server [44, 85] for functional enrichment analysis.
L. monocytogenes CodY-6His was expressed in E. coli strain BL-21 from the pET28 expression plasmid. 10 ml of overnight bacterial culture were diluted in 0.5 L of LB medium supplemented with 30 μg ml-1 of kanamycin. Bacteria were grown till O.D.600 = 0.3, and then induced with 1 mM IPTG for 4 h. The bacteria were harvested by centrifugation (2600 g, 10 min), washed in 50 ml of cold buffer A (0.3 M NaCl, 50 mM NaH2PO4, pH 8) and resuspended in 15 ml of buffer A supplemented with 10 mM imidazole and 1 mM PMSF. Bacteria were lysed by an ultra high-pressure homogenizer (Stansted Fluid Power) at 12000 psi. Cell debris were removed by centrifugation at 16,000 g for 20 min and the lysate was incubated with 1ml of Ni-NTA beads (Sigma) for 1 h at 4°C with tilting. The Ni-NTA beads were then loaded on a column and washed with 10 ml buffer A supplemented with 10 mM imidazole. The protein was eluted by 250 mM imidazole in buffer A and dialyzed overnight against 100 ml of buffer A. Protein concentration was determined using a Nanodrop 1000 (Thermo) spectrophotometer. 0.5 μg of purified protein were separated on SDS-PAGE gel followed by Commassie staining to test for the purity of the protein.
For EMSA probes of CodY target genes the upstream intergenic region up to ~100 bps into the coding sequence of the target gene was amplified using PCR and labeled using Roche DIG Gel Shift kit. Purified CodY-6His was incubated with 4 ng of DIG-labeled target DNA in binding buffer (20 mM Tris-Cl pH 8, 50 mM KCl, 2 mM MgCl2, 0.5 mM EDTA, 1mM DTT, 0.05% NP-40, 5% glycerol, 25 μg ml-1 salmon sperm DNA) for 15 min at room temperature. The samples were then loaded onto a pre-run 8% native acrylamide gel (running buffer composed of: 35 mM HEPES, 43 mM imidazole buffer, pH 7.4) and separated for 1.5 h at 200 V. Detection of DIG labeled probes was performed using Roche DIG detection kit and visualized using Super RX-N FUJIFILM. For calculations of average apparent KD values, the ratio between the free DNA probe and total DNA probe at each CodY concentration (i.e., at each lane) was quantitated by densitometry analysis using ImageJ software for each EMSA gel [86]. Regression analysis was performed for each gel and the apparent KD value was calculated. Average apparent KD values depicted in the manuscript are based on 2–3 regression analyses of each probe. To demonstrate the reproducibility of the EMSA gels, averaged quantifications of 2–3 biological repeats of each probe were fit using exponential least-squares regression analysis (shown in S4 Fig). The apparent KD values were determined as the concentration at which 50% of the DNA probes were unbound, as deduced the average from the regression analyses. List of primers used for the amplification of target DNA sequences is found in S5 Table.
BHI and LBMM soft agar plates were prepared (0.3% agar) with 1 mM IPTG. 1 μl from the overnight bacterial cultures in BHI was spotted on the soft agar plates and grown for 48 h at 30°C. For the analysis of swarming capability, diameters of the bacterial growth zones were measured using a standard ruler. Pictures were taken using Olympus camera.
L. monocytogenes strains were grown overnight in 3 ml BHI cultures at 30°C without shaking. CaCo2 cells were cultured overnight in 6-well plates in CaCo2 medium (MEM with 20% FBS, 1 mM sodium pyruvate, 2 mM L-glutamine and 5ml MEM non-essential amino acids X100 solution) supplemented with penicillin and streptomycin in 37°C incubator with 5% CO2. The next day, the cells were washed twice with PBS and replenished with fresh CaCo2 medium without antibiotics. The bacteria were washed twice with PBS and approximately 1.6x107 L. monocytogenes bacteria were used to infect 2x106 CacCo2 cells. Thirty minutes post-infection, the cells were washed 6 times with PBS and lysed with 1 ml cold water. Serial dilutions were plated on BHI plates and colony-forming units (CFUs) were counted after overnight incubation at 37°C.
|
10.1371/journal.pcbi.1005824 | Tracking urban human activity from mobile phone calling patterns | Timings of human activities are marked by circadian clocks which in turn are entrained to different environmental signals. In an urban environment the presence of artificial lighting and various social cues tend to disrupt the natural entrainment with the sunlight. However, it is not completely understood to what extent this is the case. Here we exploit the large-scale data analysis techniques to study the mobile phone calling activity of people in large cities to infer the dynamics of urban daily rhythms. From the calling patterns of about 1,000,000 users spread over different cities but lying inside the same time-zone, we show that the onset and termination of the calling activity synchronizes with the east-west progression of the sun. We also find that the onset and termination of the calling activity of users follows a yearly dynamics, varying across seasons, and that its timings are entrained to solar midnight. Furthermore, we show that the average mid-sleep time of people living in urban areas depends on the age and gender of each cohort as a result of biological and social factors.
| For humans living in urban areas, the modern daily life is very different from that of people who lived in ancient times, from which todays’ societies evolved. Mainly due to the availability of artificial lighting, modern humans have been able to modify their natural daily cycles. In addition, social rules, like those related to work and schooling, tend to require specific schedules for the daily activities. However, it is not fully understood to what extent the seasonal changes in sunrise and sunset times and the length of daylight could influence the timings of these activities. In this study, we use a new approach to describe the dynamics of human resting periods in terms of mobile phone calling activity, showing that the onset and termination of the resting pattern of urban humans follow the east-west sun progression inside the same timezone. Also we find that the onset of the low calling activity period as well as its mid-time, are subjected to seasonal changes, following the same dynamics as solar midnight. Moreover, with resting time measured as the low activity periods of people in cities, we discover significant behavioural differences between different age and gender cohorts. These findings suggest that the length and timings of the human daily rhythms, still have a sensitive dependence on the seasonal changes of the sunlight.
| The daily activity of people varies across space and time from place to place, date to date, and hour to hour as a result of biological, societal, economic, and environmental factors, shaping the society where they live. Roughly speaking, each day humans do certain activities at specific times. There are many environmental factors (cues or ‘zeitgebers’) involved in the entrainment of this clock, but as pointed out by Roenneberg et al. [1], the most dominant is light and is associated with the light-darkness cycle determined by the daily rhythm of daylight. However, mainly in places not close to the equator, the timing and duration of daylight is subject to noticeable seasonal variation due to the yearly movement of the Earth around the Sun, and these changes have a direct influence on the kind and timing of different human activities. On the other hand, humans living in urban areas are also immersed in an environment full of cues that could influence the entrainment of the circadian clock. Artificial lighting, social practices and schedules (work and school hours, workdays vs weekends), particularly for those living in big urban areas, could have a noticeable influence on the entrainment process. Social conventions impose characteristic schedules on individuals, and, at the population level we can expect people in urban areas to have periods of high activity between morning and evening, and periods of low activity (resting) during the night. The length and timings of human activity periods, specifically in urban areas, has important consequences for human health [2–5], economy and power consumption [6], and public transportation efficiency [7].
The human sleep wake cycle (SWC), and its dynamics in particular, has been studied in recent years to understand the processes and cues that govern it [8]. Generally speaking, most research on human SWC has focused on experiments with small groups under controlled conditions [9, 10], or questionnaire studies [1, 11–14] (mainly using the Morning-Eveningness Questionnaire (MEQ) [15] and the Munich Chronotype Questionnaire (MCTQ) [16]). The use of these tools for studying SWC has proved to be very fruitful and effective, though having some limits on the domain of applicability [14]. In contrast, the ever-increasing availability of information communication technologies (ICT) combined with researchers’ ability to access large-scale ICT-generated datasets (‘Big data’) has made possible the study of human behaviour using a variety of reality (data) mining techniques. In particular, there are a number of examples where mobile phone datasets have been analyzed to study social networks [17–20], sociobiology [21, 22], mental health [23], mobility [24–27], as well as social behaviour of cities [28, 29]. Over the past decade or so, the existence and accessibility of these large population-level datasets, has allowed scientists to study intrinsic human behavioural and socio-evolutionary patterns in unprecedented and complementary ways, compared to other research approaches.
Recently, datasets of mobile phone usage have also been used to study circadian rhythms, by analyzing individual’s mobile phone usage from the data captured by sensors [26, 30–35], or people’s communication patterns from their call detail records (CDRs) [31–33]. For example, one study used the mobile phone screen on-off sensor data to examine the sleep wake cycle of nine individuals, finding that most of the individuals varied their sleep time patterns between weekdays and weekends, as well as showing seasonal changes in their mid-sleep time [30]. In another study using mobile phones calls and text messages of a small number of individuals, it was shown that individuals can be classified as having morning type or evening type activity levels [31]. In our previous related work [33], we quantified the resting periods of people from their mobile phone calling activity, showing that there is a counterbalancing effect between the afternoon and night time resting periods, due to an interplay between ambient temperature and sunlight. The use of CDRs as a tool for investigating the sleep/wake circadian rhythm, is in our view a promising new line of research as of that complements the other research approaches especially the large scale survey-based studies, pioneered by Roenneberg et al [11–13].
In this study, we apply reality mining techniques to users’ call records in a mobile phone communication network to study the dynamics of the users’ calling patterns by focusing on the periods of low activity, i.e. when almost no calls are made. Users of the mobile phone network typically have specific time periods during which their calling activity ceases, and we may assume that the SWC is bounded inside this period of inactivity. We observe that the daily calling activity time displays an interesting dynamics across the year through seasons and along different geographical zones. By studying these patterns we can gain insights into human activity patterns, and the SWC, in particular. Interestingly, the calling activity pattern changes with the day of the year and it is found to depend also on the geographical location (latitude and longitude of the mobile phone user). From the circadian clocks involved in the daily rhythms of human societies, only those entrained to solar-based events depend also on the geographical location and on the day of the year.
In this work, we use mobile phone calling activity at the population level to study how the onset and termination of the urban human activity in different cities is synchronized with the East-West progression of the Sun. Also, we analyzed the annual progression of the onset and termination of the calling activity, finding that they show a strong seasonal variation. We note that this behavior is similar to the annual dynamics of solar midnight, inferring that solar midnight is an important cue entraining the human circadian clock. Finally, we determine the mid-time of the period of low calling activity, which is bounded between the termination of calling activity each day and its onset on the next day. We interpret this mid-time to correspond to the mid-sleep time, and show that it is strongly dependent on the age and gender of the individuals in the population.
Using an anonymized dataset containing details of mobile phone communication of subscribers of a particular operator in a European country described in detail in the Methods section, we investigate the calling activity of the urban population living in cities as a function of time of the day for all the dates during the year. This we do by calculating for each city the probability distribution Pall(t, d) for finding an outgoing call at time t of a day d = (1,…,365) of the year. For all the studied cities, a region of almost null activity can be found around 4:00 am. Using this natural bound to split the calling activity from one day to another, we define a ‘day’ starting from 4:00am of a calendar day and running to 3:59am of the next calendar day.
In Fig 1 we show Pall(t, d) (green line) during days d = 214−215 (marking early August) for a city with over a 500,000 inhabitants. The distribution Pall(t, d) has two high calling activity periods with the first one corresponding to the morning calls, peaking around noon, and the second related to the evening calls, peaking around 8:00 pm. This bimodal pattern is present every day across the year and all the cities included in this study. The high calling activity periods are delimited by two periods of low activity, one centered around 4:00 pm related to the time after lunch, and the second one in the middle of the night, around 4:00 am within the sleeping period. The pattern present in Pall(t, d) is similar to that reported in other studies using different CDRs [29, 36], mainly at the times when the calling activity starts and ceases. In ref. [29], where the calling activity of some Spanish cities was studied, the histogram of the number of active users at each time has a similar bimodal shape, with similar times for their onset and termination, as well as the depth in the middle located around the same time period (i.e. between 3:00 pm and 4:00 pm).
To study the specific times when the calling activity rises and falls, we analyze the ‘morning’ and ‘night’ periods separately, defining the former between 5:00 am and 3:59 pm, and latter between 5:00 pm and 3:59 am on the following calendar day, in such a way that each period is 11 hours long. During each ‘morning’, we select only the first call made by each user inside that period and construct the associated probability distribution for the time of the first call PF(t, d), directly related to the rise of calling activity. Similarly, during the ‘night’ we define the corresponding probability distribution for the time of the last call PL(t, d) by taking into account only the last call made by each user within that period. In Fig 1, it can be seen that the three defined probability distributions Pall(t, d) (green), PL(t, d) (red), and PF(t, d) (blue) for consecutive days during winter, for a particular city with a population over a 500,000. The shape of the distributions Pall(t, d), PL(t, d), and PF(t, d) depicted in Fig 1 for a specific day appear to be preserved for all the days and cities we have studied.
The mean time of the first call tF and of the last call tL of people in a city can be influenced by environmental, social, and economic factors, and their possible daily value could be distributed completely at random. However, we find that during the year and at different latitudes, despite the different factors influencing the shape of the distribution Pall, the onset and termination of calling activity follows a consistent pattern, and this characteristic behaviour allows us to compare the calling activity pattern of cities lying at different latitudes. If the onset or termination of the urban calling activity is socially driven, with fixed times for specific activities (like office working hours from 9:00am to 6:00pm), one could expect that cities lying in the same time zone and at the same latitudes have similar calling activity timings (onset and termination). However, we find that the onset and termination of calling activity synchronizes with the East-West sun progression, in such a way that cities lying in western locations start (and terminate) their calling activity after cities at eastern locations, with a delay difference corresponding to the time difference between their local meridians. In Fig 2A and 2B we show tL and tF for 5 different cities lying inside a latitudinal band centered at 42°N±40′. The region including the 5 cities spans a longitudinal angle of 10.8°, and by taking one of the cities as a reference, other cities are located at −7.8°, −4.7°, −3.7°, and +3.0° from the reference city marked here with 0.0°. Then we compare the actual distributions PL and PF of the time of the last call and of the first call, respectively, for the 5 cities in the same latitudinal band, and find that PL and PF for western cities seem shifted to later times. However, when the distributions are shifted by an amount of time corresponding exactly with the time difference between the local meridian of the corresponding city and the reference city, the distributions visibly collapse onto each other, as can be seen in Fig 2C and 2D. In this case, the time shifts are +31.2, +18.8, +14.8, and -12 minutes for the cities located at -7.7°, -4.7°, -3.7°, and +3° from the reference city at 0°, respectively.
The distribution collapse shown in Fig 2 is obtained by introducing a time shift corresponding to the sun transit differences between cities. In order to quantify the exact delay between the distributions, we calculate the required time shift that should be introduced between the calling distributions to minimize the Kullback-Leibler divergence DKL between them (see the Methods section). This measure is indicative of the similarity between the distributions, and is minimized when they are identical. We extend this analysis to include data from 30 cities, each one lying in one of the four latitudinal bands centered at 37°N (10 cities), 39.5°N (5 cities), 41.5°N (7 cities), and 42.5°N (8 cities). For each band, we choose one city lying near the mid point of the band as the reference, and calculate for all the cities in the band the average time shift between them and the reference city. This is done for each day of the week, averaging over 52 weeks of the year 2007. The results are shown in Fig 3, and it can be seen that the time shift that minimizes the divergence between the distributions corresponds to the delay between their local sun transit times. This synchronization appears stronger for the termination of the calling activity (represented by the distributions PL). As this pattern is consistently present in all of the four analyzed latitudinal bands, we conclude that it is a general behaviour of the population living in the cities. This result is consistent with those reported by Roenneberg et al. [12], obtained from MCTQ studies of people in Germany, distributed over a region that is 9° wide longitudinally. In their work, they take into account the population of the city by defining three population size categories, i.e. less than 300,000 inhabitants, between 300,000 and 500,000 inhabitants, and more than 500,000 inhabitants, while we classify each city of more than 100,000 inhabitants according to its latitudinal coordinate. Grouping the cities into latitudinal bands, we found a consistent entrainment to the East-West progression of the Sun, regardless of the population size of each city.
This result implies that the termination (last call of the day) and onset (first call of the next day) of calling activities in cities at similar latitudes follow an external cue driven by solar events, and the time difference in these solar events between two different cities is reflected in the timings of their calling activity.
We have shown that the cities located at the same latitude but at different longitudes have periods of low calling activity with different onset and termination times (Figs 2 and 3). This shift coincides with the difference between their local sun transit times, i.e. when the sun crosses the meridian of the city. This observation raises the question as to what external daily event induces such synchronization. As the delays correspond to the time period between the local sun transit times of the cities, it seems plausible to think that the sun functions as a cue for this entrainment.
At the latitudes where the studied cities are located, the time difference between the sunset in the summer and in the winter is around 3 hours, if daylight saving is not taken into account, and the same holds for the time difference between sunrises. In contrast, the time difference between the mean time of the last calls between summer and winter is at most one hour [33]. However, there is a clear synchronization between the sun transit time and the timings of calling activity. This means that there should be an external clock functioning as a cue. On the other hand, from a biological perspective, the time when the secretion of melatonin reaches its maximum [37] lies close to midpoint between sunset and sunrise (i.e. solar midnight), once the night is as dark as possible. It has been proposed that the mid-sleep time coincides with the time corresponding to maximum melatonin secretion [38, 39], and if the solar midnight shifts through the year, the time for the maximum melatonin secretion should follow a similar pattern, as well as the entrained mid-sleep time.
In their study, Allenbradt et al. [40], using the MCTQ approach, have reported that mid-sleep time (on free-days) changes from one season to another. In some of the studied populations, they found that there is a small but significant difference in the average mid-sleep time between the days when Daylight Saving Time is applied and other days. This lends support to our assumption that if the mid-sleep time shifts in response to seasons, the timings of the calling activity should be influenced by its variation. In such a case, when the human mid-sleep time occurs at later hours, the timings of the calling activity for the following days should also occur at later hours. In other seasons, when the mid-sleep time occurs earlier, the activity timings should also be shifted towards earlier hours. If this is the case, then solar midnight should be functioning as the cue to which the calling activity timings are entrained. The activity pattern is a consequence of the interplay between seasonal and geographical factors, as well as social and societal activities like work and/or school, transportation, eating and leisure activities. However, the latter require specific timings during the day, not necessarily controlled by the sleep/wake cycle. We have shown elsewhere [33] that the total period of low calling activity (that is, the period between the termination and the onset of the calling activity) is strongly correlated with the duration of daylight, showing seasonal changes similar to the mid-sleep time.
In order to find any possible synchronization between the onset (and termination) of calling activity and solar midnight, we calculate the average of the mean times of the last call t ¯ L and that of the first call t ¯ F, for three sets of cities located at the latitudinal bands ϕ = 37°30′N (seven cities), 40°20′N (six cities), and 43°0′N (eight cities). We compare t ¯ L, and t ¯ F with the yearly evolution of the solar midnight in a reference city within a given latitudinal band (see Fig 4). A detailed description of how t ¯ L and t ¯ F are calculated can be found in the Methods section. It can be seen that only t ¯ L resembles to some extent the dynamics of the solar midnight, with their two minima and at least one of their maxima occurring around the same days of those of solar midnight, although the relative amplitudes are not in correspondence. In addition, the discontinuities introduced by the daylight saving is visible in all the graphs, suggesting that the timings of the calling activity are not solely influenced by the socially-driven time, but instead are synchronized with an external (astronomical) clock.
The period of low calling activity is bounded by the mean times of the last call during the night and of the first call in the morning. The duration of this period changes across seasons [33] and is strongly influenced by the length of the day (or conversely by the length of the night). The mid-time of this low calling activity period should correspond to the average time of human low activity, i.e. when the majority of the urban population is sleeping. In chronobiology studies, the mid-sleep time, corresponding to the time when human sleep is in the middle of its cycle, has been found to vary with the age and gender of the individuals [11, 41]. Despite the fact that each individual has a distinctive sleep-wake cycle, with a chronotype ranging from advanced sleep period (morningness) to delayed sleep period (eveningness) [42], at the population level a characteristic mid-sleep time can be consistently calculated, taken simply as the average of individual mid-sleep times.
From the mean times of the last call of the day, tL and of the first call tF of the next day, we define the period of low calling activity TLCA as the elapsed time between tL and tF, as a measure of the time when cities cease their activity. In Fig 5a, the width of the low activity period TLCA of the most populated city in the dataset is shown, for 4 different days of the week (Tuesdays, Fridays, Saturdays and Sundays), as a function of the subscribers’ age and gender. There is a noticeable change of about 3 hours, moving from the age cohort of 20 to that of 40 year olds. After that rather abrupt increase, especially for Fridays and Saturdays, TLCA slightly decreases, reaching a local minimum value for the age cohort of 50 year olds, and then it increases again to reach the highest value at the age of 78 years. For the analyzed weekday (Tuesday) as well as for Sunday, TLCA increases almost monotonically with the cohort age, showing a small plateau for age cohorts between 45 and 58.
We have also tracked the midpoint of the inactivity period, defined as the mid-time between tL and tF. Due to its similarity with the average time in the middle of the sleeping period [41], we interpret this minimum calling activity time as the mid-sleep time tmid, calculated simply as tmid = (tL + tF − 24)/2. Both quantities are found to depend on the age and gender of each cohort, as can be seen in Fig 5b. We find that, for certain age groups (from 18 to 32 years old, and from 43 to 80 years old) tmid occurs at a later time for women as compared to men, while in the age group of 33 to 42 years old, tmid for the men occur later. This finding differs somewhat from the reported mid-sleep times (on free days) in the chronotype questionnaire study based on the MCTQ [11, 13], where males show a later mid-sleep time for age cohorts younger than 38 years old. Also, there is a strong dependence on age, with younger age cohorts (20–30 year old) having later tmid, i.e. around 30 minutes after that of the oldest age cohort (70-80 years old). This observation is in accordance with the observed chronotypes [41], which are attributed to biological factors or internal clock being regulated by neuronal and hormonal mechanisms. We also found an unexpected rise of tmid for the age cohort of 45–65 year old individuals, which we suspect is entirely of social origin. Hence it seems that both biological and social factors play a role in changing tmid, i.e. shifting the period of low activity to later hours.
In addition, we find that tmid varies across days of the week. On Fridays and Saturdays tmid occurs at a later hours compared with the other days. Similarly, the age cohort with the latest mid-sleep time tmid is different for different days of the week. On Saturdays, individuals in the age group 30 to 45 years old have the latest tmid, while for the other days of the week it is the 20–25 years old cohort which shows the latest mid-sleep time. The results of TLCA and tmid for the most populated city are also and consistently found in the next 5 most populated cities, as shown in the Supplementary Material (S1 and S2 Figs, respectively).
In this study, we have found that the onset and termination of the period of low calling activity for people in cities at about the same latitude but at different longitudes are shifted according to their relative longitudinal separation. Cities westward from the easternmost analyzed city stop their activity later in line with the time delay of the sun transit time. This result suggests that a solar event acts as a cue for the circadian rhythm of the period of low calling activity with the SWC bounded inside. This result is consistent with those reported by Roenneberg et al. [12], although strictly speaking the two studies cannot be compared directly as the focus of our study is on variation by latitude and theirs was on variation by population size of cities.
In addition, we found that the seasonal variation of the termination of calling activity resembles the annual variation in solar midnight (or solar noon). However, when the annual behaviour of activity termination is compared with other characteristic solar events like the sunrise and sunset, it appears to have a different functional form with different number of maxima and minima with different dates. Although, it seems likely that solar midnight (or solar noon) acts as a cue in the synchronization of the termination of the calling activity, further research is needed to confirm this. At the individual level, knowledge of the mid-sleep time and sleep duration allows the determination an individual’s chronotype [16]. However, at the population level, we could determine from the calling distributions the characteristic variation in the sleep duration and mid-sleep time as a function of the group age. The observed overall trends are in line with the earlier findings [41] and reveal an increase in the sleep duration and decline in the mid-sleep time with age. Several other intricacies are also evidenced at closer inspection. Firstly, the aspect of ‘social jetlag’ [43], defined as the difference between the mid-sleep times on free days and that of work days, becomes apparent across all age groups. Interestingly, although social jet-lag is expected to give rise to extended sleep duration on free days as a compensatory effect, for young adults (20–25) we find that the sleeping periods are comparatively less on free days (Friday and Saturday nights and the following mornings). Therefore, sleep deprivation is likely to be at a maximum for this age range. Second, previous observations suggest a monotonic decrease in the mid-sleep time from around 20 years of age, which can be attributed to endocrine factors [41]. In contrast, we observe a reversal in trend of the mid-sleep time such that at the age of 45 years it starts rising till 55 years of age, after which it decrease again.
In this study, we have analyzed a very large dataset of anonymized call detail records (CDRs) from a mobile phone service provider offering services in a a country located in the Southern Europe subregion of the United Nations geoscheme [44]. Due to a Non Disclosure Agreement associated to the dataset we are bound to keep the identity of the country unknown, and thus we have partially masked the latitude and Longitude coordinates of the cities to screen their actual location, such that each city is associated with a latitudinal band, and the latitude at the center of the band is assigned as the latitude of the city. In the analyses, depending on the measure we were focusing on, we chose the width and center of the latitudinal bands and in all cases specifying the corresponding values. The latitude coordinate associated with each band is described by ϕ ± dϕ, with ϕ the latitude in degrees at the center of the band, and dϕ the half-width of the band in degrees. On the other hand, as the latitudinal region is given, the Longitude coordinate is also screened by providing instead its angular separation from, an arbitrary point located in the same latitudinal band. Thus, for a given city, its longitude coordinate θ denotes the number of degrees from a reference point located in the same latitudinal band. The anonymization of the subscribers’ identities was performed by the service provider prior the data been given to us. The dataset contains CDRs of around 10,000,000 subscribers during 2007, with more than 3 billion calls between 50,000,000 unique identifiers. Each record contains the date, time, duration, and anonymized caller and callee identifiers. The dataset also includes demographic information of the majority of the subscribers, and, for those cases, the age, gender, postal code, and location of the most accessed cell tower (MAC-tower) are known. Thus, there are three possible locations associated to each user, namely the associated city center, the location of the MAC-tower and the center of the postal code region, and we use them to determine whether the subscriber “lives in a city”—defined by cases where their three associated locations are sufficiently close to each others. Taking as a reference point the geographical location of the associated city center, a subscriber lives there if the following three conditions are satisfied:
In this study, we chose 36 of the cities with more than 100,000 inhabitants in 2007 (see S3 Fig in the Supplementary Material), in such a way that our final analysis takes into account the calling patterns of around 1,000,000 subscribers in total. Locations of the subscribers are associated with the locations of the cities they reside. Each city is associated with the following two geographical coordinates: the latitudinal coordinate is fixed as the midpoint of a latitudinal band including the city, and the longitudinal coordinate, defined as the angular distance between the city and a reference point located in the studied region.
Calling behavior varies seasonally, particularly the mean value and the width of the distributions of the first and last call vary across the year, being pushed towards the afternoon during winter and towards midnight during the summer. In spite of this seasonal variation, for a given day the calling distributions of different cities have similar shapes, and we exploit this similarity to calculate the delays between them to identify the temporal shifts of the distributions. The Kullback-Leibler divergence [45] is a measure of similarity between two distributions, commonly used in statistical analysis, for example when comparing one distribution obtained from data and another generated by a model. It reaches zero, its minimum possible value, when the distributions are identical, and it increases in value as the distributions become more and more dissimilar. In the case of the calling activity of different cities, the distributions are not identical but have a very similar shape. Applying Kullback-Leibler divergence to a pair of these distributions, it would reach a minimum value when these distributions overlap most, falling on top of each other and collapse to one. Thus, if we measure the amount of time one distribution should be shifted in order to minimize its divergence from the second distribution. The time shift would correspond the actual time delay between them.
In order to quantify the actual time shift between the distributions PL of last calls for cities lying along different Longitudes, we proceed as follows. First, for all the cities within the band, we calculate all the distributions PL(t, d) between January 2nd and December 31st. For each day d, we fix PL(t, d)0° of the city labeled ‘0°’ as the reference distribution, and for every other city c in the band, we compared the reference PL(t, d)0° with time-shifted versions PL(t + nΔ, d)c of the distribution PL(t, d)c, with −5 ≤ n ≤ 8 and Δ = 5 min, to find the time shift n*Δ that minimizes the divergence DKL between them. Here, DKL is the Kullback-Leibler divergence measure, defined as DKL(P, Q) = ∑i Pi log(Pi/Qi), with P, Q being the two discrete distributions. Once we find for each city the set {n*Δ} with all the time-shifts across the year, we calculate its average time-shift 〈n*Δ〉, and plot it for all the cities in the band in the right column of Fig 3. As the time for the mean time of the last call is different for different days of the week [33], the average is calculated separately for each day of the week. We apply the same procedure for the time of the first call distributions PF, and the results are shown in the left column of Fig 3.
In order to find if there is any relation between tL and tF and the solar midnight, we have chosen 7, 6 and 8 cities, lying in the latitudinal bands centered at ϕ = 37°30′ N, 40°20′ N, and 43°0′ N, respectively. For each city, we shift its corresponding distributions in accordance with its longitudinal difference to collapse all into one. Then we calculate the average mean time of the last call, t ¯ L ( d ) = 〈 t ¯ L ′ ( d , c ) 〉, where, t ¯ L ′ ( d , c ) denotes the mean time of the last call for the shifted distribution for a city c belonging to the analyzed band during the day d, and 〈⋅〉 denotes the average over all cities lying within the band. Similarly, we calculate the average mean time of the first call t ¯ F ( d ) for the given latitudinal band. The quantities t ¯ L ( d ) and t ¯ F ( d ) are compared with the time at which the solar midnight occurs in the reference city of each band. It should be noted that in the original graphs there are days of national holidays and local festivities that introduce drastic pattern changes, which we filter out to construct the final graphs.
|
10.1371/journal.pgen.1003178 | MTERF3 Regulates Mitochondrial Ribosome Biogenesis in Invertebrates and Mammals | Regulation of mitochondrial DNA (mtDNA) expression is critical for the control of oxidative phosphorylation in response to physiological demand, and this regulation is often impaired in disease and aging. We have previously shown that mitochondrial transcription termination factor 3 (MTERF3) is a key regulator that represses mtDNA transcription in the mouse, but its molecular mode of action has remained elusive. Based on the hypothesis that key regulatory mechanisms for mtDNA expression are conserved in metazoans, we analyzed Mterf3 knockout and knockdown flies. We demonstrate here that decreased expression of MTERF3 not only leads to activation of mtDNA transcription, but also impairs assembly of the large mitochondrial ribosomal subunit. This novel function of MTERF3 in mitochondrial ribosomal biogenesis is conserved in the mouse, thus we identify a novel and unexpected role for MTERF3 in coordinating the crosstalk between transcription and translation for the regulation of mammalian mtDNA gene expression.
| One of the main functions of the mitochondrial network is to provide the energy currency ATP to drive a large array of cellular metabolic processes. The formation of the mitochondrial respiratory chain, which allows this energy supply, is under the control of two separate genetic systems, the nuclear and the mitochondrial genomes, whose expressions have to be tightly coordinated to ensure efficient mitochondrial function. The regulation of mitochondrial genome expression is still poorly understood despite the profound importance of this process in human physiology, disease, and aging. Here, we make one step forward by unraveling a new role for the mitochondrial transcription termination factor 3 (MTERF3), which was initially characterized as a factor able to decrease mitochondrial transcription. Using gene invalidation approaches, we show in two distinct model organisms, the fruit fly and the mouse, that MTERF3 is not only involved in mitochondrial transcription but also favors the assembly of the mitochondrial ribosome and thereby reinforces the coordination between transcription and translation events, two key steps in mitochondrial genome expression.
| There is a growing interest in molecular mechanisms regulating oxidative phosphorylation capacity because of the increasing number of diseases associated with mitochondrial dysfunction [1] and as aging is associated with mitochondrial functional decline [2], [3]. Regulation of mitochondrial gene expression has an important role in fine-tuning oxidative phosphorylation capacity because critical subunits of the respiratory chain and the ATP synthase are encoded by mitochondrial DNA (mtDNA) [4], [5]. The regulation of mtDNA expression is completely dependent on nuclear genes but the mechanisms are not fully understood [4]. The expression of mtDNA could, in principle, be controlled at many different levels, e.g. by regulation of mtDNA copy number, transcription initiation, mRNA stability, translation or stability of respiratory chain subunits. Mitochondrial transcription factor A (TFAM) is essential both for transcription initiation [6], [7], [8], [9] and mtDNA copy number control [10]. TFAM packages mtDNA into a compact protein-DNA structure termed the nucleoid [11], [12]. There is a good correlation between TFAM levels and mtDNA levels in eukaryotic cells and mtDNA cannot be stably maintained if not coated by TFAM. However, there are a large number of mtDNA molecules in any given cell and changes in copy number are a slow process that is unlikely to have a main regulatory importance. In support of this notion, experimental manipulation of TFAM expression has been used to create mouse models with moderate decrease or increase of mtDNA copy number, with no or only minor effects on oxidative phosphorylation capacity [10], [13].
The basal mitochondrial transcription machinery consists of the nuclear-encoded mitochondrial RNA polymerase (POLRMT), TFAM and mitochondrial transcription factor B2 (TFB2M), which together are sufficient and necessary for transcription initiation in vitro [6], [14], [15], [16]. A large number of nucleus-encoded proteins have been reported to directly interact with and modulate the activity of the basal mitochondrial transcription machinery [17], [18], but this whole area is lacking a consensus for the role of these putative intramitochondrial transcription factors [14].
We have recently demonstrated that the mammalian mitochondrial leucine-rich pentatricopeptide repeat containing (LRPPRC) protein [19] and its fly homolog the bicoid stability factor (BSF) protein [20] are essential and have very similar roles in controlling mRNA stability, mRNA polyadenylation and coordination of translation in metazoan mitochondria [19], [20]. Regulation of mitochondrial translation not only involves mRNA maturation and stability, but also factors regulating translation and ribosomal biogenesis [21], [22]. An example of these factors is the adenine dimethyltransferase TFB1M, which modifies the 12S rRNA of the small ribosomal subunit and is necessary for the stability of the small ribosomal subunit and ribosomal biogenesis [23].
The role of the MTERF (mitochondrial transcription termination factor)-family of proteins [24] in regulation of mtDNA expression is of particular interest, because its members have been reported to influence mtDNA expression at different levels. MTERF1 has been suggested to play a role in mitochondrial transcription termination, by binding mtDNA downstream of the two mitochondrial rRNA genes to regulate the ratio between transcription of the upstream rRNA genes and the downstream mRNA genes [25], [26], [27], [28], [29]. In addition, MTERF1 has been reported to have a role in activating mtDNA transcription [30]. Mice lacking the Mterf2 gene are viable, but have been reported to develop myopathy and memory deficits [31]. The exact molecular mechanisms of MTERF2 function remain unclear, but it has been reported to bind the mitochondrial promoter region and to stimulate transcription initiation [31], whereas another report showed that MTERF2 associates with nucleoids without displaying sequence-specific DNA binding [32]. MTERF3 and MTERF4 are both essential for embryonic survival [5], [33]. Characterization of conditional knockout mice has shown that MTERF3 functions as a negative regulator of mtDNA transcription initiation by interacting with the control region to inhibit activation of the two mitochondrial promoters [5]. Loss of MTERF3 in the mouse heart leads to a massive activation of mtDNA transcription and a severe respiratory chain deficiency, possibly caused by imbalanced amounts of mtDNA transcripts [5]. MTERF4 forms a heterodimer with the cytosine methyltransferase NSUN4 and targets this enzyme to the large ribosomal subunit [33], [34], where it likely modifies 16S rRNA to regulate mitochondrial ribosomal biogenesis.
The mitochondrial genomes of flies and mammals have the same gene content although there are differences in gene order and expression patterns [35], [36]. This high level of conservation of metazoan mtDNA suggests that important regulators of mtDNA expression also may be conserved. We therefore decided to use a cross-species comparison approach to further study the in vivo role of MTERF3. We demonstrate here that knockout and knockdown of the Mterf3 gene expression in Drosophila melanogaster leads to activation of mtDNA transcription and impaired mitochondrial translation. We further show that imbalanced transcription is not the only cause of the altered mtDNA expression because also the 16S rRNA levels are reduced and the assembly of the large (39S) mitochondrial ribosomal subunit is impaired. These findings prompted us to reinvestigate the role for MTERF3 in the mouse, where we also found a reduction in levels of the 39S mitochondrial ribosomal subunit and impaired ribosomal assembly in the absence of MTERF3. These findings identify a novel role for MTERF3 in the biogenesis of metazoan mitochondrial ribosomes and point to a close crosstalk between transcription initiation and ribosomal biogenesis in control of mtDNA expression and regulation of oxidative phosphorylation capacity.
We performed an extensive phylogenetic analysis of Mterf3 and found a single gene ortholog in Drosophila melanogaster, which we denoted DmMterf3 (Figure S1A). We used algorithms to predict the subcellular localization for the DmMTERF3 protein and found a high probability for mitochondrial localization using either Mitoprot (0.986) or TargetP (0.875) softwares. Next, we performed live imaging of cells expressing GFP-tagged DmMTERF3 after counterstaining with MitoTracker Deep Red (Figure S1B) and found a co-localization rate of 94.9±1.4% in Schneider (S2R+) cells (n = 8 analyzed cells) and 98.3±0.4% in HeLa cells (n = 10), thus experimentally verifying the predicted mitochondrial localization of DmMTERF3.
In order to analyze the in vivo function of DmMterf3 we generated knockout flies by ends-out homologous recombination [37] to replace the complete coding sequence for DmMterf3 with an attP-site and a loxP-flanked marker gene denoted white (Figure 1A). Heterozygous knockout flies (genotype DmMterf3+;white) were crossed to cre-recombinase expressing flies to remove the white gene and thereafter the third chromosome balancer Tubby (TM6B) was introduced. This balancer chromosome causes the Tubby larval phenotype, which will segregate with the wild-type DmMterf3 allele in our crosses. The homozygous removal of DmMterf3 as well as the excision of white was confirmed by PCR analysis of genomic DNA with gene-specific primers (Figure 1B). DmMterf3 knockout (DmMterf3−/−) larvae have a profoundly reduced body size and die in the third instar larval stage, whereas heterozygous DmMterf3+/− larvae pupate and develop into flies in a similar manner as wild-type larvae (Figure 1C). Quantitative reverse transcription (qRT)-PCR from DmMterf3−/− larvae showed ∼90% reduction of the DmMterf3 transcript levels at 3 days after egg-laying (ael) and ∼95% reduction at 6 days ael (Figure 1D). The residual levels of DmMterf3 transcript found in DmMterf3−/− larvae at 3 days ael are most likely due to a persisting maternal contribution because we saw further reduction of DmMterf3 transcript levels in older DmMterf3−/− larvae (Figure 1D). Loss of DmMTERF3 resulted in increased mtDNA levels in knockout larvae at 3 and 6 days ael (Figure 1E) and an increase of ND1, ND2 and ND6 steady-state transcript levels, whereas the steady-state levels of the COXIII and 12S rRNA transcripts were unchanged and levels of the 16S rRNA profoundly reduced (Figure 1F).
To summarize, there are important phenotypic similarities between DmMterf3 knockout flies and Mterf3 knockout mice [5], because in both cases the gene is essential and its inactivation leads to increased steady-state levels of most mitochondrial transcripts, as well as reduction of 16S rRNA transcript levels. The early death of DmMterf3−/− larvae prevented a detailed molecular characterization of the phenotype and we therefore proceeded to use DmMterf3 RNAi flies for the subsequent studies.
We proceeded to use a UAS-GAL4 based strategy to knock down DmMterf3 expression in flies. We first tested the RNAi construct by using it in conjunction with the eye-specific eyeless-GAL4 driver and found a massive phenotype with reduced eye-size and disorganized head structure consistent with efficient silencing of DmMterf3 expression (Figure S2A, S2B).
Next, we proceeded with the ubiquitous knockdown (KD) of DmMterf3 expression using the daughterless-GAL4 driver (da-GAL4) to produce a KD line containing transgenes encoding both the da-GAL4 transactivator and the inducible UAS-RNAi construct w;;UAS-DmMterf3-RNAi/da-GAL4. We also generated two control lines, the first line w;;da-GAL4/+ only containing the da-GAL4 transgene and the second line w;;UAS-DmMterf3-RNAi/+ only containing the transgene encoding the inducible RNAi construct. The KD line and the two control lines were analyzed in parallel for all subsequent experiments.
Ubiquitous knockdown of DmMterf3 expression led to ∼80–90% reduction of DmMterf3 transcript levels in KD larvae at 3, 5 and 6 days ael (Figure 2A). The DmMterf3 KD larvae were visibly smaller from 4 days ael and onwards, as documented by a reduced body weight in comparison with controls (Figure 2B). Eventually, DmMterf3 KD larvae displayed delayed larval development and died at the pupal stage.
In order to rule out off-target RNAi effects, we generated a transgenic fly line expressing Drosophila pseudoobscura (Dp) MTERF3. DpMTERF3 has ∼80% similarity to DmMTERF3 at the amino acid level, whereas the nucleotide sequence of the corresponding genes differs substantially (Figure S3A, S3B). We therefore hypothesized that the RNAi construct directed against DmMterf3 expression would have no effect on DpMterf3 expression and that the DpMTERF3 protein therefore would be expressed to rescue the lethal DmMterf3 KD phenotype. This prediction was indeed confirmed and qRT-PCR analysis showed loss of DmMterf3 transcripts and the presence of the DpMterf3 transcript in rescued flies (Figure 2C). Importantly, expression of DpMTERF3 fully rescued the growth phenotype of DmMterf3 KD larvae (Figure 2D), which indicates the absence of off-target effects of the RNAi construct we are using.
We proceeded to investigate the biochemical consequences of reduced DmMterf3 expression by measuring mitochondrial respiratory chain capacity in permeabilized tissue extracts from larvae. DmMterf3 KD larvae at 3–6 days ael showed a major reduction in the oxygen consumption rates in the presence of substrates that are metabolized to deliver electrons to the respiratory chain at the level of complex I (CPI) or complex I and II (CPI-SUCC-G3P) (Figure 3A). In contrast, substrates metabolized to deliver electrons at the level of complex II or glycerol-3-phosphate dehydrogenase, thereby eliciting electron transport by-passing complex I, had no major effect on oxygen consumption (Figure 3A). We also measured the activities of individual respiratory chain complexes in larvae at 6 days ael and found severely decreased enzyme activities of all complexes containing mtDNA-encoded subunits in KD larvae, whereas the exclusively nucleus-encoded complex II was unaffected (Figure 3B).
Additionally, we assessed the levels of assembled respiratory chain enzyme complexes by Blue-Native polyacrylamide gel electrophoresis (BN-PAGE). Assembled complex I and IV were markedly reduced in DmMterf3 KD larvae at 6 days ael, as indicated by a reduction of complex I and IV in-gel activity (Figure 3C and Figure S4B). In addition, a smaller and partially assembled form of complex I was present in DmMterf3 KD larvae at 6 days ael, again indicating a problem with complex I (Figure 3C, asterisk). Western blot analysis showed low levels of the NDUFS3 subunit of complex I, whereas the levels of the ATP5A subunit of complex V (ATP synthase) were unaffected in KD larvae at 6 days ael (Figure 3D) and in knockout larvae at 3 days ael (Figure S4A). Taken together, these results show that complex I and IV are the most affected of the oxidative phosphorylation complexes in the absence of DmMTERF3.
The progressive respiratory chain dysfunction induced by loss of DmMTERF3 (Figure 3) led us to investigate mtDNA levels and mtDNA expression (Figure 4 and Figure S5A). Similar to what we observed in DmMterf3 knockout embryos (Figure 1E), we found an increase of mtDNA levels in DmMterf3 KD larvae at 6 days ael (Figure S5A), possibly caused by a compensatory activation of mitochondrial biogenesis as previously observed in respiratory chain deficient flies [20] and mice [23].
We proceeded to use qRT-PCR to analyze levels of mtDNA-encoded transcripts (Figure 4A, 4B) in KD larvae at 3, 5 and 6 days ael. We observed a progressive increase in levels of the ND1, ND2 and ND6 transcripts, whereas there were no changes in the levels of ND4L, COXIII and 12S rRNA transcripts (Figure 4B). We also used Northern blots to analyze transcript steady-state levels in larvae at 6 days ael and found an increase of the ND2, ND4 and Cytb transcripts, whereas the levels of COXI, COXII and 12S rRNA transcripts were unaltered (Figure 4C, 4E). We observed decreased levels of the 16S rRNA (Figure 4C, 4E and Figure S5B, S5C). In contrast, all tRNAs analyzed, regardless of the location of the corresponding genes in the genome, showed a progressive increase of their steady-state levels (Figure 4D, 4F and Figure S5D, S5E). Interestingly, the mtDNA transcript profiles in Mterf3 knockout mice [5] and DmMterf3 KD fly larvae show many similarities, including increased levels of many, but not all, mRNAs, increased levels of tRNAs and decreased levels of the 16S rRNA.
We have previously observed that the steady-state levels of tRNAs, but not those of mRNAs, tend to correlate well with increased de novo transcription in flies [20] and mice [33]. We performed in organello transcription assays in larvae at 3–5 days ael and found no clear difference at 3 days ael, whereas there was an increase of de novo transcription of mtDNA in KD larvae at 4 and 5 days ael (Figure 5A).
The combination of a respiratory chain deficiency (Figure 3) and transcription activation (Figure 4C–4F and Figure 5A) suggested that either the activation of de novo transcription leads to a respiratory chain deficiency, e.g. by causing the observed imbalance in steady-state levels of transcripts, as previously suggested for the Mterf3 knockout mouse [5], or, alternatively, that the transcriptional activation is a secondary response to respiratory chain deficiency. The observation of decreased 16S rRNA levels (Figure 4C, 4E and Figure S5B, S5C) were interesting in this respect because the Mterf3 knockout mouse also displays such a decrease [5]. We therefore continued with a more detailed characterization of mitochondrial translation in the DmMterf3 KD larvae (Figure 5B–5D).
First, we determined the assembly states of the mitochondrial ribosomes by sedimentation gradient centrifugation of mitochondrial extracts isolated from larvae at 3 and 5 days ael (Figure 5B, 5C and Figure S6A). Fractions were collected across the gradient and analyzed for absorption at 260 nm to determine the RNA content in each fraction. Equal loading was assessed by analyzing an aliquot of the total protein extract, to be loaded on the gradient, on a SDS-PAGE gel followed by Coomassie staining (Figure S6B). Fractions were thereafter analyzed by qRT-PCR to measure levels of 12S and 16S rRNA. Control samples confirmed that we indeed were able to separate the small (28S) subunit, the large (39S) subunit and the assembled (55S) ribosome (Figure 5B, 5C). Already in DmMterf3 KD larvae at 3 days ael, we observed a reduction in levels of the assembled 55S ribosome (Figure 5B). This decrease of assembled ribosomes was even more pronounced in DmMterf3 KD larvae at 5 days ael and was, at this time point, accompanied by a marked increase of the 28S ribosomal subunit and a marked decrease of the 39S ribosomal subunit (Figure 5C).
To further study the consequences of reduced levels of the assembled ribosome, we performed assays to determine the de novo translation activity in isolated mitochondria and found a clear decrease in DmMterf3 KD larvae at 3 days ael and onwards (Figure 5D). Our results suggest that the reduced mitochondrial translation is caused by a problem with ribosome assembly and that the concomitant transcriptional response with imbalanced steady-state transcript levels may be a contributing factor, thus suggesting a link between these processes.
The suggestion that DmMTERF3 might play a direct role in mitochondrial ribosome biogenesis prompted us to re-investigate the Mterf3 knockout mice. We previously created Mterf3 heart knockout mice by crossing Mterf3loxP mice to transgenic mice expressing cre-recombinase under the control of the muscle creatine kinase promoter (Ckmm-cre) [5]. Deletion of MTERF3 in the heart leads to a severe respiratory chain deficiency, progressive increase in steady-state levels of most mitochondrial transcripts and profound increase of de novo mtDNA transcription [5]. In the knockout hearts, MTERF3 protein levels are severely reduced already at 4 weeks of age (Figure S7A), concomitant with a dramatic increase of de novo transcription (Figure S7B and [5]).
We proceeded to investigate the assembly of the mitochondrial ribosomal subunits in Mterf3 knockout mouse heart mitochondria, by gradient sedimentation and Western blot analysis. The 28S and 39S ribosomal subunits, as well as the fully assembled 55S ribosome were clearly resolved in control samples, as determined by the migration of the ribosomal subunit markers MRPS15 and MRPL13 (Figure 6A, 6B). In contrast, in Mterf3 heart knockout samples the amount of MRPL13 co-migrating with MRPS15 was severely reduced already at 4 weeks of age, suggesting a reduction of fully assembled ribosomes (Figure 6A). Concomitant with the reduction of 55S ribosomes, we observed increased levels of the free 28S ribosomal subunit (Figure 6A). The MRPL13 protein steady-state levels progressively decreased (Figure S7C) and by 13 weeks of age, no fully assembled ribosomes were detectable (Figure 6B). These results clearly suggest that the assembly of the mitochondrial ribosome is impaired in the absence of MTERF3. We performed a set of confirmatory experiments, where we used qRT-PCR to assess presence of 12S and 16S rRNA in the different fractions. As predicted, the relative levels of 16S rRNA were reduced in the fraction corresponding to the 39S ribosomal subunit already at 4 weeks of age in Mterf3 heart knockout mitochondria (Figure 6C). This reduction became even more pronounced at later stages and there was eventually a complete loss of the fully assembled ribosomes (Figure 6D). The partial co-migration of MTERF3 and MRPL13 on sucrose gradients (Figure S7D) suggests that MTERF3 is involved in the maturation of the 39S subunit. Next, we investigated whether the impaired ribosomal assembly affected mitochondrial translation by performing in organello de novo translation experiments, which showed no change at the age of 4 weeks and severely decreased translation in 13-week-old Mterf3 knockout heart mitochondria (Figure 6E). A similar global decrease in mitochondrial translation has also been reported in Mterf3 knockdown Drosophila cell lines [38].
We observed that loss of MTERF3 leads to reduced 39S ribosomal subunit assembly and a concomitant decrease of levels of the fully assembled 55S ribosome, in both flies and mice. These results are somewhat reminiscent of the findings in Mterf4 heart knockouts [33], which show accumulation of apparently normal 28S and 39S ribosomal subunits, but a severe reduction in levels of the fully assembled 55S ribosome. Loss of MTERF3 is associated with a drastic reduction of 16S rRNA and impaired assembly of the 39S large ribosomal subunit, suggesting that the 16S rRNA may be interacting with MTERF3. We therefore performed electrophoretic gel mobility shift assays (EMSA) by incubating constant amounts of DNA and RNA templates with increasing amounts of recombinant human MTERF3 protein. Non-specific double- (ds) and single-stranded (ss) DNA or RNA templates with an identical 28 bases long arbitrary sequence only interacted weakly with MTERF3 (Figure S8A, S8B), whereas mitochondrial ribosomal RNA templates showed a stronger binding (Figure S8C, S8D). These binding assays support the prediction that MTERF3 preferentially binds mitochondrial ribosomal RNA. To further characterize MTERF3 interactions with mitochondrial rRNA in vivo, we performed RNA-immuno-precipitation (RNA-IP) in mitochondrial preparations from wild-type mouse heart and fly larvae. The lack of a suitable DmMTERF3 antibody prompted us to generate a transgenic fly line expressing a Flag-tagged form of DmMTERF3 under the inducible UAS-GAL4 system. Expressing DmMTERF3 with a Flag tag directly at the C-terminus leads to an unstable protein not detectable by Western blotting. We therefore introduced a linker sequence between the C-terminus of DmMTERF3 and the Flag tag, and confirmed the expression of this tagged protein by Western blot (Figure S9A). Control experiments demonstrated that we were able to efficiently immuno-precipitate endogenous mouse MTERF3 or Flag-tagged DmMTERF3 proteins (Figure S9B, S9C). RNA-IP clearly demonstrated a very specific interaction between MTERF3 and 16S rRNA both in mouse and fly samples (Figure 7A, 7B).
Regulation of mtDNA expression is important to fine-tune oxidative phosphorylation in response to physiological demand and pathological states. This regulation may occur at many different levels and mitochondria are in essence a prokaryotic system where transcription and translation occur within the same compartment, the mitochondrial matrix, and therefore likely directly interact in a molecular crosstalk. The roles of many of the involved factors are poorly understood. The 39S ribosomal subunit MRPL12 and the posttranscriptional regulator LRPPRC have both been implicated in activation of transcription [18], [39], but these proposed roles are not supported by other studies [14], [19]. MTERF3 has previously been identified as a negative regulator of mtDNA transcription in mammals [5], but its molecular mode of action has remained difficult to assess. Based on the knowledge that the mtDNA gene content is conserved among metazoans, we hypothesized that key regulatory processes controlling mtDNA expression also should be conserved. We therefore decided to take a novel approach to investigate MTERF3 function by creating knockout and knockdown fruit flies with abolished or reduced DmMTERF3 expression. Similar to the mouse, we found that loss of DmMTERF3 results in lethality and activation of mtDNA transcription. Unexpectedly, we could also identify a novel role for DmMTERF3 in mitochondrial ribosome assembly by regulation of the biogenesis of the 39S ribosomal subunit. The biogenesis of the 28S ribosomal subunit was not affected in the absence of DmMTERF3 and instead this subunit accumulated as it could not be assembled into a functional ribosome in the absence of the 39S subunit. Reinvestigation of the Mterf3 knockout phenotype in the mouse showed a similar assembly defect of the 39S subunit, which was present already in early stage knockout animals. In summary, MTERF3 has a novel function in regulation of ribosomal biogenesis and loss of MTERF3 expression does not only impair translation but also causes activation of mtDNA transcription in both flies and mice.
The crystal structure of MTERF3 has been solved at 1.6 Å resolution [40] and predicts that the protein binds nucleic acids. The crystal structure of MTERF1 bound to mtDNA has given novel mechanistic insights how such a binding can occur [28], [29]. On chromatin immuno-precipitation analysis, MTERF3 binds the promoter region of mammalian mtDNA and depletion of MTERF3 from a human mitochondrial extract leads to activation of mtDNA transcription [5]. These effects of MTERF3 on transcription may be due to direct interaction with mtDNA, but it is also possible that MTERF3 modulates transcription by binding the nascent RNA emerging after transcription initiation. In this report, we performed a series of gel-shift analyses, which show that MTERF3 displays weak binding to single- or double-stranded RNA or DNA of random sequence. However, recombinant MTERF3 has a marked preference for binding mitochondrial rRNA fragments, containing both single and double stranded regions. RNA-IP studies further demonstrated a strong and specific interaction of MTERF3 with the 16S rRNA in vivo, suggesting that MTERF3 contributes to 16S rRNA stabilization and/or modification and thereby explaining the critical role for MTERF3 in the biogenesis of the 39S ribosomal subunit. We propose that without this putative 16S rRNA modification, the 39S ribosomal subunit cannot be properly assembled, which, in turn, leads to a severe translational defect. There is strong precedence that abolished modification of mitochondrial rRNAs can affect the assembly of the ribosome. The best understood example is TFB1M, which is an adenine methyltransferase that dimethylates two highly conserved adenines at a stem loop structure close to the 3′ end of 12S rRNA in mammalian mitochondria [23]. Another well characterized example is MTERF4, which interacts with NSUN4 and brings this cytosine methyltransferase to the large ribosomal subunit, where it is thought to modify 16S rRNA [33], [34].
In bacteria, transcription and translation are coordinated, and the rate of transcription is tightly coupled to the processivity of the translating ribosome [41]. Mitochondria may coordinate gene expression in a similar way, where transcription and translation are oppositely coordinated, because loss of assembled ribosomes leads to a massive increase in de novo transcription. Interestingly, knockout of Mterf3 [5], Tfb1m [23] and Mterf4 [33] all cause a severely defective translation and a dramatic increase in de novo transcription, with increased steady-state levels of most or all mitochondrial transcripts. These findings suggest that one of the early responses to a ribosomal assembly defect is a massive transcriptional activation. It is interesting to note that the MTERF3 protein levels are down-regulated in both Tfb1m [23] and Mterf4 [33] knockouts, suggesting further that MTERF3 may have a key role in mediating the effects on transcription, and that up-regulation of mitochondrial transcription in the absence of MTERF3 cannot simply be attributed to a passive compensatory mechanism. We propose that MTERF3 promotes translation by regulation of ribosomal biogenesis and that this process is linked to repression of mtDNA transcription activation (Figure 7C). In the absence of MTERF3, the ribosomal biogenesis is impaired and there is an increased uncontrolled activation of mtDNA transcription leading to imbalanced steady-state levels of mitochondrial transcripts (Figure 7C). Our present data in combination with previous reports [5] suggest that MTERF3 could have a dual function and be a part of a molecular checkpoint, acting to coordinate transcriptional and translational rates and thereby optimizing mtDNA expression.
Unraveling the function of specific proteins is not always easy in mammalian systems and many of the methods used to study protein functions and interactions are plagued by experimental ambiguities. Here, we describe a strategy that takes advantage of genetic manipulation of the orthologous gene in two distantly related metazoans, accompanied by a comprehensive molecular characterization. By using this strategy, we present compelling evidence that MTERF3 has a conserved role in ribosomal biogenesis in metazoans and that it also coordinates mitochondrial transcription and translation. At least two members of the mammalian MTERF family, i.e. MTERF3 and MTERF4, have now been found to have critical roles in mitochondrial ribosomal biogenesis. This makes it tempting to speculate that MTERF1 and MTERF2 could have similar, yet undiscovered, roles in ribosomal biogenesis. Future studies will have to clarify whether MTERF3 has protein interaction partners that are involved in modifying rRNA or if MTERF3 is essential for ribosomal biogenesis by some other mechanism.
This study was performed in strict accordance with the recommendations and guidelines of the Federation of European Laboratory Animal Science Associations (FELASA). The protocol was approved by the “Landesamt für Natur, Umwelt und Verbraucherschutz Nordrhein-Westfalen”.
In order to rescue phenotypes caused by DmMterf3 RNAi expression, we co-expressed the Mterf3 gene from Drosophila pseudoobscura (Dp), which is not a target of the DmMterf3 RNAi line [42]. The fosmid clone FlyFos047383 that contains the DpMterf3 gene was kindly provided by Dr. Pavel Tomancak (MPI for Cell Biology, Dresden, Germany). A 10 kb Fosmid fragment containing the DpMterf3 gene was cloned into the pBluescript II Sk+ vector (Stratagene) by ET recombination. Subsequently, the 10 kb fragment was released by NotI and BglII restriction enzyme cleavage and subcloned into the transfection vector pattB [43]. Oligonucleotide primers used for cloning are listed in Table S1. Embryo injections to achieve site-specific integration into attP40 flies were performed by Best Gene Drosophila Embryo Injection Services (Chino Hills, California, USA).
DmMterf3 null mutants were generated by ends-out homologous recombination as described [37]. Approximately 4 kb of 5′ and 3′ flanking sequences of the DmMterf3 gene were cloned into the pBluescript II Sk+ vector (Stratagene) by ET recombination, using a DmMterf3 BAC clone as template (BACPAC Resource Center, Oakland, California, USA). Both 5′ and 3′ homologous arms were sequenced to ensure the absence of base substitutions and subsequently subcloned into the pGX attP vector [44] to generate the DmMterf3 targeting plasmid. Primer sequences and restriction sites used for subcloning into the pGX attP vector are listed in Table S1. The targeting plasmid was injected into D. melanogaster embryos via P-element-mediated germ line transformation using the Best Gene Drosophila Embryo Injection Services (Chino Hills, California, USA). Crosses for ends-out homologous recombination were carried out as described [37]. Homologous recombination events were identified by PCR. Subsequently, the white (hs) marker was removed using cre-recombinase and the absence of the DmMterf3 gene was confirmed by PCR and sequencing (primers are listed in Table S1). The maintenance of the fly lines is described in Text S1.
The DmMterf3 cDNA clone LD27042 was purchased form DGRC and subsequently cloned into the transfection vector pUASTattB. A Flag tag was linked to the C-terminus of DmMTERF3 via a linker sequence (GAAAAGAAAAG), generating the DmMTERF3-linker-Flag construct. Oligonucleotide primers used for cloning are listed in Table S1. The construct was embryo injected into attP40 flies for generation of the transgenic flies.
Fly larvae (n = 3–7) were dissected in PBS and resuspended in 2 ml of respiratory buffer (120 mM sucrose, 50 mM KCl, 20 mM Tris-HCl, 4 mM KH2PO4, 2 mM MgCl2, 1 mM EGTA, 0.01% digitonin, pH 7.2). Oxygen consumption was measured at 25°C using an oxygraph chamber (OROBOROS). Complex I-dependent respiration was assessed by adding the substrates proline (10 mM), pyruvate (10 mM), malate (5 mM) and glutamate (5 mM). Succinate and glycerol-3-phosphate dehydrogenase activities were measured using 20 mM succinate (SUCC) and 15 mM glycerol-3-phosphate (G3P), respectively. The mitochondrial quality of each sample was assessed by measuring the respiratory control rate (RCR) using 1 mM ADP (state 3) or 1 mM ADP and 2.5 µg/ml oligomycin (pseudo state 4). Permeabilized control mitochondria consistently had RCR values between 4 and 7 with complex I substrates. The respiration was uncoupled by the addition of 400 µM CCCP and the rotenone-sensitive flux was measured in the presence of 200 µM rotenone. Finally, the protein content was determined by the Bradford method (BioRad) in order to normalize the oxygen consumption flux to mitochondrial protein content.
Mitochondria were isolated from fly larvae or mouse hearts and in organello transcription assays were performed as described [45] by incubating 200 µg mitochondria in a modified transcription buffer (30 µCi [α-32P]-UTP, 25 mM sucrose, 75 mM sorbitol, 100 mM KCl, 10 mM K2HPO4, 50 µM EDTA, 5 mM MgCl2, 1 mM ADP, 10 mM glutamate, 2.5 mM malate, 10 mM Tris-HCl (pH 7.4) and 5% (w/v) BSA) for 45 min. Labeled mitochondrial RNA was isolated using Totally RNA kit (Ambion), separated on a 1.2% agarose gel and blotted to Hybond-N+ membranes (GE Healthcare).
In vitro assays to study mitochondrial de novo translation with [35S]-methionine were performed as described [20] and equal amounts of total mitochondrial protein were separated on 15% SDS-PAGE gels. Gels were fixed in isopropanol-acetic solution, stained with Coomassie, destained in ethanol-acetic acid solution and treated with Amplify Solution (GE Healthcare). Afterwards gels were dried and [35S]–methionine-labeled proteins were visualized by autoradiography. For in organello transcription and translation fly mitochondria were incubated at 30°C and mouse heart mitochondria at 37°C.
Mitochondrial ribosomes from fly larvae or mouse hearts were prepared as previously described [23], [46]. Mitochondrial ribosomes were loaded onto 10–30% sucrose gradients and separated by centrifugation overnight. From sucrose gradients, fractions (∼500 µl) were collected with continuous monitoring of absorbance at 260 nm. RNA was extracted from each fraction using TRIzol LS Reagent (Invitrogen) according to manufacturer's recommendations, subsequently treated with DNase and used for cDNA synthesis. Absolute qRT-PCR analysis using a standard curve composed of an equal amount of RNA from each fraction from the control and the KD group, was performed using SYBR green master mix and primers specific for 12S and 16S rRNA, as described in Text S1.
Fractions (750 µl) were collected from the mouse heart sucrose gradients and proteins in each fraction were precipitated with trichloracetic acid and subjected to SDS-PAGE followed by immunoblotting. Sub-ribosomal particles were detected using antisera specific for individual proteins from the 28S and 39S ribosomal subunits, as described in Text S1. RNA was extracted from each fraction using TRIzol LS Reagent (Invitrogen) according to manufacturer's recommendations, subsequently treated with DNase and used for cDNA synthesis and qRT-PCR with TaqMan probes specific for 12S and 16S rRNA [19].
Mitochondria were isolated by differential centrifugation from control (w;;da-GAL4/+) and MTERF3-linker-Flag expressing (w;;UAS-DmMterf3:linker:Flag/+;da-GAL4/+) larvae and from wild-type mouse heart mitochondria. RNA-IP was performed essentially as previously described [33]. The final mitochondrial pellet was suspended in a low-salt NET-2 lysis buffer buffer (50 mM Tris-HCl [pH 7.4], 150 mM NaCl, 0.05% Nonidet P-40, 1× complete EDTA-free protease inhibitor cocktail [Roche]) supplemented with 100 U of RNasin Plus (Promega). After 20 min incubation on ice, the mitochondrial lysates were centrifuged at 10,000 g for 10 min at 4°C in order to pellet the debris. Supernatants were collected and protein concentrations determined by Bradford-based assay (Sigma). IPs were performed using ∼200 µg mitochondrial protein, and lysates were pre-cleared with agarose beads (Sigma) by rotation for 1 h at 4°C, followed by a 2 h incubation at 4°C with Anti-Flag M2 Affinity Gel (Sigma) or a mix of protein A agarose/protein G agarose (Roche) coupled to a polyclonal antibody directed against mouse MTERF3 (peptide Specialty Laboratory), for fly or mouse samples, respectively. Both, anti-Flag M2 Affinity gel and protein A/protein G agarose bead mix coupled to MTERF3 antibody, were incubated for 1 h with 100 µg yeast tRNA prior to usage. After incubation with fly or mouse mitochondria, beads were washed by rotation for 2×10 min at 4°C in low-salt NET-2 buffer, followed by 2×5 min washes in high-salt NET-2 buffer (50 mM Tris-HCl [pH 7.4], 300 mM NaCl, 0.05% Nonidet P-40, 1× complete EDTA-free protease inhibitor cocktail [Roche]), and a final wash for 4×10 min in low-salt NET-2 buffer. The washed beads were resuspended in 120 µl reversion buffer (50 mM Tris-HCl [pH 6.8], 1% SDS, 5 mM EDTA, 10 mM DTT) supplemented with RNasin Plus (Promega) and incubated for 45 min at 65°C. RNA was recovered by TRIzol extraction (Invitrogen) following manufacturer's recommendations, using 10 µg yeast tRNA (Ambion) as a carrier. RNA was subjected to DNase treatment (Turbo DNA-free kit, Ambion) and reverse transcribed to cDNA by using the High-Capacity cDNA Archive kit (ABI). Mitochondrial transcripts from the RNA-IP experiments were identified and quantified by qRT-PCR, with non-primed beads used as background controls.
|
10.1371/journal.pmed.1002425 | Perinatal mortality associated with induction of labour versus expectant management in nulliparous women aged 35 years or over: An English national cohort study | A recent randomised controlled trial (RCT) demonstrated that induction of labour at 39 weeks of gestational age has no short-term adverse effect on the mother or infant among nulliparous women aged ≥35 years. However, the trial was underpowered to address the effect of routine induction of labour on the risk of perinatal death. We aimed to determine the association between induction of labour at ≥39 weeks and the risk of perinatal mortality among nulliparous women aged ≥35 years.
We used English Hospital Episode Statistics (HES) data collected between April 2009 and March 2014 to compare perinatal mortality between induction of labour at 39, 40, and 41 weeks of gestation and expectant management (continuation of pregnancy to either spontaneous labour, induction of labour, or caesarean section at a later gestation). Analysis was by multivariable Poisson regression with adjustment for maternal characteristics and pregnancy-related conditions. Among the cohort of 77,327 nulliparous women aged 35 to 50 years delivering a singleton infant, 33.1% had labour induced: these women tended to be older and more likely to have medical complications of pregnancy, and the infants were more likely to be small for gestational age.
Induction of labour at 40 weeks (compared with expectant management) was associated with a lower risk of in-hospital perinatal death (0.08% versus 0.26%; adjusted risk ratio [adjRR] 0.33; 95% CI 0.13–0.80, P = 0.015) and meconium aspiration syndrome (0.44% versus 0.86%; adjRR 0.52; 95% CI 0.35–0.78, P = 0.002). Induction at 40 weeks was also associated with a slightly increased risk of instrumental vaginal delivery (adjRR 1.06; 95% CI 1.01–1.11, P = 0.020) and emergency caesarean section (adjRR 1.05; 95% CI 1.01–1.09, P = 0.019). The number needed to treat (NNT) analysis indicated that 562 (95% CI 366–1,210) inductions of labour at 40 weeks would be required to prevent 1 perinatal death. Limitations of the study include the reliance on observational data in which gestational age is recorded in weeks rather than days. There is also the potential for unmeasured confounders and under-recording of induction of labour or perinatal death in the dataset.
Bringing forward the routine offer of induction of labour from the current recommendation of 41–42 weeks to 40 weeks of gestation in nulliparous women aged ≥35 years may reduce overall rates of perinatal death.
| National guidelines recommend that induction of labour is carried out between 41 and 42 weeks of gestation to prevent the risks associated with prolonged pregnancy. However, women having their first baby at age 35 years or over are at increased risk of pregnancy complications, including perinatal death.
A recent randomised controlled trial demonstrated that induction of labour at 39 weeks of gestation has no short-term adverse effect on the mother or infant among nulliparous women aged 35 years or older. However, the trial was underpowered to address the effect of routine induction of labour on the risk of perinatal death.
The present study aims to answer the question ‘Does routine induction of labour at or after 39 weeks of gestation reduce the risk of perinatal mortality in first-time mothers aged 35 years or older, compared with expectant management?’
In this national cohort study of 77,327 first-time mothers aged 35 or older, induction of labour at 40 weeks of gestation was associated with a 66% lower risk of perinatal death (0.08% versus 0.26%) than expectant management.
Perinatal death is a rare outcome even in this group and 562 inductions of labour at 40 weeks would be required to prevent 1 perinatal death.
Bringing forward the routine offer of induction of labour from the current recommendation of 41–42 weeks to 40 weeks of gestation in this group of women may reduce overall rates of perinatal death.
| Across industrialised nations, the proportion of births to women aged ≥35 years is rising [1,2]. In England and Wales, births to women aged ≥35 years have increased from 6% of all births in 1975 to 21% in 2015 [3]. There has also been an increase in the number of babies born to first-time mothers aged ≥35 years, which in 2015 accounted for 14% of all first-time births and 5.4% of all births in England and Wales [4].
Older women are at increased risk of pregnancy complications, including gestational diabetes, placenta praevia, and postpartum haemorrhage [5,6], and experience higher rates of intervention during labour and delivery [5,7]. After controlling for comorbidities, the risk of antepartum stillbirth at term is higher among women aged ≥35 years than among younger women [8] and is higher still for nulliparous women aged ≥35 years [9]. Observational data indicate that induction of labour at or before term may be beneficial because the risk of perinatal death is at its lowest for births between 38 and 39 weeks of gestation [9]. However, current United Kingdom guidelines recommend that induction for prolonged gestation is offered to women between 41 to 42 weeks of gestation, when the risk of stillbirth is 2 to 3 per 1,000 deliveries [10].
A recent randomised controlled trial (RCT) has shown that among nulliparous women aged ≥35 years, elective induction of labour at 39 weeks of gestation had no significant effect on the rate of caesarean section and no adverse short-term effects on maternal or neonatal outcomes compared with expectant management [11], but the trial was underpowered to examine the effect of induction of labour on the risk of perinatal death. Well-conducted observational studies have found that induction of labour at term is associated with decreased perinatal mortality in the general pregnant population [12]; however, none has been sufficiently powered to examine the impact on this specific age group known to be at increased risk. We employed a large English routine dataset to determine the association between induction of labour at ≥39 weeks and the risk of perinatal death among nulliparous women aged ≥35 years.
We designed our methods to test the hypothesis that induction of labour at 39, 40, and 41 weeks reduced the risk of perinatal mortality among nulliparous women aged ≥35 years compared with expectant management (continuation of pregnancy to either spontaneous labour, induction of labour, or caesarean section at a later gestation).
The study is exempt from UK National Research Ethics Service (NRES) approval because it involved the analysis of an existing dataset of anonymised data for service evaluation. Hospital Episode Statistics (HES) data were made available by NHS Digital (Copyright 2015, re-used with the permission of NHS Digital. All rights reserved.) Approvals for the use of anonymised HES data were obtained as part of the standard NHS Digital data access process.
We used the HES database to identify births in English National Health Service (NHS) hospitals. The HES database contains patient demographics, diagnostic and procedure information, and administrative data for each inpatient episode of care since 1997 [13]. A unique identifier enables studies to combine episodes of care that belong to the same patient.
In the HES database, for episodes related to childbirth, supplementary fields (the ‘maternity tail’) capture details including parity, birthweight, gestational age, onset of delivery, method of delivery, and birth outcome. Mothers’ delivery episodes were defined as records containing information about the mode of delivery in either the OPCS4 codes (R17–R25) or the maternity tail.
Diagnostic information is coded using the International Classification of Diseases, 10th Revision (ICD10) [14], and operative procedures are coded using the UK Office for Population Censuses and Surveys Classification, 4th Revision (OPCS4) [15]. The level of data completeness has improved over time [16,17] but varies across NHS hospitals: in 2014, data on onset of labour and gestational age were available in 85% and 82% of all delivery episodes, respectively [18].
We included all nulliparous women aged 35 to 50 years delivering at 39 weeks of gestation or beyond, between April 2009 and March 2014. We excluded multiple births; women with preexisting comorbidities (hypertension, diabetes, and cardiac or lung disease); births complicated by fetal malpresentation, abdominal pregnancy, and placenta praevia; and pregnancies resulting in perinatal deaths due to congenital abnormality. We excluded records that were missing birth status, delivery onset, or gestational age. We also performed data quality assessments at the individual hospital level and excluded hospitals with suspected poor-quality data for these key data items (S1 Appendix, Text A). The characteristics of women excluded on the basis of these assessments were compared with those of the study cohort. Limiting birth cohorts in this way to include only hospitals with high completeness of recording has been demonstrated to be a valid way of constructing cohorts from routine hospital data [16,17]. The mothers’ delivery records were linked to the hospital records of their babies using probabilistic linkage [19] to obtain data on perinatal outcomes using the babies' birth records (e.g., for in-hospital perinatal death) and any subsequent hospital inpatient or A&E records (e.g., emergency neonatal readmission within 28 days). Induction of labour was defined as either surgical induction by amniotomy; medical induction, including the administration of agents either orally, intravenously, or intravaginally; or a combination of surgical and medical induction.
For an observational study to appropriately examine the outcomes of induction of labour at different gestational ages, it is important to compare outcomes of women who have an induction of labour at a particular week of gestation (week n) with women who are expectantly managed, i.e., go on to deliver at a later gestation by any mode of onset, and not with women who labour spontaneously at the same gestation. There are 2 possible ways to define the expectant management group using observational data in which gestational age is recorded in weeks (see Stock et al. [12] for discussion):
For the primary analysis, we adopted the first approach, including women delivering at weeks >n following spontaneous or induced labour or prelabour caesarean. The robustness of this approach was then tested using a secondary analysis that used the alternative definition.
For each week of gestation examined, we excluded women if their labour was induced following an antepartum intrauterine death or prelabour rupture of membranes because, in both conditions, if labour does not begin spontaneously within 24 hours, the standard management is induction of labour [10]. However, we did not exclude women with these complications from the expectant management group if the event occurred after the week of gestation of induction in the induction group (at weeks >n). This followed a similar approach used in previous studies [12] and is supported by evidence that delays in the delivery of antepartum stillbirths or PROM are uncommon in UK hospitals [20]. Intrapartum stillbirths in all weeks ≥n were included in both groups.
We examined the following neonatal outcomes: stillbirth, in-hospital perinatal death, birth injury, shoulder dystocia, hypoxia in labour, meconium aspiration syndrome, neonatal seizures, and emergency readmission to hospital within 28 days of birth. In-hospital perinatal death was defined as stillbirth or in-hospital neonatal death within 7 days of birth. We also recorded the following maternal outcomes: emergency caesarean section, instrumental delivery, third or fourth degree perineal tear, and emergency readmission to hospital within 28 days of delivery. To calculate readmission rates, births that occurred in the last 28 days of the study period were excluded. Details of the definitions used are given in S1 Appendix, Table A.
All analyses were prespecified as described in the Methods section, with the exceptions of the exclusion of women with preexisting comorbidities and the use of Poisson rather than logistic regression. These modifications were made prior to the commencement of any statistical analyses; the rationale for each change is provided in S1 Appendix, Text B. We did not publish or pre-register a plan for this analysis.
We used proportions to summarise the distribution of pregnancy characteristics of induced and non-induced women and the chi-squared test for comparisons of variables between the groups. For each week of gestation, univariable and multivariable Poisson regression with robust standard errors was used to estimate the crude and adjusted effects of induction of labour compared with expectant management on each maternal and perinatal outcome. We chose not to use logistic regression because odds ratios overestimate the risk ratio for common outcomes [21]. The confounding variables included in all models were maternal age (years); ethnicity (white, Asian, black, other, or unknown); year of birth; baby’s sex (male or female); birthweight percentile according to UK 1990 fetal growth charts (<10th, 10th–90th, or >90th) [22]; and socioeconomic quintile according to the a Index of Multiple Deprivation (IMD) score, a measure that combines economic, social, and housing indicators [23]. The year of the birth was recorded as a linear variable to take into account changes in clinical practice over time. Estimates were adjusted for pregnancy-related conditions (pregnancy-induced hypertension, preeclampsia, or oedema; gestational diabetes; and fluid abnormalities) when these were found to have significant coefficients. No formal tests of interaction were done, and no adjustments were made for multiple comparisons. For both our primary and secondary analyses, we estimated the number of inductions of labour needed to prevent 1 perinatal death: number needed to treat (NNT) = 1/([induction of labour event risk]−[expectant management event risk]). All statistical tests were 2-sided, and the level of significance was set at P < 0.05. All analyses were performed in STATA version 14.1 (StataCorp, College Station, TX, United States).
There were 77,327 women aged 35–50 years who met the inclusion criteria and gave birth in hospitals that passed the data quality assessments for key data items (Fig 1). Of these women, 25,583 (33.1%) were induced and 51,744 (66.9%) were not. Induction of labour rates among this group of women increased each year during the time period from 30.2% in 2009–2010 to 35.7% in the 2013–2014 cohort. Medical induction of labour was the principal method of induction throughout the time period (57.7% of inductions), with surgical and combined methods used less frequently (19.7% and 19.4% of inductions, respectively).
There were 50,761 eligible women who gave birth in hospitals that failed the data quality assessments for key data items required in the analysis, and these women were therefore not included in the study. The women included in the study shared similar characteristics to the excluded women who gave birth in hospitals with poor data quality (S1 Appendix, Table B). Hospitals that failed the data quality assessments were missing data on gestational age, birth status, and onset of labour in 40%, 25%, and 24% of records, respectively, compared with 11%, 10%, and 11% of records in hospitals that passed these assessments.
Women who had labour induced were more likely to be over 40 years old, of white ethnicity, and to deliver infants in less than the 10th centile for birthweight (Table 1). They were also more likely to have acquired complications of pregnancy. Fig 2 describes the composition of the cohorts used for the primary and secondary analyses.
Labour induction from 40 weeks onwards was associated with a significantly reduced rate of both in-hospital perinatal death and stillbirth when compared with expectant management (Fig 3).
In the primary analysis, the adjusted risk ratio (adjRR) for in-hospital perinatal death associated with induction compared with expectant management was 0.33 (95% CI 0.13–0.80) at 40 weeks and 0.24 (95% CI 0.09–0.65) at 41 weeks (Table 2). However, there was no difference in the estimated adjRR for in-hospital perinatal death associated with induction at 39 weeks (0.37; 95% CI 0.12–1.15). Similar magnitudes of effect were observed for stillbirth, with an adjRR of 0.25 (95% CI 0.09–0.79) at 40 weeks and 0.18 (95% CI 0.05–0.65) at 41 weeks. The results of the secondary analysis were also consistent with those in the primary analysis for both outcomes (Fig 3 and S1 Appendix, Table C). The unadjusted risk ratios did not materially differ from the risk ratios adjusted for maternal characteristics (Table 2 and S1 Appendix, Table C).
The NNT analysis indicated that 562 (95% CI 366–1,210) and 658 (95% CI 421–1,506) inductions of labour at 40 weeks would be required to prevent 1 perinatal death, for the primary and secondary analysis, respectively.
In the primary analysis, labour induction from 39 weeks onwards was associated with a significantly reduced rate of meconium aspiration syndrome, when compared with expectant management (Fig 3). Induction at 39 weeks was also significantly associated with reduced rates of hypoxia in labour (adjRR 0.74; 95% CI 0.65–0.85). However, this association was not significant at later weeks of gestation. Labour induction at 40 weeks was associated with higher rates of neonatal readmission to hospital within 28 days of birth (adjRR 1.30; 95% CI 1.03–1.50). Induction at 41 weeks was associated with reduced rates of birth injury (adjRR 0.47; 95% CI 0.29–0.78) and neonatal seizures (adjRR 0.50; 95% CI 0.26–0.99). No differences were found in the rates of shoulder dystocia in association with induction. Similar observations were seen in the secondary analysis, although the association with neonatal seizures was not replicated in secondary analysis (adjRR 0.67, 95% CI 0.38–1.16), and the association with higher neonatal readmission was seen at 39 as well as 40 weeks (Fig 3 and S1 Appendix, Table C).
In the primary analysis, no differences in the rates of emergency caesarean section or instrumental delivery were found in association with induction at 39 weeks when compared with expectant management (Fig 3). Induction at 40 weeks was associated with a slightly increased risk of instrumental vaginal delivery (adjRR 1.06; 95% CI 1.01–1.11) and emergency caesarean delivery (adjRR 1.05; 95% CI 1.01–1.09). Induction at 41 weeks was associated with a slightly reduced risk of emergency caesarean section (adjRR 0.94; 95% CI 0.90–0.97) compared with expectant management. No differences were found in the rates of severe perineal tears in association with induction. Induction at 39 weeks was associated with higher risk of maternal readmission within 28 days of delivery (adjRR 1.38; 95% CI 1.13–1.60). In the secondary analysis, induction from 39 weeks onwards was associated with a 20%–30% increased rate of emergency caesarean section when compared with expectant management and a 10% increased rate of instrumental delivery at 39 and 40 weeks (S1 Appendix, Table C).
The key finding of the present study is that induction of labour at 40 weeks of gestation was associated with a third of the risk of perinatal death compared with expectant management in a national cohort of nulliparous women aged ≥35 years. At this stage in pregnancy, the risk of perinatal death with expectant management was 26 per 10,000 pregnancies, whereas the risk among women induced at 40 weeks was 8 per 10,000 pregnancies. If these associations are causal, these data indicate that 562 (95% CI 366–1,210) inductions of labour would be required to prevent each perinatal death. Induction of labour was also associated with a significantly reduced risk of meconium aspiration syndrome.
A recent RCT has demonstrated no short-term adverse effect on the mother or infant of routine induction of labour at 39 weeks for women of advanced maternal age, and an associated economic evaluation has suggested that such a policy was associated with lower healthcare costs, principally through reduced rates of maternal readmission [24]. These findings could be taken together with those of the present study and used to support a policy of actively offering all women aged ≥35 years in their first pregnancy to have labour induced around their due date.
However, it could be argued that, because the data in this study are observational, the findings may be due to bias. We do not feel that this is a plausible explanation. First, the women being induced had higher proportions of risk factors, such as very advanced age and acquired complications of pregnancy. Second, we were able to adjust for these and other confounding factors in our statistical models, and this was without material effect. Third, while the risk adjustment models did not include all potential confounders, if there was an imbalance in their distribution across the 2 groups, it is likely that the confounders would be more prevalent in the induction group. Fourth, other well-conducted observational studies using other datasets have found similar effects on perinatal mortality, albeit not in this specific maternal age group [12].
The major cause of perinatal death at term is antepartum stillbirth. It is biologically plausible that stopping pregnancy at week 40 prevents the possibility of an antepartum stillbirth at 41 weeks. It may be argued that induction of labour in this context should only be widely recommended when these results are confirmed by an RCT, but the rarity of the outcome means a trial would be difficult. A sample size calculation based on the observed rates of perinatal death and effect size in the present study indicates that around 15,000 women would be required for a trial with 90% power to detect similarly large effects at 40 weeks of gestation. A well-funded multicentre RCT managed to recruit just over 600 such women [11], which is 4% of the required sample size.
In our primary analysis, we observed a 5% increase in the rate of emergency caesarean section and a 6% increase in the rate of operative vaginal delivery in the induction group. Although the 35/39 trial demonstrated no statistically significant association between induction and these outcomes, the point estimates for the effects in the present study are within the 95% CIs reported in the RCT [11]. Given the higher-risk nature of the women being induced, the adverse associations with induction of labour may be due to unmeasured confounders, as discussed above. These issues may be resolved by a larger-scale RCT of routine induction of labour in 6,000 nulliparous women aged ≥35 years, which is currently in progress [25]. However, that study will not be powered to detect a reduction in the risk of perinatal death, on the basis of the sample size calculation above.
The present study had a number of methodological strengths. The cohort is large and was drawn from an unselected population-based database which records the necessary data items for the appropriate comparison groups to be defined for this study. The use of a novel technique to link mothers’ and babies’ hospital records [19] enabled an examination of both maternal and perinatal outcomes, including morbidity as well as mortality. We were also able to follow up newborns after hospital discharge to examine emergency hospital readmission rates. A disadvantage of hospital administrative birth data is varying data quality between hospitals, which led to the records from some hospitals being excluded from this study. However, this approach has been demonstrated by others to be a valid way of constructing birth cohorts [16], and validation studies have found that the data are nationally representative for key variables, including sex, gestation, birth weight, maternal age, stillbirth, and multiple birth [19].
Despite the large sample size in the present study, it could be argued that because the number of observed perinatal deaths in the induced group is very small, under-recording of labour induction in the dataset could have a major impact on the results. However, since HES is used to guide the reimbursement of maternity care expenses and labour induction is recognised within the national pricing framework [26], we would expect hospitals not to overlook this procedure when coding. To reduce the risks associated with under- and over-coding of induction, we also excluded hospitals that appeared to have divergent coding practices from the analysis (S1 Appendix, Table A).
As with other studies using routine data to examine this issue, we addressed the limitation of gestational age being recorded in weeks rather than in days by testing the robustness of our primary definition of the expectant management group using a secondary analysis that used the alternative definition [12]. Nevertheless, we were not able to control for some important possible confounders, such as maternal obesity, nor to examine some important outcomes, such as postpartum haemorrhage. It is also possible that a small number of antepartum stillbirths were inappropriately included in the expectant management group in the case of delay between death and delivery following induction. Inclusion of these women in the expectant management group would overestimate the risk of perinatal death rate in this group. However, this bias is more likely to affect the results of our secondary analysis, and we found that the magnitude of effect was similar in both analyses.
In summary, our results suggest that among women aged ≥35 years, induction of labour at term is associated with a lower rate of perinatal mortality and morbidity. Hence, bringing forward the routine offer of induction of labour from the current recommendation of 41–42 weeks to 40 weeks of gestation in this group may reduce overall rates of perinatal death. It is, however, important to note the potential for downsides to a policy which would significantly increase the use of labour induction, and further studies should examine the impact of such a policy on resource utilisation and patient satisfaction.
|
10.1371/journal.pntd.0003501 | Minipool Caprylic Acid Fractionation of Plasma Using Disposable Equipment: A Practical Method to Enhance Immunoglobulin Supply in Developing Countries | Immunoglobulin G (IgG) is an essential plasma-derived medicine that is lacking in developing countries. IgG shortages leave immunodeficient patients without treatment, exposing them to devastating recurrent infections from local pathogens. A simple and practical method for producing IgG from normal or convalescent plasma collected in developing countries is needed to provide better, faster access to IgG for patients in need.
IgG was purified from 10 consecutive minipools of 20 plasma donations collected in Egypt using single-use equipment. Plasma donations in their collection bags were subjected to 5%-pH5.5 caprylic acid treatment for 90 min at 31°C, and centrifuged to remove the precipitate. Supernatants were pooled, then dialyzed and concentrated using a commercial disposable hemodialyzer. The final preparation was filtered online by gravity, aseptically dispensed into storage transfusion bags, and frozen at <-20°C. The resulting preparation had a mean protein content of 60.5 g/L, 90.2% immunoglobulins, including 83.2% IgG, 12.4% IgA, and 4.4% IgM, and residual albumin. There was fourfold to sixfold enrichment of anti-hepatitis B and anti-rubella antibodies. Analyses of aggregates (<3%), prekallicrein (5-7 IU/mL), plasmin (26.3 mU/mL), thrombin (2.5 mU/mL), thrombin-like activity (0.011 U/g), thrombin generation capacity (< 223 nM), and Factor XI (<0.01 U/mL) activity, Factor XI/XIa antigen (2.4 ng/g) endotoxin (<0.5 EU/mL), and general safety test in rats showed the in vitro safety profile. Viral validation revealed >5 logs reduction of HIV, BVDV, and PRV infectivity in less than 15 min of caprylic acid treatment.
90% pure, virally-inactivated immunoglobulins can be prepared from plasma minipools using simple disposable equipment and bag systems. This easy-to-implement process could be used to produce immunoglobulins from local plasma in developing countries to treat immunodeficient patients. It is also relevant for preparing hyperimmune IgG from convalescent plasma during infectious outbreaks such as the current Ebola virus episode.
| Plasma-derived immunoglobulin G (IgG) is on WHO’s Essential Medicines List, yet developing countries face severe shortages of this critical treatment. Infusion of IgG prepared from locally-collected plasma provides an advantageous mix of antibodies to viral and bacterial pathogens found in the living environment, and this can reduce recurrent infections in immune-deficient patients. We developed a simple manufacturing process using disposable equipment (blood bags, hemodialyzer, and filters) to isolate immunoglobulins from minipools of 20 plasma donations. This process yields a ca. 90% pure virally-inactivated immunoglobulin fraction at 50–60% recovery. Anti-hepatitis B and anti-rubella immunoglobulins were enriched fourfold to sixfold. The product was free of in-vitro thrombogenic and proteolytic activity, confirming its expected clinical safety profile. Virus validations showed caprylic acid treatment robustly inactivated or removed infectivity of lipid-enveloped viruses, including human immunodeficiency virus (HIV) and hepatitis C virus model. This simple and cost-effective process is implemented in Egypt to prepare experimental batches for clinical evaluation. It can enhance immunoglobulin supplies to treat immunodeficient patients through passive transmission of antibodies directed against local pathogens. The method requires minimal training and reasonable infrastructure, and is a practical means to prepare convalescent hyperimmune IgG during infectious outbreaks such as the current Ebola episode.
| Plasma products to treat congenital bleeding and immunological diseases are made in industrialized countries using complex technologies unavailable in the developing world [1]. Low- to medium-income countries may have little or no access to these life-saving products; these nations urgently need practical processing methods to produce them affordably. We have introduced the concept of small-scale (“minipool”) plasma processing methods implementable with minimum infrastructural requirements. We developed viral inactivation and protein purification technologies in single-use equipment to prepare virally safe solvent/detergent-filtered (S/D-F) plasma for transfusion as well as minipool S/D-F cryoprecipitate to treat bleeding disorders [2–4]. Similarly simple technologies are desperately needed to make safe immunoglobulin G (IgG), a product on the Essential Medicine List of the World Health Organization, to treat immune-deficient patients. Thus we describe here a small-scale caprylic acid IgG fractionation process that requires minimal capital investment and uses disposable equipment. This production approach could increase the supply of IgG in developing countries and improve treatment of immunodeficient patients. It is also a realistic approach to consider in the preparation of convalescent immunoglobulins during infectious outbreaks such as the current Ebola virus epidemic [5,6].
Whole blood was collected with CPD-A anticoagulant/preservative solution (ratio: 14ml/100ml of blood) from regular volunteer non-remunerated donors at Shabrawishi Hospital Blood Bank (Giza, Cairo, Egypt). Donors received information prior to donation in compliance with national regulations. The procedure was approved by the Institutional Review Board from Cairo University (Number N-5–2014). The blood bank is licensed (license number N°7) by the General Directorate for Blood Transfusion Affairs, Ministry of health and is ISO certified (ASR number 1230).
Non-leuco-reduced blood was centrifuged at 3600x g for 12 minutes within 4 hours of collection. Plasma was transferred into storage bags, frozen in a -40°C freezing room, and stored at ≤-25°C for a maximum of 12 months.
The preparation of the IgG fraction is summarized in Fig. 1. Plasma from 20 blood donations tested for anti-A and anti-B titer < 32 (Micro Typing Cards with NaCl; DiaMed AG, Cressier sur Morat, Suisse), or from the same blood group, were subjected to in-bag cryoprecipitation [2,7]. The cryoprecipitate-poor supernatant (approximately 200mL) was transferred into a transfusion bag, frozen and stored at <-30°C. Supernatants were thawed at 30–35°C. After thawing, caprylic acid (Merck, Darmstadt, Germany) was added within one minute to each bag under constant manual stirring to 5% (v/v) final concentration, pH 5.5 +/- 0.1, and the mixture incubated at 31+/- 0.5°C for 90 minutes at 150 rpm in a temperature-controlled shaker-incubator (Lab Therm LT-W, Kühner, Switzerland) [2]. Precipitated proteins were removed by centrifugation (KR4i, Jouan, St Herblain, France) at 3500x g for 45 minutes. The clear supernatants (approximately 2.8 L) were pooled under laminar flow into a SD Virus Inactivation Bag Cascade (VIPS SA, Colombier, Switzerland) and concentrated (typically 60 g/L) using a sterile single-use hemodialyzer (F6HB, Fresenius, Bad Homburg, Germany), a hemodialysis pump and monitoring equipment (Terumo BCT, Lakewood, CO, United States). The solution was progressively diluted with 5 volumes of sterile pyrogen-free saline solution and subjected to diafiltration to remove caprylic acid and concentration. The Ig fraction was centrifuged (Jouan) at 3500x g for 45 minutes at 2–4°C to remove any particulates, then filtered by gravity through a pyrogen-free pharmaceutical-grade BC0025L60SP03A cartridge (3M Cuno, Cergy-Pontoise, France) and a 0.2 μm Mini-Kleenpak sterilizing filter (Pall Corporation, Dreieich, Germany) and directly dispensed under laminar flow into sterile plastic storage bags and stored frozen at <-25°C.
Total protein was determined by Biuret (Protein Kit 110307, Merck Millipore, MA, USA). Zone electrophoresis was performed on agarose gels (Hydragel 7 protein kit, Sebia, Evry, France), staining with amidoblack and densitometric analysis by a semi-automated Hydrasys instrument (Sebia). Sodium dodecylsulfate polyacrylamide gel electrophoresis (SDS-PAGE), under non-reducing and reducing conditions, used 4%~12% Bis-Tris Gel (NuPAGE, Novex Life Technologies, CA, USA) as before [8]. Albumin was measured photometrically using bromocresol green (DiaSys Diagnostic Systems, Holzheim, Germany). IgG, IgA, and IgM were determined by immunoturbidimetry [9]. Anti-hepatitis B and anti-rubella immunoglobulins G titres were determined using Architect Anti-HBs Reagent and Architect Rubella IgG reagent, respectively (Abbott Laboratories, North Chicago, IL, USA). Molecular size distribution was analyzed by size exclusion chromatography on a TSKGel G3000SWXL column (7.8mm ID X 30 cm L) protected with a TSKGel guard column (6.0mm ID X 4.0 cm L), equipped with isocratic pump model SDS 9414, UV-VIS detector model S3210, Rhiodyne manual injector and PeakSimple Chromatography Data System SRI Model333 as data integrator (Schemback, Germany). The mobile phase was 0.1 M sodium sulfate, 50 mM sodium acetate, 0.05% sodium azide, pH5, flow-rate was 0.5 ml/min, and the detection wavelength was 280nm. Thrombin generation assay (TGA) used Technothrombin fluorogenic substrate and RC High reagent (Technoclone, Vienna, Austria), and prekallikrein activator (PKA), plasmin, thrombin, thrombin-like amidolytic activities used S-2302, S-2251, S-2238, and S-2288 chromogenic protease substrates (Chromogenix, Milan, Italy), respectively [10]. Factor XI coagulant activity was measured by one-stage thromboplastin time coagulation assay with human factor XI—deficient and reference plasma (DiaMed, Cressier, Switzerland), and FXI/FXIa antigen with Human Factor XI quantitative sandwich ELISA (Abcam, Cambridge, UK) as before [10]. Endotoxins were determined by the LAL assay. A licensed human IgG preparation produced by a combined ethanol fractionation-chromatography process was used as a control. Data are expressed as the mean ± standard deviation.
The 0.5-mL samples were mixed with 1mL ice-cold methanol and incubated overnight in a deep freezer at -80°C. Samples were centrifuged at 4000x g for 20 minutes and 1 mL of the supernatant was taken and filtered through a 0.45μm syringe filter to a 1.5 ml clean tube. Samples were processed and analyzed by HPLC (Schemback SFD GmbH, Bad Honnef, Germany) equipped with analytical pump (SFD 9414), UV/VIS detector (S 3210), Rheodyne manual injector model 7725i, Peak Simple Data System model 333 (SRI, Torrance, California, USA), and Luna 5u, C8(2) 100Å (150 mm x 4.6 mm) column chromatography (Phenomenex, Torrance, USA); 0.1% trifluoroacetic acid (TFA) in a 80:20 mixture of methanol (Fisher Scientific, UK) and water was used as mobile phase at 0.8 mL/min. Caprylic acid detection was done at 214nm wavelength
Di (2-ethylhexyl) phthalate (DEHP) was assessed on the starting plasma and final IgG. Samples were processed and analyzed as before [2] by HPLC (Schemback SFD GmbH, Bad Honnef, Germany) equipped with analytical pump (SFD 9414), UV/VIS detector (S 3210) at a wavelength of 202 nm, using a Lichrospher 100 RP 18–5μ (250 mm x 4.6 mm) column (CS-Chromatographie Service GmbH, Langerwehe, Germany). A mixture of 85:15 of acetonitrile and methanol (Merck) was used as mobile phase at 1.5 mL/min for 8 minutes analysis time.
The capacity of the caprylic acid treatment to inactivate/remove viruses was assessed at Texcell (Evry, France), a specialized laboratory working under GLP compliance awarded by the ANSM, France’s National Agency for Medicines and Health Products Safety. The process was scaled-down by a factor of 10 (40 mL). Validations were performed in duplicate under worst-case conditions using cryo-poor plasma as starting material, 4.8% caprylic acid, and a temperature of 28.5–30.5°C. The study followed Good Laboratory Practices and CPMP recommendations [11]. Plasma tested negative for HBsAg; HIV-1/HIV-2 Ab+P24 Combo assay; Anti-HCV by Abbott Architect Chemiluminescence (Abbott Laboratories); HBV, HIV and HCV individual-donor-Nucleic Acid Test (NAT) (Tigris; Grifols Diagnostic Solutions Inc., Emeryville, CA, USA) using the Procleix Ultrio assay. Cryo-poor plasma samples were prepared at the Shabrawishi Hospital Blood bank, frozen at -30°C and shipped with dry ice to Texcell. HIV-1 (Lai strain), bovine viral diarrhoea virus (BVDV; NADL strain; ATCC VR-534), and pseudorabies virus (PRV; Aujeszky disease virus; Kojnock strain; ATCC VR-135) were used for spiking, and P4-CCR5, MDBK cell lines (ATCC CCL-22) and Vero (Molecular Virology Laboratory, Institut Pasteur, Paris, France), respectively, for titration assays. Cryo-poor plasma was transferred into the reaction container. When temperature reached 28.5–30.5°C, the material was spiked with virus-inoculum (2.0% [v/v]). Spiked starting material was homogeneized and positive controls were collected; 4.8% (v/v; final concentration) caprylic acid was added in less than one minute. The spiked solution was kept at 28.5–30.5°C under continuous transversal agitation. Samples were taken right after caprylic acid addition (T0) and at 15, 45, 60, and 120 minutes after, and were immediately centrifuged at 3500x g for 45 minutes at 4°C. Supernatants were recovered and diluted 30 folds (BVDV) or 50 folds (PRV and HIV) with culture medium to stop the reaction, and were frozen and stored at -70°C. Control samples prepared under these conditions were verified not to induce cellular toxicity. Spiked samples were titrated by validated end-point dilution assay, and the viral clearance of the steps was assessed in terms of infectivity. Viral titers were calculated and expressed as 50% tissue culture infective dose per milliliter (TCID50/mL) using the Sperman Kaber formula. Infectious titers were calculated at the non-interfering dilution after large volume titration.
The safety of the purified IgG was evaluated in Sprague—Dawley rats as described before [2], after approval from the National Cancer Institute (Cairo, Egypt), where the study was performed. It was conducted according to institution guidelines on animal studies and following recommendations specified in the Code of Federal Regulations Title 21, except that the injection was done intravenously, not intraperitonally, and the observation period was 14 days instead of 7. Twenty-one healthy rats weighing 80–100 g and not used previously for any test purpose were divided into three groups of seven rats. Animals received a dose of 6.5 mL/kg of saline, commercial IgG (control) or minipool IgG. The rats were observed once daily for abnormal behavior or clinical signs. Body weight, and consumption of water and food was recorded at 10 time-points over the observation period. The study was only observational and did not require anesthesia, sacrifice nor dissection.
The main characteristics of the final preparations are summarized in Table 1. The product was clear, with a slight bluish color, and not turbid. Sodium content and osmolality were close to physiological value. Mean protein concentration was 60.2 g/L, with a content of gamma-globulins close to 90%, and traces of albumin, alpha-1, alpha-2, and bêta-proteins revealed by zone electrophoresis (Fig. 2A and B). The relative abundance of IgG was 82–85%, IgA 11–13%, and IgM 4–5%, close to the physiological proportion. Albumin was less than 3 g/L. Content of high molecular weight proteins/aggregates was less than 3% and monomers and dimers more than 90% by HPLC. Anti-A and anti-B isoagglutinin titer in all batches was less than 1/32. Titer of anti-hepatitis B and anti-rubella immunoglobulin G showed enrichment factors compared to plasma of 5.8 and 4.1, respectively. Proteolytic and thrombogenic activity were also assessed. Mean PKA was 6.1 ± 1.1 IU/ml (control: 3 IU/ml), well below the maximum limit of 35 IU/ml in the European Pharmacopoeia. TGA data showed a peak thrombin of 0–223 nM (control: 56.8 nM) below the threshold value of 350nM associated with thromboembolic activity in some IVIG preparations. Mean plasmin was 26.3 mU/mL (control: 20.3 mU/mL), thrombin 2.5 mU/mL (control: 24 mU/mL), thrombin-like proteolytic activity 0.011 U/g protein (control: 0.06 U/g protein), Factor XI activity < 0.01 IU/mL, and Factor XI/XIa 2.4 ng/g protein. SDS-PAGE (Fig. 3) under non-reducing conditions (A) evidenced that most proteins migrated at a MW close to 150–160 kDa (immunoglobulins G and A). Minor protein bands were detected at MW close to 25, 60, and 80–90 kDa. Under reducing conditions (B), two major protein bands with MW of 50 and 25 kDa (immunoglobulin G heavy and light chains, respectively) were detectable. The minipool IgG pattern was similar to control apart for additional protein bands with MW of approximately 150–160, 90, and 70kDa under reducing conditions. Caprylic acid in the final preparation was <750ppm, and DEHP <5ppm. Endotoxin content was less than 0.5 EU/ml.
Viral validation data (Table 2) showed that (a) viral infectivity was not affected after spiking to the starting material, and (b) HIV-1, BVDV, and PRV inactivation was fast and complete within 15 minutes after caprylic acid addition. Reduction factors (duplicate experiments) were > 5.69 and > 5.74 log for HIV-1, > 5.23 and > 5.35 log for BVDV, and > 5.10 and >5.10 log for PRV.
General safety tests did not induce rat mortality nor behavioral changes in the three groups and there was no significant difference in body weight increase rate over 7 days (Table 3). The water and food consumption was not noticeably different among the three groups, ranging between 4–13% and 9–14% of total available, respectively, over the 14 days of observation post infusion.
Polyvalent IgG, the leading plasma product [12], is manufactured from thousands of liters of US or European plasma pools fractionated in highly sophisticated facilities using complex and highly regulated large-scale technologies [1,13]. This IgG product is in short supply and has a mix of antibodies that is not adapted to the treatment of patients in tropical areas, who are exposed to different pathogens. It is therefore crucial to develop small-scale, easy-to-use technologies adapted to process the plasma available in non-Western countries. We show here that a minipool human IgG-enriched fraction can be prepared using a simple process run in disposable CE-marked equipment. The technique relies on caprylic (octanoic) acid precipitation of non-Ig molecules [14,15]. Caprylic acid is already used to prepare licensed IgG or IgM preparations from precipitates II+III [16,17], II [18], or III [18]; 5%-pH 5.5 caprylic acid precipitation is also used to produce horse plasma-derived therapeutic antivenom immunoglobulins [19,20], and this process may become an alternative to chromatography for monoclonal antibodies production [21].
The minipool Ig fraction contained approximately 90% Ig, with a ratio of IgG/IgA/IgM similar to plasma. The content in IgA does not constitute a risk for primary immunodeficient patients who cannot develop anti-IgA. The preparation should not be infused to IgA-deficient patients, although such risks were recently highlighted as being not evidence-based in many patients [22]. Residual proteins included albumin, alpha-1, alpha-2, and bêta-proteins. Aggregate content was low (below 3%) in compliance with requirements for commercial intravenous IgG. It is important to ensure that the anti-A /-B titers of the minipool IgG is consistently low, especially in a situation where the preparation would be used in hematologically challenged recipients, as is the case of Ebola patients. All batches prepared here had a titer below 1:32, less than the limit of 1:64 in the European Pharmacopeia. One means to reduce the ABO isoagglutinin titer is to mix plasma donations for each batch based on a ratio of 30% group A, 30% group B, 20% group AB and 20% group O, an approach already done for a universal pooled plasma for transfusion [23]. Another means we have now implemented is the preparation of minipool IgG from specific blood group plasma donations allowing transfusion to matched blood group patients. The preparation was free of proteolytic activity and had PKA levels within European Pharmacopoeia limits. Recent thromboembolic events associated with intravenous or sub-cutaneous IgG [24,25] prompted us to check most particularly for procoagulant markers [26] especially in situations when such minipool Ig preparation would be used in individuals (e.g. ebola patients) experiencing disseminated intravascular coagulation. Results using TGA, the current preferred assay for assessing the thrombogenicity of IgG [25,27,28], chromogenic substrates for thrombin and thrombin-like activity, and ELISA and coagulant assays to detect FXI and FXIa strongly suggested that the preparation is devoid of relevant in vitro thrombogenic and proteolytic activity. Previous spiking experiments using cryoprecipitate-poor plasma had shown that 5%-pH5.5 caprylic acid incubation inactivates/precipitates FXI/FXIa [9], confirming our findings. We could not assess Fc fragment integrity and anticomplementary activity, but the fact that immunoglobulins are not precipitated during the whole procedure argues in favor of unlikely molecular alteration and aggregation [12]. The disposable hemodialyzer was effective in removing caprylic acid to undetectable levels. It may also contribute to removal of DEHP plasticizer, together with the adsorption filter used prior to sterile filtration and dispensing, as previously found for a charcoal filter [2].
Viral safety of the preparation relies on proper donor screening, serological testing of donations [29,30] and, when feasible, single-donor multiplex NAT testing for HIV, HBV and HCV [31], as was done in this research. Most importantly, manufacturing processes should include one or two dedicated viral inactivation steps, a major tripod of viral safety [30,32]. Caprylic acid treatment is known to be a robust viral reduction treatment for both human [16,18,33] and horse-derived IgG [34,35]. Our study confirms that at the concentration and pH used in this work, this is highly effective against lipid-enveloped viruses, as > 4 log of HIV, BVDV and PRV were inactivated within 15 minutes of treatment. Implementing a pH4 incubation step is readily possible as a second inactivation step for lipid-enveloped viruses [30]. The small size of the pool (20 donations) and the neutralizing activity of potential antibodies against hepatitis E or A (HAV) viruses and parvovirus B19 reduce the risks from non-enveloped viruses. Although we did not perform such evaluation, some removal of non-enveloped viruses may occur during caprylic acid precipitation, as reported for parvovirus and HAV by others [16,17,36]. In addition, dedicated virus removal by 20–35nm nanofiltration, as well as duplex nucleic acid testing for HAV and Parvovirus B19 could be considered to improve the safety margin, especially if larger batches are produced, to make these additional steps more cost-effective [37]. Reproducible IgG recovery (55–65%, corresponding to about 4.5 to 5 g of IgG/L plasma) was achieved, consistent with recovery of antivenom immunoglobulins from horse plasma [19].
This process could be implemented readily in blood establishments or national service centers after appropriate operator training and basic equipment acquisition. The caprylic acid treatment is performed in a closed-bag system under continuous gentle transversal agitation in a thermostated shaker incubator. Laminar flow cabinets are used for additions of reagents to protect against bacterial contamination. Blood bank centrifuges are used for centrifugation. Concentration of IgG and caprylic acid removal are done using commercially available single-use pyrogen-free hemodialyzer. These devices are easy to use, do not present cleaning-related cross-contamination risks, and are affordable (about US$10 per unit). IgG concentration and dialysis fraction is fast (typically 120 minutes to concentrate 3.1 to 3.2 L of supernatant IgG, and 90 minutes to perform 5 dialysis cycles), and yields a clear solution free of particles. The bluish color is typical of processes using caprylic acid and is likely due to the presence of residual ceruloplasmin [38]. The resulting concentrate is clarified by simple gravity, without pumps, on a pyrogen-free single-use adsorptive filter connected online to a sterilizing 0.2μm filter, as is also done for S/D-F plasma and S/D-F cryoprecipitate [2]. The process can use whole plasma, cryoprecipitate-poor plasma or prothrombin complex-depleted cryoprecipitate-poor plasma [39] thanks to the robustness of the 5%/pH 5.5 caprylic acid step to precipitate non-Ig proteins [19]. Thus, the process does require training and basic equipment and facility, but is more feasible than current fractionation technologies for implementation in low or medium income countries.
The enrichment factor found for the two specific immunoglobulins monitored (anti-hepatitis B and anti-rubella IgG) shows the capacity of the process to concentrate hyperimmune IgG. This shows the feasibility of applying this production concept to the preparation of IgG from convalescent plasma. This purification/viral inactivation process, based on small volume and disposable equipment, could be ideal for the preparation of hyperimmune IgG from convalescent plasma in infectious outbreaks, as seen currently in West African countries exposed to the Ebola virus [6,40].
Producing a 90% pure immunoglobulin fraction in disposable, single-use devices is feasible. This method could be used to produce immunoglobulins from local plasma in developing countries to protect immunodeficient patients against infectious agents, and could be of interest for preparing hyperimmune IgG from convalescent plasma collected during infectious outbreaks such as the current Ebola virus episode. Clinical evaluations of this preparation in immunodeficient children are on-going and indicate good tolerance and normal IgG half-life.
|
10.1371/journal.pntd.0007647 | Negligible exposure to nifurtimox through breast milk during maternal treatment for Chagas Disease | Treatment with nifurtimox (NF) for Chagas disease is discouraged during breast-feeding because no information on NF transfer into breast milk is available. NF is safe and effective for paediatric and adult Chagas disease. We evaluated the degree of NF transfer into breast milk in lactating women with Chagas disease.
Prospective study of a cohort of lactating women with Chagas disease. Patients were treated with NF for 1 month. NF was measured in plasma and milk by high performance liquid chromatography (HPLC). Breastfed infants were evaluated at admission, 7th and 30th day of treatment (and monthly thereafter, for 6 months).
Lactating women with chronic Chagas disease (N = 10) were enrolled (median age 28 years, range 17–36). Median NF dose was 9.75 mg/kg/day three times a day (TID). Six mothers had mild adverse drug reactions (ADRs), but no ADRs were observed in any of the breastfed infants. No interruption of breastfeeding was observed.
Median NF concentrations were 2.15 mg/L (Inter quartil range (IQR) 1.32–4.55) in milk and 0.30 mg/L (IQR 0.20–0.95) in plasma. Median NF milk/plasma ratio was 16 (range 8.75–30.25). Median relative infant NF dose (assuming a daily breastmilk intake of 150 mL/kg/day) was 6.7% of the maternal dose/kg/day (IQR 2.35–7.19%).
The low concentrations of NF in breast milk and the normal clinical evaluation of the breastfed babies imply that maternal NF treatment for Chagas disease during breastfeeding is unlikely to lead to clinically relevant exposures in the breastfed infants.
Clinical trial registry name and registration number: ClinicalTrials.gov NCT01744405.
| It is not known whether Nifurtimox, a drug for Chagas disease, is significantly transferred into breast milk, and no clinical trials were conducted to evaluate this topic. Treatment with nifurtimox is safe and effective in children and newborns with Chagas disease. Treatment of young women before pregnancy prevents congenital transmission of Chagas disease. This is the first study to measure nifurtimox concentrations in breast milk. We found that presence of nifurtimox into breast milk is limited, and that breastfed babies had normal clinical evaluations with no observable adverse events. None of the mothers had to discontinue breastfeeding due to adverse events. The exposure of nifurtimox through breast milk during the treatment of mothers with Chagas disease does not seem to pose significant risks for the breastfed infants.
| Chagas disease (CD) or American trypanosomiasis, is a parasitic zoonosis caused by infection with Trypanosome cruzi of worldwide distribution, endemic to the Americas, predominantly affecting the poor and medically underserved [1, 2, 3]. Most CD patients are asymptomatic during the acute phase, but progress to a chronic phase that if the disease is left untreated can lead to cardiac and/or digestive complications in almost 30% of the patients [4, 5]. More than 10 million people are infected in South and Central America. Most people acquire CD during childhood. If girls are not treated, they can transmit the infection to their babies. Recently, CD has become a global health problem expanding to virtually all regions of the world via immigration, with many cases reported in Europe and North America [2, 6].
The only two drugs available for the treatment of CD are nifurtimox (NF) and benznidazole (BNZ), Both drugs have similar effectiveness and limitations [3, 7]; Their mechanisms of action, pharmacokinetics or toxicokinetics are still unclear, but they have been used since decades nonetheless, in spite of a high risk of toxicity in adults, especially dermatological reactions [3, 7]. However, the reported incidence of adverse drug reactions (ADR) is much lower in infants and children [8,9,10,11].
Most physicians rarely give women treatment for CD during lactation due to a perceived risk of infant exposure to these drugs through breastmilk. On the other hand, discontinuation of breastfeeding to allow for maternal treatment is not advisable, given that breast milk is the ideal food for newborns, as well as a source of multiple benefits [12]. However, in areas with high birth rates and limited access to health care, the postpartum breastfeeding period may be the only, and brief, period of time when a woman has consistent contact with health services, and may be amenable for CD treatment.
The aim of this study was to prospectively study, for the first time, NF transfer into breast milk in a cohort of lactating women with CD in order to clarify the safety of this practise. The secondary objective is to provide support for evidence-based recommendations for the management of CD during lactation.
The study protocol was approved by the Research and Teaching Committee, and Bioethics Committee of the Buenos Aires Children’s Hospital “Dr. Ricardo Gutierrez”. The approval number is 2011-06-22.1LAC. Written informed consent was obtained from all participants. All parents provided informed consent on behalf of all minor participants evaluated in this trial. The protocol was registered in ClinicalTrials.gov (#NCT01744405).
Women with chronic CD who were breastfeeding and their infants were enrolled in this prospective cohort study at the Parasitology and Chagas Service, Buenos Aires Children´s Hospital “Ricardo Gutierrez”, between October 2011 and February 2012. CD diagnosis we performed with at least two independent serological tests for T.cruzi antibodies, as per routine clinical care.
Exclusion criteria included: None of the included patients were taking any co-medications. Patients with known medical conditions that could affect result interpretation, a positive pregnancy test, history of NF hypersensitivity or previous NF treatment were excluded from the study.
Treatment: Lactating women with CD received 8 to 12 mg/kg/day, TID NF (120 mg NF tablets) p.o. (Lampit, Bayer, El Salvador), for 30 days [13, 10].
A detailed clinical history, physical examination and routine laboratory tests were obtained at diagnosis, at the end of the first week and 30 days into NF treatment. Patients were then followed as per CD treatment and follow up guidelines.
Treatment response was evaluated by T.cruzi specific real-time Polymerase Chain Reaction (PCR) performed at diagnosis and at the end of treatment [14,15]. Patients were instructed to use contraception during treatment; a pregnancy test was performed before enrolment.
Diagnosis of congenital Chagas disease: All infants under 8 months old were monitored for CD using microhematocrit test. Infants with negative parasitemia were later tested by serology at 8 months of age [10]. Children older than 8 months of age were evaluated with two serological tests for T.cruzi antibodies [10]. Growth and psychomotor development was assessed in children by experienced paediatricians. Pediatric evaluations were performed at days 0, 7 and 30 of maternal treatment, and monthly thereafter for at least 6 months.
Breast milk samples (approximately 30 mL) were collected before the start of NF treatment, and on the 7th (+/- 3 days) and 30th (+/- 3 days) day of treatment. Each milk sample was mixed, total volume recorded and an aliquot stored at -20 C° until analysis. Breastmilk lipid content was not measured.
Venous blood was sampled in heparinized tubes, centrifuged at 3,000 g for 10 min and plasma stored at -20 C° and lyophilized prior to analysis.
A high performance liquid chromatography (HPLC) method was used to determine NF concentration in plasma and milk, as described previously [16]. Briefly, plasma samples were deproteinized with 100 μL tricloroacetic acid (30% w/v), vortexed 20 seconds, sonicated for five minutes and then centrifuged at 8,000 rpm for 5 minutes. Supernatants were mixed with 500 μL of ethyl acetate, precipitated with 100 mg of anhydrous sodium sulfate (to a concentration near saturation) and vortexed for one minute. The mixture was centrifuged at 8,000 rpm for 5 minutes and the organic phase of three consecutive liquid/liquid extraction procedures were recovered together and rotoevaporated to dryness at 40°C and 40–80 bars. The residue was resuspended in 250 μL of methanol, vortexed for 20 seconds and centrifuged 2 minutes before injection in the HPLC.
Breast milk samples (1000 μL) were deproteinized by adding 100 μL of trichloroaceticacid (30% w/v), vortexed for 1 minute and sonicated for 10 minutes, after which the samples were filtered through a 0.45 micron membrane by centrifugation at 8,000 rpm for 20 minutes to obtain an ultrafiltrate of breast milk. The ultrafiltrate was directly injected into the HPLC [17,18].
The limit of detection (LOD) and limit of quantitation (LOQ) for plasma and breastmilk were 0.01 mg/L and 0.1 mg/L, respectively.
Milk-to-plasma (MP) ratios were calculated from single milk and plasma concentration measurements. In those patients that had plasma concentrations below the LOQ (but above LOD), a value equal to half the LOQ (i.e. 0.05 mg/L) was imputed in order to provide a realistic estimate of plasma concentrations that would overall avoid under- or overestimating MP ratios in these patients.
Single-point maximum observed milk concentration for each individual was multiplied by 0.15 L/kg/day (i.e. estimated median milk intake for an infant) to yield the absolute infant daily NF dose (in μg/kg/day) that the infant would ingest per day through breastfeeding. The absolute infant daily NF dose was then divided by the weight-normalized maternal NF dose (in μg/kg/day) and multiplied by 100 to estimate the percent Relative Infant Dose (RID) [19,20,21]. In cases where more than one RID estimate was available for the same patient, the highest RID was chosen for the statistical calculations. The RID represents the percentage of the therapeutic dose (usually taken from the maternal dose) that a baby would be exposed during breastfeeding. The NF dose used for calculations (i.e. 10–15 mg/kg/d) is the actual pediatric dose used in clinical practice [10,22].
Ten women and their 10 babies were enrolled in the study. All mothers were in the chronic CD stage; six of them had acquired the infection in Bolivia, 3 in Argentina and 1 in Paraguay. Median age and weight of the mothers were 28 years (range 17–36 years) and 58, 5 kg (range 52–73 kg), respectively. Median infant age at the start of maternal treatment was 6.8 months (range 1 month-11 months), and median weight 7.6 kg (range 5–9.5 kg). All infants were healthy, within 25th to 95th percentiles for weight and height for their respective ages. Three babies were exclusively breastfed and seven also received solid foods. Median maternal daily dose of NF was 9.82 mg/kg/day (range 8.3–12 mg/kg/day). [Table 1]
Six mothers (60%) had adverse drug reactions (ADR) to NF: 4 were mild (1 vomiting and fever, 1 headache and dizziness, 1 eosinophilia and 1 mild leukopenia) and were able to continue treatment, and 2 were moderate (psychomotor agitation and headache) and led to medication discontinuation by patient decision after 9 and 19 days of treatment, respectively. There were no serious ADRs and no infant had to stop breastfeeding. No ADRs were observed in the breastfed infants, nor any changes in their behaviour, weight progress or other effects potentially attributable to NF. All infants were healthy during and after the study, as assessed by paediatricians skilled in the evaluation of paediatric patients with CD.
Breast milk samples, a total of 17, were taken at a median 9.4 days (range 4–21) after start of NF treatment, so that all patients are assumed to have been at steady state for NF plasma concentrations at the time of sampling. Post-treatment breast milk samples were taken within 24 hours after the last dose. Median plasma NF concentration was 0.30 mg/L (9 samples were LOQ) (IQR of samples that were not LOQ, 0.20–0.95 mg/L). Median milk concentration was 2.15 mg/L (IQR 1.32–4.55). Median milk/plasma NF concentration ratio (MPR) was 16 (IQR 8.75–30.25).
Assuming a 150 mL/kg daily milk intake, the estimated median NF daily infant dose was 0.50 mg/kg/day (IQR 0.20–0.69), representing a median RID of 6.70% of the maternal weight-corrected daily dose (IQR 2.35–7.19%).
Among the 10 infants enrolled in the study, 8 turned out not to have congenital CD, as confirmed by serology at 9 months of age; the remaining 2 were diagnosed with congenital CD and were treated accordingly; Both had a serological response and negative conversion of PCR; None of these infants had any medication related ADRs.
Only one mother showed positive qPCR at the end of treatment. The measured NF concentrations for this mother in blood and milk were below LOD. After re-evaluation, this patient admitted to not taking the drug correctly (and therefore her data were left out of the analysis). A new 60 days NF treatment course was started and the qPCR was negative at the end of treatment and during posttreatment follow-up.
CD transmission can take place by contact with the vector (i.e. known as “kissing bugs”), congenitally, and via transfusions or organ transplantation. Every year an estimated 1,300 children are born with congenital CD in Argentina, but less than half are offered access to treatment.
Recently, small outbreaks have also been linked to ingestions of parasite-contaminated food [3,23]. T. cruzi has rarely been detected in human milk, only in mothers with bleeding nipples during acute CD infection. In a previous study of 21 lactating women, our group found no presence of T. cruzi in human milk using qPCR [24]. Even though risks for parasite exposure from breastmilk are unclear, they are unlikely to be significant and CD in the mother is not considered a reason to avoid breastfeeding [25,26,27].
In a previous study by our group, we observed limited transfer of benznidazole (the other drug available for CD) into breastmilk, and no significant risks to the infants [24], and Vela et al later confirmed that benznidazole used during postpartum in women with CD had no negative impacts on the breastfed child, suggesting that there is no need to interrupt breastfeeding [28]. Unfortunately, benznidazole is not consistently available in all endemic countries, which led us to study NF during breastfeeding, encouraged by a theoretical pharmacokinetic model that suggested that the transfer of NF into breastmilk was likely to be very limited [29].
In rural Latin America young women may only sporadically interact with the health system except during pregnancy, delivery and the early postpartum period. Also, short inter-pregnancy intervals may leave few opportunities for CD treatment beyond breastfeeding periods. The heretofore lack of data supporting safety of NF during breastfeeding put health care professionals in the uncomfortable position of deciding between supporting breastfeeding or CD treatment for the mother, thus forgoing widespread recommendations to support exclusive breastfeeding, and to treat CD [12, 30]. However, this choice between Chagas disease treatment and breastfeeding implies risks such as losing the opportunity to treat the mother and hopefully prevent congenital infections in future babies, as well as preventing long term cardiac complications in the mother, or, if treatment is chosen over breastfeeding, increased risks of infant diarrhea, infections and other formula-associated problems.
This study describes the first prospective study of NF transfer to breastmilk in CD patients, suggesting that infants’ exposure to NF via breastmilk would amount to less than 5% of the usual infant weight-corrected NF dose (i.e. 10–15 mg/kg/day). This exposure is below the 10% cut-off commonly used as threshold evaluate risk for exposure to maternal drugs during breastfeeding [31,32,33]. Taking into account the known safety of NF in children, observed NF milk concentrations (i.e. ~10 times lower than therapeutic doses) would not be expected to produce exposures associated to infant ADRs or any other risks. Furthermore, treatment with NF is better tolerated in infants and children with CD than in adults [10,34]. No ADRs were observed in the breastfed infants in our study, and careful evaluation by experienced paediatricians found no behavioural, growth or weight impacts potentially attributable to NF. The potential difficulties of detecting adverse events in children and infants (especially central nervous system events in small infants) have not escaped our attention. However, even if specific ADRs may be hard to pinpoint (e.g. headache), these events do have detectable manifestations that trained pediatricians can detect. Our group also participated in a multidisciplinary study in children using NF to treat Chagas disease and an incidence of 19% of NF related ADRs were observed, the most common being weight decrease, decreased appetite, headache and rash. All ADRs were readily identified by the pediatricians evaluating these children, many of which participated in this study. The overall observed incidence of ADRs in adults in our cohort (60%) is in agreement with the rate previously described in adults [37,13].
Transfer of drugs into breastmilk is a function of molecular weight (MW) and maternal plasma level [31,32,35]. NF is a small molecule (MW = 287) with high oral bioavailability and moderate plasma protein binding (50%) [31]. Our results show clear evidence that the milk concentrations are a function of the plasma concentrations (Table 1). These results follow the general rule stating that drug concentration in human milk are usually low and will seldom lead to levels that could produce a pharmacological response in the nursing infant [31,36]. The MP ratio estimates in our patients was hampered by the fact that many plasma concentrations were below LOD (i.e. detectable but not measurable), thus forcing us to estimate a concentration in order to calculate MP ratios. We chose the value of 50% LOQ (i.e. the median plasma level that is detectable but not measurable) as a good overall estimate for the observed but non-measurable NF concentrations. MP ratios are intended to provide a general (over) estimate of drug transfer into breastmilk for medications taken by the mother, but contain limited information to judge potential exposure of the baby through breastmilk. MP ratios are, in fact, not generally the preferred estimator of potential for infant drug exposure if other, better; indicators of degree of exposure risk are available such as RID. There is an abundance of examples in the literature of drugs that have high MP ratios but negligible infant exposures due to very low milk concentrations [37]. In the case of NF, the median MP ratio of 16 suggests a significant accumulation of NF in breastmilk. Many potential explanations can account for this, but the main possible reason is that NF is a substrate of breast cancer resistance protein (BCRP), which may be responsible for actively transferring it into the breast (and other tissues)35. One patient (P1, Table 1) had an estimated MP ratio of 190. This large MP ratio may be related to BCRP polymorphisms, or other factors. Unfortunately, we do not have enough data to explore this interesting observation further [32].
Given the nature of the design of this study (e.g. in many cases, mothers expressed milk at home and brought it to the clinic the next day), we cannot ascertain whether fore or hind milk was obtained in most occasions, as the main objective was to obtain leftover milk and in no way interfere with infants’ breastfeeding. NF concentrations do not vary significantly depending on fat content, and therefore we did not expect to see much variation between hind and fore milk.
A limitation of this study is the small number of infants enrolled, which makes it impossible to rule out uncommon ADRs. However, relatively large numbers of paediatric CD patients, including infants and neonates, have been treated with NF at therapeutic doses (approximately 8 to 10 times higher than the expected exposure through breast milk based on our data) for the past few decades in many centres in Latin America, and no significant developmental problems or other significant ADRs have been identified to date [10,38,39,29]. We have no reason to believe that a significantly lower exposure would lead to ADRs not observed at therapeutic doses.
The results of this study, the first of its kind in CD, suggest that NF may be compatible with breastfeeding due to limited drug transfer into breast milk, and low overall infant exposure. The currently perceived contraindication to NF treatment during lactation, so far unsubstantiated by any evidence, may lead to lost opportunities to treat lactating women. This conclusion is further supported by the complete absence of ADRs attributable to NF in the breastfed infants. Our study provides, for the first time, support for continuation of breastfeeding during maternal CD treatment with NF, a practice that can potentially benefit many women and their breastfed infants in settings where maternal treatment during breastfeeding may be advantageous.
|
10.1371/journal.ppat.1005688 | Nonclassical MHC Ib-restricted CD8+ T Cells Recognize Mycobacterium tuberculosis-Derived Protein Antigens and Contribute to Protection Against Infection | MHC Ib-restricted CD8+ T cells have been implicated in host defense against Mycobacterium tuberculosis (Mtb) infection. However, the relative contribution of various MHC Ib-restricted T cell populations to anti-mycobacterial immunity remains elusive. In this study, we used mice that lack MHC Ia (Kb-/-Db-/-), MHC Ia/H2-M3 (Kb-/-Db-/-M3-/-), or β2m (β2m-/-) to study the role of M3-restricted and other MHC Ib-restricted T cells in immunity against Mtb. Unlike their dominant role in Listeria infection, we found that M3-restricted CD8+ T cells only represented a small proportion of the CD8+ T cells responding to Mtb infection. Non-M3, MHC Ib-restricted CD8+ T cells expanded preferentially in the lungs of Mtb-infected Kb-/-Db-/-M3-/- mice, exhibited polyfunctional capacities and conferred protection against Mtb. These MHC Ib-restricted CD8+ T cells recognized several Mtb-derived protein antigens at a higher frequency than MHC Ia-restricted CD8+ T cells. The presentation of Mtb antigens to MHC Ib-restricted CD8+ T cells was mostly β2m-dependent but TAP-independent. Interestingly, a large proportion of Mtb-specific MHC Ib-restricted CD8+ T cells in Kb-/-Db-/-M3-/- mice were Qa-2-restricted while no considerable numbers of MR1 or CD1-restricted Mtb-specific CD8+ T cells were detected. Our findings indicate that nonclassical CD8+ T cells other than the known M3, CD1, and MR1-restricted CD8+ T cells contribute to host immune responses against Mtb infection. Targeting these MHC Ib-restricted CD8+ T cells would facilitate the design of better Mtb vaccines with broader coverage across MHC haplotypes due to the limited polymorphism of MHC class Ib molecules.
| Tuberculosis, the disease caused by Mycobacterium tuberculosis (Mtb), remains a major global health burden. As T cells are crucial to the control of Mtb infection, it is imperative to decipher the role of different T cell subsets in anti-Mtb immunity for the development of more effective vaccines. While the contribution of conventional T cells to protective immunity against Mtb is well established, the involvement of unconventional T cells is less clear. In this study, we used mutant mice that lack distinct MHC I molecules to characterize immune responses mediated by unconventional T cells during Mtb infection. We found that unconventional CD8+ T cells preferentially expanded in the lung after Mtb infection. These CD8+ T cells responded to numerous mycobacterial antigens, produced multiple cytokines, and contributed to protection against Mtb. A large proportion of unconventional T cells induced by Mtb infection are Qa-2 restricted CD8+ T cells, suggesting this group of T cells may play a greater role in anti-mycobacterial immunity than other unconventional T cell populations that have been characterized previously. Targeting these unconventional T cells may facilitate the design of TB vaccines that are universally effective in ethnically diverse populations due to the non-polymorphic nature of MHC Ib molecules.
| Tuberculosis (TB), an infectious disease caused by Mycobacterium tuberculosis (Mtb), remains one of the world’s deadliest communicable diseases, with 1.5 million deaths annually [1]. Due to the emergence of multidrug-resistant Mtb strains, co-infection with HIV, and the failure of BCG vaccine to control adult pulmonary TB [1, 2], there is an urgent need for new and more effective TB vaccines. However, achieving this goal relies on further investigation of the properties of protective T cells during Mtb infection [3]. It is well established that immune protection against Mtb infection is dependent on a robust Th1 response, mediated by CD4+ T cells [4–7], while CD8+ T cells are required for optimal immunity [8–11]. The cytokines IL-12, IFN-γ and TNF-α are critical for the control of Mtb infection [12]. Current subunit vaccine candidates target conventional CD4+ and MHC Ia-restricted CD8+ T cells [13]. However, increasing evidence shows that unconventional T cells restricted by MHC Ib molecules can recognize distinct types of microbial antigens and may contribute to host defense against microbial infection [14, 15]. Yet, it remains unclear whether MHC Ib-restricted CD8+ T cells play a protective role during Mtb infection and which MHC Ib molecules may be involved in anti-mycobacterial immunity.
MHC Ib molecules are structurally similar to MHC Ia molecules and associated with β2-microglobulin (β2m) [14]. Unlike MHC Ia molecules, MHC Ib molecules exhibit limited polymorphism, making them attractive targets for vaccine development [14, 15]. The mammalian genome encodes many MHC Ib molecules though only a few are known to have immunological function. These include H2-M3 (M3), Qa-1/HLA-E, Qa-2/HLA-G, CD1 and MHC-related gene 1 (MR1) in mice and/or humans [14]. T cells restricted by MHC Ib molecules have been implicated in host defense against Mtb in humans and mice [15]. In particular, M3-restricted CD8+ T cells recognize several N-formylated peptides derived from Mtb [16] and vaccination of mice with dendritic cells pulsed with N-formylated Mtb peptides conferred protection against Mtb in mice [17]. CD1d-restricted iNKT cells, which recognize self and/or microbial lipids [18, 19], can be activated by Mtb-infected macrophages and lead to the control of intracellular mycobacteria through the production of GM-CSF [20, 21]. In addition, Mtb lipid-specific group 1 CD1-restricted T cells were detected in patients with active or latent TB infection [22, 23] and participated in host adaptive immune responses to Mtb in human group 1 CD1 transgenic mice [24, 25]. Recently, MR1-restricted mucosal-associated invariant T cells (MAIT), which recognize vitamin B metabolites during bacterial infection [26], were also shown to contribute to anti-mycobacterial immunity [27, 28]. In addition, HLA-E, the human homolog of mouse Qa-1, presents Mtb-derived peptides to cytotoxic CD8+ T cells [29, 30]. Although these MHC Ib-restricted T cells were detected or induced following immunization or infection, their relative contribution during Mtb infection has yet to be defined. Furthermore, it is unclear whether other MHC Ib molecules are involved in antigen presentation to T cells during Mtb infection.
Previous studies have shown that MHC Ia-deficient mice were more resistant to Mtb infection than β2m-deficient mice [31, 32]. However, it is unclear whether MHC Ib-restricted CD8+ T cells contribute to the observed protection in these studies because β2m-/- mice have aberrant iron metabolism and impaired innate immunity aside from lacking both MHC Ia and Ib-restricted CD8+ T cells [33]. In addition, it is unclear whether M3 plays a dominant role in MHC Ib-mediated immune responses against Mtb infection, as was the case in Listeria infection [34]. In this study, we used mice that lack MHC Ia (Kb-/-Db-/-), MHC Ia/M3 (Kb-/-Db-/-M3-/-) [34] or β2m (β2m-/-) to study the role of M3-restricted and other MHC Ib-restricted CD8+ T cells in immunity against Mtb aerosol infection. We found Mtb-infected Kb-/-Db-/-M3-/- mice do not have a significantly reduced number of CD8+ T cells as compared to Kb-/-Db-/- mice, suggesting that MHC Ib molecules other than M3 are responsible for the development of a robust CD8+ T cell response to Mtb in the absence of MHC Ia molecules. These non-M3, MHC Ib-restricted CD8+ T cells recognized Mtb-derived protein antigens, expanded preferentially in the lungs of Kb-/-Db-/-M3-/- mice and contributed to protective immunity against Mtb. Furthermore, a large proportion of these expanded CD8+ T cells were restricted to MHC Ib molecule Qa-2, suggesting that this is a new T cell population that participate in immune responses against Mtb infection.
We have previously demonstrated that M3-restricted CD8+ T cells expanded extensively during Listeria infection and played a prominent role in host defense against Listeria [34]. To test whether M3-restricted CD8+ T cells play similar roles during Mtb aerosol infection, we first compared CD8+ T cell responses in the lungs and spleens of Mtb-infected C57BL/6 (B6), Kb-/-Db-/-, Kb-/-Db-/-M3-/- and β2m-/- mice. Consistent with previous reports, almost no CD8+ T cells were detected in Mtb-infected β2m-/- mice. Interestingly, we found that CD8+ T cells in the lungs and spleens of Kb-/-Db-/-M3-/- and Kb-/-Db-/- mice expanded to a similar extent following Mtb infection, although the total numbers of CD8+ T cells in both mouse strains remained lower than those in B6 mice (Fig 1A–1C). Furthermore, Mtb antigen-specific CD8+ T cell responses in Kb-/-Db-/- mice were not reduced upon stimulation with Kb-/-Db-/-M3-/- BMDCs compared to stimulation with Kb-/-Db-/- (M3-sufficient) BMDCs (S1 Fig). These data suggest that among the β2m-dependent CD8+ T cells, H2-M3-restricted CD8+ T cells only represent a small proportion of the CD8+ T cells responding to Mtb infection and that other MHC Ib-restricted CD8+ T cells expand significantly during infection.
To evaluate the protective capacity of non-M3 MHC Ib-restricted CD8+ T cells during Mtb infection, we compared the bacterial burden in both the lungs and spleens of infected B6, Kb-/-Db-/-and Kb-/-Db-/-M3-/- mice with that of infected β2m-/- mice. Mtb-infected Kb-/-Db-/- and Kb-/-Db-/-M3-/- mice showed no differences in bacterial burden in the lungs and spleens over the course of infection, supporting the notion that M3-restricted T cells do not play a prominent role in anti-mycobacterial immunity. However, Kb-/-Db-/-M3-/- mice had significantly lower bacterial loads than β2m-/- mice at day 30 and day 60 post-infection (Fig 1D and 1E), suggesting that non-M3, MHC Ib-restricted CD8+ T cells play a protective role during Mtb infection. Moreover, Kb-/-Db-/-M3-/- and B6 mice had comparable bacterial burden in the lungs and spleens over the course of infection (Fig 1D and 1E), suggesting that the presence of MHC-Ib restricted T cells can compensate for the lack of MHC-Ia-restricted CD8+ T cells to control Mtb infection.
To determine whether non-M3, MHC Ib-restricted CD8+ T cells mediated the protective effect observed in Mtb-infected Kb-/-Db-/-M3-/- mice, Mtb-infected Kb-/-Db-/-M3-/- mice were repeatedly treated with anti-CD8β mAb or control Ab and the bacterial burden in the lungs of these two groups of mice were compared at day 30 post-infection. As shown in Fig 1F, depletion of CD8+ T cells in Mtb-infected Kb-/-Db-/-M3-/- mice resulted in a significant increase in bacterial burden in the lungs compared to mice that received control rat IgG antibody. These results demonstrated that non-M3, MHC Ib-restricted CD8+ T cells contribute to the protective immunity against Mtb.
Since there were no significant differences in the total number of CD8+ T cells and bacterial loads between Kb-/-Db-/- and Kb-/-Db-/-M3-/- mice after Mtb infection, we used Kb-/-Db-/-M3-/- mice in following experiments to further characterize non-M3, MHC Ib-restricted CD8+ T cell responses during Mtb infection. Kinetic analysis of CD8+ T cells in different organs showed that CD8+ T cells from Kb-/-Db-/-M3-/- mice expanded almost 50 fold in the lung but only 3–4 fold in the spleen by day 60 post-infection (Fig 2A), suggesting that non-M3, MHC Ib-restricted CD8+ T cells expand more in the lung than in the spleen. Differential expansion of CD8+ T cells in the lung was also observed in Mtb-infected B6 mice (mostly MHC Ia-restricted), but the degree of expansion (10–12 fold) was less vigorous than that in Kb-/-Db-/-M3-/- mice (Fig 2A). In addition, the absolute numbers of CD8+ T cells with CD44hiCD62Llo effector phenotype (CD8+ TEFF) were comparable between Mtb-infected Kb-/-Db-/-M3-/- and B6 mice, even though in naïve animals, Kb-/-Db-/-M3-/- mice had significantly fewer CD8+ TEFF cells than B6 mice (Fig 2B). These results suggest that Mtb infection induces a more robust expansion CD8+ TEFF cells in Kb-/-Db-/-M3-/- mice than CD8+ Teff cells in wildtype mice.
Conventional effector CD8+ T cells were shown to gradually lose function or become exhausted in the chronic LCMV infection model [35]. The elevated expression of several cell surface receptors including programmed death-1 (PD-1) and killer cell lectin-like receptor G1 (KLRG1) has been correlated with functional exhaustion and terminal differentiation of effector CD8+ T cells [35]. Most of the MHC class Ib-restricted CD8+ T cells in naïve Kb-/-Db-/- or Kb-/-Db-/-M3-/- mice exhibit an activated T cell phenotype (CD44hi, CD11ahi, and CD122+) [31, 34]; however, it is not clear whether MHC Ib-restricted CD8+ T cells would develop a similar phenotype during chronic Mtb infection. We found that at the chronic infection stage at day 60 post-infection, both Kb-/-Db-/-M3-/- and B6 mice had increased expression of KLRG-1 and PD-1 on CD8+ TEFF cells compared to naïve controls. In addition, CD8+ TEFF in Kb-/-Db-/-M3-/- mice expressed a higher level of KLRG1 than CD8+ TEFF in B6 mice, whereas no difference in PD-1 expression on CD8+ TEFF cells was detected between Kb-/-Db-/-M3-/- and B6 mice (Fig 2C and 2D). To further address the question whether activated CD8+ T cells in Mtb-infected Kb-/-Db-/-M3-/- mice can develop into long-lived memory precursor cells, we compared the expression of CD127, the IL-7 receptor α, on CD44hiCD8+ T cells in Mtb-infected Kb-/-Db-/-M3-/- and B6 mice. We found a significantly lower proportion of memory (CD44hiCD62Lhi) CD8+ T cells in Mtb-infected Kb-/-Db-/-M3-/- mice expressed CD127 as compared to those in B6 mice (Fig 2E–2G). These results suggest that MHC Ib-restricted CD8+ T cells acquire a terminally differentiated phenotype during Mtb infection and most of these CD8+ T cells are short-lived effectors.
To compare the magnitude of Mtb antigen-specific CD8+ T cell responses in B6 and Kb-/-Db-/-M3-/- mice, we isolated lymphocytes from the lungs of B6 and Kb-/-Db-/-M3-/- mice at different time points after infection and determined the number of IFN-γ-producing CD8+ T cells upon stimulation with Mtb whole cell lysate (WCL)-pulsed B6 and Kb-/-Db-/-M3-/- BMDCs, respectively. CD8+ T cells in the lungs of Kb-/-Db-/-M3-/- and B6 mice at day 30 after infection both produced IFN-γ in response to stimulation with Mtb WCL (Fig 3A). However, the percentage of CD8+ T cells producing IFN-γ from Mtb-infected Kb-/-Db-/-M3-/- mice was substantially higher than that seen among CD8+ T cells from B6 mice (Fig 3A and 3B). Consequently, there was no significant difference in the absolute number of Mtb-specific IFN-γ producing CD8+ T cells between Kb-/-Db-/-M3-/- and B6 mice during the course of Mtb infection (Fig 3C). These results further support the notion that MHC Ib-restricted CD8+ T cell responses can compensate for the lack of MHC Ia-restricted CD8+ T cell responses during Mtb infection. Importantly, CD8+ T cells from Mtb-infected B6 mice were able to produce IFN-γ when stimulated with Kb-/-Db-/-M3-/- BMDCs (Fig 3D), suggesting that Mtb-specific MHC Ib-restricted CD8+ T cells are induced in B6 mice after Mtb infection.
Previous studies have shown that different MHC Ib molecules present different types of antigens. For instance, H2-M3, Qa-1/HLA-E and Qa-2 molecules were shown to present peptide antigens [30, 36, 37] while MR1 and CD1d present vitamin B metabolites and lipid antigens, respectively [18, 19, 38]. To determine what kinds of antigens MHC Ib-restricted CD8+ T cells from Kb-/-Db-/-M3-/- mice recognize during Mtb infection, CD8+ T cells isolated from the lung of Mtb-infected Kb-/-Db-/-M3-/- mice were stimulated with BMDCs pulsed respectively with Mtb WCL, culture filtrate proteins (CFP) and purified protein derivatives (PPD) and total lipids. We expected that the Mtb lipid fraction would contain antigens presented by CD1d while WCL would contain antigens presented by various MHC Ib molecules, including MR1. Eighteen hours after stimulation with the different Mtb antigen preparation, cytokine production was detected by intracellular cytokine staining (ICS). As shown in Fig 4, MHC Ib-restricted CD8+ T cells from Mtb-infected Kb-/-Db-/-M3-/- mice produced IFN-γ in response to WCL, PPD and CFP stimulation but not total lipids (Fig 4A and 4B). Meanwhile, when stimulated with proteinase K-treated WCL, the frequency of IFN-γ producing CD8+ T cells was reduced by more than 50% (Fig 4A and 4B), suggesting that the majority of Mtb-specific MHC-Ib-restricted CD8+ T cells recognize proteinase K-sensitive antigens. As CFP induced a stronger cytokine response than WCL in CD8+ T cells from Mtb-infected Kb-/-Db-/-M3-/- mice, we used CFP as antigens in following experiments to study the properties of Mtb-specific MHC-Ib-restricted CD8+ T cells.
To determine whether β2m or TAP is required for the presentation of Mtb antigens to MHC Ib-restricted CD8+ T cells, we stimulated CD8+ T cells from Mtb-infected Kb-/-Db-/-M3-/- mice with CFP-pulsed BMDCs derived from B6, Kb-/-Db-/-M3-/-, TAP-/- and β2m-/- mice and measured cytokine production by ICS. We observed a significant reduction in the percentage of IFN-γ+TNF-α+CD8+ T cells upon stimulation with β2m-/- BMDCs compared to stimulation with B6 or Kb-/-Db-/-M3-/- BMDCs (Fig 4C). In contrast, no reduction in the percentage of cytokine-producing cells was observed when stimulated with BMDCs from TAP-/- mice (Fig 4C). These data indicated that MHC Ib-restricted CD8+ T cells recognize Mtb protein antigens in a β2m-dependent and TAP-independent manner. As some MHC Ib-restricted T cells can be activated by pro-inflammatory cytokines in an antigen-independent manner [39, 40], we also examined whether MyD88-mediated signaling affected cytokine-production by CD8+ T cells from Mtb-infected Kb-/-Db-/-M3-/- mice. A similar percentage of cytokine-producing CD8+ T cells was observed upon stimulation with MyD88-deficient BMDCs compared to stimulation with Kb-/-Db-/-M3-/- BMDCs (Fig 4C), suggesting that MyD88 pathway likely does not play a major role in the activation of MHC Ib-restricted CD8+ T cells.
Polyfunctional T cells simultaneously producing multiple cytokines (IFN-γ, TNF-α and IL-2) have been shown to either correlate with protective immunity against Mtb [41, 42] or TB disease activity in humans [43, 44]. To examine whether MHC Ib-restricted CD8+ T cells displayed multi-functionality during Mtb infection, CD8+ T cells isolated from the lungs of Mtb-infected Kb-/-Db-/-M3-/- mice were stimulated with CFP-pulsed BMDCs, and responses were analyzed using ICS for IFN-γ, TNF-α and IL-2. We found Kb-/-Db-/-M3-/- mice had a higher percentage of double-cytokine-producing (IFN-γ+TNF-α+) and triple-cytokine-producing (IFN-γ+TNF-α+IL-2+) CD8+ T cells compared to B6 mice (Fig 5A and 5B). In addition, enriched CD8+ T cells from lungs and spleens of Mtb-infected Kb-/-Db-/-M3-/- mice also produced IL-17A in the presence of CFP (Fig 5C). Taken together, these results indicate that after Mtb infection, the MHC Ib-restricted CD8+ T cell responses in Kb-/-Db-/-M3-/- mice are antigen-specific and polyfunctional.
Several immunodominant antigens recognized by Mtb-specific MHC Ia-restricted CD8+ T cells have been identified [10]. However, little is known about the Mtb antigens recognized by MHC Ib-restricted CD8+ T cells. To address this issue, we examined the reactivity of CD8+ T cells in Mtb-infected Kb-/-Db-/-M3-/- mice to several representative immunogenic Mtb antigens, including CFP, CFP10, ESAT6, Ag85A, Ag85B, Ag85C, PstS1, MPT32 and TB10.44−11 peptide, in an IFN-γ Elispot assay. As shown in Fig 6, enriched CD8+ T cells from lungs and spleens of Kb-/-Db-/-M3-/- mice at day 30 post-infection responded to most Mtb antigens tested except Ag85A and the TB10.44−11 peptide, which is known to be restricted by MHC Ia molecule H2-Kb [45] (Fig 6A and 6B). The Mtb antigen-specific CD8+ T cell responses were higher in the lungs compared to the spleens, but the pattern of reactivity was similar between the organs. Among these Mtb antigen-specific MHC Ib-restricted CD8+ T cells, the frequencies of CFP- and PstS1-specific CD8+ T cells were highest (Fig 6A and 6B). In contrast, CD8+ T cells from B6 mice recognized most of these Mtb antigens at a lower frequency but exhibited strong reactivity to TB10.44−11 peptide (Fig 6C and 6D). These results indicate that MHC Ia and MHC Ib-restricted CD8+ T cells recognize distinct Mtb antigens during infection. It also suggests that PstS1 might be an immunodominant antigen recognized by MHC Ib-restricted CD8+ T cells.
To investigate which MHC class Ib molecules might present Mtb antigens to CD8+ T cells during infection, we enriched CD8+ T cells from Kb-/-Db-/-M3-/- mice and stimulated them with CFP-pulsed as well as Mtb-infected BMDCs derived from various mouse strains (i.e. B6, Kb-/-Db-/-M3-/-, Qa-1-/-, Qa-2null, MR1-/-, CD1d-/- and β2m-/- mice), and examined their cytokine production by ICS. Compared to stimulation with B6 BMDCs, there was no difference in the percentage of cytokine-producing CD8+ T cells when stimulated with BMDCs from Qa-1-/-, MR1-/-, CD1d-/- and Kb-/-Db-/-M3-/- mice in either the CFP-pulsed or Mtb-infected BMDCs (Fig 7A and 7B). This suggested that most of the cytokine-producing CD8+ T cells in Kb-/-Db-/-M3-/- were not restricted by these MHC I molecules. However, when stimulated with BMDCs from Qa-2null mice, the percentage of cytokine-producing CD8+ T cells decreased to the same level as seen with β2m-/- BMDCs stimulation (Fig 7A and 7B), suggesting that a substantial fraction of Mtb-specific CD8+ T cells in Kb-/-Db-/-M3-/- mice are restricted to Qa-2.
To substantiate this finding, we examined the cytokine production of these CD8+ T cells in response to stimulation with Mtb antigen-pulsed BMDCs derived from mice that express different levels of Qa-2, i.e. C57BL/6 (Qa-2hi), BALB/cJ (Qa-2lo), and B6.C-H2d/bByJ (Qa-2null) [46, 47] (S2 Fig). We found the percentage of Mtb-specific IFN-γ-producing CD8+ T cells was higher when CD8+ T cells were co-cultured with BMDCs from C57BL/6 than from BALB/cJ mice and almost undetectable when stimulated with Qa-2null BMDCs, indicating that IFN-γ production was correlated to the expression level of Qa-2 on the BMDCs (Fig 7C). Since Qa-2null BMDCs and B6 BMDCs induced a similar IFN-γ response from LemA-specific M3-restricted CD8+ T cells (S2 Fig), which suggested that Qa-2null BMDCs are capable of presenting antigens to other MHC Ib-restricted T cells. Furthermore, pretreatment of Kb-/-Db-/-M3-/- BMDCs with an anti-Qa-2 antibody (clone 20-8-4)[48], but not control IgG, resulted in a significant reduction of Mtb-specific IFN-γ secretion by CD8+ T cells isolated from Mtb-infected Kb-/-Db-/-M3-/- mice (Fig 7D). Collectively, these results indicate that a large proportion of unconventional Mtb-specific CD8+ T cells found in Kb-/-Db-/-M3-/- mice are Qa-2-restricted. To determine if Qa-2-restricted T cell responses can be detected in Mtb-infected B6 mice, we performed Qa-2 blocking experiment in an ELISPOT assay. We found that the percentage of IFN-γ producing CD8+ T cells in Mtb-infected B6 mice was significantly reduced when stimulated with CFP-pulsed Kb-/-Db-/-M3-/- BMDCs in the presence of anti-Qa-2 blocking antibody (Fig 7E). This suggested the existence of Mtb-specific Qa-2-restricted T cells in B6 mice following Mtb infection.
A recent study showed that MR1-/- mice had higher bacterial load in the lung compared to wild-type mice after BCG infection [26], suggesting a role for MAIT cells during mycobacteria infection. In addition, MAIT cells were found to be highly enriched in the bronchoalveolar lavage fluid of patients with active TB [49]. Although we did not detect significant number of Mtb-specific MR1-restricted T cells when we stimulated CD8+ T cells from Mtb-infected Kb-/-Db-/-M3-/- mice with Mtb-infected BMDCs (Fig 7B), it remains possible that in vitro Mtb-infected BMDCs do not contain sufficient amounts of MR1 ligands. To address this issue, culture supernatants from logarithmic phase of Mtb culture (Mtb sup) were used as antigen(s) since they contain vitamin B metabolites known to activate MAIT cells [38, 50]. As both CD8+ and CD4-CD8- (DN) MAIT cells are present in the lung of naïve mice [51], we stimulated CD8+ and DN T cells from Mtb-infected Kb-/-Db-/-M3-/- mice with Mtb sup-pulsed MR1-/-, β2m-/- and B6 BMDCs, and measured cytokine production by ICS. We found that IFN-γ and TNF-α production by CD8+ T cells was β2m-dependent but not MR1-dependent, while the cytokine production by DN T cells was neither MR1-dependent nor β2m-dependent (Fig 8A). These results suggest that MAIT cells do not constitute a significant proportion of CD8+ and DN T cells expanded in Kb-/-Db-/-M3-/- mice during Mtb infection. Consistent with this finding, no significant differences in the expression of Vα19-Jα33 mRNA, the canonical TCR α chain of MAIT cells [52], were detected in the lungs, spleens and mediastinal lymph nodes at day 30 after Mtb infection, compared to naïve mice (Fig 8B). Furthermore, unlike MAIT cells, which preferentially use Vβ6 or Vβ8, MHC Ib-restricted CD8+ T cells found in Mtb-infected Kb-/-Db-/-M3-/- mice use diverse TCR Vβ chains (Fig 8C). Taken together, our data suggest that MAIT cells do not represent a significant population of MHC-Ib restricted T cells in the mouse model of TB infection.
In this study, we defined the relative contribution of MHC Ia-, M3- and other MHC Ib-restricted CD8+ T cells in Mtb infection by comparing CD8+ T cell responses in B6, Kb-/-Db-/-, Kb-/-Db-/-M3-/- and β2m-/- mice upon aerosol infection with virulent Mtb. Unlike their role in Listeria infection, M3-restricted CD8+ T cells do not play a dominant role in the MHC Ib-restricted CD8+ T cell responses to Mtb infection. This finding highlights the differential roles of various MHC Ib-restricted responses in immunity against distinct microbial pathogens. While CD1-restricted and MR1-restricted T cell responses have been characterized in the context of Mtb infection, our data showed that neither CD1d nor MR1 serve as major restriction elements for the Mtb-specific MHC Ib-restricted CD8+ T cells found in Kb-/-Db-/-M3-/- mice. In fact, we found a substantial fraction of these Mtb-specific unconventional CD8+ T cells were restricted by Qa-2, which is known to present a more diverse array of peptides than other MHC Ib molecules [53]. Qa-2-restricted T cell responses have been implicated in anti-tumor immunity [48] and antiviral immunity [37] [54]. However, this is the first study to describe a role for Qa-2 in host defense against bacterial infection. As HLA-G is a possible functional homolog of Qa-2 [55], it will be of great interest to explore whether HLA-G-restricted Mtb-specific T cell responses can be detected in patients with active TB or BCG-vaccinated individuals.
Previous studies have shown that Kb-/-Db-/- mice were more resistant to Mtb infection than β2m-/- mice [31, 32]. However, these studies did not address whether CD8+ T cells in Mtb-infected Kb-/-Db-/-mice contribute to β2m-dependent resistance to Mtb-infection. The observation from our study that in vivo depletion of CD8+ T cells in Kb-/-Db-/-M3-/- mice resulted in increased susceptibility to Mtb infection, clearly demonstrates a role for non-M3, MHC Ib-restricted CD8+ T cells in the control of Mtb infection. It has been shown that some TCRβ+ CD4-CD8- (DN) T cells or TCRγδ+ T cells recognize MHC Ib molecules. However, we did not detect significant expansion of γδ T cells in Kb-/-Db-/-M3-/- mice after Mtb infection (S3 Fig). In addition, we found that the expansion of DN T cells in Mtb-infected mice was largely β2m-independent (S4 Fig), suggesting that MHC Ib molecules do not play a critical role in the expansion of DN T cells during Mtb infection. Thus, the major contribution of MHC-Ib molecules in Mtb infection is likely mediated through MHC Ib-restricted CD8+ T cells. In contrast to a previous report [31], we did not observe significant differences in bacterial burdens between B6 and Kb-/-Db-/- mice following aerogenic Mtb infection at the time points examined. This discrepancy could in part due to the use of different backcross generations of Kb-/-Db-/- mice (i.e. 6 times backcrossed versus 10 times backcrossed) and/or animal housing environments.
Unlike MHC class Ia-restricted CD8+ T cells, MHC class Ib-restricted CD8+ T cells in naïve mice exhibit activated/memory T cell phenotype that is CD44hi, CD11ahi, CD122hi and Ly6Chi [31, 56]. Our labs and others have shown that MHC class Ib-restricted CD8+ T cells rapidly respond to Listeria infection [57, 58]. Similarly, during Mtb infection, though the frequency of CD8+ T cells in naïve Kb-/-Db-/-M3-/- mice was much lower than that in naïve B6 mice, we eventually detected similar number of effector CD8+ T cells (Fig 2B). In addition, we detected comparable number of IFN-γ-producing CD8+ T cells (Fig 3C) between Mtb-infected Kb-/-Db-/-M3-/- and B6 mice upon ex vivo stimulation with Mtb antigens. These results suggest that MHC Ib-restricted CD8+ T cells mount a more robust response than conventional CD8+ T cells following Mtb infection. It is noteworthy that Mtb-specific MHC Ib-restricted CD8+ T cell responses can be readily detected in the lung of B6 mice following Mtb infection, albeit the magnitude is lower compared to those observed in Mtb-infected Kb-/-Db-/-M3-/- mice. Therefore, it is possible that the presence of MHC Ia-restricted CD8+ T cells could affect the precursor frequency and peripheral expansion of MHC Ib-restricted CD8+ T cells.
Memory-like CD44hi CD8+ T cells [39] and innate-like T cells [40] have been shown to be activated either through the engagement of TCR with microbial antigens or cytokine-driven signals during infection. Some of the Mtb antigens used in our study could enhance the production of inflammatory cytokines by BMDCs in part through the MyD88-dependent pathway (S5 Fig). However, we found CD8+ T cells isolated from Mtb-infected Kb-/-Db-/-M3-/- mice recognize proteinase K-sensitive antigens and their activation is independent of MyD88-mediated signaling, suggesting these MHC Ib-restricted CD8+ T cells are activated through TCR recognition of Mtb-derived antigens. This notion is further supported by our results, which showed MHC Ib restricted CD8+ T cells produced IFN-γ in response to stimulation with several recombinant Mtb protein antigens. Nevertheless, it remains possible that some MHC Ib-restricted CD8+ T cells can be activated in vivo via a bystander mechanism during Mtb infection.
Several studies showed that antigens present in Mtb culture filtrate proteins are highly effective in inducing protective immunity against Mtb infection [10]. We found that non-M3 MHC Ib restricted CD8+ T cells responded to several of these antigens. Among the Mtb antigens tested, the phosphate-binding transporter lipoprotein PstS1 seems to induce the strongest response. In contrast, CD8+ T cells from Mtb-infected B6 mice exhibited strong reactivity to TB10.44−11 peptide, but had a weaker response to other Mtb CFP antigens tested. These data suggest that immunodominant antigens targeted by MHC Ia and MHC Ib-restricted CD8+ T cells are substantially different. Mtb antigen-specific CD8+ T cell responses detected in Kb-/-Db-/-M3-/- mice were TAP-independent. As a large proportion of these Mtb-specific MHC Ib-restricted T cells are Qa-2-restricted CD8+ T cells, it is possible that Qa-2 might present Mtb antigens in a TAP-independent manner, similar to HLA-E in humans [29, 30, 59]. However, it is not clear whether the TAP-independent pathway is preferentially used by Mtb-infected cells presenting antigens to MHC Ib-restricted CD8+ T cells in vivo.
In an effort to further define the role of Qa-2 during Mtb infection, we infected two sub-strains of BALB/c mice, BALB/cJ (Qa-2-sufficient) and BALB/cByJ (Qa-2-deficient), with Mtb and compared the bacterial burden and Mtb-specific T cell responses between these two sub-strains. We found that BALB/cByJ mice lacked Mtb-specific Qa-2-restricted CD8+ T cells and had a higher bacterial burden in the spleen compared to BALB/cJ mice (S6 Fig). These results suggest that Mtb-specific Qa-2-restricted CD8+ T cells contribute to protection against Mtb infection. However, it is known that BALB/cByJ and BALB/cJ mice have additional genetic differences besides Qa-2 expression. Thus, further experiments involving Qa-2 knockout mice in the B6 background (this mouse strain is yet to be generated) are needed to definitively confirm the role of Qa-2-restricted CD8+ T cells during Mtb infection.
In summary, TB remains a challenge for human health and protective cellular immunity is extremely desirable for the design of better Mtb vaccines. Our finding that nonclassical CD8+ T cells, largely Qa-2 restricted CD8+ T cells, can provide protection against Mtb revealed the presence of a potentially novel cellular population in combating tuberculosis. In humans, a large proportion of Mtb-specific CD8+ T cells appear to be restricted by MHC Ib molecules [60]. Targeting these MHC Ib-restricted CD8+ T cells would facilitate the design of better vaccines against Mtb that can induce broader immune protection than targeting MHC Ia-restricted CD8+ T cells in genetically diverse human populations due to the limited polymorphism of MHC class Ib molecules.
This study was carried out in strict accordance with the recommendations in the Guide for the Care and Use of Laboratory Animals of the National Institutes of Health. The protocol was approved by the Animal Care and Use Committee of the Northwestern University (Protocol number: IS00000985).
Kb-/-Db-/-, Kb-/-Db-/-M3-/- [56], and CD1d-/- [61] on the B6 background mice were generated or maintained in house. C57BL/6, BALB/cJ, BALB/cByJ, TAP-/-, β2m-/- and B6.C-H2d/bByJ mice (Qa-2 null) [62] were purchased from The Jackson Laboratory (Bar Harbor, ME). MyD88-/- mice were obtained from Mutant Mouse Resource and Research Centers. MR1-/- [63], Qa-1b-/- [64] mice were provided by Dr. Ted Hansen (Washington University School of Medicine, St Louis, MO) and by Dr. Harvey Cantor (Dana-Farber Cancer Institute, Boston, MA), respectively.
Mtb fractions and recombinant Mtb proteins (CFP10, PstS1, Ag85A, Ag85B, Ag85C, MPT32 and ESAT6) were obtained through BEI Resources (Manassas, Virginia). Mtb fractions and antigens were dissolved in either DMSO or PBS and stored as aliquots at -20°C.
For Mtb aerosol infection, frozen aliquots of Mtb H37Rv were thawed and diluted in PBS with 0.05% Tween 80. Mice were infected with 100–200 CFU using a nose-only aerosol exposure chamber (In-Tox Products, NM), equipped with a Lovelace nebulizer, as previously described [24]. A day 1 count was performed to determine the infecting dose. At indicated time-points after infection, bacterial loads in lungs and spleens were determined by plating serial dilutions of homogenate on Middlebrook 7H11 agar plates (BBL, BD), and colonies were counted after 2–3 weeks of incubation at 37°C.
Single-cell suspensions were prepared from the lung, spleen and mediastinal lymph node by mechanical disruption in HBSS/2% FBS. Lung was digested with collagenase IV (1mg/ml) and DNase I (30μg/ml) for 30 min at 37°C before disruption. To enrich CD8+ T cells, splenocytes and lung leukocytes were labeled with biotinylated mAb specific to CD19 (6D5), CD4 (GK1.5), CD11b (M1/70), CD49b (DX5), TCRγδ (GL-3) and I-A/I-Eb (M5/114.15.2) (Biolegend, San Diego, CA) followed by streptavidin-conjugated magnetic beads (Dynabeads, Invitrogen). The purity and composition of enriched CD8+ T cells were confirmed by flow cytometry. Bone marrow-derived dendritic cells (BMDCs) were derived from mouse bone marrow progenitors using GM-CSF and IL-4 (PeproTech, Rocky Hill, NJ) as previously described [56].
BMDCs were plated in 96-well flat-bottom plate at a centration of 1x105 cells/well and infected at a multiplicity of 1 or 3 with Mtb for 2h at 37°C. Cultures were washed three times and treated with 20 μg/ml gentamycin for 2 h to remove extracellular bacteria. Enriched CD8+ T cells from Mtb-infected mice were added 24h later to the indicated wells for co-culture. To determine bacterial uptake, some BMDCs were lysed with sterile distilled water containing 0.01% SDS and plated on 7H11 plate.
Monoclonal antibodies against mouse CD8α (53–6.7), CD8β (YTS156.7.7), CD44 (1M7), CD62L (MEL14), CD127 (A7R34), PD-1 (29F.1A12), KLRG1 (2F1/KLRG1), B220 (RA36B2), CD4 (GK 1.5), TCRβ (H57-597), Gr-1 (1A8), NK1.1 (PK136), anti-Qa-2 (1-1-2), Vβ2 (B20.6), Vβ3 (KJ25), Vβ4 (KT4), Vβ5.1/5.2 (MR9-4), Vβ6 (RR4-7), Vβ7 (TR310), Vβ8.1/8.2 (MR5-2), Vβ8.3 (1B3.3), Vβ9 (MR10-2), Vβ10 (B21.5), Vβ11 (RR3-15), Vβ12 (MR11.1), and Vβ13 (MR12-3) with different fluorochrome conjugate, were purchased either from BioLegend or eBioscience or BD Bioscience (San Diego, CA). Anti-Qa-2 mAb (20-8-4) was purified from hybridoma culture supernatant using the protein A column. For cell surface staining, cells were incubated with 2.4G2 Fcγ RII/RIII blocking mAb for 15 min, then stained with the appropriate combinations of mAbs in HBSS/2% FBS for 30 min at 4°C. Flow cytometry was performed with a FACS CantoII (BD Biosciences, San Jose, CA) and analyzed using FlowJo software (Tree Star, Ashland, OR).
T cells or CD8+ T cells from the spleen and lung of infected mice were stimulated with unpulsed or Mtb antigen-pulsed BMDCs or Mtb-infected BMDCs. After two hours of incubation, Brefeldin A (10μg/ml, Sigma, St. Louis, MO) was added and cells were cultured for additional 16 hours. After incubation period, cells were harvested, stained for cell surface markers, fixed with 4% paraformaldehyde, permeabilized with 0.2% saponin and stained with APC-conjugated anti-IFN-γ (eBioscience), FITC-conjugated anti-TNFα (Biolegend) and PE-conjugated anti-IL-2 mAbs (eBioscience). Flow cytometry was performed as described earlier.
IFN-γ Elispot assay was performed as previously described [24], with some modifications. Briefly, multiscreen-IP plates (Millipore, Bedford, MA) were coated with anti-IFN-γ mAb (An-18, eBioscience) at 5μg/ml in PBS. BMDCs were pre-pulsed with Mtb fractions or recombinant protein antigens overnight. In a blocking assay, Mtb antigen-pulsed BMDCs were pre-incubated with mouse IgG or anti-Qa-2 mAb (20-8-4) [48] for 30 min before the assay. Enriched CD8+ T cells from infected mice (5×103−2×104) were mixed with BMDCs stimulator cells (5×104/well) in RPMI 10 medium and plated in triplicate wells. After 18h incubation at 37°C, plates were washed using PBS-Tween (PBS and 0.05% Tween 20) and incubated for 2h at room temperature with biotinylated anti-IFN-γ mAb (R4.6A2, eBioscience). Plates were then washed and incubated with streptavidin-conjugated alkaline phosphatase (Jackson ImmunoResearch Laboratories, West Grove, PA). After 1 h incubation at room temperature, plates were developed with a BCIP/NBT substrate kit (Bio-Rad, Hercules, CA) according to the manufacturer’s instructions. Spots were counted using an ImmunoSpot reader (Cellular Technology, Shaker Heights, OH).
Mice were given 100 μg of anti-CD8β mAb (53–5.8) or normal rat IgG (controls) by the i.p. route on days -3, 0, 7, 14, and 21 post infection with Mtb, as described by Wang et al.[65]. Mice were killed on day 28 post infection and the depletion efficiency was determined by flow cytometry.
Total RNA was extracted using Trizol reagent (Invitrogen), and first-strand cDNA was synthesized using the Superscript III reverse transcriptase (Invitrogen) according to the manufacturer’s instructions. Real-time PCR was performed by using IQ5 instrument (Bio-Rad systems). The expression of Vα19-Jα33 transcripts was quantified using primer pairs: Vα19F, 5'-GGTACAGTTACCTGCTTCTGAC-3', and Jα33R, 5'-GATAGTTGCTATCCCTCACAGC-3' [52], and the results were normalized to TCRα constant region mRNA.
Statistical analysis was performed with GraphPad Prism software (GraphPad, La Jolla, CA). When comparing experimental values from two groups of mice, one- or two-tailed student's t-tests were routinely used. When comparing experimental values from multiple groups, two-way ANOVA Bonferroni post-tests were used. Statistically significant differences are noted (***P < 0.001; **P < 0.01; *P < 0.05).
|
10.1371/journal.pgen.1002675 | The DSIF Subunits Spt4 and Spt5 Have Distinct Roles at Various Phases of Immunoglobulin Class Switch Recombination | Class-switch recombination (CSR), induced by activation-induced cytidine deaminase (AID), can be divided into two phases: DNA cleavage of the switch (S) regions and the joining of the cleaved ends of the different S regions. Here, we show that the DSIF complex (Spt4 and Spt5), a transcription elongation factor, is required for CSR in a switch-proficient B cell line CH12F3-2A cells, and Spt4 and Spt5 carry out independent functions in CSR. While neither Spt4 nor Spt5 is required for transcription of S regions and AID, expression array analysis suggests that Spt4 and Spt5 regulate a distinct subset of transcripts in CH12F3-2A cells. Curiously, Spt4 is critically important in suppressing cryptic transcription initiating from the intronic Sμ region. Depletion of Spt5 reduced the H3K4me3 level and DNA cleavage at the Sα region, whereas Spt4 knockdown did not perturb the H3K4me3 status and S region cleavage. H3K4me3 modification level thus correlated well with the DNA breakage efficiency. Therefore we conclude that Spt5 plays a role similar to the histone chaperone FACT complex that regulates H3K4me3 modification and DNA cleavage in CSR. Since Spt4 is not involved in the DNA cleavage step, we suspected that Spt4 might be required for DNA repair in CSR. We examined whether Spt4 or Spt5 is essential in non-homologous end joining (NHEJ) and homologous recombination (HR) as CSR utilizes general repair pathways. Both Spt4 and Spt5 are required for NHEJ and HR as determined by assay systems using synthetic repair substrates that are actively transcribed even in the absence of Spt4 and Spt5. Taken together, Spt4 and Spt5 can function independently in multiple transcription-coupled steps of CSR.
| Class switch recombination (CSR) in B cells is required for interaction with different effector molecules while retaining the affinity for the same antigens. CSR mechanism involves the orchestrated steps of transcription, DNA break, and repair of the target loci. Within the cells, these processes occur at the chromatin level—involving DNA, histones, and their associated post-translational modifications (PTMs). Transcription factors associated with RNA Polymerase II complex often have regulatory roles in chromatin maintenance, which in turn might regulate the process of DNA cleavage and repair. Here we report that the transcription factor DSIF complex (Spt4 and Spt5) is critically required for CSR. The absence of either Spt4 or Spt5 blocked CSR. Interestingly, Spt4 and Spt5, although previously thought to work as a complex, can function independently of each other at several nodes of CSR, namely transcription regulation, DNA break formation, and histone PTM maintenance, exemplified by H3K4me3. The importance of H3K4me3 unifies three programmed recombinations—CSR, VDJ, and meiotic—in their reliance on this modification for their respective DNA cleavage formations. Moreover, Spt4 and Spt5 are required for DNA repair, another critical aspect of CSR, suggesting that the DNA repair steps of CSR may be coupled with transcription.
| Immunoglobulin (Ig) class switch recombination (CSR), which takes place in activated B lymphocytes, alters antibody effector functions by changing the Ig heavy-chain constant region (CH) from Cμ (IgM) to other CHs (namely, IgGs, IgE, or IgA). CSR is initiated by the cleavage of two DNA switch (S) regions, a donor and an acceptor locus, located 5′ to each CH region [1]. S region double-strand breaks (DSBs), which are generated by staggered nicks, are paired and recombined by the general repair mechanisms; non-homologous end joining (NHEJ) or alternative end joining [2], [3], [4], [5]. Simultaneously, the intervening CH region between the donor and acceptor S regions is looped out and deleted [1].
CSR absolutely depends on three critical events: (a) the expression of activation-induced cytidine deaminase (AID), a master regulator of Ig gene diversification processes including CSR, somatic hypermutation (SHM), and gene conversion [6], [7], [8], [9]; (b) the active transcription of S regions, which contain highly repetitive sequences [10], [11]; and (c) the repair and joining of the cleaved DNA ends [5]. The requirement of AID in CSR was convincingly shown by the finding that both Aicda knockout model mice and human patients with AICDA mutations fail to produce Ig isotypes other than IgM [6], [7]. Subsequent AID-mutant studies showed that AID controls two CSR intermediate steps; mutations in the AID N-terminal region affect cleavage of both the S and variable (V) regions, while mutations in the C-terminal domain are capable of cleaving DNA but incapable of recombining the cleaved S regions [12], [13], [14].
Active S region transcription has been associated with efficient CSR [15], [16], and its absolute requirement was demonstrated by gene-targeting experiments [10], [11]. S region transcription is initiated from the I promoter located upstream of each S region, and terminates downstream of the CH region. The mature transcripts, designated as germline transcripts (GLTs), contain the I and CH exons after splicing out the S region and the CH intronic sequences [15], [16]. Active S region transcriptions are predicted to readily form non-B DNA structures due to the repetitive nature of the S region sequences [17],[18],[19],[20]. Indeed, decreasing the amount of topoisomerase 1 (Top1) protein appears to cause excessive negative supercoiling to accumulate behind the transcription machinery, which could facilitate the formation of non-B DNA structures within repetitive sequences in the S regions and triplet repeats; the latter is implicated in Huntington's disease [19], [21], [22]. These unusual DNA structures are proposed to be suitable substrates for Top1-mediated irreversible cleavage [19], [21]. According to the DNA deamination model, transcription-induced DNA structural alteration such as R-loop formation is proposed to be critical for AID to directly deaminate S region cytosine [23], [24], [25].
Transcription by RNA polymerase II (RNAPII) through chromatin is associated with various histone post-translational modifications (PTMs) that are largely regulated by transcription factors and histone chaperones. Indeed, it was recently shown that a member of the histone chaperone FACT complex is involved in Ig class switching through histone PTM modulation, especially H3K4me3 [26]. The requirement of this particular histone modification in CSR is reminiscent of other programmed recombination events, namely meiotic and VDJ recombinations, in which H3K4me3 is essential for their respective DNA cleavages [27], [28], [29].
Following S region DNA breaks, sealing of the two S region ends is mediated by general DNA repair mechanisms. The major pathway is the error-prone, non-homology-mediated end joining (NHEJ), in which the joining is mediated without long nucleotide microhomology between the paired DNA ends. This type of repair relies heavily on the involvement of the Ku70/80 proteins, which act as anchors for other downstream NHEJ factors to assemble [2], [5]. Another pathway to recombine the cleaved ends is microhomology-mediated joining which depends on the homology of single-stranded overhangs of the two S regions [4].
The DSIF (DRB sensitivity-inducing factor) complex, a transcription factor composed of Spt4 and Spt5, was initially discovered as a factor that rendered RNAPII transcription sensitive to the nucleoside analog 5,6-dichloro-1-β-d-ribofuranosylbenzimidazole (DRB) [30]. DSIF's interaction with the RNAPII complex has been widely reported [31], [32], [33]. DSIF was also shown to be distributed across the body of transcribed genes [34] and to facilitate RNAPII transcription elongation [35], indicating that this complex has a positive effect on transcription. Consistent with its regulatory role in transcription, Spt5 has been reported to regulate histone PTMs that are intimately linked with transcription, such as H2B mono-ubiquitination (ubH2B) and subsequently H3K4me3, through the trans-histone modification pathway [36], [37]. Moreover, DSIF has the potential to inhibit RNAPII progression, notably at the early elongation step through promoter-proximal pausing [31], [38], [39]. However, most of these conclusions are based solely on studies of Spt5, the larger DSIF subunit.
While it was recently reported that Spt5 guides AID to its target sites and is required for CSR [40], the importance of Spt4, the smaller DSIF subunit, in CSR has never been addressed. Our independent screening for transcription elongation factors required for both efficient CSR and histone PTM modulations led us to identify both Spt4 and Spt5 as critical factors for CSR. Using the CH12F3-2A B cell line, which robustly switches to IgA when CIT (CD40L, IL4, and TGFb) is added [41], we showed that depletion of either DSIF subunit abolished CSR. Unexpectedly, however, we found that Spt4 and Spt5 function independently in various phases of CSR, including histone PTM regulation, S region DNA breakage and Sμ cryptic transcript suppression. We also found evidence that Spt4 and Spt5 regulate DNA repair, suggesting that these components play diverse roles to regulate CSR though transcription-coupled processes.
First, to examine whether both components of the DSIF complex are involved in CSR, we introduced RNAi oligonucleotides into CH12F3-2A cells to knockdown either Spt4 or Spt5. By using multiple siRNA oligonuclotides recognizing different sequences of the target transcripts, flow cytometry analysis showed that depletion of either Spt4 or Spt5 dramatically reduced IgA switching (Figure 1A and 1B, top) without significant cell death (Figure 1B, bottom). RT-qPCR analysis and immunoblotting confirmed that both factors were significantly reduced following the introduction of the specific RNAi oligonucleotides (Figure 1C and 1D).
As all of the RNAi oligonucleotides specific for Spt4 and Spt5 affected CSR dramatically, we selected oligonucleotide #125 and #16 (recognizing Spt4 and Spt5 transcript, respectively) for the subsequent analyses. Spt5 and Spt4 both remained robust even in the absence of their respective counterparts Spt4 and Spt5, although slight reductions in their expressions were visible (Figure 1E).
The absence of either Spt4 or Spt5 did not negatively affect the levels of other transcripts critical for CSR, such as μ-GLT, α-GLT, or AID; the total VH transcripts were also unperturbed (Figure 1F). On the other hand, the V-Cα transcripts, the final CSR products in CH12F3-2A cells, were reduced drastically, confirming that CSR is indeed blocked if either of the two DSIF subunits, Spt4 or Spt5, is depleted (Figure 1F).
Coincidentally, microarray analysis showed augmented IGHG3 (Cγ3) transcript in CH12F3-2A cells in the absence of Spt4 or Spt5 (see below). Using a primer pair recognizing Iγ3 and Cγ3, we confirmed that γ3GLT was indeed enhanced by depleting either of the two DSIF components (Figure 1G). We therefore examined the class-switching to IgG3 upon CIT stimulation in CH12F3-2A cells. Unexpectedly, we observed significant switching to IgG3 within 24 hours of adding CIT, even though the overall IgG3-positive population was much smaller than the IgA-positive population generated within the same time period (Figure 1H and 1I). Nonetheless, the absence of either of the two DSIF components, while augmenting γ3GLT, significantly reduced the IgG3-positive population. We therefore concluded that both components of the DSIF complex are required for efficient CSR to IgG3 as well as IgA in CH12F3-2A cells.
To ascertain that the CSR suppression in the absence of the DSIF complex is not due to the reduced expression of known critical CSR factors, we performed a microarray analysis of global transcripts. This analysis indicated that only a very small number of genes were transcriptionally affected within 48 hours of introducing either Spt4 or Spt5 RNAi oligonucleotides into CIT-stimulated CH12F3-2A cells (Figure 2A). Moreover, the affected transcripts did not seem to code for any critical proteins presently known to be required for efficient CSR (Table S1). Out of about 23,000 transcripts examined, only 40 and 22 transcripts were up- or down-regulated, respectively, by at least 2-fold in the absence of Spt4 (P<0.05) (Figure 2A, left circles and Table S1). On the other hand, Spt5 knockdown suppressed 111 transcripts and increased 128 transcripts (Figure 2A, right circles and Table S1). This is consistent with previous studies in HeLa cells, which showed that Spt5 depletion affects only a very small subset of genes [42].
Strikingly, among this small subset of genes whose expression was affected by the depletion of either of the two DSIF components, only 20 and 11 transcripts were commonly down- or up-regulated, respectively, by at least 2-fold (Figure 2A). Several transcripts that show significant difference in the absence of either Spt4 or Spt5 were further quantified by RT-qPCR to confirm that the array result was valid (Figure 2B and Figure S1). These data indicate that Spt4 and Spt5 independently regulate the transcription of a limited number of genes.
Another difference found between Spt4 and Spt5 in the transcriptional regulation came from analysis of the cryptic transcripts that arise within the intronic S region using primer sets detecting intronic Sμ-Cμ and Sα-Cα sequences (Figure 2C, top). These transcripts are generally present at a very low level as compared to GLTs and initiated mainly from cryptic initiator elements (Stanlie, A. unpublished data). Interestingly, Spt4 knockdown dramatically increased the cryptic Sμ transcripts, whereas the effect of Spt5 knockdown is much less (Figure 2C, bottom). However, Spt4 or Spt5 depletion only slightly augmented transcripts initiated from the Sα intronic region; this indicates that there are distinct properties between the donor and acceptor loci. Sμ cryptic transcript suppression is therefore critically dependent on Spt4 but much less on Spt5, further confirming the differential roles of Spt4 and Spt5 in transcriptional regulation.
To analyze the involvement of Spt4 or Spt5 in AID-induced S region DNA cleavage, we conducted two independent assays – ChIP assay of γH2AX focus formation, and in situ DNA-end labeling with biotinylated-dUTP to directly measure cleaved DNA ends. Spt4 knockdown did not reduce, but modestly enhanced, γH2AX level and biotinylated-dUTP-labeled DNA fragments in both the Sμ and Sα regions (Figure 3A, 3B). On the other hand, depleting Spt5 significantly reduced both γH2AX foci and biotinylated-dUTP-labeled DNA fragments in the Sα, but not Sμ region (Figure 3A, 3B). These results indicate that Spt4 is dispensable for the DNA cleavage step in both the Sμ and Sα regions, whereas Spt5 is especially critical for introducing DNA breaks in the acceptor Sα region.
Our previous studies of the FACT complex revealed that S region chromatin modifications, particularly H3K4me3, play a critical role in generating AID-induced DNA breaks [26]. Spt5 has been reported to regulate H3K4me3 though the trans-histone modification pathway [36]. To further delineate the importance of Spt4 and Spt5 in H3K4me3 maintenance, we performed ChIP assays using anti-H3K4me3 antibody in the absence of either Spt4 or Spt5. Spt4 knockdown slightly augmented the H3K4me3 modification in both the Sμ and Sα regions as compared to the control, whereas Spt5 depletion drastically reduced the H3K4me3 formation specifically in the Sα but not Sμ region (Figure 3C). H3K4me3 modification levels thus correlated well with DNA breakage efficiency. On the other hand, neither Spt4 nor Spt5 depletion altered the core histone H3 in the Sμ and Sα regions (Figure 3C), suggesting that H3K4me3 depletion in the absence of Spt5 is not ascribed to histone H3 loss, but rather attributed to the inefficiency of the histone modification itself.
We also analyzed the total H3K4me3 and H3 by immunoblotting and found that depletion of Spt4 or Spt5 did not significantly change the cellular levels of these histones (Figure 3D), indicating that the H3K4me3 loss in the absence of Spt5 was not due to a global event, but rather limited to specific loci. These results are consistent with our previous conclusion that the presence of H3K4me3 is a critical determinant for the introduction of S region DNA cleavage in CSR [26]. Furthermore, the present results indicate that distinct transcription factors are required by different S regions to regulate chromatin status and the eventual DNA break. It is striking that Spt4 or Spt5 depletion has opposite effects on both DNA cleavage and H3K4me3 formation at the Sα region.
Since the absence of Spt4 inhibited CSR but did not reduce the DNA breakage in either donor or acceptor S regions, we suspected a possible CSR repair-phase defect. To investigate possible defects in S region DNA repair in the absence of Spt4 or Spt5, we performed ChIP assays in CH12F3-2A cells using an antibody against Ku80, a protein implicated in the initial phase of NHEJ DNA repair [5]. CIT stimulation enhanced Ku80 accumulation in both the Sμ and Sα regions, while Ku80 was minimally detectable in the Cμ and Cα region (Figure 4A). Depleting either Spt4 or Spt5 did not significantly affect the Ku80 accumulations in the Sμ region, which confirms our observation that DNA cleavage is still detected in the absence of either of the two components. In contrast, depletion of Spt4 or Spt5 was associated with the reduction of Ku80 accumulation in the Sα region. While this result is consistent with the reduction of Sα DNA cleavage by Spt5 knockdown, it was somewhat unexpected because Spt4 knockdown still gave rise to robust DNA breakage. This finding led us to suspect that Spt4 may be involved in the DNA repair phase of CSR.
To confirm that Spt4 is really required for DNA repair but not cleavage in CSR, we used an artificial NHEJ substrate construct to directly assay Spt4 and Spt5 function in DNA repair as CSR utilizes the general DNA repair pathway [5]. The construct can be cleaved by I-SceI endonuclease at two I-SceI sites flanking a TK cassette, which is actively transcribed under the control of the pCMV promoter (Figure 4B). The joining of two DNA ends with minimal homology sequences brings a GFP cassette close to the promoter, thus initiating its expression. Repair efficiency can be measured by monitoring the GFP-positive cell population by FACS [43].
The introduction of I-SceI expression plasmids into H1299dA3-1 human lung cancer cells carrying this NHEJ construct induced GFP expression (Figure 4C and 4D). The absence of Spt4, Spt5 or both drastically reduced the GFP-positive cell population as compared to the control level (Figure 4C and 4D). RT-qPCR analysis confirmed that the Spt4 and Spt5 gene knockdown was efficient (Figure 4E).
To ascertain that the reduction in GFP-positive cells was not due to transcription inhibition in the pCMV reporter construct, GFP-positive cells expressing I-SceI (indicating they had accomplished NHEJ) were subjected to Spt4, Spt5 or double knockdown. As shown in Figure 4F, the GFP expression was unaltered in the absence of Spt4 and/or Spt5, suggesting that Spt4 and Spt5 do not regulate the GFP reporter's transcription efficiency. Moreover, direct PCR measurement of the ligated DNA ends revealed that the junction signal intensity was reduced in samples treated with Spt4 and/or Spt5 RNAi oligonucleotides (Figure 4G). However, in addition to the major band expected, bands with slightly higher or lower molecular weights were also present, specifically in samples depleted of Spt4 or Spt5 (Figure 4H). These bands were excised and sequenced to identify their origins. While the majority of the PCR products corresponded to the appropriately repaired band, we also obtained clones with long insertions and extensive resections in the products derived from Spt4 or Spt5 knockdown samples (Figure S2), suggesting that the NHEJ repair was compromised. These results revealed that both Spt4 and Spt5 are involved in NHEJ repair of the transcribed locus.
We next examined whether Spt4 or Spt5 plays a role in the repair by homologous recombination, again using a synthetic recombination substrate (Figure 5A). This construct is composed of two tandem GFP genes with different mutations: the upstream SceGFP contains the I-SceI recognition site and two in-frame stop codons, and the downstream internal GFP (iGFP) has 5′ and 3′ deletions. I-SceI expression introduces DSBs in SceGFP and triggers repair by gene conversion, using the downstream iGFP as a template [44], [45]. Successful repair can be monitored by GFP expression. The introduction of I-SceI-expression plasmids into GM7166VA cells (derived from a NBS patient) carrying the recombination assay GFP construct robustly induced GFP, which we assayed by FACS (Figure 5B and 5C). However, the GFP-positive population was significantly reduced in the absence of Spt4, Spt5 or both (Figure 5B). RT-qPCR analysis confirmed the gene knockdown efficiency (Figure 5D). We also confirmed that the absence of Spt4 and/or Spt5 did not negatively regulate the transcriptional status of the repaired GFP reporter substrate, using a method similar to the above-mentioned NHEJ assay (Figure 5E). The identical repair assay system conducted in CH12F3-2A cell line gave a similar effect by Spt4, Spt5 or double knockdown (Figure S3A and S3B). Moreover, in the absence of either Spt4 or Spt5, CH12F3-2A cells were more sensitive to ionizing radiation (Figure S3C), whose repair grossly depends on NHEJ [46]. The result is consistent with studies conducted in yeast [47]. Taken together, our findings indicate that Spt4 and Spt5 have essential functions in the DNA repair phase of CSR.
In the present study, we showed that while both components of the DSIF complex, Spt4 and Spt5, are essential for Ig class switching, they have distinct roles in transcriptional regulation, histone PTM maintenance and S region DNA cleavage. Spt5 has been reported to regulate H3K4me3 through the trans-histone modification pathway initiated by H2B ubiquitination (ubH2B) [36], [37]. Interestingly, depletion of a ubiquitin ligase Bre1, which is required for ubH2B leading to efficient H3K4me3 formation, affects the H3K4me3 level specifically in the Sα region but not the Sμ region [26]. Consistent with this finding, we showed that Spt5 depletion strongly affected H3K4me3 and DNA breakage in the Sα region but had relatively little effects in the Sμ region. Spt4 knockdown, on the other hand, slightly augmented DNA breakage in both the Sμ and Sα regions, and slightly increased the H3K4me3 levels. Taken together, these findings indicate that Spt4 and Spt5 differentially regulate histone PTMs, most likely those associated with the ubH2B trans-histone modification cascade. Interestingly, we found that Spt5, but not Spt4 knockdown slightly reduced the AID protein amount without affecting its transcript (Figure S5). As Spt5 and AID have been shown to interact with each other [40], it is possible that Spt5 might directly regulate the stability of this mutator protein. However, since Sμ DNA break formation is unperturbed in the absence of Spt5, the slight depletion of AID expression per se is unlikely to be the cause of CSR blockade.
The results presented in this report further strengthen our previous conclusion that H3K4me3 in the S regions provides the histone mark for DNA cleavage sites in CSR [26]. It is intriguing that both VDJ and meiotic recombination also utilize the same histone modification for their respective DNA cleavages [27], [28], [29]. It should be stressed that H3K4me3 is required, but may not be sufficient for the cleavage target determination. An attractive explanation is that a combinatorial readout between H3K4me3 and other histone PTMs or variants, in conjunction with non-B DNA or other specific DNA structures, might be required to recruit a CSR recombinase (i.e. either Top1 or AID itself). In any case, our data support the idea that the presence of nucleosomes and their associated modifications are critical for DSB formation in vivo.
Interestingly, even though Spt4 and Spt5 are well-established transcription factors, their depletion in CH12F3-2A cells did not affect known critical transcripts and key players of CSR. Spt4 or Spt5 knockdown affected a relatively small subset of transcripts of mostly non-overlapping loci, causing either up- or down- regulation. Moreover, only Spt4 depletion strongly enhanced the Sμ cryptic transcripts, which are initiated from the intronic region; this is reminiscent of the pervasive transcripts observed in the yeast system [48]. It is therefore likely that Spt4 and Spt5 function separately in most of the actively transcribed loci, although they may form the DSIF complex for other functions or loci.
Our data point out distinctive features of the donor Sμ and acceptor Sα switch loci, as summarized in Table S2. Although this distinction has never been clearly defined, previous research has hinted at this possibility. Suv39h1, the methyltransferase responsible for H3K9 trimethylation, specifically promotes switching to IgA but not to other Ig isotypes [49]. Moreover, it was recently reported that a deficiency of the transcription factor Ikaros increases class switching to IgG2b and IgG2a, with a concomitant reduction in all other isotypes; this is achieved by modulating the chromatin status and the transcriptional competency of the γ2b or γ2a genes [50].
These results suggest that different S regions are distinctly regulated. Indeed, using the underlying sequences of Sμ and Sα to predict the nucleosomal occupancy of the loci, we observed a stark contrast in nucleosome distribution in the two regions, especially at the promoter; Iμ promoters have a clearly demarcated nucleosome-free region (NFR), while Iα promoters have no clear NFR boundary (Figure S4). This analysis correlates perfectly with our ChIP data of histone H3 (Figure 3C, compare position ‘a’ and ‘f’). The distinct nucleosomal occupancies, coupled with discrete promoters and chromosomal architecture, might eventually recruit different transcription factors and, in turn, histone PTMs, subsequently dictating the genes' regulatory strategy. It is therefore useful to think of different switch loci as independent entities, regulated distinctly and modulated by various transcription-associated processes.
We unexpectedly demonstrated the important roles for Spt4 and Spt5 in DNA repair, using assays with artificial constructs specific to NHEJ and homologous recombination. While Spt4 and Spt5 are not equally necessary for DNA cleavage, they seem to be similarly required for DNA repair; therefore, it is possible that Spt4 and Spt5 may function as a complex for efficient DNA repair. Since Spt4 depletion strongly reduces CSR but not S region cleavage, CSR inhibition in the absence of Spt4 is likely due to the inhibition of the repair phase.
Interestingly, earlier studies found that mutants of Spt4 and Spt5 resulted in methyl methanesulfonate (MMS) sensitivity in yeast, indicating a possible role in DNA repair and recombination [47], [51]. Consistent with this possibility, several repair factors, including BRCA1 and Ku80, have been reported to interact with the DSIF complex in human cells [47], [52]. Moreover, DSB repair proteins involved in NHEJ or homologous recombination such as Ku70/80, DNAPKcs, and RAD51 associate with the RNAPII complex [53]. The interaction between these repair factors and RNAPII transcription elongation machineries suggests that DNA breaks are repaired through transcription-coupled processes at some loci. Such mechanisms have been widely studied in the context of transcription-coupled repair (TCR), in which DNA damage leads to RNAPII arrest, followed by specific factors recruitment to the arrest site, and the lesions are removed by nucleotide excision repair (NER) [54]. Interestingly, deletion of the NER genes ERCC1-XPF is also reported to reduce CSR efficiency [55]. Therefore, the basic idea of TCR might be extended to transcription-coupled homologous recombination and NHEJ, in which efficient repair complexes may interact dynamically with various transcription-associated factors such as Spt4 and Spt5, possibly by modulating the chromatin or histone status. Ultimately, this could increase the stability and residence time of repair factors at the site of DNA damage. It is worth noting that the assay systems employed in our study involve DNA joining in the presence of active transcription, and that efficient CSR repair is likely to be coupled with the transcriptional activity.
CSR's repair phase uses both NHEJ and alternative end-joining systems [2], [3], [4], while homologous recombination is involved in Ig gene conversion. In addition, while AID's exact function in DNA cleavage step is still debated [56], AID is known to be involved in the post-cleavage step of CSR because C-terminally truncated AID mutants can cleave the S regions but cannot complete CSR [12], [13], [14]. If this is the case, it is most likely that AID plays a role in pairing the appropriate ends of the S regions in cis. This event requires bending the DNA to bring the two S regions close to each other [57]. It is interesting to note that similar cis pairing is required for gene conversion [58]. On the other hand, C-terminally truncated AID can still carry out c-myc-IgH translocation, suggesting that this particular mutation does not inhibit non-specific joining [12]. The present finding that Spt4 and Spt5 are required for NHEJ and homologous recombination suggests that these factors probably remain at the cleaved sites to recruit repair factors until the ends have been successfully joined in CSR.
A recent study by Pavri et al. showed that Spt5 is required for efficient CSR and associates with more than 9000 AID-targeted loci; these authors assumed that Spt5 is the factor that guides AID to its targets [40]. While our data essentially converge to the conclusion that Spt5 is indeed required for efficient switching, we wondered if this extremely large number of loci bound by Spt5 reflects AID's physiological relevance as a mutator. Such a conundrum is best exemplified by RAG1, which is directly involved in VDJ DNA cleavage and binds preferentially to recombination signal sequences (RSSs) [59]. In addition to these putative RSSs, both mouse and human genomes contain millions of cryptic RSSs that are recognized by RAG proteins [60], [61]. However, current evidence indicates that meaningful RAG-mediated cleavage can occur at some, but most likely not all, of these cryptic sequences [62], [63]. In other words, the propensity for a particular protein to bind to a particular DNA region does not always reflect the true essence of the physiological outcome. More importantly, the target loci of AID and Spt5, while expected to correlate well based on the guiding-factor model do not in fact show a good correlation; by comparing the top 50 targets from both lists only 9 loci are commonly targeted [40], [64]. Our data, on the other hand, suggest that transcription factors like FACT [26], Spt5 (current report) and Spt6 (Begum, N.A; unpublished data) are intimately associated with the S regions, acting primarily as chromatin landscape regulators that in turn promote efficient DNA cleavage. Spt4 and Spt5 are also critical for NHEJ, which is required for CSR.
Collectively, our current data suggest that the DSIF subunits Spt4 and Spt5 can function independently to modulate histone PTM, DNA breakage, and transcription. It remains to be examined whether this dissociation can be observed in contexts other than CSR. It is of note that DSIF's role, especially in RNAPII stalling, was previously studied primarily by focusing on Spt5. Therefore, it remains to be seen whether stalling can be regulated independently at various loci. Finally, since the CSR mechanism is intimately associated with DNA breakage and recombination, often leading to off-target mutations and translocations [65], it is worth investigating the dynamic interactions of various transcription elongation factors that regulate chromatin architecture at non-IgH loci.
CH12F3–2A cells expressing Bcl2 were cultured and stimulated to induce class switch, as previously described [41]. Cells were subjected to FACS analysis after 24 hours of CIT (CD40L, IL4, and TGFβ) stimulation. FITC-conjugated anti-IgM and PE-conjugated anti-IgA antibodies were used for surface IgM and IgA staining, respectively. To analyze IgG3 switching, cells were prepared by staining with biotinylated anti-IgG3 and allophycocyanin-labeled streptavidin. Dead cells were excluded by propidium iodide staining. All analyses were performed on a FACSCalibur (Becton Dickinson). Electroporation (Amaxa) was used for knockdown experiments and the transfection of various RNAi oligonucleotides (Invitrogen) into CH12F3-2A cells; the cells were cultured for 24 hours, stimulated by CIT, and further cultured for another 24 hours.
Total RNA was extracted from cells using TRIzol (Gibco BRL), cDNA was synthesized using Superscript II and Oligo (dT) Primer (Invitrogen), and the real-time PCR reaction was performed using SYBR Green Master Mix (Applied Biosystems) and specific primer pairs.
CH12F3-2A cells were lysed in 1× RIPA lysis buffer containing 10% glycerol and 1% Triton-X-100, and were subjected to immunoblotting following standard protocols.
ChIP assays were performed using ActiveMotif ChIP-IT Express Kit according to the manufacturer's instructions. In brief, 5×106 cells were fixed in the presence of 1% formaldehyde for 5 minutes at room temperature. Glycine was added to a final concentration of 0.125 M to stop the reaction. Cell lysis and sonication yielded a soluble chromatin fraction containing fragmented DNA of 200–500 bp. The lysate was immunoprecipitated by incubation with 2–3 µg of antibody. The pulled-down DNA was detected by real-time PCR normalized to the input. The maximal value in each data set was set as 100%, as described elsewhere [66].
DNA break assays were performed as described previously [12].
Lipofectamine 200 reagent (Invitrogen) was used to co-transfect I-SceI-expressing plasmid (pCBASce) and gene-specific RNAi oligonucleotides into either human lung cancer cells (H1299dA3-1) or cells derived from NBS patient repleted with full length human NBS1 (GM7166VA), carrying either NHEJ or homologous recombination artificial repair constructs, respectively (ref. 43, 44). Cells were incubated for 48 hours and analyzed by FACS. For homologous recombination repair analysis in CH12F3-2A cells, the linearized repair construct (linearized by Xho I restriction enzyme) was introduced into the cells by electroporation (Amaxa). Stably transfected colonies were picked up after 2 week of selection with puromycin. I-SceI-expressing plasmid was introduced to the cells 24 hour after RNAi introduction. Cells were further incubated for 48 hours and analyzed by FACS.
For the DNA microarray analysis, RNA samples were derived from either Spt4- or Spt5-knockdown CIT-stimulated CH12F3-2A cells. A 3D-Gene Mouse Oligo chip 24k (Toray Industries Inc., Tokyo, Japan) was used (23,522 distinct genes). For efficient hybridization, this microarray has 3 dimensions; that is, it is constructed with a well as the space between the probes and cylinder-stems, with 70-mer oligonucleotide probes on the top. Total RNA was labeled with Cy5 using the Amino Allyl MessageAMP II aRNA Amplification Kit (Applied Biosystems, CA, U.S.A.). The Cy5-labeled aRNA was hybridized for 16 hours using the supplier's protocol (www.3d-gene.com). Hybridization signals were scanned using a ScanArray Express Scanner (PerkinElmer) and were processed by GenePixPro version 5.0 (Molecular Devices). The raw data of each spot was normalized by subtracting the mean intensity of the background signal (determined by all the blank spots' signal intensities with 95% confidence intervals). Raw data intensities greater than 2 standard deviations (SD) of the background signal intensity were considered valid. The signals detected for each gene were subjected to global normalization (the median of the detected signal intensity was adjusted to 30).
For measurement of ionizing radiation sensitivity, CH12F3-2A cells that have been subjected to knockdown for 24-hour period were irradiated with the indicated doses of γ-ray. Cell survival was measured after 2 days by PI staining.
The expression array data derived from Spt4 and Spt5 knockdown samples are deposited in GEO under accession number GSE33206.
Information about the antibodies, RNAi oligonucleotides, and primers used are available in Table S3.
|
10.1371/journal.pbio.1000584 | Clusters of Temporal Discordances Reveal Distinct Embryonic Patterning Mechanisms in Drosophila and Anopheles | Evolutionary innovations can be driven by spatial and temporal changes in gene expression. Several such differences have been documented in the embryos of lower and higher Diptera. One example is the reduction of the ancient extraembryonic envelope composed of amnion and serosa as seen in mosquitoes to the single amnioserosa of fruit flies. We used transcriptional datasets collected during the embryonic development of the fruit fly, Drosophila melanogaster, and the malaria mosquito, Anopheles gambiae, to search for whole-genome changes in gene expression underlying differences in their respective embryonic morphologies. We found that many orthologous gene pairs could be clustered based on the presence of coincident discordances in their temporal expression profiles. One such cluster contained genes expressed specifically in the mosquito serosa. As shown previously, this cluster is redeployed later in development at the time of cuticle synthesis. In addition, there is a striking difference in the temporal expression of a subset of maternal genes. Specifically, maternal transcripts that exhibit a sharp reduction at the time of the maternal-zygotic transition in Drosophila display sustained expression in the Anopheles embryo. We propose that gene clustering by local temporal discordance can be used for the de novo identification of the gene batteries underlying morphological diversity.
| Linking genotype to phenotype is a major undertaking in modern biological research. A variety of strategies are used but have generally failed to explain the maintenance and acquisition of new phenotypic traits in changing populations. We propose that whole-genome cross-species comparisons can be used to identify gene clusters underlying phenotypic variation. In the present study we used gene expression datasets collected during fruit fly and mosquito embryogenesis to identify temporal changes in gene expression. We found that differentially represented tissue types (such as extraembryonic serosa) were clearly manifested by clusters of local discordances in gene expression. Discordances were also observed for a suite of maternally expressed genes, consistent with the notion that the abrupt maternal-zygotic transition seen in Drosophila is an evolutionary innovation of higher Diptera. We propose that gene clustering by expression discordance can be used to determine the genetic basis of phenotypic variation.
| During the 1980s and 1990s methods of molecular genetics were used to determine the contributions of individual genes to different developmental processes, such as the segmentation of the Drosophila embryo [1]. However, during the past decade post-genome technologies have opened the door to identifying all of the genes engaged in such processes. In particular, collections of microarray data accumulated in public databases now cover a variety of different conditions and sometimes even the full life cycles for a range of evolutionarily distant species. These data provide new opportunities to identify complete ensembles of genes engaged in the specification of body plans and morphological diversification. Recently a number of such studies in different model systems have been conducted [2]–[9]. The initial reports as exemplified by [7], where the comparative analysis was extended across different phyla, provide evidence for the existence of deeply conserved co-regulated gene sets (kernels [10]) responsible for fundamental cellular functions. More recent studies focused on a set of fairly proximal yeast species have revealed that the regulation of even the most essential processes such as the cell cycle may not be conserved [2]–[4]. It was found instead that temporarily similar engagement of multi-component biological machines could be achieved by the species-specific regulation of different subunits within these complexes [11],[12]. Also, the regulation of genes, which are differentially regulated between species, was shown to be primarily driven by TATA-box containing promoters [5].
The aforementioned studies were done by “co-expression meta-analysis” extensively reviewed in [13] and in [14]. This method employs gene lists independently precompiled by condition-dependent clustering in individual species. The benefit of this approach is its ability to compare gene expression independently of species-specific experimental conditions. In contrast, “expression meta-analysis” [13] employs direct comparisons of the expression profiles of orthologue pairs. This approach can specifically pick up a condition or a time point when orthologous genes are differentially regulated between species. However, its use is restricted to a set of similar conditions, such as a time-series. This method requires data preprocessing, or in the case of temporal data, matching the corresponding time points, which due to differences in metabolism may not relate directly between the species. The computational frameworks as well as data sources for direct cross-species gene expression studies have only recently become available.
In a recent study we used datasets spanning several cell cycles of synchronized culture of fission and baker's yeast as a comparative training data source to develop a computational platform for expression meta-analysis [15]. Using so-called “time warping” [16]–[20] enhanced by noise suppression [15], we created alignment paths and successfully predicted the comparative duration of the cell cycle phases in baker's and fission yeast. In the present study we have employed the time warping method to compare mosquito and fruitfly embryos.
Thus far, the phylogenetic comparisons of gene expression have identified differential expression of orthologous genes implicated in similar processes. However, most such comparisons have been restricted to unicellular organisms [2]–[4] or to closely related multicellular species that manifest relatively few phenotypic differences. Moreover, such studies have examined excessively divergent species (yeast versus humans) [7], thereby complicating the association of discordant gene expression profiles with phenotypic variation.
We have selected divergent species that are nonetheless members of a common order of insects, the Diptera. We reasoned that mosquitoes and fruitflies possess a number of distinctive phenotypic traits, but are sufficiently similar to link such differences with orthologous clusters of discordantly expressed genes. These two flies belong to separate suborders (Nematocera and Brachycera) of Diptera and are thought to be separated by an evolutionary distance of ∼200 million years [21]. Mosquito and fruitfly larvae possess a number of striking differences, such as the representation of larval head, which is involuted in Drosophila but fully extended in mosquitoes. Another prominent feature, which was recently examined at the level of gene expression by our group [22],[23], is the presence in the mosquito embryo of a double-layered extraembryonic envelope (amnion and serosa), which is substituted by single amnioserosa in fruitflies (reviewed in [24]). We were interested in whether these and additional features could be detected by the comparative analysis of gene expression datasets.
The duration of embryogenesis in flies and mosquitoes differs significantly. At 25°C it takes ∼22 h for the fruit fly embryo and ∼50 h for the mosquito embryo to fully develop. For this reason, the temporal gene expression datasets for fly [25] and mosquito embryogenesis [22] were first aligned using time warping [15] and then analyzed for discordant gene expression. Since all the distinct developmental events taking place in the course of embryogenesis are expected to have specific timing and duration, the aim of the analysis was to cluster the orthologous gene pairs based on timing and duration of local discordances in their temporal expression profiles.
These studies reveal that a major gene cluster matches those known to be involved in the function of the mosquito serosa. This cluster is shown to be re-engaged later in development during cuticle synthesis. Significant discordances were also observed for a number of maternally expressed genes in flies and mosquitoes, consistent with the evolution of a sharp maternal-to-zygotic transition in gene expression higher Diptera. These studies provide a framework for the identification of the genetic circuits underlying embryonic diversity.
Previous time-lapse microscopy [22] suggests that there is a simple linear correspondence in the embryonic development of fruitflies and mosquitoes, despite a ∼2-fold difference in duration (∼50 h at 25°C for Anopheles versus 24 h for the fruitfly). We used the time warping algorithm [15] to compare the Anopheles temporal microarray datasets and the available Drosophila dataset [25]. The term “time warping” is generally used to describe a set of computational procedures that allow matching the similar regions of numerical data, corresponding to the processes occurring at different time scales. Anopheles and Drosophila genomes are represented in genome databases by accordingly 12,604 and 13,781 annotated genes, most of which are represented on the corresponding microarray platforms (10,873 Anopheles genes and 13,056 Drosophila genes). Ortholog mapping (8,126 Drosophila genes matched to 8,047 Anopheles genes resulting in 10,708 orthologue pairs in ENSEMBL database) and ANOVA filtering (removing the genes showing no change in gene expression across the developmental time course) limited the total amount of data available for the time-warping and comparative gene expression analysis to 4,839 profile pairs, nearly 40% of all protein coding genes in the Drosophila genome (see Methods).
Alignments of the Drosophila and Anopheles datasets suggest a near linear correspondence in embryonic development, despite the different rates of development (see Figure 1A,B and UCB web resource). There is a high level of concordance in the gene expression profiles: 1,172 of 4,839 profiles (24%) display a strong correlation (r>0.9), another 1,757 profiles (36%) exhibit a good correlation (0.9>r>0.6), and 744 pairs (15%) show a moderate correlation (0.6>r>0.3). Thus, ∼75% of the orthologous gene pairs exhibit very similar temporal profiles of expression in the divergent fly and mosquito embryos. Clustering the 1,538 most concordant expression profiles (based on Drosophila melanogaster expression data) reveal a striking correspondence during development (see Figure 1C and UC Berkeley web resource). The large, almost rectangular, regions of similarity in the beginning and in the end of the time courses but not in the middle (see the heatmap on Figure 1A,B) suggest that both organisms use related gene repertoires during late stages of embryogenesis that are considerably different from those used at earlier stages.
We expected that the developmental events, which differ between the two species, would correlate with the discordant expression of numerous genes rather than just one or two. Moreover, we did not expect the individual members of such gene batteries to have similar expression profiles even within the same organism (Figure 2C), in contrast with previous models that assume similar expression of related genes [25]. To maximize the chance of identifying genes manifesting local as opposed to global differences in gene regulation, a scoring scheme was used that maximized discordance inside and the similarity outside the sliding window across the temporal axis of the comparative gene expression dataset (for details see Methods). The major parameters of this function are time (window position), duration of developmental event (window size), and the score (reflecting the amplitude of discordance). These parameters are sufficient to explicitly describe the pattern of discordance (Figure 2A).
For each relative time point and window length a list of genes was obtained with discordance scores above a chosen threshold. In this way, a comparison of the overall gene expression profiles between the two organisms could be represented as a two-dimensional heatmap, whereby the x-axis shows the relative developmental time and the y-axis indicates the window length and the color corresponds to the number of the genes with a similar discordance pattern (Figure 2B). The map patterns reveal territories (gene clusters) corresponding to groups of discordant genes, which exhibit variability at a specific time-point in development. This map could be constructed for any arbitrarily chosen discordance cutoff. We therefore went further to check whether these clusters might represent evolutionary innovations (e.g., distinct morphological structures) involving the coincident deployment of a large set of genes in one of the organisms.
We used the dissimilarity in the structure of extraembryonic membranes of Anopheles and Drosophila, investigated in our prior study [22], namely the absence of the serosa in the fruitfly embryo, as a model of differentially represented trait/organ/tissue type. To determine whether any of the major hot spots (gene clusters) on the discordance heatmap corresponds to serosa, we built the maps selectively for the serosal genes (249 genes expressed higher in serosa than in the embryo proper with Log2Fold >0.7; see [22]) (Figure 3C) as well as for the whole dataset with (Figure 3A) or without the serosal genes (Figure 3B). Taken separately, the serosal genes manifest a prominent cluster at the position corresponding approximately to 12 h of mosquito development and 6 h of fruitfly development. While the general discordance map pattern remained the same, the depletion of the serosal genes from the dataset resulted in the disappearance of this cluster (cluster 15, see the UCB web resource) from the map (Figure 3B). On the example of the serosal cluster, we therefore conclude that the hotspots on the discordance heatmap may correspond to conspicuous or cryptic differences in anatomy or regulatory programs employed during development.
We further used the serosal cluster to define the best discordance cutoff that would be applicable for a wide range of morphological differences. Specifically, we composed a training set of positives (serosal genes) and negatives (non-serosal genes) from the discordant genes that could be extracted from the cluster 15 region at discordance cutoff = 0 and built the Receiver Operator Characteristic (ROC; see Figure 3E), which describes the True Positive Rate (TPR) and False Positive Rate (FPR) of this training set within the gene sets extracted from the cluster 15 at different discordance cutoffs (0 to 4). Using the classic analysis of the ROC curve we identified the best discordance cutoff value, as the one, which corresponds to the point when the number of true positives among the genes that would augment the list, if the cutoff was further increased, is higher than the number of false negatives. Practically it is the point on the ROC curve when the first derivative of True Positive Rate as a function of False Positive Rate rises above 1 (TPR∼0.4, FPR∼0.2 matching the discordance cutoff 1.7; Figure 3E). The 1.7 cutoff represents a compromise between the sensitivity and the specificity of detection, resulting in 30% of serosal within the 102 genes of the cluster. Yet at higher cutoffs (2.9) further increase in specificity (8 serosal out of 12 total genes) despite a significant loss in sensitivity could be achieved (Figure 3D). The clusters extracted from the discordance heatmaps can be further processed by standard genomics tools, such as GO enrichment analysis (see below).
The successful identification of the serosal genes led us to analyze additional discordant gene clusters. There are about 23 discordant hotspots (Figure 2B). Functional assignment of the newly identified clusters was investigated using FlyBase GO terms and controlled vocabulary annotations from the BDGP in situ database (see UCB web resource for annotations).
One of the discordant clusters (clusters 12–13, see Figure 2B and UCB web resource) corresponds to a set of maternal genes in both flies and mosquitoes. In flies, these genes exhibit a sharp reduction in expression during the maternal to zygotic transition but display continuous expression in mosquito embryos (Figure 4A). RNA in situ hybridization assays (Figure 4E,F, and G,H) show that these genes are ubiquitously expressed. In Drosophila these transcripts are rapidly lost at the onset of gastrulation, while in Anopheles they persist throughout the periods of gastrulation and germband elongation without significant changes in levels (Figure 4E,F, and G,H top versus bottom panels). Removing the maternal gene battery from the comparative datasets revealed that it indeed corresponds to clusters 12 and 13 on the discordance heatmap (Figure 4C,D).
Another unique cluster (cluster 8, see Figure 2B) was enriched for genes expressed in yolk in Drosophila. Among the genes expressed in the yolk we noticed another example of coordinate difference in gene expression. A number of metabolic genes such as CG9232 (galactose metabolic process) and others expressed in Drosophila yolk at mid-embryogenesis show maternal expression in mosquito, suggesting that the dynamics of yolk metabolism differs dramatically between the two branches of Diptera represented by these insects (Figure 5A–H).
This study provides a conceptual and computational framework for cross-organismal temporal alignments and comparisons of species-specific transcriptional datasets. The resulting heatmaps of discordant gene clusters identified distinctive patterning properties in Dipteran embryos, including the organization of extra-embryonic tissues and the maternal-zygotic transition (see Figure 2B and UCB web resource). There is surprising concordance in the development of Drosophila and Anopheles embryos, despite different rates of growth and distinctive patterning features. The heatmaps and alignments of gene expression datasets (representing more than 4,000 orthologous gene pairs) obtained with the time warping algorithm suggest a near linear relationship in their embryonic development. Interestingly the alignment path manifested a subtle deviation from linearity at stage 9, presumably shortly after serosa completion in Anopheles (Figure 1B). The removal of serosal genes from the datasets did not result in change in the automatic alignment path, suggesting that this time point may indeed represent temporarily local slow down of gene expression programs in Anopheles embryo proper compared to Drosophila.
The heatmaps obtained with the discordance mining algorithms implemented in Peak-mapper and Glob-mapper identify clusters of discordant genes that represent either unique tissue types (such as serosa) or changes in regulatory pathways such as that governing the turnover of maternal transcripts. The percentage of false positive genes (as judged by the analysis of the serosa gene cluster) was expected to be relatively high. One approach to this problem is the application of annotation methods such as GO-term enrichment or controlled vocabulary annotations (see UCB web resource).
“Gene sharing” or “co-option” manifested in redeployment of gene batteries in a diverse set of tissues is a recurrent theme in the evolution of animal embryos [26]. One of the mechanisms for gene co-option, specifically when it happens in the absence of gene duplication, is the diversification of regulatory regions resulting in acquisition of new territories of expression. For example, the skeletogenic mesoderm of the sea urchin embryo employs a conserved gene battery that is used for the secretion of the adult exoskeleton [27],[28]. Previous studies showed that the serosa gene battery is reengaged later in development, during the time of embryonic cuticle production [22]. The fact that an identical gene set is used twice during development at different time points or in different tissues suggests that the genes don't fall within this set due to random co-occurrence. In fact this reengagement is observed due to similarities of the cuticle secreted by the serosa at early stages [29]–[32] and the embryonic cuticle produced at the end of embryogenesis. We propose that the incidence of gene battery reengagement might be used as a further indicator of functional relevance of the gene within the battery or otherwise a filter that could remove the false positives from the genes extracted from the discordance cluster. Indeed, serosa genes can be identified by their biphasic expression in the mosquito as compared with the single late peak of expression in the fruitfly embryo (Figure 4B). Perhaps the low threshold discordance clustering in combination with verification of the reengagement patterns can be used as an alternative strategy for the extraction of functional gene sets from comparative datasets.
Previous whole-genome comparisons focused on ancestral gene networks based on phylogenetic conservation of gene expression patterns. Here, evidence was presented that discordant clusters provide a means for identifying gene batteries involved in evolutionary diversification and novelty. This approach was used to identify de novo the changes in the turnover of maternal transcripts in the Drosophila and Anopheles embryos and in temporal expression of a subset of yolk genes. The rapid turnover of maternal genes correlates with increased rates of embryogenesis in higher Dipterans.
Anopheles gambiae population was reared at 27°C, 75% humidity, with a 12-h light/dark cycle. Adults were maintained on a 10% sucrose solution and females were blood-fed on anesthetized hamsters. For synchronized embryo collection the females were placed in the dark at 27°C for 1 h inside a 15 cm petri dish lined up with circles of wet Whatman paper. The developmental time was counted starting from the moment the Whatman paper was moisturized with water. Mosquito embryo fertilization happens at the moment of egg laying. Due to constraints in the experimental setup, after 2 h following the start of collection, the eggs were shifted from 27°C and constantly kept at 25°C.
Drosophila-Anopheles ortholog pairs were downloaded using BIOMART interface of ENSEMBL website. Drosophila developmental microarray data-course was obtained from [25]. Mosquito developmental data-course (http://www.ncbi.nlm.nih.gov/geo/, accession number GSE15001) was taken from our prior studies [22]. Low level microarray data treatment involved standard quantile normalization of microarray data in log2 space [33] and Z-score normalization of time points (Gene_expr_value-mean(across development))/stdev(across development)) [33]. Both Drosophila and Anopheles datasets were filtered based on variation between biological triplicate's points using standard ANOVA analysis with very mild thresholds (p<0.05). In cases when several probesets were corresponding to a single gene, a probeset with a better (smaller) ANOVA p value (i.e., more consistent change within replicates and higher change across the time points) was chosen. In addition, the Drosophila expression array data [25] was filtered based on correlation with the Drosophila tiling array data [34] with mild thresholds as well (r>0.25). Finally, the expression profiles were superimposed based on the table of orthologs taken from ENSEMBL. Profiles corresponding to genes with multiple orthologs (paralogues) were multiplied, where necessary. These procedures produced 4,839 profile pairs, corresponding to 4,072 unique Drosophila genes.
Global alignment was constructed based on Kruskal-Liberman time warping algorithm [16],[35] as described in our previous study [15]. In brief, using the RZ-smooth filter from our software package (see web resource), the microarray data were resampled to 100 points in each dataset and smoothened using Gaussian function with a standard deviation corresponding to ∼2 original time points (15 relative points in resampled datasets). To compensate for the impact of terminal regions, after the resampling both datasets were truncated for less than 1 original time point at the beginning and the end. Similarity matrices and the corresponding global alignment path based on resampled and truncated datasets were constructed using time-warping algorithm. Specifically, local uncentered Pearson correlation between two sliding windows of l (l can be any natural number, we used l = 20) points each was calculated. This method, which was introduced in [15], is less sensitive to the interspecific noise in the data and better captures the subtle similarities between the datasets. Following the alignment, datasets were smoothened once more (using half of the original smoothing window) to compensate for step-like patterns in alignment curve.
Concordant genes were identified in the aligned datasets based on the global uncentered Pearson correlation. For the identification of the discordant genes, for each orthologous pair of resampled and aligned expression profiles the sums of square differences F (between the matching points on profiles) were calculated for the time point i and window length l—inside (Fwin) and outside of the window (Fext). Conditional probabilities p were computed based on distribution of the F values in the datasets A and B (for all genes). α is a pseudocount, limiting the probability values (α = 0.01 was used in this study).(1)
For every orthologous gene pair the discordance score S(i, l) was calculated as a function of i (window position) and window length (duration) l.
For any single comparison two heatmaps representing upregulated and downregulated genes (with reference to expression in Drosophila) were built. Batteries of discordant genes were identified as “hot spots” in the (i, l) parameter space, showing high numbers of genes with scores exceeding a threshold (S>1.5) for a given set of parameters i, l. For the further analysis, such as evaluation of annotation enrichment in the local discordance clusters (hotspots), we extracted the genes from an arbitrarily defined region (i1–i2, l1–l2) on the map, corresponding to the hotspot.
Different stages of dataset processing as well as the detailed help for the programs can be downloaded from http://flydev.berkeley.edu/cgi-bin/GTEM/dmap_dm-ag/index_dmap.htm. In short, initial dataset parsing, specifically the resampling and smoothing, were accomplished by RZ-smooth. The similarity matrices and alignment paths were built by Time-warp. At the next step the datasets were matched by M-align and a database of discordances was built by Peak-mapper. Finally the threshold-dependent heatmaps were constructed by Glob-mapper, which as part of its interface also allowed extracting the content of the discordance clusters (hotspots).
The images of whole-mount in situ hybridizations of Drosophila embryos were taken from BDGP in situ database (http://www.fruitfly.org/cgi-bin/ex/insitu.pl). Mosquito embryos were collected and fixed as described previously [23]. The hybridization dig-labeled anti-sense RNA probes against specific A. gambiae genes were generated by RT-PCR amplification from embryonic RNA and reverse transcription. A 26 bp tail encoding the T7 RNA polymerase promoter (TAATACGACTCACTATAGGGAGA) was included on the 5′ side of the reverse primer.
|
10.1371/journal.pntd.0003974 | Determinants of Health Service Responsiveness in Community-Based Vector Surveillance for Chagas Disease in Guatemala, El Salvador, and Honduras | Central American countries face a major challenge in the control of Triatoma dimidiata, a widespread vector of Chagas disease that cannot be eliminated. The key to maintaining the risk of transmission of Trypanosoma cruzi at lowest levels is to sustain surveillance throughout endemic areas. Guatemala, El Salvador, and Honduras integrated community-based vector surveillance into local health systems. Community participation was effective in detection of the vector, but some health services had difficulty sustaining their response to reports of vectors from the population. To date, no research has investigated how best to maintain and reinforce health service responsiveness, especially in resource-limited settings.
We reviewed surveillance and response records of 12 health centers in Guatemala, El Salvador, and Honduras from 2008 to 2012 and analyzed the data in relation to the volume of reports of vector infestation, local geography, demography, human resources, managerial approach, and results of interviews with health workers. Health service responsiveness was defined as the percentage of households that reported vector infestation for which the local health service provided indoor residual spraying of insecticide or educational advice. Eight potential determinants of responsiveness were evaluated by linear and mixed-effects multi-linear regression. Health service responsiveness (overall 77.4%) was significantly associated with quarterly monitoring by departmental health offices. Other potential determinants of responsiveness were not found to be significant, partly because of short- and long-term strategies, such as temporary adjustments in manpower and redistribution of tasks among local participants in the effort.
Consistent monitoring within the local health system contributes to sustainability of health service responsiveness in community-based vector surveillance of Chagas disease. Even with limited resources, countries can improve health service responsiveness with thoughtful strategies and management practices in the local health systems.
| Elimination of domiciliated vectors led to a decreased prevalence of Chagas disease in parts of Latin America. In Central America, where the domiciliated vector Rhodnius prolixus has been almost eliminated, Triatoma dimidiata, which cannot be eliminated, continues to threaten the population in vast areas. To maintain the risk of transmission of Trypanosoma cruzi at lowest levels despite limited resources, Guatemala, El Salvador, and Honduras integrated community-based vector surveillance into local health systems. One challenge to sustaining surveillance is to ensure continuous responsiveness to reports of household infestation from the community. Our research in 12 study areas in the three countries over a five-year period investigated eight potential determinants of health service responsiveness, including volume of vector notifications, local geography, demography, manpower, and managerial approach. We found that consistent (quarterly) monitoring by departmental personnel within the local health services was associated with high response rates. Results of interviews added additional insight.
| The prevalence of Chagas disease in Central America decreased from 1.7 million in the 1990s to 0.4 million in 2010 as a result of successful vector control [1, 2]. Of the two main vectors, Rhodnius prolixus is almost eliminated, but Triatoma dimidiata remains widespread in the region despite greatly reduced rates of household infestation [3–8]. To prevent transmission of Chagas disease resulting from re-infestation of houses by T. dimidiata in areas with limited resources, Guatemala, El Salvador, and Honduras implemented community-based surveillance, in which community members report the presence of bugs in houses to trigger a response by local health services of the Ministry of Health [5, 9, 10].
Community-based surveillance has been shown to be effective and cost-effective, but can be challenging to sustain [11–14]. Household infestation with vectors can be detected readily by inhabitants, and in turn, health services are expected to respond to every vector report by visiting houses to spray insecticide and provide educational advice [9, 11]. However, little is known about the extent to which vector reports from the community are met with appropriate responses and the factors that determine responsiveness of health services to vector reports.
Research on responsiveness of health services may provide insights that help sustain and strengthen vector surveillance throughout the region. We retrospectively analyzed health services’ response rates and underlying determinants in community-based vector surveillance of Chagas disease in Guatemala, El Salvador, and Honduras.
We selected 12 areas with community-based vector surveillance in Guatemala, El Salvador, and Honduras–four from each country (Fig 1). Each area was a conglomerate of villages and defined as being under the jurisdiction of a particular health center. For inclusion of an area in the study, the Ministry of Health had to have completed the attack phase in all villages by conducting multiple cycles of extensive insecticide spraying of at-risk houses to reduce household vector infestation, implemented community-based vector surveillance, and recorded data from 2008 to 2012. To compare management styles in unevenly decentralized health systems, we included one area per Department. The selected study areas were rural and in the most endemic districts of the Departments.
The study areas varied in population size (1,160 to 33,579 persons), geographic area (6 to 150km2), entomological situation, and human resources (Table 1 and Fig 2). The main target vector for surveillance in the study areas was Triatoma dimidiata, although in six areas (two in Guatemala and four in Honduras) surveillance also focused on Rhodnius prolixus because of previous history of infestation.
All 12 health centers had physicians, nurses, and operational technicians except San José de la Reunión in Honduras, which had no physicians, and Ojo de Agua in Guatemala, which had no nurses (Table 1). Operational technicians had different qualifications or responsibilities. Vector control was carried out by vector control specialists and occasionally assisted by unspecialized rural health technicians in Guatemala; and was jointly conducted by vector control specialists, health promoters, and environmental sanitation inspectors in El Salvador. In Honduras, environmental health technicians were responsible for food security, environmental sanitation, and zoonoses as well as vector control. Some technicians belonged to neighboring health centers or a departmental office, and covered the health centers through regular visits. Community health volunteers were present in all 12 health centers and insecticide sprayers were present in nine.
We defined a health service’s response rate as the percentage of the number of households sprayed or visited for advice by the local health services divided by the number of households infested with Chagas disease vectors as reported by the community. The annual response rate was calculated for each study area between 2008 and 2012, so that a total of 60 response rates (12 areas x 5 years) potentially were available for analysis. If the response occurred during the year following notification, it was considered as an action of the year of notification. A household with consecutive notifications of vector infestation was counted as a single infested household until the health service responded, regardless of interval length between notification and response.
Taking in account factors that might influence demand, supply, and work process in community-based vector surveillance, we selected for analysis the following eight variables as potential determinants of health service responsiveness: number of infested households as reported by the community; distance from health centers to departmental capitals; number of operational technicians per 1,000 households; numbers of community volunteers and sprayers per 1,000 households; interval between receipt of vector reports from the community and response by health services, i.e. <3 months, 3–12 months or >12 months; degree of decentralization of response to vector reports, i.e. by health center or departmental office; presence of consistent monitoring by departmental technicians; and presence of technical assistance by JICA.
We collected data on surveillance activities, local demography, geography, and human resources during visits to the departmental health offices and health centers during 2013. We interviewed personnel responsible for Chagas disease vector surveillance in each facility to identify any perceived factors or circumstances that might have influenced responsiveness during the five year period.
Communities reported a total of 2,630 households with T. dimidiata infestation in the 12 study areas between 2008 and 2012. Of these, the Ministry of Health responded to 2,041 households (response rate 77.6%, Table 2 and S1 Table). Of the 2,041 responses, 68.4% were by insecticide spraying and the reminder by providing education and advice.
Values of the eight variables that potentially influenced health service’s response rates differed among the health centers, but remained relatively constant within health centers over the 5-year period, with the exception of number of infested households reported, consistent monitoring by departmental technical officials, and technical assistance by the JICA project (Table 3). Numbers of health workers fluctuated according to trainees’ temporary assignments, and the population size of areas grew over time, but we treated these data as constant over the five year period.
Of the eight variables analyzed, two were found by linear regression and mixed-effects multi-linear regression to be significantly associated with health service responsiveness: consistent monitoring by departmental technicians and technical assistance by JICA (Table 4). In both regression analyses, consistent monitoring from the departmental level was correlated positively with health service responsiveness to a moderate degree (r = 0.48–0.55) whereas the correlation of assistance from JICA was weak and negative (r = -0.13).
Health centers in Dolores, Honduras and Comapa, Guatemala reported large numbers of infested households in 2009 and 2012, following campaigns in schools to promote bug searches as explained during interviews with health center staff (S1 Table).
Response rates followed four general patterns over the 5-year period: 1) nearly 100% response for most of the period, 2) nearly 100% for years 1 to 3 but then falling, 3) fluctuating moderately (between 50% and 100%), and 4) fluctuating substantially (between 0% and 100%) but with a tendency towards improvement.
When mixed-effects multi-linear regression was clustered by response pattern, similar associations between health services’ response rates and regular monitoring by departmental technicians (r = 0.71, p<0.01) and assistance from JICA (r = -0.15, p<0.01) were seen as in earlier models (S2 Table).
Interviews with health center personnel offered insight into the reasons underlying the different patterns (Table 5). Centers with higher response rates appeared to be more prepared to react to reports of infested houses; had better trained and more engaged workers; had superior management skills for coordinating and solving problems; and had greater support from higher institutional levels and local stakeholders such as community health volunteers and municipalities.
Interviews with the personnel of health centers and departmental health offices identified the persons responsible for different surveillance functions (Table 6). Bug detection was performed by the population in all study areas. Operational technicians or clinical staff of health centers were responsible for analysis, decision making and planning of response in 7 of the 12 study sites, while personnel at the Department level carried out this function in the other 5 areas (Table 6). Health promotion, bug reporting, and response to reports were conducted by distinct combinations of stakeholders in the different study areas. Overall Honduras recorded higher degrees of involvement by community personnel and clinical staff and lesser involvement by operational technicians than Guatemala and El Salvador (Fig 4).
We found that regular (quarterly) monitoring by departmental health offices was a significant determinant of health service responsiveness in community-based vector surveillance of Chagas disease in Guatemala, El Salvador, and Honduras. Perhaps surprisingly, response rates were significantly higher among health centers without presence of technical assistance by the donor (JICA). However, this finding can be explained by the presence of JICA at early stages of planning and implementation of the surveillance program at each study area, during which time response rates were low or fluctuated but subsequently improved. Three-year bilateral projects to establish community-based vector surveillance began in 2008 in El Salvador and Honduras and in 2009 in Guatemala.
Health service responsiveness was independent of the volume of bug reports from the community, distance between health centers and departmental offices, numbers of operational technicians in the local health service and community workers, intervals between vector report and institutional response, and degree of decentralization of response.
Interviews with health center staff demonstrated the effectiveness of regular monitoring on responsiveness and a decline in response rates following the departure of departmental supervisors in two health centers. This finding confirms previous research on primary health care services in low-resource settings, which showed that work performance was not motivated by written guidelines but by monitoring [15]. Because monitoring in this study provided an opportunity for departmental technicians and health center staff to review surveillance data, check equipment and supplies, participate in meetings with community health volunteers, and exchange information and experiences, continuation of quarterly visits should maintain or improve work performance over time. On the contrary, the consequences of inadequate monitoring can be serious in the long run, as reported in Gran Chaco in Argentina, where failure to supervise community personnel caused dysfunction of vector surveillance and reemergence of Chagas disease transmission [12].
Interviews also shed light into the lack of association between the other potential determinants and health service responsiveness. Although greater numbers of vector reports, for example following campaigns at school, increased the workload of local health services, response rates did not decline because manpower was augmented to meet the demand and tasks were reassigned among local stakeholders. Departmental technicians temporarily increased response capacity by mobilizing operational staff from other districts (as often occurs in reaction to dengue outbreaks) and by organizing extensive spraying operations with health center staff and community sprayers from different villages in the jurisdiction. Stakeholder analysis showed that surveillance tasks normally carried out by health specialists were simplified by the National Chagas Program and shifted to less specialized personnel through training, as we and others have reported previously [9, 16, 17]. In short, such combinations of short-term and long-term strategies reinforced responsiveness of health services.
Managerial responsibility for response at the departmental office rather than the health center did not appear to affect the response rate. Although the departmental response approach was more vertical and less integrated into primary health care services, interviews showed that both departmental health offices and health centers with high responsiveness were able to find solutions for difficult situations. For instance, departmental vector teams assigned a data collection technician to health centers, concentrated response efforts in time and space, and travelled by motorcycles to reduce transportation costs. A physician and a nurse at one health center posted a large map of the jurisdiction on a billboard in the waiting room and used thumbtacks to represent the number of households reporting vectors in each village and removed them following the appropriate response. Such strategies reinforced the management capacity of the local health services.
Longer intervals between receipt of vector reports and health service response did not lead to either higher response rates because of greater efficiency from economies of scale, or to lower response rates due to increased demands to deal with greater number of bug notifications. However, longer intervals are worrisome because of extended time of exposure of the population to the vector and thus greater risk of transmission of infection. Another potential negative impact is that the community may become reluctant to participate in bug notification if the interval is perceived as too long.
While portraying the reality of vector surveillance in Guatemala, El Salvador, and Honduras, this observational study has important limitations. The sensitivity of the analysis may have been affected by the limited number of infested households in certain areas and during specific years, and by lack of data at the individual household level, which would have detected repeatedly infested and responded households. Our resources were insufficient to measure outcomes such as household vector infestation rates and incidence and prevalence of Chagas disease. These data would enable analysis of the consequences of not achieving 100% response rate; the effect of spraying vs. educational advice; and the impact of variable quality of responses by specialized vs. lay workers. We were unable to conduct cost analyses that would allow us to compare the effectiveness of the different styles and approaches to integrated surveillance, which varied substantially among the 12 study areas [12]. Further research is needed to address these limitations as well as long-term effects of monitoring on community-based surveillance where stakeholders may be changing.
The greatest challenges to control of Chagas disease in Central America are non-eliminable, widespread vectors and underfunded and irregularly decentralized health systems. Although the disease has been targeted for elimination [18], a more realistic approach is to prepare for permanent control in the region [19]. The success of vector control efforts in reducing household infestation and disease prevalence have made vector bugs and patients less visible and made the interventions less likely to be prioritized for government budgets in the future. Prospects for external funding are not good, since international aid agencies are often attracted to health problems which are eliminable or reducible to a great extent in a short time. Thus, Chagas disease control strategies need to be extraordinarily cost-effective and sustainable, and intervention models should be simple enough to be readily integrated and monitored in local health systems at different stages of decentralization. Although in Guatemala, El Salvador, and Honduras community-based vector surveillance for Chagas disease is part of the local health systems and functions with existing human resources and minimum costs, reductions in budget could affect availability of transportation and insecticide, and consequently health service responsiveness.
In the control of non-eliminable vectors, such as T. dimidiata, the roles of continued spraying of infested houses and alternative interventions must be determined. In our study, 33.5% of responses to infested households was by insecticide spraying in Guatemala, versus 95.8% in El Salvador and 84.0% in Honduras. This partly reflects periodic scarcity of insecticides in the Guatemalan Ministry of Health, but also a deliberate shift towards house improvement. Multiple cycles of insecticide spraying are effective in reducing household infestation [20], but are costly and difficult to sustain in the long run. Moreover, continuous application of insecticide might promote emergence of resistance in vectors. On the other hand, risk factors such as cracked mud walls, dirt floors, thatched roofing, and improperly tiled roofing [21] can be mitigated using locally available materials [22, 23]. The cost-effective approach for improving house structures and living conditions innovated by Guatemalan researchers was adapted by the country’s Ministry of Health [22, 23]. Also, local operational technicians developed an effective community organization approach which promotes engagement by the population and local government, and efficient implementation and scale-up of the house improvement method [5]. Evaluation of these efforts should also be part of the future research agenda.
This research found that consistent monitoring at the departmental level of the Ministry of Health makes a significant difference in health service responsiveness in community-based vector surveillance of Chagas disease. Other potential factors, such as the number of infested households, numbers of health personnel and community workers, distance from departmental health offices to health centers, and degree of decentralization of response seemed to have limited impact on health service responsiveness. Challenges related to these factors were met largely because of managerial efforts of the local health services in implementing short-term and long-term strategies. Basic management practices such as monitoring and supervision combined with thoughtful strategies can improve health service responsiveness in resource-limited settings.
|
10.1371/journal.ppat.1007974 | Humoral immunity prevents clinical malaria during Plasmodium relapses without eliminating gametocytes | Plasmodium relapses are attributed to the activation of dormant liver-stage parasites and are responsible for a significant number of recurring malaria blood-stage infections. While characteristic of human infections caused by P. vivax and P. ovale, their relative contribution to malaria disease burden and transmission remains poorly understood. This is largely because it is difficult to identify ‘bona fide’ relapse infections due to ongoing transmission in most endemic areas. Here, we use the P. cynomolgi–rhesus macaque model of relapsing malaria to demonstrate that clinical immunity can form after a single sporozoite-initiated blood-stage infection and prevent illness during relapses and homologous reinfections. By integrating data from whole blood RNA-sequencing, flow cytometry, P. cynomolgi-specific ELISAs, and opsonic phagocytosis assays, we demonstrate that this immunity is associated with a rapid recall response by memory B cells that expand and produce anti-parasite IgG1 that can mediate parasite clearance of relapsing parasites. The reduction in parasitemia during relapses was mirrored by a reduction in the total number of circulating gametocytes, but importantly, the cumulative proportion of gametocytes increased during relapses. Overall, this study reveals that P. cynomolgi relapse infections can be clinically silent in macaques due to rapid memory B cell responses that help to clear asexual-stage parasites but still carry gametocytes.
| Plasmodium vivax contributes significantly to global malaria morbidity and remains a major obstacle for malaria elimination due to its ability to form dormant stages in the liver. These forms can become activated to cause relapsing blood-stage infections. Relapses remain poorly understood because it is difficult to verify whether P. vivax blood-stage infections in patients are due to new infections or relapses in most cases. Here, we use a nonhuman primate model of Plasmodium vivax malaria in concert with state-of-the-art immunological and molecular techniques to assess pathogenesis, host responses, and circulating gametocyte levels during relapses. We found that relapses were clinically silent compared to initial infections, and they were associated with a robust memory B cell response. This response resulted in the production of antibodies that were able to mediate clearance of asexual parasites. Despite this rapid immune protection, the sexual-stage gametocytes continued to circulate. Our study provides mechanistic insights into the host-parasite interface during Plasmodium relapse infections and demonstrates that clinically silent relapses can harbor gametocytes that may be infectious to mosquitoes.
| Due to their ability to establish dormant forms in the liver called hypnozoites, relapsing malaria parasites pose a significant obstacle to malaria elimination [1]. Hypnozoites can become activated, resulting in repeat blood-stage infections, and these relapses account for the majority of P. vivax blood-stage parasitemias [2, 3]. This raises critical questions about the relative importance of relapses in causing illness, and their importance in transmission. However, there are very few studies that have directly examined relapse biology, especially in the context of pathogenesis, host immunity, and transmission. An improved understanding of relapses and their immunological and epidemiological implications is needed to design effective control and elimination strategies for relapsing malaria parasites.
There are no rodent malaria parasites species that form hypnozoites, making these models inadequate for studying relapses. Human studies in endemic areas have limited utility because it is generally difficult to determine whether a blood-stage infection resulted from new, relapsing, or recrudescent infections [4, 5]. Although approaches such as parasite genotyping, relocation of individuals from P. vivax endemic areas to non-endemic areas, and mass drug administration provide more confidence that a P. vivax infection is due to a relapse, these approaches have caveats [6–10]. For example, new genotypes detected in sequential blood-stage infections could be due to the activation of hypnozoites that did not originally activate and circulate in the blood at the time of an initial sample collection; thus, these parasites could be mistaken as parasites from a new infection even if they originated from hypnozoites in the liver. Additionally, the inability to control for an individual’s infection history complicates the investigation of immune responses during human relapse infections.
Nonhuman primate (NHP) models lack these barriers and present several advantages for studying relapses. In particular, rhesus macaques infected with Plasmodium cynomolgi, a simian malaria parasite closely related to P. vivax, recapitulate the biological, clinical and pathological features of P. vivax malaria, including hypnozoite formation and relapses [11–15]. Further, P. cynomolgi is now recognized as a zoonotic species in South East Asia, making understanding its biology a priority [16–18]. These species have similar 48-hour intraerythrocytic developmental cycles, antigenic makeup, and infected red blood cell (iRBC) modifications that include abundant caveolae vesicle complexes [19–22]. Moreover, genomes and immunological tools are available to support the study of host-parasite interactions and NHP immune responses. Hence, these factors make the rhesus macaque–P. cynomolgi model valuable for the study of relapse biology.
We previously showed that P. cynomolgi relapses in rhesus macaques have substantially reduced parasitemias compared to primary infections, and they do not result in anemia or other clinically detectable disease manifestations [12]. The present study explored the host-pathogen interactions that underpin clinically silent P. cynomolgi relapses to garner insights into asymptomatic P. vivax relapses in humans [7, 23, 24]. We hypothesized that humoral immunity could explain the lack of clinical disease during relapses since passive transfer of antibodies has been shown to control parasitemia and ameliorate disease during human, NHP, and rodent Plasmodium infections [25–27]. Here, we demonstrate that lack of clinically detectable disease during P. cynomolgi relapses is associated with rapid memory B cell responses and the swift rise of anti-parasite IgG1 antibodies that can mediate clearance. The same humoral immune responses were also associated with protection against subsequent challenge with the same parasite strain about 60 days after radical cure. Interestingly, the immune response reduced the number of sexual stage gametocytes present compared to the primary infection, but the cumulative proportion of gametocytes increased during relapses. This suggests that the immune response generated by the infection primarily targeted the asexual stages. Concordantly, mature gametocyte gene expression was not significantly different between primary infections and relapses. Together, these data show that clinically silent, P. cynomolgi relapses carry gametocytes despite a significant reduction in parasitemia associated with an effective humoral immune response. Overall, this study broadens our understanding of relapsing malaria parasite pathogenesis and infections, with important epidemiological implications relevant to malaria elimination strategies.
Six rhesus macaques were inoculated intravenously with 2,000 P. cynomolgi M/B strain sporozoites on Day 0, and the parasitemia and clinical status of each individual monkey was evaluated daily for up to 100 days by light microscopy and complete blood count (CBC) analysis, respectively. Blood specimen collection time points are indicated in Fig 1A and 1B. A summary of the clinical and parasitological criteria for each specimen collection is provided in S1 Table.
The infections reached patency between days 10–12 (mean ± SE = 11.16 ± 0.3) after inoculation. Parasitemia peaked between 276,981–540,156 parasites per microliter of blood (mean ± SE = 400,563 ± 33,028 parasites/μl) between days 17–19 post-inoculation (Fig 1B). After collecting blood samples at the peak, all animals were administered a sub-curative dose of artemether to reduce but not eliminate the blood-stage parasites. This treatment ensured that the animals would remain parasitemic but not develop severe disease. Following the administration of curative blood-stage treatment, relapses were observed at different time intervals for each individual (Fig 1B). As anticipated based on earlier studies, relapse infections had approximately 200-fold lower parasitemia than the primary infections (p < 0.05; Fig 1B and 1C). Notably, relapse parasitemias declined below patency 7 to 15 days after their detection in the blood without treatment (Fig 1B).
The primary infection induced a myriad of clinical manifestations with elevated temperatures (mean ± SE = 102.3 ± 0.44°F) although not statistically significant and was accompanied by significant decreases in hemoglobin levels and platelet counts compared to when the animals were naïve (Fig 1D). Temperatures remained elevated post-peak for approximately one week after the administration of subcurative artemether treatment (Fig 1B and 1D). The anemia severity was moderate to severe with hemoglobin nadirs ranging from 5.7–9.0 g/dl (mean ± SE = 7.3 ± 0.4 g/dl) after the peak of parasitemia (S1A Fig). Platelet counts dropped from a mean of 421,250 platelets/μl when naïve to an average of 119,000 platelets/μl at the peak of the primary infection (Figs 1D and S1B). In stark contrast to the primary infections, anemia, thrombocytopenia, and fever were not observed in the animals during relapses (Fig 1D, S1A and S1B Fig).
Clinical severity of malaria has been correlated with inflammatory responses to infecting Plasmodium parasites [28–31]. Congruent with elevated temperatures and clinical presentation, inflammation was highest at the peak of parasitemia when 22 out of 45 cytokines, chemokines, or growth factors tested were significantly increased in the plasma compared to malaria naïve values (Figs 1E, 1F and S2). Pyrogenic cytokines such as IL-1β, TNFα, IL-6 and IFNγ were significantly increased in plasma at peak parasitemia when the animals were presenting with clinical illness (Fig 1F). Notably, only 6 out of 45 analytes remained significantly elevated after the administration of sub-curative artemether treatment (S2 Fig). In concordance with the clinical presentations, cytokine responses were subdued during relapses compared to the primary infection (Fig 1E). IL-1β, TNFα, IL-6 and IFNγ were not significantly elevated during relapses compared to pre-infection values, and Monokine Induced by Interferon Gamma (MIG) was the only cytokine significantly increased during relapses (Figs 1F and S2). The decrease in inflammation was likely due to the significant reduction in parasitemia that was observed between the initial infections and relapses (Fig 1B and 1C). Overall, these results reaffirmed that P. cynomolgi relapses had substantially reduced parasitemia and did not cause clinical signs of malaria.
To identify host responses that may reduce parasitemia and disease during relapses, we performed RNA-Seq on whole blood samples collected at the time points indicated in Fig 1A. The sequencing reads were mapped to concatenated host and parasite genomes, normalized by library size, log2 transformed, and finally, variance due to inter-individual variability was removed via the SNM transformation as previously described [32, 33].
Unsupervised hierarchical clustering of transcriptional profiles identified three major clusters as indicated by blue, purple, and yellow shading; these clusters captured approximately 30% of the variance associated with changes in gene expression (Fig 2A). Generally, the clinical presentation appears to drive the clustering pattern. The blue cluster consists of samples collected when the animals were not experiencing clinical symptoms, including prior to infection, post peak when the animals were recovering from illness, and during relapses (Fig 2A). In contrast, the yellow cluster consists of samples collected at the peak when clinical signs of malaria were evident, and the purple cluster consists of samples acquired after parasites were detected in the blood but prior to the onset of clinically detectable disease (Fig 2A). These results demonstrated that distinct changes in the host transcriptome occurred during the course of P. cynomolgi blood-stage infections. Samples collected as relapses were resolving (i.e. relapse resolution) formed a distinct subcluster compared to the early and peak relapse samples within the blue cluster (Fig 2A). Importantly, this subclustering showed that the transcriptional changes associated with resolving relapse infections were similar across individuals and distinct from the transcriptional changes during primary infections.
In agreement with the clustering pattern, differential gene expression analysis identified major transcriptional changes during the primary infection with more muted changes during relapses. Differentially expressed genes (DEGs) were determined via ANOVA followed by a t-test post-hoc analysis with Benjamini-Hochberg false discovery rate (FDR) correction. Genes with an FDR adjusted p-value of less than 0.05 were considered significantly differentially expressed. The DEG analysis was focused on the identification of genes that are differentially expressed between the malaria naïve time point and each subsequent time point (e.g. malaria naïve vs. peak, malaria naïve vs. relapse resolution, etc.). Compared to when the animals were naïve, the largest number of DEGs were identified during the peak of parasitemia, with approximately 3,350 and 4,713 DEGs upregulated and downregulated, respectively (Fig 2B). The pre-peak and post-peak time points induced comparable changes in host gene expression with over 2,000 genes upregulated and over 2,000 genes downregulated for each (Fig 2B). In contrast, relapses caused substantially less changes in host-gene expression compared to when the animals were naive (Fig 2B). Only 44 upregulated and 74 downregulated DEGs were identified during the early relapses, and 291 upregulated and 316 downregulated DEGs during the relapse peaks (Fig 2B). In contrast, the relapse resolution infection points had the most DEGs during relapses with 1,294 upregulated and 1,459 downregulated DEGs (Fig 2B).
Next, we focused on genes that were only differentially expressed during relapses. All upregulated and downregulated DEGs during the primary and relapse infections were compared. Six hundred and thirty-five upregulated and 607 downregulated DEGs were determined to be unique to relapses (Fig 2C and 2D). Metacore pathway enrichment analysis based on these upregulated and downregulated DEG sets revealed nine and over 50 significantly enriched pathways, respectively (S2 and S3 Tables). The nine upregulated pathways identified during the relapses are related to B cells, T cells, cell signaling, and antigen presentation (Fig 2C). The pathway with the highest enrichment score was related to B cells, and three out of the nine pathways are related to B cell responses (e.g., B cell responses in SLE, B cell signaling in hematological malignancies, and the B cell receptor pathway) (Fig 2C). The genes that were responsible for the enrichment of these pathways are composed of B cell surface proteins such as the BAFF-R, CD79A, and CD79B in addition to signaling molecules like Btk and VAV-2 (S2 Table). In contrast, the pathways related to T cells, signaling, and antigen presentation are composed of genes such as Akt, MAP kinases, CD86, T-bet, MHC-II, etc., which are expressed by a variety of immune cells, including B cells (S2 Table). The top downregulated DEGs during relapses belong to pathways related to inflammation, cell damage responses, and innate immune cells, such as dendritic cells, macrophages and neutrophils (Fig 2D and S3 Table).
Together, this analysis showed that relapses induced relatively minor, but unique changes in the host transcriptome compared to the primary infections that were predominantly related to the downregulation of pathways involved in innate immune responses and upregulation of pathways related to B cells.
Since the transcriptome data suggested that B cell responses may be involved in ameliorating disease during relapses, we next performed flow cytometry analysis on peripheral blood mononuclear cells (PBMCs) isolated during the primary and relapse infections. We utilized a B cell immunophenotyping strategy previously developed for human immunology studies and optimized it for specimens collected from rhesus macaques [34]. With this strategy, B cell subsets in PBMCs are CD19 and CD20 positive and further classified into four subsets based on the surface expression of IgD and CD27 (Figs 3A and S3). Similar to humans, four B cell subpopulations were identified in PBMCs from rhesus macaques: naïve (IgD+CD27-), unswitched memory (USM: IgD+CD27+), switched memory (SM: IgD-CD27+), and double-negative B cells (DN: IgD-CD27-). Surface IgM was present in all subsets, as previously shown for humans, albeit at low frequencies in the SM compartment (S3B and S3C Fig; [34, 35]). IgG surface staining was evident in the SM and DN populations but not in naïve B cells (S3B and S3C Fig). In summary, this immunophenotyping strategy yielded comparable results for samples acquired from either humans or rhesus macaques.
All B cell subsets decreased in the blood at pre-peak (Fig 3B). This decrease was consistent with the pan-lymphopenia observed during the initial infection (S4 Fig). At the peak, naïve and DN B cells stabilized in the periphery whereas SM and USM B cells remained significantly reduced compared to when the animals were naive (Fig 3B). Although USM and SM B cell numbers were reduced, the percentage of Ki67+ USM and SM B cells was significantly increased at the peak, suggesting these cells were activated (Fig 3C). Following subcurative treatment, USM B cells stabilized and SM B cells significantly increased from 263 ± 75 cells/μl (mean ± SE) when the animals were uninfected to 761 ± 242/μl (mean ± SE) post-peak (Fig 3B). During relapses, USM B cell numbers expanded from 153 ± 187/ μl (mean ± SE) when naïve to 686 ± 89/μl (mean ± SE), and SM B cells rose to 1982 ± 590 cells/μl (mean ± SE) (Fig 3B). The increase in the absolute numbers was accompanied by an increase in the frequency of Ki67+ USM and SM B cells, consistent with a new expansion of these cells during the relapses (Fig 3C). In contrast, the absolute numbers of naïve and DN B cells were unchanged (Fig 3B and 3C). Notably, there was an increase in frequency of Ki67+ DN B cells during relapses, although the absolute numbers of this population did not increase (Fig 3B and 3C).
Parasite-specific IgG and IgM are important for controlling parasitemia. Early during relapses IgG+ SM B cell frequencies were 330 ± 49/μl, which is similar to the naïve numbers (225 ± 71/μl), but these rapidly increased approximately 8-fold to 1,768 ± 567/μl as the relapse infections resolved (Fig 3D). Thus, the expansion in IgG+ SM B cell numbers was induced by the relapse. IgM+ SM B cells also increased in response to the relapses, but the absolute numbers of these cells were smaller in comparison to the IgG+ SM B cells (Fig 3D).
Finally, we evaluated if the changes in B cell subsets were correlated with parasitemia during primary and relapse infections. There was not a significant correlation between the number of naïve, DN, USM, and SM B cells and parasitemia during the primary infections (Figs 3E and 3F and S5). In contrast, USM (Spearman’s ρ = -0.57, p = 0.04) and SM (Spearman’s ρ = -0.57, p = 0.03) B cells were inversely correlated with parasitemia during relapses (Fig 3E and 3F). Notably, naïve and DN B cell numbers were not inversely correlated with parasitemia during relapses (S5 Fig). Overall, these data illustrate that USM and SM B cells dramatically expand during relapses in response to a new blood-stage parasitemia and may be involved in controlling parasitemia and ameliorating disease.
Since B cell responses and antibodies are critical for suppressing parasitemia, we determined whether anti-parasite IgM and IgG were increased during relapses using an ELISA with infected RBC (iRBC) and uninfected RBC (uRBC) lysates. Total IgM (tIgM) increased from 0.34 ± 0.02 mg/ml to 1.93 ± 0.18 mg/ml (mean ± SE) at the peak of the primary infections and remained elevated post-peak (S6A Fig). P. cynomolgi-specific IgM increased at the peak and remained increased post-peak (Fig 4A). Total and parasite-specific IgG also increased at post-peak (Figs 4B and S6B). The IgG subclass of the iRBC-specific IgG produced during the primary infections was IgG1 (S6C Fig). Neither IgM nor IgG were inversely correlated with parasitemia during the primary infections (Fig 4C and 4D).
Notably, we did not observe a difference between the reactivity of IgM that recognized iRBC versus uRBC lysates during the primary infections (S7A Fig). Similar to IgM, uRBC-specific IgG also increased during the primary infections, and a significant difference between IgG recognizing iRBC versus uRBC was also not discernable (S7B Fig). uRBC lysate-specific IgM and IgG were both inversely correlated with hemoglobin levels during the primary infections, suggesting that these antibody responses may be linked with the loss of uRBCs and the development of anemia during P. cynomolgi infections (S7C and S7D Fig) [36].
In contrast with the primary infections, relapses did not result in significant changes in tIgM levels (S6A Fig). However, there was an increase in iRBC-specific IgM at the relapse resolutions, but this increase was approximately half of what was detected during the primary infections (Fig 4A). In contrast, iRBC-specific IgG was rapidly produced during relapses and was significantly increased during the peak relapse and the relapse resolution periods (Fig 4B). These values are five-fold higher than those observed during the primary infections (Fig 4B). The increase in IgG occurred alongside the expansion of IgG+ SM B cells that peaked during relapse resolutions, strongly suggesting that this response was important for controlling parasitemia during a relapse (Figs 3D and 4B). As in the primary infections, the iRBC-specific IgG was IgG1 (S6C Fig).
Next, we determined if the iRBC specific IgM and IgG were inversely correlated with parasitemia during relapses. Notably, the early relapse time points for monkeys RBg14 and RIb13 were excluded from this analysis because these time points were taken one to two days before a relapse was patent. Including these time points would confound our analysis since the relapse resolution time points were also taken when the parasitemia was below patency. As expected, iRBC-specific IgM and IgG were inversely correlated with parasitemia during relapses (Fig 4C and 4D)
To determine the functionality of the antibodies generated during the primary and relapse infections with respect to clearance of iRBCs, we performed a phagocytosis assay using a THP-1 monocyte cell line, as previously described [37]. The percentage of THP-1 monocytes that phagocytosed iRBCs after being opsonized with heat-inactivated plasma increased from 5.1 ± 0.31% from naïve animals to 9.3 ± 1.1% and 10.8 ± 0.7% at the peak and post-peak of the initial infections, respectively (Fig 4E). During relapses, the percentage of THP-1 monocytes that phagocytosed iRBCs was even higher (16.0 ± 1.4%) (Fig 4E). Interestingly, the opsonization of iRBC by the heat-inactivated plasma was positively correlated with the amount of iRBC-specific IgG (Spearman’s ρ = 0.53, p = 0.01) and inversely correlated with iRBC-specific IgM (Spearman’s ρ = -0.65, p = 0.0017) (Fig 4F). Collectively, these data demonstrate that the anti-parasite antibodies produced during relapses can mediate iRBC clearance, consistent with a key role for IgG+ SM B cells in providing rapid host immunity to suppress parasitemia during relapses.
We questioned whether the animals would remain protected months later against a homologous parasite challenge infection and, if so, if the immune responses would be similar to those observed during a relapse. The same cohort of macaques was administered two rounds of radical cure to best ensure elimination of all liver- and blood-stage parasites. Approximately 60 days later they were re-challenged with 2,000 P. cynomolgi M/B strain sporozoites. Specimen collections were performed according to the schematic shown in Fig 5A and S4 Table. The reinfections with the homologous strain reached patency between days 9–12 post infection (mean ± SE = 11 ± 0.54 (Fig 5B). Peak parasitemias were substantially reduced compared to the initial infection and ranged from 269–5,742 parasites/μl (Fig 5B and 5C). There was no evidence that the decrease in parasitemia was directed against the sporozoite or liver stages since the number of days to patency was similar between the initial infections and homologous reinfections (S8 Fig). Like relapses, homologous reinfections did not cause clinical signs of malaria (Figs 5D and S9). Changes in cytokine profiles were also minimal with only IL-7 and RANTES differing significantly from pre-homologous values (Fig 5E). The pyrogenic cytokines IL-6, TNF-α, IL-1β, or IFN-γ did not increase during homologous reinfection and were not significantly different from values obtained during relapses (Fig 5F). Together, this homologous challenge experiment demonstrated that non-sterilizing immunity persisted for at least 60 days after radical cure and that this immunity could control peripheral parasitemia and the clinical manifestations of malaria in the P. cynomolgi model.
Since relapses and homologous reinfections had similar clinical presentations, we next employed RNA-Seq analysis on whole blood collected during the homologous challenges to determine if the host responses were similar. To identify DEGs during the homologous reinfections, the primary and post primary time points were compared to the pre-homologous challenge time point by ANOVA followed by a t-test post-hoc analysis with Benjamini-Hochberg false discovery rate (FDR) correction. Genes with an FDR adjusted p-value of less than 0.05 were considered significantly differentially expressed. As with the relapses, the homologous reinfections induced minimal changes in the host transcriptome (Fig 6A). Forty-five percent (18/40) of the upregulated DEGs during the homologous reinfections overlapped with the upregulated DEGs during relapses (Fig 6B). Pathway enrichment analysis of all upregulated DEGs during the homologous reinfections again identified pathways related to B cells (Fig 6C). The genes that were enriched in these pathways include those encoding B cell surface proteins such as CD19 and CD20 and B cell signaling molecules such as AKT, BTK, PLC-gamma2, and VAV-2 (S5 Table). Similar to relapses, these data indicate that the changes in the host responses during homologous reinfections were characterized by pathways involving B cells.
The changes in B cell subsets during the homologous challenge experiment were similar to those measured in relapse infections. There was an increase in SM B cells when the homologous infection was resolving at the post-primary time point (Fig 7A and 7B). However, unlike relapses, there was not a significant increase in USM B cells, and DN B cells significantly decreased during the homologous challenges (Fig 7B). The frequency of Ki67+ SM, USM, and DN B cells increased during the homologous reinfections like in the relapses although the USM and DN B cells did not increase in number (Fig 7C). Notably, only IgG+ SM B cells increased during the homologous reinfections whereas both IgG+ and IgM+ SM increased during relapses (Fig 7D). Importantly, USM, SM, and DN B cell numbers were inversely correlated with parasitemia during the homologous infections (Figs 7E, 7F and S10).
Although total IgM was unchanged during the homologous reinfections, IgM recognizing both iRBCs and uRBCs was significantly increased as observed during relapses, albeit at much lower levels than the initial primary infections (Figs 8A, S11A and S12A). Consistent with relapses, total IgG and iRBC-specific IgG were also increased during homologous reinfections (Figs 8B and S11B). Again, the IgG subclass was predominantly IgG1 (S11C Fig). Similar to the relapses, iRBC-specific IgG was inversely correlated with parasitemia during the homologous reinfections, but iRBC-specific IgM was not (Fig 8C and 8D). Notably, the IgG reactivity with iRBC versus uRBC lysates was significantly higher in the homologous reinfections, like the relapses (S12B Fig).
As with the relapses, the humoral response during the homologous reinfections was highly effective at opsonizing iRBCs (Fig 8E). The increase in phagocytic activity was again correlated with iRBC-specific IgG (Spearman’s ρ = 0.55, p = 0.0005), but unlike relapses, iRBC-specific IgM was also correlated (Spearman’s ρ = 0.48 p = 0.02) with opsonic phagocytosis activity during homologous reinfections (Fig 8F). Altogether, these data are consistent with B cell mediated immune responses conferring protection during relapses and homologous reinfections.
Next, we questioned how the immunity during a relapse may affect the number and proportion of asexual and sexual parasite stages during relapses and, thus, enumerated the sexual and asexual parasites by microscopy during the primary and relapse infections. During the primary infections, the parasite differentials were performed from patency until the peak of parasitemia. Samples after the peak were excluded from the analysis since these were collected after sub-curative blood-stage drug treatment. The parasites were enumerated during relapses for all days showing patent parasitemia.
As expected, the number of gametocytes were significantly reduced during the relapses given the significant reduction in parasitemia during relapses compared to the primary infections (Fig 9A). While the absolute number of gametocytes decreased, the cumulative proportion of circulating iRBCs that developed into gametocytes was significantly increased in the relapses (Fig 9B). In contrast, there was no significant difference between primary and relapse infections in the cumulative proportion of iRBCs containing ring, trophozoite, and schizont stages (Fig 9B). Notably, the percentage of days out of the primary and relapse infections that gametocytes were observed in the blood was also similar (Fig 9C). These data suggested that the immunity during relapses may disproportionately affect asexual stages as opposed to gametocytes.
To validate the microscopy results, we examined gametocyte gene expression using parasite transcriptomes obtained from whole blood RNA-Seq data. Samples with less than 100,000 parasite reads were removed from the analysis; these included some relapse samples. The post-peak time points from the primary infection were also excluded since these were collected after sub-curative antimalarial treatment. We limited the analysis to P. cynomolgi genes that are homologous to P. vivax gametocyte genes that have been associated with P. vivax transmission in vivo [38–40]. In concordance with the microscopy data, P. cynomolgi homologues of the mature gametocyte genes pvs25, pvs28, and pvlap5 had similar gene expression across the primary infections and relapses (Fig 9D). Overall, these results demonstrate that despite the development of effective B cell immunity and reduction of parasitemia during relapses, the relapses maintained detectable levels of gametocytes.
In this study, single, sporozoite-initiated infections with P. cynomolgi in a cohort of rhesus monkeys resulted in the establishment of immunity that was capable of suppressing parasitemia during relapses or homologous reinfections initiated 60 days after radical cure. Whole blood RNA-Seq analysis showed that the host response during relapses and homologous reinfections was associated with distinct changes in the host transcriptome related to B cells. This finding was corroborated by flow cytometry and antibody ELISAs demonstrating that class-switched memory B cells rapidly responded during relapses and homologous reinfections along with a concomitant increase in anti-iRBC IgG. Collectively, these data demonstrate that protective, but non-sterilizing, humoral immunity can form after a single P. cynomolgi infection and may be key for preventing disease during relapses. Similarly, studies have found P. vivax-specific B cells can persist for years after an initial infection, and investigations using the P. chabaudi mouse model of malaria have also confirmed that humoral immunity can form and protect against subsequent challenges [41–44]. Therefore, we speculate that additional factors such as the genetic diversity of local P. vivax populations may be responsible for circumventing immunity, leading to symptomatic relapse infections. This seems likely when considering the high genetic diversity of P. vivax relapse infections in endemic areas [45–48]. Additional factors that may contribute to the lack of or subversion of immunity during a relapse include the age of first exposure or the presence of co-infecting pathogens.
Antibody responses during Plasmodium infection have been studied for decades, yet the roles of antibody isotypes other than IgG are currently under investigation [49, 50]. The inverse association of P. cynomolgi-specific IgM antibodies with relapse parasitemia indicate these antibodies may be involved in suppressing parasitemia and preventing the development of disease. USM B cells can differentiate to produce IgM in secondary responses, and these cells expanded during relapses and were inversely correlated with parasitemia during relapses and homologous reinfections [51, 52]. IgM+ SM B cells also expanded during relapses. Together, our data and recent evidence from rodent malaria models and human samples support a role for IgM+ memory B cells in anti-Plasmodium immunity [53]. Future studies should aim to evaluate neutralizing IgM responses to identify the origin of the B cell subsets that are responsible for their production. Such experiments are needed to delineate if anti-parasite IgM antibodies arise from the memory B cell compartment during recall responses or if they originate from naïve B cells that are stimulated to differentiate and secrete IgM during each blood-stage infection. Identification of the B cell compartment where protective antibodies originate and persist is needed to understand naturally acquired immunity against relapsing malaria parasites and may help to advance the development of a P. vivax vaccine.
While IgM may play a role in neutralizing blood-stage parasites during relapses, parasite-specific IgG1 was the predominant isotype produced during both relapses and homologous reinfections. Typically IgG3 rather than IgG1 has been reported as being the dominant subclass in human malaria [54, 55]. This discrepancy is likely due to differences between the NHP and human immune systems whereby IgG1 mediates the majority of effector functions in NHPs compared to humans where IgG1 and IgG3 contribute similarly [56]. In the data reported here, IgG levels strongly correlated with opsonic phagocytosis activity of P. cynomolgi iRBCs, an important mechanism of peripheral parasite control during blood stage infection. IgG+ SM B cells were the most significantly expanded memory subset during relapses and homologous reinfections and were inversely correlated with parasitemia in both cases, showing the potential importance of the SM B cell compartment during P. cynomolgi relapses and homologous reinfections. This should be taken into consideration with the ‘anti-relapse’ vaccine strategies currently being considered [57].
Despite the beneficial roles of IgG and IgM during P. cynomolgi infection, these antibodies may also contribute to pathogenesis. Removal of uRBCs by the immune system is a substantial contributor to the development of malarial anemia in humans and NHP models [36, 58–61]. Malarial anemia has been associated with the production of anti-self antibodies that tag uRBCs for elimination in rodent models and in human P. falciparum and P. vivax infections [62–64]. In this study, IgM and IgG antibodies that recognized uRBC lysates were detected during the primary infections, relapses, and homologous reinfections. The peak of the anti-uRBC IgM response occurred when parasitemia plateaued, and these antibodies were inversely correlated with hemoglobin levels when anemia was observed. Together, these data are consistent with a role for anti-uRBC IgM and IgG antibodies in the development of malarial anemia in addition to parasite control. Production of anti-uRBC antibodies could be due to non-specific, polyclonal activation of B cells in response to inflammatory stimuli released by iRBCs. Alternatively, the production of these antibodies may be an adverse, yet necessary, component of the normal immune response against P. cynomolgi that provides benefits like parasite neutralization. Either way future studies should identify the origin and function of ‘anti-self’ antibodies produced during longitudinal Plasmodium infections in NHPs.
Human studies with P. vivax have documented that gametocytes are present in symptomatic and asymptomatic relapse infections [7, 23, 24]. However, it has remained unclear how gametocytes are affected in the face of ongoing immune responses during a relapse [65]. Our study showed that P. cynomolgi gametocytes are substantially reduced during asymptomatic P. cynomolgi relapses, but the cumulative proportion of gametocytes increases relative to asexual stages of the parasite. The asexual stages predominated in circulation, and their relative proportions remained similar between the primary and relapse infections. These results argue that the reduction in the number of gametocytes is likely due to removal of asexual parasites, thereby, preventing their development into gametocytes, rather than anti-gametocyte immunity during relapses. In essence, our data are consistent with the host developing immunity to reduce parasitemia and prevent disease while the parasite manages to produce gametocytes that remain in circulation for ingestion by mosquitoes. This situation is advantageous for relapsing malaria parasites because the establishment of non-sterilizing immunity minimizes the chances of the host succumbing to infection while allowing for continued opportunities for transmission. Importantly, this scenario is fitting with the biology of P. vivax gametocytes since these are detectable soon after patency and, thus, could be transmitted before relapse parasitemia is substantially reduced [66]. Future studies should address whether the potent humoral immune response during P. cynomolgi relapses in rhesus macaques may also possess transmission-enhancing properties as shown previously in Toque monkeys (Macaca sinica) infected with P. cynomolgi [67].
While our study provides the most comprehensive analysis of P. cynomolgi relapses to date, it is not without limitation. Although our data strongly support the premise that humoral immunity is important in suppressing parasitemia during relapses to ameliorate disease, other cell types are likely involved. For example, future analysis of monocytes and T cells would be useful, particularly since the decreased inflammation observed during relapses and homologous reinfections could result in improved T cell help for B cells, thereby, increasing the effectiveness of the humoral immune response in subsequent exposures. Nonetheless, our study demonstrates the importance of humoral immunity because the increase in IgG during relapses and homologous reinfections occurs nearly twice as fast as in primary infections. Also, the lower increase in the IgM response during relapses and reinfections compared to the primary infections is consistent with a strong secondary response. Second, this study was not designed to test the infectiousness of gametocytes to mosquitoes. However, this would be an important addition to P. cynomolgi relapse investigations based on our results. Lastly, although P. cynomolgi infections of rhesus macaques are a valuable experimental surrogate for human P. vivax infections, there are differences that may influence the development of immunity. For example, P. cynomolgi parasitemias in rhesus are typically higher than P. vivax parasitemias in humans, which could lead to establishment of durable immunity faster in the rhesus macaque—P. cynomolgi infection model. On the other hand, experimental infections with P. vivax in neurosyphalitic patients have demonstrated appreciable homologous immunity months to years after one infection [68–70]. The results from those studies argue that the data presented here have a high degree of relevance to P. vivax infections in humans.
In conclusion, our studies with P. cynomolgi in rhesus macaques and studies on human P. vivax infections collectively provide strong evidence that relapses and homologous reinfections do not necessarily result in clinically detectable disease [41, 68–70]. Instead, it is becoming clear that relapses and potential reinfections with the same parasite variant can be clinically silent, and we have shown that this is, at least in part, due to potent humoral immunity that forms after an initial infection. This is highly significant considering that we have shown that clinically silent P. cynomolgi relapses continue to harbor gametocytes. If individuals in endemic communities have clinically silent relapses, they will not seek treatment. Meanwhile, they may serve as a source of gametocytes that may remain infectious to mosquitoes. The number of clinically silent relapse infections and their infectiousness to mosquitoes remains largely unknown and should be evaluated carefully in the future. As a next step on the path to eliminating P. vivax and other relapsing malaria parasites, empirical studies that identify the factors that influence relapse pathogenesis, immunity, and infectiousness to mosquitoes are needed, and the P. cynomolgi-macaque models can be used for investigations in each of these areas.
Nonhuman primate cohort infections were performed at the Yerkes National Primate Research Center (YNPRC) at Emory University, an Association for Assessment and Accreditation of Laboratory Animal Care (AAALAC) international-certified institution. Freshly isolated, salivary gland sporozoites were generated for each infection using additional rhesus monkeys at the Centers for Disease Control and Prevention (CDC). Rhesus monkeys utilized for experiments were of Indian origin, male, 7–13 kg, and 5–6 years of age. All male animals were used for experiments to eliminate the female menstrual cycle as a contributor to the development of anemia. All procedures including blood collections, infections with malaria parasites, clinical interventions, etc. were reviewed and approved by Emory University’s and/or CDC’s Institutional Animal Care and Use Committees. During the experimental procedures at YNPRC, the animals were housed socially in pairs in compliance with Animal Welfare Act regulations as well as the Guide for the Care and Use of Laboratory Animals.
Plasmodium cynomolgi M/B strain parasites were used for all experiments as previously described [12].
Two thousand freshly isolated, salivary gland sporozoites were generated, isolated, and administered intravenously as previously described to initiate the initial infections with relapses and homologous reinfections [12].
Subcurative antimalarial treatments consisted of a single dose of artemether at 1 mg/kg administered intramuscularly (IM). Curative blood-stage treatments consisted of a 7 day regimen of artemether administered IM with the first dose at 4 mg/kg and subsequent doses at 2 mg/kg. Radical cure consisted of a combination treatment with artemether and primaquine. Artemether was first administered IM for 8 days using an initial dose of 4 mg/kg followed by subsequent doses at 2 mg/kg. At the conclusion of the artemether treatment, primaquine was administered orally in peanut butter at 2 mg/kg for 7 days.
Blood was collected in EDTA at pre-defined time points as indicated in the experimental schematics in Figs 1A and 5A. Parasitemia and hematological parameters were evaluated daily by light microscopy and complete blood counts (CBC) analysis, respectively. For the daily parasitemia and CBC assays, blood was collected into a pediatric capillary tube using a standardized ear-prick procedure as previously described [12]. Blood specimens utilized for transcriptomic analysis were bled directly into Tempus tubes according to the manufacturer’s suggested protocol. Plasma was collected from each time point prior to isolation of PBMCs for flow cytometry analysis as described below. Bone marrow samples were not utilized for the experiments presented here.
Daily parasitemia was determined as reported in Joyner et al. 2016 [12]. Briefly, thick and thin blood film preparations from capillary or venous blood were prepared and allowed to dry. Thin films were fixed with 100% methanol, and thick films left unfixed. The thick and thin films were then stained using a Wright’s-Gurr stain. For thick film preparations, parasites were enumerated by determining the number of parasites that were observed within 500–2000 white blood cells (WBCs) depending on the parasite density. The number of parasites per the number of WBCs was then calculated and multiplied with the leukocyte count as determined by the CBC to yield parasites per microliter. For days where parasitemia was too high to enumerate using thick blood films, the thin blood film was used; this was typically when parasitemia was greater than 1%. The number of parasites out of 1000–2000 RBCs were determined and the percent parasitemia calculated by dividing the number of parasites counted by the number of total RBCs counted and multiplying by 100. The percentage of infected RBCs was then multiplied against the RBC concentration as determined by the CBC analysis to determine parasites/μl. Parasitemia was determined by two expert microscopists independently through the course of each infection. If discrepancies were observed between the two readers, a third, independent microscopist counted the slides. The two most similar values were then averaged to determine the parasitemia at any point during infection.
The parasite stages present at the selected times during the infections were determined by thick or thin film microscopy by counting 10 to 100 parasites and noting their stage. Each slide was examined for at least 15 minutes before stopping. The proportion of rings, trophozoites, schizonts, and gametocytes, were then calculated and used to determine the frequency of each parasite-stage per microliter of blood using the number of parasites per microliter as determined above or as proportions (%) for area under the curve (AUC) analysis as described below.
Area under the curve (AUC) for parasitemia (parasites/μl) and proportions (%) of parasite stages during the primary, relapse, and/or homologous reinfections were calculated using Riemann sums by determining the trapezoidal area between each data point during the indicated time periods. The formula used to calculate AUC was as follows: (DataPoint1+DataPoint22)×Δtime.
Complete blood counts (CBCs) were performed prior to infections and daily after inoculation using capillary and/or venous blood. If values from the CBC were considered abnormal (e.g. low platelet counts or observation of nucleated RBCs), the values obtained by the hematology analyzer were either confirmed or adjusted based on a manual differential or manual platelet count. For manual differentials, the phenotype (e.g. monocyte, neutrophil, nucleated RBC, etc.) was determined, and the percentage of each subset calculated. If there was a discrepancy with the CBC based on the differential, the percentage of monocytes, lymphocytes, and granulocytes was adjusted to ensure accuracy. If nucleated RBCs were present, the number of nucleated RBCs was determined and subtracted from the leukocyte count and added to the RBC count. Rectal temperatures were also obtained when animals were sedated for sample collections. Notably, two pre-infection values were collected prior to the initial infections to ensure accurate naive measurements were obtained since abnormal values may occur before an NHP becomes used to daily interaction. These values were averaged to obtain the malaria naïve values used for the analysis of the clinical data.
A custom, nonhuman primate multiplex cytokine assay was designed and purchased from eBioscience/Affymetrix, which is now a part of Thermofisher. These kits were performed according to the manufacturer’s suggested protocol except for one modification. Instead of diluting plasma 1:1 with sample dilution buffer, the samples were not diluted prior to running the assay. This was altered after initial experiments demonstrated that many analytes were not within the dynamic range of the standard curves if additional dilutions were performed. Samples were fully randomized prior to performing the multiplex kit to minimize plate- and well-specific effects. All multiplex data was analyzed using the ProcartaPlex Analyst software available through Thermofisher. Concentrations of cytokines in the plasma were determined and used for downstream analyses.
Total RNA was extracted using the Tempus RNA isolation kit (Fisher Scientific; Cat#:4380204). Globin transcripts were depleted using GLOBINclear Human Kit (Fisher Scientific; Cat#:AM1980) according to the manufacturer's instructions. Libraries were prepared using the Illumina TruSeq mRNA stranded kit (Illumina Inc.; Cat#:20020595) as per manufacturer’s instructions. 1 ug of Globin depleted RNA was used for library preparation. ERCC (Invitrogen; Cat#:4456740) synthetic spike-in 1 or 2 was added to each Globin depleted RNA sample. The TruSeq method (high-throughput protocol) employs two rounds of poly-A based mRNA enrichment using oligo-dT magnetic beads followed by mRNA fragmentation (120–200 bp) using cations at high temperature. First and second strand cDNA synthesis was performed followed by end repair of the blunt cDNA ends. One single “A” base was added at the 3’ end of the cDNA followed by ligation of barcoded adapter unique to each sample. The adapter-ligated libraries were then enriched using PCR amplification. The amplified library was validated using a DNA tape on the Agilent 4200 TapeStation and quantified using fluorescence based method. The libraries were normalized and pooled and clustered on the HiSeq3000/4000 Paired-end (PE) flowcell on the Illumina cBot. The clustered PE flowcell was then sequenced on the Illumina HiSeq3000 system in a PE 101 cycle format. Each sample was sequenced to a target depth of 100 million pairs (50 million unique fragments) with exception of Time point 2 samples that were sequenced to 200 million pairs (100 million unique fragments).
Raw FASTQ files from the RNA-Seq experiments of all animals at all time points were aligned to the P. cynomolgi [21] and M. mulatta (version 7.8) [71] reference genomes using the Spliced Transcripts Alignment to a Reference tool (STAR, version 2.4.1c). The aligned features were further quantified and mapped using the High-Throughput Sequencing tool version 0.6.1p1[72] using only the P. cynomolgi reference to select parasite-specific transcripts. All sequencing and transcript mapping results were deposited to the NCBI GEO and SRA databases under the accessions GSE104223 (E23) and GSE104101 (E24).
Raw count data from MaHPIC Experiments 23, 24, and 25 were all library size normalized using the ‘DESeq2’ package for R [73]. Prior to normalization for library size, genes of extremely low read count (<10 reads across all samples) were filtered. RNA-Sequencing data taken during initial infections, homologous reinfections, and heterologous reinfections (not presented here) were normalized together. Data structure was then examined with principal component analysis. Individual animal effects were removed using Supervised Normalization of Microarrays with the ‘SNM’ package for R [32]. LIMMA was then used to assess gene expression changes during each infection and between infection stages [74]. Fraction of gene expression variance explained by unsupervised Ward’s hierarchical cluster analysis was determined by finding the ratio of between-cluster variance (B) to total variance (T), which is in turn the sum of between-cluster and within-cluster variance (W).
Where j = 1, 2, … are the different clusters, i = 1, 2, … are the different genes measured in each sample, k = 1, 2, … are the different samples, nj is the number of samples in cluster j, gj¯(i) is the average value of gene i in cluster j, g¯(i) is the average value of gene i across all samples, and gk(i) is the value of gene i in sample k. For the within-cluster variance equation, the cluster j that is used for each step of the summation is the cluster to which sample k belongs.
Only samples in which parasites were detected by microscopy and at least 100,000 total reads (corresponding to at least 90 parasite/ μL) were analyzed. For the comparison of initial infections versus relapses, the pre-peak and peak infection stages were used for comparison with the relapse time points. The early relapse or peak relapse points were used to represent relapses. For animals in which both early and peak relapse samples had sufficient parasitemia and parasite reads the earlier sample was selected to represent relapse. Notably, if there was not an early relapse infection stage, the peak relapse point was used for analysis. ROh14 did not have a relapse and thus was not analyzed. All samples were library size normalized together and log2-transformed using DESeq2. Changes in gene expression were then assessed by using a linear mixed effect model with a Tukey-Kramer HSD post hoc analysis. P-values < 0.05 were considered statistically significant.
Plasma was isolated prior to performing the PBMC isolation by centrifuging the blood samples at 400 × g followed by pipetting off the plasma. After removing the plasma, the blood pellet was resuspended in two times the original volume of blood that was received. This modification of the procedure did not appear to alter the viability or yield of PBMCs. After this step, the manufacturer’s suggested protocol was followed. After each isolation, each monkey’s PBMCs were washed two times in sterile PBS followed by enumeration on a Countess II fluorescent cell counter. The viability of the PBMCs was simultaneously assessed by Trypan Blue exclusion assay. PBMC viability was always ≥ 90%.
5×105–2×106 PBMCs were aliquoted into flow cytometry tubes for staining with fluorescently conjugated antibodies. The variation in number of PBMCs used for each staining procedure was due to leukopenia that developed during the acute, symptomatic infections. After aliquoting into individual FACS tubes, cells were washed once more in PBS prior to re-suspending in antibody cocktails comprised of the antibodies indicated in S6 Table for the initial infections with relapses and S7 Table for the homologous reinfections. Notably, some markers listed are not presented in the manuscript, but are provided to convey the complete panel configurations used in each experiment.
All staining procedures were two-step. For surface IgG staining, the IgG was prepared in a separate cocktail and added first, followed by a 30-minute incubation, washing PBS by centrifugation at 400 × g, and then resuspending in a cocktail that contained the other antibodies in the panels.
For intracellular staining, cells were initially surface-stained with the cocktail followed by incubation in eBioscience FoxP3 fix perm buffer (Thermofisher) overnight at 4°C for intracellular markers. After fixing overnight, the cells were washed according to the manufacturer’s procedure and then incubated for 45 minutes at 4°C with antibodies against intracellular markers. The cells were then washed twice in the fix/perm buffer provided by the kit and resuspended in 100–200 μl of PBS depending upon cell yield. All samples were acquired on an LSR-II flow cytometer using standardized acquisition templates and rainbow calibration particles for voltage control. Compensation controls were run at each acquisition. Data were initially compensated in FlowJo version 10.1 followed by exporting to Cytobank for gating. Cell population level statistics were then exported from Cytobank for further analysis.
Absolute numbers for each B cell subset was determined by calculating the percentage of each subset out of the mononuclear cells in the sample and multiplying the percentage with the mononuclear cells/μl value obtained from the CBC at each time point. The mononuclear cells/μl was obtained by adding the lymphocyte/μl value to the monocyte/μl value. If two values for absolute numbers were obtained due to a population being present in both panels (e.g. switched memory), the values were treated as technical duplicates and averaged to obtain the final value used for each analysis.
Corning high-binding microtiter plates were coated with Anti-Monkey IgG+IgA+IgM (Rockland Immunochemicals) or Anti-Monkey IgM (Life Diagnostics) diluted in ELISA coating buffer (Abcam) to 0.6 ug/ml and 5 ug/ml, respectively. The plate was incubated overnight at 4°C followed by washing four times with PBS containing 0.05% Tween-20 (PBS-T). After the final wash, the plate was blotted dry and blocked using serum-free Sea Block (Abcam) for two hours at RT followed by four washes in PBS-T. Plasma samples from the different infection points were diluted 1:100,000 for total IgG or 1:10,000 for total IgM in 10–33% serum-free Sea Block and then added to each well. The plate was then incubated at RT for 2 h followed by washing four times with PBS-T. After blotting dry, HRP-conjugated anti-IgG (Jackson Immunoresearch) or HRP-conjugated anti-IgM (Jackson Immunoresearch) diluted 1:30,000 or 1:20,000 in 10–33% Sea Block in PBS, respectively, were added to each well and incubated for 1 h at RT in the dark. After incubating, the plate was washed four times with PBS-T and 100 μl of High Sensitivity TMB Substrate (Abcam) was added to each well and allowed to develop for 3–5 minutes. One hundred microliters of Stop Solution (Abcam) was added to stop the reaction. The absorbance at 450 nm was then measured, and total IgM or total IgG antibody concentrations were calculated based on a 4-PL standard curve using purified IgM calibrators from Abcam’s Monkey Total IgM ELISA kit and using purified IgG Monkey Calibrators from Rockland Immunochemicals. Concentrations of total IgG and IgM were used for downstream analyses.
Rhesus macaques were inoculated with cryopreserved, blood-stage P. cynomolgi B/M strain parasites to generate schizonts for lysate preps described below. Briefly, a vial of cryopreserved, blood-stage parasites were removed from the liquid nitrogen and quickly thawed in a 37°C water bath. After thawing, saline solutions of different concentrations were added drop-wise to slowly change the osmotic pressure while preventing RBC lysis. The number of ring-stage parasites were then enumerated using light microscopy as described above and inoculated intravenously into a rhesus macaque. The infections were followed daily for each monkey until parasitemia reached 3–10% ring-stage parasites. At this time, blood containing predominantly rings was collected in sodium heparin, washed, and depleted of leukocytes and platelets by passing over a glass bead column and through a Plasmodipur filter. The parasites were then matured ex vivo to 3–8 nucleated schizonts under blood-gas conditions (5%:5%:90%;O2:CO2:N2) in RPMI supplemented with L-glutamine, supplemented with 0.25% sodium bicarbonate, 50 μg/ml hypoxanthine, 7.2 mg/ml HEPES, 2 mg/ml glucose, and 10–20% Human AB+ serum. When mature, the schizonts were isolated by a 1.093 g/ml Percoll density gradient. The parasite layer was then isolated and washed 4 times in sterile RPMI, aliquoted, and stored in vapor phase liquid nitrogen until needed.
Aliquots of parasite or uninfected RBC pellets were removed from liquid nitrogen storage, thawed quickly in a 37°C water bath and placed back into the liquid nitrogen tank for ten minutes. This procedure was repeated three more times. After the final thaw, 1 volume of PBS was added followed by vigorous vortexing for 1–2 minutes. The aliquot was then centrifuged at 3,000 × g for 10 minutes at 4°C. The supernatant was then removed and placed into another sterile tube and the pellet discarded. This process was repeated three more times. After the final centrifugation, the protein concentration was determined using a Pierce BCA assay according to the manufacturer’s protocols. The lysates were then diluted to optimal concentrations for ELISAs in PBS, aliquoted, and stored at -80°C until needed.
Corning high-binding microtiter plates were coated with schizont lysate or uninfected RBC lysate diluted in ELISA coating buffer (Abcam) to 5 ug/ml. The plate was incubated overnight at 4°C followed by washing four times with PBS containing 0.05% Tween-20 (PBS-T). After the final wash, the plate was blotted dry and blocked using serum-free Sea Block (Abcam) for two hours at RT followed by four washes in PBS-T. Plasma samples from the different infection points were diluted 1:100 in 10–33% serum-free Sea Block and then added to each well. The plate was then incubated at RT for 2 h followed by washing four times with PBS-T. After blotting dry, horseradish-peroxidase (HRP) conjugated anti-IgG (Jackson Immunoresearch) or HRP-conjugated anti-IgM (Jackson Immunoresearch) diluted 1:30,000 or 1:20,000 in 10–33% Sea Block in PBS, respectively, were added to each well and incubated for 1 h at RT in the dark. After incubating, the plate was washed four times with PBS-T and 100 μl of High Sensitivity TMB Substrate (Abcam) was added to each well and allowed to develop for 3–5 minutes. One hundred microliters of Stop Solution (Abcam) was then added to stop the reaction. The absorbance at 450 nm was then measured, and the OD450 of iRBC and uRBC-specific IgG and IgM were used for downstream analyses.
Corning high-binding microtiter plates were coated with schizont lysate diluted in ELISA coating buffer (Abcam) to 5 ug/ml. As a positive control, duplicate wells were coated with recombinant expressed rhesus IgG1, IgG2, or IgG3 (NHP Reagent Resource) diluted to 1 ug/ml in ELISA coating buffer. The plate was incubated overnight at 4°C followed by washing four times with PBS containing 0.05% Tween-20 (PBS-T). After the final wash, the plate was blotted dry and blocked using serum-free Sea Block (Abcam) for two hours at RT followed by four washes in PBS-T. Plasma samples from the different infection points were diluted 1:100 in 10% serum-free Sea Block and then added to each well. The plate was then incubated at RT for 2 h followed by washing four times with PBS-T. After blotting dry, mouse anti-rhesus IgG1, IgG2, or IgG3 (NHP Reagent Resource) diluted 1:10,000, 1:1,000, and 1:10,000 in 10% Sea Block in PBS, respectively, was added to each well and incubated at RT for 1 h. Following four washes in PBS-T and blotting dry, HRP-conjugated anti-mouse IgG (Jackson Immunoresearch) diluted 1:10,000 in 10% Sea Block in PBS was added to each well and incubated for 1 h at RT in the dark. After incubating, the plate was washed four times with PBS-T and 100 μl of High Sensitivity TMB Substrate (Abcam) was added to each well and allowed to develop for 30 minutes. One hundred microliters of Stop Solution (Abcam) was then added to stop the reaction. The absorbance at 450 nm was then measured, and optical densities (ODs) were used for downstream analyses.
We adapted a previously established phagocytosis assay for P. cynomolgi [75]. Briefly, the THP-1 monocytic cell line was obtained from ATCC and maintained in vented 75cm2 culture flasks at 10% CO2 in RPMI-1640 supplemented with 10% fetal bovine serum, 2mM L-glutamine, 10mM HEPES, 1mM sodium pyruvate, 4500 mg/L glucose, and 1500 mg/L sodium bicarbonate. The cells were maintained at a density of 1 × 105 cells/ml of culture and were not allowed to exceed 1 × 106 cells/ml. Plasmodium cynomolgi strain M/B were thawed, matured in vitro to schizonts, and isolated as described above. Purified schizonts were incubated with 5 ug/ml dihydroethidium (DHE) for 20 min at 37°C, followed by 3 washes in THP-1 media before use in the assay. After labeling with DHE, the schizonts were opsonized in heat-inactivated plasma from different specimen collections for 45 minutes at RT in the dark. While the parasites were opsonizing, THP-1 cells were harvested and an aliquot of THP-1 cells was incubated with 5 uM Cytochalasin D for 1 h at 37°C to serve as a negative control. THP-1 cells were then added to each well to a final Effector Target ratio of 1:20 and incubated at 37°C for 3 h. The cells were then transferred to FACS tubes, washed twice with THP-1 media, and then lysed with ACK lysing solution for 10 min at RT in the dark. Cells were then re-suspended in PBS and acquired immediately on a BD LSR-II using a standardized acquisition template.
All statistical analyses were performed using a linear mixed-effect model with Tukey-Kramer HSD post-hoc analysis. For the statistical model, each animal was treated as a random effect with time points as fixed effects. All data were transformed as necessary to ensure the best model fit, and the best model fits were typically obtained with a log10, log2, or arcsin transformation.
All data went through rigorous validation protocols and are publicly deposited in public repositories. All clinical data associated with the experiment have been publicly released on PlasmoDB http://plasmodb.org/plasmo/mahpic.jsp (see http://plasmodb.org/common/downloads/MaHPIC/Experiment_23/ and http://plasmodb.org/common/downloads/MaHPIC/Experiment_24/ for the datasets in this manuscript). All RNASeq results have been publicly released as described in the Materials and Methods. Flow cytometry, multiplex cytokine assays, and ELISA are publicly available at ImmPort as part of study SDY1409.
|
10.1371/journal.pntd.0000621 | Identification by PCR of Non-typhoidal Salmonella enterica Serovars Associated with Invasive Infections among Febrile Patients in Mali | In sub-Saharan Africa, non-typhoidal Salmonella (NTS) are emerging as a prominent cause of invasive disease (bacteremia and focal infections such as meningitis) in infants and young children. Importantly, including data from Mali, three serovars, Salmonella enterica serovar Typhimurium, Salmonella Enteritidis and Salmonella Dublin, account for the majority of non-typhoidal Salmonella isolated from these patients.
We have extended a previously developed series of polymerase chain reactions (PCRs) based on O serogrouping and H typing to identify Salmonella Typhimurium and variants (mostly I 4,[5],12:i:-), Salmonella Enteritidis and Salmonella Dublin. We also designed primers to detect Salmonella Stanleyville, a serovar found in West Africa. Another PCR was used to differentiate diphasic Salmonella Typhimurium and monophasic Salmonella Typhimurium from other O serogroup B, H:i serovars. We used these PCRs to blind-test 327 Salmonella serogroup B and D isolates that were obtained from the blood cultures of febrile patients in Bamako, Mali.
We have shown that when used in conjunction with our previously described O-serogrouping PCR, our PCRs are 100% sensitive and specific in identifying Salmonella Typhimurium and variants, Salmonella Enteritidis, Salmonella Dublin and Salmonella Stanleyville. When we attempted to differentiate 171 Salmonella Typhimurium (I 4,[ 5],12:i:1,2) strains from 52 monophasic Salmonella Typhimurium (I 4,[5],12:i:-) strains, we were able to correctly identify 170 of the Salmonella Typhimurium and 51 of the Salmonella I 4,[5],12:i:- strains.
We have described a simple yet effective PCR method to support surveillance of the incidence of invasive disease caused by NTS in developing countries.
| The genus Salmonella has more than 2500 serological variants (serovars), such as Salmonella enterica serovar Typhi and Salmonella Paratyphi A and B, that cause, respectively, typhoid and paratyphoid fevers (enteric fevers), and a large number of non-typhoidal Salmonella (NTS) serovars that cause gastroenteritis in healthy hosts. In young infants, the elderly and immunocompromised hosts, NTS can cause severe, fatal invasive disease. Multiple studies of pediatric patients in sub-Saharan Africa have documented the important role of NTS, in particular Salmonella Typhimurium and Salmonella Enteritidis (and to a lesser degree Salmonella Dublin), as invasive bacterial pathogens. Salmonella spp. are isolated from blood and identified by standard microbiological techniques and the serovar is ascertained by agglutination with commercial antisera. PCR-based typing techniques are becoming increasingly popular in developing countries, in part because high quality typing sera are difficult to obtain and expensive and H serotyping is technically difficult. We have developed a series of polymerase chain reactions (PCRs) to identify Salmonella Typhimurium and variants, Salmonella Enteritidis and Salmonella Dublin. We successfully identified 327 Salmonella isolates using our multiplex PCR. We also designed primers to detect Salmonella Stanleyville, a serovar found in West Africa. Another PCR generally differentiated diphasic Salmonella Typhimurium and monophasic Salmonella Typhimurium variant strains from other closely related strains. The PCRs described here will enable more laboratories in developing countries to serotype NTS that have been isolated from blood.
| In industrialized countries, non-typhoidal Salmonella (NTS) constitute a well recognized public health problem that in healthy subjects is overwhelmingly encountered clinically as self-limited gastroenteritis [1],[2]. In immunocompromised and debilitated hosts, NTS can become invasive, leading to bacteremia, sepsis and focal infections (e.g., meningitis) [2],[3]. Among infants less than three months of age who become infected with NTS in industrialized countries, invasiveness is also occasionally observed, resulting in bacteremia and focal infections [4].
Interestingly, whereas systematic blood culture-based surveillance of febrile pediatric patients in Asia has clearly highlighted the high incidence of bacteremia associated with Salmonella enterica serovars Typhi and Paratyphi A in children residing in crowded urban settings [5]–[7], isolation of NTS has not been common. In striking contrast, systematic blood culture-based surveillance and clinical studies of hospitalized and ambulatory pediatric patients <60 months of age with fever or focal infections in sub-Saharan Africa have documented the important role of NTS as invasive bacterial pathogens [8]–[17]. NTS constituted one of the three most common invasive bacterial pathogens in all these studies. Importantly, two serovars, Salmonella Typhimurium (and Typhimurium variants) and Salmonella Enteritidis have been reported to account for 79–95% of all bacteremic non-typhoidal Salmonella infections in sub-Saharan Africa [9], [11]–[13],[15],[16],[18],[19]. Salmonella Dublin has been associated with a few percent of cases in some studies [12],[13] but with a more substantial proportion in Mali [18], where a fourth serovar, Salmonella Stanleyville, also accounted for a notable proportion of all isolates [18], bringing the cumulative total to >95% of all strains.
We previously developed a multiplex polymerase chain reaction (PCR)-based approach to identify the three main pathogens responsible for typhoid (Salmonella Typhi) and paratyphoid (Salmonella Paratyphi A and Salmonella Paratyphi B) fevers [18]. Three sequential PCRs identify strains of Salmonella serogroups A, B or D (and Vi positive or negative); strains that express Phase 1 flagellar (H) antigen types H:a, H:b or H:d; and strains incapable of fermenting d-tartrate (d-T). By means of this PCR technology, Salmonella Typhi (O serogroup D, Vi+; H:d), Salmonella Paratyphi A (O serogroup A; H:a) and Salmonella Paratyphi B (O serogroup B; H:b; d-T non-fermenter) strains were identified with 100% sensitivity and 100% specificity.
Classical Salmonella serotyping methods identified the serovars of 336 NTS isolates from blood cultures of febrile children <16 years of age in Bamako, Mali, obtained in the course of systematic surveillance of children admitted to hospital or seen in the Emergency Department with fever or invasive infection syndromes [20]–[22]. Salmonella Typhimurium and “variants” (mainly I 4,[5],12:i:-), Salmonella Dublin, Salmonella Enteritidis and Salmonella Stanleyville were the most commonly isolated NTS [18]. Herein, we describe PCRs that when used in conjunction with the O serogrouping PCR described by Levy et al. [18] can identify Salmonella Typhimurium and variants (O serogroup B; H:i), Salmonella Enteritidis (O serogroup D, Vi-; H:g,m), Salmonella Dublin (O serogroup D, Vi+ or Vi-; H:g,p) and Salmonella Stanleyville (O serogroup B; H:z4,z23) with 100% sensitivity and 100% specificity. We anticipate that this methodology will be useful in reference laboratories and major clinical microbiology laboratories to identify Salmonella isolated from blood and other sterile sites in developing countries where robust PCR-based typing techniques are becoming increasingly popular and because high quality H typing sera are difficult to obtain, expensive and technically demanding to use.
The surveillance protocol and consent form were reviewed by the Ethics Committee of the Faculté de Médecine, Pharmacie et Odonto-Stomatologie, Université de Bamako, and by the Institutional Review Board of the University of Maryland, Baltimore. For any patient eligible for laboratory surveillance to detect invasive bacterial disease, informed consent was obtained prior to their enrollment; ∼95% of eligible subjects agreed to participate. Since the literacy rate in Bamako is <30%, as is customary practice for CVD-Mali clinical studies [20]–[22], the consent form was translated into Bambara and several other local languages and the translations recorded on audiotape [20]. CVD-Mali personnel explain the study, including the objectives and risks and benefits associated with participation. The audiotaped version of the consent form is then played and any questions posed are answered. Once the parent or patient has had all questions answered and agrees to participate, this is documented on a printed consent form written in French. If the participant is illiterate, a witness who is present throughout the consent procedure completes the necessary portions and signs the consent form; the parent/participant marks the consent form (either fingerprint or other notation). If the person is literate, then he/she may read and sign the consent form. This standard method of obtaining consent practiced by CVD-Mali was approved by ethics commitees in Mali and at the University of Maryland.
Since July 2002, clinical staff of the Centre pour le Développement des Vaccins du Mali (CVD-Mali) and l'Hôpital Gabriel Touré (HGT) have been conducting systematic surveillance to detect invasive bacterial disease among hospitalized children <16 years of age [20]–[22]. Age-eligible children presenting to the emergency department with fever (≥39°C) or focal clinical findings suggestive of invasive bacterial infection (meningitis, septic arthritis, etc.) and requiring hospitalization are referred to CVD-Mali staff by the evaluating clinicians. A CVD-Mali physician obtains informed consent, records clinical and epidemiologic data, and obtains blood (and other relevant fluids) for culture in the HGT Clinical Bacteriology Laboratory. The child's clinician is promptly notified when a culture yields a bacterial pathogen.
Salmonella Typhimurium strain 81.23500, Salmonella Enteritidis strain CVD SE and Salmonella Dublin strain 06-0707 were used to develop the multiplex PCR. Twenty-four control strains which came from the Salmonella Reference Laboratory of the Centers for Disease Control and Prevention (CDC), Atlanta, GA or the Center for Vaccine Development, Baltimore, MD have previously been described [18]. These strains were Salmonella serovars of various O serogroups (B, C1, C2, D, E1, O28 and O38) and H types (b, c, d, h, i, g, k, l, m, p, s, t, v, y, z10 and z29). Nine O serogroup B, Phase 1 flagella antigen H:i reference strains from the CDC were used to develop a PCR that discriminates between Salmonella Typhimurium and I 4,[5],12:i:- (Table 1). The NTS test strains consist of 327 Salmonella serogroup B and D isolates that were originally obtained from the blood cultures of febrile patients at l'Hôpital Gabriel Touré in Bamako, Mali. These strains were identified by conventional microbiological and classical serotyping methods by the CVD and CDC, as previously described [18]; 69 isolates were O serogroup D, including 37 Salmonella Dublin and 32 Salmonella Enteritidis, and 258 isolates were O serogroup B.
PCR was performed in 1× PCR buffer, 3.5 mM MgCl2, 0.2 mM of dNTPs and 0.2 U of Invitrogen Taq DNA polymerase (final volume of 25 µl) in an Eppendorf Mastercycler®. The primer mixes contained primers at a concentration of 5 µM each (final concentration of 0.2 µM) except for FFLIB and RFLIA that were used at a concentration of 10 µM each and the positive control primers (16SF and DG74) that were used at a concentration of 2.5 µM each. For each PCR reaction, 1.0 µl of primer mix was used. Crude DNA was prepared by suspending 3 colonies in 100 µl water and boiling for 10 min followed by centrifugation at 16,000×g for 30 sec and purified DNA was prepared using a GNOME DNA kit (QBIOgene, Irvine, CA) according to the manufacturer's instructions, and 5 µl of DNA was used in each PCR. The cycling parameters of the multiplex PCR that detects H:i, H:g,p and Sdf I involved denaturation at 95°C for 2 min, followed by 25 cycles comprised of heating to 95°C for 30 sec, 64°C for 30 sec and 72°C for 15 sec, and a final step of 72°C for 5 min. The cycling parameters of the PCR that discriminates between Salmonella Typhimurium and I 4,[5],12:i:- involved denaturation at 95°C for 2 min, followed by 25 cycles of 95°C for 30 sec, 64°C for 30 sec and 72°C for 1.5 min, and a final step of 72°C for 5 min. PCR products were separated on 2% (w/v) agarose gels, stained with ethidium bromide and visualized using a UV transilluminator.
Figure 1 shows that the primers within the multiplex PCR were able to clearly identify the appropriate NTS alleles. A 779-bp product was amplified from Salmonella Dublin (fliC-gp), a 551-bp product was amplified from Salmonella Typhimurium (fliC-i) and a 333-bp product was amplified from Salmonella Enteritidis (Sdf I). The internal positive control primers (universal 16S rRNA gene primers) amplified a 167-bp product from each strain.
To preliminarily assess the specificity of the multiplex PCR assay, we tested 24 control Salmonella strains consisting of a range of serovars (previously described in [18]) in a blinded fashion (Figure 2). The multiplex PCR correctly identified Salmonella Typhimurium and Salmonella Cotham as H:i, Salmonella Dublin as H:g,p and Salmonella Enteritidis as containing Sdf I (Figure 2). Faint products of the size of Sdf I were observed for Salmonella Meleagridis and Salmonella Livingstone. However, Salmonella Meleagridis is O serogroup E1 and Salmonella Livingstone is O serogroup C1, so when also tested by our previously described O serogrouping PCR [18], these serovars would not be mistaken as Salmonella Enteritidis. The same is true for Salmonella Cotham, which although it possesses fliC-i, is not O serogroup B and would not be mistaken as Salmonella Typhimurium. Therefore, the new multiplex PCR was sensitive in terms of its ability to identify serovar Cotham as H:i and was specific, when combined with the O-serogrouping PCR, in showing that the strain was not serovar Typhimurium. We also blind-tested a sample of Salmonella Typhi and Salmonella Paratyphi A and B strains to ensure that the PCR would not detect these strains. The multiplex PCR correctly identified fliC-i of six Salmonella Typhimurium, Sdf I of four Salmonella Enteritidis, and fliC-g,p of five Salmonella Dublin strains but only the 16S rRNA gene was amplified from five strains each of serovars Typhi, Paratyphi A and Paratyphi B (data not shown).
We blind-tested 69 non-Typhi serogroup D Salmonella and 258 serogroup B strains that were originally obtained from the blood cultures of febrile patients at l'Hôpital Gabriel Touré in Bamako, Mali [18] with the multiplex PCR designed to identify Salmonella Typhimurium (I 4,[5],12:i:1,2) and variants (monophasic I 4,[5],12:i:- and non-motile (NM) I 4,[5],12:NM), Salmonella Enteritidis and Salmonella Dublin. This PCR was performed in parallel to serotyping. We correctly identified all the serogroup D isolates (37 Salmonella Dublin and 32 Salmonella Enteritidis) and all 232 Salmonella Typhimurium and variant strains (Table 3). If the Salmonella Typhimurium-like strains (i.e., I 4,[5],12:i:- and I 4,[5],12:NM) are included in the target group then the PCR is 100% sensitive and 100% specific in identifying Salmonella Typhimurium, Salmonella Enteritidis, Salmonella Dublin and Salmonella Typhimurium-like organisms. The remaining 26 serogroup B isolates were negative for the tested targets.
During the course of this study, we decided to determine the prevalence of I 4,[5],12:i:- in Mali. Levy et al. [18] identified 220 Salmonella Typhimurium, four I 4,[5],12:i:- and eight I 4,[5],12:NM strains. However, in this previous study, Phase 2 flagella typing was not performed on all of the strains. We re-examined the 220 Salmonella O serogroup B, H:i isolates that had been previously been presumptively identified as Salmonella Typhimurium and used classical methods (i.e., sera against the Phase 2 H1,2 flagella) to determine that 48 isolates were in fact I 4,[5],12:i:- (bringing the total number of isolates of this serovar to 52) and one isolate was I 4,[5],12:NM (bringing the total number of isolates of this serovar to nine). The remaining 171 strains were confirmed as Salmonella Typhimurium.
We have combined previously described primers in a PCR to discriminate between Salmonella Typhimurium and I 4,[5],12:i:-. Primers FFLIB and RFLIA amplify the fliB-fliA intergenic region of the flagellin gene cluster [24]. Salmonella Typhimurium strains possess an IS200 fragment in this region [25]. Burnens et al. [25] showed that 21 of 23 isolates of Salmonella Typhimurium and none of 85 isolates of 37 other Salmonella serovars contained IS200 in this region. Primers FFLIB and RFLIA have been reported to amplify a 1-kb product from Salmonella Typhimurium and I 4,[5],12:i:- strains and a 250-bp product from all other serovars [24]. However, when validating these primers, we found that a 1-kb fragment was amplified from Salmonella Farsta (not tested by Echeita et al. [24]) suggesting that this serovar also possesses IS200 in the fliB-fliA intergenic region (Figure 3).
Primers Sense-59 and Antisense-83 amplify the fljB allele [26]. Primer Sense-59 binds at position 258 and primer Antisense-83 binds at position +100 of the 5′-3′ consensus fljB1,2 sequence. These primers amplify a 1389-bp product from strains that possess a Phase 2 flagellar antigen and no product from strains that lack a Phase 2 flagellar antigen such as I 4,[5],12:i:-. As shown in Figure 3, the PCR was able to discriminate between Salmonella Typhimurium and I 4,[5],12:i:- strains and other serogroup B, H:i serovars except Salmonella Farsta.
We tested all the Salmonella Typhimurium, I 4,[5],12:i:- and I 4,[5],12:NM strains identified in Mali and found that 170 of 171 Salmonella Typhimurium strains were correctly identified (i.e., possessed a 1-kb fliB-fliA intergenic region product and fljB1,2), and 51 of 52 I 4,[5],12:i:- strains were correctly identified (i.e., possessed a 1-kb fliB-fliA intergenic region product and lacked fljB1,2) (Table 4). The nine I 4,[5],12:NM strains produced mixed results in that all nine strains produced a 1-kb fliB-fliA intergenic region product but three strains possessed fljB1,2.
Since Salmonella Stanleyville was found to be fairly common among the Mali NTS isolates, we added primers to detect fliC-z4,z23 of Salmonella Stanleyville to the multiplex PCR containing primers H-for, Hi, sdfF, sdfR, 16SF and DG74. The primers were first tested on Salmonella Stanleyville by themselves and produced a 427-bp amplicon. The fliC-z4,z23 primers were then added to the multiplex primer mix and PCR was performed (using the previously optimized conditions) on all 26 Salmonella Stanleyville strains, and a sample of 10 Salmonella Typhimurium, 10 Salmonella Dublin and 11 Salmonella Enteritidis strains. Correct amplicons were observed for all the strains tested. Figure 4 shows amplicons from a sample of three Salmonella Stanleyville strains and the control Salmonella Typhimurium, Salmonella Enteritidis and Salmonella Dublin strains.
We have combined published primers and new primers in a multiplex PCR that, following the application of a previously described O serogrouping multiplex PCR [18], can identify Salmonella Typhimurium (and variants), Salmonella Enteritidis, Salmonella Dublin and Salmonella Stanleyville. Detection of Salmonella Typhimurium, Salmonella Dublin and Salmonella Stanleyville is based on amplification of the respective fliC alleles. We were unable to design primers to detect fliC-g,m of Salmonella Enteritidis due to the high nucleotide identity between fliC-g,m and fliC-g,p (of Salmonella Dublin). We therefore used primers to detect “Salmonella difference fragment I” (Sdf I), a segment of Salmonella Enteritidis DNA that was reported to be absent from 73 non-Enteritidis Salmonella enterica isolates comprising 34 different serovars as determined by PCR [27]. We confirmed the utility of Sdf I, with the exception of serovars Meleagridis and Livingstone. We found that Salmonella Livingstone yielded a weak PCR product using the same Sdf I primers that were previously reported [27]. The disparity could be due to a difference in the amplification method (different polymerases and cycling conditions were used).
From the epidemiologic and public health perspective, being able to detect strains that are genetically similar to Salmonella Typhimurium yet that constitute distinct serovars (i.e., I 4,[5],12:i:- and I 4,[5],12:NM) is important (e.g., for outbreak investigations). In the USA and Europe such strains are increasingly being reported [28]–[30]. In Spain, I 4,[5],12:i:- was the fourth most commonly isolated Salmonella serovar from humans from 1998–1999 [29] and several studies suggest that this monophasic serovar is a variant of Salmonella Typhimurium [24], [31]–[33]. The PCR that we have described can generally discriminate the diphasic Salmonella Typhimurium serovar (I 4,[5],12:i:1,2) from monophasic (I 4,[5],12:i:-) variants. Only one Salmonella Typhimurium was misidentified as I 4,[5],12:i:- and vice versa. It is possible that our PCR will not be able to detect some serologically monophasic I 4,[5],12:i- strains as lack of Phase 2 flagellar antigen expression can be due to a variety of mechanisms ranging from point mutations to partial or complete deletions in fljB1,2 and adjacent genes. Additionally, if there is a deletion in the first 250 bp of fljB1,2, the primers we have chosen will not identify the strain as I 4,[5],12:i-. Furthermore, our PCR scheme cannot differentiate between Salmonella Typhimurium and Salmonella Farsta. However, in practical terms, this is unlikely to pose a problem as Salmonella Farsta is extremely rare.
One small set of strains where our PCR gives differing results from traditional serological methods are Salmonella Typhimurium-like non-motile variants (I 4,[5],12:NM). Notably, all nine Malian strains identified by serotyping methods as I 4,[5],12:NM were found to possess the fliC-i allele and three of the strains also possessed the fljB1,2 gene. Two quite distinct explanations can account for these observations. One is that in some strains lack of motility is not due to loss of flagellar genes but rather to other factors (e.g., regulation) that keep expression turned off. Alternatively, it may be that our genetic identification of these strains is correct and that the failure to detect flagella phenotypically is merely a consequence of not knowing how to grow the bacteria under conditions optimal for expression of those flagella. We assume that the I 4,[5],12:NM strains from Mali are Salmonella Typhimurium variants as they possess fliC-i and IS200 in the fliB-fliA intergenic region. It is also possible, albeit unlikely, that they could be the very rarely isolated Salmonella Farsta.
Soyer et al. [34] have reported that there are at least two common clones of I 4,[5],12:i:- with different genomic deletions (an ‘American’ deletion genotype and a ‘Spanish’ deletion genotype). Both I 4,[5],12:i:- clones completely lack fljB and fljA. Preliminary analysis of the deletion using a variety of primers that amplify different sections of the fljB1,2 gene indicates that the I 4,[5],12:i:- strains from Mali appear to possess the 3′ end of fljB and the entire fljA ORF. At least 250 bp of fljB (including the Sense-59 binding site) has been deleted at the 5′ end (data not shown). This suggests that these strains are genetically different from both the Spanish and American I 4,[5],12:i:- isolates. We are sequencing the deletion in several Malian I 4,[5],12:i- strains to determine the exact deletion. It will be interesting to see whether I 4,[5],12:i:- strains from other African countries are genetically similar to the Malian strains.
Several other DNA-based Salmonella typing methods have been described [35]–[41]. However, some of these do not identify the breadth of enteric fever and NTS serovars of our multistep, multiplex PCR or fail to include an internal positive control. An O serogroup-specific Bio-Plex assay to detect serogroups B, C1, C2, D, E and O13 and serovar Paratyphi A [42] and a DNA sequence-based approach to serotyping have also been described [43]. However, these methods require greater financial and technical resources over those required for our method. Our PCRs are novel because they use as few primers as possible to identify the most common non-typhoidal Salmonella serovars isolated from blood and other invasive sites in sub-Saharan Africa, including Salmonella Typhimurium (and several variants), Salmonella Enteritidis, Salmonella Dublin and Salmonella Stanleyville. Since the late 1980s, the majority (85 to 95%) of NTS associated with invasive disease in sub-Saharan Africa belong to these serovars [9], [11]–[13],[15],[16],[18],[19]. Therefore, we do not believe that there is a need for multiplex PCRs that detect more serovars unless the epidemiologic picture changes. We have tried to keep the PCRs as simple as possible so that they can be performed easily and the results interpreted correctly in laboratories in Africa that may be new to PCR. If a large outbreak or otherwise frequent isolation occurred of a serovar not presently recognized or contained within our multiplex, this serovar would not be identifiable using our PCR and would have to be identified in a reference laboratory using antisera or by molecular serotyping.
We are currently evaluating various PCR reagents that are stable at room-temperature and can be readily obtained by laboratories in Africa. Depending on the prevalence of certain serovars in a given country, either typhoidal or non-typhoidal Salmonella (or both) can be identified using our primer sets (Table 5 and Figure 5). For example, one may wish to test all Salmonella isolates in the O serogrouping PCR, then screen serogroup A, B and D Vi+ strains using the first H typing multiplex PCR to identify Salmonella Typhi, Salmonella Paratyphi A and Salmonella Paratyphi B. The d-tartrate fermentation PCR can be performed to differentiate Salmonella Paratyphi B sensu stricto strains from Salmonella Paratyphi B Java. Any serogroup B isolates not identified by the 1st H typing PCR can be tested along with non-Typhi O serogroup D strains in the second H typing/Sdf I multiplex PCR to identify serovars Typhimurium (and related strains), Dublin (which can be Vi+ or Vi- [44]), Enteritidis and Stanleyville. The O serogroup B H:i strains can be tested using the Typhimurium/I 4,[5],12:i:- PCR to identify Salmonella Typhimurium and I 4,[5],12:i:-. It should be stressed that the O serogrouping PCR described by Levy et al. [18] needs to be performed in conjunction with the PCRs described here to ensure that Salmonella Enteritidis and Salmonella Typhimurium are identified correctly and not mistaken as Salmonella Meleagridis and Salmonella Livingstone; and Salmonella Cotham, respectively.
The surveillance experience in Mali is the first to show that Salmonella Dublin and Salmonella Stanleyville can constitute important serovars associated with invasive non-typhoidal Salmonella disease, along with Salmonella Typhimurium (and variants) and Salmonella Enteritidis. Previously, Salmonella Dublin and Salmonella Stanleyville were recovered only occasionally from blood cultures of patients in Africa [12],[13],[45],[46]. We thought it useful to be able to detect these serovars by PCR in future surveillance studies in Africa.
In conclusion, we have described a series of PCRs based on O serogrouping and H typing that can identify the causative agents of enteric fever (Salmonella Typhi and Salmonella Paratyphi A and Salmonella Paratyphi B), the three most commonly isolated serovars that cause invasive disease in young children in sub-Sahara African (Salmonella Typhimurium [and Typhimurium-like], Salmonella Enteritidis and Salmonella Dublin) and Salmonella Stanleyville, an invasive pathogen that may be of regional importance in West Africa.
|
10.1371/journal.pgen.1003762 | Two Portable Recombination Enhancers Direct Donor Choice in Fission Yeast Heterochromatin | Mating-type switching in fission yeast results from gene conversions of the active mat1 locus by heterochromatic donors. mat1 is preferentially converted by mat2-P in M cells and by mat3-M in P cells. Here, we report that donor choice is governed by two portable recombination enhancers capable of promoting use of their adjacent cassette even when they are transposed to an ectopic location within the mat2-mat3 heterochromatic domain. Cells whose silent cassettes are swapped to mat2-M mat3-P switch mating-type poorly due to a defect in directionality but cells whose recombination enhancers were transposed together with the cassette contents switched like wild type. Trans-acting mutations that impair directionality affected the wild-type and swapped cassettes in identical ways when the recombination enhancers were transposed together with their cognate cassette, showing essential regulatory steps occur through the recombination enhancers. Our observations lead to a model where heterochromatin biases competitions between the two recombination enhancers to achieve directionality.
| The state of chromatin, heterochromatin or euchromatin, affects homologous recombination in eukaryotes. We study mating-type switching in fission yeast to learn how recombination is regulated in heterochromatin. Fission yeast exists as two mating-types, P or M, determined by the allele present at the expressed mat1 locus. Genetic information for the P and M mating-types is stored in two silent heterochromatic cassettes, mat2-P and mat3-M. Cells can switch mating-type by a replication-coupled recombination event where one of the silent cassettes is used as donor to convert mat1. Mating-type switching occurs in a directional manner where mat2-P is a preferred donor in M cells and mat3-M is preferred in P cells. In this study, we investigated factors responsible for these directed recombination events. We found that two portable recombination enhancers within the heterochromatic region compete with each other and direct recombination in a cell-type specific manner. We also found that heterochromatin plays an important role in directionality by biasing competitions between the two enhancers. Our findings suggest a new model for directed recombination in a heterochromatic domain and open the field for further studies of recombination regulation in other chromatin contexts.
| Fission yeast cells switch mating type by directed recombination events where the information in the expressed mat1 locus is replaced with information copied from one of two silent loci, mat2 or mat3 (reviewed in [1]). The system allows investigating multiple facets of recombination, including effects of chromatin structure on recombination and mechanisms of donor choice: how is a particular DNA template selected for recombination when several are available in a cell?
The mat1, mat2 and mat3 loci are linked in the mating-type region (Figure 1). mat1 determines the mating type of the cell by expressing two divergent regulatory genes, Pi and Pc in P cells (mat1-P allele), Mi and Mc in M cells (mat1-M allele; [2]). Silent information for the P and M mating types is stored at respectively mat2 ∼17 kb centromere-distal to mat1, and mat3 ∼29 kb centromere-distal to mat1 [3]–[5]. The mating-type specific information at mat1, mat2 and mat3 is flanked by short homology boxes, the centromere-distal H1 box and the centromere-proximal H2 box [2]. Other elements are specific for mat2 and mat3 [2],[6],[7] (Figure 1). mat2 and mat3 are furthermore embedded in a ∼20 kb heterochromatic domain that spans the mat2-mat3 interval and extends on both sides to inverted repeat boundaries [8],[9]. This domain has been studied extensively. It provides one of the best characterized model systems for how heterochromatic regions can be established and maintained. In this domain, histones are hypoacetylated, histone H3 is methylated at lysine 9 (H3K9me) in an RNA interference-dependent manner, and chromodomain proteins of the HP1 family are associated with the modified histones [8],[10]–[15]. The HP1-like chromodomain protein Swi6 interacts with numerous protein complexes believed to modulate heterochromatin formation, gene silencing and recombination, in ways that remain to a large extent undefined in particular regarding roles in recombination [14],[16]–[19].
Interconversions of the mat1 locus between mat1-P and mat1-M lead to mating-type switching (reviewed in [1]). The conversions are coupled to DNA replication which reaches mat1 from a centromere-distal origin [20],[21]. Switching is initiated by the introduction of a strand-specific imprint in the lagging strand, resulting from the incorporation of two ribonucleotides or a nick between the mat1 H1 homology box and the mating-type specific information [20],–[28]. In the following rounds of DNA replication, the imprint is placed again on the chromatid made by lagging-strand synthesis, generating a lineage of imprinted, switchable cells [24],[29]. While lagging-strand synthesis propagates the imprinted mat1 locus in this lineage, leading-strand synthesis produces switched progeny (Figure 1B). At each division, leading-strand synthesis proceeds through the mat1 H1 homology box and stops at the imprint creating a single-ended double-strand break (DSB) or other recombinogenic molecule with a free 3′end [25],[30]. The free 3′end invades the H1 box of one of the silent loci which is then used instead of mat1 as template for leading-strand synthesis [29],[31]. This heals the break. Resolution of the recombination intermediate occurs within the H2 homology box with the help of the Swi4/8 and Swi9/10 gene products, producing a switched mat1 locus [5],[32]–[36]. The newly-switched mat1 locus does not carry an imprint hence it does not switch at the following S phase, however the chromatid made by lagging-strand synthesis acquires an imprint and starts a new lineage of switchable cells.
A choice of information is made in all switchable cells such that either mat2-P or mat3-M is used as donor to replicate and convert mat1. At this step, mat2-P and mat3-M are not picked at random. Switchable mat1-M cells preferentially use mat2-P whereas switchable mat1-P cells use mat3-M [23],[37],[38]. Coupled with the mechanism of switching outlined above, these directed choices produce a reproducible pattern of mating-type switching where (1) one out of four grand-daughters of a newly-switched cell has a high probability of having a switched mating-type (80–90%; one-in-four rule) and (2) once a cell becomes switchable the probability of recurrent switches in its lineage is very high (80–90%; recurrent-switching rule; Figure 1C).
Previous studies have revealed the importance of donor location and chromatin structure in donor choice [39]–[42]. Strains in which the mat2 and mat3 contents were swapped from mat2-P mat3-M (h90 configuration) to mat2-M mat3-P (h09 configuration) switch inefficiently to the opposite mating-type [39]. Mutations in the chromodomain protein Swi6, the H3K9 methyltransferase Clr4, or the Clr4-complex subunits Clr7 and Clr8 have opposite effects on switching in the h09 and h90 mating-type regions indicating chromatin structure favors use of mat2 in M cells and use of mat3 in P cells [39],[42]. The phenotypes of these mutants suggest that unproductive homologous switching occurs in h09 cells where mat1-M is converted by mat2-M and mat1-P by mat3-P instead of the productive heterologous switching occurring in h90 cells.
The strand exchange taking place at H1 is likely to be a determining step in donor choice. The Swi2 and Swi5 proteins are believed to facilitate this step together with Rad22, the fission yeast RAD52 homolog [43] and Rhp51, the S. pombe RecA homolog [44],[45]. The imprint, detected as a chromosomal fragile site, is formed at mat1 in swi2 and swi5 mutants but a subsequent step in the conversion process fails [5]. Consistent with this step being strand-invasion at the donor loci Swi2 interacts physically with both Swi5 and Rhp51 [17] and Swi5 facilitates Rhp51-mediated strand exchange in vitro [46]–[49]. Combined with the observation that Swi2 interacts with Swi6 [17], the properties of Swi2 and Swi5 place these factors close to the point where donor selection takes place.
A model for the directionality of mating-type switching combining effects of chromatin structure and targeted recruitment of recombination proteins was proposed in [41] (Figure 1; referred to as 2004 model below). In this model, the search for a donor starts at mat2. If the Swi2/Swi5 recombination-promoting complex (RPC) is encountered at mat2, mat2 is used to convert mat1. If RPC is not at mat2, the search proceeds to mat3 and mat3, which is constitutively associated with RPC, is used to convert mat1. The constitutive association of mat3 and RPC observed in both P and M cells is proposed to occur through a DNA element named SRE [41] that we will call SRE3 below to reflect its proximity to mat3. RPC is localized at SRE3 in P cells – ensuring that mat3 is used in P cells - but spreads from SRE3 all the way to mat2 in M cells – ensuring that mat2 is used in M cells. Spreading of RPC from SRE3 to mat2 requires Swi6. A recent addition to the model proposes that the spreading is driven by a greater abundance of the Swi2 and Swi5 proteins in M cells resulting from the positive regulation of the swi2 and swi5 genes by the M-specific transcription factor Mc [50]. Alternatively, Mc might stimulate the production of a shorter form of Swi2 expressed in P cells through alternative promoter usage [51].
The directionality model summarized above [41] provides a framework for investigations of mating-type switching, although several critical steps in it have no documented mechanism. One unexplained feature is that the search for a donor should start at mat2. This step is important because it accounts for mat2 being used in M cells when Swi2/Swi5 is present at both mat2-P and mat3-M. The model proposes that a higher-order chromatin structure helps choosing mat2 by placing it near mat1 but how this occurs remains unknown. Another aspect of the model that has not been documented experimentally is the physical interaction between SRE3 and Swi2. This is also a crucial element because the model is centered on SRE3 being the sole entry point for Swi2 in the mating-type region, accounting for the observation that Swi2 was not detected in the mating-type region of SRE3Δ strains by ChIP [41]. Here, we report further investigations on the directionality of mating-type switching bearing on these and other points. Our results challenge the notions that SRE3 is the sole entry point for Swi2, that Swi2 reaches mat2 by spreading from SRE3, and that the search for a donor starts at mat2. Instead, our results show that directionality requires two recombination enhancers, SRE2 and SRE3, whose ability to stimulate recombination in a cell-type specific manner is not tied to a specific location in the mating-type region. We present evidence that directionality results from competitions between SRE2 and SRE3, governed by cell-type specific chromatin structures.
The SRE3 element was described as the entry point at which the Swi2/Swi5 complex associates with the mating-type region [41]. Following this initial association, proposed to take place in both cell types, Swi2 could remain at SRE3 in P cells or spread to mat2 in M cells. Support for this mechanism is provided by ChIP experiments that failed to detect Swi2 anywhere in the mating-type region in strains lacking SRE3 [41]. We examined the model further through a simple genetic test. If the directed association of Swi2 with the mating-type region is abolished in SRE3Δ cells the mating-type bias in SRE3Δ cells should not be altered by deletion of swi2.
In S. pombe, the efficiency of mating-type switching can be estimated by staining sporulated colonies with iodine vapors. Efficiently-switching strains produce colonies that are stained darkly by iodine vapors because they contain many spores while poorly-switching strains produce lightly-stained colonies [52]. The predominant mating-type in cell populations can be further determined by quantifying the content of mat1 by Southern blot or PCR. In addition, we developed here a reporter system in which M cell express YFP and P cells express CFP allowing typing individual cells with a fluorescence microscope (Figure 1D). Sporulated SRE3Δ colonies were stained lightly by iodine and colonies did not stain at their junctions indicating preferential use of one donor (Figure 2B). Southern blotting, competitive PCR, and fluorescent typing all showed that SRE3Δ cells contain predominantly the mat1-P allele (P∶M = 82∶18 by Southern blot; P∶M = 88∶12 by cell count; Figure 2; competitive PCR not shown). The SRE3Δ strain used for these analyses was made in our laboratory [7] hence these results confirm the observations in [41] with an independent strain and support the conclusion of these authors that mat2-P is the preferred donor in SRE3Δ cells.
We assessed the effect of deleting swi2 in both wild-type h90 cells and SRE3Δ cells (Figure 2). Iodine staining indicated that deletion of swi2 severely affected switching efficiency in both backgrounds: h90 swi2Δ and SRE3Δ swi2Δ cells formed streaky colonies staining at their junctions showing that cells are predominantly of the M mating-type in some colonies and predominantly of the P mating type in other colonies (Figure 2B). We measured P∶M ratios in nine independent cultures of respectively h90 swi2+; h90 swi2Δ; SRE3Δ swi2+; and SRE3Δ swi2Δ cells by Southern blot (Figure 2C). While all h90 swi2+ cultures had balanced P∶M ratios and all SRE3Δ swi2+ cultures were predominantly of the P mating-type, large fluctuations in P∶M ratios were observed in h90 swi2Δ and SRE3Δ swi2Δ cultures. The strong phenotypic variability observed in h90 swi2Δ cultures disagrees with the report [41] that h90 swi2Δ cells contain predominantly mat1-P and that the switching pattern of h90 swi2Δ cells is indistinguishable from switching in h90 SRE3Δ cells. Further, the clear phenotypic differences we observed between SRE3Δ swi2+ (P≫M in all colonies; 81% P cells averaged over 9 cultures) and SRE3Δ swi2Δ strains (variegated phenotype; 40% P cells averaged over 9 cultures) is not predicted in [41]. Similarly, we observed culture-to-culture variations with, if any, a bias towards M cells in h90 swi5Δ cultures (72% M cells averaged over 9 cultures; Figure S1) in contrast to [50] who found that h90 swi5Δ cells are predominantly P. As for the deletion of swi2+, deletion of swi5+ abrogated the preferential use of mat2-P in SRE3Δ cells (Figure S1). We conclude that the RPC is necessary for the efficient and preferential choice of mat2-P in SRE3Δ cells. Since this represents a situation where RPC cannot reach mat2-P by spreading from SRE3, this result does not support the spreading model and suggests instead that other DNA element(s) or factors attract Swi2 and Swi5 independently of SRE3.
While systematically introducing deletions in the mating-type region we found that a set of nested deletions on the centromere-distal side of mat2-P affected switching, defining a ∼500 bp element adjacent to the H1 box, SRE2. Deletion of SRE2 caused a pronounced switching defect (Figure 3). Sporulated SRE2Δ colonies were stained very lightly by iodine vapors and they did not stain at their junctions; a Southern blot testing mat1 content in nine independent cultures indicated a large predominance of M cells in all cultures; and the existence of a strong bias towards M cells was also supported by fluorescence microscopy (Figure 3). Identical phenotypes were obtained when SRE2Δ colonies were seeded from P or M spores confirming efficient asymmetric switching favoring mat3-M (data not shown). The location of SRE2 relative to mat2 is similar to the location of SRE3 relative to mat3 but no extensive sequence similarities were noted between SRE2 and SRE3. Both elements are A-T rich (75% for SRE2 and 72% for SRE3 over 492 bp). The authors of a recent study [51] noticed like us that a deletion adjacent to mat2-P prevented efficient use of mat2-P however the study did not characterize the element further. Several observations reported below argue against deletion of SRE2 simply preventing use of mat2 as a donor. They suggest instead that SRE2 regulates donor choice.
As for the strains examined above, deleting swi2+ affected switching in SRE2Δ cells. Two types of sporulated colonies were observed following iodine staining, light colonies with a few dark streaks containing mostly M cells, and more rare darker colonies containing a greater proportion of P cells (Figure 3; 80% M cell averaged over nine colonies). Deletion of swi5 produced a similar phenotype (77% M cell averaged over nine colonies; Figure S1; Southern blot quantifications are summarized in Figure S2.). These phenotypes are consistent with mat3-M remaining a preferred donor in SRE2Δ cells even when RPC is not present in the cells. This again fails to support the 2004 model, where mat2-P is the preferred donor when RPC is not present due to higher-order chromatin structure. Alternatively, SRE2 might be responsible for the higher-order structure postulated by the model.
We investigated the association of Swi2 with parts of the mating-type region by ChIP (Figure S3). In unswitchable mat1-M cells, where mat2-P is perhaps poised for switching, Swi2 was detected at mat2-P and at SRE2 as previously reported [41]. In our experiments, Swi2 was also detected at these locations in SRE3Δ cells consistent with an SRE3-independent mode of recruitment to the mating-type region. This recruitment appeared facilitated by SRE2 since the association of Swi2 with mat2 was reduced in SRE2Δ cells (primer pairs 44, 46 and ‘SRE2Δ’ in M cells, Figure S3).
A deletion reducing the use of a donor cassette is not on its own evidence that the deletion removed a directionality element. We explored the possibility that SRE2 and SRE3 are genuine directionality elements by engineering h09 cells. The donor loci are mat2-M mat3-P in the h09 mating-type region [39]. The cassette contents are precisely exchanged between the H2 and H1 homology boxes placing mat2-M near SRE2 and mat3-P near SRE3. This arrangement results in inefficient switching to the opposite mating-type ([39]; Figure 4). The h09 strain provides a useful tool to study directionality since it allows designing experiments in which the tested outcome is improved switching rather than loss of switching. We tested whether swapping SRE2 and SRE3 in h09 cells improved heterologous switching.
Figure 4 shows that h09 cells with swapped elements switched mating-type very efficiently and produced populations with equal proportions of P and M cells. Their sporulated colonies were undistinguishable from h90 colonies. Their mat1 content examined by Southern blot was evenly balanced and fluorescence microscopy confirmed equal proportions of P and M cells in colonies (Figure 4). Conversely h90 cells with swapped SRE elements switched mating-type poorly, produced mainly mat1-M cells as h09 cells with unswapped elements do, and formed colonies very similar to h09 colonies (Figure 4). Together these experiments show that the PSRE2 MSRE3 combination (whether mat2-PSRE2 mat3-MSRE3 in wild-type h90 cells with native elements or mat2-MSRE3 mat3-PSRE2 in h09 cells with swapped elements) leads to balanced use of the two cassettes while the PSRE3 MSRE2 combination (whether mat2-MSRE2 mat3-PSRE3 in h09 cells with native elements or mat2-PSRE3 mat3-MSRE2 in h90 cells with swapped elements) leads to inefficient heterologous switching. We conclude from these observations that SRE2 and SRE3 behave as directionality elements responsible for the balanced heterologous switching observed in h90 cells. P cells select the cassette adjacent to SRE3 while M cells select the cassette adjacent to SRE2 and SRE2 and SRE3 can both be recognized ectopically when their location relative to mat1 has been swapped.
Should SRE2 and SRE3 be the sole determinants of directionality and should their action be fully symmetrical, h09 cells with native SRE elements and h90 cells with swapped SRE elements would engage in futile cycles where mat1-P selects mat3-PSRE3 (in h09) or mat2-PSRE3 (in h90 cells with swapped elements) and mat1-M selects mat2-MSRE2 (in h09) or mat3-MSRE2 (in h90 cells with swapped elements; Figure 4). Two types of colonies would be formed, one type containing predominantly P cells, the other predominantly M cells. This is not what is observed. Both h09 cells with native elements and h90 cells with swapped elements form populations where M cells predominate (Figure 4) indicating preferential choice of the cassette adjacent to SRE2. The fact that the bias is towards the MSRE2 cassette in both cases even though the MSRE2 cassette occupies different locations in the two strains shows that a preponderant cause for the bias is location independent.
The mechanisms responsible for directionality are likely to fail occasionally, allowing the ‘wrong’ donor to be selected. We reasoned that a small error rate would not have a strong impact on the overall composition of h90 cell populations, but the same error rate could have more profound consequences in h09 populations because the mistakes would lead to changes in mating-type that would subsequently be stably propagated through homologous switching. We modeled a situation where P cells use predominantly SRE3 to select a donor for switching, while M cells select predominantly SRE2 (Figure 4). We allowed a low occurrence of mistakes in both cell types, where P cells occasionally use SRE2 (20% of attempted switches) while M cells use SRE3 more rarely (10% of attempted switches). As expected such a bias leads to an accumulation of M cells in both h09 cells with native elements and h90 cells with swapped elements supporting the hypothesis that aberrant choices contribute to the preponderance of M cells in these strains. We note that in addition to SRE2 being more promiscuous than SRE3, the cassette content in the MSRE2 combination might facilitate use of MSRE2 over PSRE3 in P cells.
As a way of testing the extent to which P cells can use SRE2 we replaced SRE3 with SRE2 (mat2-PSRE2 mat3-MSRE2 strain referred to as 2×SRE2). The 2×SRE2 strain switched mating-type efficiently as judged from its dark iodine staining and balanced ratio of P and M cells (Figure 5). 2×SRE2 populations contained 48% P cells according to Southern blot, 56% P cells according to microscopy. The phenotype of the 2×SRE2 strain shows that P cells are proficient in the use of the SRE2 element in mat3-MSRE2 otherwise P cells would accumulate in the population of 2×SRE2 cells. To illustrate this further SRE3Δ colonies are shown near the 2×SRE2 strain for comparison in Figure 5. SRE2 at mat3-M considerably improves heterologous switching showing that P cells use SRE2. Even though 2×SRE2 cells switch mating-type efficiently, mating-type selectivity in 2×SRE2 is not as in wild-type leading us to propose that donor choice is randomized in 2×SRE2 rather than directional. A differential behavior of 2×SRE2 and wild-type mating-type region is shown for example in the next section where h90 swi6Δ and 2×SRE2 swi6Δ strains clearly differ from each other. Similarly we replaced SRE2 with SRE3 (mat2-PSRE3 mat3-MSRE3 strain referred to as 2×SRE3). The 2×SRE3 strain produced a mixture of P and M cells which shows that M cells can use SRE3, however with a bias towards M cells (Figure 5). 2×SRE3 populations contained 23% P cells as estimated from Southern blot, 25% estimated from microscopy. Together with the results presented above for the 2×SRE2 strain these ratios indicate that M cells are not as proficient at using mat2-PSRE3 as P cells are at using mat3-MSRE2. In summary SRE2 can stimulate recombination of donor loci with mat1 efficiently not only in M cells but also in P cells whereas SRE3 is more active in P cells than in M cells. The ability of each element to function in both cell types shows that these elements are not strictly dependent on cell-type-specific factors to stimulate recombination.
A remarkable aspect of mating-type switching is that the donor loci are in heterochromatin. We asked whether and how the ability of the SRE elements to stimulate recombination was affected by heterochromatin through epistasis analyses using cells lacking the chromodomain protein Swi6.
Deletion of swi6+ in h09 or h90 cells with native or swapped elements radically altered donor choice (compare Figure 4 and 6). Populations of h90 cells or h09 cells with swapped elements went from balanced P∶M ratios (49% and 50% P resp.) to containing predominantly M cells (87% and 84% resp.). Conversely the M bias in populations of h09 cells or h90 cells with swapped elements was abrogated by swi6Δ. In all cases, the changes reflected that use of the cassette adjacent to SRE2 was greatly decreased in favor of the cassette adjacent to SRE3 following deletion of swi6+, as indicated in Figure 6. These phenotypes show that Swi6 biases donor choice towards the cassette controlled by SRE2, or away from the cassette controlled by SRE3, whether the cassette contains the P or M information, and whether it is located at mat2 or mat3.
Reduced selection of SRE2 in h90 swi6Δ cells depended on the presence of SRE3 in the same cells. No change in preferred donor was observed in h90 SRE3Δ cells following deletion of swi6+, SRE2 kept being used (compare SRE3Δ in Figure 2 to SRE3Δ swi6Δ in Figure 6; mat1-P predominates in both). This indicated that SRE2 could stimulate recombination at mat2-PSRE2 in the absence of Swi6 when SRE3 was not present. Very inefficient switching in SRE2Δ SRE3Δ swi6Δ cells confirmed that the selection of mat2-PSRE2 in SRE3Δ swi6Δ cells depended on SRE2 (Figure 6; inefficient switching in the SRE2Δ SRE3Δ swi6Δ strain produces colonies staining at their junctions and large fluctuations in P/M ratios). Similarly, use of mat3-MSRE3 in SRE2Δ swi6Δ cells required SRE3 (compare SRE2Δ swi6Δ with SRE2Δ SRE3Δ swi6Δ in Figure 6). In summary these phenotypes show that both SRE2 and SRE3 can stimulate recombination in the absence of Swi6. Competitions between the two enhancers drive donor selection both in the absence and presence of Swi6. In the absence of Swi6 SRE3 is preferred over SRE2. When present, Swi6 biases donor selection towards SRE2.
Some forms of recombination occur with an extraordinary efficiency in heterochromatin as illustrated by fission yeast mating-type switching. In mating-type switching, a euchromatic locus, mat1, undergoes productive recombinogenic interactions with a heterochromatic partner in every other dividing cell. Not only are these recombination events frequent, they are also exquisitely fine-tuned such that a specific donor is selected for each conversion of mat1. Our work identifies small, portable, DNA elements responsible for donor choice and provides new insights into the mechanisms responsible for the directionality of switching. Some of our observations differ from previous reports [41],[50]. We discuss here these discrepancies and use our findings to build a new model for the directionality of mating-type switching.
Cells in which the silent-cassette contents are swapped (h09) switch mating-type inefficiently, indicating cells fail to choose heterologous donors when the donors are not at their endogenous location [39]. Here, we find that a crucial aspect of donor location is proximity of the donors to their respective recombination enhancers, SRE2 and SRE3. The determining role of SRE2 and SRE3 in donor selection was revealed by the high efficiency of switching in h09 cells when the SRE elements were swapped concomitantly with the contents of mat2 and mat3 (Figure 4). Heterologous donors could be found efficiently even when they were not at their endogenous location, provided the coupling with their cognate recombination enhancers was maintained.
The fact that h09 cells with swapped SRE elements switch well has strong implications for the 2004 model. The 2004 model is a two-component model integrating effects of donor positioning relative to mat1 (in the model the recombinogenic DSB at mat1 encounters mat2 before it encounters mat3) and presence of RPC (the first RPC-associated donor encountered is used; Figure 1). In h09 cells with swapped elements a search starting at mat2 would encounter SRE3 first. SRE3 being the proposed nucleation site for RPC, constitutively associated with RPC in both cell types, mat2-MSRE3 should be selected preferentially in both cell types which is clearly not the case. Our observations show instead that M cells choose SRE2 and P cells choose SRE3 when these elements are present, independently of their location.
One way of reconciling the portability of the SRE elements with the 2004 model is to propose that SRE2 is responsible both for the higher-order chromatin structure that brings mat2 close to mat1 in this model and also for directing the spreading of Swi2 away from SRE3 in M cells. While such roles for SRE2 should be envisioned and tested, other observations we made suggest that Swi2 does not reach mat2 by spreading from SRE3.
ChIP experiments reported in previous publications have detected large, cell-type specific variations in the association of RPC with the mating-type region [41],[50]. RPC was detected over the entire mat2-mat3 interval in M cells but the association was restricted to SRE3 in P cells [41]. In cells lacking SRE3, RPC was not detected at all [41]. While these strikingly differential associations hint to some relevance for directionality, how the associations lead to the selection of a specific donor is not straightforward. RPC associations do not on their own determine which cassette will be used since the association of RPC with SRE3 is cell-type independent. Here, we found that RPC catalyzes switching even in situations where RPC was not previously detected by ChIP [41] and in the absence of SRE3. In our experiments, the pronounced bias towards the P mating-type displayed by SRE3Δ cells was abolished in SRE3Δ swi2Δ cells and in SRE3Δ swi5Δ cells, showing RPC is necessary for the preferential use of mat2-P in SRE3Δ cells (Figure 2 and Figure S1). Not only is this epistatic relationship not predicted by the 2004 model – the model predicts that the SRE3Δ swi2Δ double mutant should switch like SRE3Δ – but the 2004 model specifically relies on swi2Δ and SRE3Δ cells having identical phenotypes, which is also contradicted by our results (Figure 2).
Based on our genetic evidence we suggest that ChIP has failed to detect interactions between Swi2 and the mating-type region that are relevant to directionality. Difficulties in detecting the association of Swi2 with the mating-type region might be due to the fact that Swi2 is not an abundant protein, that relevant interactions occur in a short window of the cell cycle, or to the fact that ChIP experiments have been conducted in switching-defective cells lacking elements at mat1 that might participate in directionality as indicated in [24]. Unlike [41], we observed that in M cells Swi2 remained associated with mat2-P and SRE2 in the absence of SRE3 (Figure S3). A core feature in the 2004 model is that Swi2 spreads from SRE3 to mat2 in P cells. Spreading of a protein along the chromatin fiber can be difficult to distinguish from other mechanisms that might lead to the same final associations. Binding at multiple sites might give the appearance of spreading from one of the sites. Here, we suggest that Swi2 does not have to spread from SRE3 to facilitate switching at SRE2.
We observed that both SRE2 and SRE3 can stimulate recombination in the absence of Swi6. While populations of SRE2Δ swi6Δ cells were predominantly M and populations of SRE3Δ swi6Δ cells were predominantly P these biases were lost in populations of SRE2Δ SRE3Δ swi6Δ cells (Figure 6C–D) showing SRE3 stimulates recombination with mat3-M in SRE2Δ swi6Δ cells and SRE2 stimulates recombination with mat2-P in SRE3Δ swi6Δ cells. We furthermore observed that competitions between SRE2 and SRE3 take place in swi6Δ cells when both elements are present. While SRE3Δ swi6Δ populations were predominantly P (Figure 6C–D), reflecting choice of SRE2, h90 swi6Δ populations were predominantly M (Figure 6A–B), reflecting choice of SRE3, from which we conclude that SRE3 outcompetes SRE2 in h90 swi6Δ cells. The switching phenotypes of h09 swi6Δ; h09 with swapped elements swi6Δ; and h90 with swapped elements swi6Δ cells all show that SRE3 is preferred over SRE2 in swi6Δ cells when both elements are present (Figure 6A–B).
Swi6 exerts major effects on mating-type switching through SRE2 and SRE3. Comparing Figure 4 and Figure 6A–B shows that Swi6 tilts the relative efficiency of the two elements, allowing SRE2 to be preferred over SRE3 in M cells. We suggest that this effect is key to directionality. Several lines of evidence have established that heterochromatin differs in the mating-type region of P and M cells making heterochromatin a good candidate for providing cell-type specificity in mating-type switching. Ectopic reporters are more strongly repressed in M cells than in P cells, whether the reporters are near mat2 or mat3 [7],[53], (G. Thon unpublished data) and consistently more Swi6 is detected over the entire mat2-mat3 region in M cells than in P cells [8]. These differences between P and M cells are likely to reflect differential associations of various protein complexes over the entire mating-type region, including but not limited to Swi6, Swi2 and Swi5 [14],[16]–[19],[54]. Global changes over the entire region would account for our observation that the effects of Swi6 on SRE2 and SRE3 were independent of donor location (Figure 6). The model for the directionality of switching outlined below proposes that differences in the chromatin structure of P and M cells determine which recombination enhancer is used in each cell type.
We propose a simple model for the directionality of mating-type switching that takes into account our observations (Figure 7). This model is an alternative to models where the recombination enhancers favor cell-type specific interactions between the donor loci and mat1 through DNA looping but it is not incompatible with looping models.
In the proposed model SRE2 and SRE3 compete to capture the free DNA end generated at mat1. When Swi6 and associated factors are in comparatively low abundance in the mating-type region as is the case in P cells, SRE3 stimulates recombination at its adjacent H1 homology box more efficiently than SRE2. When Swi6 and associated factors are in greater abundance in the mating-type region, as is the case in M cells, SRE2 is more efficient than SRE3.
Several mechanisms can be envisioned for how SRE2 and SRE3 might facilitate strand invasion at their adjacent H1 box in a chromatin-dependent manner. SRE2 and SRE3 might have an intrinsic ability to facilitate D-loop formation as suggested by their low melting temperature (predicted from 72–75% AT content and data not shown). Indeed, evidence has been presented that SRE2 can form a heteroduplex with DNA adjacent to the mat1 H1 box [33]. Swi6 could modulate the ability of SRE2 and SRE3 to stimulate strand-invasion at H1 through changes in chromatin structure. Swi6 binds to nucleosomes methylated at H3K9 and it oligomerizes. The association of Swi6 with chromatin per se might constrain the topology of DNA around H1 and the SRE elements in a way that would alter D-loop induction by the SRE elements and depend on the concentration of Swi6. Other, non mutually-exclusive effects of SRE2 and SRE3 could be through the positioning of nucleosomes. Swi6 might induce the local sliding or displacement of nucleosomes through one of its associated ATP-dependent chromatin remodeling complexes (RSC, Ino80, FACT; [18],[19]) thereby altering the ability of a recombination enhancer to increase H1 accessibility. Finally, direct interactions might take place between the recombination enhancers and recombination factors such as Swi2 as suggested in the case of Swi2 and SRE3 [41]. Directionality would occur if SRE3 had a higher affinity for Swi2 than SRE2 but a lower peak efficiency than SRE2 when stimulating recombination in the context of heterochromatin. At low concentration of Swi2, SRE3 but not SRE2 would be associated with Swi2, promoting invasion of its adjacent cassette. At high concentrations of Swi2, SRE2 would not only be associated with Swi2 but it would use its associated Swi2 more efficiently than SRE3, leading to preferred choice of SRE2 over SRE3. Low association of Swi6 and Swi2 in the mating-type region of P cells would promote invasion of the SRE3-adjacent cassette. High association of Swi6 and Swi2 in the mating-type region of M cells would promote invasion of the SRE2-adjacent cassette.
How pre-existing chromatin structures affect recombination and DSB repair is poorly understood in spite of a great relevance for the maintenance of genome integrity in all eukaryotes. Competitions between donors for gene conversions [55],[56] and regional, cell-type specific, control of recombination [57]–[59] were observed in S. cerevisiae similar to what we observed here. Indeed, much of our knowledge on the effects of chromatin on recombination was acquired using S. cerevisiae [59]–[61]. Our characterization of the fission yeast SRE elements opens the field for further in vivo and in vitro studies of recombination regulation in other chromatin contexts.
Standard procedures were used to construct and examine S. pombe strains. The details of the strain constructions, Southern blots and microscopy are presented in Text S1 (Extended experimental procedures). Strain genotypes are listed in Table S1 and oligonucleotide sequences in Table S2.
|
10.1371/journal.pbio.2007097 | Serotonin receptor HTR6-mediated mTORC1 signaling regulates dietary restriction–induced memory enhancement | Dietary restriction (DR; sometimes called calorie restriction) has profound beneficial effects on physiological, psychological, and behavioral outcomes in animals and in humans. We have explored the molecular mechanism of DR-induced memory enhancement and demonstrate that dietary tryptophan—a precursor amino acid for serotonin biosynthesis in the brain—and serotonin receptor 5-hydroxytryptamine receptor 6 (HTR6) are crucial in mediating this process. We show that HTR6 inactivation diminishes DR-induced neurological alterations, including reduced dendritic complexity, increased spine density, and enhanced long-term potentiation (LTP) in hippocampal neurons. Moreover, we find that HTR6-mediated mechanistic target of rapamycin complex 1 (mTORC1) signaling is involved in DR-induced memory improvement. Our results suggest that the HTR6-mediated mTORC1 pathway may function as a nutrient sensor in hippocampal neurons to couple memory performance to dietary intake.
| Optimized dietary intake is crucial for maintaining cognitive performance. A mild reduction (between 20% and 40%) in food intake—called dietary restriction (DR) or calorie restriction—has been shown to extend life span and to improve cognitive ability in various species through a mechanism that is not fully understood. Here, we investigate the nutritional basis of DR and find that protein—in particular a single amino acid, tryptophan—is important in limiting the improved memory performance induced by DR. We also observe that DR induces reduced serotoninergic signaling through the suppression of serotonin receptor–mediated mechanistic target of rapamycin complex 1 (mTORC1) signaling in the mouse hippocampus. Whereas genetic and pharmacological inhibition of the 5-hydroxytryptamine receptor 6 (HTR6)–mTORC1 pathway mimics DR-induced structural and electrophysiological changes that are associated with memory enhancement, activation of this pathway diminishes these effects. Our results provide a mechanistic connection linking HTR6 and the mTORC1 pathway that mediates the favorable adaptation of improved memory to reduced dietary intake.
| Nutritional status is closely linked to cognitive performance. Whereas high-calorie intake increases the risk for neurodegenerative diseases [1, 2], food shortage can disable costly memory formation in order to favor survival [3]. An adequate but reduced dietary intake, such as dietary restriction (DR; 20%–40% reduction in total daily caloric intake without malnutrition), has been recognized to be the most effective anti-aging intervention, not only improving cognitive performance in elderly humans but also prolonging healthy life span in several model organisms [4, 5]. Studies investigating the nutritional basis of DR benefits have revealed that reduced dietary intake of protein as well as of certain amino acids, including tryptophan, can improve surgical stress resistance and extend life span in rodents [6, 7]. Although the underlying mechanism remains elusive, altered serotonergic signaling is thought to be involved.
Serotonin receptor 5-hydroxytryptamine receptor 6 (HTR6) has been shown to regulate neuronal migration and differentiation during development [8–10] and is also implicated in mental disorders, such as anxiety and depression [11]. HTR6 stimulates Gs and adenylyl cyclase, which are generally known to have a positive influence on cognitive functions [12]. However, accumulating evidence in both rodent and human studies suggests that pharmacological blockade of HTR6 signaling improves memory performance [13–18]. This discrepancy has highlighted alternative pathways that mediate the procognitive effect of HTR6 inhibition. Despite an enhanced corticolimbic release of acetylcholine, glutamate, and monoamines that favors cognitive processes [19], the disrupted recruitment of mechanistic target of rapamycin (mTOR) signaling that occurs upon HTR6 inhibition is postulated as a mechanism mitigating cognitive deficits in animal models of schizophrenia [20].
The mTOR pathway has been shown to integrate signals from nutrients and growth factors, and it further initiates downstream pathways via two distinct protein complexes, mTOR complex 1 (mTORC1) and mTOR complex 2 (mTORC2), which regulate various cellular processes related to growth, differentiation, and metabolic homeostasis [21]. Through its inhibition of the mTOR pathway, rapamycin is one of a few molecules that have been demonstrated to promote memory performance and extend life span in animals [22, 23]. The interplay between dietary manipulation and neuronal control of memory performance has not been thoroughly investigated. We therefore sought to examine the effects of DR on memory performance, and our results demonstrate that dietary tryptophan is a major contributor in limiting DR-induced memory enhancement. The serotonin receptor HTR6 is indispensable for this process, through its modulation of downstream mTORC1 signaling. Our findings thus establish a mechanistic connection between DR and improved memory performance.
To explore the molecular mechanism underlying DR regulation of memory function, we began by examining the changes in memory performance of young mice (4 months old, Fig 1A–1F) and aged mice (24 months old, S1A–S1F Fig) that had been fed a normal (ad libitum [AL]) or DR diet for 8 weeks. DR mice were fed once a day with a daily food allotment that was 60% of that eaten by the AL animals (Fig 1A and 1B, S1A and S1B Fig). Mice on a DR diet exhibited reduced body weight (Fig 1C and S1C Fig) but did not show noticeable changes in general locomotor activity compared to the AL group in the open field test (Fig 1D and S1D Fig).
Using the novel object recognition (NOR) test, we evaluated the recognition memory of mice at 1 hour or 24 hours after a training session, for both short-term and long-term memory, respectively. The NOR test is based on the innate preference of animals for exploring a novel object over a familiar one, and it is independent of emotional cues [24]. Animals achieving a higher object discrimination index have enhanced memory performance. We found that both young AL and DR mice showed significant memory retention 1 hour after the training session (Fig 1E and 1F). However, young AL mice showed significant memory decline 24 hours after the training session, whereas young DR mice showed sustained memory retention, suggesting that DR mice have improved long-term memory formation (Fig 1E and 1F). Although we also found similar effects of DR on aged mice (S1E and S1F Fig), we focused on young animals for the subsequent mechanistic studies.
We reasoned that appetite might contribute to DR-induced memory enhancement, since DR mice were fed once a day and were fed following training and memory sessions on the days of behavioral tests. To address this issue, we performed 18-hour acute fasting on AL mice and fed the DR mice 2 hours prior to the training session (Fig 1G). Neither 18-hour acute fasting in AL mice nor 2-hours-prior prefeeding in DR mice affected general locomotor activity (Fig 1H) or memory performance (Fig 1I), compared to normal AL and DR mice, respectively. Moreover, we found that the DR-induced memory enhancement required at least 2 weeks of dietary manipulation and that memory performance returned to the normal AL level within 2 weeks following a shift of DR mice to an AL diet (Fig 1J–1L). Thus, chronic DR is essential for enhanced memory performance.
The major sources of calories in standard mouse food are carbohydrates, proteins, and fats. To identify which of these nutrients from the diet contributes to DR-induced memory enhancement, we compared the recognition memory of mice under DR and under DR supplemented with carbohydrate, protein, or fat to a level equivalent to that of AL feeding (S2 Table). We found that adding back carbohydrate or fat did not affect DR-induced memory retention of mice in the NOR test (Fig 2A), indicating that these constituents do not limit memory performance during DR. In contrast, addition of protein to the DR diet attenuated DR-induced memory performance, bringing it back to a level comparable to that for AL mice (Fig 2A). We next determined which amino acid affected memory function in this context. Glutamate was tested, since it is an amino acid neurotransmitter; tryptophan, tyrosine, and cysteine are precursors for serotonin, catecholamine, and sulfur-containing amino acid biosynthesis in the brain, all of which are known to regulate cognitive function [25, 26]. Our results demonstrated that adding back tryptophan, but not glutamate, cysteine, or tyrosine, attenuated DR-induced memory retention (Fig 2B). Tryptophan alone is thus limiting for memory retention during DR. Although adding back carbohydrate and fat, but not protein or individual amino acids, significantly increased the body weight of DR mice (S2A and S2D Fig), none of these dietary manipulations affected general locomotor activity or total time spent on object exploration, compared to either the AL or DR group in the open field test or NOR test, respectively (S2B, S2C, S2E and S2F Fig). These observations implicated dietary tryptophan in modulation of memory processing, through altered serotonergic signaling in the brain. To test this hypothesis, animals were given an intraperitoneal (IP) injection of fenfluramine, a serotonin-releasing agent [27], approximately 30 minutes prior to the NOR training session. We found that fenfluramine abolished DR-induced memory enhancement (Fig 2C), suggesting that DR may induce reduced serotonergic signaling and that an acute increase in serotonin transmission interferes with DR improvement of memory performance.
To investigate whether DR could indeed affect serotonergic signaling, we measured the levels of serotonin (5-hydroxytryptamine [5-HT]), serotonin metabolite 5-hydroxyindoleacetic acid (5-HIAA), and mRNA expression of serotonin receptors in mouse hippocampal tissues. We found that while DR induced a trend toward reduced serotonin levels, it significantly down-regulated the levels of 5-HIAA in the mouse hippocampus (Fig 2D). The 5-HIAA/5-HT ratio was therefore lower, implying reduced serotonergic activity in the DR mouse hippocampus (Fig 2E). We found that mRNA expression of serotonin receptor HTR6 was significantly lower among all HTRs examined (Fig 2F) and that reduced HTR6 protein expression was also observed in the hippocampus of DR mice (Fig 2G and 2H). Similar results were found in other brain regions involved in memory formation, such as prefrontal cortex (S3A–S3C Fig). DR-induced down-regulation of HTR6 mRNA expression is likely to work through an elevated level of circulating corticosterone, as suggested in previous studies [28, 29]. In agreement with this, we found a higher level of serum corticosterone in DR (compared with AL) mice (Fig 2I), and chronic supplementation with a low dose of corticosterone (10 μg/ml in drinking water for 6 weeks) significantly repressed hippocampal HTR6 mRNA expression (Fig 2J) and improved memory performance (Fig 2K).
We further examined HTR6 involvement in DR-induced memory enhancement in mice by performing IP injection of an HTR6-specific agonist (WAY 208466 dihydrochloride [WAY]) or antagonist (SB 399885 hydrochloride [SB]) approximately 30 minutes before the NOR training session. We found that IP injection of the HTR6 agonist abrogated DR-induced memory enhancement, whereas IP injection of the HTR6 antagonist did not further enhance memory performance of DR mice (Fig 2L and 2M). Moreover, HTR6 antagonist administration alone improved the memory performance of AL mice (Fig 2M), suggesting that HTR6 blockade is beneficial for memory formation. In addition, we obtained HTR6 knockout (KO) mice and confirmed that the HTR6 transcript and protein were undetectable in the hippocampal tissues (S4A and S4B Fig). mRNA expression of the other known HTRs was unchanged in the HTR6 KO mice, indicating absence of any compensatory effect by the other HTRs (S4A Fig). When performing behavioral analyses, we found that HTR6 KO mice showed a higher memory performance under both AL and DR conditions, an improvement comparable to that seen in the wild-type (WT) DR mice, and that the HTR6 KO mice did not exhibit the tryptophan supplementation–or fenfluramine-induced memory attenuation seen in WT DR mice (Fig 2S and 2V). Our data show that the profoundly enhanced memory performance of HTR6 KO mice cannot be attributed to alterations in food intake, body weight, or physical activity (Fig 2N–2R, 2T and 2U). HTR6 KO–induced memory enhancement was also not associated with peripheral glucose metabolism (S4C and S4D Fig) or ketogenesis (S4E Fig), which are known to be involved in DR- or intermittent fasting–induced longevity and neuroplasticity [30, 31].
Two independent gene expression databases, the Allen Brain Atlas of mRNA in situ (www.brainatlas.org) and Gene Expression Nervous System Atlas (GENSAT) of bacterial artificial chromosome–enhanced green fluorescent protein (BAC–eGFP) transgenic mice (www.gensat.org) [32, 33], both indicate that HTR6 is highly expressed in the mouse hippocampus (S5A and S5B Fig). We therefore examined the neuronal morphology of the mouse hippocampus under dietary or genetic manipulations. We used the Golgi-Cox impregnation method and reconstructed the dendritic profile using Neurolucida software. Overall, we found that DR reduced the complexity and dendritic length of the CA1 pyramidal neurons (Fig 3A–3E, 3H and 3I) and dentate gyrus (DG) granule cells (S5C–S5E Fig) of the mouse hippocampus. The spine density of both CA1 pyramidal neurons (Fig 3F, 3G, 3J and 3K) and DG granule cells (S5F and S5G Fig) was significantly increased in the DR mice. However, these DR-induced structural alterations were not observed in the HTR6 KO mice (Fig 3A–3K, and S5C–S5G Fig), suggesting that HTR6 is required for the observed DR-induced structural alterations in hippocampal neurons. Neither DR nor HTR6 KO induced any changes in the neuronal density (Fig 3L and 3M, and S5H and S5I Fig).
Long-term potentiation (LTP) is a well-recognized synaptic plasticity that reflects higher-order brain functions such as memory. We performed field recordings of Schaffer collateral-CA1 synapses in the hippocampus and found that DR mice had a significantly enhanced LTP compared to AL mice (Fig 4A–4C). This effect was also achieved in AL mice receiving chronic supplementation with corticosterone (Fig 4D–4F). These mice showed improved memory performance similar to that seen in the DR mice (Fig 2K). DR-induced LTP enhancement, however, was attenuated by bath application of the HTR6 agonist during the recordings, and this attenuation was prevented when both the HTR6 agonist and antagonist were present during the recordings (Fig 4A–4C). Consistent with this observation, HTR6 KO mice also exhibited a higher magnitude of LTP, similar to that of DR mice, and DR did not further enhance LTP in HTR6 KO mice (Fig 4G–4I). The DR-induced LTP enhancement seen in both WT and HTR6 KO mice was independent of basal synaptic transmission, as input–output curves obtained in hippocampal slices were similar for AL and DR groups (S5J Fig). We also performed a rescue experiment in the HTR6 KO mice by bilateral injection of HTR6–green fluorescent protein (GFP) or GFP plasmid into the CA1 region of the hippocampus (Fig 4J–4L). We confirmed that HTR6 mRNA was reexpressed in the hippocampus of the HTR6 KO mice (Fig 4M), and the LTP level measured in HTR6–GFP–transfected hippocampal slices from HTR6 KO mice returned to a lower level, similar to that seen in WT AL mice (Fig 4A–4C and 4N–4P). These data together suggest that HTR6 may act downstream of DR to mediate DR-induced synaptic plasticity.
HTR6 is a Gs-coupled receptor that activates cAMP production upon serotonin stimulation [34]. To identify the downstream effector mediating HTR6 KO–induced memory enhancement, we first examined the cAMP–protein kinase A (PKA)–cAMP-responsive element-binding protein 1 (CREB-1) axis, known to regulate synaptic plasticity and memory [35], in mouse hippocampal tissue. In agreement with previous hypotheses, we found that DR significantly reduced PKA phosphorylation but increased CREB-1 phosphorylation (Fig 5A, 5D and 5E) in mouse hippocampal tissues [36, 37]. However, we observed normal levels of PKA and CREB-1 phosphorylation in the hippocampal tissues of HTR6 KO mice compared to WT mice (Fig 5B, 5H and 5I), suggesting that PKA and CREB-1 activities are uncoupled from HTR6 KO–induced memory enhancement. Since activation of HTR6 has also been shown to recruit and regulate mTOR signaling in transfected human embryonic kidney cells [20], we therefore further explored the activity of mTOR pathways in mouse hippocampal tissues. We found that hippocampal tissue from both DR and HTR6 KO mice showed reduced S6 kinase (S6K; downstream of mTORC1) phosphorylation, but not Akt (downstream of mTORC2) phosphorylation, compared to tissue from AL or WT mice (Fig 5A, 5B, 5F, 5G, 5J and 5K). Additionally, DR did not further reduce S6K phosphorylation in the HTR6 KO mice (Fig 5C and 5L), indicating that DR and HTR6 affect mTORC1 activity through a common pathway. To demonstrate whether mTORC1 signaling may mediate DR- and HTR6 KO–induced memory enhancement in mice, we performed behavioral tests and found that supplementation of food with an mTORC1 activator (phosphatidic acid [PA]) [38] attenuated DR- and HTR6 KO–induced memory performance (Fig 5M). On the other hand, supplementation of food with an mTORC1 inhibitor (everolimus, rapamycin analog [RA]) mimicked but did not further enhance the memory performance of DR and HTR6 KO mice (Fig 5M). Supplementation with the mTORC1 inhibitor also prevented mTORC1 activator–induced memory impairment in the DR mice (Fig 5M). Neither mTORC1 activator nor inhibitor treatment affected feeding behavior, body weight, or locomotor activity of mice (S6A–S6C Fig).
Similar to the results in DR mice, WT AL mice fed a diet supplemented with an mTORC1 inhibitor also showed reduced dendritic complexity, reduced dendritic length, increased spine density, and normal neuronal density of the CA1 pyramidal neurons (Fig 6A–6G). However, these mTORC1 inhibitor–induced structural alterations were not observed in the HTR6 KO mice (Fig 6A–6F), suggesting that reduced HTR6-mediated mTORC1 signaling is essential for imitating the DR-induced structural alterations in hippocampal neurons.
In this study, we examined the nutritional basis and mechanistic regulation of DR in enhancing normal brain function. We found that a chronic and constant regimen of DR can effectively improve memory performance of mice through negative modulation of HTR6- and mTORC1-mediated serotonergic signaling. Our data show that tryptophan supplementation limits DR-induced memory enhancement, and it is therefore reasonable to assume that reduced serotonergic signaling in the brain of DR mice may be responsible for this effect. Although we observed only a trend toward low levels of 5-HT or 5-HIAA in DR brain tissues, a significantly reduced 5-HIAA/5-HT ratio could indicate a dampened serotonin turnover rate, a reflection of reduced serotonergic activity in the brain [39]. These results are consistent with previous findings, which showed that DR suppresses serotonergic activity in the brain [40, 41]. Our data also showed that fenfluramine administration abolished the beneficial effect of DR on memory performance of mice, further implicating reduced serotonergic signaling in the DR brain tissues. This hypothesis is strongly supported by our findings of reduced HTR6-mediated mTORC1 signaling in the hippocampal tissue of DR mice. Extracellular and intracellular levels of serotonin will need to be measured in future studies in order to confirm the role of DR in serotonin metabolism in the brain.
Current opinion regards DR as a form of intermittent metabolic switching (IMS) in which the brain experiences transitional cycles of utilizing carbohydrates and ketones as major energy sources. Although the mechanisms for IMS-induced neuroplasticity remain to be established, both peripheral circulating signals and intrinsic neuronal network pathways are proposed [31]. Our results suggest that the elevated level of circulating corticosterone, but not altered glucose metabolism or ketogenesis, mediates DR-induced memory enhancement through repressed hippocampal HTR6 expression and consequently reduced mTORC1 signaling. The brain is considered to be an important target for corticosterone, since two types of receptors, the type I high-affinity mineralocorticoid receptor (MR) and the type II lower-affinity glucocorticoid receptor (GR), are highly expressed in the hippocampus and many other brain regions that are involved in multiple cognitive processes [42, 43]. Previous studies have demonstrated a biphasic effect of corticosterone on cognitive function. Whereas enhanced memory and LTP occur when the level of corticosterone is mildly increased (predominantly MR activation), impaired memory and LTP appear when the corticosterone level is greatly increased (both MR and GR activation) [44–46]. This biphasic effect of corticosterone is also observed in the regulation of hippocampal neurogenesis, which is known to be associated with DR-induced memory formation [47, 48]. A low dose of corticosterone treatment, similar to that used in our study, does not induce stress responses in mice [49], but the improved memory performance and LTP observed in our study are largely in agreement with previous studies discussed above.
HTR6 was originally identified as a Gs-coupled receptor that activates cAMP production upon serotonin stimulation [34], and recent characterization of intracellular binding partners for HTR6 have also revealed other ligand-dependent and ligand-independent pathways that regulate a number of cellular functions [10, 20, 50]. Our data are in line with the notion that mTORC1 functions as an alternate downstream effector of HTR6 [20] and that this pathway may further regulate structural alteration and neuronal plasticity in aiding memory performance. A possible linkage between DR, HTR6, and the cAMP–PKA–CREB-1 axis in memory regulation is suggested by a previous study demonstrating a role for CREB-1 in mediating DR-induced neuronal plasticity, memory, and social behavior [36]. However, the concept of an up-regulated HTR6–cAMP–PKA–CREB-1 axis in memory enhancement contradicts the generally accepted idea that HTR6 inactivation is promnemonic [13–20]. Our findings that chronic DR reduced PKA phosphorylation but increased CREB-1 phosphorylation in mouse hippocampal tissue indicate that PKA and CREB-1 have more dynamic interactions during memory formation that require further investigation. The key feature of the cAMP–PKA–CREB-1 pathway is its transient activation in response to stimulation of Gs-coupled receptors, which regulate transcription, with rates peaking between 30 minutes and 1 hour [51, 52]. In this study, we measured the steady state of PKA and CREB-1 phosphorylation following chronic dietary manipulation in both WT and HTR6 KO mice. These data do not reflect acute cellular responses of PKA and CREB-1 phosphorylation during memory behaviors. A lower intrinsic PKA phosphorylation level in the brain tissues of DR mice could imply a higher efficacy in activating downstream signaling molecules upon transient stimulation, which may be critical during memory formation in living animals. In addition to the PKA pathway, CREB-1 can also be activated through other kinases to mediate neuronal activity, growth factor signaling, and stress responses [53–58]. It is of particular interest that DR up-regulates N-methyl D-aspartate receptor and brain-derived neurotrophic factor/tropomyosin receptor kinase B signaling, which may result in the increased CREB-1 activity observed in the brain of the DR mice [30, 59]. Accordingly, creating an in vitro cell culture environment mimicking chronic DR and a real-time monitor for protein (kinase) activity during memory behaviors will be important for future studies to establish which signaling networks are influenced by DR.
mTOR exerts a critical role in the regulation of dendritic protein synthesis, which is essential for long-lasting synaptic plasticity [60]. Current concepts of neuroplasticity and memory regulation refer to mTOR as a rheostat rather than an on–off switch. Whereas acute and complete inhibition of mTOR abolishes synaptic plasticity, chronic partial reduction of mTOR signaling may result in DR-mimicking effects, leading to the enhanced memory performance in animals observed in this study and consistent with previous reports [61, 62]. At a structural level, mTOR is involved in the regulation of dendritic formation and axon elongation, as well as synaptic pruning, all of which are critical for normal brain development [63, 64]. Our analyses of Golgi staining revealed HTR6-dependent structural alterations in DR hippocampal neurons, including reduced dendritic complexity and dendritic length but increased spine density. These results suggest that nutritional restriction, such as DR, with reduced HTR6-mediated mTOR signaling may minimize the dendritic size and complexity of neurons but that increased spine density may compensate for these changes by permitting more efficient communication among neurons. It is also worth noting that HTR6 has been shown to regulate neuronal differentiation through constitutive interaction with cyclin-dependent kinase 5 (Cdk5), which is known to control cytoskeletal dynamics involved in dendritic spine morphogenesis and neurite growth, as well as neuronal migration [65]. Although examination of the role of HTR6 in neuronal morphogenesis is not our primary focus, the current findings provide added evidence for a connection between diet and dendritic organization in neurons. Future examination of the interactions among HTR6, mTOR, and Cdk5 would certainly broaden our understanding of nutritional control in dendritic arborization and spine formation.
In summary, we propose a mechanism that explains DR-induced memory enhancement and identify serotonin receptor HTR6, in association with the mTORC1 pathway, as playing a pivotal role in this process (Fig 6H). Our dietary, pharmacological, and genetic manipulations point to attenuated HTR6-mediated mTORC1 signaling in the brain of DR mice, and our results show that interventions that interfere with this pathway compromise the favorable adaptation of memory functions to reduced dietary intake. These results are also supported by previous findings of increased hippocampal spine density and LTP formation in DR animals [36, 66], even in the presence of reduced dendritic complexity and dendritic length, as observed in this study.
All experimental protocols followed local animal ethics regulations and were approved by the National Taiwan University College of Medicine and College of Public Health Institutional Animal Care and Use Committee (approval no. 20120262).
C57BL/6 mice were obtained from the Laboratory Animal Center, National Taiwan University College of Medicine. HTR6 KO mice (B6;129S5-Htr6tm1Lex/Mmucd) were obtained from the Mutant Mouse Resource & Research Centers at University of California, Davis and were backcrossed to the C57BL/6 mouse background for 10 generations. Each mouse was genotyped using gene-specific primers (S1 Table) for polymerase chain reaction (PCR) and gel electrophoresis as described in our previous study [67]. Male mice were used in this study, and all mice were group-housed (2–5 mice per cage) and maintained in an animal room at a controlled temperature of 22–24 °C and 50%–55% humidity, under a 12-hour light/dark cycle. Mice were fed once per day, at the beginning of the dark phase, with purified rodent diet AIN-93G powder (MP Biomedicals) supplemented with different nutrients as indicated in each experiment (S2 Table). PA (30 g/kg; Avanti) and RA (15 mg/kg; everolimus, Tocris Bioscience) were added to AIN-93G as a daily diet for some experiments. Corticosterone (10 μg/ml, Sigma) was added to the drinking water of mice. Food intake and change in body weight of mice were monitored regularly. All of the following behavioral, morphological, and electrophysiological analyses were done blind with respect to the diet and genotype of the mice.
All behavioral tests were performed in the dark phase. The open field test and NOR test were performed as described previously [68, 69]. We calculated the object discrimination index, in order to measure the memory performance of mice, by subtracting the time spent on exploring the familiar object from the time spent on exploring the novel object and dividing by total time spent exploring both objects. Fenfluramine (5 mg/kg, Sigma), WAY (10 mg/kg; Tocris Bioscience), and SB (10 mg/kg; Tocris Bioscience) were IP injected into mice 30 minutes before the NOR training session.
Total RNA was prepared from hippocampal tissue of each mouse using the NucleoSpin RNA Kit (Macherey-Nagel), and cDNA was prepared using oligo-d(T)15 (Invitrogen) and SuperScript III reverse transcriptase (Invitrogen), as described previously [70]. Quantitative PCR was carried out using a StepOnePlus Real-Time PCR System (Applied Biosystems), SYBR Green Master Mix (Fermentas), and gene-specific primers (S1 Table).
Four-month-old mice were fasted for 6 hours, and blood glucose concentration was measured at 0, 30, 60, 90, 120, and 180 minutes following an IP injection of glucose (2 g/kg, Sigma) or insulin (0.75 unit/kg, Sigma). Blood samples from nonfasted and 6-hour-fasted mice were collected for beta-hydroxybutyrate measurements. The concentrations of blood glucose and beta-hydroxybutyrate were measured using FreeStyle Optium Neo Blood Glucose and Ketone Monitoring System (Abbott Diabetes Care).
Serotonin (Abcam) and 5-HIAA (BioVision) concentration in the mouse brain tissues were measured using enzyme-linked immunosorbent assays, following the manufacturer’s instructions. Total protein level was quantified using the Bradford protein assay (Bio-Rad). Corticosterone concentration in the mouse serum samples was determined using a corticosterone enzyme-linked immunosorbent assay, following the manufacturer’s instructions (Enzo Life Sciences). Blood samples were collected during the dark phase (around 6 to 8 PM).
Morphologic features of CA1 pyramidal neurons and granule cells in the DG were visualized using the FD Rapid GolgiStain kit following the manufacturer’s protocol (FD NeuroTechnologies). Dendritic morphology and spine density were reconstructed and analyzed using Neurolucida software (MBF, Bioscience). For neuronal density analysis, 7-μm coronal sections of mouse brain were stained with 0.1% cresyl violet, and every 20th section from dorsal to ventral hippocampus was examined using a 63x oil-immersion objective lens on a photomicroscope (Zeiss Axio Imager 2). The neuronal density of the CA1 pyramidal neurons and DG granule cells was quantified using Image J software.
Extracellular recordings of field excitatory postsynaptic potentials (fEPSPs) at Schaffer collateral-CA1 synapses in mouse hippocampal slices were performed with a MED64 multichannel recording system equipped with a data acquisition and analysis program (Alpha MED sciences), as described in our previous study [71]. LTP was induced by tetanic stimulation (TS) at 100 Hz for 1 second. In each slice, fEPSPs were monitored for at least 30 minutes to obtain stable fEPSPs. The slopes of fEPSPs recorded for the following 10 minutes were averaged and taken as the baseline. LTP magnitudes were obtained by the average slope of the least 20 fEPSPs (10 minutes) recorded following TS and expressed as the percentage of baseline fEPSP slope. Brain slices were treated with saline, 20 μM WAY, and/or 30 μM SB 10 minutes prior to TS. For rescue experiments in the HTR6 KO mice, pCMV–GFP or pCMV–HTR6–GFP plasmid (0.5 μg) was mixed with BrainFectIN transfection regent and bilaterally injected into the CA1 regions of the mouse hippocampus (ML: ±1.5 mm, AP: −2 mm, DV: −1.5 mm). Murine HTR6 was cloned into pCMV–GFP, a gift from Connie Cepko (Addgene plasmid #11153). LTP measurements were performed 5–7 days after transfection.
Brain tissues were lysed in radioimmunoprecipitation assay buffer (Thermo Fisher Scientific). Proteins were then separated by SDS-PAGE and transferred to PVDF membranes (Millipore) using standard procedures [72]. The antibodies used were rabbit anti-HTR6 (1:500, Abcam #ab103016), rabbit anti-phospho-S6K (Thr389, 1:1,000, Cell Signaling Technology #9205), rabbit anti-S6K (1:1,000, Cell Signaling Technology #2708), rabbit anti-phospho-Akt (Ser473, 1:1,000, Cell Signaling Technology #9271), rabbit anti-Akt (1:1,000, Cell Signaling Technology #9272), rabbit anti-phospho-PKA (Thr197, 1:1,000, Cell Signaling Technology #4781), rabbit anti-PKA (1:1,000, Cell Signaling Technology #4782), rabbit anti-phospho-CREB (Ser133, 1:1,000, Millipore #06–519), rabbit anti-CREB (1:1,000, Millipore #AB3006), and mouse anti-α tubulin (1:10,000, GeneTex #GTX628802). Protein signals were visualized with horseradish peroxidase–conjugated secondary antibodies and ECL reagent (Thermo Fisher Scientific). Quantification of immunoblots was conducted with Image J software.
All data are expressed as mean ± SEM and were examined by Student t test, one-way ANOVA, or two-way ANOVA with Fisher’s LSD post hoc test (StatPlus:mac). The statistical details of experiments can be found in the figure legends.
|
10.1371/journal.pgen.1004003 | The CCR4-NOT Complex Mediates Deadenylation and Degradation of Stem Cell mRNAs and Promotes Planarian Stem Cell Differentiation | Post-transcriptional regulatory mechanisms are of fundamental importance to form robust genetic networks, but their roles in stem cell pluripotency remain poorly understood. Here, we use freshwater planarians as a model system to investigate this and uncover a role for CCR4-NOT mediated deadenylation of mRNAs in stem cell differentiation. Planarian adult stem cells, the so-called neoblasts, drive the almost unlimited regenerative capabilities of planarians and allow their ongoing homeostatic tissue turnover. While many genes have been demonstrated to be required for these processes, currently almost no mechanistic insight is available into their regulation. We show that knockdown of planarian Not1, the CCR4-NOT deadenylating complex scaffolding subunit, abrogates regeneration and normal homeostasis. This abrogation is primarily due to severe impairment of their differentiation potential. We describe a stem cell specific increase in the mRNA levels of key neoblast genes after Smed-not1 knock down, consistent with a role of the CCR4-NOT complex in degradation of neoblast mRNAs upon the onset of differentiation. We also observe a stem cell specific increase in the frequency of longer poly(A) tails in these same mRNAs, showing that stem cells after Smed-not1 knock down fail to differentiate as they accumulate populations of transcripts with longer poly(A) tails. As other transcripts are unaffected our data hint at a targeted regulation of these key stem cell mRNAs by post-transcriptional regulators such as RNA-binding proteins or microRNAs. Together, our results show that the CCR4-NOT complex is crucial for stem cell differentiation and controls stem cell-specific degradation of mRNAs, thus providing clear mechanistic insight into this aspect of neoblast biology.
| Although transcriptional regulation in stem cells is a very active subject, much less is known about how post-transcriptional mechanisms of gene expression affect stem cells. Here, we use freshwater planarians in order to address this question. Planarians have a striking regenerative capacity driven by a population of pluripotent stem cells, the neoblasts. Control of both proliferation and differentiation is thought to rely heavily on post-transcriptional mechanisms, but their precise role is unknown. Poly(A) tail length regulation is an important mechanism of post-transcriptional control of gene expression as changes can be very rapid, and longer poly(A) tails are linked to increased mRNA stability and translational activity. We investigated the role of the CCR4-NOT complex, the major deadenylating complex in eukaryotes, by knocking down its main scaffolding subunit called Not1. Neoblasts in knock down animals are unable to differentiate and accumulate mRNAs with longer poly(A) tails. Our results show that the CCR4-NOT complex is needed for the targeted degradation of specific mRNAs expressed in stem cells, and the failure of this process likely prevents neoblasts from differentiating. These results reveal a new functional aspect of the CCR4-NOT complex and offer a mechanistic insight into the regulation of planarian stem cells.
| Post-transcriptional control is central for the regulation of gene expression in stem cells [1]. A key post-transcriptional process is mRNA degradation [2]–[4] the control of which is believed to be as important as transcriptional regulation [5], [6]. Although transcriptional regulation has been extensively studied, less is known about the developmental and physiological roles of mRNA degradation in stem cells, which are thought to involve the same RNA binding proteins [7] that act together to coordinate many complex aspects of mRNA biology, one of which is degradation.
mRNA degradation starts with deadenylation (i.e. shortening of the poly(A) tail) [2], [8]. This affects gene expression both by decreasing translational activity and mRNA stability [9], [10]. The major deadenylase in eukaryotes is the CCR4-NOT complex [11]–[13], which is also involved in regulating several other aspects of mRNA metabolism, such as mRNA export [14], [15], translation [15] and transcription itself [11], [16]–[18].
In yeast, the CCR4-NOT complex is composed of nine different subunits [11]: two deadenylases (Ccr4p and Pop2p/Caf1p), five Not proteins (Not1p–Not5p), Caf40p and Caf130p. Among them, Not1p, a 240 kDa protein, is thought to act as a scaffold and is the only subunit required for yeast viability [11], [19]. Most of the subunits of the yeast complex are conserved across metazoans [20]–[22], with the exception of Not5p and Caf130p. In mammals two paralogous genes with mutually exclusive expression patterns encode each deadenylase of the complex [23]. Furthermore, the two deadenylase subunits Caf1 and Ccr4 regulate distinct sets of mRNAs [24], [25].
A number of translational repressors interact with the CCR4-NOT complex to repress their targets. For instance, Nanos proteins [26], [27], PUF proteins [28], [29], Smaug [30], and Bicaudal-C [31] all repress their target mRNAs via interaction with different subunits of the CCR4-NOT complex. Furthermore, the CCR4-NOT complex mediates the deadenylation of miRNA-targeted and piRNA-targeted mRNAs, executing the repressive functions of some small RNAs [32]–[35]. The CCR4-NOT complex directly binds to GW182, a component of miRNA repression complexes through evolutionary conserved motifs [36], [37].
Little is known, however, about the role of the CCR4-NOT complex in stem cells. It was found, for instance, that different components are important in maintaining mouse and human ESC identity [38], but the mechanisms remain largely unexplored. The freshwater planarian Schmidtea mediterranea is an emerging model for stem cell biology [39]–[41]. Its striking regeneration capacities are sustained by the presence of the neoblasts, a population of pluripotent stem cells that not only drive regeneration but sustain constant homeostatic cell turnover as well [42]. Planarian neoblasts can be eliminated by irradiation and are amenable to RNAi-mediated gene knock down. Furthermore, their abundance in the organism allows the quantitative evaluation of phenotypes.
The regulation of neoblasts and their pluripotency involves both transcriptional and post-transcriptional regulation, and a number of putative post-transcriptional regulators have been described to affect neoblast function [43]–[46]. Neoblasts also contain chromatoid bodies, RNA rich granules similar to germ granules which are thought to constitute hubs for mRNA processing and post-transcriptional regulation [39], [45], [47]–[49]. Recently, transcriptomic profiles of neoblasts have become available [50]–[54], confirming their long known resemblance to germ line stem cells [55], but also highlighting the importance of post-transcriptional mechanisms for their regulation [50] and the conservation of pluripotency determinants between planarian and mammalian stem cells [51], [52], [56], [57].
Here, we use S. mediterranea to investigate the role of the CCR4-NOT complex in stem cell regulation by characterizing the function of the Smed-not1 gene, which encodes for the homologue of Not1p in yeast and CNot1 in mammals. We report that its knock down specifically affects the differentiation and self-maintenance capabilities of neoblasts. Smed-not1 knock down results in a progressive increase in the levels of several neoblast transcripts, and we demonstrate that this increase is stem cell specific. Finally, we observe that these same mRNAs have stem cell specific increases in the frequency of long poly(A) tails after Smed-not1 knock down, showing that the observed increases in mRNA levels in stem cells are likely a consequence of decreased targeted deadenylation by the CCR4-NOT complex. Our findings highlight a likely central role for poly(A) tail length regulation in orchestrating pluripotent stem cell differentiation.
We identified the different subunits of the CCR4-NOT complex in the planarian species S. mediterranea by TBLASTN searches in the S. mediterranea genome and in our reference transcriptome assembly [50], [53] and other genomic and transcriptomic resources [58]–[60] (Table S1). We identified homologues of all metazoan genes known to encode subunits of the CCR4-NOT complex. Two components of the CCR4-NOT complex, the deadenylases Ccr4 and Caf1 were previously described in the planarian species Dugesia japonica [45]. However, no phenotype was reported for these enzymatic components after RNAi-mediated knock down and we also observed no strong phenotype (Figure S1) for Smed-not6 (the Ccr4 orthologue, Table S1), Smed-not7A and Smed-not7B (the two paralogues of Caf1 in S. mediterranea, Table S1). We instead chose to focus on Smed-not1, as Not1 is believed to act as the central scaffolding protein in the complex and it is the only component of the complex essential for viability in yeast [11], [19].
We investigated the expression pattern of Smed-not1 by whole mount in situ hybridization (WMISH). We observed broad expression of this key component of the core deadenylation complex, including expression throughout the parenchyma and the central nervous system (CNS) (Figure 1A, top panel). This pattern suggested to us that Smed-not1 may be expressed in neoblasts, since they are distributed throughout the parenchyma. To test this we monitored parenchymal expression after irradiation to remove neoblasts and observed that the parenchymal component of expression disappeared over a period of 5 days after irradiation (Figure 1A, middle and bottom panels). All neoblasts disappear 24–48 hours after lethal irradiation, and consequently the expression of neoblast specific genes disappears over a similar period. As controls we analyzed the expression of Smedtud-1, the orthologue of the previously described Schmidtea polychroa Tudor gene Spoltud-1 [48] (Figure 1B), Smed-vasa-1, a Vasa orthologue of S. mediterranea [54], [61] (Figure 1C), and Smed-pcna, the orthologue of the PCNA gene described in Dugesia japonica [62] (Figure 1D). As expected, the neoblast-specific signals of all three disappeared by day 3 post-irradiation, while irradiation insensitive expression in differentiated cells of the CNS remained for Smedtud-1 and Smed-vasa-1 (Figure 1B–D). Smed-eye53, a marker expressed in differentiated cells [63], was used as a control to demonstrate that gene expression in post-mitotic cells is not ablated by irradiation (Figure 1E). Since Smed-not1 hybridization signals present in the parenchyma were reduced but did not completely disappear by day 3 post-irradiation, it is likely that Smed-not1 is expressed in neoblasts and their recent progeny, but also other post-mitotic cells. These data indicates that Smed-not1 is broadly expressed throughout the planarian body, consistent with a housekeeping function, in a pattern that includes neoblasts, their progeny as well as differentiated cells. Investigation of Not1 expression in previous transcriptome based studies of mRNAs expressed in neoblasts is consistent with this expression pattern (Figure S2).
We then analyzed the function of Smed-not1 by RNAi experiments. All Smed-not1(RNAi) animals displayed abrogated regeneration capacities and eventually died. They were able to produce both anterior and posterior regeneration blastemas, but never completed the regenerative process (Figure 2B, vs. Figure 2A). In order to test if the formation of a regeneration blastema depended on the time of transection, we cut animals at 1, 3, 5, 10 and 15 days after Smed-not1(RNAi) treatment (Figure S3). We found that animals were able to produce a regeneration blastema at all-time points, however animals cut earlier produced larger blastemas. All blastemas of Smed-not1(RNAi) worms eventually regressed. The ability of Smed-not1(RNAi) animals to produce a large regeneration blastema at early time points after RNAi suggests that mitotic neoblasts, the source of blastema cells, are still present and proliferating.
We then analyzed the phenotype of intact Smed-not1(RNAi) animals. We found that Smed-not1(RNAi) worms presented head regression (Figure 2D, vs. Figure 2C) [43], [44], 64,65, a symptom of interrupted homeostatic cell turnover. A survival curve during which the onset of tissue homeostasis defects was recorded (N = 40) demonstrated temporal phenotypic variability (Figure 2E). In Smed-not1(RNAi) animals these defects were seen first at 15 days after RNAi, and in the majority of animals after 20 days (Figure 2C). Variable degrees of head regression were also observed after 20 days of RNAi (Figure 2C, also see Figure S3). By day 22 all animals had defects, showing complete penetrance of Smed-not1 RNAi. All animals died by day 36 after dsRNA delivery, with the majority of deaths occurring between day 26 and day 34 (Figure 2E). All control(RNAi) animals survived without any defect for >35 days, however. Together, these results demonstrate that Smed-not1 is needed for regeneration and homeostatic cell turnover in S. mediterranea.
Next, we analyzed the mitotic marker phospho-histone-H3 (h3p) in Smed-not1(RNAi) animals at 5, 10, 15 and 20 days after RNAi. Up to 15 days all (N = 7 per time point) had mitotic neoblasts comparable in numbers to those of control(RNAi) animals (Figure 3A). After 20 days all animals still had mitotic cells, but at variable levels, in agreement with our earlier phenotypic characterisation (Figure 3B–E). Most had normal levels (Figure 3C, vs. Figure 3B), although animals with more severe head regression defects showed a visible reduction in the mitotic levels (Figure 3D–E, vs. Figure 3B–C), but overall the reduction in mitoses was not statistically significant. These experiments show that Smed-not1(RNAi) worms have mitotic cells, even as head regression defects progress. Significantly, similar defects are seen in irradiated worms only weeks after complete loss of mitotic activity. Therefore, we interpret our data as showing that effects on stem cell proliferation are not the primary cause of the Smed-not1(RNAi) phenotype, instead implicating neoblast differentiation impairment as responsible for regenerative failure, head regression and other defects.
By performing FACS experiments we found that Smed-not(RNAi) led to a moderate reduction of the sorted irradiation sensitive X1 cells, which primarily contains neoblasts, at 15 days but not at 10 days (Figure 3F. Figure S4A–B). Even though no significant decrease of h3p cells was detected at this time point we interpret our FACS data as more sensitive and conclude that both methods consistently detect large numbers of proliferating neoblasts 15 days after Smed-not1 dsRNA administration, further implicating neoblast differentiation defects instead.
In order to monitor the behaviour of neoblasts and their post-mitotic progeny during progression of the Smed-not1 knock down phenotype we analysed the expression of neoblast and progeny markers [66]. Smedwi-1, a marker of neoblasts, Smed-nb.21.11e, a marker of early neoblast progeny, and Smed-agat-1, a marker of late neoblast progeny, were analyzed in control(RNAi) worms (Figure 4A–C) and Smed-not1(RNAi) worms after 10 (Figure 4D–F), 15 (Figure 4G–I), and 20 (Figure 4J–L, Figure S5A–F) days of RNAi. Only one time point (10 days) is shown for control(RNAi) worms, since no differences were observed between time points. Smedwi-1 expression was qualitatively the same after 10 and 15 days of Smed-not1(RNAi), (Figure 4D, G, vs. Figure. 4A), but clearly reduced to a variable extent after 20 days (Figure 4J vs. Figure 4A, D and G; Figure S5A vs. Figure S5D); some animals had nearly normal expression while Smedwi-1 expression was severely reduced in those with the most severe head regression defects. These results, like those above, suggest that prominent stem cell loss occurs only when Smed-not1(RNAi) animals begin to regress the head and to die, again implicating differentiation impairment instead of proliferation or self-renewal as a primary cause for regenerative failure.
Expression of Smed-nb.21.11e, a marker of early neoblast progeny, looked broadly equivalent to that of control worms in animals fixed 10 days after Smed-not1 dsRNA delivery (Figure 4E, vs. Figure 4B) but a clearly visible progressive decrease in the number of Smed-nb.21.11e-positive cells was detected after 15 (Figure 4H, vs. Figure 4B) or 20 (Figure 4K, vs. Figure 4B) days after RNAi. After 20 days of Smed-not1 knock down animals had only a few remaining Smed-nb.21.11e-positive cells (Figure S5E). These results show that clearly visible decreases in early progeny cell number precede the prominent decrease in neoblasts themselves. When we checked the expression of Smed-agat-1, a marker of late neoblast progeny, we observed a consistent qualitative increase in Smed-agat-1-positive cells in Smed-not1(RNAi) worms after 10 days (Figure 4F vs. Figure 4C). At later time points, however, the number of Smed-agat-1-positive cells also progressively declined (Figure 4I, L, vs. Figure 4C). Again, we observed a considerable variability in Smed-not1(RNAi) worms after 20 days (Figure S5F). However, in all animals with a clear decrease in the number of Smed-agat-1-positive cells this was particularly apparent in the anterior region, a characteristic feature of Smed-agat-1-positive cell depletion upon neoblast elimination by irradiation or perturbation by RNAi [50], [64], [65], [67].
Therefore, neoblasts are abundant 15 days after RNAi, and only clearly start to be depleted later, coinciding with the onset of head regression defects. Similar defects are observed after irradiation, however, these take >10 days to manifest after elimination of mitotic activity. In contrast Smed-not1(RNAi) animals display these defects when neoblasts are still present. These results suggest that a primary defect in neoblast differentiation, rather than neoblast maintenance, causes failure in tissue homeostasis. In support of this idea, altered stem cell progeny numbers precede the disappearance of Smedwi-1 signals and mitotic activity. This alteration can be observed as early as 10 days after RNAi in the case of Smed-agat-1-positive cells and 15 days for Smed-nb.21.11e-positive cells, which are clearly depleted at this time point.
To further test this, we compared the dynamics of the neoblast and progeny cell markers in Smed-not1(RNAi) animals to those of Smedwi-2(RNAi) animals, in which neoblast differentiation is abrogated [44]. Smedwi-1, Smed-nb.21.11e and Smed-agat-1 were expressed in Smedwi-2(RNAi) animals with very similar dynamics to Smed-not1(RNAi) animals (Figure S5G–R). Taken together, these results show that, similar to Smedwi-2 RNAi, Smed-not1 RNAi impairs neoblast differentiation with proliferation only affected in a later time point.
In order to achieve a quantitative measure of stem cell progeny mRNA levels in Smed-not1(RNAi) animals we performed quantitative real time PCR experiments (qRT-PCR) of Smed-nb.21.11e and Smed-agat-1 transcripts in Smed-not1(RNAi) worms. We focused on earlier time points of 10 and 15 days after RNAi. qRT-PCR showed that Smed-nb.21.11e levels in whole animals were similar to those of control(RNAi) after 10 and 15 days of dsRNA administration (Figure 5A). However, Smed-agat-1 mRNA levels increased by day 10 and were almost two-fold higher after 15 days. Since this result did not correlate with what we observed by colorimetric WMISH, we quantified Smed-agat-1-positive cells by fluorescent WMISH (FWMISH). A significant increase in the numbers of Smed-agat-1-positive cells was found in Smed-not1(RNAi) animals 10 days after dsRNA administration (Figure 5B), but their numbers declined to control levels after 15 days. These data confirmed the qualitative data from colorimetric WMISH (Figure 4). The distribution of Smed-agat-1-positive cells in Smed-not1(RNAi) animals 15 days after dsRNA delivery was different from controls, with less Smed-agat-1-positive cells in the anterior part of the worm (Figure 5C). This revealed a stark discordance between the number of Smed-agat-1-positive cells and the level of Smed-agat-1 mRNA. We conclude that Smed-agat-1 transcripts are accumulating in decreasing numbers of Smed-agat-1-positive cells after 15 days of Smed-not1 knock down, and that each Smed-agat-1-positive cell contains an increased average number of Smed-agat-1 transcripts. A similar, but less pronounced, process could also explain reduced numbers of Smed-nb.21.11e positive cells and discordant stable levels of this transcript, which do not drop by day 15.
Given that the CCR4-NOT complex is known to regulate gene expression through its deadenylating activity we wished to ascertain if it could be directly responsible of the discordance between Smed-agat-1 positive cell number and mRNA levels. If an increase in mRNA levels is caused by impaired deadenylation and subsequent degradation we would expect to observe increased frequency of long poly(A) tail lengths. Using a poly(A) tail length (PAT) assay [68]–[71], in whole worms we observed that this was the case for both Smed-agat-1 and Smed-nb.21.11.e (Figure 5D), with the first giving starker differences, while control mRNAs Smed-eif-3, Smed-mhc, or Smed-ef-2 were only mildly affected or not affected at all. In addition a spike-in control of exogenous mRNA showed equal poly(A) tail length distribution across samples (Figure 5D), showing that the differences observed are present in our different planarian mRNA samples and not introduced by the PAT assay technique.
Next, we found that WMISH analysis of neoblast markers, Smedtud-1, Smed-vasa-1 and Smed-pcna, was suggestive of increased levels of these transcripts in Smed-not1(RNAi) worms, with qualitatively visible differences after both 10 and 15 days after RNAi treatment (Figure 6A–C). These data further demonstrate that neoblast maintenance is not affected by Smed-not1 knock down at these time points.
Since WMISH does not provide a quantitative measure of mRNA levels we quantified these differences by qRT-PCR experiments on RNA from Smed-not1(RNAi) animals. Smedtud-1, Smed-vasa-1 and Smed-pcna all progressively increased to levels approximately 50% and 100% higher than those of control(RNAi) after 10 and 15 days, respectively (Figure 6D). Smedwi-1 transcript levels were also significantly increased in whole animals after 15 days (Figure 6D). Collectively, these results demonstrate that Smed-not1 RNAi knock down leads to an increase in mRNA levels in several genes expressed in neoblasts and their progeny.
Given the known conserved function of the CCR4-NOT complex in regulating mRNA levels through targeted deadenylation we performed PAT assays on the set of neoblast markers and on the samples above (Figure 5D). In all cases Smed-not1 knock down resulted in increased average poly(A) tail length (Figure 6E), demonstrating that increased transcript levels correlate with a failure in deadenylation. These data confirm that knock down of the CCR4-NOT complex subunit Smed-not1 leads to increased transcript levels of genes known to be key to neoblast function, by blocking their deadenylation and subsequent degradation.
Since many neoblast mRNAs are also expressed in differentiated cells (e. g. Smedtud-1 and Smed-vasa-1 are also prominently expressed in the CNS, Figure 1B–C), the increased mRNA levels we detected could arise from a response in differentiated cells alone, stem cells alone or from both differentiated and stem cells. To distinguish these possibilities, we planned to use an irradiation approach to remove all neoblasts and then measure transcript levels in Smed-not1(RNAi) and control(RNAi) worms by qRT-PCR. We reasoned that if transcript accumulation was indeed limited to stem cells then mRNA levels of these transcripts after irradiation would be equal in both irradiated control(RNAi) and irradiated Smed-not1(RNAi) samples.
Smedtud-1 and Smed-vasa-1 are expressed in the CNS to levels that amount respectively to roughly 70% and 40% of their total expression, according to our previous neoblast profiling by a combinatorial RNA-seq, RNA interference and irradiation approach [50]. We confirmed this by qRT-PCRs (Figure S5A) and observed that Smed-pcna expression amounts to roughly only 10% of its normal expression 24 hours after irradiation (Figure S6A), indicating that 24 hours of irradiation suffice to eliminate around 90% of neoblasts. Conversely, most neoblast progeny survive beyond 1 day post-irradiation, as measured by our qRT-PCR experiments with the markers Smed-nb.21.11e and Smed-agat1 (Figure S6B) and consistent with previously published data [50], [66], making then 24 hours after irradiation the ideal time point to perform our experiment.
Therefore, in order to find out if the increased levels of neoblast mRNAs originate in neoblasts or instead in the CNS or elsewhere, we then used this irradiation approach in Smednot-1(RNAi) animals and compared them to control(RNAi) animals. We irradiated both Smednot-1(RNAi) and control(RNAi) animals at 9 and 14 days after dsRNA administration, 24 hours before the data collection time points of 10 and 15 days respectively. We then examined the expression pattern of Smedtud-1 by WMISH. This confirmed our qRT-PCR experiments, and was consistent with our previous WMISH results (Figure 7A). Non irradiated Smed-not1(RNAi) animals showed qualitatively more intense expression of Smedtud-1, However, all animals that were irradiated 24 hours prior to fixation showed an identical expression pattern of Smedtud-1, with similar levels of signal detected only in the CNS to their control(RNAi) irradiated counterparts (Figure 7A). This experiment shows that Smedtud-1 is not ectopically expressed in other tissues or organs after Smed-not1 knock down, since this ectopic expression should be visible either in non irradiated or irradiated samples, and suggests instead that the increased levels of Smedtud-1 come from an accumulation of this transcript in neoblasts.
We performed qRT-PCR and confirmed that Smedtud-1, Smed-vasa-1 and Smed-pcna mRNA levels increased progressively in Smed-not1(RNAi) animals (Figure 7B, left) but both control(RNAi) and Smed-not1(RNAi) animals irradiated 24 hours previously contained similar levels of all three transcripts (Figure 7B, right). Similar to results for wild type worms (Figure S6A), the levels of the Smedtud-1, Smed-vasa-1 and Smed-pcna transcripts were respectively around 70%, 40% and 10% of the expression in non-irradiated control(RNAi) animals. Therefore we conclude that the overexpression observed in Smed-not1(RNAi) animals for these three transcripts disappears 24 hours after irradiation, and is therefore located in irradiation-sensitive neoblasts, rather than elsewhere in the body.
To implicate the CCR4-NOT complex-mediated targeted mRNA degradation directly in this effect we performed PAT assays on non irradiated and irradiated samples and observed that the distribution of poly(A) tail lengths of all these neoblast mRNAs was increased in Smed-not1(RNAi) animals, but this effect was removed by irradiation (Figure 7C and D). In the case of neoblast specific Smedwi-1 and Smed-pcna no signal was detected after irradiation (Figure 7C). In the case of Smed-vasa-1 and Smedtud-1 signal was detectable after irradiation, reflecting expression in the CNS, but these transcripts did not show any increase in poly(A) tail length distribution after Smed-not1 knock down (Figure 7D). These data confirm that effects on mRNA levels for these important neoblast genes are confined to stem cells. As an additional control we also measured Smed-nb.21.11e and Smed-agat-1 poly(A) lengths (Figure 7E). 24 hours after irradiation cells expressing these transcripts are still present and the increase in poly(A) tail length caused by Smed-not1 knock down is still evident. This confirms that irradiation itself does not cause the absence of detected differences in poly(A) tail length per se, but by removing the cycling neoblasts. The poly(A) tail length distribution of control mRNAs were not affected by irradiation (Figure 7F).
As a further proof that the effects we observe are confined to stem cells we looked at gene expression, poly(A) tail length and the efficacy of Smed-not1 knock down across FACS cellular compartments [51], [72]. Progressive increases in neoblast gene mRNA levels were confined to X1 and X2 populations of sorted cells from Smed-not1(RNAi) animals (Figure 8A, X1 and X2). Both of these sorted fractions contain neoblasts to different extents. Levels of these mRNAs were not increased in irradiation insensitive (Xins) sorted fractions from Smed-not1(RNAi) animals compared to controls (Figure 8A, Xins). This fraction contains primarily differentiated cells including CNS cells which also express Smedtud-1 and Smed-vasa-1.
One possibility for the specificity we observe is that Smed-not1 knock down has a high efficacy in stem cells but not in post-mitotic cells, as it has been suggested recently for Smed-bruno-like knock down [38], [73]. We performed qRT-PCR measurement of Smed-not1 transcript levels and found that Smed-not1 mRNA is consistently expressed in all cellular fractions (consistent with Figure S2 and Figure 1A) and that Smed-not1 knock down significantly depletes Smed-not1 in all sorted compartments compared to control(RNAi) animals (Figure 8B). Consequently the observed neoblast specificity is unlikely due to an absence of knock down in differentiated cells.
To finally link the mechanism of CCR4-NOT complex-mediated deadenylation to increased mRNA levels we also checked poly(A) tail lengths in FACS sorted cells. For Smedtud-1, Smed-vasa-1, Smed-pcna and Smedwi-1 we observed a progressive increase in long poly(A) tails in the X1 compartment of stem cells (Figure 8C and D). This trend was also observed for these transcripts in X2 cells (Figure 8C and D). No poly(A) tail signal was detectable in Xins fractions for Smed-pcna and Smedwi-1, consistent with their low abundance in this fraction, and poly(A) tail length was not affected for Smedtud-1 and Smed-vasa-1 in Xins cells (Figure 8C and D). These data are in agreement with our irradiation based experiments (Figure 7). The poly(A) tail lengths of control mRNAs Smed-ef-2 and Smed-mhc were only mildly or not affected by Smed-not1 knock down (Figure 8E). These mild differences are likely due to CCR4-NOT mediated deadenylation, but again seem to be restricted to X1 and X2 cells, whereas Xins cells remain unaffected. To which extent all, most or only a subset of neoblast transcripts are affected after Smed-not1 knock down remains an open question.
Taken together, these results clearly demonstrate that Smed-not1 knock down induces a prominent increase of key transcripts expressed in neoblasts, that this accumulation occurs specifically in neoblasts, and that it is associated with an increased frequency of long poly(A) tails of these transcripts specifically in neoblasts. Furthermore, these results strongly suggest that the neoblast-specific increase of mRNA levels of genes such as Smedtud-1, Smed-vasa-1 and Smed-pcna may be responsible for the impaired differentiation capacities of neoblasts observed in Smed-not1(RNAi) animals. It is likely that other genes expressed in neoblasts are similarly upregulated in Smed-not1(RNAi) animals and contribute to differentiation defects. Our results suggest a mechanism by which the differentiation capacities of neoblasts are dependent on CCR4-NOT mediated degradation of specific neoblast mRNAs.
Planarians are an emerging in vivo model for stem cell biology because of their unique stem cell population. In this study we used the planarian S. mediterranea as a model system to establish a function for the CCR4-NOT complex in stem cell regulation. We found that Smed-not1 knock down abrogated regeneration and impaired homeostatic cell turnover. Interestingly, Smed-not1 knock down primarily affects the stem cell compartment of S. mediterranea rather than inducing more widespread effects, even though the CCR4-NOT complex is the major deadenylating complex in eukaryotes and regulates at least 85% of mRNAs in yeast [74].
While we observed a stark and specific effect of Smed-not1 knock down on deadenylation, it is still possible that other functions of the CCR4-NOT complex might also be impaired. The CCR4-NOT complex is involved in several steps of RNA metabolism [12], [13], [16] and further work is therefore needed to elucidate which ones are also at work in stem cells.
We observed effective gene knock down of Smed-not1 even in differentiated cell fractions, but specific effects on neoblast transcripts were limited to stem cells. This fact suggests that targeted deadenylation by either RNA-binding proteins or miRNAs is providing specificity and is therefore central to stem cell differentiation and self-renewal properties. Consistently, several studies have implicated the CCR4-NOT complex in mRNA-specific deadenylation via targeted recruitment of the CCR4-NOT complex by RNA-binding proteins [26], [27], [30], [31], which are in turn known to be highly enriched and functionally important in neoblasts [43]–[46], [48]–[50], [52], [75]. It is possible to hypothesize that disruption of the CCR4-NOT complex via knock down of its scaffolding subunit might impair the protein-protein interactions needed to tether deadenylating activity to transcripts targeted by RNA-binding proteins or miRNAs for degradation, while general non-targeted deadenylation would remain relatively unaffected, therefore not causing a broader metabolic failure at earlier time points. In fact, the miRNA silencing complex protein GW182 specifically interacts with the Not1 subunit of the CCR4-NOT complex through specific and conserved domains [36], [37].
Despite the specificity of the effects seen in neoblast transcripts, mRNAs expressed elsewhere were also found to be affected. Both Smed-agat-1 and, to a lesser extent, Smed-nb.21.11e were also affected. These results show that CCR4-NOT deadenylating activity is present in cell types other than neoblasts and that specificity is not due to restriction of activity to stem cells. Furthermore, the effects on well described neoblast progeny markers suggest that Smed-not1 knock down likely influences several steps of cellular differentiation that may all contribute to the observed effects on homeostasis and regeneration.
The Smed-not1 phenotype is progressive with respect to both the decreasing capacity of the animals to produce blastema cells and by the accumulation of mRNAs in stem cells. The phenotype results from a drop in neoblast progeny numbers, followed by stem cell loss. Similar neoblast and progeny dynamics have been shown by us in Smedwi-2(RNAi) organisms here and by another group in Smed-CHD4(RNAi) organisms [64]. Both Smedwi-2 and Smed-CHD4 knock downs initially affect neoblast differentiation rather than their self-renewal and proliferative capabilities [44], [64] which are only affected at later time points, similar to Smed-not1 knock down. The ultimate cause for stem cell loss after the impairment of neoblast differentiation is unknown, and likely to be a broad failure in homeostasis as organs and tissues fail. However, for more than 15 days after Smed-not1 dsRNA administration proliferating neoblasts are detectable in large numbers while regeneration is abrogated, suggesting neoblast loss is not a primary cause for the regeneration defect, instead implicating impaired neoblast differentiation capacities.
Smed-not1 knock down induces an increased frequency of long poly(A) tails of Smedtud-1, Smed-vasa-1, Smed-pcna and Smedwi-1 mRNAs as these same mRNAs accumulate. This regulation occurs specifically in neoblasts, rather than in the CNS, where two of these transcripts are also expressed. However, it is likely that Smed-not1 knock down affects many more transcripts that contribute to failure in stem cell differentiation. Due to the lack of specific antibodies it is difficult to evaluate if the differences observed in transcript abundance and polyadenylation state affect the abundance of the proteins that these transcripts encode. However, taking into account that all four transcripts expressed in neoblasts analyzed accumulate in these stem cells, it is reasonable to conclude that Smed-not1 knock down induces transcriptome-wide changes in stem cell expression patterns, and that these changes will likely affect protein levels.
Our results offer a new mechanistic insight into post-transcriptional regulation in neoblasts and its targets. After depletion of a post-transcriptional regulator many transcripts accumulate without being degraded, and this likely prevents neoblast differentiation, which needs the effective removal of these transcripts. The deadenylating activity of the CCR4-NOT complex is clearly central to this process. For genes like Smedtud-1 and Smed-vasa-1, which are expressed in neoblasts and the CNS, we observe accumulation and increased frequency of long poly(A) tails of the transcript only in neoblasts. This suggests that the deadenylation of these mRNAs is regulated specifically during the onset of differentiation and requires the targeted recruitment of the CCR4-NOT complex by RNA-binding proteins, as has been described in other organisms [26], [27], [30], [31], [76]. Interestingly, several RNA binding proteins and post-transcriptional regulators have already been described as crucial for neoblast function [43], [44], [46], [48], and some of them have been already been functionally linked with the CCR4-NOT complex in other model organisms. Future research will help elucidate the mechanisms by which these proteins orchestrate planarian stem cell processes.
The CCR4-NOT complex has been shown to mediate the repressive function of both miRNAs and piRNAs [32]–[35]. Small RNAs are believed to be very important regulators of mammalian stem cells [77] and neoblasts [78], [79]. The Smed-not1 RNAi phenotype is very similar to those of Smedwi-2 and Smedwi-3, two Piwi proteins involved in piRNA regulation [44], [80]. Furthermore, several studies have highlighted the presence of miRNAs highly enriched in stem cells [78], [79], [81]. Future research will help in understanding if these phenotypic similarities reflect a functional link between Piwi proteins, piRNAs, miRNAs and the CCR4-NOT complex in planarian stem cells. Our results highlight the importance of the CCR4-NOT complex in the regulation of stem cells, the fact that post-transcriptional regulation of gene expression is a key element in the regulation of pluripotency, and that planarians will provide an excellent platform for these studies.
Planarians of the asexual strain of S. mediterranea were kept and used as previously described [82].
The putative members of the S. mediterranea CCR4-NOT complex and other transcripts were identified in published S. mediterranea transcriptomic and genomic sequences [53], [58]–[60] and the longest transcripts for Smed-not-1 confirmed by PCR and sequencing. The putative members of the S. mediterranea CCR4-NOT complex were identified by TBLASTN searches in the current assembly of the S. mediterranea genome and in the available transcriptomic data. In order to determine the number of loci for each of the components, the different transcripts identified were mapped to the S. mediterranea genome. The genomic sequence encoding Smed-not1 was found split in two contigs (v31.001778 and v31.002774), the existence of one single transcript for these two genomic contigs was confirmed by PCR using the primers 5′-CATCGCAACAATGGAGAGAA-3′ and 5′-ATTTGAGCTGTATGGGCGAT-3′, each mapping to one of the two contigs. These PCR experiments revealed as well the existence of a 3 kb region not present in the S. mediterranea genomic data.
The full sequence of Smed-not1 was obtained by de novo assembling the transcript from the raw transcriptomic data, using the known transcriptomic and genomic data as a guide. The Smed-not1 sequence has been deposited in Genbank (accession KF781122).
The sequences of Smedtud-1, Smed-vasa-1 (accession JQ425140) and Smed-pcna (accession EU856391) were found in our reference transcriptomic data, encoded by the transcripts AAA.454ESTABI.16133, AAA.454ESTABI.18605 and AAA.454ESTABI.22122 respectively. Smed-ef-2 is encoded by the transcript AAA.454ESTABI.17328. The Smedtud-1 sequence has been deposited in Genbank (accession KF781126).
RNAi experiments were performed as previously described [82]. Control(RNAi) worms were injected with dsRNA encoding for GFP, a gene not present in the S. mediterranea genome. dsRNA encoding for Smed-not1 was prepared by in vitro transcription of a region of the Smed-not1 gene. Briefly, an amplicon was generated from S. mediterranea reverse transcribed RNA with the primers 5′-GGCCGCGGTGTCCAAGAAAAAGCAAGTCAG-3′ and 5′-GCCCCGGCCAGCTGGCGTCAGTTTAGTGAA-3′, containing a 5′ adaptor sequence for the subsequent addition of T7 promoter. The product of this PCR was gel purified and subjected to another step of amplification with the primers 5′-GAGAATTCTAATACGACTCACTATAGGGCCGCGG-3′ and 5′-AGGGATCCTAATACGACTCACTATAGGCCCCGGC-3′ with the purpose of attaching T7 promoter sequences to both ends of the amplicon. The product of this PCR was further purified with Purelink DNA purification columns (Invitrogen) and used as a template for in vitro transcription using T7 RNA polymerase (Roche). The product of the in vitro transcription was treated with Turbo DNAse (Ambion), phenol extracted, precipitated in ethanol in presence of sodium acetate, glycogen and EDTA and re-suspended in water. dsRNA encoding GFP for use as a negative control was similarly prepared from a vector encoding the GFP gene. The final concentration of the injected solution was 1 µg/ µl. Animals were injected with a Nanoject II (Drummond) for three consecutive days and monitored or used for experiments in the subsequent days. Day 1 after RNAi is considered to be in all experiments the first day after the third dsRNA injection. Alive animals were imaged in a Zeiss Discovery V8 with a Zeiss AxioCam MRC camera.
Irradiation was performed as previously described [79]. Animals were placed in a sealed γ-ray source and administered an irradiation dose of 100 Gy.
WMISH, FWMISH and IHC were performed and imaged as previously described [82]. When qualitative differences are shown, animals were processed and monitored in parallel. Riboprobes were generated by in vitro transcription of PCR products generated as described above, with only one T7 promoter linked to the 3′ end of the amplicon, and in the presence of digoxigenin-labelled UTP (Roche). The products of in vitro transcription reactions were then treated with Turbo DNAse (Ambion). Riboprobes were then precipitated in ethanol in the presence of LiCl and glycogen and resuspended in 50% formamide in TE buffer, 0.01% Tween.
The following primers were used:
Smed-not1: 5′- GGCCGCGGTGTCCAAGAAAAAGCAAGTCAG-3′
and 5′- GCCCCGGCCAGCTGGCGTCAGTTTAGTGAA-3′
Smedtud-1: 5′-GGCCGCGGCTAATGCCAGTTGACTGTCC-3′
and 5′-GCCCCGGCCCGAAAAAGTTCCGCATCACTT-3′
Smedvas-1: 5′-GGCCGCGGAGCTGTTGGAGTTGTTGGCTCAG-3′
and 5′-GCCCCGGCCCTAATCTTCGAGCCATTCAG-3′
Smed-pcna: 5′-GGCCGCGGATGGACTTGGATGGAGATCACT
Smedwi-1: 5′-GGCCGCGGAAGTGGTGGTATTCGAGAAGGA-3′
and 5′-GCCCCGGCCACGAATCGTAATCGGTTGTTCT-3′
Smed-agat-1: 5′-GGCCGCGGGAAATGATTGAGTCCACCATGA-3′
and 5′-GCCCCGGCCTGCAATATCTGGATAAGGAGCA-3′
Smed-nb.21.11e: 5′-GGCCGCGGGTGATTGCGTTCGCGTATATT-3′
and 5′-GCCCCGGCCATTTATCCAGCGCGTCATATTC-3′
Briefly, animals were killed in a 2%HCl solution, fixed in Carnoy's solution, bleached in a 8% H2O2/methanol, rehydrated, permeabilized with Proteinase K (Sigma), treated with 0.25% and 0.5% acetic anhydride in 0.1M triethanolamine pH 7.6, prehybridized and hybridized with digoxigenin labelled riboprobes (0.2 ng/µl, O/N at 56°C) They were then washed in buffers of increasing stringency, immunolabelled with anti-digoxigenin-alkaline-phosphatase antibody (Roche) and developed in the presence of NBT and BCIP (Roche). For FWMISH, an anti-digoxigenin-peroxidase antibody (Roche) was used and the signal was developed with the Tyramide Signal Amplification kit (Perkin Elmer). WMISH specimens were imaged on a Zeiss Discovery V8 equipped with a Zeiss AxioCam MRC camera. FWMISH specimens were imaged in a Leica SP3 confocal.
Animals (N = 8 per time point and treatment) were stained by FWMISH of Smed-agat-1 and monitored by confocal microscopy. Cell counts were performed in z-projections of both the dorsal and ventral sides of the animals. 130 squares corresponding to 250 µm2 in both the dorsal and ventral parts of the animals were selected for counting along the length of the animals. Counts were performed using ImageJ software. Error bars represent standard error of the mean. Statistical significance was analyzed by Student's T test by comparing values from each sample to its respective control sample.
Whole mount immunohistochemistry of phospho-histone-3 was performed as previously described [83]. Essentially, animals were fixed as above, blocked in a 1% BSA/PBS 0.3% triton X-100 solution, incubated overnight with anti-phospho-histone-3 (Millipore, 1/500 dilution) and anti-phospho-tyrosine (Cell Signaling, 1/200 dilution), washed and incubated in Alexa Fluor 488 and 568 secondary antibodies (Molecular Probes, 1/400 dilution). Animals were then washed, mounted in 70% glycerol/PBS and imaged with a Leica SP3 confocal and a Leica MZ16F fluorescence stereomicroscope and a Leica DFC 300Fx camera. Countings of phospho-histone-3 were performed with ImageJ software and normalized to the total area of the sample.
qRT-PCR experiments were performed as previously described [48] with modifications. Essentially, total RNA from samples of 5 animals was extracted with Trizol reagent (Invitrogen) according to the manufacturer's instructions, and cDNAs were synthesized with SuperScriptIII Reverse Transcriptase (Invitrogen). qRT-PCR experiments were then performed using the Absolute qPCR SYBR Green Master Mix (Thermo Scientific). Experiments were performed on three biological replicates per time point and treatment. Each biological replicate was technically replicated three times in each reaction, each reaction was repeated three times. The gene Smed-ef-2 was used for normalization.
The following primers were used:
Smed-not1: 5′-GACAGCGATTATGAACTGCC-3′ and 5′-CTGCTATGTTACTGGTGTTGAG-3′
Smedwi-1: 5′-AGTTCCTGTTCCAACGCATTATG-3′ and 5′-CTGGAGGAGTAACACCACGATGA-3′
Smed-nb.21.11e: 5′-GTCTCCCGCCAAATCAAGTA-3′ and 5′-TTTCATGCAATCTGCTTTCG-3′
Smed-agat-1:5′-TCCATCCAGAACCGATTGAT-3′ and 5′-CTCCCAAGTCATGGTGGACT-3′
Smedtud-1:5′-TGATGAAGGAACTTCGGGTGAT-3′ and 5′-TCTGAGCAACCGATTGAAACC-3′
Smedvas-1:5′-TGAAATGAACAAATCCCGAC-3′ and 5′-GAGAGCCAAACTAATTCCAG-3′
Smed-pcna:5′-GGCGCTTGGTAGTAATGATTCCCTA-3′ and 5′-TACCTAAGTGATCTCCATCCAAGTCC-3′
Smed-ef-2:5′-CAGCCAGTAGCTTTAAGCGATGA-3′ and 5′-ACTCTCAACGCTGCTGTCACTTC-3′
Statistical significance was measured by Student's T test by comparing values from each sample to their respective control sample.
Fluorescence activated cell sorting (FACS) of planarian samples was performed as previously described [51], [72]. Analysis was performed with FlowJo.
PAT assays were performed as previously described [69], [71] with minor modifications. A total of 400–1000 ng of total RNA extracted from five animals were used per each time point, replicate and treatment, except for FACS samples, in which 40 animals were used per time point and treatment and PAT reactions were performed with 100 ng of total RNA. Three biological replicates were analyzed per time point and treatment, and technically replicated at least twice, except for FACS samples, which were only technically replicated. C. elegans total RNA was spiked-in as a control and assayed with a primer for cpg-2.
RNA samples were incubated with 0.3 µg of 5′-phosphorylated oligo d(T) in a total volume of 8 µl, and heat denatured for 5 min at 65°C. Then, the following mixture was added: 4 µl of Super Script II First Strand Buffer, 0.5 µl of 0.1 M DTT, 2.25 µl of 10 mM ATP, 0.125 µl of RNAsin (Promega), 1.25 µl of 1 mM (each) dNTPs (Promega), and 1 µl of 2000 units/µl T4 DNA Ligase (New England Biolabs). The volume was then brought to a total of 20 µl with water, and the samples were incubated at 42°C for 30 min. in order to allow the saturation of polyA tails with oligo d(T). Then, 0.5 µl of an oligod(T)-anchor 100 µM was added (5′-GCGAGCTCCGCGGCCGCGTTTTTTTTTTTT-3′), and the mixture was incubated for 2 hours at 12°C to allow ligation of oligo d(T) molecules. Then, samples were pre-warmed at 42°C and 1 µl of SuperScript II (Invitrogen) was added. Finally the samples were incubated for 1 hour at 42°C and the SuperScript enzyme was heat inactivated at 70°C for 30 min.
These cDNAs were then used in PCR reactions, with fresh aliquots of the anchor primer and gene specific forward primers designed close to the 3′ region of the mRNA. The sequences of the mRNAs tested were obtained in our reference transcriptome (4). The following primers were used:
Smedtud-1: 5′-TGATGAAATAATGCTACCCGCGCAAT-3′
Smed-vasa-1: 5′-AGCCGACTTCTGAATGGCTCGAAGA-3′
Smed-pcna: 5′-CAAAGGCTGCACCTCTTTCTTCTCA-3′
Smedwi-1: 5′-CGTTGGCAAGATTCATCGTGGTGTT-3′
Smed-agat-1: 5′-TCGGATGTTAGAAGGCGAGGAGACC-3′
Smed-nb.21.11e: 5′- GACGGCCACTGTGACGCAGAAT-3′
Smed-ef-2: 5′-AACCCACTGGATCCCACAACGAAAC-3′
Smed-eif-3: 5′-GTTGCCCCATCGATTGGATACTTCG-3′
Smed-mhc: 5′- CGAGGAGCAAGTTCTGGACCTGGAA-3′
PCR reactions were carried away with DreamTaq (Fermentas) and amplified for 28–32 cycles of 94°C for 20 s., 65°C for 20 s. and 72 for 30 s. The products of the PCR reactions were analyzed on 1.5% agarose gels.
|
10.1371/journal.pcbi.1005158 | Machine Learning for Characterization of Insect Vector Feeding | Insects that feed by ingesting plant and animal fluids cause devastating damage to humans, livestock, and agriculture worldwide, primarily by transmitting pathogens of plants and animals. The feeding processes required for successful pathogen transmission by sucking insects can be recorded by monitoring voltage changes across an insect-food source feeding circuit. The output from such monitoring has traditionally been examined manually, a slow and onerous process. We taught a computer program to automatically classify previously described insect feeding patterns involved in transmission of the pathogen causing citrus greening disease. We also show how such analysis contributes to discovery of previously unrecognized feeding states and can be used to characterize plant resistance mechanisms. This advance greatly reduces the time and effort required to analyze insect feeding, and should facilitate developing, screening, and testing of novel intervention strategies to disrupt pathogen transmission affecting agriculture, livestock and human health.
| Insect vectors acquire and transmit pathogens causing infectious diseases through probing on host tissues and ingesting host fluids. By connecting insects and their food source via an electrical circuit, computers, using machine learning algorithms, can learn to recognize insect feeding patterns involved in pathogen transmission. In addition, these machine learning algorithms can show us novel patterns of insect feeding and uncover mechanisms that lead to disruption of pathogen transmission. While we use these techniques to help save the citrus industry from a major decline due to an insect-transmitted bacterial pathogen, such intelligent monitoring of insect vector feeding will engender advances in disrupting transmission of pathogens causing disease in agriculture, livestock, and human health.
| The invention of an electronic method for monitoring the feeding behavior of sucking insects [1–4] provided a potentially powerful tool to describe the cryptic behavior of the mouthparts of fluid-feeding phytophagous insects inside a host plant (Fig 1). Coupled with histological studies to correlate specific waveforms with the mouthparts’ position within the host [5, 6], electronic monitoring allows researchers to follow the sequence of events that lead to ingestion and, in the case of insect vectors, to acquisition and transmission of pathogens. The method, variously referred to as electronic feeding monitor or electrical penetration graph (EPG), has been applied to various studies of host plant resistance and pathogen transmission [6–12]. It has also been used to characterize feeding by blood-feeding mosquitoes and ticks [11, 13, 14].
A major constraint to the utility of the method is the amount of time required to interpret the waveforms produced. Currently, a trained human observer is required to characterize each waveform and assign the corresponding feeding state on a second-by-second basis. During a typical experiment, EPG recordings generate gigabytes of data. Classification of these data into insect feeding states corresponding to intercellular passage, cell sampling, salivation, phloem ingestion, xylem ingestion and other activities associated with feeding or pathogen transmission is typically accomplished by comparison to published standards [6]. Computer classification methods based on motif recognition have been devised, but suffer from low accuracy [15]. Most analysis currently requires expert training and manual annotation that preclude high-throughput analysis. This onerous and time-consuming process is a major limitation to the broader and more in-depth application of this otherwise useful technique.
We focused on removing the data analysis bottleneck through application of machine learning algorithms designed to teach a computer program to recognize and learn from insect feeding states with little or no human input [16]. To do so, we relied on EPG recordings from an insect-plant-pathogen model system where automated processing and analysis of insect feeding data could have an immediate and measurable impact on development of effective intervention strategies through screening of plant varieties resistant to pathogen transmission. In this system, the Asian citrus psyllid, Diaphorina citri (Hemiptera: Liviidae) transmits the phloem-limited and persistently propagated bacterium Candidatus Liberabacter asiaticus (CLas), implicated as the causative agent of citrus greening disease [17–19]. Citrus trees infected with this pathogen rapidly develop debilitating symptoms affecting tree health and fruit quality; the pathogen kills the tree within three to five years [20].
Since the first report of this pathogen in Florida in 2005, this vector-pathogen complex has devastated the United States citrus industry. The Florida citrus industry alone has seen five years of unprecedented decline resulting in billions of dollars of lost revenue and jobs [21]. In 2015, the U.S. Department of Agriculture predicted a precipitous drop in citrus production in 2016 to 69 million boxes in Florida, well below a peak of 242 million boxes as recently as 2004 [22]. All citrus varieties are susceptible to CLas. Citrus production in Florida including fresh fruit and juice is facing a complete collapse if significant progress is not achieved soon [23].
Management of this pathogen-insect vector complex has been extremely challenging. Intensive pesticide management has done little to halt the spread [24] and currently it is believed that 100% percent of Florida citrus groves are infected with the disease [25]. Critical to reversing the spread of this pathogen and recovering productivity of Florida citrus groves is development of pathogen transmission intervention strategies such as development of resistant citrus genotypes that prevent or reduce insect feeding [26].
Here we use random forests, hidden markov models, and heirarchical cluster analysis to reduce the time required to analyze EPG data. In addition, these analyses point to the presence of additional undescribed feeding states suggesting that the behavior of psyllid stylets within the host plant is more complex than has been recognized.
To evaluate such pathogen transmission intervention strategies, we first sought to remove the data analysis bottleneck present in the current paradigm for monitoring feeding of insects using EPG recordings. To do so, we taught the computer to recognize insect feeding states using pattern recognition algorithms. Specifically, we developed high-throughput automated classification of insect feeding states using supervised classification of Fourier-transformed raw EPG data with random forests models. Random forests models are an ensemble machine learning method that relies on bootstrap aggregation of decision trees [27]. These models have been successfully applied for diverse classification tasks including land cover classification and 3D facial recognition [28, 29].
The computer learned to recognize patterns of insect feeding remarkably well. Overall classification accuracy of random forests models trained on the six human recognized feeding states (Table 1) can reach 97.4±0.1% (95% CI) when compared with human expert annotation (Fig 2; confusion matrix and accuracy statistics in S1, S2 and S3 Tables). Accuracy improved, and can reach close to 100%, when these models are simply asked to identify phloem feeding. Phloem feeding was our primary interest in this case because ingestion and salivation in phloem sieve elements are when pathogen acquisition and inoculation of CLas are presumed to occur (Fig 3). Importantly, these supervised classification algorithms achieved high accuracy when trained on a random 5% subsample of the full dataset. This obviated the need for human manual annotation of 95% of the data and engenders timesavings that begin to enable high-throughput analysis.
Ideally, automated classification of EPG recordings would obviate all human input and allow for real-time monitoring of insect feeding states within the plant or vertebrate subject. This may be possible. Greater than 95% accuracy was achieved using a leave-one-out classification scheme wherein a supervised random forest classifier was trained on a random 5% subsample of 26 of 27 available recordings and then used to classify the remaining recording (S1 Fig). In some cases, accuracy decreased due to variation in waveform patterns generated by insect feeding on different varieties (S2 Fig). Further development of more sophisticated machine learning algorithms should enhance our ability to accurately classify insect feeding and pathogen transmission in real time to more precisely follow stylet behavior within the host.
In addition to the abilities of machine learning algorithms to enable high-throughput screening of pathogen transmission intervention strategies, such models can be used to extend our understanding of the dynamics of insect feeding. We can learn from the computer how to recognize additional patterns of insect feeding. Currently, six distinct feeding states are recognized from EPG recordings of the Asian citrus psyllid based on human observation of waveform patterns correlated with histological studies [6]. We wondered if unsupervised pattern recognition models could identify additional, as yet unrecognized, feeding states.
To do so, we applied hidden Markov models to Fourier-transformed raw EPG data without supplying the algorithm any information about human-annotated insect feeding states. Hidden Markov models use Markov processes to model and uncover hidden states affecting given observations [30] and are used in natural language processing and in predicting protein topology [31–33]. We provided the model with Fourier-transformed time series data from EPG recordings and asked it to classify the data into as many as 12 feeding states (Fig 4). By doing this, the computer could recognize and highlight additional feeding states not discerned through histological studies. Eight-state hidden Markov models successfully resolved phloem feeding states (when pathogen transmission occurs in this system) and identified two additional feeding states within the human-recognized C feeding state thought to correlate with insect stylet passage through plant tissue [6] (Fig 4). These two additional feeding states suggest that the insect is performing two rapidly alternating tasks during passage of the stylets through nonvascular tissues. Additionally, Bayesian information criterion scores from multistate hidden Markov models [34] suggest that there may be many more than the six currently recognized feeding states further emphasizing the dynamic nature of phloem, xylem, and potentially blood feeding in piercing/sucking arthropods (Fig 4).
More information regarding insect feeding patterns can be obtained by applying pattern recognition algorithms to the six human-recognized waveforms identified by histology [6]. Applying hierarchical cluster analysis to frequency distributions extracted from Fourier-transformed EPG data for each feeding state revealed similarities within ingestion (G, E1, and E2) feeding states (Fig 5: left dendrogram) [35]. The frequencies (Fig 5: density plots) produced by psyllid ingestion from xylem (feeding state G), were not significantly different (P > 0.05, from heirarchical cluster analysis) from those produced by phloem salivation and ingestion (E1 and E2, respectively). In contrast, probing and non-probing feeding states (NP, C, and D, respectively) during which ingestion does not occur, produced significantly different frequency patterns compared with those of states associated with pathogen transmission (G, E1, and E2). These results suggested that ingestion from xylem and phloem by the Asian citrus psyllid is accomplished by mechanically similar means.
Further analysis of feeding states provided insight into the nature of pathogen transmission and allowed identification of characteristics that render certain plant varieties more resistant to pathogen infection. Development of resistant citrus genotypes is of primary interest to citrus growers as other methods of controlling citrus greening have proved unsuccessful [24]. Trifoliate genotypes (Table 2), such as Poncirus trifoliata and its hybrids, are under consideration for commercial development. These have been noted for their tolerance to citrus greening [18]. The level of tolerance is yet to be determined, however. When directly inoculated with CLas by graft inoculation with infected buds, trifoliate varieties displayed symptoms of disease progression similar to susceptible Citrus trees [36]. In contrast, under field conditions where trifoliate varieties were only subjected to infection by insect transmission, trifoliate varieties displayed reduced or delayed symptoms. [37].
To compare and contrast insect feeding on different genotypes of trifoliate and non-trifoliate citrus varieties, we applied a hierarchical cluster analysis to 27 recordings of Asian citrus psyllid feeding on nine citrus genotypes [35]. Despite receiving no information on human-annotated feeding states, the computer recognized differences in insect feeding across genotypes. Cluster analysis tended to group recordings of the same variety (Fig 5: top dendrogram). Poncirus (trifoliate) citrus genotypes in particular were more similar to each other and grouped together; multidimensional Euclidean distances within trifoliate genotypes were on average 8.1% (95% CI: 2.2, 13.3%) less than between-variety differences.
These groupings of genotypes correspond to patterns of insect feeding (Fig 5: Heatmap). Genotypes that experienced little to no phloem feeding (states E1 and E2) were grouped together (Fig 5: red box). Those genotypes with limited opportunity for pathogen transmission tended to be trifoliates or trifoliate hybrids that experienced significantly (α = 0.05) less phloem feeding by the psyllid compared with other genotypes (Fig 6). The observed low incidence of phloem feeding on P. trifoliata and trifoliate hybrids suggests a mechanism to explain the observed tolerance of citrus genotypes in the field, despite demonstrated susceptibility to the pathogen by graft inoculation [36, 37]. Poncirus trifoliata may possess physical traits that confer resistance to transmission by interfering with the vector’s ability to attain the phloem. Our results suggest that psyllid feeding may be hindered by physical barriers to stylet passage conferred by fibrous rings of sclerenchyma cells associated with vascular tissue in P. trifoliata [38].
These analyses hold direct implications for prevention of transmission of CLas by its hemipteran insect vector, the Asian citrus psyllid. The low incidence of phloem feeding on varieties of P. trifoliata genotypes and Poncirus x Citrus hybrids confirms these genotypes as sources of resistance for cultivar development, and suggests a potential mechanism for their resistance to infection that can be selected for in the future through traditional breeding or genetic modification [26]. Further development of these strategies and resistance mechanisms will benefit from high-throughput screening and analysis using machine learning algorithms.
While this type of analysis provides insights directly applicable to preventing the spread of greening disease in citrus through high-throughput screening and identification of resistance mechanisms, analysis of insect feeding as described here holds implications for all insect vector-pathogen systems. These results are broadly applicable to development of resistant varieties [39, 40] and management of other plant diseases, including Zebra chip that affects the staple crop potato and is caused by a bacterium closely related to citrus greening disease [41]. Insights into the dynamics of insect feeding gained from machine learning analysis of electrical penetration graphs can be used to design novel intervention strategies to disrupt transmission of insect-transmitted pathogens of agricultural crops, livestock, and humans. Testing and screening of strategies such as genetic manipulation, RNAi, or chemical deterrents to feeding and transmission will benefit from high-throughput, human independent classification via machine learning. These electrical penetration graph analyses that extend human insight and reduce time investment will engender advances in both basic and applied investigation of insect transmitted pathogens and advance discovery of tools to prevent the spread of disease in agricultural crops, livestock, and humans.
EPG recordings were performed using a Giga-8 DC-EPG system (Wageningen, the Netherlands) to record the feeding activities of adult Asian citrus psyllids on nine trifoliate and citrus varieties. Psyllids were tethered to recording equipment using fine gold wire and silver conducting glue then settled on the adaxial midrib of a leaf (Fig 1). To complete the circuit, a second electrode electrode (ground electrode) was inserted into the saturated soil (70–80% moisture content) of the pot containing the citrus plant. EPG recordings were conducted within a Faraday cage in a climate-controlled laboratory (25 ± 1°C, 60 ± 5% RH) for 8 to 21 h under lighted conditions. Waveforms were classified by visual inspection by a trained expert according to previous reports [6, 42] into six feeding states: salivary sheath secretion and stylet passage (C), first contact with phloem (D), salivation at phloem (E1), phloem ingestion (E2), xylem ingestion (G) or no probing (NP). Twenty-seven EPG recordings totaling 470 hours on nine different citrus varieties were used to explore machine learning for waveform recognition.
Raw voltage data from psyllid feeding were recorded using WinDaq Data acquisition and Playback software (DataQ Instruments). Data were classified by visual inspection and annotated using the WinDaq data browser then exported to comma separated value files. Raw data from comma-separated values were then loaded in the R version 3.2.2 computing environment [43] and converted from the time domain to the frequency domain using fast fourier transform [44]. The six frequencies with the highest magnitudes, often harmonics, were extracted for use in machine learning algorithms.
Fast Fourier transformed data were randomly split into training and test sets for each recording. A random five percent subset of each recording was used to train a supervised random forests model with 3 repeated 10 fold cross validation and was then tested on the remaining ninety five percent of the recording. This procedure was used to classify all six human recognized feeding states, and to differentiate between phloem (E1 and E2) and nonphloem (C, D, NP, and G) feeding states. Out of sample accuracy, based on comparison to human expert classification of the test set, and ninety-five percent confidence intervals averaged for all feeding states are reported. 50:50, and 95:5 training to test set schemes were also considered for the analysis and did not produce differences in overall accuracy. A 5% training to 95% test set was considered most advantageous in terms of reducing human labor while maintaining high accuracies. Using randomly sampled training sets less than 5% of the overall dataset increased the likelihood of missing certain feeding states and lowered classification accuracy accordingly.
A leave one out classification scheme was pursued to determine the possibility of classification without additional human input. To that end, a random five percent subsample of each feeding state from each of 26 human annotated recordings was used to train a random forests model with 3 repeated 10 fold cross validation. The model was then asked to classify the 27th recording; results of such classification were compared with human expert annotation to determine out of sample accuracy. This procedure was then repeated and used to classify each of the 27 recordings, one of which was left out each time.
To explore the possibility of additional insect feeding states beyond those six currently recognized by humans, hidden Markov models were applied to the dominant frequencies extracted from Fourier transformed data and asked to separate the electrical penetration graph time series into up to 12 feeding states. Parameter estimation for the hidden Markov models was accomplished through use of the expectation maximization algorithm and the posterior state sequence was recovered by the Viterbi algorithm [45–47]. Bayesian information criterion was used to penalize additional feeding states [34].
To explore similarities between varieties and insect feeding states, hierarchical cluster analysis was applied to density distributions of dominant frequencies extracted from Fourier transformed electrical penetration graph recordings. Variety similarity was determined through bootstrapping 1000 times the difference in Euclidean distance among and between frequency density distributions of trifoliate varieties. Comparison of unsupervised classification using hierarchical clustering to human annotated states was accomplished through construction of a heatmap presenting the percent median feeding bout time scaled within each feeding state. Comparison of phloem feeding between trifoliate and non-trifoliate varieties was accomplished through bootstrapping 1000 times the difference in median phloem (feeding states E1 and E2) feeding time.
After exportation from the WinDaq data collection and browser software, all data were loaded into R version 3.2.2 for further analysis [43]. RStudio was used as a development environment [48]. Packages provided additional functionality and facilitated analysis: data.table [49], dplyr [50], tidyr [51], and pryr [52] for data management, caret [53] and randomForest [54] for implementation of random forest models, foreach [55], doParallel [56], and doMC [57] for parallel implementation of analysis, pvclust [58] and ggdendro [59] for hierarchical cluster analysis, depmixS1 [60] for implementation of Hidden Markov Models, and ggplot2 [61] for developing graphics.
|
10.1371/journal.pgen.1005764 | Histamine Recycling Is Mediated by CarT, a Carcinine Transporter in Drosophila Photoreceptors | Histamine is an important chemical messenger that regulates multiple physiological processes in both vertebrate and invertebrate animals. Even so, how glial cells and neurons recycle histamine remains to be elucidated. Drosophila photoreceptor neurons use histamine as a neurotransmitter, and the released histamine is recycled through neighboring glia, where it is conjugated to β-alanine to form carcinine. However, how carcinine is then returned to the photoreceptor remains unclear. In an mRNA-seq screen for photoreceptor cell-enriched transporters, we identified CG9317, an SLC22 transporter family protein, and named it CarT (Carcinine Transporter). S2 cells that express CarT are able to take up carcinine in vitro. In the compound eye, CarT is exclusively localized to photoreceptor terminals. Null mutations of cart alter the content of histamine and its metabolites. Moreover, null cart mutants are defective in photoreceptor synaptic transmission and lack phototaxis. These findings reveal that CarT is required for histamine recycling at histaminergic photoreceptors and provide evidence for a CarT-dependent neurotransmitter trafficking pathway between glial cells and photoreceptor terminals.
| Neurotransmitter transporters that remove neurotransmitters and recycle them after their release have particular importance at visual synapses, which must signal at high frequencies and therefore required rapid clearance of neurotransmitters from the synaptic cleft. In this study, we identified a SLC22 family transporter, CarT, in the visual system of Drosophila, which is exclusively located to photoreceptor terminals in the lamina neuropil and is responsible for taking up carcinine, an inactive histamine metabolite, from surrounding glia. Loss of CarT disrupts the regeneration of histamine and blocks neurotransmission at photoreceptor cell synapses. Our work provides direct evidence for a local histamine recycling pathway between glial cells and photoreceptor terminals, and shows that a CarT-dependent histamine/carcinine shuttle pathway is critical for maintaining the normal histamine content of neurons.
| Histamine is an important chemical messenger known to be involved in a broad spectrum of biological processes such as inflammation and gastric acid secretion. It is also recognized as an important neurotransmitter [1]. Recycling histamine at histaminergic synapses is a key event both in maintaining synaptic transmission and in terminating histamine’s action on postsynaptic neurons. The Drosophila visual system uses histamine as the neurotransmitter at photoreceptor synapses, and provides a good genetic model for studying histamine, its metabolism and recycling [2]. The compound eye of Drosophila is composed of ~800 ommatidia, each of which contains eight photoreceptor cells. Of the latter, R1-R6 photoreceptors in each ommatidium project axons from the retina to the underlying lamina neuropil, where they are organized into synaptic modules called cartridges. R7/R8 photoreceptors project axons to the second neuropil, the medulla [3–6]. In lamina cartridges, three epithelial glial cells normally envelop six photoreceptor terminals [7].
Although the synthesis of histamine from histidine occurs de novo under the action of histidine decarboxylase (Hdc) in photoreceptor cells, recycling of histamine is reported to be the dominant pathway for maintaining the histamine content in photoreceptors [8,9]. Both pathways, de novo synthesis and recycling, are required to maintain an adequate content of histamine in photoreceptor cells. Disrupting either pathway affects visual synaptic transmission in Drosophila in the long term [8,10]. Upon light stimulation, photoreceptor terminals release histamine as a neurotransmitter, which activates histamine-gated chloride channels (HisClA) on large monopolar cells (LMCs) in the lamina and hyperpolarizes these postsynaptic neurons [2,4,11]. After its release, histamine is taken up by lamina glia and conjugated to β-alanine, converting it to carcinine by the N-β-alanyl-dopamine synthase, Ebony, which is expressed in epithelial glia [10,12,13]. The metabolized histamine conjugate, carcinine, is then transported back into the photoreceptors and hydrolyzed back to histamine by Tan, an N-β-alanyl-dopamine hydrolase[10,14]. Despite knowledge of these pathways, little is known about the critical step by which carcinine is transported back to the photoreceptors. It has been proposed that the gene inebriated (ine) might encode a carcinine neurotransmitter transporter in photoreceptor cells to take up carcinine from synaptic cleft [15]. However, in this study, we show that Ine fails to function in any clear way in photoreceptor cells. In addition, the cellular location for carcinine uptake, the trafficking route by which it is returned to the photoreceptor cells where the Tan enzyme has to act, and the transporters responsible for carcinine uptake, all remain controversial. Recently, it has been suggested that metabolites of histamine are transported between glia and the cell bodies of photoreceptors through networks of intercellular gap junctions [9].
We identified a photoreceptor cell-enriched neurotransmitter transporter, CarT, which is able to transport carcinine across the membranes of photoreceptors. CarT is predominantly localized to photoreceptor terminals. The cart mutant flies are defective in photoreceptor synaptic transmission, and as a result lack phototaxis. In addition, we found that a human homologue of CarT, Organic Cation Transporter (OCT2), can also transport carcinine in vitro and is thus able to reverse synaptic transmission defects in cart mutant flies. We therefore propose the presence of a novel pathway for histamine recycling, in which the carcinine transporter CarT efficiently takes up carcinine that is released locally from glial cells lying in close vicinity to photoreceptor terminals.
Given that the histamine/carcinine shuttle in the visual system occurs between photoreceptors and surrounding glia cells [7], and that the enzyme Tan responsible for hydrolyzing carcinine to release histamine is exclusively expressed in photoreceptor cells, we assumed that the neurotransmitter transporter responsible for taking up carcinine must be enriched in photoreceptor cells. The gene glass (gl) gene encodes a zinc finger transcription factor, and glass mutations specifically remove photoreceptor cells, but leave other cell types intact. Mutations of glass specifically remove photoreceptor cells, and thus largely abolish the expression of mRNA transcripts of photoreceptor-enriched genes, such as the gene encoding major rhodopsin neither inactivation nor afterpotential E (ninaE). Expression of ninaE is greatly reduced in the heads of gl3 flies relative to wild-type (w1118) heads (Fig 1A).
By comparing mRNAs isolated from wild-type heads with gl3 heads or wild-type bodies, we identified a list of genes that are expressed predominantly in photoreceptor cells. We examined both this RNA-seq data and a DNA microarray data set, which screened for genes expressed predominantly in photoreceptor cells and the compound eyes respectively [16]. This enabled us identify candidate genes that might encode the carcinine transporters. CG9317 and CG3790 are both candidate genes for eye-enriched neurotransmitter transporters. Both proteins share significant amino acid identities with the mammalian solute carrier family 22 (SLC22) family proteins, including the mouse OCT2 and OCT3 (S1 Fig). The expression of CG9317 mRNA was greatly reduced in gl3 fly heads, indicating that CG9317 is expressed predominantly in photoreceptor cells (Fig 1A and 1B). In contrast, the expression levels of the retinal pigment cell marker gene retinol dehydrogenase B (rdhB) remain unchanged for both gl3 flies and wild-type flies (Fig 1A) [17].
We next conducted in vitro experiments to examine whether CG9317 and CG3790 can transport carcinine. We expressed mCherry-tagged proteins in S2 cells, and used immunolabeling to examine the intracellular signals for histamine or carcinine. Carcinine or histamine was added to the medium to yield final concentrations of 20μM. After three-hour incubations, the intracellular carcinine or histamine signal was examined. No transporter activity for either carcinine or histamine was detected in S2 cells expressing mCherry alone (Fig 1C and S2A Fig). There was no immunosignal for either carcinine or histamine after expressing Ine, which indicates the probability that Ine does not transport either carcinine or histamine under the conditions tested (Fig 1D and S2B Fig). We next examined the candidate carcinine transporters that are highly expressed in eyes, including CG9317 and CG3790 [16]. CG3790 failed to transport either carcinine or histamine (Fig 1E and S2C Fig). We confirmed these results by using a specific rat anti-carcinine antibody from a different source [18] (S3 Fig). In contrast, a clear immunosignal for carcinine but not histamine was detected in cells expressing CG9317 (Fig 1F and S2D Fig). When we expressed histidine decarboxylase (Hdc) in S2 cells, immunosignal for histamine was observed, which served as a positive control, validating our in vitro histamine immunolabeling method (S2E Fig). These findings suggest that CG9317 encodes a carcinine transporter, which we therefore named CarT (Carcinine Transporter).
To characterize the requirement for CarT in transmitting visual signals, we generated two different null mutations in the cart gene using the CRISPR-associated single-guide RNA system (Cas9)(Fig 2A)[19]. We identified fly lines containing these cart1 and cart2 mutations by PCR using genomic primers outside of the deleted regions (Fig 2A). Full-length PCR products were detected in wild-type flies, whereas shorter PCR products were detected in the cart1 and cart2 mutant lines, indicating the disruption of the cart locus in cart1 and cart2 flies (Fig 2B). The cart genomic region in both mutations was sequenced, and 1112 and 2344 bp fragments were deleted in cart1 and cart2 mutants respectively (S4 Fig).
As cart mutants were not lethal, so we undertook electroretinogram (ERG) recordings directly. ERG recordings are extracellular recordings that measure the summed responses of all retinal cells in response to light. Upon exposure to light, an ERG recording from a wild-type fly contains a sustained depolarizing response from the photoreceptors, and “on” and “off” transients originating from synaptic transmission to the lamina [20] (Fig 2C). Mutations with defective synaptic transmission have obvious reductions in their “on” and “off” transients [6]. As in mutants of genes involved in histamine recycling, ERG transients were not observed in cart1, cart2, or cart1/ cart2 mutant flies (Fig 2C). Phototaxis is a visual behavior that requires the integrity of the neuron circuits of the visual system [21], and defective synaptic transmission of visual signals results in poor phototaxis [22]. Significantly reduced phototaxis was associated with all the cart1, cart2, and cart1/cart2 mutations (Fig 2D). To further confirm that the loss of visual synaptic transmission resulted from mutations of the cart locus, we generated a Pcart-cart transgenic fly line expressing the cart cDNA under the control of the cart promoter. The Pcart-cart transgene reversed the loss of “on” and “off” transients and restore phototaxis in the cart1 mutant flies (Fig 2C and 2D).
Tan, the hydrolase that deconjugates carcinine and releases histamine, localizes to photoreceptor cells and functions downstream of the transport of carcinine. A carcinine transporter coupled with Tan’s action should therefore be expressed and should function in photoreceptor cells. It has been suggested that ine encodes a putative carcinine neurotransmitter transporter in photoreceptor cells [15]. We used the eyeless-GAL4 UAS-FLP (EGUF)/hid technique to generate genetically mosaic flies [23]. The compound eyes of these mosaic flies comprise cells homozygous for a selected mutation, but forming part of an entire mosaic fly that is elsewhere heterozygous for the mutation. Therefore, if Ine functions in the compound eye of Drosophila, eye-specific mutations of ine in ine mosaic flies should mirror at least the same ocular defects in synaptic transmission as those present in the ine mutants.
We observed that ERG recordings from wild-type eyes have normal “on” and “off” transients (Fig 3A and 3A’). Mutations in both the ebony (e1) and the tan (tan1) genes disrupt histamine recycling and this results in the loss of “on” and “off” transients in their ERG recordings (Fig 3B and 3C) [10]. As expected, e1 mosaic flies in which all photoreceptors were homozygous mutant for ebony had wild-type “on” and “off” transients. This is because Ebony is not required in the photoreceptors but is required in glial cells lying outside them (Fig 3B’). As Tan functions in the photoreceptor cells of the compound eye, the tan1 mosaic which lacks tan expression in the photoreceptors displayed reduced “on” and “off” transients (Fig 3C’).
The ERG responses of ine mutants (ineMI05077) contain prominent oscillations superimposed on the sustained depolarizating response and they also have reduced “on” and “off” transients (Fig 3D) [15]. The latter phenotype indicates impaired photoreceptor synaptic transmission. However, as with the ebony mutants, heterozygous flies with homozygous ine mutant compound eyes (ine mosaic flies) had wild-type ERG responses with normal “on” and “off” transients (Fig 3D’), indicating that Ine does not function obligatorily in photoreceptor cells. Therefore, it is unlikely that Ine is directly or necessarily responsible for carcinine uptake at the photoreceptor cell membrane, as previously suggested. Its possible role as a transporter elsewhere is not addressed by these experiments.
Given that expression of the cart gene is enriched in photoreceptor cells, we assumed that CarT is required in photoreceptor cells for synaptic transmission. As expected, homozygous cart1 mutant eyes lacked “on” and “off” transients despite the heterozygous background elsewhere (Fig 3E and 3E’). This finding indicates that CarT functions in the compound eyes. Photoreceptor cells and retinal pigment cells are the two major cell types in the compound eye. To confirm the retinal cell type in which CarT functions, we expressed CarT specifically in photoreceptor cells using the ninaE promoter or in retinal pigment cells using the rdhB promoter [17] [24]. Photoreceptor-enriched expression of CarT by PninaE-cart restored both the “on” and “off” transients and phototaxis in cart1 mutant flies, whereas expression of CarT in pigment cells through PrdhB-cart did not (Fig 3F and 3G). These results strongly support the interpretation that CarT functions in photoreceptor cells to maintain synaptic transmission.
In addition, we extended these ERG results by phototaxis assays. Wild-type flies displayed positive phototactic behavior, whereas flies that were homozygous mutant for ebony, tan, ine, or cart all displayed poor phototaxis, indicating that these genes are required for visual synaptic transmission (Fig 3G). Consistent with the ERG results, phototaxis was significantly reduced in the mosaic eyes of tan1 and cart1 compared with wild-type flies, whereas phototaxis of both the e1 and the ineMI05077 mosaic flies did not differ from that in wild-type flies (Fig 3G). These results suggest the possibility that CarT rather than Ine functions as a carcinine transporter in photoreceptor cells.
Trafficking of carcinine into photoreceptors is a key step in histamine recycling in Drosophila. As we have proposed here that CarT functions as a carcinine transporter acting at the photoreceptor cell membrane, we examined the localization of CarT to photoreceptor cells to evaluate the cellular location of carcinine transport. Since multiple attempts to generate an anti-CarT antibody failed, we eventually generated transgenic flies that expressed mCherry-tagged CarT driven by the cart promoter. Importantly, the Pcart-cart-mcherry transgene completely reversed the loss of “on” and “off” transients in cart mutant flies (Fig 2C). Although CarT was expressed throughout the photoreceptor neurons, the CarT signal was predominantly detected in the lamina layer where it was marked by the Ebony immunosignal, and not appreciably in the region of the retina (Fig 4A). In cross sections at high magnification we observed that CarT was not co-localized with Ebony to the epithelial glial cells (Fig 4B), but rather co-localized with the photoreceptor cell axon marker Tan to both the lamina and medulla neuropils, to which the R1-R6 and R7/R8 photoreceptors project their axons respectively (Fig 4C and 4D). The finding that CarT expression is enriched in photoreceptor terminals is consistent with the assumption that photoreceptor cells take up carcinine mainly from the local synaptic cleft in the lamina, rather than by a long-distance histamine recycling pathway which is mediated by lamina glia and a retinal pigment cell network [9]. However, we cannot exclude the existence of a long-term trafficking pathway for carcinine.
Given that the evidence so far suggests that cart acts to transport carcinine into the photoreceptor, where tan then acts to hydrolyze it and release histamine, we next sought to examine whether loss of cart would decrease histamine labeling. We labeled head cross sections from the cart1 mutant and from the w1118 control with anti-histamine antibody. The distribution of histamine signal in cart1 mutant flies relative to their w1118 controls reveals a clear loss of photoreceptor signal (Fig 5A and 5B), compatible with the mutant’s inability to take up carcinine and so liberate histamine. In the enlarged images, it is clear that cart1 mutants showed a dramatic decrease in labeling for histamine in R1-R6 photoreceptor terminals in the lamina, and in R7/R8 photoreceptor terminals in the medulla (Fig 5C and 5D). In contrast to the weak label in R1-R6 photoreceptor terminals in the lamina, a strong label was seen in the underlying marginal glia at the proximal lamina in the cart1 mutant (Fig 5C and 5D) [3,9,25]. The labeling of this region suggests that histamine might be accumulated at an ectopic site in the cart mutant.
CarT belongs to the SLC22 protein family and is highly homologous to the mammalian OCT2 protein. We therefore wondered whether heterologous expression of OCT2 in cart mutant flies would restore the synaptic transmission of photoreceptors. OCT2 is known to mediate low affinity transport of some monoamine neurotransmitters [26]. However, it is not known whether OCT2 is able to transport carcinine. We performed in vitro assays to determine whether OCT2 can transport carcinine. After expressing OCT2 in S2 cells, carcinine was taken up by the OCT2-positive cells (Fig 6A). These results indicated that OCT2 can indeed transport carcinine. We next generated a PninaE-oct2 transgene to express OCT2 in photoreceptor cells only, and introduced this transgene into the cart1 mutant background. We found that the expression of human OCT2 in cart1 mutant fly photoreceptor cells fully restored both the “on” and “off” transients and phototaxis in cart1 flies (Fig 6B–6E). These results demonstrated a conserved function for OCTs in both a mammal and Drosophila.
We used high-performance liquid chromatography (HPLC) to examine the in vivo contents of histamine as well as carcinine and β-alanine, the major metabolites in histamine recycling [18,27]. As expected, in the heads of the tan1 mutant flies, which are defective in their capacity to hydrolyze carcinine into histamine and β-alanine, the head contents of both histamine and β-alanine were significantly reduced (Fig 7A and 7B). The lack of carcinine uptake by photoreceptor cells in cart1 mutant flies ultimately depletes carcinine in these cells, which reduces the production of histamine and β-alanine mediated by Tan (Fig 7A and 7B). The reduced head contents of histamine and β-alanine are therefore in agreement with the hypothesis that CarT transports carcinine.
In contrast to histamine and β-alanine, the content of carcinine in the tan1 mutant heads was approximately three fold higher than the content in wild-type heads, which we interpret to result from diminished hydrolysis of carcinine in photoreceptor cells (Fig 7C). If the flies were not able to transport carcinine into photoreceptor cells for hydrolysis, there should be a greater amount of carcinine in fly heads. As expected, in cart1 mutants, the head content of carcinine was significantly increased. However, the content of carcinine in the cart1 mutant was not increased to the same extent as in the tan1 mutant flies.
Although histamine is an important neurotransmitter known to regulate multiple physiological processes, the mechanism by which histamine content is regulated in the nervous system still remains to be elucidated. Our study identifies a mechanism and pathway for the uptake of a primary metabolite of histamine, which has hitherto defied analysis in any nervous system.
Insofar as histamine is the primary neurotransmitter released by photoreceptors in flies [28], the ease with which photoreceptor function and anatomy can be assayed has made the compound eye the preferred system to study histamine recycling. In particular the eye lends itself readily to the identification of genes that regulate neurotransmission, by enabling comprehensive genetic screens [23,29]. Studies in flies have previously identified a histamine/carcinine recycling pathway that involves two enzymes, Ebony, expressed in the epithelial glia, and Tan, expressed in the photoreceptor cells [13,30]. However, the key neurotransmitter transporters required for the histamine/carcinine shuttle pathway have not been identified. Conceptually, the putative carcinine transporter should be functionally coupled to Tan for the uptake of carcinine into photoreceptor cells and its subsequent hydrolysis. For this, both should colocalize to photoreceptors as we have shown in this study.
In this study, we identified a new SLC22A family protein CarT and provided evidence that it is functionally coupled with Tan as a photoreceptor cell-enriched carcinine transporter. CarT is predominantly localized to photoreceptor terminals and is able to transport carcinine in vitro. The decrease in head histamine and β-alanine and the increase in head carcinine in cart1 support this hypothesis. The reduction in histamine content in cart1 mutants is ~60%. This amount corresponds rather closely to the reduction in head histamine seen in the mutant sine oculis, which lacks compound eyes and has 28% of the histamine found in the wild-type [27]. The reduction in sine oculis suggests that residual head histamine is not located in the compound eye visual system. In the same way, the reduction in cart1 is not accessible to photoreceptor synaptic transmission. Moreover, mutant ebony, in which head histamine content is reduced by 50%, has an abnormal ERG and phototaxis, corresponding to the strong ERG and phototaxis defects seen in the cart1 mutant. Consistent with these HPLC data, we also found a clear difference in the immunosignal for histamine between cart1 and control fly photoreceptors. The cart head increase in carcinine is not as high as that observed in the head of tan mutants, which raises the question of why the increases in carcinine in tan and cart mutant flies are not similar. This may be because carcinine released in the synaptic cleft in cart1 mutants is removed by other cells, or alternatively it may enter the hemolymph and be excreted. Consistent with the latter, the carcinine content in the abdomen is increased by 43% in cart1 compared with control w1118 flies. We cannot address the carcinine transport by other cells in the lamina, in particular the epithelial and marginal glia, which surround the photoreceptor terminals and which contain carcinine [18], which we propose must therefore express other carcinine transporters.
All of the known plasma membrane neurotransmitter transporters are members of the Solute Carrier (SLC) family of proteins [31]. The most extensively studied of these transporters are members of the SLC6 subfamily, a group of Na+/Cl-—dependent transporters for serotonin, dopamine, norepinephrine, GABA and glycine [32]. OCTs, which belong to the SLC22 subfamily, are known to mediate sodium-independent transport of positively charged organic compounds [33]. The expression of OCT2 in neurons has been evaluated previously, but the neuronal function of OCT2 has not been explored sufficiently [26,33,34]. Carcinine has been identified as a native metabolite related to histamine in multiple tissues in mammals, where it may serve as an antioxidant for scavenging toxic active oxygen species, especially in retinal photoreceptors [35,36]. Our findings that OCT2 can transport the inactive histamine metabolite carcinine both in vitro and in vivo suggests a possible new mechanism for OCTs to function in neurotransmitter recycling and cell protection.
The histamine/carcinine shuttle pathway plays a dominant role in maintaining an adequate level of histamine in photoreceptors. Evidence for the direct uptake of histamine into photoreceptor cells is lacking, insofar as Ebony is necessary to rescue ERG transients in histamine-fed hdc mutant flies [37]. In addition, in our model S2 cells expressing CarT fails to transport histamine, providing further support for the hypothesis that direct uptake of histamine into the photoreceptor terminals may not occur. Although the enzymatic deconjugation of carcinine to yield histamine has been well established, the route through which carcinine is then trafficked back to the photoreceptor has not been established. It has been suggested that recycling of carcinine to photoreceptor cells involves a long-distance pathway mediated by a gap-junction dependent network of lamina and retinal pigment cells [9]. In our study, we observed that CarT is predominantly localized to the terminals of photoreceptor neurons, rather than to their cell bodies in the retina layer, which suggests that carcinine is transported back to photoreceptor cells mainly from the synaptic cleft in the lamina (Fig 7D). It is also possible that this local pathway works in parallel with the long-distance neurotransmitter recycling pathway. Finally, data from the current study together with previous reports now provide evidence for a more complete histamine/carcinine recycling pathway, one which is critical for maintaining the normal histamine content of neurons (Fig 7D).
To complete the model of how histamine is recycled in the fly’s eye (Fig 7D), the remaining question concerns how histamine is transported into the epithelial glia and how carcinine is then transported out of the glia. No specific histamine transporter has been found, in either insects or vertebrates. In insects a mechanism for the fast removal of histamine from the synaptic cleft is essential to maintain the rapid signaling required for insect vision. One transporter may be White [38], but the problem is that in eukaryotes all known ABC transporters move substrates in the opposite direction i.e. out of the cell. To complete the return path for the carcinine will require us to identify how carcinine is exported out of the epithelial glia. To identify the transporter for this function it will be necessary to identify genes, for example from the expression of mRNA transcripts of genes, such as ebony [12, 13], that are enriched in the epithelial glia, in an approach that parallels the one we have adopted here to identify CarT. The transport of β-alanine, the other substrate needed for carcinine synthesis, seems to be of minor importance because this amino acid is present in the head in concentrations greatly exceeding those needed for histamine recycling and it can be also easily synthesized on demand from aspartate or uracil. Answering these questions is necessary to complete the current scheme (Fig 7D) for the recycling of histamine, to which our findings now identify CarT as the photoreceptor uptake transporter.
The following stocks were obtained from the Bloomington Stock Center: (1) 122, e1; (2) 130, tan1; (3) 38094, ineMI05077; (4) 3605, w1118; and (5) 24749, M(vas-int.Dm)ZH-2A;M(3xP3-RFP.attP)ZH-86Fb. The (nos-Cas9)attP2 flies were obtained from the lab of Dr. J. Ni at Tsinghua University, Beijing, China. The ey-flp;GMR-hid CL FRT40A/Cyo, ey-flp;FRT42D GMR-hid CL/Cyo, GMR-hid CL FRT19A/FM7;ey-flp, and ey-flp;FRT82B GMR-hid CL /TM3 flies were maintained in the lab of Dr. T. Wang at the National Institute of Biological Sciences, Beijing, China.
The cart, CG3790, ine, and Hdc cDNA sequences were amplified from GH05908, GH20501, LP16156, and LD44381 cDNA clones obtained from DGRC (Drosophila Genomics Resource Center, Bloomington, IN, USA). The oct2 cDNA sequences were amplified from IOH56335 cDNA clones obtained from Ultimate™ ORF clones (Thermo Fisher Scientific, Waltham, USA). Their entire CDS sequences, excluding the stop codon, were subcloned into the pIB-cmcherry vector (Invitrogen, Carlsbad, USA) for expression in S2 cells. To construct PninaE-cart, PrdhB-cart, and PninaE-oct2, the entire coding region of cart and oct2 was amplified from cDNA clones and cloned into the pninaE-attB and prdhB vectors (both gifts from the lab of Dr. C. Montell at the University of California, Santa Barbara, USA)[17,24,39]. To construct Pcart-cart-mcherry, the promoter region (-2579 to +11 base pairs 5' to the transcription start site) of the cart gene was amplified from genomic DNA, and cart-mcherry was amplified from pIB-cart-mcherry. These constructs were injected into M(vas-int.Dm)ZH-2A;M(3xP3-RFP.attP)ZH-86Fb embryos, and transformants were identified on the basis of eye color. The (3xP3-RFP.attP) locus was removed by crossing with P(Crey) flies.
The cart1 and cart2 mutations were generated using the Cas9/sgRNA system as described previously [19]. Three recognition sequences of guiding RNA to the cart locus were designed with tools available at the following website http://www.flyrnai.org/crispr2/ (sgRNA1: AAAACCGCACGGTATGCAGG, sgRNA2: CCTGTCCGGCGTCACTTATC, sgRNA3: TGAGCGTCATGGACACCCAG). These were cloned into the U6b-sgRNA-short vector. The pU6-sgRNA1 and pU6-sgRNA2 plasmids were used to generate the cart1 mutant flies, while pU6-sgRNA1 and pU6-sgRNA3 were used to generate the cart2 mutant flies. Plasmids were injected into the embryos of (nos-Cas9)attP2 flies. The F1 progeny were screened by PCR to identify the cart1 and cart2 deletions, using the following primers:
pF: 5’-TGTCGCTACAAATCTTAGATCCAA-3'
pR: 5’-CCATGTCAGATATTGAGGACAACG-3’
Two glass microelectrodes filled with Ringer’s solution were inserted into small drops of electrode cream (Sigma, New Jersey, USA) placed on the surfaces of the compound eye and the thorax. A Newport light projector (model 765) was used for stimulation. The source light intensity was 2000lux, and the light color was orange (the source light was filtered by FSR-OG550 filter). ERG signals were amplified with a Warner electrometer IE-210 and recorded with a MacLab/4 s A/D converter and the clampelx 10.2 program (Warner Instruments, Hamden, USA). All recordings were carried out at 23°C.
S2 cells were grown in Schneider’s Drosophila medium with 10% Fetal Bovine Serum (Gibco, Carlsbad,USA), and transfected with vigofect reagent (Vigorous Biotechnology, Beijing, China). Carcinine or histamine was added to the medium to yield final concentrations as indicated in the Figure legends. After incubation for 3h, S2 cells were transferred to poly-L-lysine-coated slices, fixed with 4% paraformaldehyde(for carcinine immunolabeling) or 4% 1-ethyl-3-(3-dimethylaminopropyl) carbodiimide (EDAC)(for histamine immunolabeling) for 30min at 25°C, and incubated with rabbit anti-carcinine/histamine (1:100, ImmunoStar, USA)[9] or rat anti-carcinine antibodies (1:100, raised by Dr. Gabrielle Boulianne, from the lab of Dr. I. A Meinertzhagen) [18]. Goat anti-rabbit lgG conjugated to Alexa 488 (1:500, Invitrogen, CA) and goat anti-rat lgG conjugated to Alexa 488 (1:500, Invitrogen, CA) were used as secondary antibodies, and images were recorded with a Nikon A1-R confocal microscope.
Fly heads were fixed with 4% paraformaldehyde for 2h at 4°C or 4% EDAC (for histamine staining), and immersed in 12% glucose overnight at 4°C. The heads were embedded in O.C.T™ compound (Tissue-Tek, Torrance, USA), and 10μm thick cryosections were cut. Immunolabeling was performed on cryosections sections with mouse anti-24B10 (1:100, DSHB, http://dshb.biology.uiowa.edu/), rat anti-RFP (1:200, Chromotek, Martinsried, Germany), rabbit anti-Ebony (1:200, lab of Dr. S. Carroll, University of Wisconsin, Madison, USA), and anti-Tan (1:200, lab of Dr. B. Hovemann, Ruhr Universität Bochum, Germany) [30] as primary antibodies. For histamine staining, rabbit anti- histamine (1:100, ImmunoStar, USA) was used as a primary antibody. The antibody was preadsorbed with carcinine as previously reported [9]. Goat anti-rabbit lgG conjugated to Alexa 488 (1:500, Invitrogen, USA), goat anti-rat lgG conjugated to Alexa 568 (1:500, Invitrogen, USA) and goat anti-mouse lgG conjugated to Alexa 647 (1:500, Jackson ImmunoResearch, USA) were used as secondary antibodies. The images were recorded with a Nikon A1-R confocal microscope.
A transparent glass tube of 20 cm long and 2.5 cm in diameter was used in this assay. A white light source (with a light intensity of 6000lux) was put at one end of the glass tube, and dark-adapted flies were collected and gently tapped into the other end of the tube. The tube was placed horizontally in the dark, and we counted the number of flies that walked past an 11-cm mark on the tube within 90s after turning the light on. Phototaxis was calculated by dividing the number of flies that walked past the mark as a proportion of the total number of flies. These assays were performed under dark conditions. To quantify the phototactic behaviors of each genotype, three groups of flies were collected for each genotype and three repeats made for each group. Each group contained ≥ 20 flies. Results were expressed as the mean of the mean values for the three groups.
Total RNA was prepared from the heads of three-day-old flies using Trizol reagent (Invitrogen, Carlsbad, USA), followed by TURBO DNA-free DNase treatment (Ambion, Austin, USA). Total cDNA was synthesized using an iScript cDNA synthesis kit (Bio-Rad Laboratories, USA). iQ SYBR green supermix was used for the real-time PCR (Bio-Rad Laboratories, USA). Three different samples were collected from each genotype. The primers used for qPCR were as follows:
ninaE-fwd, 5’-ACCTGACCTCGTGCGGTATTG-3’
ninaE-rev, 5’-GGAGCGGAGGGACTTGACATT-3’
gpdh-fwd, 5’-GCGTCACCTGAAGATCCCATG-3’
gpdh-rev, 5’-CTTGCCATACTTCTTGTCCGT-3’
rdhB-fwd, 5’-TTGAGGCACTCAGGGATCAAG-3’
rdhB-rev, 5’-CACCACATTCGTGTCGAACAG-3’
cart-fwd, 5’-TACAGCACAAGGGTCTCATCC-3’
cart-rev, 5’-AGACCATCCTAATCACGCTGAG-3’
To measurement the total head contents of histamine, β-alanine, and carcinine, flies were decapitated and their heads collected as previously reported [10]. The heads were then processed and analyzed using HPLC with electrochemical detection, all as previously reported [18,27]. Each sample contained ~50 Drosophila heads, and the mean values from five samples were calculated.
|
10.1371/journal.pntd.0004228 | Trachoma and Relative Poverty: A Case-Control Study | Trachoma is widely considered a disease of poverty. Although there are many epidemiological studies linking trachoma to factors normally associated with poverty, formal quantitative data linking trachoma to household economic poverty within endemic communities is very limited.
Two hundred people with trachomatous trichiasis were recruited through community-based screening in Amhara Region, Ethiopia. These were individually matched by age and gender to 200 controls without trichiasis, selected randomly from the same sub-village as the case. Household economic poverty was measured through (a) A broad set of asset-based wealth indicators and relative household economic poverty determined by principal component analysis (PCA, (b) Self-rated wealth, and (c) Peer-rated wealth. Activity participation data were collected using a modified ‘Stylised Activity List’ developed for the World Bank’s Living Standards Measurement Survey. Trichiasis cases were more likely to belong to poorer households by all measures: asset-based analysis (OR = 2.79; 95%CI: 2.06–3.78; p<0.0001), self-rated wealth (OR, 4.41, 95%CI, 2.75–7.07; p<0.0001) and peer-rated wealth (OR, 8.22, 95% CI, 4.59–14.72; p<0.0001). Cases had less access to latrines (57% v 76.5%, p = <0.0001) and higher person-to-room density (4.0 v 3.31; P = 0.0204) than the controls. Compared to controls, cases were significantly less likely to participate in economically productive activities regardless of visual impairment and other health problems, more likely to report difficulty in performing activities and more likely to receive assistance in performing productive activities.
This study demonstrated a strong association between trachomatous trichiasis and relative poverty, suggesting a bidirectional causative relationship possibly may exist between poverty and trachoma. Implementation of the full SAFE strategy in the context of general improvements might lead to a virtuous cycle of improving health and wealth. Trachoma is a good proxy of inequality within communities and it could be used to target and evaluate interventions for health and poverty alleviation.
| Trachoma has long been considered a disease of poverty. However, there is surprisingly little direct data that formally quantifies the relationship between trachoma and economic poverty, and none that specifically focuses on trichiasis. We compared 200 people with trachomatous trichiasis (TT) to 200 people (controls) without the condition, who were matched on age and sex, living in the same community, in Amhara Region, Ethiopia. We measured household relative poverty using three measures: household assets, self-rated wealth and peer-rated wealth. We also measured activity participation. We found TT case households were poorer by all relative economic measures. We found cases less likely to participate in economically productive activities regardless of visual impairment and other health problems, more likely to report difficulty and need assistance performing activities. The results suggest that the causative relationship between poverty and trachoma could possibly be bidirectional: poor households are more affected by trachoma and trichiasis reduces productivity even prior to development of visual impairment, which may exacerbate poverty. Implementation of the SAFE strategy in the context of general socioeconomic improvements might lead to a virtuous cycle of improving health and wealth. Trachoma could be used as proxy of inequality and to target and evaluate interventions for health and poverty alleviation.
| Trachoma is leading infectious cause of blindness worldwide [1]. Trachomatous trichiasis (TT) is the late stage consequence of repeated conjunctival Chlamydia trachomatis infection in which eyelashes turn towards the eye, causing pain and eventually irreversible blinding corneal opacification (CO). About 229 million people live in trachoma endemic areas, and approximately 7.3 million have untreated TT [2,3]. More than 2.4 million people are visually impaired from trachoma worldwide, among which between 439,000 and 1.2 million are estimated to be irreversibly blind [2,4]. The WHO recommends the SAFE Strategy for trachoma control [5].This involves Surgery for trichiasis, Antibiotics for infection, Facial cleanliness and Environmental improvements to suppress chlamydial infection and transmission.
Trachoma has long been considered a disease of poverty [6]. It is believed that the decline in trachoma observed in Europe, North America and elsewhere over the last century, in the absence of specific control measures, was largely attributable to general improvements in socio-economic status [7,8]. Trachoma remains prevalent in developing and marginalised communities, particularly in Africa, where crowded living conditions are common and access to clean water, sanitation and health care are often limited [6,8,9]. However, not all people living in such settings acquire active or scarring trachoma. It is possible that, within apparently homogeneous communities, the individuals who are most vulnerable to developing the blinding complications of trachoma are the poorest members of the poorest communities, although this has not been adequately investigated [10]. Moreover, the disability that TT causes may lead to reduced productivity, unemployment and loss of income, putting additional financial pressure on an already strained household [11–13]. The effect of trachoma on income may begin prior to the visual impairment, with the pain and the photophobia from trichiasis limiting function [13,14]. Of note, blindness has generally been associated with lower socio-economic status [15–17].
In low and middle income countries (LMICs) resources are often shared within households. Therefore, relative wealth or poverty in LMICs needs to be measured at household level, as the economic impact of a medical condition or intervention potentially affects the whole family [18]. In low-income settings estimating income can be difficult, as many people are self-employed and incomes are subject to significant short-term fluctuations [18,19]. In addition, people may earn from sources that they do not wish to disclose. Consumption expenditure data are considered more reliable than income data [16,19]. However, this method is subject to recall bias and requires detailed questionnaires, which are time consuming and costly to administer [19]. An alternative approach is to use a range of asset and housing characteristics as proxy indicators for household wealth and socio-economic status [19,20]. A key advantage of this approach is that it measures the long-term financial status of a household, and is less vulnerable to short-term fluctuations than income and consumption expenditure [19,20]. On the other hand, asset score only measure relative poverty, which may preclude regional or international comparability.
There is surprisingly little direct data that formally quantifies the relationship between trachoma and economic poverty, and none that specifically focuses on the scarring sequelae. The aim of this study was to investigate in detail the relationship between poverty and trachomatous trichiasis through an asset-based analysis, self-rated and peer-rated wealth measures, and participation in productive activities.
This study was reviewed and approved by the Food, Medicine and Healthcare Administration and Control Authority of Ethiopia, the National Health Research Ethics Review Committee of the Ethiopian Ministry of Science and Technology, Amhara Regional Health Bureau Research Ethics Review Board Committee, the London School of Hygiene and Tropical Medicine (LSHTM) Ethics Committee, and Emory University Institutional Review Board. Written informed consent in Amharic was obtained prior to enrolment from participants. If the participant was unable to read and write, the information sheet and consent form were read to them and their consent recorded by thumbprint.
This case-control study was nested within a clinical trial of two alternative surgical treatments for trichiasis. From the 1000 trichiasis cases recruited into the trial, every fifth consecutive case was also enrolled into this economic poverty study and matched to a non-trichiasis control. This approach was chosen for logistical and methodological reasons, in order to identify and collect data from controls within the shortest possible time period following case recruitment. Cases were defined as individuals with one or more eyelashes touching the eyeball or with evidence of epilation in either or both eyes in association with tarsal conjunctival scarring. People with trichiasis of other causes, recurrent trichiasis and those under 18 years were excluded. Trichiasis cases were identified mainly through community-based screening. Trichiasis screeners and counsellors (Eye Ambassadors) visited every household in their target village, identified and referred trichiasis cases to health facilities where surgical services were provided. Some individuals self-presented or were referred by local health workers. Recruitment was mainly from three districts of West Gojam Zone, Amhara Region, Ethiopia between February and May 2014. This area has one of the highest burdens of trachoma worldwide [21].
Controls were individuals without clinical evidence or a history of trichiasis (including surgery and epilation), and who came from households without a family member with trichiasis or a history of trichiasis, as we wanted to measure household level relative poverty, which requires comparison of trichiasis case households with households without trichiasis cases. One control was individually matched to each trichiasis case by location, sex and age (+/- two years). The research team visited the sub-village (30–50 households) of the trichiasis case requiring a matched control. A list of all potentially eligible people living in the sub-village of was compiled with the help of the sub-village administrator. One person was randomly selected from this list using a lottery method, given details of the study and invited to participate if eligible. If a selected individual refused or was ineligible, another was randomly selected from the list. When eligible controls were not identified within the sub-village of the case, recruitment was done in the nearest neighbouring sub-village, using the same procedures.
Data on detailed demographic characteristics were collected. Household economic poverty was measured through (a) Asset based wealth indicators, (b) Self-rated wealth, and (c) Peer-rated wealth. Activity participation data was collected using a modified ‘Stylised Activity List’ developed for the World Bank’s Living Standards Measurement Survey [22]. Visual acuity of both cases and controls were measured and cases underwent detailed trachoma examination.
To detect a difference in asset-based principal component analysis (PCA) similar to that found in the Cataract Impact Study (mean and standard deviation of asset based PCA score in cataract cases and their controls 0.6 and 2.0; and 0.3 and 2.6, respectively) with an alpha of 0.05 and 95% power, at least 346 (173 in each group) participants were required [16]. We recruited 200 trichiasis cases and 200 age, sex and location matched non-trichiasis controls.
Data were double-entered into Access (Microsoft) and transferred to Stata 11 (StataCorp) for analysis. Conditional logistic regression was used to compare basic characteristics of matched cases and controls.
Cases and controls were well matched in terms of location, gender and age and had similar levels of literacy, household size and household occupation (Table 1). Compared to the controls, the trichiasis cases were less likely to be married, more likely to be either unemployed or work as daily labourers, less likely to have a family member with formal education and more likely to have experienced a health problem during the last month. As expected, cases were more likely to be visually impaired than the controls (37.0% v 3.0%, respectively; OR = 69.0; 95%CI 9.58–496.82; p<0.0001)
The asset variables used in the PCA are described in Table 2 and their summary statistics are shown in S1 Table. The PCA was based on a combination of 28 asset values. The other 32 measured assets were excluded as they were present in less than 5% or more than 95% of the participants’ households. Households were generally poor. About 67% had a latrine, among which 65% were of the “non-improved” pit latrine type without a concrete slab. About half (54%) had their cattle dwelling within the main house. Ownership of durable assets such as mobile phones and radio was low (<30%). Only 17% of the households had access to electricity. About 12% of the households had taken a government loan. Overall, cases had fewer household and agricultural assets than controls and were more likely to have a government loan (Table 2). There was no difference in the ownership of the house they were living in (92.0% vs 94%, p = 0.22), or access to electricity (18·5% v 16·5%, p = 0·40). Case households had fewer rooms (1.22 vs 1.55, p<0·0001), and had a higher density of persons per room than the controls: 4.0, 95%CI 3.6–4.4 vs 3.3, 95%CI 3.0–3.6 respectively (P = 0.020).
The overall asset index accounts for 21% of the total variance (S1 Table). Among the three subset asset indices, the agricultural asset indicators had the highest factor scores and accounted for the highest weights in measuring wealth in this population. In contrast, the housing characteristics and utilities index, except for the number of metal roof sheets, had generally lower factor scores and contributed lower weights in estimating wealth than the other two subset indices. Among all indices, number of oxen and cows owned (0.324), the number of metal roof sheets (0.320) and amount of land owned in hectares (0.319) had the highest weights in estimating wealth. In contrast, access to electricity (-0.096) having cattle dwelling within the main house (-0.024) and having a government loan (-0.038) had negative weights. Fig 1 illustrates the distribution of the subset and overall asset indices, in order to determine whether clumping or truncation were present in this data. Overall, there was evidence of truncation and clumping when the three subset indices (Fig 1A to 1C) are used separately. However, the distribution of the overall combined factor scores was much smoother; and clumping and truncation were not observed (Fig 1D).
There was a strong association between being a trichiasis case and asset based household economic poverty: OR = 2.79; 95%CI, 2.06–3.78; p<0.0001 (Table 3). This relationship persisted after adjusting for marital status, and highest family education (OR = 2.78; 95%CI, 2.00–3.87; p<0.0001). For stratified analyses we combined “richest” and “rich” with “middle” because of small numbers, to create a “middle & above” category with three levels of socio-economic status measure to facilitate data modelling. Compared to the controls, trichiasis cases were more likely to be from the poorest (OR = 2.65; 95%CI, 2.05–3.42; p<0.0001) households than from the middle & above households (Table 4). In the stratified analysis, the association between asset based household economic poverty and trichiasis persisted regardless of age, gender, marital status, and in people with normal visual acuity after adjusting for the matching variables and family education (Table 4).
On both the self-rated and peer-rated scores, the households of trichiasis cases were rated poorer than controls (Table 3). This association persisted in both self-rated (OR = 3.99; 95%CI, 2.43–6.54; p<0.0001) and peer-rated (OR = 9.10; 95%CI, 4.79–17.27; p<0.0001) wealth measures after adjusting for marital status and highest family education. Compared to the controls, the trichiasis case households were more likely to be rated as poorest and poor rather than middle or affluent by themselves (OR = 3.74; 95%CI, 2.55–5.49; p<0.0001) and their peers (OR = 10.57; 95%CI, 6.42–17.41; p<0.0001) compared to the other households in their villages (Table 4). Using the 0 to 100 scale (poorest to richest), the mean self-rated scores for cases and controls were 34.1 v 49.1 (p<0.0001) and for peer-rated scores they were 27.5 v 50.3 (p<0.0001). The association of lower self-rated and peer-rated wealth with trichiasis persisted regardless of age, gender, marital status, and in people with normal visual acuity after adjusting for the matching variables and family education (Table 4).
The asset based socio-economic classification of households was found to be robust and produced similar ranking of households when the overall index was compared with the different subset indexes; the Spearman rank correlation coefficient ranged between 0.88 and 0.94. A Spearman rank correlation coefficient between asset index and self-rated wealth index, asset index and peer-rated wealth index, and self and peer-rated wealth indexes were 0.58, 0.70 and 0.63, respectively.
Trichiasis cases were significantly less likely to participate in household, outdoor, agricultural and leisure activities, even after controlling for the presence of other health problems during the preceding month, (Table 5). However, the trichiasis cases were slightly more likely to participate in daily labouring and self-employment activities such as selling goods. These associations persisted in multivariable analysis after controlling for self reported health problems during the preceding month, except for leisure activities. In stratified analyses by vision, trichiasis cases with normal vision were significantly less likely to participate in processing of agricultural products and in productive outdoor activities such as fetching wood and travelling compared to controls with normal vision (Table 5).
After adjusting for the matching variables and self reported health problems, trichiasis cases were significantly more likely to report difficulty in performing all productive and leisure activities than the controls: >66% of the cases reported difficulty in all productive activities in contrast to <5% of controls (Table 6). Similarly, trichiasis cases were significantly more likely to report receiving assistance in doing all productive activities compared to controls. In contrast to other activities, higher proportions of trichiasis cases received assistance particularly in agricultural activities such as farming, animal husbandry and processing agricultural products (Table 6).
In a univariable analysis (Table 7), being a household head with trichiasis had a strong association with economic poverty (OR = 3.29; 95%CI, 1.89–5.75; p<0.0001) while visual impairment had a borderline association (OR = 1.71; 95%CI, 0.98–2.97; p = 0.058). Not having a marriage partner (OR = 9.41; 95%CI, 4.16–21.31; p<0.0001), no family member with formal education (OR = 4.95; 95%CI, 1.73–14.16; p = 0.0028) and a main family job of daily labouring (OR = 19.64; 95%CI, 2.32–166.49; p = 0.0063) as opposed to farming were independently associated with economic poverty (Table 7). Families in which there were more people of a productive age were less likely to be poor than their counterparts (OR = 0.32; 95%CI, 0.16–0.60; p = 0.0005) (Table 7). In a multivariable analyses, participating in animal husbandry (OR = 0.05; 95%CI, 0.02–0.12; p<0.0001) and agricultural product processing (OR = 0.50; 95%CI, 0.27–0.91; p = 0·024) activities were independently associated with wealthier households while house cleaning (OR = 2.05; 95%CI, 1.03–4.08; p = 0.042) and self employment (OR = 2.77; 95%CI, 1.25–6.18; p = 0.012) activities were associated with poorer households.
Poverty is a complex multidimensional issue that encompasses not only deprivation of material possessions but also wider issues such as nutrition, health and education [32,33]. Many different approaches have been taken to measuring “poverty”, both in absolute and relative terms [34]. In general, these involve a survey methodology to capture estimates of income or consumption and methods that take into account broader issues of health and education such as the Multidimensional Poverty Index [34].
According to the 2011 World Bank estimates, 29.6% (Urban, 25.7%; Rural, 30.4%) of Ethiopians live below the national absolute poverty line (defined as 3781 Birr) and 30.7% live on less than US$1.25 PPP (purchasing power parity) a day [35]. Using asset indicators, the World Bank defines a household as being deprived “when none of these assets are owned by the household: fridge, phone, radio, TV, bicycle, jewelry, or vehicle” [35]. According to these criteria, 53% of rural households in Ethiopia were in deprivation in 2011. However, these are narrowly defined assets and most of these would not be commonly found in a rural Ethiopian community, irrespective to the level of wealth [35].
In this study we compared individuals with trichiasis to matched controls from within the same communities in Amhara Region, Ethiopia using three different measures of relative poverty: Asset Index, Self-Rated Wealth Index and Peer-Rated Wealth Index. These measures allow us to understand whether people with TT were relatively poorer than their neighbours, even within these very poor communities. We performed a PCA of household assets to stratify the participants into economic groupings. The variance explained by the first principle component was similar to the range reported in other similar studies (between 11% and 27%) [19,20,27,36]. The asset index used in this study is probably a reasonable proxy for consumption expenditure as we collected data on a sufficiently broad set of asset indicators that are capable of capturing living standards and wealth inequalities based on local values [37].
The age distribution, gender profile and literacy status of the trichiasis cases in this study were comparable with those reported in our earlier studies in Ethiopia as well as other studies of trichiasis patients elsewhere in Sub-Saharan Africa [31,38–40]. This suggests that the results are probably generalizable for this region of Ethiopia at least. The households of trichiasis cases were significantly less well off than controls in terms of ownership of almost all asset indicators measured. Consistent with the literature, trichiasis cases had significantly smaller and more crowded households [6,41]. Cases had less latrine access and more kept their cattle within the house, which is consistent with observations that active trachoma is associated with poor sanitation access [41–43]. These differences reflect a gap in the implementation of the “E” component of the SAFE strategy, which needs on-going emphasis in this region.
We have found clear evidence from each measure that even within trachoma-endemic communities individuals and households affected by trichiasis are significantly economically poorer than those that are not. Within endemic communities some individuals or families appear to be more severely affected by the disease and develop sight-threatening complications. This raises the important question of whether the association between poverty and trichiasis arises from a general state of impoverishment or whether there are a number of critical factors that primarily drive the relationship that might be amenable to focused intervention. The data we present here suggest that the relationship between poverty and trachoma could possibly be bidirectional.
Poverty may contribute to trachoma. This study provides evidence that even within superficially homogeneous endemic communities relative poverty plays a major part in the vulnerability of families to scarring disease. Firstly, trichiasis cases were more likely than the controls to come from households where the main family job is daily labouring and from families with no or lower formal education. Both of these factors have a major influence on income and health awareness, which in turn increase the vulnerability of the family to trachoma. Consistent with this, studies from Malawi, Tanzania and Ethiopia identified that children from lower socio-economic households had a higher prevalence of active trachoma than their counterparts indicating an association between poverty and active trachoma [10,44,45]. Secondly, previously described risk factor associations for active trachoma such as crowding and poor access to latrine, characterised the households of the trichiasis cases in this study. Such conditions are believed to promote the transmission of Chlamydia trachomatis within endemic communities, sustaining higher prevalence levels. Poorer households and communities may be less likely to have either the resources or the awareness to access treatment and sustain a sufficiently hygienic environment to control trachoma [8,17,46,47]. Households with higher income were more likely to have a latrine than their counterparts in a study conducted in the same area [48].
Trachoma may also contribute to poverty. Poor health frequently results in loss of productivity through disability and diversion of resources [11]. Trichiasis and its associated visual impairment probably lead to a loss of income, exacerbating pre-existing poverty in a “vicious cycle” [12,13]. Previously healthy and productive adults can be rendered dependent on others, unable to work or fully care for themselves due to pain, photophobia or visual impairment [13]. We found clear evidence of reduced activity and participation among trichiasis cases. Trichiasis cases were less likely than the controls to participate in productive household activities, outdoor activities (shopping/marketing, fetching wood and water) and agricultural activities (farming, animal husbandry and processing agricultural products). The stratified analysis found trichiasis cases with normal vision are less likely to participate in outdoor and agricultural activities than controls. This is consistent with a study of Tanzanian women with trichiasis without visual impairment, who had a degree of functional limitation which was comparable to those with visual impairment [14]. We found evidence that households with fewer economically productive adults and where the family head had trichiasis tended to be poorer. Conversely, households where trichiasis cases participated in agricultural activities were better off. Even where the trichiasis cases were undertaking specific activities, they reported much more difficulty and greater need for assistance than the controls. Similarly in another study, trichiasis cases reported difficulty in performing day-to-day farming activities [49]. These observations all point towards households with someone with trichiasis being under greater financial strains through reduced income contribution and greater needs and dependence of the person with trichiasis. The burden of disability caused by trachoma has been estimated between 171,000 and 1.3 million DALYs, with economic losses of 5–8 billion USD/year [4,12,13]. The economic loss from trichiasis alone due to lost productivity was estimated to be 3 billion USD/year [12,13].
This study comprehensively assesses the relationship between trachoma and economic poverty using four different measures, with a robust process to select suitable community controls. The asset index quantifies the long-term economic welfare of trachoma affected communities, which is important as trachoma and its sequelae are probably related to long-term SES [19,20]. The asset index has the practical advantage that it is much less affected by recall or measurement bias during data collection [19]. Most of the housing characteristics, utilities and durable assets were collected through direct observation minimising miss-measurement. Broad ranges of asset data were collected increasing the power of the study in the following ways. Clumping and truncation, potential problems that can arise with PCA of asset data and compromise its suitability for defining socio-economic strata, did not occur when all asset indices were combined into a single index. This indicates that the data from this study is sufficient to measure economic status and effectively infer inequality between different socio-economic strata and that in this region assessment of economic status by asset measurement requires a wider pool of parameters, particularly including agricultural assets. Encouragingly, the asset based poverty measure was moderately and strongly correlated with the self-rated and peer-rated wealth measures.
Poverty is a complex multidimensional problem with many causes and manifestations. Therefore there are many ways in which poverty can be measured. Here we only examined the economic aspect using relative measures such as low asset ownership. We use the first principal component (PC1) to measure socio-economic status. However, there is no clear description of the number of principal components to use and often the factor scores derived from the other principal components are difficult to interpret [27]. Despite the comparability of the amount of variance explained by PC1 with other studies, there is uncertainty whether the first component alone sufficiently explains all the pertinent variation. Asset scores are usually developed to be locally relevant, to allow ranking of people within the same community with respect to poverty. Unfortunately, socio-economic classifications based on asset ownership quintiles measure relative poverty within a given context and face the limitation of lacking international comparability. Therefore, between region or country comparison of SES should be done with caution [28]. We did not collect consumption or expenditure data, and so were not able to assess absolute poverty levels.
Although a community based screening method was used to identify trichiasis cases, it is possible that some cases might have been missed, which could potentially introduce non-response bias. Similarly, it is possible that some potential controls were not listed by the sub-village administrators. Self and peer-rated wealth are subjective measures, which might have suffered from the tendency to favour ranking households in the middle of the distribution. The activity participation data relied on the participant’s recall ability on what s/he had done in the last week. Finally, our results suggest that a bidirectional relationship may possibly exist between trachoma and poverty. However, the authors recognise that inference about causality is speculative as it is not possible to draw firm conclusions from a cross-sectional observational study such as this.
In this study we found a clear association between trichiasis and household economic poverty by all three economic measures. Trichiasis cases were more likely to have economically poor households and less likely to participate in productive activities regardless of visual impairment, more likely to report difficulty in performing productive activities and more likely to need assistance in performing activities than controls. These suggest that the causative relationships between poverty and trachoma may possibly involve bidirectional interaction: poor households are more affected by trachoma and the scarring sequelae of trachoma and trichiasis reduces productivity even prior to the development of visual impairment, which might lead to additional poverty.
These data are anticipated to be useful in advocacy and to support programme leaders and funders to secure resources to promote trachoma prevention linked to socio-economic development in trachoma-endemic communities. Implementation of the full SAFE strategy in the context of general improvements might lead to a virtuous cycle of improving health and wealth. Trachoma is a good proxy of inequality within communities and it could be used to target and evaluate interventions for health and poverty alleviation. Measuring the effect of trichiasis surgery on household economic poverty through longitudinal studies would provide an indication of the relative contribution of trichiasis to poverty, as improved health potentially leads to improved productivity and income.
|
10.1371/journal.pcbi.1000574 | Dynamic Allostery in the Methionine Repressor Revealed by Force Distribution Analysis | Many fundamental cellular processes such as gene expression are tightly regulated by protein allostery. Allosteric signal propagation from the regulatory to the active site requires long-range communication, the molecular mechanism of which remains a matter of debate. A classical example for long-range allostery is the activation of the methionine repressor MetJ, a transcription factor. Binding of its co-repressor SAM increases its affinity for DNA several-fold, but has no visible conformational effect on its DNA binding interface. Our molecular dynamics simulations indicate correlated domain motions within MetJ, and quenching of these dynamics upon SAM binding entropically favors DNA binding. From monitoring conformational fluctuations alone, it is not obvious how the presence of SAM is communicated through the largely rigid core of MetJ and how SAM thereby is able to regulate MetJ dynamics. We here directly monitored the propagation of internal forces through the MetJ structure, instead of relying on conformational changes as conventionally done. Our force distribution analysis successfully revealed the molecular network for strain propagation, which connects collective domain motions through the protein core. Parts of the network are directly affected by SAM binding, giving rise to the observed quenching of fluctuations. Our results are in good agreement with experimental data. The force distribution analysis suggests itself as a valuable tool to gain insight into the molecular function of a whole class of allosteric proteins.
| Proteins carry out most of the cellular processes, from metabolic reactions to the regulation and expression of genes. Tight and effective regulation of the executing protein machinery is commonly achieved by allostery. The only general requirement for allosteric communication is the transmission of a signal, e.g., the binding of a cofactor, from the ligand binding site to the allosteric (active) protein site; in other words an internal propagation of strain. Based on molecular dynamics simulations, we recently presented a method that allows visualization of force distribution in proteins. We here applied this method to MetJ, a transcription factor whose activity is regulated by a co-repressor. Interestingly, co-repressor binding does not cause visible structural changes, yet increases DNA binding affinity manyfold. We were able to reveal a network linking fluctuations of distal parts of MetJ, including the DNA binding interface. Mechanical strain caused by SAM binding propagates to certain key residues, thereby altering fluctuations and finally resulting in increased DNA binding affinity. By directly monitoring ligand induced strain, instead of conformational changes, which might be absent or slow, our force distribution analysis suggests itself suitable to detect the mechanically crucial motifs in allosterically regulated protein machineries.
| Protein allostery plays a key role in the regulation of cellular functions such as transcription or enzymatic action [1]. It crucially governs the formation of protein or protein-DNA complexes as well as the functional activity of individual proteins. Allosteric signals used by nature are diverse, ranging from ligand binding to reversible covalent modifications such as phosphorylation, or changes in the environment like pH or temperature. Intriguing examples are allosteric proteins in which effector molecules bind distal to the active site [2],[3].
A fundamental question is how the allosteric perturbation is transmitted through the protein to the active site for functional regulation. Can we understand and predict the mechanism and the network of interactions that propagate an allosteric signal? Answering this question is a prerequisite for functional mutagenesis and rational design of allostery. Sequence-based statistical analysis has proven highly successful to detect signal propagation pathways within and between allosteric proteins on the basis of evolutionary constraints [4],[5]. On the theoretical side, various thermodynamic concepts for inter-domain communication in allosteric proteins have been established [6],[7]. As yet, the molecular basis for long-range allosteric coupling between the regulatory and active site of a protein remains a matter of debate. This is why a range of experimental and computational techniques to monitor conformational changes involved in allostery have been developed and applied [7],[8], among others NMR [9], molecular dynamics (MD) simulations [10],[11], normal mode analysis and elastic network models [12],[13].
The basic premise of the above approaches is a conformational transition between two distinct states or a shift in a pre-existing conformational ensemble upon allosteric perturbation. In a commonly accepted picture, allosteric signals cause a perturbation at the regulatory site of the protein, analogous to an externally applied force. The perturbation then dissipates as internal strain or energetic coupling through the protein to the active site [14]. Signal propagation in turn causes conformational rearrangements, inducing an enhancement or decrease in the protein's activity. However, examples of long-range allosteric communication in the absence of any obvious conformational changes [9],[15] question this picture, showing that allostery does not necessarily rely on a change in mean atomic coordinates. Instead, allosteric strain can dissipate through rigid scaffolds without detectable conformational rearrangements.
A more fundamental understanding of allostery would thus require a way to directly follow strain propagation through proteins. This could reveal the allosteric network in a protein even in the absence of - or prior to the occurrence of - conformational changes. We recently presented a method termed force distribution analysis (FDA), based on MD simulations, that allows to detect propagation of internal strain caused by an external signal through proteins. The high sensitivity of the method makes it possible to even detect propagation through stiff materials, where a signal will propagate causing only minimal conformational changes that are below the threshold of experimentally accessible resolution. We have previously demonstrated the feasibility of FDA to detect force propagation in two mechanically perturbed proteins, namely the highly robust titin immunoglobulin domain I27 [16], and silk crystalline units [17]. While classical approaches focus on conformational changes or ensemble redistributions as a consequence of the signal-induced strain, such as normal mode analysis or essential dynamics [18], FDA sets out from the strain distribution itself. This renders FDA a perfectly fitted tool to elucidate the mechanism underlying allosteric signaling in proteins in general, be it with or without the involvement of structural rearrangements.
We here chose to test the feasibility of FDA to detect allosteric networks in proteins using the classical textbook example of the methionine repressor protein MetJ [19]. MetJ is a challenging candidate, as it features long-range allosteric communication, yet without any noticeable changes in protein structure upon effector binding. MetJ is a transcription factor in the met regulon of Escherichia coli, the gene regulatory control system for methionine biosynthesis [20]. MetJ regulates the transcriptional levels of its own gene and those of several other proteins. Repressor activity results from binding to its operator, a specific 8 bp DNA sequence (the “metbox”), located in the promoter regions of genes regulated by MetJ. Changes in sequence of the metboxes are supposed to explain different regulatory activity [20],[21]. MetJ forms a homodimer in its native state [22]. In case of multiple adjacent metboxes it may form complexes of several homodimers arranged in a wheel-like structure around the DNA [23]. DNA binding of MetJ is regulated by its co-repressor, S-adenosylmethionine (SAM), an end product of methionine biosynthesis, Fig. 1A. Sensitivity for DNA is increased several-fold [24],[25] upon co-repressor binding. Of special interest is that SAM binds distant from the DNA binding site, with a minimal SAM-DNA distance of in crystal structures [26]. Holo and apo structures do not show significant structural changes [15]. For this reason it remains controversial how SAM influences DNA binding.
S-adenosylhomocysteine (SAH), a SAM analogue, binds MetJ with a binding affinity similar to SAM, but has no effect on its affinity for the operator (S. Philipps, Leeds Univ, 2009, personal communication). The main difference between SAM and SAH (Fig. 1B) is a positive charge on the sulfur atom of SAM, and it has been suggested that the increased sensitivity upon co-repressor binding is of purely electrostatic nature [27]. On the other hand, introduction of positive charges by a series of point mutations could not substitute the need for co-repressor [28]. Based on the force distribution pattern observed within the MetJ homodimer, we here propose a new model for MetJ activation upon cofactor binding. We measure directed propagation of internal strain from the SAM binding site to distinct residues in the DNA binding interface, through a specific network of a few key residues. The consequence is a wide-spread quenching of slow fluctuations and relocation and stiffening of specific side chains at the MetJ-DNA interface, leading to increased protein - DNA interaction. A distinct interaction pattern of individual residues with the co-repressor allows MetJ to fine-tune its response to co-repressor binding, explaining the inability of SAH to act as a co-repressor. Our results yield a molecular basis for MetJ allosteric function and are consistent with previous experimental studies.
We carried out extensive MD simulations to elucidate the force distribution and conformational properties of MetJ. We used crystal structures of MetJ (PDB code 1CMC [15]) and MetJ in complex with DNA (PDB code 1CMA [26]) as starting point for our simulations. Throughout the manuscript, we will use the terms MetJ for the system without DNA and MetJ-dna for the MetJ-DNA complex. In both cases, simulations of the holo and apo forms were performed for comparison. Apo forms were created by deleting the bound SAM molecules from the crystal structures. An apo structure of MetJ is available, but as force distribution analysis is very sensitive to structural changes we decided to use the same crystal structure as basis for our simulations. Structures for Q44K, a mutant not relying on cooperativity to be functional [29], exist as well. Yet, as the altered charge distribution alters the DNA recognition pattern, though not the allosteric effect itself, we decided not to further investigate Q44K. For each of the five systems, 10 independent 30 ns MD simulations were performed, totaling 300 ns of simulation time, respectively. In agreement with crystallographic data [15], our simulations do not show major deviations between holo and apo forms. The overall backbone root mean square deviation (RMSD) of average structures is 0.66 Å for MetJ-dna and 0.64 Å for MetJ. This compares well with the crystal structures where we find a backbone RMSD for holo and apo structures of 1.63 Å which lowers to 0.59 Å after excluding poorly resolved loop regions having different conformation (residues 12–20 and 77–84), Fig. 1C. Crystal waters in the protein-DNA interface of 1CMA were found to quickly move into the bulk solvent and are thus unlikely to bridge specific interactions.
To elucidate distribution of the allosteric signal induced by co-repressor binding, we directly calculated forces between each pair of atoms and from our MD trajectories. We here analyze scalar pair-wise forces, which in contrast to the vectorial representation are unaffected by rotation of the system during the simulations. Observing pairwise forces has the advantage that forces do not average to zero over time, thus being the measure of choice for internal strain in systems equilibrated under a perturbation. Forces are calculated individually for bonded and non-bonded (electrostatic and van der Waals) interactions below the cutoff distance using the interaction potential defined by the Amber03 [30] force field. Long-range interactions as well as solvation effects such as screening of electrostatic forces and hydrophobic forces are not directly included in , which is calculated only for the solutes and within the non-bonded cut-off. We however indirectly accounted for these effects by calculating forces from a system simulated in explicit solvent and with full electrostatics. Details are given in Methods. Propagation of the mechanical perturbation caused by SAM binding is measured as the difference in pairwise force, , between the apo and holo forms of MetJ/MetJ-dna. For convergence, forces for each system were averaged over all ten equilibrium trajectories, each 30 ns in length. To reduce noise further, mainly resulting from slow side chain fluctuations that cannot equilibrate during simulation time, data were normalized as described in Methods. Dimensionless normalized changes in force are denoted .
The MetJ homodimer has a high degree of symmetry, and we thus expect the force distribution pattern to be highly symmetric as well. We checked this by calculating correlation coefficients between residue wise forces , see Methods. Indeed, we find the force propagation pattern for the monomers to be very similar in all systems. For MetJ, residue wise forces correlate with , Fig. S2A. The MetJ-dna structure shows a less symmetric pattern, with , Fig. S2B. The lower symmetry of MetJ-dna might be a result of the lower resolution of the 1CMA crystal structure (2.8 Å for 1CMA vs. 1.8 Å for 1CMC) or of the only partially resolved DNA.
Force distribution at the DNA binding site (Fig. 2A) reveals that remote MetJ binding induces a high degree of strain at distinct regions of the MetJ-DNA interface. In particular, Arg40 and a loop formed by residues 50–53 are subjected to high strain. The presence of the co-repressor thus is sensed by the DNA binding site, apparently via a long-range propagation of force from the bound SAM molecule through the protein scaffold to the MetJ-DNA interface. Importantly, the force distribution pattern was equally observed in the absence of DNA, Fig. 2B. In fact, forces in MetJ and MetJ-dna distribute in a very similar way, yielding a correlation of , Fig. 2C. First, this is strong evidence that the observed change in forces is a result of SAM binding, independent from the presence of DNA. Second, as the initial crystal structures differ in resolution and conformation, the significant correlation highlights that the distribution pattern is robust with regard to the starting structure.
On the basis of FDA, we next investigated which protein structural elements are key to the strain distribution, allowing communication between SAM and the protein-DNA interface over a distance of more than 1 nm. Within the protein scaffold, we observe force propagation through helix B (B′) and forces are transmitted via side chain interactions onto helix A (A′), which in turn forms various side chain contacts with the DNA, Fig. 3A+B. Force propagation is highly non-isotropic and directed. This is to say, when compared to helix A and B, we see relatively little changes in pair-wise forces for the and the loops formed by residues 12–20, both in direct contact with the DNA, as well as for helix C (C′), Fig. 3C. In agreement with the low allosteric strain in the , this motif, even though binding to the major groove of the metbox, has been found to play a role in DNA sequence specificity, but not in the allosteric regulation of DNA binding affinity [31]. Only a few side chains of helix A show significant changes in pair wise force, the strongest of which is observed for Glu39, Arg40, Arg42 and Arg43. Out of these residues only Arg40 is in direct contact with the DNA. This observation is remarkable as an almost complete loss of binding affinity was reported for mutation of Arg40 and its spacial neighbor Thr37, but not for mutation of others in direct contact with DNA [31]. Thr37, however, has been suggested to be involved in enhancing cooperativity, thereby only indirectly regulating DNA affinity. In agreement, we do not find Thr37 to be under SAM-induced strain.
We find two inter-related mechanisms of force propagation responsible for the specific targeting of the above mentioned structural elements. First, SAM strongly exerts a direct strain onto a set of MetJ residues, as reflected by extra-ordinarily high forces between the co-repressor and these residues, , Fig. 3D. Most importantly, the adenosyl group of SAM strongly interacts with Glu39 and Arg42 in helix A, influencing their dynamics (see below and Fig. 3D).
Second, SAM features repulsive forces with helix B, inducing a high strain in the helix backbone hydrogen bonds. This apparently involves slight helix bending, Fig. 4A. Indeed, measuring the angle defined by the atoms of residues Ala64, Cys58 and Asn53 shows a bending upon SAM binding of for MetJ and for MetJ-dna. We note that it is the significant difference in hydrogen bond forces, not in the mere atomic coordinates, between apo and holo form, that serves as robust indication for SAM-induced signal propagation. Helix bending in turn imposes strain on the salt bridge between Glu59 in helix B and Arg43 in helix A by minor conformational rearrangements, Fig. 4A+B. We measure high change in force between these residues, suggesting this electrostatic interaction, buried in the protein core, to propagate force between helix B and helix A.
Both mechanisms, direct forces imposed from SAM onto key residues in helix A, and propagation of forces from SAM via bending of helix B, result in inconspicuous rearrangements at the DNA binding interface; most notably in the loop linking helices A and B (residues 50–53) and Arg40, as described above, Fig. 2A+B. Repositioning of Arg40 upon SAM binding is accompanied by an adjustment of the side chain packing with its direct neighbors, Thr37 and Asn53, Fig. 4C. Again, pairwise forces here served as a measure for signal propagation, rather than the only minor, yet reproducible coordinate changes (as for example a change of the angle in Asn53 between , and found for both, MetJ-dna as well as MetJ).
The described rearrangement of Arg40 caused by propagation of strain entails a strengthening of its saltbridge with DNA. From FDA, we measured an increase in attraction between Arg40 and DNA of . Overall, the potential energy between MetJ and DNA decreases by from in the holo to in the apo form, as a result of allosteric signaling by the co-repressor.
The loops formed by residues 12–20 (referred to as loop 1) suggest themselves to be involved in the allosteric mechanism, as they strongly differ in conformation between the 1CMC (MetJ) and 1CMA (MetJ-dna) crystal structures and are in direct contact with the DNA, Fig. 1C. NMR data for these loops shows a strong quenching of ns time-scale fluctuations upon co-repressor binding (Steve Homans, Leeds University, 2009, personal communication). In good agreement with these experimental findings our simulations of the MetJ-dna system show a strong decrease in backbone RMSF for loop 1 residues upon SAM binding, as well as stiffening of helix C, Fig. 5A–C. Quenching is observed for both the MetJ and MetJ-dna system, though less pronounced for the former (see below). Remarkably, principal component analysis (PCA) on the trajectory data reveals the dynamics of the distal loop 1 and helix C regions to be highly coupled, Fig. 5D, and the dynamics of both MetJ monomers to be highly cooperative. The lowest frequency mode (Eigenvectors 1–3) for apo and holo structures of MetJ-dna describe highly similar fluctuations, yet at very different amplitudes. Strong quenching of fluctuations is reflected by a decrease of the highest Eigenvalue from 120 (apo) to 28 (holo), Fig. S1. These observations are supported by entropy calculations based on Schlitter's formula [32]. Upon SAM binding, we find a decrease in entropy of for MetJ-dna and for MetJ, see also Table 1. The quantitatively different, yet qualitatively equivalent, results might be caused by the different crystal structures used, i.e the differences for loop 1 and adjacent residues. Overall, we find the stiffening effect of SAM to be independent from the presence of DNA.
The question arises how the distal helix C and loop 1 regions are dynamically linked through a largely rigid core of the MetJ-DNA system. To elucidate the communication pathway, we performed PCA on residue averaged pair-wise forces, , here termed force-PCA. Again, observing forces directly has the unique advantage to allow for following the complete propagation pathway, including parts showing only subtle coordinate changes. Force-PCA on MetJ-dna revealed a network of correlated changes in pair-wise forces, Fig. 5E. The network spans through the protein core, linking helix C and loop 1, the latter of which is connected to the rest of the network via residues Tyr11, Ile28, Lys31 and Glu55. Synchronization of the fluctuations between both monomers is achieved by force propagation along helix A and the . We found the allosteric signal caused by SAM binding to target large parts of helix A, in particular Glu39, Arg40 and Arg42, resulting in wide-spread stiffening, Fig 4C. Helix A accounts for a large part of the network propagating fluctuations, moreover it directly is part of the link between helix C and loop 1, Fig. 5E. In summary, SAM binding alters correlated forces linking loop 1 and helix C thus affecting the dynamics of these regions.
The SAM analogue SAH has no regulatory function, i.e. no impact on the MetJ activity for binding to DNA, yet has the same binding mode and similar binding affinities as SAM (S. Philipps, Leeds University, 2009, personal communication). Based on this observation, an entirely electrostatic activation of MetJ by the positively charged SAM has been suggested [27]. We decided to elucidate differences between SAM and SAH binding, and to this end performed simulations of MetJ-dna in complex with SAH as co-repressor. We modeled the MetJ-SAH structure by removing the group from the sulfur atom of SAM in the 1CMA crystal structure used as template.
The overall conformation of MetJ-dna is not affected when replacing SAM by SAH, both structures have a backbone RMSD of only 0.42 Å. Also, the potential energy between protein and DNA is quasi identical to the energy measured for MetJ-SAM and DNA . As for the co-repressor, our simulations show strong quenching of fluctuations upon SAH binding, yet quenching is less distinct. This is reflected by higher backbone-RMSF for MetJ-SAH throughout the protein, Fig, 5B, as well as a higher eigenvalue of 48 for the first eigenvector, what is significantly above 28, the value measured for SAM. Both eigenvectors describe a very similar mode of fluctuation, Fig. S1C. The flexibility of the bound ligand itself is increased as well. We measured an almost twofold increase in RMSF for SAH when compared to SAM (0.89 Å vs. 0.57 Å), apparently due to the loss of backbone interactions with SAM's positive charge.
Indeed, and unsurprisingly, the changes in direct interactions between the co-repressor and individual residues are significant, Fig. 6A. Removing the positive charge alters the charge distribution of SAM's whole methionine group, and we see changes in interaction even for residues as far as in helix A (residues 39 to 43), though most of the observed changes affect residues in direct proximity to the sulfur atom (residues 59 to 67). These changes lead to wide-spread alterations in the overall force propagation pattern, which are most pronounced in helix C and the proceeding loop, Fig. 6C. Interestingly, we find high changes in forces for Tyr11 and Ile28, both of which were found to link fluctuations of loop 1 with helix C by force-PCA. However, this effect is only present in the domain with the full DNA fragment resolved (residues 106–209 in the 1CMA structure), and thus further validation is necessary.
As the differences in binding affinity between SAM and SAH are of primarily entropic nature, we performed entropy calculations on MetJ-dna based on Schlitter's formula [32]. Vibrational entropies were calculated on the whole trajectory data totaling 300 ns per system and are sufficiently converged to allow semi-quantitative comparisons between SAM and SAH, Fig. S3. We found an entropy difference of between SAM and SAH as co-repressor, of which the protein dynamics with accounts for the major contribution. All values are given in Table 1. The absolute conformational entropies of (apo) and (holo) per residue are in agreement with previous estimates for other proteins [33],[34]. The values clearly show that there is a significant increase in entropy when substituting SAM by SAH, consistent with the observed difference in regulatory function. Both, the overall RMSF and the entropies suggest SAM to reduce MetJ flexibility more efficiently than SAH.
We have analyzed force distribution and dynamics in MetJ, a stiff allosteric protein regulated by SAM, its co-repressor. FDA allowed us to identify the network of interactions guiding force modulation within MetJ by cofactor binding. Experimental data, among others the inactivity of SAH as a co-repressor, suggest that a long range electrostatic interaction between DNA and the positive charge on SAM may exclusively explain MetJ activation [35]. Notwithstanding, there is evidence from mutagenesis experiments that charge alone cannot explain MetJ activation [28]. We here suggest strain propagation by subtle alterations of the MetJ structure as an important mode of allosteric signal propagation. The highly anisotropic distribution of internal strain leads to conformational re-adjustments at the interaction interface, mainly of Glu39, Arg40, Arg42, Arg43 and residues 50–53. Our simulations thus predict adjustments of these specific protein-DNA interactions to be an important factor for efficient DNA binding. Such a mechanism would allow MetJ to easily move along or between DNA strands until the target side is found, thereby speeding up target site location as recently proposed [36].
While the importance of this communication pathway has been experimentally probed by the loss of allosteric function upon mutation of residues identified as key residues by FDA, it is independent of the positive charge on SAM, as we find it similarly for SAH. This pathway therefore apparently causes or is complemented by an additional allosteric mechanism unique to SAM. We find the major SAM-dependent allosteric function of MetJ to come from an entropic contribution due to quenching of slow backbone and fast side chain dynamics. Only for SAM, the force network communicating the allosteric signal between loop 1 and helix C can substantially reduce correlated fluctuations. This is supported by theoretical models [37] as well as NMR data that suggest dynamics to play an important role (Steve Homans, Leeds University, 2009, personal communication). The major correlated motion that is quenched involves parts distant to each other as well as to the co-repressor binding site. Again, measuring correlated forces instead of coordinates revealed the role of the protein core in this long-range communication and allosteric regulation. We find a strong increase in entropy when substituting SAM by SAH, suggesting that the regulatory difference between SAM and SAH is of entirely entropic nature. It is the differential effect of SAM and SAH on the correlated forces involved in this motion that is likely to be responsible for the observed difference in allostery.
Dynamics are increasingly revealed as a regulatory driving force [38]–[40] and have recently been found for another transcription factor, the CAP protein [9]. We here find a similar mechanism for MetJ, suggesting that changes in dynamics upon cofactor binding may be a commonly used regulation pattern. Long-range allostery in the absence of any noticeable conformational change as featured by MetJ has remained a challenge for structure-based experimental and theoretical approaches. In combination with conventional analysis of the MetJ dynamics, we find FDA an optimal tool to track an allosteric pathway in MetJ. Signal propagation was found to be largely hidden in unremarkable shifts in atomic coordinates. Yet, these mere conformational shifts, as revealed by FDA, can involve large changes in forces for strongly interacting atom pairs, resembling “stiff springs” in the protein interaction network. Monitoring forces instead of coordinates therefore renders FDA highly sensitive. Pure conformational analysis would simply overlook rearrangements of the magnitude reported here, especially as properties such as root-mean square deviations or fluctuations are easily dominated by slow sub-domain movements, as it is the case for MetJ, Fig. 5C. By considering pair-wise forces which are, by definition, dominated by strong and relatively short-ranged interactions, such large fluctuations have only minor influence. Pair-wise interactions have the additional advantage of being independent from any fitting scheme, as conventionally used for RMSD or RMSF calculations, thereby not introducing any bias by the arbitrary choice of a reference structure. The same multivariate statistical methods, such as PCA, that are used for the analysis of coordinate based trajectory data can be applied to pair-wise forces. Again, one has the advantage of being able to observe relations that would otherwise be below the sensitivity of the method.
We recently determined the force bearing scaffold in a titin immunoglobulin domain, a protein mainly designed to withstand mechanical load by means of FDA [16]. Here, we present the first successful application to a stiff allosteric protein, opening the road to better understand the function of a whole class of proteins, including enzymes, by examining their internal force network. We note that FDA does not require extensive sampling of an allosteric conformational transition, which at current simulation time-scales is out of reach for most proteins. This is an unique advantage over other MD based simulation techniques used for studying protein allostery. FDA is content with monitoring the development of internal strain prior to the eventual shift in the protein conformational ensemble. We predict forces averaged over a total simulation time in the sub-microsecond range to suffice for the analysis of much slower allosteric signaling pathways. Importantly, while we here modified the Gromacs simulation suite to add FDA functionality, virtually any MD simulation package can be easily modified to include FDA at practically no additional computational expense, as pair-wise forces are anyways calculated at each time step.
Our results highlight the strength of FDA as a tool supporting experimental design, as it can straightforwardly be verified by experimental studies. In particular, our results suggest Arg40, Thr37 and Asn53 at the MetJ-DNA interaction interface to be important for allosteric function. Mutations of Arg40 and Thr37 have indeed been previously shown to abolish SAM-dependent allosteric regulation of MetJ [31]. In addition, we predict mutation of Glu59 and Arg43, forming the salt bridge between helix A and B, and the crucial SAM interaction partners Glu39 and Arg42 to lower the co-repressor activity of SAM.
|
10.1371/journal.pcbi.1000318 | Accurate Detection of Recombinant Breakpoints in Whole-Genome Alignments | We propose a novel method for detecting sites of molecular recombination in multiple alignments. Our approach is a compromise between previous extremes of computationally prohibitive but mathematically rigorous methods and imprecise heuristic methods. Using a combined algorithm for estimating tree structure and hidden Markov model parameters, our program detects changes in phylogenetic tree topology over a multiple sequence alignment. We evaluate our method on benchmark datasets from previous studies on two recombinant pathogens, Neisseria and HIV-1, as well as simulated data. We show that we are not only able to detect recombinant regions of vastly different sizes but also the location of breakpoints with great accuracy. We show that our method does well inferring recombination breakpoints while at the same time maintaining practicality for larger datasets. In all cases, we confirm the breakpoint predictions of previous studies, and in many cases we offer novel predictions.
| In viral and bacterial pathogens, recombination has the ability to combine fitness-enhancing mutations. Accurate characterization of recombinant breakpoints in newly sequenced strains can provide information about the role of this process in evolution, for example, in immune evasion. Of particular interest are situations of an admixture of pathogen subspecies, recombination between whose genomes may change the apparent phylogenetic tree topology in different regions of a multiple-genome alignment. We describe an algorithm that can pinpoint recombination breakpoints to greater accuracy than previous methods, allowing detection of both short recombinant regions and long-range multiple crossovers. The algorithm is appropriate for the analysis of fast-evolving pathogen sequences where repeated substitutions may be observed at a single site in a multiple alignment (violating the “infinite sites” assumption inherent to some other breakpoint-detection algorithms). Simulations demonstrate the practicality of our implementation for alignments of longer sequences and more taxa than previous methods.
| Recombination is the process by which a child inherits a mosaic of genes or sequences from multiple parents. Though most species participate in some form of genetic mixing or recombination, the mechanics by which this occurs varies greatly among them. In higher order organisms, crossing over occurs in meiosis along the parent-child relationship, whereas in bacteria, viruses, and protozoans, homologous exchange of DNA material can occur from one individual to another without the need for sexual reproduction [1]. The diversity with which recombination occurs motivates the need for different models and methods, each ideally suited to its biological situaion. We have developed a probabilistic approach to recombination detection that we believe to be superior for analyzing situations of admixture of pathogen subspecies with a high mutation/recombination ratio.
The situation we concern ourselves with has been termed phylogenetic recombination inference (PRI) by [2], and works by inferring phylogenetic tree topology changes over a multiple alignment. Though it has been shown that under a neutral coalescent model, the number of recombination events which will lead to a tree topology change is very small, [3] in situations of admixture following geographical separation a greater proportion of topology-changing recombinations are expected. Abandoning the infinite-sites model of sequence evolution and instead using a continuous-time Markov chain makes direct inference intractable, and so we instead employ a phylo-HMM which models an effect of recombination, rather than modeling the process explicitly.
While recombination detection is an interesting mathematical challenge, fast, flexible, and reliable computational methods are also motivated by a multitude of biological reasons. We see our method not as being able to answer all of these biological questions on recombination, but rather a potentially valuable tool for furthering recombination research.
We give here an outline of previous methods which are related to our phylo-HMM approach. For a more thorough survey or recombination detection methods, see [1].
The rationale for phylogenetic recombination inference is motivated by the structure of the Ancestral Recombination Graph (ARG), which contains all phylogenetic and recombination histories. The underlying idea is that recombination events in the history of the ARG will, in certain cases, lead to discordant phylogenetic histories for present-day species.
Various approaches to learn the ARG directly from sequence data have been developed, such as [8] and [9]. We recognize that PRI is in a somewhat different category both in goal and approach as compared these methods, though they are motivated by the same underlying biological phenomenon. Rather than aiming to reconstruct the ARG in its entirety, our emphasis is on modeling fast-evolving organisms with the goal of accurately detecting breakpoints for biological and epidemiological study.
The most widely-used program for phylogenetic recombination detection is SimPlot [10] (on MS Windows). Recombination Identification Program (RIP), a similar program, [11] runs on UNIX machines as well as from a server on the LANL HIV Database site. This program slides a window along the alignment and plots the similarity of a query sequence to a panel of reference sequences. The window and step size are adjustable to accommodate varying levels of sensitivity. Bootscanning slides a window and performs many replicates of bootstrapped phylogenetic trees in each window, and plots the percentage of trees which show clustering between the query sequence and the given reference sequence. Bootscanning produces similar output to our program, namely a predicted partition of the alignment as well as trees for each region, but the method is entirely different.
In [12], Husmeier and Wright use a model that is similar to ours except for the training scheme. Since they have no scalable tree-optimizing heuristic, their input alignment is limited to 4 taxa so as to cover all unrooted tree topologies with only 3 HMM states, making their method intractable for larger datasets. They show they are able to convincingly detect small recombinant regions in Neisseria as well as simulated datasets limited to 4 taxa [12].
The recombination detection problem can be thought of as two inter-related problems: how to accurately partition the alignment and how to construct trees on each region. This property is due to the dual nature of the ARG: it simultaneously encodes the marginal tree topologies as well as where they occur in the alignment. Notice that if the solution to one sub-problem is known, the other becomes easy. If an alignment is already partitioned, simply run a tree-inference program on the separate regions and this will give the marginal trees of the sample. If the trees are known, simply construct an HMM with one tree in each state and run the forward/backward algorithm to infer breakpoints. Previous methods have used this property by assuming one of these problems to be solved and focusing on the other. For example, in Husmeier and Wright's model, there were very few trees to be tested, and so the main difficulty was partitioning the alignment, which they did with a HMM similar to ours. In SimPlot, windows (which are essentially small partitions) passed along the alignment and trees/similarity plots are constructed. This allows the program to focus on tree-construction (usually done with bootstrapped neighbor-joining) rather than searching for the optimal alignment partition.
By employing a robust probabilistic model with a novel training scheme, we find a middle ground between the heuristic approach of SimPlot [10] and the computational intractability of Husmeier and Wright's method [12], where we are essentially able to solve the recombination inference problem a whole, rather than neglecting one sub-part and focusing on the other. We use a HMM to model tree topology changes over the columns of a multiple alignment. This is done much in the same way as Husmeier and Wright, but our use of a more sophisticated tree-optimization (the structural EM heuristic) method allows searching for recombination from a larger pool of sequences. By modifying the usual EM method for estimating HMM parameters in a suitable way, we are able to simultaneously learn the optimal partitioning of the alignment as well as trees in each of these partitions. We are able to detect short recombinant regions better than previous methods for several reasons. First, we do not use any sliding windows which may be too coarse-grained to detect such small regions of differing topology. Second, our method allows each tree after EM convergence to be evaluated at every column, and so small recombinant regions are not limited by their size; they must only ‘match’ the topology to be detected or contribute to the tree training. By embedding trees in hidden states of an HMM, the transition matrix allows us to essentially put a prior on the number of breakpoints, as opposed to considering each column independently. Furthermore, since the counts in the E-Step are computed using all columns of the alignment, distant regions of the alignment with similar topology may contribute their signal to a single tree, whereas in a window-sliding approach each window is analyzed independently.
Since it is difficult to experimentally verify predictions of recombination, we test our methods on previously-analyzed data from earlier studies on Neisseria and HIV-1 as well as simulated datasets. Individual cases highlighting various facets of our method can be seen in Text S1, whereas statistics summarizing simulations with respect to several simulation parameters are included in Figure 1.
When comparing real and simulated data, one must keep in mind that real data may have complications such as rate heterogeneity and structural features that are not present in simulations, which are carried out using a simple independent-sites Markov chain model of nucleotide evolution, such as the HKY85 model [13]. While this is currently the only model we use in our program, it is straightforward to extend this to other models.
In analyzing real data, when there were several alignments in the original analysis, we include only those in which we recover new breakpoints, or otherwise demonstrate our method's utility. It is implicitly assumed that in the analyses which we don't include, we came to similar or identical conclusions as the original authors.
In our analysis of simulated data, we aim to quantitatively characterize the strengths and weaknesses of the recHMM method by varying several simulation and analysis parameters. In each simulation case, ARGs (and hence of trees) were generated with RECODON [14], which uses a coalescent-based simulation approach (for exact simulation parameters, see Text S1). In keeping with the above discussion of PRI vs coalescent modeling, we filter out ARGs whose marginal trees are identical in topology using the treecomp program [15]. Thus, in all of the simulations, a perfect detector of topology change would find every breakpoint. After tree-simulation, we simulate alignments using Seq-Gen, which generates multiple alignments according to simple independent sites Markov chain models. The reason for this decoupling of tree and sequence simulation is that Seq-Gen allows for easier manipulation of the variables we're testing, namely length of region, divergence since recombination, and overall divergence (by way of branch-scaling.) [16].
After running our program on the simulated data to estimate parameters, recombinant regions are determined by a posterior decoding algorithm which we describe briefly in the Methods section and is fully outlined in Text S1. (We use posterior decoding as opposed to the Viterbi algorithm since we are primarly concerned with maximizing the expected number of correct column labelings as opposed to maximizing the probability that our state path is exactly correct.) As the notion of a ‘true negative’, a column which was correctly annotated as a non-breakpoint, is not meaningful in this case, we instead examine positive predictive value: , where a true positive (TP) is defined as a predicted breakpoint which occurs within 10 positions of a true breakpoint. Similarly, a false positive (FP) is a predicted breakpoint which has no true breakpoint within 10 positions. In plotting sensitivity: , we define a false negative (FN) to be a true breakpoint for which we have no predicted breakpoint within 10 positions.
We vary the following parameters with regard to simulation of data, the results of which are depicted in Figure 1:
We examined the effect of the following parameters in data analysis:
In our analysis of real data, we cover a range of data sizes and types, ranging between 4 and 9 taxa, with length ranging from 700 bp to circa 10,000 bp. We find that in each case, we are able to recover the previous authors' predictions for breakpoints. In many cases, we find compelling evidence for additional, often shorter recombinant regions that the original analysis either missed completely or registered as minor ‘spikes’ in their plot. In each example we highlight the aspects of our method that contribute to its sensitivity, flexibility and utility. In the case where we had no additional predictions to add to a dataset, we omitted that analysis for brevity. For example, we analyzed the data from [4], but the low mutation rate enabled their simpler approach to adequately determine breakpoints. In this situation we acknowledge that our method is able, but not necessary, to analyze the data.
We used our program to analyze data from Neisseria data that consisted of single gene regions suspected of recombination. In these analyses, recombinant regions were quite short and we demonstrate that our method is capable of handling this situation.
In their 2001 work, Husmeier and Wright use a similar tree-topology HMM to detect recombination. Since each EM iteration involves searching over all possible tree topologies for the optimal trees for each region, they were limited to alignments of 4 taxa, where there are only 3 unrooted phylogenies [12]. As mentioned earlier, both this and window-based methods assume one part of the recombination inference problem to be solved. In this case, the method allocates one tree per HMM state, and so estimation of the trees is no longer necessary, leaving only the alignment partitioning problem to be solved. Our results on this dataset are shown in Figure 3. The previous predictions are shown in red dashed lines. The horizontal axis refers to the position within the alignment, and the vertical axis is partitioned according to posterior state probability of the HMM. The posterior state probability can intuitively be thought of as the probability that a certain column was generated by a certain phylogenetic tree, taking into account the model structure and all the alignment data. At each position, the posterior probabilities for the three trees must sum to one, and hence the different colors partition the vertical axis. We were able to closely replicate their results (namely the state probabilities depicted in Figure 15 of [12]).
In comparing our results to theirs, we note that our program, which does a probabilistic tree-updating step, rather than providing a hidden state for all possible topologies, recovers all the breakpoints of the previous study. At positions 202, 507, and 538 there are clearly points at which different colors have high posterior probabilites. In regions such as 0–50, it is difficult to make reliable inferences because with so few bases, phylogenetic tree construction is unreliable. As mentioned in the Methods section, we employ a simple length cutoff heuristic whereby all recombinant regions smaller than a certain length are removed. Though this is less sophisticated than, say, explicitly specifying a prior distribution over state paths which takes length into account, it performs reasonably well for the situations we analyzed. In considering putative crossovers, points where a tree with high posterior probability changes abruptly in favor of a different tree should be considered most closely. Also, topology changes that are extremely short could be the result of spurious convergent mutations, in which two phylogenetically distant species undergo mutations to the same base, making it seem as if they had exchanged information. Note also that our method is better able to characterize the regions 537–560 and 750–780. In [12], 537–560 is classified as “irregular”, and 750–780 shows only a spike in probability in Figure 13 of [12], and not at all in their Figure 15. We predict the 500–600 region to be composed of two separate topologies, and 750–780 to be a possibly newly characterized recombinant region, having the same topology as the 100–202 region.
In [17], Bowler et al. discovered a mosaic structure in the PenA gene of Neisseria Meningitidis which conferred resistance to Penicillin. Analyzing a DNA multiple alignment between 9 species, they were able to manually determine estimates for recombination breakpoints. Constructing phylogenetic trees for each of the regions gave them clues as to the donors of the acquired regions, after which these predictions were experimentally verified. In contrast, our method is able to simultaneously partition and estimate the trees of a recombinant alignment. In Figure 3 we see that our method predicts nearly the same breakpoints with high posterior probabilities. The alignment analyzed was composed of 9 species, covering the range of virulent and commensal Neisseria subtypes, with length 1950 bp. This analysis demonstrates the ability of our phylo-HMM to effectively make use of alignments with relatively many taxa, a notable advantage over Husmeier et al.'s method. For a quantitative look at how detection power varies with taxa number, see Figure 1. By using so many subtypes for comparison, Bowler et al. were able to precisely determine which species were the donors and recipients of the recombinant regions, and subsequently verified these predictions in a laboratory setting. If they had been limited to 4 taxa, the analysis would have had to be repeated many times to cover all the possibilities. Biologically, the results in [17] motivate a search for recombination within genes implicated in resistance, in contrast to the multiple resistance gene transfer that has been previously studied, and this is a possible application of our method.
In order to determine the effectiveness of our method on longer alignments, we analyzed several datasets of entire genomes (10,000 bp) of HIV-1 strains suspected of inter-subtype recombination. Our method is equally able to perform on the genome scale as it is on the single-gene scale. In Neisseria argF, one of the predicted recombinant regions was only 30 bp long, whereas in HIV they range from 100 bp to 6 kb. This is a notable advantage over sliding-window methods which have a fixed resolution to be used over the whole scan. We demonstrate here that we are able to determine breakpoints between both large and small recombinant regions, making our method a promising tool for comparative analysis of HIV and similar genomes. In analyzing data from previous studies, we recovered all the breakpoints found by the previous authors. In cases in which we found additional breakpoints, we describe them below, but otherwise we omit the plots for brevity.
Figure 4 depicts our results on a new Malaysian HIV strain previously analyzed by Lau et al. [18]. We recover all six of the breakpoints inferred by the original authors, who used a SimPlot/Bootscanning approach, and also we find two new breakpoints whose significance appears equal to those found previously. In Figure 4, we show for comparison the results from bootscanning, which Lau et al. used for their inference of recombination breakpoints. Lau et al. provided precise breakpoint positions, and these are plotted in our diagram as red dashed lines. Since bootscanning typically removes gaps from multiple alignments before analysis, the breakpoint positions do not align with Lau et al.'s plot very well, and we provide rough mapping between plots. All six of their breakpoint predictions are well-represented in our analysis. Note the ‘spike’ in likelihood at around nt 5800 in Lau et al's plot. This region registered as strongly recombinant in our analysis, depicted as the grey region in region nt 6415–6594. Lau et al.'s characterization of the 1500 to 2000 region ( 2141 to 2856 in ours ) is marked somewhat by uncertainty in the optimal tree topology; their “% trees” line wavers and is never very close to 100%, in contrast to their inference of region 3000 to 5500, where the line remains constant and close to 100%. This uncertainty suggests that there may be additional recombination points within that region, as is more conclusively shown in our diagram. We venture that the region between nt 2141 and 2856 can be further partitioned by two more breakpoints, at nt 2360 and 2553, shown in Figure 4.
When using bootscanning, there is a lower limit to the size of the recombinant region that can be found, which depends closely on the size of the sliding window. The 3283–3617 region, at just over 300 nt, is clearly found, but smaller regions registered only as spikes or showed uncertainty of the region. Our method is probabilistic, and instead of defining sharp partitions of the alignment, we allow the parameter training to gradually decide which regions to use to train different trees. In our analyses, we consistently found that our program is able to find small recombinant regions better than others' methods. In this case, as the posterior probabilities become more certain of the alignment partitioning, each of the grey regions contributes its information in updating the grey tree. If a sliding window was being passed over the alignment, each region would have to ‘fend for itself’ in conveying its phylogenetic signal, and small regions would go undetected. Because we use information from the entire alignment during tree-optimization and sum over all possible tree-column assignments, our approach is computationally more expensive but allows collaboration between small recombinant regions, and, consequently, improved detection.
We examined data from the original SimPlot paper by Lole et al. In this work, the authors test five newly-sequence HIV-1 strains from India, and find one of them to be recombinant [10]. We examine this strain and confirm all five of their breakpoints and offer one new prediction. SimPlot detects mosaic strains by plotting the similarity of a query strain to other subtype reference strains. The similarity is computed within a sliding window of predefined size, according to various criteria (eg Hamming distance, Jukes-Cantor distance, etc). The result is a visualization of the closest relative of different regions of the query sequence. This is similar in effect to bootscanning distance-based phylogenetic methods (eg Neighbor-joining), and suffers from many of the same pitfalls. For example, in their whole-genome analysis of strain 95IN21301, Lole et al. used a window size of 600 bp, severely limiting the resolution of recombination detection. They conclusively found breakpoints around nt 6400 and 9500 (since gaps were removed, it is difficult to determine exact breakpoint predictions from their plot alone). They then did a second, finer-scale analysis with window size 200 bp on just the env and nef genes which were suspected to be recombinant. Within each of these single-gene regions they found an additional breakpoint in which the query sequence more closely represented subtype C.
In our analysis, we confirmed all five of these breakpoints by using our method (again, gap-stripping made exact comparison somewhat limited), and our result is shown in Figure 5; breakpoints previously found are in red, and new predictions in green. Since we do not have to specify a window and use instead a probabilistic weighting scheme, we are able to detect large regions (eg the break at position 6402) just as well as shorter regions (eg 6969–7073, 9431–9585). Furthermore, the method uses information from the entire alignment, rather than partitioning it by windows. In this case, it's possible that attempting to train a phylogenetic tree on the region between nt 6969 and 7073 wouldn't have yielded conclusive results. If region has high posterior probability of being generated by the ‘black’ tree that is dominant from positions 1–6402, the following M-Step will incorporate more counts from region , and so when region is examined, the inference algorithm recognizes that these columns ‘fit’ perfectly with the black topology, which corresponds to having high emission probabilities from the black HMM state. Also, a short 83 bp region is found supporting grey topology, in which 95IN21301 clusters with subtype A. This region is short, and its posterior probability never reaches 1, but a neighbor-joining tree on this section, 4328–4401 supports this clustering. In this way, our method is able to take into account information from the entire alignment, rather than defining a rigid window which can skip over small recombinant regions.
We considered data from Filho et al. [19]. Their data was composed of 10 newly sequenced strains from Brazil determined to have varying levels and structure of mosaicism, as determined by bootscan analysis. We confirm their predictions (from Figure 2 of [19]) in strains PM12313, BREPM11871, and BREPM16704 and we find several more small recombinant regions. Each of the new recombinant regions we find share breakpoints with other strains we analyzed as well as strain CRF12_BF [20], suggesting they could be hotspots for recombination activity.
As seen in Figure 6, strain BREPM12313 showed a clear recombinant region from nt 1322–2571, previously characterized by Filho, et al. Also, a region around 4700–5000 showed some evidence of recombination, having the same topology as 1322–2571. As this region's posterior probability is more ‘spike-shaped,’ rather than having sharp borders between colors, it is difficult to say whether or not it is an ambiguous region or a genuine recombinant. It does share one crossover point with strain BREPM11871, giving it somewhat more credibility. Performing neighbor-joining on nt 4784–4945 (eg positions where posterior probability is higher for grey) showed BREPM12313 clustering with subtype F. At the end of the genome, another possible recombinant region is seen, at around 9700. This region includes only gaps for the query sequence, and thus the inference is not reliable. Our method treats gaps as missing information, and when they are present in small numbers reliability is not affected, but in places like this where only gaps are present it can hinder the tree-inference.
Strain BREPM16704 was previously predicted to have four breakpoints, which we recovered with remarkably high posterior probabilities for the tree-states. Figure 7 shows our results with previous predictions in red. A new region, at 9281–9405, shows high posterior probability and is common to BREPM11871 and CRF12BF [20], making a strong case for a recombination hotspot.
In strain BREPM11871, all four breakpoints predicted by Filho et al. were found, as well as a new crossover region, common to BREPM 16704, at 9238–9361 (shown in green in Figure 8). The break previously described at nt 5462 bp was predicted by our method to be at 5277. To determine the more likely crossover point, we performed 1000 bootstrapping trials on each of the following regions: 4782–5277 (our prediction), 4782–5462 (Filho et al.'s prediction), and 5277–5462 (the disputed region). We found that the 5277–5462 region strongly supported BREPM11871 clustering with subtype B, with 98.2% bootstrap support. Moreover, bootstrap support for query-F clustering appears higher for our predicted region (99.9%) than the previous prediction (85.1%). We conclude that our algorithm often outperforms previous methods in accurately determining recombination breakpoints.
Recombination is an important force driving genome evolution, and in several cases it is the primary force for diversity. As such, methods which can detect and characterize recombination events are crucial to the successful utilization of new sequence data. On the single-gene level, recombination has been shown, in at least one case, to be able to confer antibiotic resistance [17]. It could be possible that inter-subtype recombination conferring drug resistance is a common phenomenon, which could be investigated using our methods. On the multiple-gene scale, Chlamydia trachomatis has been shown to undergo frequent inter-subtype recombination resulting in mosaic genomes [4] which complicate subtype definition and classification. On the genome scale, HIV-1 has long been known to participate in recombination leading to several identified circulating recombinant forms (CRFs). For these clinically relevant pathogens, accurate detection of recombination following introgression is important not only to guide disease treatment methods, but also for tracing the epidemiological history of the virus. In this work we present a novel method for recombination detection which we believe to be more sensitive, flexible, and robust in the aforementioned evolutionary scenario. We combine two long-standing concepts, phylogenetics and hidden Markov models, in a maximum-likelihood framework to model topology changes over an alignment of related sequences.
We present a training scheme which attempts to solve the two problems embedded within recombination detection simultaneously. We model evolution probabilistically with a continuous-time Markov chain which directs the likelihood-based tree construction algorithm [21]. Furthermore, our alignment-partitioning is handled with posterior probabilities which take into account each hidden state tree. By summing over all possible tree-column assignments and not using sharp window cutoffs, we are able to perform more precise breakpoint determination. We can adjust the specificity and sensitivity of the model with the transition matrix of the HMM, which dictates how much of a likelihood change should cause the model inference to change states. We believe this likelihood comparison to be superior to adjusting the size of the window because it enables distant sections of the alignment to combine their phylogenetic signal in training a hidden state of the HMM.
The goal of this study is to start with a multiple alignment of sequence data and find positions where recombination events have occurred. This is done by recovering a set of phylogenetic trees and a map that assigns a tree to each column. The points at which neighboring columns have different tree assignments will indicate possible locations of recombination events in evolutionary history.
We use a hidden Markov model with tree topologies as hidden states and alignment columns as observed states. The usual method to train HMM parameters is by the specialization of the EM algorithm known as the Baum-Welch algorithm. Our transitions can be optimized in the usual way, but the emissions are more difficult since their likelihood is governed by the tree topologies in the hidden states, which are not easily optimized. For this problem, we employ an EM method for trees, developed by Friedman et al. [21], within our M-step for emission probabilities. Their phylogenetic inference algorithm is implemented with such improvements as simulated annealing and noise injection in SEMPHY, available from their website at http://compbio.cs.huji.ac.il/semphy/. We instead implemented our own ‘bare-bones’ version of this algorithm, without these improvements. To progressively assign columns to trees, at each M-step, we weight the expected counts of the tree-EM by the posterior counts of the phylo-HMM. Intuitively, this guides the tree-maximization by providing comparatively more information from regions which fit a particular tree. We give a high-level description of our method here and more detail is provided in Text S1. Figure 10 gives a graphical representation of the overall task-flow. After running our program on the alignment data to estimate parameters, a state path through the HMM is computed from the final matrix of posterior probabilities. We would like this path to represent a balance between being biologically reasonable and highly probable under our model. Thus, we propose maximizing the sum of posterior state probabilities subject to a length cutoff by the following method.
Let be a path of length through the state space of the HMM (discounting start and end states), emitting the first columns of the alignment, with one state per column. Let be the total number of columns in the alignment. We say that the state path is partial if , and complete if .
Let denote the state in path . The score of path is defined as where is the posterior probability that column was emitted by state .
We say a state path is valid if all its breakpoints are more than apart. That is, there exist no with such that and .
Finding the maximal valid is solved by a simple dynamic programming procedure, outlined in Text S1.
Since the EM algorithm has a tendency to converge to local likelihood maxima which are not global maxima, especially when initialized randomly, we ran the algorithm several times for each dataset, took the highest-likelihood result for the set of trials, and performed the above posterior decoding on the final distribution. We show plots of our program's performance when various aspects of the model and input alignment are varied in Figure 9.
The hidden data for each tree are defined as , the number of transitions from nucleotide to nucleotide from node to node , for all pairs of nodes. For neighboring nodes, this can be computed by Elston and Stewart's Peeling algorithm [24], and for non-neighboring nodes, exact computation requires a dynamic programming routine described in the original Structural EM paper [21]. In this work, the authors showed that the likelihood contribution of an edge between two nodes can be summarized as a function of this count .
If we arrange these summaries in a weight matrix in which represents the expected likelihood contribution resulting from placing an edge from node to node , it is easy to see that the maximum expected likelihood tree will be the topology which maximizes the sum of its edge scores. Finding such a topology is trivial if we do not require that the tree is binary, for instance by maximum spanning tree algorithms. This (possibly non-binary) tree is then transformed to a binary tree by operations which do not alter the tree's likelihood. In this way, Structural EM allows for iteratively constructing higher likelihood trees by choosing the next tree which maximizes the expected likelihood based on the current tree. The reader is referred to Text S1 and the original Structural EM paper [21] for more detailed discussions of this algorithm which is crucial to our method.
In our methods, instead of allowing Structural EM to converge, we allow two iterations using the same set of transition and posterior probabilities, as a heuristic substitute for finding the true hidden tree topologies.
We outline here a number of extensions which could grow directly from this work. One of the strengths of the method is its generality and flexibility, and so we believe it is ideally suited for continued development.
At each training iteration:
- If the trees maximized in each hidden state are not consistent:
A simpler way of estimating would be to run a coarser heuristic method (e.g., SimPlot) and seed the HMM with the number of states that it finds.
All sequence data used in this study was downloaded from public databases (GenBank and LANL HIV Database). The sequences were aligned with MUSCLE [27] with the default parameters. Gaps in the alignments were treated as missing information. Bootstrap analyses were performed with CLUSTAL W [28] with 1000 replicates and the default parameters. The GenBank identifiers for sequences used are as follows, grouped by figure: Figure 3: argF: X64860, X64866, X64869, X64873; penA: X59624–X59635; Figure 5: AF067158, AB253429, AF067159, M17451, AF005494; Figure 4: AB032740, AB023804, AY713408, EF495062; Figures 6–8: AF286228, AF005494, AY173956, AB098332, AY173956, DQ085867, DQ085876, DQ085872.
The source code for our programs, though still being developed, is available upon request or through CVS. For documentation, contact, and download information See http://biowiki.org/RecHMM
|
10.1371/journal.pgen.1005916 | Cdkn1c Boosts the Development of Brown Adipose Tissue in a Murine Model of Silver Russell Syndrome | The accurate diagnosis and clinical management of the growth restriction disorder Silver Russell Syndrome (SRS) has confounded researchers and clinicians for many years due to the myriad of genetic and epigenetic alterations reported in these patients and the lack of suitable animal models to test the contribution of specific gene alterations. Some genetic alterations suggest a role for increased dosage of the imprinted CYCLIN DEPENDENT KINASE INHIBITOR 1C (CDKN1C) gene, often mutated in IMAGe Syndrome and Beckwith-Wiedemann Syndrome (BWS). Cdkn1c encodes a potent negative regulator of fetal growth that also regulates placental development, consistent with a proposed role for CDKN1C in these complex childhood growth disorders. Here, we report that a mouse modelling the rare microduplications present in some SRS patients exhibited phenotypes including low birth weight with relative head sparing, neonatal hypoglycemia, absence of catch-up growth and significantly reduced adiposity as adults, all defining features of SRS. Further investigation revealed the presence of substantially more brown adipose tissue in very young mice, of both the classical or canonical type exemplified by interscapular-type brown fat depot in mice (iBAT) and a second type of non-classic BAT that develops postnatally within white adipose tissue (WAT), genetically attributable to a double dose of Cdkn1c in vivo and ex-vivo. Conversely, loss-of-function of Cdkn1c resulted in the complete developmental failure of the brown adipocyte lineage with a loss of markers of both brown adipose fate and function. We further show that Cdkn1c is required for post-transcriptional accumulation of the brown fat determinant PR domain containing 16 (PRDM16) and that CDKN1C and PRDM16 co-localise to the nucleus of rare label-retaining cell within iBAT. This study reveals a key requirement for Cdkn1c in the early development of the brown adipose lineages. Importantly, active BAT consumes high amounts of energy to generate body heat, providing a valid explanation for the persistence of thinness in our model and supporting a major role for elevated CDKN1C in SRS.
| Silver Russell syndrome is a severe developmental disorder characterised by low birth weight, sparing of the head and neonatal hypoglycemia. SRS adults are small and can be extremely thin, lacking body fat. Numerous genetic and epigenetic mutations have been linked to SRS primarily involving imprinted genes, but progress has been hampered by the lack of a suitable animal model. Here we describe a mouse model of the rare micro duplications reported in some SRS patients, which recapitulated many of the defining features of SRS, including extreme thinness. We showed that these mice possessed substantially more of the energy consuming brown adipose tissue (BAT), driven by a double dose of the imprinted Cdkn1c gene. We further show that Cdkn1c is required for the postranscriptional accumulation of the BAT determinant PRDM16 and that these proteins co-localise to the nucleus of in a rare label-retaining cell within BAT. These data suggest that Cdkn1c contributes to the development of BAT by modulating PRDM16 and supports a major role for this gene in SRS.
| Silver-Russell syndrome (SRS; MIM 180860), Beckwith Weidemann Syndrome (BWS; MIM 130650) and IMAGe syndrome (MIM 614732) are all rare imprinted developmental disorders that occur as a result of genetic or epigenetic alterations primarily at human chromosome 11p15 [1, 2]. Recent studies highlight the potential involvement of one maternally expressed imprinted gene, CYCLIN DEPENDENT KINASE INHIBITOR 1C (CDKN1C), in all three disorders [3]. Loss-of-function or loss-of-expression of CDKN1C is a common feature of BWS, either through direct DNA mutation, epigenetic misregulation or loss of the maternal chromosome [4]. The rare IMAGe syndrome, which has the major features of fetal growth restriction, metaphyseal displasia, adrenal hypoplasia congentia and genital abnormalities, is associated with genetic mutations in the CDKN1C gene [5, 6]. The changes associated with growth restriction are gain-of-function mutations of the PCNA domain, limited to a handful of rare familial cases highlighted in a recent review [3], that may increase the stability of the protein [6, 7]. SRS is characterised by severe pre and post natal growth restriction combined with some of the following: neonatal hypoglycaemia, excessive sweating, triangular shaped face, head circumference of normal size but disproportionate to a small body size, clinodactyly, feeding problems, low body mass index manifesting as extreme thinness, no catch up growth and increased risk of delayed development and learning disabilities [8]. Numerous genetic and epigenetic alterations have been reported in SRS patients but identifying the causal gene mutation(s) has been challenging. Some studies suggest loss of function of the paternally expressed growth factor INSULIN-LIKE GROWTH FACTOR 2 (IGF2) [9]. However, there are SRS patients that carry an extra copy of maternally derived 11p15 without loss-of-function of IGF2 [10]. Maternal duplications spanning the complex imprinted domain at 11p15 have been independently reported in a number of studies [11–16] and the majority are associated with unbalanced translocations suggesting that increased dosage of a maternally expressed imprinted gene may be important in SRS. The minimal region of maternal microduplication in SRS encompasses CDKN1C and three other maternally expressed protein coding genes POTASSIUM CHANNEL, VOLTAGE GATED KQT-LIKE SUBFAMILY Q, MEMBER 1 (KCNQ1), PLECKSTRIN HOMOLOGY-LIKE DOMAIN, FAMILY A, MEMBER 2 (PHLDA2) and SOLUTE CARRIER FAMILY 22, MEMBER 18 (SLC22A18) [17, 18]. Since CDKN1C is a maternally expressed gene [19, 20], these SRS patients are predicted to have twice the normal level of CDKN1C expression.
We, and others, have shown that loss of Cdkn1c in mice results in a late fetal overgrowth and disrupted placental development [21–25] consistent with a key role for this gene in BWS. CDKN1C, which is maternally expressed in both humans and mice [19, 20], belongs to the Kip cyclin dependent kinase inhibitor family that induce cell cycle arrest and limit proliferation [26]. In mice, Cdkn1c is widely expressed during embryonic development in cells exiting differentiation [27–29]. Cdkn1c also functions to orchestrate cell fate determination targeting key transcription factors [30–36] and in stem cell self-renewal and quiescence in a number of embryonic [32–34, 37–40] and adult [41–43] stem/progenitor cells. These multiple roles may account for complex phenotypic consequences in response to alterations in the dosage of this gene.
We previously reported growth restriction in mice carrying a bacterial artificial chromosome (BAC) transgene spanning Cdkn1c, Phlda2 and Slc22a18 [24]. This alteration essentially models the minimal microduplicated region observed in some SRS patients [17, 18]. The mice exhibited significant fetal growth restriction from embryonic day (E) 13.5 with the absence of catch-up growth. We were able to attribute the fetal growth restricting properties of this microduplication to two-fold expression of Cdkn1c consistent with the phenotype observed in SRS. Fetal growth restriction per se is a relatively generic phenotype and more specific features of SRS would lend greater support to the hypothesis that altered expression of CDKN1C contributes significantly to SRS in human patients. To provide further evidence for or against a key role for CDKN1C in SRS, we examined the microduplication mice for additional SRS-associated phenotypes. This work revealed low birth weight with a relative sparing of the head, neonatal hypoglycaemia, small sized adults with substantially less white adipose tissue, all of which were genetically attributable to just two-fold expression of Cdkn1c. These findings support a major role for elevated CDKN1C in SRS. Importantly, we identified a novel function for Cdkn1c in directly promoting the development of brown adipose tissue early in life, a finding that could account for the prevalence of thinness in SRS.
The minimal microduplicated region in SRS spans four imprinted, protein-coding genes: KCNQ1, CDKN1C, PHLDA2 and SLC22A18 (Fig 1A). Our mouse BAC transgene spans three of these genes, Cdkn1c, Phlda2 and Slc22a18 (Fig 1B). Previously we showed that a single copy of the transgene (BACx1) drove significant fetal growth restriction on a mixed strain background which was more severe in a two copy line (BACx2) and absent in a reporter line in which transgenic Cdkn1c was replaced by a β-galactosidase gene (BAC-lacZ) attributing growth restriction to elevated expression of Cdkn1c [24]. The transgene was lethal on a pure 129S2/SvHsd (129) background but preliminary breeding into C57BL/6J (BL6) suggested improved viability with the retention of growth restriction, albeit attenuated [24]. For this study, three transgenic lines (BACx1, BACx2 and BAC-lacZ) were bred further onto a BL6 strain background, to >12 generations. BACx1 and BACx2 fetuses were significantly lighter at embryonic day (E) 18.5 while BAC-lacZ fetuses were similar in weight to wild type controls confirming the fetal growth restricting properties of Cdkn1c. Importantly, fetal viability was not compromised on this genetic background (S1 Fig).
Children with SRS are born low birth weight and are prone to develop spontaneous hypoglycaemia, particularly if they are not fed both frequently and regularly [44]. Newborn BACx1 and BACx2 mice were lighter than their wild type littermates (Fig 1C). There was a significant difference in the relative proportion of the brain to body weight in the two copy line (Fig 1D). Marked hypoglycaemia in the fed state was evident for both the single copy and the two copy line (Fig 1E). Glucose levels, birth weight and brain weights were normal in the control line BAC-lacZ genetically attributing these phenotypes to elevated Cdkn1c expression. These data were consistent with the observed phenotypes of low birth weight, head sparing and neonatal hypoglycaemia reported in young SRS patients.
As adults, SRS patients commonly display short stature with a low body mass index and a lack of subcutaneous fat [8]. At 10 weeks of age BACx1 and BACx2 male and female adult mice were significantly lighter than their wild type littermates (Fig 2A). An exploratory magnetic resonance image of an adult BACx1 male mouse alongside a wild type co-housed littermate suggested less white adipose tissue (WAT; S2 Fig). Dissection and weighing of individual WAT depots revealed a disproportionate reduction in the weight of the mesenteric (mes), inguinal (ing) and retroperitoneal (rp) WAT depots relative to total body weight (Fig 2B and 2C). Mice carrying a single copy of the transgene (BACx1) consumed a similar daily weight of standard chow whereas mice carrying two copies (BACx2) consumed significantly less (Fig 2D). The basal body temperature of both BACx1 and BACx2 mice was significantly elevated (Fig 2E). Histological examination of rpWAT revealed an abundance of smaller adipocytes with a multilocular appearance, which were less apparent in wild type rpWAT and depots from the reporter line BAC-lacZ (Fig 2F). BEIGE cells are a recruitable form of brown adipose tissue (BAT) that develops postnatally within some WAT depots defined by a smaller cell size, multilocular lipid droplet morphology, a high mitochondrial content and the expression of brown fat–specific genes [45–50]. In addition to the elevated expression of Cdkn1c two key markers of brown adipose tissue, uncoupling protein-1 (UCP1) and elongation of very long chain fatty acids (FEN1/Elo2, SUR4/Elo3 and yeast)-like 3 (Elovl3), were significantly elevated in rpWAT from adult BACx1 mice as compared to matched wild type rpWAT (Fig 2G). Importantly, neither Cdkn1c nor these markers were elevated in rpWAT from the reporter line BAC-lacZ (Fig 2G). These data suggested an increased representation of BEIGE cells, sometimes referred to as the “browning” of WAT, driven by increased expression of Cdkn1c.
Cdkn1c is expressed from the BAC in a number of tissues including the pituitary, the hypothalamus and the pancreas [51] that may stimulate the browning of WAT. However, elevated Cdkn1c expression within transgenic rpWAT (Fig 2G) suggested the potential for a direct role for Cdkn1c in brown adipogenesis. Expression of Cdkn1c has been reported in the epididymal white adipose tissue of adult mice [52]. At postnatal day 7 (P7), Cdkn1c expression was detectable within several adipose depots with levels positively correlating with the brown adipose-like nature of these depots [53]. Cdkn1c was found to be most highly expressed in the interscapular-type brown fat depot (iBAT), which is composed of a classical or canonical type of BAT sharing a developmental origin with myoblasts [48], with moderate expression in rpWAT and subcutaneous (sc) WAT and lowest expression in mesenteric (mes) WAT (Fig 3A). At E16.5, when iBAT is discernable as a discrete depot, a few Cdkn1c-positive cells were identifiable by both in situ hybridisation and immunohistochemistry (Fig 3B). At P7, Cdkn1c was more widely expressed within the iBAT depot (Fig 3C, left panel). Cdkn1c was also expressed within a few discrete niches in P7 rpWAT (Fig 3D, left panel). Importantly, BACx1 and BACx2 rpWAT and iBAT displayed a similar expression pattern to WT depots by in situ hybridisation (Fig 3C and 3D, middle panels) indicating that Cdkn1c was not ectopically expressed from the transgene in these depots. β-galactosidase staining of dissected intact depots from BAC-lacZ pups revealed blue staining niches consistent with expression originating from the transgene in both depots (Fig 3C and 3D, far right panels).
To determine whether expression of Cdkn1c was imprinted in adipose tissue, we made use of the Cdkn1c restriction fragment length polymorphism (RFLP) assay [54]. Mus musculus domesticus BL6 mice possess an AvaI restriction enzyme site within an exon of Cdkn1c that is absent in Mus spretus mice (Fig 3E). P7 pups were generated from crosses between pure BL6 females and BL6 males carrying a copy of the Mus spretus Cdkn1c region. AvaI digestion of a PCR product amplified across the polymorphic region from genomic DNA demonstrated that both alleles were physically present. Digestion of the PCR product amplified from cDNA revealed the predominant presence of only the maternally inherited BL6 allele (lower band) in both P7 iBAT and P7 rpWAT (Fig 3E). Similarly, depots from adult mice displayed predominantly maternal-allele expression (Fig 3E). Differential DNA methylation spanning the predicted Cdkn1c promoter region [55] was also discernable in both adipose depots at P7 and in the adult (Fig 3F). These data demonstrated that Cdkn1c was both expressed and imprinted in post-natal adipose tissue, and that both expression and imprinting was maintained into adulthood.
The in situ hybridisation analysis (Fig 3D) and further histological examination of rpWAT at P7 suggested that the phenotypic differences present in adult mice were apparent at this much earlier timepoint (Fig 4A). Electron microscopic imaging of BACx1 P7 rpWAT depots revealed clusters of cells that possessed BEIGE characteristics including a larger volume of cytoplasm, numerous mitochondria and smaller, multilocular, lipid droplets not readily observed in matched WT depots (Fig 4B). QPCR demonstrated that Cdkn1c expression was significantly elevated in BACx1 and BACx2 P7 rpWAT, by 1.5- and 2.2-fold respectively (Fig 4C). Several markers of BAT were also elevated including peroxisome proliferator-activated receptor gamma, coactivator 1 alpha (Ppargc1a), cell death-inducing DFFA-like effector a (Cidea), Ucp1, Elovl3 and PR domain containing 16 (Prdm16) in BACx1 and BACx2 P7 rpWAT (Fig 4D). Consistent with 10-fold higher expression of Ucp1 mRNA, UCP1 protein was more readily detectable in BACx1 rpWAT than in WT rpWAT in a within litter comparison (Fig 4E). PRDM16, a brown fat determinant [48], was also more readily detectable in BACx1 rpWAT depots than wild type depots (Fig 4E). Importantly, P7 rpWAT from the reporter line BAC-lacZ had a normal appearance and neither Cdkn1c nor key BAT markers were elevated (S3A Fig). The presence of BEIGE-like cells in the BACx1 and BACx2 rpWAT and their absence in the BAC-lacZ model, in which Cdkn1c was expressed at a normal level, identified Cdkn1c as a gene that promotes the “browning” of WAT. Importantly, this phenotype was apparent when WAT first emerged as a distinct depot in very young mice.
Elevated expression of Cdkn1c also had an effect on iBAT. P7 BACx1 and BACx2 iBAT depots were heavier as a proportion of total body weight, by 30% and 60% respectively, than WT iBAT depots (Fig 5A). This was not due to increased lipid deposition as BACx1 and BAC2 iBAT depots displayed increased cellularity, confirmed by cell counting (Fig 5B). As in rpWAT, Cdkn1c expression was significantly elevated in BACx1 and BACx2 iBAT, by 1.5- and 2.2-fold respectively (Fig 5C). QPCR analysis revealed near wild type expression of the adipogenesis regulators retinoblastoma 1 (Rb1), peroxisome proliferator-activated receptor-γ (PPARγ) and CCAAT-enhancer-binding protein-α (C/EBPα) but elevated expression of CCAAT-enhancer-binding protein-β (C/EBPβ) in both BACx1 and BACx2 depots (Fig 5C). BAT markers Ppargc1a, Ucp1 and Elovl3 were significantly elevated in BACx1. All five BAT markers examined were significantly elevated in the higher dosage line, BACx2 (Fig 5C). Critically, BAC-lacZ iBAT appeared morphologically normal and neither Cdkn1c nor key BAT markers were elevated (S3B Fig) genetically assigning these alterations to the increased dosage of Cdkn1c in BACx1 and BACx2. The ratio of mitochondrial DNA to nuclear DNA can be used as an estimate of mitochondrial load. Both BACx1 and BAC2 P7 iBAT depots contained significantly increased mitochondrial DNA content compared to WT (Fig 5D). Consistent with a greater mitochondrial load, expression of the nuclear mitochondrial marker cytochrome c, somatic (Cycs) was significantly elevated in BACx2 and both Cycs and the mitochondrion-encoded cytochrome c oxidase subunit II (Cox2) were elevated in BAC1 and BACx2 iBAT (Fig 5C).
Fully functional iBAT at birth is important for maintaining newborn body temperature. 33°C approaches thermoneutrality and corresponds to the temperature within litters of newborn mice in contact with their mother [56] whereas 26°C elicits near-maximal thermogenesis by brown adipose tissue [57]. P2 WT and BACx2 pups kept at 33°C and then exposed to room temperature (22°C) for a 20 minute period both lost body heat at the same rate despite significant differences in their body weights (Fig 5E), consistent with functional iBAT at this timepoint.
Taken together, these data identified a novel function for Cdkn1c in boosting the development of both BEIGE and iBAT early in post-natal life, with increasing expression of Cdkn1c associated with the increased development of brown adipose.
In contrast to the increase in classic iBAT in response to elevated Cdkn1c, loss-of-expression of Cdkn1c resulted in the loss of iBAT. Mice inheriting a targeted deletion of Cdkn1c maternally (loss-of-function) die in the neonatal period [21]. Cdkn1c-/+ (KOMAT) embryos examined a day prior to neonatal demise, at E18.5, possessed poorly discernable iBAT depots lacking the characteristic butterfly shape normally observed at this stage of development (Fig 6A). H&E staining of the isolated KOMAT depots revealed a disorganised morphology with large areas of lipid (Fig 6B). QPCR analysis confirmed considerably reduced expression of Cdkn1c. KOMAT expressed relatively normal levels of Rb1, PPARγ and C/EBPα (Fig 6C). C/EBPβ was expressed at 50% the wild type level reciprocal to the increased expression observed in response to elevated Cdkn1c. Pan-adipocyte markers fatty acid binding protein 4 (Fabp4) and perilipin 1 (Plin1) were also markedly reduced. Prdm16 was expressed at wild type levels while expression of peroxisome proliferator-activated receptor gamma, coactivator 1 alpha (Ppargc1a), which encodes a transcriptional coactivator that is involved in the activation of brown fat cells, was markedly elevated indicating that the initiation of brown adipocyte commitment was not prevented by loss of Cdkn1c. Nonetheless, there was a marked reduction in expression of downstream genes required for brown adipose development and function including the cAMP-inducible gene, Ucp1, and the cAMP insensitive genes Cidea and Elovl3 (Fig 6C). Expression of the nuclear mitochondrial marker Cycs and the mitochondrion-encoded Cox2 were also diminished, by 25–30% (Fig 6C). Mitochondrial DNA content was 40% less than the wild type level (Fig 6D). UCP1 and PRDM16 proteins were barely detectable in Cdkn1c KOMAT iBAT (Fig 6E and 6F), all indicative of severely compromised BAT development. Loss of function of PRDM16 in iBAT early in life results in a switch from an iBAT identity towards a skeletal muscle identity [48]. Consistent with the loss of PRDM16, Cdkn1c KOMAT iBAT expressed two-fold higher levels of the skeletal muscle-selective genes myogenic factor 5 (Myf5) and myogenic differentiation 1 (Myod1) (Fig 6C) further supported by western analysis for MYOD1 (Fig 6G). These data identified a requirement for Cdkn1c in the development of classic BAT.
To determine whether Cdkn1c could function intrinsically to boost brown adipogenesis, we performed an ex-vivo adipogenesis assay. Mouse embryonic fibroblasts (MEFs) are multipotent and have the potential to differentiate into brown adipocytes. MEFs were isolated from E12.5 BACx1 and WT fetuses and induced to differentiate using a standard adipogenic protocol [58]. The expression profile of Cdkn1c in both WT and BACx1 MEFs followed similar pattern of up regulation by day 2 and down regulation by day 8 of differentiation, with BACx1 MEFs expressing consistently higher levels of Cdkn1c at each time point (Fig 7A). Having confirmed elevated expression of Cdkn1c in the differentiating MEFs, a single copy of the Cdkn1c transgene was genetically combined with a maternally inherited targeted deletion of Cdkn1c (KOMAT) to generate MEFs of four genotypes: WT, BACx1, KOMAT and KO+BACx1. After 8 days of adipocyte-directed differentiation Cdkn1c was expressed 1.4-fold the WT level in BACx1 MEFs and at barely detectable levels in KOMAT MEFs (Fig 7B). KO+BACx1 MEFs, which carried both the transgene and the targeted allele, expressed Cdkn1c at WT levels (Fig 7B). All four genotypes differentiated into lipid-containing cells, as evidenced by Oil-Red O staining and mRNA levels for general adipogenic markers (Fig 7C and S4 Fig). As in vivo, key markers of BAT fate and function Cidea, Ucp1 and Elovl3 were elevated in BACx1 D8 MEFs (Fig 7D). Importantly, KO+BACx1 MEFs, in which Cdkn1c was expressed at WT levels, did not display altered expression of these markers (Fig 7D). After 8 days of adipocyte-directed differentiation BACx2 D8 MEFs displayed 2.4-fold elevated expression of Cdkn1c and further elevated expression of several BAT markers (Fig 7E) consistent with the dosage-related function of Cdkn1c in inducing a BAT-like gene program. Confocal imaging suggested more mitochondria in the BACx1 differentiated samples (Fig 7F), consistent with in vivo data (Fig 4B). UCP1 protein was detectable in BACx1 D8 MEFs but not WT MEFs, a difference further highlighted by exposure to the positive regulator of Ucp1 gene transcription, retinoic acid [59, 60] (Fig 7G). Taken together, these data demonstrated that Cdkn1c can drive a BAT-like cell fate in adipocyte-differentiated fibroblast cells ex-vivo, and in a dosage-sensitive manner.
There was a considerable loss of PRDM16 protein in Cdkn1c KOMAT iBAT (Fig 6F) but Prdm16 mRNA levels were relatively unaltered (Fig 6C) suggesting a function for Cdkn1c in the post transcriptional regulation of PRDM16. Consistent with this role, CDKN1C protein co-localised with the brown fat determinant, PRDM16, to the nucleus of rare cells present within P7 iBAT (Fig 8A). Acute loss of CDKN1C, driven by siRNA transfection of the brown preadipocyte cell line HIB-1B [61], resulted in a reduction of PRDM16 protein (Fig 8B). Prdm16 and Cdkn1c are both known to be functionally important for adult haematopoietic stem cells [41, 62] and adult neural stem cells [63, 64]. 5-bromo-2-deoxyuridine (BrdU) label retention has been defined as a characteristic attributed to slow-cycling adult stem cells [65]. In two independent experiments, pregnant females were injected with BrdU (single dose at E16.5 or four doses of BrdU from E16.5). Within the adult iBAT from offspring of these pregnancies CDKN1C/PRDM16 double positive cells retained BrdU for six to eight weeks after embryonic labeling (Fig 8C). Taken together, all our data suggest that CDKN1C functions to support the post transcriptional accumulation of PRDM16 in a progenitor cell, and thus promotes the development of brown fat.
Here we provide in vivo evidence for key features of SRS in a novel mouse model of the minimal microduplicated region reported in some patients with this syndrome [66, 67] including low birth weight, head sparing, neonatal hypoglycaemia, smallness as adults and an extreme lack of body fat. Critically, we show that these phenotypes were due solely to the two-fold increased dosage of Cdkn1c consistent with the predicted expression levels in SRS patients. In our model, Cdkn1c was not ectopically expressed nor was Cdkn1c expressed at excessively high levels thus our findings are physiologically relevant. In addition to providing compelling evidence for a major role of elevated CDKN1C is SRS, we demonstrated in vivo and ex-vivo that Cdkn1c promotes the formation of brown adipose tissue, both the classic form exemplified by the iBAT depot and also the BIEGE form that emerges within WAT depots which persists into adulthood. Moreover, our data suggest that Cdkn1c functions to boost BAT, in part, by supporting protein accumulation of the brown fat determinant, Prdm16. This work has implications both for the diagnosis of SRS and the clinical management of SRS patients.
Microduplication mice were born low birth weight with a relative sparing of the head and neonatal hypoglycaemia. As adults the mice failed to catch-up in weight with their littermates and possessed substantially less white adipose tissue. We were able to exclude a role for two other genes present on the BAC (Phlda2 and Slc22a18) in driving these phenotypes by using a reporter line in which expression of the BAC copy of Cdkn1c was replaced by lacZ [51]. While fetal growth restriction and low birth weight are relatively common complications of pregnancy that can have numerous origins, the more specific features of SRS support a major role for elevated CDKN1C expression in SRS. Currently the diagnosis of SRS is hampered by the complexity of alterations reported in different patients and the variable presentation of phenotypes. Moreover, some alterations may have an epigenetic origin not detectable by traditional DNA based approaches or not present in accessible tissues. The greater certainty that CDKN1C is a major contributor to SRS should lead to the development of better diagnostic tools and potentially the improved sub-classification of patients. It would now seem pertinent to examine BAT in SRS patients and, conversely, to assess individuals with a diagnosis of fetal growth restriction followed by extreme thinness for alterations in the expression of CDKN1C.
In addition to observing several defining features of SRS in our microduplication model, we identified Cdkn1c as a gene that functions in vivo, and in a dosage sensitive manner, to boost the amount of brown adipose tissue that develops early in life. Elevated Cdkn1c was associated with an increased amount of BEIGE adipose (non-classic BAT) located within the rpWAT depot in very young and in adult mice. Transgenic rpWAT depots had a marked appearance of BAT-like niches and expressed much higher levels of several BAT markers including Elovl3 and Cidea, markers that are insensitive to cAMP. Both UCP1 and PRDM16 protein were readily detectable in BACx1 rpWAT depots in comparison to wild type depots in within litter comparisons. Elevated Cdkn1c also resulted in a larger iBAT depot relative to body weight in young mice and augmented the existing brown adipose gene program. The function of Cdkn1c in boosting the formation of BAT early in life would explain neonatal hypoglyceamia and the failure of our mice to lay down sufficient stores of white adipose tissue into adulthood manifesting as thinness.
While Cdkn1c was expressed from the BAC in a number of tissues including the pituitary, the hypothalamus and the pancreas [51] that may stimulate the browning of WAT, Cdkn1c was expression and imprinted within both rpWAT and iBAT depots. Importantly, elevated Cdkn1c enhanced the expression of brown adipose marker genes in adipogenically-differentiated MEFs. Normalising Cdkn1c by combining a single copy of the transgene with maternal inheritance of the targeted Cdkn1c allele in this same experiment resulted in wild type levels of both Cdkn1c and the BAT markers. This experiment demonstrated the intrinsic ability of Cdkn1c to drive a BAT-like gene program ex-vivo. Our findings that Cdkn1c plays a key role in promoting BAT development is novel and has important implications both for our understanding of BAT development.
Elevated Cdkn1c boosted the development of BAT while loss-of function of Cdkn1c resulted in abnormal morphology of the iBAT depot alongside a striking reduction in the expression of several brown adipose markers, and loss of UCP1 and PRDM16 protein. Classic iBAT derives from a common progenitor to skeletal muscle and a switch between these two lineages is thought to be controlled by Prdm16 [48]. Consistent with loss-of-function of Prdm16, Cdkn1c KO iBAT expressed elevated levels of two muscle-specific genes. While PRDM16 protein was barely detectable, Cdkn1c KO iBAT expressed normal levels of Prdm16 mRNA. Acute knock-down of CDKN1C in a brown fat cell line resulted in the loss of PRDM16 protein suggesting that Cdkn1c acts to regulate the post transcriptional accumulation of PRDM16. A precedent exists for Cdkn1c in regulating the post-transcriptional accumulation of several other transcription factors [30–36, 68, 69]. Moreover, co-expression of CDKN1C and PRDM16 in the nucleus of a rare, BrdU label-retaining cell in iBAT suggests that regulation takes place with an adult brown adipose progenitor cell. Prdm16 and Cdkn1c are both already known to be functionally important for adult HSC [41, 62] and adult NSC [63, 64]. However, it remains controversial whether label retention is a definitive feature of stem cells and further work is required demonstrate that the PRDM16/CDKN1C double positive cells are indeed brown fat progenitors. What is clear is that both Prdm16 and Cdkn1c are required for the proper determination of BAT cell fate, as evidenced by elevated expression of the myogenic markers Myf5 and MyoD in response to loss-of-function of Prdm16 [48] and Cdkn1c (Fig 6). Rather than participating in cell fate decisions, we propose that Cdkn1c modulates the accumulation of PRDM16 to promote “brownness”, acting downstream of cell fate choice.
Our mouse model recapitulated several defining features of SRS but there are potential limitations with this study. Firstly, the human and mouse CDKN1C predicted proteins share amino acid sequence conservation in the cyclin-dependent kinase inhibitory domain and in the QT domain, but the internal proline-rich and an acidic repeat domains found in the mouse sequence are replaced by a single PAPA repeat in the human sequence [27]. A key question that therefore arises is whether CDKN1C functions in humans to regulate brown adipogenesis? Although a low body mass index is consistent with more brown adipose tissue, we can find no report examining brown adipose tissue in SRS patients. However, recent data suggest that increased methylation at CDKN1C is associated with a higher BMI in a normal population [70] which holds promise. Secondly, while the transgenic model partially recapitulates the minimal microduplication observed in SRS, some key Cdkn1c enhancers located at a distance from the gene are absent from the mouse transgene [51]. We have examined the consequences of increased dosage in only a subset of tissues in which Cdkn1c is normally expressed, which excludes skeletal muscle and cartilage. However, this is likely to also be true for the SRS syndrome patients with smaller microduplications as the human CDKN1C enhancers are also located at a distance from the gene body [55].
Loss-of-function of CDKN1C in humans has been reported in cases of BWS. Excessive weight gain, which might be anticipated from a lack of BAT, is not a feature of BWS. BWS children can display neonatal hypoglycemia and one recent study reported early onset diabetes in a family with a mutation in CDKN1C [71], all of which could suggest a metabolic function for CDKN1C in humans. There are differences in the epigenetic regulation between humans and mice with some expression from the paternal allele in humans [55] which may attenuate the phenotype in BWS. Our findings may therefore have implications for several rare human imprinting disorders.
There is now sufficient evidence from animal models and human studies to indicate a key role for the imprinted CDKN1C gene in SRS, BWS and IMAGe syndrome (Fig 9). This knowledge will undoubtedly improve our understanding of these complex childhood growth disorders and their longer term implications. From an evolutionary perspective, our finding that Cdkn1c acts early in life to promote the formation of brown adipose tissue in mice is also intriguing. Thermogenesis is critical for the survival of young mammals before the development of subcutaneous fat and hair but comes at an energetic cost to the individual. Cdkn1c is both a BAT-promoting gene and one that negatively regulates embryonic growth [24, 25]. Our data predict that silencing of Cdkn1c by the paternal genome, which occurred after mammals diverged from marsupials [72], would result in larger offspring with the simultaneous reallocation of resources away from maintaining body temperature towards supporting the enhanced growth, providing a clear competitive advantage and lending support to the hypothesis that thermogenesis is an arena for genomic conflict in mammals [73].
In conclusion, this work provides genetic evidence from a physiological relevant animal mode that Cdkn1c functions to boosts the development of BAT in mice. This work fundamentally establishes that Cdkn1c gene dosage, rather than gene function per se, plays a key role in this process. We critically show that relatively small (< two-fold) changes in gene expression can have a dramatic consequence for development in mice with long lasting consequences. If these functions hold true in humans, this information will provide a step change in our understanding of the pathologies that occur in SRS and potentially other disorders including BWS and IMAGe syndrome, leading to improved diagnosis and the clinical management of patients.
All animal studies and breeding was approved by the University of Cardiff ethical committee and performed under a UK Home Office project license (RMJ). Mice were housed on a 12 hour light–dark cycle with lights coming on at 06.00 hours with a temperature range of 21°C +/- 2 with free access to tap water and standard chow. BAC transgenic lines Cdkn1cBACx1, Cdkn1cBACx2 and Cdkn1c BAC-lacZ, were bred onto a C57BL/6J (BL6) background for >12 generations and genotyped as described [51]. The Cdkn1ctm1Sje allele [21] for historical reasons was maintained on the 129S2/SvHsd (129) background. Cdkn1c-RFLP mice were generated by crossing a M. m. spretus male with a BL6 female and selecting for the Cdkn1c AvaI RFLP for >8 generations. Basal body temperatures of group-housed, experimentally naive female transgenic mice were monitored with a rectal probe (IN005A, Vet Tech solution). Surface temperature of P2 pups was recorded using a thermal imaging camera (Optris P1200). Glucose concentrations in whole blood were determined in neonatal pups in the fed state with the HemoCue system.
β-galactosidase (lacZ) staining, H&E staining and in situ hybridisation were performed as previously described [51]. CDKN1C immunohistochemistry: 10 μm sections were prepared from E16.5 fetuses fixed overnight in 4% PFA at 4°C and paraffin embedded. Slides were dewaxed in xylene and rehydrated through graded ethanols, submerged in 1X Citrate Buffer (DAKO) and heated in a pressure cooker for 20 minutes. Slides were cooled and blocked for 20 minutes in Peroxidase Block (Envsion), then 30 minutes in 10% normal rabbit serum and 1% BSA in PBS, and then incubated in primary antibody (Santa Cruz P57 M-20; SC-1039) overnight at room temperature diluted 1:50 in 10% rabbit serum and 1% BSA in PBS, washed in PBS, incubated with 1:200 dilution of HRP-conjugated rabbit anti-goat IgG secondary (DAKO) for 1 hour at room temperature, washed in PBS 3 x 5 minutes at room temperature and visualized with DAB (DAKO). Slides were counterstained in Mayers Haematoxylin, dehydrated, cleared and mounted in DPX mounting medium. For electron microscopy, rpWATs from P7 mice were fixed overnight in 2% PFA/2% gluteraldehyde in 0.1M Sorensons PB, post fixed in 1% osmium tetraoxide for 2 hours and stained in uranyl acetate overnight at 4°C. After sequential dehydration, samples were embedded in pure araldite and ultra-thin sections were visualised under Philips TEM 208 transmission electron microscope (Phillips). Cryosections were incubated with primary antibodies (1:100 dilution) for 3 hours at room temperature, washed in PBS before incubation with fluorescent secondary antibodies (1:1000 dilution) for one hour at 4°C followed by 4’,6-diamidino-2-phenylindole (DAPI) staining. Slides were mounted using Fluoromount aqueous media (Sigma) and imaged using Leica TCS SP2 AOBS laser confocal microscope, and Leica Confocal software. Samples were scanned with appropriate excitation and emission settings (S2 Table). To identify label-retaining cells in iBAT, we performed two BrdU pulse chase experiments. WT pregnant mice were intraperitoneally administered injections of BrdU at 80 mg/kg/time (Sigma, USA) either once at E16.5 or twice daily from E16.5 for two days. Offspring born from these pregnancies were euthanised 6–8 weeks after the last BrdU injection. iBAT was harvested and cryosections were incubated with the primary antibodies to CDKN1C, PRDM16, BRDU and fluorescent secondary antibodies as described above. Samples were scanned with appropriate excitation and emission settings (S3 Table).
Genomic DNA was bisulphite treated using an EZ DNA Methylation Kit (Zymo Research). Sodium modification treatments were carried out in duplicate for each DNA sample and at least three independent amplification experiments were performed for each individual examined. The region spanning the Cdkn1c was amplified by PCR using primers 5’-tgggtgtagagggtggatttagtta-3’s and 5’- cccacaaaaaccctaccccc-3’ and hemi-nested primer 5’- gtattgttaggattaggatttagttggtagtagtag. The PCR products were cloned into pGEM-T (Promega, Madison, WI, USA) and an average of 20 clones per sample were sequenced using M13 reverse primer and an automated ABI Prism 3130xl Genetic Analyzer (Applied Biosystems, Foster city, CA, USA) as previously described [74]. Quantitative RT-PCR was performed in duplicate on four independent samples obtained from two litters as described [75]. Mitochondrial DNA was quantitated by comparing the nuclear mitochondrial marker cytochrome c, somatic (Cycs) with the mitochondrion-encoded cytochrome c oxidase subunit II (Cox2) by quantitative PCR. Primers are given in S1 Table. RFLP analysis was performed on cDNA prepared from iBAT and rpWAT obtained from crossing a BL6 female with a Cdkn1c-RFLP male. Western blot analysis: total proteins (30 μg) were resolved by SDS-PAGE, transferred to PVDF (Millipore Corp., Bedford, MA), blocked in TBS-T (10 mM Tris, 150 mM NaCl, 0.05% Tween 20, 5% skimmed milk), incubated with primary antibodies (Sigma SAB4500071 CDKN1C; Sigma SAB1300006 PRDM16; Abcam ab10983 UCP1; R&D sytems MAB5966 MYOD; Sigma A5316 β-ACTIN) and visualised using secondary horseradish peroxidase-linked antibodies (Invitrogen) and ECL.
For differentiation experiments, MEFs isolated from E12.5 embryos and cultured in DMEM/F12 (Invitrogen), 10% fetal bovine serum (Invitrogen), 2 mM glutamine (Sigma) and 50 μg/ml penicillin/streptomycin (Sigma) for two passages were used. Differentiation of two-day-post confluent MEFs (D0) was performed by incubation with 170 nM insulin (Sigma), 250 nM dexamethasone (Sigma), 2.5 nM rosiglitazone (Axxora ALX-350-125-M025) and 0.5 mM isobutylmethylxanthine (IBMX) (Sigma) for 2 days and medium containing only 170 nM insulin and 2.5 nM rosiglitazone for 6 additional days, changing the medium every 48 hours. For ORO, cells were fixed for 20 minutes in paraformaldehyde vapour and stained for 15 minutes with Oil Red O solution (0.6% (w/v) in isopropanol:water 60:40), washed and photographed. For UCP1 western blots, cells differentiated for 8 days were harvested, or treated with vehicle (dimethyl sulfoxide) or 9-cis-retinoic acid (1 μM in dimethyl sulfoxide) over 48 hours with protein extraction at intervals. For confocal microscopy, MEFs underwent differentiation in 5 cm glass bottom plates (Mat Tek). After 8 days of differentiation, cells were stained with 5 μg/ml Hoechst 33342 (Invitrogen) and Rhodimine-123 (Sigma Aldrich) for 30 minutes at 37°C. Dyes were removed, and cells were washed for 5 minutes in media. Further staining with 7.5 μg/ ml HCS CellMask Red (Invitrogen) for 10 minutes was performed followed by three washes in ddH20. Samples were imaged using Leica TCS SP2 AOBS laser confocal microscope and Leica Confocal software. HIB-1B cells were maintained in DMEM/F12 (Invitrogen) supplemented with 10% fetal bovine serum (Invitrogen), 2 mM glutamine (Sigma) and 50μg/ml penicillin/streptomycin (Sigma). The siRNA sequence for Cdkn1c-depletion was p57 siRNA (m) (Santa Cruz Biotechnology sc-37621). Control siRNA-A (Santa Cruz Biotechnology sc-37007) was used as the scrambled sequence. Lipid complexes were prepared and reverse transfected according to manufacturer instructions (INTERFERin, Polyplus) in 12-well plates with 10 pmole of the siRNAs complexed with 2 μl of INTERFERin in OPTIMEM with a repeat transfection performed at 24 hours. Cells were harvested 48 hours after transfection and analysed by western blotting. Experiments were performed in three separate occasions in duplicate (ECL) or triplicate (fluorescent) wells.
Statistical significance (Probability values) was determined using the Student’s t-Test (two tailed distribution and two sample unequal variance). For qPCR analysis, Mann-Whitney test was performed on ∆Ct values between groups.
|
10.1371/journal.pmed.1002818 | Iron deficiency, elevated erythropoietin, fibroblast growth factor 23, and mortality in the general population of the Netherlands: A cohort study | Emerging data in chronic kidney disease (CKD) patients suggest that iron deficiency and higher circulating levels of erythropoietin (EPO) stimulate the expression and concomitant cleavage of the osteocyte-derived, phosphate-regulating hormone fibroblast growth factor 23 (FGF23), a risk factor for premature mortality. To date, clinical implications of iron deficiency and high EPO levels in the general population, and the potential downstream role of FGF23, are unclear. Therefore, we aimed to determine the associations between iron deficiency and higher EPO levels with mortality, and the potential mediating role of FGF23, in a cohort of community-dwelling subjects.
We analyzed 6,544 community-dwelling subjects (age 53 ± 12 years; 50% males) who participated in the Prevention of Renal and Vascular End-Stage Disease (PREVEND) study—a prospective population-based cohort study, of which we used the second survey (2001–2003)—and follow-up was performed for a median of 8 years. We measured circulating parameters of iron status, EPO levels, and plasma total FGF23 levels. Our primary outcome was all-cause mortality. In multivariable linear regression analyses, ferritin (ß = –0.43), transferrin saturation (TSAT) (ß = −0.17), hepcidin (ß = −0.36), soluble transferrin receptor (sTfR; ß = 0.33), and EPO (ß = 0.28) were associated with FGF23 level, independent of potential confounders. During median (interquartile range [IQR]) follow-up of 8.2 (7.7–8.8) years, 379 (6%) subjects died. In multivariable Cox regression analyses, lower levels of TSAT (hazard ratio [HR] per 1 standard deviation [SD], 0.84; 95% confidence interval [CI], 0.75–0.95; P = 0.004) and higher levels of sTfR (HR, 1.15; 95% CI 1.03–1.28; P = 0.01), EPO (HR, 1.17; 95% CI 1.05–1.29; P = 0.004), and FGF23 (HR, 1.20; 95% CI 1.10–1.32; P < 0.001) were each significantly associated with an increased risk of death, independent of potential confounders. Adjustment for FGF23 levels markedly attenuated the associations of TSAT (HR, 0.89; 95% CI 0.78–1.01; P = 0.06), sTfR (HR, 1.08; 95% CI 0.96–1.20; P = 0.19), and EPO (HR, 1.10; 95% CI 0.99–1.22; P = 0.08) with mortality. FGF23 remained associated with mortality (HR, 1.15; 95% CI 1.04–1.27; P = 0.008) after adjustment for TSAT, sTfR, and EPO levels. Mediation analysis indicated that FGF23 explained 31% of the association between TSAT and mortality; similarly, FGF23 explained 32% of the association between sTfR and mortality and 48% of the association between EPO and mortality (indirect effect P < 0.05 for all analyses). The main limitations of this study were the observational study design and the absence of data on intact FGF23 (iFGF23), precluding us from discerning whether the current results are attributable to an increase in iFGF23 or in C-terminal FGF23 fragments.
In this study, we found that functional iron deficiency and higher EPO levels were each associated with an increased risk of death in the general population. Our findings suggest that FGF23 could be involved in the association between functional iron deficiency and increased EPO levels and death. Investigation of strategies aimed at correcting iron deficiency and reducing FGF23 levels is warranted.
| Iron deficiency, one of the most common nutritional disorders worldwide, is known to be associated with increased risk of death in the general population, although a clear mechanism has not been identified.
In patient populations, it has been shown that higher levels of erythropoietin (EPO) are associated with adverse outcomes, but its significance in the general population has not been elucidated yet.
Emerging data from the chronic kidney disease (CKD) field suggest that both iron deficiency and EPO lead to an up-regulated expression and concomitant cleavage of a phosphate-regulating hormone, fibroblast growth factor 23 (FGF23), another risk factor for mortality.
We used the Prevention of Renal and Vascular End-Stage Disease (PREVEND) study, a prospective population-based cohort study from 2001 to 2003 involving 6,588 subjects and for which follow-up has been performed for a median of 8 years.
We found that iron status parameters and EPO were the strongest determinants of FGF23 in the general population, even more than such well-described determinants as phosphate, kidney function, and calcium levels.
Functional iron deficiency and EPO were associated with an increased risk of death in the general population, and FGF23 explained a substantial part of these associations.
The current findings suggest that FGF23 is important in the pathophysiology of iron-deficiency–and EPO-mediated mortality in the population.
This study provides the rationale to assess the potential benefits of iron supplementation to reduce mortality in the general population, possibly by lowering FGF23 levels.
| Iron deficiency is one of the most common nutritional disorders worldwide [1]. In addition to its effect on quality of life and functional capacity, iron deficiency has previously been associated with an increased risk of all-cause and cardiovascular mortality in the general population [2]. However, the underlying mechanisms linking iron deficiency with mortality in this setting remain unknown. Emerging data from disease populations such as chronic kidney disease (CKD) patients suggest that iron deficiency stimulates the expression and concomitant cleavage of the osteocyte-derived, phosphate-regulating hormone fibroblast growth factor 23 (FGF23). Elevated levels of FGF23, in turn, have been shown to be associated with an increased risk of mortality in both community-dwelling individuals [3] and across stages of CKD [4–6].
Similarly, higher circulating endogenous erythropoietin (EPO) levels have been associated with an increased all-cause and cardiovascular mortality risk in various disease populations, including chronic heart failure patients, kidney transplant recipients, and elderly individuals [7–10]. In addition, higher doses of exogenous EPO increase the risk of cardiovascular events in CKD patients [11–13]. To date, it is unknown whether EPO levels are associated with adverse outcomes in the general population. Similar to iron deficiency, recent studies from our group and others have established that both endogenous and exogenous EPO may influence FGF23 production and metabolism in CKD patients [14–16].
Currently, the clinical implications of iron deficiency and high EPO levels in the general population, as well as the potential downstream role for FGF23, are unclear. Hence, in the current study, we investigated whether iron deficiency and EPO influence FGF23 levels, whether iron deficiency and elevated EPO levels are associated with an increased risk of death, and whether such associations could be explained by variation in FGF23 levels.
Details from the Prevention of Renal and Vascular End-Stage Disease (PREVEND) study have been described previously [17]. This study was planned following prior results indicating that iron deficiency and high EPO levels are associated with higher FGF23 levels and subsequently contribute to an increased mortality risk in kidney transplant recipients [16,18]. Between January and September 2018, we addressed the hypothesis that similar relationships would exist in the general population. For this study, we did not have a prespecified analysis plan, but we performed hypothesis-driven analyses in which no data-driven changes have taken place. This study is reported as per the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) guideline (S1 Checklist). In brief, from 1997 to 1998, all inhabitants of the city of Groningen, the Netherlands, who were 28 to 75 years old received a questionnaire on demographics, disease history, smoking habits, use of medication, and a vial to collect an early morning urinary sample (n = 85,421). Of these individuals, 40,856 (47.8%) responded. After exclusion of subjects with insulin-dependent diabetes mellitus and pregnant women, subjects with a urinary albumin concentration ≥10 mg/L (n = 6,000) and a randomly selected control group with a urinary albumin excretion <10 mg/L (n = 2,592) completed the screening protocol and, as such, formed the baseline PREVEND cohort (n = 8,592). For the current analyses, we used data from the second survey, which took place between 2001 and 2003 (n = 6,894), since for this visit, blood samples were also available. We excluded 436 subjects due to missing data on FGF23, resulting in 6,544 subjects (visually depicted in a flowchart in S1 Fig). The PREVEND study protocol was approved by the institutional medical review board and was in accordance with the Declaration of Helsinki. Written informed consent was obtained from all subjects.
Fasting blood samples were drawn in the morning from all subjects from April 24, 2001, to December 3, 2003. All hematologic measurements were measured in fresh venous blood. Aliquots of these samples were stored immediately at −80°C until further analysis. Serum creatinine was measured using an enzymatic method on a Roche Modular analyzer (Roche Diagnostics, Mannheim, Germany). For predicting the estimated glomerular filtration rate (eGFR), the Chronic Kidney Disease Epidemiology Collaboration (CKD-EPI) was applied [19]. Serum iron was measured using a colorimetric assay, ferritin using immunoassay, and transferrin using an immunoturbidimetric assay (all Roche Diagnostics). Transferrin saturation (TSAT, %) was calculated as 100 × serum iron (μmol/L) ÷ 25 × transferrin (g/L) [20]. Serum EPO was measured using an immunoassay based on chemiluminescence (Immulite EPO assay, DPC, Los Angeles, CA). Serum hepcidin was measured with a competitive enzyme-linked immunosorbent assay (ELISA), as described elsewhere with intra- and interassay coefficients of variation (CVs) of 8.6% and 16.2%, respectively [21]. Soluble transferrin receptor (sTfR) was measured using an automated homogenous immunoturbidimetric assay with intra- and interassay CVs <2% and <5% [22]. Total FGF23 levels were measured in plasma EDTA samples with a human FGF23 ELISA (Quidel Corp., San Diego, CA) directed against 2 different epitopes within the C-terminal part of the FGF23 molecule. Hence, the assay measures both the intact hormone and the C-terminal fragments and, as such, measures total FGF23 levels. In our hands, this ELISA had intra- and interassay CVs of <5% and <16% in blinded replicated samples, respectively [23]. We assessed prospective associations of iron status parameters, EPO, hepcidin, and FGF23 levels with death. In the PREVEND cohort, data on mortality were received through the municipal register, and follow-up was available until January 1, 2011.
Data were analyzed using IBM SPSS software, version 23.0 (SPSS, Chicago, IL), R version 3.2.3 (Vienna, Austria), and STATA 14.1 (STATA Corp., College Station, TX). Baseline characteristics are described as means with standard deviation (SD) for normally distributed variables, as medians with interquartile range (IQR) for skewed variables, or as numbers with corresponding percentages for categorical variables. Differences in baseline characteristics across tertiles of FGF23 were tested with a one-way ANOVA for normally distributed variables, Kruskal-Wallis test for skewed variables, and a chi-squared test for categorical variables. Hereafter, we assessed the relationship between iron status parameters and EPO as determinants of FGF23 by calculating Spearman correlation coefficients. We subsequently performed multivariable analyses adjusted for age, sex, eGFR, urinary albumin excretion, systolic blood pressure, alcohol use (5 categories), smoking status (never, former, or current smoker), hemoglobin, mean corpuscular volume (MCV), and plasma/serum levels of high-sensitivity C-reactive protein (hs-CRP), phosphate, calcium, 25-hydroxyvitamin D (25D), and parathyroid hormone (PTH). In all analyses, skewed variables, i.e., FGF23, ferritin, hepcidin, sTfR, EPO, urinary albumin excretion, PTH, and hs-CRP, were naturally log transformed prior to inclusion. Second, we performed stepwise backward multivariable linear regression analyses in which we determined whether iron status parameters and EPO remained major determinants of FGF23 levels. Of note, because the different iron status parameters are highly correlated with each other (S1 Table), we included all iron status parameters individually in multivariable linear regression analyses. To visualize the cross-sectional associations between the different iron status parameters—EPO and FGF23—plots were generated using locally weighted scatterplot smoothing.
We subsequently assessed the associations of iron status parameters and EPO levels with all-cause mortality using Cox proportional hazard regression analyses. Assumption of proportionality in Cox regression analyses was checked using Schoenfeld residuals plots. We constructed multivariable models adjusted for age and sex (model 1), and additionally for eGFR, urinary albumin excretion, body mass index (BMI), systolic blood pressure, hs-CRP, presence of diabetes, smoking, alcohol use, and use of antihypertensives (model 2). Subsequently, we adjusted for FGF23 (model 3). Similarly, we assessed whether FGF23 levels were associated with mortality, independent of potential confounders, including adjustment for iron status parameters and EPO in the final models. Finally, we performed mediation analyses with the methods described by Preacher and Hayes, which is based on logistic regression. These analyses allow for testing significance and magnitude of mediation [24,25]. Overall, 3.4% of demographic data were missing and were imputed using regressive switching [26]. Five datasets were multiply imputed, and results were pooled according to Rubin’s rules [27]. In all analyses, a two-sided significance level <0.05 was considered significant.
We included 6,544 community-dwelling subjects (age 53 ± 12 years; 50% males) with a median (IQR) FGF23 level of 70 (57–87) RU/mL. Mean hemoglobin concentration was 13.7 ± 1.2 g/dL; median ferritin concentration was 96 (47–172) μg/L; mean TSAT was 25.0 ± 9.5%; median sTfR level was 2.5 (2.1–3.0) mg/L; median hepcidin level was 8.5 (4.6–13.8) ng/L; and median EPO level was 7.8 (5.9–10.3) IU/L. Iron deficiency based on ferritin levels (i.e., ferritin < 15 μg/L for women and < 30 μg/L for men) was present in 448 (7%) individuals [28]. Iron deficiency based on low TSAT levels (i.e., TSAT < 20%) was present in 1,806 (28%) individuals [2]. EPO-stimulating agents (ESAs) and oral or intravenous iron were not used by any of the subjects. Across tertiles of FGF23, we observed inverse associations with ferritin, TSAT, and hepcidin and a positive association with sTfR and EPO levels. Additional demographic, clinical, and laboratory parameters are depicted in Table 1.
In univariate analyses, higher FGF23 levels correlated with lower levels of ferritin (ρ = –0.21, P < 0.001), TSAT (ρ = –0.21, P < 0.001), sTfR (ρ = 0.20, P < 0.001), and hepcidin (ρ = –0.18, P < 0.001) and with higher levels of serum EPO (ρ = 0.19, P < 0.001) (Fig 1A–1E). After adjustment for age, sex, eGFR, urinary albumin excretion, systolic blood pressure, alcohol use, smoking status, hemoglobin, MCV, hs-CRP, phosphate, calcium, 25D, and PTH, all iron parameters—including ferritin (ß = –0.43, P < 0.001), TSAT (ß = –0.17, P < 0.001), sTfR (ß = 0.33, P < 0.001), and hepcidin (ß = –0.36, P < 0.001)—remained strongly and independently associated with FGF23 levels. Similarly, EPO (ß = 0.28, P < 0.001) remained independently associated with FGF23. In stepwise backward linear regression analyses, ferritin (ß = –0.38, P < 0.001), sTfR (ß = 0.27, P < 0.001), hepcidin (ß = –0.31, P < 0.001), and EPO (ß = 0.21, P < 0.001) levels were identified as the strongest determinants of FGF23 levels, with higher standardized regression coefficients than more established determinants of FGF23, including eGFR (ß = –0.20, P < 0.001), phosphate (ß = 0.13, P < 0.001), and calcium (ß = 0.17, P < 0.001) (S2 Table).
During a median (IQR) follow-up of 8.2 (7.7–8.8) years, 379 (6%) subjects died. After 1, 5, and 8 years of follow-up, 6,437, 6,040, and 3,993 subjects, respectively, were still followed up in the study. We assessed the associations of the individual iron status parameters, EPO and FGF23, with mortality (Table 2). After adjustment for age and sex, neither ferritin nor hepcidin was associated with risk of death (hazard ratio [HR] per 1 SD higher ln[ferritin], 0.95; 95% confidence interval [CI], 0.84–1.07; P = 0.38; HR per 1 SD higher ln[hepcidin], 0.99; 95% CI 0.88–1.11; P = 0.80, respectively). In contrast, a lower TSAT was strongly associated with an increased risk of death (age- and sex-adjusted HR per 1 SD higher TSAT, 0.79; 95% CI 0.70–0.88; P < 0.001). After further adjustment for eGFR, urinary albumin excretion, BMI, systolic blood pressure, hs-CRP, presence of diabetes, smoking, alcohol use, and use of antihypertensives (model 2), the association between TSAT and mortality persisted (HR, 0.84; 95% CI 0.75–0.95; P = 0.004). However, subsequent adjustment for FGF23 attenuated the association, such that TSAT was no longer significantly associated with death (HR, 0.89; 95% CI 0.78–1.01; P = 0.06).
Similarly, a higher plasma sTfR level, indicating functional iron deficiency, was associated with a higher risk of mortality in a model adjusted for age and sex (HR per 1 SD higher ln[sTfR], 1.17; 95% CI 1.06–1.30; P = 0.004). In multivariable analyses (model 2), the association between sTfR and mortality remained significant (HR, 1.15; 95% CI 1.03–1.28; P = 0.01). However, adjustment for FGF23 strongly attenuated the association, rendering the association between sTfR and mortality nonsignificant (HR, 1.08; 95% CI 0.96–1.20; P = 0.19).
In mediation analyses, we analyzed whether the significant associations between iron status parameters (i.e., TSAT and sTfR) and mortality were mediated by FGF23 (Table 3). FGF23 was identified as a significant mediator (indirect effect P < 0.05; 31% of the association between TSAT and mortality was explained by FGF23 and 32% of the association between sTfR and mortality).
In age- and sex-adjusted analyses, higher serum EPO levels were associated with an increased risk of death (HR per 1 SD higher ln[EPO], 1.22; 95% CI 1.10–1.34; P < 0.001, Table 2). In multivariable analyses (model 2), the association between EPO and mortality persisted (HR, 1.17; 95% CI 1.05–1.29; P = 0.004). However, adjustment for FGF23 abrogated the association between EPO and mortality, rendering the association nonsignificant (HR, 1.10; 95% CI 0.99–1.22; P = 0.08).
In mediation analyses, we analyzed whether the significant association between EPO and mortality was mediated by FGF23 (Table 3). FGF23 was identified as a significant mediator (indirect effect P < 0.05; 48% of the association between EPO and mortality was explained by FGF23). Because functional iron deficiency often occurs in states of EPO-mediated erythropoiesis, we also analyzed whether the positive association between EPO and FGF23 might be, at least in part, mediated by TSAT and sTfR. Both parameters were found to be significant mediators in the positive association between EPO and FGF23 (indirect effect P < 0.05; 12% by TSAT and 33% by sTfR, independent of potential confounders, Table 3).
In age- and sex-adjusted analyses, higher plasma FGF23 levels were strongly associated with an increased risk of death (HR per 1 SD higher ln[FGF23], 1.29; 95% CI 1.20–1.34; P < 0.001, Table 2). In multivariable analyses (model 2) with inclusion of bone mineral parameters (i.e., calcium, phosphate, 25D, and PTH), the association between FGF23 and mortality remained (HR, 1.20; 95% CI 1.10–1.32; P < 0.001, Table 2). Further adjustment for TSAT, sTfR, and EPO did not materially change the association between FGF23 and mortality (HR, 1.15; 95% CI 1.04–1.27; P = 0.008).
In this study, we found that markers of iron deficiency, especially lower iron availability (i.e., lower TSAT and higher sTfR), as well as elevated serum EPO, were associated with an increased risk of mortality in community-dwelling individuals. Notably, we identified FGF23 as a potential mediator of iron-deficiency–and EPO-related mortality. Iron deficiency and elevated levels of EPO were major determinants of FGF23 levels, to a greater extent than more established determinants such as renal function and serum calcium, PTH, and phosphate. FGF23 in itself was strongly associated with mortality independent of adjustment for iron status parameters and EPO. Our findings suggest that FGF23 could be involved in the pathophysiology of iron-deficiency–and EPO-mediated mortality in the population.
We first addressed the relationship between iron deficiency, measured by 4 different parameters, and all-cause mortality in the general population. Our results are consistent with a previous study identifying low TSAT as a predictor of mortality in the general population [2]. Of note, the prospective associations with mortality differed among the various iron parameters. While TSAT and sTfR were strongly associated with mortality, ferritin and hepcidin were not. This discrepancy is most likely explained by the fact that these markers reflect different aspects of iron metabolism. Serum ferritin is a surrogate for body iron stores, but as an acute-phase reactant, it is also strongly up-regulated by inflammation, malignancy, and alcohol intake. Hepcidin is also an acute-phase reactant and is highly correlated with serum ferritin. In contrast, TSAT is more a marker of iron availability for erythropoiesis. Elevated levels of sTfR reflect an increased tissue iron demand, but not body iron stores, and are less affected by concomitant chronic disease and inflammation [29]. In the setting of increased metabolic requirements for iron, transferrin receptors are overexpressed on erythroid precursors in the bone marrow and are shed, resulting in increased sTfR levels in the circulation. Hence, an increased sTfR level reflects both erythroid activity and functional iron deficiency. Because functional iron deficiency occurs in patients with significant EPO-mediated erythropoiesis or as a response to treatment with ESAs [30], we also aimed to assess the association between endogenous EPO levels, as a reflection of tissue hypoxia, and mortality. Prior studies conducted in various populations, including in elderly individuals, kidney transplant recipients, and in patients with chronic heart failure, found that higher EPO levels are associated with an increased risk of death, even independent of hemoglobin levels [8–10]. Furthermore, large randomized trials in chronic heart failure and CKD patients striving for stronger correction of anemia with ESAs were associated with an increased risk of mortality [12,13,31,32]. In the current study, we identified for the first time a strong and independent association between higher serum EPO levels and increased mortality in community-dwelling individuals.
The associations we observed between functional iron deficiency, high EPO levels, and mortality led us to explore FGF23 as a potential downstream factor mediating these associations, given accumulating evidence supporting a direct relationship between iron status, EPO, and FGF23 metabolism [16,18,33–35]. Recently, our group and others demonstrated that iron deficiency is a strong determinant of total FGF23 levels in CKD and kidney transplant recipients [18,36]. Mechanistically, it has recently been shown that iron deficiency stabilizes hypoxia-inducible factor 1-alpha, which in turn up-regulates furin, promoting cleavage of the intact FGF23 (iFGF23) molecule into C-terminal FGF23 fragments [34,37,38]. Furthermore, EPO-induced up-regulation of FGF23 production has been demonstrated in CKD patients, kidney transplant recipients, and in animal models in which circulating EPO levels are elevated due to either endogenous or exogenous sources [14,33]. Currently, the exact mechanism by which EPO increases bone and bone marrow FGF23 transcription and FGF23 post-translational cleavage is unknown; however, Rabadi and colleagues observed in bled mice decreased polypeptide N-acetylgalactosaminyltransferase 3 (GalNT3) bone marrow mRNA expression, which protects iFGF23 from proteolysis by furin, allowing increased FGF23 cleavage [14]. Our group, together with collaborators, recently also showed that in mice with chronically elevated endogenous EPO levels, a decreased GalNT3 bone marrow mRNA expression was present, without differences in family with sequence similarity 20, member C (Fam20C) or furin expression [16]. In the current study, multivariable cross-sectional analyses also demonstrated strong associations of iron parameters and EPO with FGF23, independent of more established determinants of FGF23, including serum phosphate and eGFR.
We subsequently found that the associations between functional iron deficiency—reflected by low TSAT or high sTfR levels—and mortality were mediated by FGF23. Moreover, EPO-related mortality was also for a considerable part explained by variation in FGF23 levels. Of interest, the positive association between EPO and FGF23 was in part mediated by functional iron deficiency. These findings support our hypothesis that FGF23 is closely related to erythropoiesis and that up-regulation of FGF23 induced by iron deficiency or high EPO levels may subsequently lead to a higher mortality risk. The association between FGF23 and death has been previously demonstrated in different patient populations, including CKD, renal transplant recipients, acute kidney injury, chronic heart failure, and in the general population [3,4,39,40]. The downstream consequences of elevated levels of FGF23 have not been fully elucidated yet. Many reports have revealed that iFGF23 has biologic activity through binding to several FGF23 receptors, including FGFR1, FGFR2, and FGFR4. Besides the classic functions of iFGF23 in regulating renal phosphate handling and vitamin D metabolism, recent studies have demonstrated several “off-target” effects of iFGF23. Preclinical studies demonstrated that FGF23 can induce left ventricular hypertrophy by binding to FGF23 receptor 4 in cardiac myocytes, and promote endothelial dysfunction [41,42]. Furthermore, FGF23 stimulates fibrosis in the kidneys [43], exerts proinflammatory effects by up-regulation of interleukin-6 production [44], and impairs immune function [45]. Finally, it is known that abnormal FGF signaling can promote tumor development by stimulating cancer cell proliferation, invasion, and survival and by supporting tumor angiogenesis [46,47]. Feng and colleagues showed that exogenous FGF23 promotes prostate cancer proliferation, invasion, and independent growth in vitro, whereas FGF23 knockdown in model systems decreased tumor progression both in vitro and in vivo [48]. Given the strong and independent association of FGF23 with mortality and the emerging pathophysiological implications of FGF23, it seems important to unravel major determinants of FGF23 in order to be able to reduce FGF23 levels. In the current study, we have shown that, in the general population, iron deficiency is a major determinant of FGF23 levels, implicating that iron deficiency could potentially be an easily modifiable driver of high FGF23 levels. The potential benefits of iron supplementation to reduce mortality in the general population remain to be addressed in prospective studies.
Our study has several strengths as well as limitations. Major strengths include the availability of FGF23 along with multiple iron status parameters (including hepcidin and sTfR) and EPO in a large population-based cohort. On the other hand, we were unable to measure iFGF23 levels, because samples were not stored with protease inhibitors, and iFGF23 has been shown to be susceptible to degradation with long-term storage [49]. This precludes us from discerning whether the elevated levels of total FGF23 that we observed are attributable to increased circulating levels of intact, biologically active FGF23 or due to increased levels of C-terminal fragments, which are not biologically active. The latter could be due to increased production of FGF23 matched by a concomitant increase in FGF23 cleavage. This pattern has been observed in previous studies in various disease populations, and we speculate that the same pattern might occur in the general population as well. Furthermore, due to the observational design, residual confounding might have biased the results despite the substantial number of potentially confounding factors for which we adjusted. For example, it is known that TSAT is also influenced by inflammation (similar to ferritin and hepcidin); it might be that, despite adjustment for hs-CRP, inflammation has influenced this specific association. Finally, although we have strong evidence from the literature that FGF23 is a mediator in the association of iron deficiency and EPO with mortality, we cannot exclude that an unmeasured cause of mortality or alternative potential mediators have influenced the current results.
In conclusion, we have shown that functional iron deficiency and higher EPO levels are strongly associated with increased all-cause mortality in the general population. Furthermore, we demonstrated that the increased mortality risk in individuals with diminished iron availability or increased levels of EPO seems to be substantially attributable to variation in FGF23 levels. Future studies will be needed to delineate in more detail the underlying mechanism for the currently identified associations.
|
10.1371/journal.pcbi.1005568 | On the effects of alternative optima in context-specific metabolic model predictions | The integration of experimental data into genome-scale metabolic models can greatly improve flux predictions. This is achieved by restricting predictions to a more realistic context-specific domain, like a particular cell or tissue type. Several computational approaches to integrate data have been proposed—generally obtaining context-specific (sub)models or flux distributions. However, these approaches may lead to a multitude of equally valid but potentially different models or flux distributions, due to possible alternative optima in the underlying optimization problems. Although this issue introduces ambiguity in context-specific predictions, it has not been generally recognized, especially in the case of model reconstructions. In this study, we analyze the impact of alternative optima in four state-of-the-art context-specific data integration approaches, providing both flux distributions and/or metabolic models. To this end, we present three computational methods and apply them to two particular case studies: leaf-specific predictions from the integration of gene expression data in a metabolic model of Arabidopsis thaliana, and liver-specific reconstructions derived from a human model with various experimental data sources. The application of these methods allows us to obtain the following results: (i) we sample the space of alternative flux distributions in the leaf- and the liver-specific case and quantify the ambiguity of the predictions. In addition, we show how the inclusion of ℓ1-regularization during data integration reduces the ambiguity in both cases. (ii) We generate sets of alternative leaf- and liver-specific models that are optimal to each one of the evaluated model reconstruction approaches. We demonstrate that alternative models of the same context contain a marked fraction of disparate reactions. Further, we show that a careful balance between model sparsity and metabolic functionality helps in reducing the discrepancies between alternative models. Finally, our findings indicate that alternative optima must be taken into account for rendering the context-specific metabolic model predictions less ambiguous.
| Recent methodological developments have facilitated the integration of high-throughput data into genome-scale models to obtain context-specific metabolic reconstructions. A unique solution to this data integration problem often may not be guaranteed, leading to a multitude of context-specific predictions equally concordant with the integrated data. Yet, little attention has been paid to the alternative optima resulting from the integration of context-specific data. Here we present computational approaches to analyze alternative optima for different context-specific data integration instances. By using these approaches on metabolic reconstructions for the leaf of Arabidopsis thaliana and the human liver, we show that the analysis of alternative optima is key to adequately evaluating the specificity of the predictions in particular cellular contexts. While we provide several ways to reduce the ambiguity in the context-specific predictions, our findings indicate that the existence of alternative optimal solutions warrant caution in detailed context-specific analyses of metabolism.
| Genome-scale metabolic models (GEMs) have proven instrumental in characterizing the activity of metabolic pathways in different biological scenarios. The activity of all metabolic reactions is specified by the flux distribution, which can be readily inferred from GEMs through the usage of constraint-based approaches [1,2]. Such approaches often infer fluxes as solutions to a convex optimization problem in which an objective function is optimized under specified constraints. Two types of constraints can generally be considered: The first is due to the stoichiometry, thermodynamic viability (i.e., if a reaction is irreversible or reversible under normal physiological conditions) and mass-balance conditions. These constraints are included in every constraint-based approach. The second type comprises constraints specific to each approach, and usually reflects the context-specific knowledge or data to be integrated. Flux distributions which satisfy the set of constraints are called feasible. A convex optimization problem is guaranteed to render a unique optimal value [3]. However, it is not always guaranteed that there is a unique flux distribution realizing the optimal objective value, leading to alternative optimal flux distributions. Indeed, such a space of alternative optima arises even in the case of flux balance analysis (FBA), as a classical representative of constraint-based approaches [4–9].
Experimental systems biology studies have generated comprehensive atlases of transcript, protein, and metabolite levels from different context, such as: cell types, developmental stages, and environments, across different species from all kingdoms of life [10–15]. Analyses of these data sets have already pointed that context-specific differences in the levels of molecular components often affect the activity of metabolic pathways. Additionally GEMs allow constraint-based approaches to integrate such data sets through the so-called gene-protein-reaction rules, which relate metabolic reactions with the enzymes involved and their coding genes [16–19]. These approaches address two aims: (i) obtaining context-specific flux distribution and (ii) determining context-specific GEMs; we refer to the respective approaches as flux- and network-centered, respectively. Alternative optima may also result from the integration of context-specific data. In both settings, the existence of alternative optima leads to ambiguity in context-specific flux distributions and/or network reconstructions, since alternative solutions may substantially differ. This is particularly important in the case of context-specific network reconstructions, where further investigations conducted on a single optimal network could lead to erroneous conclusions.
To our knowledge, only three studies considered the space of alternative optimal solutions arising from flux-centered approaches: The approach termed iMAT [20] proposed a procedure to classify the flux state of reactions into active, inactive or uncertain across the alternative optima space. Another approach, abbreviated as EXAMO [21], later used the set of active reactions obtained from the iMAT alternative optima space as input to the approach referred to as MBA [22], a network-centered method, to reconstruct a context-specific network. Additionally, the Flux Variability Sampling [23] was used to sample the alternative space of flux values that are equidistant to the data integrated. Finally, we note that alternative optimal context-specific models have not been recognized in the case of network-centered approaches, and currently, there is no available method for their analysis.
In the present study, we propose a method to quantify the variability of alternative optimal flux values of a flux-centered approach. Additionally, we quantify the effect in the alternative optima of including an additional constraint in the flux values, minimize the total sum of absolute flux values, which has been proposed to obtain unique solutions in a flux-centered method [24]. Furthermore, we investigate, for the first time, the space of alternative optimal context-specific models that arise from several network-centered approaches, and analyze the potential impact on further metabolic predictions and biological conclusions drawn. The study is organized in two parts. The first part is dedicated to explaining the mathematical and computational logic of both (i) the context-specific data integration approaches herein evaluated, and (ii) the methods that we propose to analyze the respective alternative optima. The second part presents the findings obtained from applying the previously described methods to two particular case studies: a leaf-specific reconstruction from the model plant Arabidopsis thaliana, and a human liver reconstruction. This second part serves as an illustration of the impact that alternative optima have in context-specific metabolic reconstructions, and may be followed independently from the first part—which is primary addressed to the specialized reader.
In this section, we present the mathematical formulation of the computational methods that we developed to investigate the alternative optima of three selected data integration approaches. In all three cases, we first provide an overview of the approach, which is followed by a description of the method to explore its alternative optima space. We start by a representative of a flux-centered approach—a modified version of RegrEx [25]—and the method that we propose to explore its alternative optima, termed RegrEx Alternative Optima Sampling (RegrExAOS). We then focus on Core Expansion (CorEx), also developed in this study, which we take as representative of a network-centered approach. In addition, we show how the optimization program behind CorEx can be adapted to evaluate not only its alternative optima space, but that of FastCORE [26] and CORDA [27], two state-of-the-art network-centered approaches.
Given a GEM and (context-specific) gene or protein expression data, the Regularized metabolic model Extraction (RegrEx) method reconstructs a context-specific metabolic model, along with the corresponding flux distribution. To this end, RegrEx finds a feasible flux distribution that is closest to a given experimental data set, and is, therefore, considered a flux-centered approach.
The original RegrEx approach relied on a regularized least squares optimization in which the Euclidean distance between the given gene expression data vector, d, and a feasible flux distribution, v, i.e., the squared ℓ2 norm of the difference vector ϵ = d – v, was minimized [25]. The regularization was implemented by also considering the (weighted) ℓ1 norm of v in the minimization problem, as a means to select the reactions in the GEM that are most important for a given metabolic context. However, here we used a slightly modified version of RegrEx: Instead of minimizing the sum of square errors, we minimize the sum of absolute errors, i.e., the ℓ1 norm of ϵ. Except for this substitution, the modified RegrEx version, called RegrExLAD (for Least Absolute Deviations), follows the same formulation as the original RegrEx (see S1 Appendix for detailed comparison).
The minimization problem behind RegrExLAD considers a set of constraints required to handle reversible reactions: In this case, absolute flux values must be taken into account when minimizing the distance to the (non-negative) associated gene expression (i.e., for a reversible reaction i, ϵi = |vi|–di). This is accomplished by splitting reversible reactions into the forward and backward directions, each constrained to have non-negative flux value, and introducing a vector of binary variables, x, to select only one of them during the optimization. Altogether, these particularities are captured in the mixed integer linear program (MILP),
vopt=argminϵ+=[ϵirr+;ϵfor+;ϵback+],ϵ−=[ϵirr−;ϵfor−;ϵback−],v=[virr;vfor;vback]∈ℝ0+,x∈{0,1}nwT(ϵ++ϵ−)+λ||v||1s.t.1. Sextv=02. virr i+(ϵirr+−ϵirr−)=dirr3. vfor i+(ϵfor+−ϵfor−)+xdrevRxns=drevRxn4. vrev i+(ϵback+−ϵback−)−xdrevRxn=0}, i∈RD 5. virrmin≤virr≤virrmax6. vfor+xvformin≥vformin7. vback−xvrevmin≥08. vfor+xvformax≤vformax9. vback−xvrevmax≤0
(OP1)
In OP1, the flux distribution, v, is partitioned into the sets of irreversible (virr), and reversible reactions proceeding into the forward (vfor) and backward directions (vback), and the (reaction) columns of the stoichiometric matrix, Sext, are ordered to match the partition of v. In addition, the components of the error vector, ϵi = ϵ+i−ϵ–i, ϵ+i, ϵ–i ≥ 0, are split into two non-negative variables, ϵ+i, ϵ–i, as a way to computationally treat the otherwise required absolute values of the components of ϵ. Thus, the ℓ1 norm ||ϵ||1 = Σi |ϵi| is replaced by ϵ+i + ϵ–i in the objective function. (ϵ is defined only over the set of reactions with associated data, RD in OP1). Finally, the λ parameter corresponds to the weight of the ℓ1 norm in the objective function, and is chosen during the optimization as to maximize the Pearson correlation between data and flux values [25].
The convexity of OP1 guarantees finding the minimum distance between experimental data and a feasible flux distribution that is allowed by the constraints. However, it does not guarantee that the resulting flux distribution is the only feasible one that is optimal with respect to a particular context-specific data. This variability in optimal flux distributions may be attributed to two factors. On the one hand, as mentioned above, not all reactions in a GEM are typically associated to data. In contrast to data-bounded reactions, there is a set of data-orphan reactions comprising non-enzymatically catalyzed reactions, reactions without gene-protein annotation or without associated data for a particular context. Data-orphan reactions do not contribute to the error norm in the RegrExLAD objective function, described in OP1, and their flux value can vary as long as v satisfies the imposed constraints and its ℓ1 norm is preserved. This situation is depicted in Fig 1, where the search for a flux distribution v that is closest to the data vector, d, is carried out in the projection of the flux cone, F = {v: Sv = 0, vmin ≤ v ≤ vmax}, where d resides. On the other hand, the geometry of F may preclude certain reactions to obtain an exact match with the data value, when d remains outside the projection of F. In this case, a set of flux distributions may be equidistant to d, thus generating variability also in the optimal flux value of data-bounded reactions.
The general approach followed by RegrExAOS, depicted in Fig 2, is similar to the Flux Variability Sampling [23] (here adapted to RegrExLAD, see S1 Appendix). RegrExAOS first creates a random flux vector, vrand, which is bounded by the maximum and minimum flux values previously calculated by Flux Variability Analysis (using only upper and lower bounds as constraints, see Methods). It then searches for the closest flux vector, v, to vrand that belongs to the alternative optima space, i.e., it is at the same distance to the data vector, d, and has the same ℓ1 norm as the previously calculated RegrExLAD optimum. This is performed by solving the MILP given in OP2:
minϵ+=[ϵirr+;ϵfor+;ϵback+],ϵ−=[ϵirr−;ϵfor−;ϵback−],δ+=[δirr+;δfor+],δ−=[δirr−;δfor−],v=[virr;vfor;vback]∈ℝ0+,x∈{0,1}n||δ++δ−+δback||1s.t.1−9 (OP1)10. ϵ++ϵ−=ϵopt++ϵopt−11. ||v||1=||vopt||112. virr−(δirr+−δirr−)=vrand(irr)13. vfor−(δfor+−δfor−)−xvrand(revRxn)=014. −vback+δback+xvrand(revRxn)=vrand(revRxn)
(OP2)
Finally, RegrExAOS iterates this routine n times to obtain a sufficiently large sample; here we used n = 2000, which is sufficient sample size for the subsequent statistical analyses.
OP2 inherits constraints 1–9 from OP1 and incorporates two sets of new constraints. Constraints 10 and 11 are added to guarantee that v renders the same similarity to data and the same ℓ1 norm of the previously found RegrExLAD optimum, vopt, respectively. In addition, constraints 12–14 introduce the auxiliary variables δirr, δfor and δback quantifying the distance of an optimal flux distribution to the randomly generated vrand. More specifically, δirr(i) = δ+irr(i)−δ–irr(i) = vrand(i)−virr(i), i ∈ IR, acts over the set of irreversible reactions (IR) and δfor(i) = δ+for(i)−δ–for(i) = vrand(i)−vfor(i), δback(i) = vrand(i)−vback(i), i ∈ RR, over the set of reversible reactions (RR). Note that both δirr, δfor, are defined as the difference of two non-negative components, which enables us to formulate a linear objective function that renders OP2 computationally tractable. In contrast, δback does not require this treatment since it always takes non-negative values (see Fig 2). This is because in OP2, the stoichiometric matrix, S, corresponding to the GEM is first modified in the following way: we change the sign of the columns, as well as the entry in vrand, corresponding to reversible reactions that were randomly assigned a negative flux value in vrand. In this manner, all reversible reactions in vrand operate in forward direction (i.e., are non-negative) which facilitates the optimization process. In addition, δfor and δback are constrained to be mutually exclusive by the same binary variable, x, introduced to select only one of the directions in reversible reactions (i.e. either forward or backward). In this manner, OP2 will select the direction of reversible reactions that minimizes the overall distance to vrand. Finally, reversible reactions whose sign was originally changed in vrand are altered back to their original directions and their sampled flux values are modified accordingly.
In this section, we analyze the alternative optimal solutions of CorEx, a method that we designed in this study to represent the network-centered approaches. In a general sense, network-centered approaches first partition the set R = C∪P of reactions in the original GEM into a core set, C, that must be present in the final context-specific model, and a non-core set, P, which does not necessarily have to be in the final model. These approaches find then a subset PA⊆ P of non-core reactions that renders C consistent, i.e., all reactions in the core are able to carry a non-zero flux in at least one steady-state solution. The final context-specific subnetwork is then defined as RA = C∪PA. Some approaches, like MBA [22], mCADRE [28] and FastCORE [26], aim at minimizing the size of PA, as to obtain a parsimonious final model. In contrast, CORDA [27] relaxes the parsimony condition as a way to prevent eliminating important reactions for a given context. In this respect, CorEx aims at obtaining a parsimonious model, although, as shown in the following, it can be easily adapted to allow increasing the size of PA if desired.
CorEx follows the MILP displayed in OP3, which minimizes the number of reactions with non-zero flux in P while constraining all reactions in the core to carry at least a small positive flux (ϵ in constraints 2–3). This is achieved by minimizing the norm (Z in OP3) of the vector, x, of binary variables (constraints 4–7) which selects the set PA that renders the MILP feasible. Note that the selected non-core reactions are forced to carry a small positive flux (constraints 5, 7) to guarantee that they are active in the final context-specific model. Finally, like in RegrEx, reversible reactions are split into the forward and backward directions, to operate only with non-negative flux values. In addition, another vector of binary variables, y in constraints 8–9 of OP3, is introduced to select the direction of reversible reactions (i.e., imposing vfor > XOR vback > 0, when the reaction is selected to be active).
To identify alternative optimal CorEx extracted networks, we developed the MILP displayed in OP4. The general idea behind OP4 is to find the most dissimilar context-specific network, RA* = C∪PA*, to a previously found optimal RA, that maintains the set C consistent. Namely, it maximizes the number of differences between the reactions contained in PA and PA*. Note that OP4 inherits constraints 1–9 from OP3, and incorporates three new constraints. Constraint 10 guarantees that the cardinality of PA* equals that of the previous optimal PA in OP3. Constraint 11 introduces two additional binary variables, δ+, δ–, which measure the mismatches between the vectors x, selecting the reactions in PA*, and the optimal vector xopt, selecting the reactions in PA and previously found by OP3. Finally, constraint 12 is added to impose a δ+ XOR δ− relationship to avoid the trivial optimal solution in which δ+ = δ–,
maxv=[virr;vfor;vback]∈ℝ0r+,x=[xirr;xrev],δ+,δ−∈{0,1}Py∈{0,1}rev ||δ++δ−||1s.t.1−9. (OP3)10. || x||1=Z11. x+δ+−δ−=xopt12. δ++δ−≤1
(OP4)
However, besides CorEx, OP4 can be used to generate alternative optimal networks to other network-centered approaches. We just need to set xopt, in constraint 11, to be the optimal x vector of the particular approach under study; in addition, we need to update Z, in constraint 10, to the corresponding number of non-core reactions added by this approach (i.e., the size of PA). Note that xopt can be easily constructed from the set PA, which is derived from a particular context-specific model. In addition, a similar constraint to the constraint 10 of OP4, namely ||x||1 ≥ Zlb, may be included in OP3, as a lower bound to its objective function, where Z* ≤ Zlb ≤ R, and Z* is the unconstrained optimum of OP3. It is in this manner that CorEx allows relaxing the parsimony condition, as commented before, although in this study we did not constrain the CorEx optimum.
Noteworthy, the main advantage of using OP4 to obtain alternative optimal networks lies in its MILP formulation. This is because, with the exception of CorEx, which also relies on a single MILP, all existing network-centered approaches require iteratively solving a convex optimization problem. For instance, the linear programs behind the consistency testing step of FastCORE [26], or the ones behind the flux balance analysis, iterated over each reaction of the GEM, in CORDA [27]. Alternative optima may arise in each one of these iterations, thus exploring the alternative optima space in each case would require an extensive computational effort. In contrast, we circumvent this problem with OP4 by analyzing the alternative solutions of a single MILP. However, OP4 only generates a single, maximally different, alternative optimal network. To generate a sample of alternative networks, here we applied OP4 in an iterative way. We first used OP4 to obtain a maximally different network to a given optimal context-specific network, and then repeated this process of feeding OP4 with the successively generated alternative networks until no additional one was found. At that point, we randomly perturbed the last network by changing the state (active or inactive) of 1% of the reactions, and repeated this process until no additional network was found (an implementation of the procedure is provided in S1 File). We note that with this iterative process, which we term the AltNet procedure, we do not guarantee an exhaustive enumeration of all maximally different alternative networks. However, as shown in the next section, it sufficed to illustrate the variety found across optimal context-specific extracted networks in this study.
Finally, we use the AltNet procedure to analyze the alternative optima space of CorEx, FastCORE and CORDA. In the latter case, however, OP4 had to be slightly modified. The reason for the modification is that CORDA divides the reactions in the GEM into four categories, in contrast to CorEx and FastCORE, where only the core, C, and the non-core set, P, are considered. Concretely, reactions are separated into three groups based on experimental evidence: reactions with high (HC), medium, (MC) and negative (NC) confidence, and an additional group collecting the remaining reactions (OT) in the GEM, for which experimental evidence is not available. In this case, the group HC corresponds to the core set of reactions (i.e., all reactions in HC must be included in the final model), and the other three groups constitute the non-core set P, although reactions in MC are preferentially added over NC and OT reactions. To account for the different reaction groups, we partitioned the vector x in OP4 into the sets of MC, NC and OT reactions, and evaluated constraint 10 for each of the three sets. In this manner, we guaranteed that an alternative optimal network contained, besides all HC reactions, the same number of MC, NC and OT reactions than the original CORDA optimum.
Here, we illustrate the ambiguity found during the extraction of context-specific flux distributions and metabolic networks due to the alternative optima. To this end, we apply the methods described in the previous section to two case studies: a leaf-specific scenario, the model plant Arabidopsis thaliana, and a human, liver-specific reconstruction. In the first case, we used the AraCORE model, which includes the primary metabolism of Arabidopsis thaliana [29], and a leaf-specific gene expression data set, obtained from [30] (Methods). In the second case, we employed Recon1, a well-established human metabolic model [31]. Moreover, we considered two different core sets of reactions that were identified as liver-specific by experimental evidence (taken from [19] and [20]), and upon which the liver reconstructions were built. In addition, we reduced the original metabolic models by taking only the consistent part of them. The resulting models are termed here Recon1red and AraCOREred, and contain a total number of 2469 and 455 reactions, respectively (see Methods for details).
We first analyzed the alternative optima space of RegrExLAD—as a representative of a flux-centered approach—and evaluated the ability of the ℓ1-regularization of RegrExLAD to reduce this space. To this end, we focused on the leaf-specific scenario; however, we also applied these methods to the liver-specific scenario, to verify if our main conclusions held in the case of a larger genome-scale model. We then applied CorEx, a network-centered representative, to extract and analyze the alternative optima for the leaf- and the liver-specific reconstructions, and compare its performance with that of FastCORE [19], a well-established approach. In addition, we evaluated the alternative optimal liver-specific networks generated by CORDA, a recently published approach [20]. Finally, we also investigated the alternative optima of iMAT to the leaf- and liver-specific scenario with both, the original approach proposed in [16] and our own complementary method.
After applying RegrExLAD with λ = 0, we obtained an optimal, leaf-specific flux distribution. We then applied RegrExAOS to evaluate the alternative optima space of the previously obtained optimum. The results from this evaluation confirmed the existence of an alternative optima space for RegrExLAD. However, the variability of the fluxes at the optimal objective value was not uniform across different reactions. As expected, data-orphan reactions exhibited more broadly distributed flux values at the alternative optima than data-bounded reactions. We quantified this property by the Shannon entropy (Methods), as a measure of uncertainty of flux value prediction associated to a data integration problem. In this sense, data-orphan reactions showed a larger mean entropy value of 1.64 in comparison to the value of 0.95 found for the data-bounded reactions (one-sided ranksum test, p-value = 1.95x10-5). However, we found reactions with particularly low or high entropy values in both sets, data-bounded and data-orphan (S1 Table).
This last observation suggests that reactions with low entropy values may be of special importance under the leaf-specific metabolic state. On the other side, high entropy values suggest that the corresponding reactions could operate more freely in the leaf context. For instance, we found that the majority of transport reactions showed large entropy values, in accord with the fact that most transport reactions are data-orphan. Nevertheless, there were some transport reactions with particularly low entropy values, such as: the TP/Pi translocator (reaction index 327 in AraCOREred, H = 0.07) interchanging glyceraldehyde 3-phosphate and orthophosphate between the chloroplast and cytoplasm, the P5C exporter (index 363, H = 0.01) exporting 1-Pyrroine-5-carboxylate from mitochondria to cytoplasm and the ADP/ATP carrier (index 320, H = 0.01), interchanging ATP and ADP also between mitochondria and cytoplasm (for a comparison, the highest entropy value in the rank is H = 2.92, corresponding to the Proline uniporter, see the complete list in S1 Table). Therefore, the leaf data integration constrains these transport reactions to take a small range of different flux values due to the network context in which they operate, since they are not directly bounded by experimental data. This observation is contrasted by the high entropy values that these same three reactions when no experimental data are integrated, i.e., when a similar sampling procedure is performed in which only mass balance and thermodynamic constraints are imposed (Methods). In this case, all three entropy values are markedly larger (H > 2, S1 Table).
We next focused on the entropy values of reversible reactions in the AraCOREred model. Reversible reactions in a GEM usually correspond to reactions for which no thermodynamic information is available (leaving aside the set which is known to operate close to equilibrium). Therefore, it would be informative to evaluate whether integrating context-specific experimental data in a GEM could be used to fix the direction of such reactions. Interestingly, we found that a large proportion (75.81%) of the reversible reactions carrying a non-zero flux (including data-orphan) had a fixed direction, either forward or backward, in the alternative optima (Table 1). This finding indicates that, even though there is variation in the flux value of reversible reactions, integration of expression data can determine their direction in a given context. Therefore, the proposed approach and findings provide valuable information on how metabolism could be operating under the particular condition.
For the analyzed sequence of increasing λ-values, the table includes: The sum of entropy values for the subset of data-bounded, HData, and data-orphan, HOrphan, reactions, as well as for all reactions, HTotal, the mean entropy value across all reactions,H¯Total, and the proportion of reversible reactions with fixed direction in the alternative optima sample, FixedRev.
We next evaluated the RegrExLAD alternative optima space for a sequence of increasing λ-values. This was motivated to test whether the inclusion of ℓ1-regularization, besides imposing sparsity in optimal flux distributions, could also reduce the variability found in individual reaction flux values across the alternative optima space. This property could serve as a way to decrease the uncertainty, as measured by the Shannon entropy, associated to a context-specific data integration problem. To this end, we first applied RegrExLAD on AraCOREred and the same leaf data set, but using three increasing λ-values (λ1 = 0.1, λ2 = 0.3 and λ3 = 0.5). We then applied RegrExAOS to sample the alternative optima space of each of the three RegrExLAD data integrations.
We found that the entropy tended to decrease with increasing λ-values, although the effect was more pronounced for the data-orphan reactions (Table 1, Fig 3). For instance, the sum of entropy values among data-orphan reactions decreased from a value of HOrphan = 86.82 for λ = 0, to HOrphan = 36.50 with λ = 0.5. In contrast, for the data-bounded reactions, it only decreased from a value of 73.17 with λ = 0 to 65.46 with λ = 0.5, and even led to a transient increase at λ = 0.3 (Table 1, Fig 3). These findings suggest that the inclusion of regularization can reduce the uncertainty associated to a context-specific data integration problem. Naturally, there is a trade-off between decreasing uncertainty and increasing sparsity of the obtained models, since greater λ-values also produce smaller models that may exclude reactions that are relevant to a particular context (S1 Fig). However, a mild regularization (λ = 0.1) already had a substantial effect in reducing the uncertainty of the RegrExLAD data integration in this analysis. Specifically, it decreased the total model entropy, defined as the sum of entropy values over all reactions, by 16.54% (from a value of HTotal = 159.99 for λ = 0, to HTotal = 133.52 with λ = 0.1, Table 1).
Finally, we focused on the effect that regularization had on reversible reactions. We found that the number of reversible reactions with fixed direction increased monotonically with increasing λ-values (Table 1). Hence, this finding suggests that a mild regularization can further constrain the direction in which a reversible reaction can proceed under a particular metabolic context.
We next analyzed the alternative optima space of RegrExLAD in the liver scenario. Specifically, we focused on evaluating whether the qualitative results obtained in the leaf context remained unchanged when using Recon1red, a larger genome-scale model. To this end, we used a liver-specific and publicly available gene expression data set [32], and mapped it to the reactions in Recon1red following the same procedure as in the leaf-scenario (Methods). Obtaining samples in a larger model is more challenging, due to the increased computational time required to solve the MILP of OP2. Therefore, we restricted our sample to 100 random points for each of the four λ-values evaluated here, as to avoid an excessively large computational time (the total sample time remained under 41 hours, see Methods for details). In this case, we observed a general qualitative agreement between the leaf and the liver scenarios throughout the increasing λ sequence (Fig 3E–3H). More specifically, data-orphan reactions showed a monotonic decrease in their median entropy values; however, this effect was less apparent in the case of data-bounded reactions. Specifically, although the total entropy values of data-bounded reactions tended to decreased with increasing λ, with the exception of λ = 0.5 (Table 1), these differences were not significant (one-sided ranksum test, α = 0.05). However, we observed marked differences when looking at the proportion of fixed reversible reactions. In general, this fraction was smaller in the liver scenario, 61.78% versus 75.81% with λ = 0 (Table 1), and, in contrast to the leaf case, it did not show an increasing trend with increasing λ-values. We conclude that, while the sample size was smaller than that in the leaf case, these results again suggest that a mild ℓ1-regularization of RegrExLAD can be of help in reducing the ambiguity of context-specific flux values.
We first applied CorEx and FastCORE to reconstruct two leaf-specific networks, LeafCorEx and LeafFastCORE. To this end, we used the AraCOREred model and a core set of 91 reactions, which was previously obtained by considering reactions for which the associated gene expression data had a value greater than the 70th percentile (Methods). Both LeafCorEx and LeafFastCORE, contained the core set and were consistent, i.e., all reactions were unblocked. However, we noticed that LeafCorEx was more compact than LeafFastCORE, containing 236 versus 254 non-core reactions, respectively (Table 2). We next reconstructed the two liver-specific networks in a similar way. To this end, we used the Recon1red model, and the core set of 1069 reactions defined in the original FastCORE publication [26]. In this case, CorEx added 593 non-core reactions to the core set, obtaining the liver-specific reconstruction LiverCorEx. FastCORE, on the other hand, added 677 non-core reactions to generate LiverFastCORE. Hence, CorEx was able to extract a more compact liver-specific network, resembling the behavior found in the leaf-specific case. After obtaining these context-specific metabolic reconstructions, we searched for alternative optimal networks to all of them, using the AlterNet procedure describe in the previous section. To quantify the uncertainty of the leaf- and liver-specific reconstructions, we looked at the number of reaction mismatches between all pairs of alternative networks in each case (computed as the Hamming distance, see Methods). This metric was normalized by the total number of reactions in each metabolic model to allow fair comparison between the two case studies.
This table summarizes the results of the evaluation of the CorEx alternative optima space. It includes the number of added non-core reactions, P, the maximum, MRmax (within brackets the percentage of reaction in P), and the mean number, MR¯ (CV stands for coefficient of variation), of reaction mismatches (i.e., Hamming distance) across the alternative networks for the leaf- and the liver-specific scenarios evaluated by two methods, CorEx and FastCORE. The last column displays the p-value resulted from a one-sided ranksum test comparing the distributions of Hamming distances between any pair of the alternative networks of CorEx and FastCORE (the null hypothesis states that the distribution generated by CorEx is bigger than that of FastCORE).
We found marked differences between alternative optimal networks in both approaches and metabolic scenarios. In the case of LeafCorEx, alternative networks differed on average in 29 non-core reactions, with a maximum value of 52 reactions (22% of the added non-core reactions). In LeafFastCORE, networks differed on average in 66.78 reactions, and had a maximum number of 118 discrepant reactions (46.5%, Table 2). This situation was even worsened in the liver-specific reconstructions. Between alternative networks to LiverCorEx, we found a maximum of 156 discrepant reactions among the 593 in the added non-core (26.3%), with an average of 108.3. In the case of LiverFastCORE, the maximum number of discrepant reactions was as high as 398 out of the 677 (58.8%) added non-core reactions, with an average of 246.93 between alternative optimal networks (Table 2).
As a complementary analysis, we also determined the frequency of occurrence of every non-core reaction across the alternative optimal networks. In this manner, we could identify: (i) a set of non-core reactions that were always included, termed the active non-core set, (ii) a set of non-core reactions that were excluded from all alternative networks, termed the inactive non-core set, and (iii) a set of non-core reactions that were included in some of the networks, referred to as the variable non-core set. In this case, we took the size of the variable non-core set as a measurement of the uncertainty of a context-specific network extraction; 28% and a 47% of the total non-core reactions were in the variable set in the cases of LeafCorEx and LeafFastCORE. On the other hand, a 12% and a 58% were found in LiverCorEx and LiverFastCORE, respectively (Fig 4A–4D).
The previous results quantify the structural differences among the generated alternative optimal networks. However, these structural differences do not consider which kind of reactions (i.e., in which pathways in the GEM) are more or less frequent (i.e., ambiguous), in the alternative optima space. To address this issue, we assigned a score (between 0 and 1) to each metabolic pathway based on its representation in the active, variable or inactive non-core set. Specifically, the score represents the fraction of reactions of a given pathway that are assigned to a non-core subset with respect to the total size of the non-core set (Methods). Pathways with high score values in the active and inactive non-core are consistently over- and under-represented, respectively, among the alternative optimal networks. Therefore, these pathways should be more important (the opposite in the inactive non-core case) to maintain the core active and hence the assumed context-specific metabolic function. In contrasts, pathways with high-score values in the variable non-core tend to be represented only in certain alternative optimal networks, thus being more ambiguous in the context-specific reconstruction.
For instance, in the leaf scenario, we found among the pathways with highest score in the active non-core: the Calvin-Benson cycle, light reactions and photorespiration. All of these pathways showed a maximum score value of 1 in both cases LeafCorEx and LeafFastCORE, which agrees with key roles of these pathways in a photosynthetic tissue. Additionally, alongside these photosynthetic pathways, we also found housekeeping pathways for the synthesis of AMP, CTP, GMP, UMP, Acetyl-coA or Fatty acid, among others, with the maximum score value in both cases. More interestingly, among the pathways with the highest scores in the variable set we also found primary pathways like the Tricarboxylic acid cycle, Alanine synthesis, the Pentose Phosphate Pathway and Pyruvate metabolism. However, we also found pathways that are usually linked to active photosynthetic tissues like Starch and sucrose degradation and sucrose synthesis (see S9 Table for a complete list containing the ranked pathways).
Moreover, in the liver scenario, we also found typical liver-specific pathways like Cholesterol Metabolism and Fatty acid oxidation [33] with the maximum score value in the active non-core in the case of LiverCORDA. However, we also found a variety of other pathways with high scores in the variable non-core like CoA catabolism, ROS detoxification or Vitamin A metabolism, which indicates that the variable non-core set contains a diverse set of metabolic functions that may be important to the canonical liver physiology (see S9 Table for a complete list of the ranked metabolic pathways).
Finally, we analyzed the alternative optima space of CORDA, a recently published network-centered approach [27]. As explained in the previous section (Computational methods) CORDA differs to CorEx and FastCORE in two ways. On one hand, CORDA does not aim at obtaining compact or parsimonious models, but rather emphasizes the metabolic functionality of the final context-specific reconstructions. On the other hand, CORDA considers four groups of reactions based on experimental evidence, out of which only one, the high confidence core set (HC), has to be fully included in the final model (thus being equivalent to the core set of CorEx and FastCORE). In this case, a suitable alternative optimal network must contain not only the entirety of the HC set, but exactly the same number of reactions added by CORDA in each one of the three remaining groups: the medium (MC) and the negative confidence (NC) groups, and the reactions without experimental data (OT). Therefore, it is reasonable to expect that this additional constraint may reduce the uncertainty of the CORDA reconstructions.
To test this idea, we searched for alternative networks to the CORDA liver reconstruction (here LiverCORDA) provided in [27]. LiverCORDA was obtained from Recon1 and experimental evidence from the Human Protein Atlas [13], and contains 279 HC, 369 MC, 11 NC and 1147 OT reactions. We used again our AltNet procedure, Recon1red (since blocked reactions, by definition, can never be included in a final network), and the classification of the reactions in the four groups also provided in [27]. We were indeed able to find alternative networks to the original LiverCORDA with marked differences among them. Concretely, a maximum number of 992 discrepant reactions between two alternative networks, out of the total 1527 distributed among the MC, NC and OT groups (65%, Table 2), with a mean number of 545.22. Similarly, 51% of the non-core reactions (MC, NC and OT) in Recon1red were assigned to the variable non-core set (Fig 4E).
The examples presented here show that the context-specific reconstructions are more ambiguous than specific, especially in the human liver scenario. This latter case is of special concern, given the implications of obtaining accurate context-specific reconstructions in biomedical research. In fact, most, if not all, of the network-centered approaches have focused on human metabolism [22,26–28]. There are ways, however, to cope with this ambiguity or uncertainty of context-specific reconstructions. For instance, as commented before, CORDA aims at obtaining functional reconstructions. In fact, the authors in [27] tested the capability of the LiverCORDA model to conduct a basic set of liver metabolic functions, including aminoacid, sugar and nucleotide metabolism.
We evaluated the alternative LiverCORDA models with the same metabolic test (Methods), and extracted the subset that passed it. Among these networks, we found that the number of discrepancies and the size of the variable non-core were significantly reduced, as compared to the total set of alternative networks (Table 2, Fig 4E and 4F). This is not surprising, since requiring the alternative networks to fulfill certain metabolic functions indirectly imposes an additional constraint to the optimal solution. On the other hand, this additional constraint can also be realized by augmenting the core set, as to guarantee that certain key reactions are present in the final context-specific network. This relates to an additional way to reduce the ambiguity of the reconstruction. In the case studies evaluated here, we found that the CorEx alternative networks tended to be more similar among each other than that of FastCORE or CORDA, as quantified by the (normalized by non-core size) number of discrepancies (Table 2). These differences may be explained by the number of non-core reactions included in the optimum: CorEx obtained more compact models than FastCORE in the Leaf- and the Liver-specific case. This imposes a more stringent constraint when searching for alternative optimal networks. However, there is a tradeoff between model parsimony and functionality. In fact, the LiverCorEx model was not able to pass the metabolic function test, while LiverFastCORE was able to pass it. In this particular case, LiverCorEx did not contain the 9 basal exchange reactions (Methods) required to perform the metabolic functions in the test. However, including these 9 reactions in the liver core set sufficed to generate a LiverCorEx model that passed the test.
The analysis of the alternative optima space can be employed to cope with the ambiguity of a context-specific network reconstruction. Notably, the authors of EXAMO (EXploration of Alternative Metabolic Optima) [21] proposed a first step in this direction. In this case, EXAMO first generates a sample of alternative optimal flux distributions of iMAT [20]. It then focuses on the activity state of each reaction across the sample, for which it binarizes the flux values through the usage of an arbitrary threshold value. A reaction is included in the High Frequency Reaction (HFR) set if it is active throughout the alternative optima sample. Finally, EXAMO uses the HFR set as a core set to MBA [22], a network-centered method, which reconstructs the minimal network that renders the HFR set consistent. EXAMO directly addresses the problem of alternative optima. However, the final context-specific model is again subject to the effects of alternative optima, since a set of alternative networks, all containing the HFR set as a core, could be found for the MBA method.
A possible way to circumvent this problem in the case of iMAT could be the following: i) similar to EXAMO, obtain samples of alternative optimal flux distributions, binarize flux values and rank the reactions according to the number of times that they appear as active in the sample, ii) include the reactions that are always active (the HFR set) in a core set and the rest in a non-core set, and iii), add non-core reactions in decreasing order of frequency until consistency of the core is reached. In this manner, this ranking provides a way to select which non-core reactions are included in the final model. This idea parallels that of mCADRE [28], although in the latter, reactions are ranked following an heuristic approach that considers experimental evidence from several databases, which may be difficult to obtain for certain metabolic contexts. Finally, to generate the sample of alternative optima flux distributions of iMAT, we propose a sampling method similar to RegrExAOS that allows drawing arbitrarily large samples, as opposed to the one used in EXAMO which generates samples of restricted size. Details about this method, here called iMATAOS, can be found in S2 Appendix.
In the case of the network-centered approaches here evaluated, establishing a ranking of non-core reactions could also be a way to deal with the ambiguity during network reconstructions. Non-core reactions that occur with high frequency in the alternative optima space should be preferentially included in the final network, while reactions with a low frequency should be discarded. To guarantee that the final network is consistent (i.e. the core set is active), non-core reactions could be again added in decreasing order of frequency to the core set until consistency is reached. Naturally, this requires the development of competent methods to sample the alternative space of network-centered approaches. In this sense, we consider our proposed AltNet procedure a first step towards this goal.
We analyzed the space of alternative optima resulting from the integration of context-specific data into GEMs. To this end, we evaluated a representative set from the flux- and network-centered approaches. We selected RegrEx [25] as a representative of flux-centered approaches and CorEx, as a network-centered approach, proposed in this study. In addition, we adapted CorEx to obtain alternative optimal networks for FastCORE [26] and CORDA [27], two state-of-the-art network-centered approaches. We compared the developed approaches and implemented tools on two illustrative case studies: (i) a medium size GEM of the primary metabolism of Arabidopsis thaliana [29] and a leaf-specific gene expression data set [30], and (ii) a larger GEM collecting a reconstruction of a human metabolic network [31], two liver-specific core sets of reactions [26,27] and a liver-specific gene expression data set [32].
Our findings demonstrated the existence of a space of alternative optima for all evaluated approaches integrating context-specific data. Consequently, this space of alternative optima induces ambiguous context-specific reconstructions. In the case of flux-centered approaches, RegrExLAD in this study, we proposed the usage of a mild regularization to mediate the uncertainty of the resulting context-specific fluxes. In network-centered approaches, our results showed the existence of markedly disparate alternative context-specific networks in CorEx, FastCORE and CORDA. A delicate balance between model parsimony and metabolic functionality seems key to reducing the ambiguity of the context-specific reconstructions. Additionally, an evaluation of the alternative optima space followed by a ranking of the reactions according to their frequency may serve as a way to determine their context-specificity. On this line, we proposed the AltNet procedure to generate alternative optimal context-specific networks.
As a concluding remark, we acknowledge the utility of the existent experimental data integration methods, since they allow a fast and automated generation of context-specific flux distributions and metabolic networks. However, our findings indicated that the interpretation and further usage of their results warrant caution. Specially, since the existence of alternative optima is likely linked to the nature of the context-specific data integration problem, and thus is independent of the approach used. The latter claim is supported by our evaluation across qualitatively different approaches. We advocate the view that an analysis of alternative optimal solutions should be performed, whenever possible, if context-specific data are integrated in metabolic models. In the case of context-specific networks reconstructions, more reliable results could be obtained from subsequent careful knowledge-based curation.
This section contains the details about the implementation of the methods described in this study, the GEMs and context-specific data employed in the case examples, and the computation of the distance metric between alternative optimal networks. In addition to this section, the MATLAB code containing the entire workflow followed in this study can be found in the Supplementary Information.
All optimization programs used in this study, (i.e., OP1-6) were implemented in MATLAB and solved using Gurobi (version 7.1) [34] on a desktop machine with an Intel Core i7-4790 @3.6 GHz processor and 16GB of RAM. We used default Gurobi parameter values except for: i) reduced feasibility tolerance to 10−9 when solving OP3-4, ii) increased MIPGap parameter to 1% when solving the MILP of OP2. All generated code with the implementations is available as Supplementary Information.
A reduced version of the original AraCORE model [29] was used in this study: AraCORE contains 549 reactions and 407 metabolites assigned to four subcellular compartments, whereas the herein used version (AraCOREred) contains 455 reactions and 374 metabolites. The reactions that were removed correspond to exchange reactions that directly connect organelles to the environment (circumventing the cytoplasm), and were eliminated to avoid bias in the obtained flux distributions. AraCOREred can be found in the Supplementary Material.
Leaf-specific gene expression values were taken from [30], stored in the GEO database under the accession numbers GSM852923, GSM852924 and GSM852925 corresponding to Arabidopsis thaliana Col-0 lines with no treatment. The corresponding CEL files were normalized using the RMA (Robust Multi-Array Average) method implemented in the affy R package [35]. In addition, probe names were mapped to gene names following the workflow described in [36], where probes mapping to more than one gene name are eliminated. Gene expression values were then scaled to the maximum value and mapped to reactions in the AraCOREred model following the included Gene-Protein-Reaction rules and a self-developed MATLAB function, mapgene2rxn, which is available in S1 File. This process was repeated for the three samples in the dataset and mean values were taken as representative values to obtain the final leaf-specific data used in this study.
Liver-specific gene expression values were obtained from [32], which is accessible under: http://medicalgenomics.org/rna_seq_atlas/download. In this case, we used the RPKM values corresponding to the liver (normal tissues). Since the RPKM values are already normalized we used them directly as input of the mapgene2rxn procedure, already described.
We removed blocked reactions from the original Recon1 model to get the Recon1red model used in this study. To this end, we performed a Flux Variability Analysis (see next section) and removed reactions with a maximum absolute flux, |vi| < 10−6. The Flux Variability Analysis was implemented in the MATLAB function reduceGEM which also extracted the reduced model, Recon1red, in a COBRA compatible MATLAB structure. The function is available in S1 File.
The minimum and maximum allowed values of each reaction in AraCOREred were determined through Flux Variability Analysis [4]. Although only the mass balance and the thermodynamic constraints were imposed (i.e., no reaction was forced to take a fraction of a previously calculated optimal value). This was accomplished through the following linear program,
min/max vvi, ∀i∈vs.t.Sv=0vmin≤v≤vmax ,
which was implemented in MATLAB and solved with the Gurobi solver (version 6.04). The own-developed MATLAB function can be found in Supplementary Material under the name of FVA.
To evaluate to what extent the Leaf data integration affected the entropy values of the reactions in the AraCOREred model, we also sampled the space of feasible flux distributions (i.e., the flux cone) when no experimental data was been integrated. To this end, and to allow direct comparability of the results, the flux cone was sampled following a similar approach as in RegrExAOS: first, we generated a random vector of flux values, vrand, within the minimum and maximum values obtained by regular Flux Variability Analysis. The closest flux vector v to vrand within the flux cone was then obtained by minimizing the Euclidean distance between the two vectors. The following quadratic program was used to this end:
minv12‖v−vrand‖22s.t.Sv=0vmin≤v≤vmax .
This procedure was iterated to obtained a sample of size n = 2000. After the sample was generated, we obtained the Shannon entropy values of the samples in the same way as when evaluating the alternative optima space of RegrExLAD (described in the next section). The MATLAB function implementing this sampling procedure can be found in S1 File under the name coneSampling.
The Shannon entropy of the sampled alternative optima distribution, Hi, was used to quantify the extent to which the flux values of a reaction, i, varied across the alternative optima space. It was calculated as follows:
Hi=−∑k=1nfi,klog(fi,k).
Where fi,k represents the frequency (i.e., number of counts relative to sample size) of the k interval in the distribution, for n = 20 equally spaced flux value intervals within the flux range of i. In addition, the total entropy of an alternative optima space, HT, was defined as the sum of the entropies corresponding to the r reactions in AraCOREred, i.e.,
HT=∑i=1rHv(i),
and was taken as a measure of the total flux variability found in a particular alternative optima space.
In the case of CorEx, we generated the set of alternative optimal metabolic networks from the set of sampled alternative optimal flux distributions. To this end, we first generated the binary vector representations of the flux distributions. The binary vector representations were generated by assigning a value of 1 to the entries corresponding to reactions with a flux value v ≥ 10−6, and 0 otherwise. This process was repeated for each sampled alternative optimal flux distribution. In addition, repeated vector representations were removed from the generated set. After the binary representations were obtained, we calculated the number of mismatches between any pair, a,b, of binary vectors, with a ≠ b, i.e., the Hamming distance,
MR(a,b)=∑k=1n|a(i)−b(i)|.
In this way, we obtained a distribution of MR values whose characteristics were reported and compared.
We computed a score, ranging between 0 and 1, to quantify the ambiguity found in individual metabolic pathways (subsystems in the GEM) across the space of alternative optimal networks. Concretely, the score of a pathway, M, represents the fraction of the reactions in the (total) non-core set, P, belonging to the pathway that are assigned to the active, variable or inactive non-core (thus producing a score value for each case). That is, in general,
SX(M)=XMP,
where XM ∈ {AM, VM, IM} represents the number of reactions assigned to M that are included in the active, variable or inactive non-core, respectively.
We performed the same metabolic test proposed in [27] and applied to the original Liver-specific CORDA reconstruction. This test consists of a list of metabolic tasks that a metabolic model has to perform, including parts of the aminoacid, sugar and nucleotide metabolism. Concretely, there a total of 48 metabolic tasks, divided into the production of different aminoacids from minimal metabolic sources and the excretion on urea (19 tasks), the ability to synthetize glucose from 21 different sources (including some aminoacids), and the production of all 5 nucleotides and nucleotide precursors (8 tasks). The details about these tasks can be found in the original CORDA publication [27], while the MATLAB code of our implementation of this test is provided in S1 File. In this study, we used the fraction of performed tasks as measure of the ability of a given liver-specific model to pass this test. For instance, the liver-specific model provided in [27] (under the name of liverCORDAnew), was able to pass 89.58% of the tasks (43 out of 48). In this study, however, we required to pass all tasks in the test to consider an alternative liver-specific network as functional. We realized that, in the liverCORDAnew model, some reactions were slightly different to the analogous reactions in the Recon1red model that we used throughout this study (likely due to different versions of the Recon1 model, which is periodically updated [37]). When we reconstructed our LiverCORDA model, using the same reaction identifiers in liverCORDAnew but extracting the reactions from our Recon1red version, we found that the generated model passed all metabolic 48 tasks in the test. Hence, for consistency of the results, we considered that all proper alternative optimal networks to LiverCORDA had to pass all 48 tasks as well.
|
10.1371/journal.pgen.0030214 | Rrp1b, a New Candidate Susceptibility Gene for Breast Cancer Progression and Metastasis | A novel candidate metastasis modifier, ribosomal RNA processing 1 homolog B (Rrp1b), was identified through two independent approaches. First, yeast two-hybrid, immunoprecipitation, and functional assays demonstrated a physical and functional interaction between Rrp1b and the previous identified metastasis modifier Sipa1. In parallel, using mouse and human metastasis gene expression data it was observed that extracellular matrix (ECM) genes are common components of metastasis predictive signatures, suggesting that ECM genes are either important markers or causal factors in metastasis. To investigate the relationship between ECM genes and poor prognosis in breast cancer, expression quantitative trait locus analysis of polyoma middle-T transgene-induced mammary tumor was performed. ECM gene expression was found to be consistently associated with Rrp1b expression. In vitro expression of Rrp1b significantly altered ECM gene expression, tumor growth, and dissemination in metastasis assays. Furthermore, a gene signature induced by ectopic expression of Rrp1b in tumor cells predicted survival in a human breast cancer gene expression dataset. Finally, constitutional polymorphism within RRP1B was found to be significantly associated with tumor progression in two independent breast cancer cohorts. These data suggest that RRP1B may be a novel susceptibility gene for breast cancer progression and metastasis.
| Metastasis, which is defined as the spread of malignant tumor cells from their original site to other parts of the body, accounts for the vast majority of solid cancer-related mortality. Our laboratory has previously shown that host germline-encoded variation modifies primary tumor metastatic capacity. Here, we detail how germline-encoded Rrp1b variation likely modulates metastasis. In mice, constitutional Rrp1b variation correlates with ECM gene expression, which are genes commonly differentially regulated in metastasis prone tumors. Furthermore, we demonstrate that Rrp1b expression levels are modulated by germline variation in mice with differing metastatic propensities, and that variation of Rrp1b expression in a highly metastatic mouse mammary tumor cell line modifies progression. Differential RRP1B functionality also appears to play an important role in human breast cancer progression. Specifically, we demonstrate that a microarray gene expression signature indicative of differential RRP1B expression predicts breast cancer-specific survival. Furthermore, we show that germline-encoded RRP1B variation is associated with markers of outcome in two breast cancer populations. In summary, these data suggest that Rrp1b may be a germline-encoded metastasis modifier in both mice and humans, which leads to the possibility that knowledge of RRP1B functionality and variation in breast cancer might facilitate improved assessment of prognosis.
| Most cancer-related mortality is a consequence of metastasis, and the vast majority of deaths from breast cancer, the most common malignancy of women in the United States [1], are attributable to disseminated disease. Disseminated breast cancer is still considered incurable in spite of therapeutic advances [2], and a more comprehensive understanding of the biology of tumor progression is therefore necessary to facilitate development of improved treatments. This includes the ability to spare women at low risk of metastasis from needless additional therapy, while allowing earlier initiation of aggressive treatment to reduce the incidence and extent of metastasis in women with poorer prognoses.
We previously demonstrated the significant influence of germline variation on tumor progression [3,4], which allowed us to identify the first known heritable mouse gene that modulates metastasis [5,6], the Rap-GTPase activating protein (GAP) Sipa1 [7]. Subsequent human studies demonstrated that SIPA1 polymorphisms are associated with metastatic cancer [7] and poor outcome in breast cancer [8], validating the utility of the highly metastatic polyoma middle-T (PyMT) transgenic mouse model to identify relevant human metastasis modifiers. The current study represents the convergence of two parallel strategies to enhance our understanding of the role of heritable factors in metastasis. Using in vitro, genetic, and epidemiologic analyses, we have identified ribosomal RNA processing 1 homolog B (Rrp1b) as a factor that physically interacts with the metastasis modifier gene, Sipa1, modulates elements of metastasis predictive gene expression signatures, suppresses tumor progression in animal models, and is associated with progression and survival in pilot human breast cancer epidemiology cohorts. This integrated approach suggests that Rrp1b is a novel tumor progression and metastasis susceptibility locus in both mice and humans.
Previous mouse studies demonstrated that a polymorphism in Sipa1 in the region encoding a PDZ protein–protein interaction domain is associated with metastasis [7]. Yeast two-hybrid screening of Sipa1 was therefore performed to identify additional genes potentially involved in metastasis (Table S1). Following sequence alignment, 29 clones were found to bind to at least one of the SIPA1 baits (Table S2). One of these was RRP1B (the human homolog of Rrp1b), which was identified by a probe spanning the PDZ domain.
To confirm the interaction, HEK293 cells were cotransfected with epitope-tagged mouse Rrp1b and Sipa1. AQP2, which also interacts with the PDZ domain of Sipa1, was cotransfected with Sipa1 as a positive control. Cell extracts were then immunoprecipitated with Sipa1 antibodies and blotted with V5-antibodies (V5 was the epitope fused to Rrp1b in this experiment), revealing an Rrp1b-specific band (Figure 1A, upper panel, lane 5). Conversely, when HA-tagged Rrp1b was cotransfected with V5-tagged Sipa1, immunoprecipitation with an HA-antibody followed by western blotting yielded a Sipa1-specific band (Figure 1B, upper panel, lane 3).
As further validation, the functional consequence of the Rrp1b–Sipa1 interaction on the Rap-GTPase enzymatic activity of Sipa1 was examined. HEK293 cells were cotransfected with a Rap exchange factor, Epac, and Sipa1 in the presence of AQP2 or Rrp1b (Figure 1C). AQP2, which has been shown previously to interfere with the RapGAP activity of Sipa1 [7], was used as a positive control. In the absence of Sipa1, Epac induced an increase in Rap-GTP, regardless of whether the cells also expressed AQP2 or Rrp1b (upper panel, lanes 1–3), indicating that Rrp1b did not directly affect Rap-GTP levels. As expected, the presence of Sipa1 reduced Epac-induced Rap-GTP levels (upper panel, lane 4). This reduction was partially inhibited by AQP2 or Rrp1b (upper panel, lanes 5 and 6, respectively). Thus Rrp1b, like AQP2, inhibits the RapGAP activity of Sipa1.
Examination of published reports describing primary human breast tumor expression profiles predicting metastasis or disease outcome reveals a common association with the expression levels of extracellular matrix (ECM) genes [9–11]. Similar ECM-rich metastasis-predictive signatures also exist in PyMT-induced mouse mammary tumors [12]. The consistent association of ECM gene expression levels with outcome suggests that differential ECM gene expression is either a causative factor or marker of metastatic potential.
In a set of experiments performed concurrently with our efforts to identify proteins interacting with SIPA1, the relationship between ECM gene expression and metastasis susceptibility was further characterized by investigating the inherited origins of metastasis-predictive gene signatures. Specifically, expression quantitative trait locus (eQTL) mapping of ECM gene expression was performed to identify genomic regions associated with ECM gene expression (see Text S1). To achieve this, we analyzed microarray data derived from PyMT-induced primary tumors in the AKXD RI panel recombinant inbred (RI) mice [13,14], a panel of RI mice derived from high metastatic potential AKR/J and low metastatic potential DBA/2J strains [3]. An Internet-based analytical package/repository that allows for analysis of RI microarray expression data called WebQTL [15,16] was used to map ECM eQTLs. Reproducible, statistically significant or suggestive ECM eQTLs were discovered on chromosomes 7, 17, and 18 (see Figure 2 and Text S1), implying that loci from these three regions regulate much of the metastasis-predictive ECM gene expression in AKXD tumors. The chromosome 17 locus (Figure S1) was of particular interest, as its peak linkage region (∼29.5 Mb) colocalizes with a previously described metastasis efficiency and tumor growth kinetics QTL [6], and encompasses the physical location of Rrp1b (∼29.9 Mb).
Based on the current prevailing hypothesis that most modifiers are likely to result from modest variations in gene expression levels or mRNA stability [16,17], identification of potential candidates for the chromosome 17 ECM modifiers was performed by correlation analysis. Genes were correlated with ECM gene expression on a genome-wide scale using the Trait Correlation function of WebQTL. The results were subsequently filtered to examine only those genes that were present under each of the ECM eQTL peaks. Thirty genes located within a genomic region spanning the peak likelihood ratio statistic (LRS) score (physical locations on chromosome 17 ∼ 18–40 Mb) displayed both a high degree of correlation and a low p value with regard to expression of two or more of the nine probes within metastasis-predictive ECM genes (Table S3). Rrp1b was one of those genes that displayed high levels of expression correlation with ECM gene probes (Table S3). This gene was selected for further analysis for the following reasons: (a) its physical proximity to the peak eQTL linkage; (b) its apparent correlation with expression of various metastasis-predictive ECM genes; and (c) that it had also been identified as interacting with the metastasis efficiency modifier Sipa1. The effects of other genes within the eQTL linkage region upon metastasis remain under investigation, but their potential role in the modulation of metastasis efficiency (if any) is beyond the scope of the current study.
To confirm the role of Rrp1b in the regulation of ECM expression, cell lines stably over-expressing Rrp1b were generated in the highly metastatic mouse mammary tumor cell lines Mvt-1, which is derived from FVB/NJ mice [18] and 4T1, which is derived from BALB mice [19]. Multiple individual clones were generated by clonal dilution, and ectopic expression of Rrp1b confirmed by quantitative real-time PCR (qPCR) (ratio of Rrp1b expression in Mvt-1/Rrp1b versus controls = 3.28 ± 0.41; p = 0.001 and ratio of Rrp1b expression in 4T1/Rrp1b versus controls = 4.55 ± 0.95; p = 0.015). Metastasis-predictive ECM gene expression was then quantified, and expression of eight of the 12 quantified ECM genes (Col1a1, Col3a1, Col6a2, Fbln2, Fbn1, Mfap5, Serpinf1, and Serping1; see Table 1) was significantly changed in response to ectopic Rrp1b expression in the Mvt-1 cell line. Expression analysis of five of six significantly dysregulated genes in the 4T1 cell lines changed in the same direction as the Mvt-1/Rrp1b cells (Table 1), suggesting that Rrp1b modulation of ECM genes was not a unique characteristic of the Mvt-1 epithelial cell line. To confirm that these results were not an artifact of the Mvt-1 and 4T1 tumor cell lines, the experiment was repeated in NIH-3T3 fibroblasts. Of the six metastasis-predictive ECM genes dysregulated in all three cell lines, four showed the same profile in response to ectopic Rrp1b expression. The remaining discrepancies are likely due to experimental variability, differently regulated genes in epithelial (i.e., Mvt-1, 4T1) and mesenchymal cells (i.e., NIH-3T3), or the effects of the differing genetic backgrounds of the three cell lines.
Growth curves were plotted for the Mvt-1 cell lines to confirm that the observed changes in gene expression in did not result from differential cellular growth rates (Figure S2). The growth of Rrp1b-transfected cell lines did not significantly differ from the growth rate of the control cell lines implying that the observed differences in metastasis-predictive gene expression are intrinsically related to the effects of Rrp1b rather than secondary to reduced growth kinetics. Similarly, no differences in growth rates were observed with the 4T1 and NIH-3T3 cells ectopically expressing Rrp1b compared to control cell lines (unpublished data).
Spontaneous metastasis assays were performed by subcutaneously implanting equal amounts of either Mvt-1/Rrp1b or Mvt-1/β-galactosidase clones into virgin FVB/NJ female mice. Specifically, the in vivo growth characteristics of four Mvt-1/Rrp1b clones were compared to that of one Mvt-1/β-galactosidase cell line. Since previous experiments have demonstrated that the in vitro and in vivo growth characteristics of the multiple independent isolates of the control cell line are virtually identical to those of the wild-type cell line, only one such cell line was used to minimize the number of animals in accordance with National Cancer Institute Animal Care and Use guidelines. Tumor weight and lung surface metastasis count were quantified following a four-week incubation period. Both tumor growth and lung surface metastasis were significantly reduced in Mvt-1/Rrp1b clonal isolates. Average tumor weight was 240 mg ± 200 mg for the Rrp1b clones compared to 600 mg ± 270 mg for the β-galactosidase clone (p < 0.001) (Figure 3A), and average lung surface metastasis count being 5.7 ± 6.0 for the Rrp1b clones compared to 12.6 ± 8.4 for the β-galactosidase clone (p = 0.010) (Figure 3B).
Genomic sequencing of Rrp1b was performed from high (AKR/J, FVB/NJ) and low (I/lnJ, DBA/2J, and NZB/B1NJ) metastatic inbred strains to identify polymorphisms that might account for the differential ECM gene expression. Of particular interest were polymorphisms identified in the AKR/J and DBA/2J strains, since these are the progenitors of the AKXD RI panel. In addition to multiple intronic polymorphisms, the AKR/J Rrp1b proximal promoter contained two adenosine insertion polymorphisms located 1,132 bp and 1,540 bp upstream of the transcription initiation site. Rrp1b polymorphisms in all strains are listed in Table S4.
To examine the functional consequences of the AKR/J promoter polymorphisms, a region 1.67 kb upstream of the transcription initiation site of Rrp1b from AKR/J and DBA/2J was cloned into pBlue-TOPO (Invitrogen). Following normalization for transfection efficiency it was found that the AKR/J proximal promoter activity was reduced 30% relative to its DBA/2J counterpart (p < 0.001) (Figure S3; Table S5) implying a subtle functional difference in Rrp1b functionality between the high and low metastatic genotypes.
Given the outcome of the promoter activity experiments, we would expect to observe differential expression of Rrp1b in tissue derived from the AKXD ancestral AKR/J and DBA/2J strains. To test this hypothesis, we quantified expression of Rrp1b in normal mammary tissue from AKR/J or DBA/2J genotype mice. Following total RNA extraction from the mammary tissue from three individual AKR/J mice and three DBA/2J mice, reverse-transcription PCR (RT-PCR) was used to synthesize cDNA, and Rrp1b expression determined using qPCR. Rrp1b expression was found to be ∼30% less in normal mammary tissue derived from high metastatic potential AKR/J mice compared to mammary tissue from the low metastatic potential DBA/2J genotype (AKR/J normalized relative Rrp1b quantity = 1.21 ± 0.20, DBA/2J relative expression = 1.70 ± 0.30; Mann Whitney U-test p = 0.0495; Table S6). Combined, the in vitro and in vivo data imply that germline polymorphism, in the form of proximal promoter polymorphism, is causing differential functionality of Rrp1b in genetic backgrounds of differential metastatic capacity.
If RRP1B is at least partially responsible for the presence of the ECM components of metastasis predictive gene signatures, it would suggest that a signature of RRP1B activation or expression [20] might also be predictive of breast cancer survival. To test this hypothesis, Affymetrix microarrays were used to compare gene expression in four Mvt-1/Rrp1b clonal isolates and three Mvt-1/β-galactosidase clonal isolates. An Rrp1b expression signature was identified using the Class Comparison tool of BRB ArrayTools was performed, using a two-sample t-test with random variance univariate test. p-Values for significance were computed based on 10,000 random permutations, at a nominal significance level of each univariate test of 0.0001. A total of 1,739 probe sets representing 1,346 genes passed these conditions. Significantly upregulated and downregulated probes according to these criteria are listed in Tables S7 & S8, respectively.
A human RRP1B gene expression signature was generated by mapping the differentially regulated genes from mouse array data to human Rosetta probe set annotations [10]. One hundred ninety six genes from the mouse data could be mapped to the available Rosetta Hu25K chip annotations. The 295 samples of the Rosetta data set [10] were clustered into one of two groups representing high and low levels of RRP1B activation in primary tumor samples in an unsupervised manner based on the 196 significantly differentially expressed RRP1B signature genes on the Hu25K chip. Kaplan-Meier survival analysis was performed to investigate whether there was a survival difference between groups. A significant survival difference was observed implying that the level of activation of RRP1B or RRP1B-associated pathways within a tumor, presumably because of either somatic mutation or germline polymorphism, may be an important determinant of the overall likelihood of relapse and/or survival (Figure 4A). Further analysis indicated that survival was associated primarily because of the effects of 33 genes (Table S9). The degree of survival difference represented by the 33-gene RRP1B-induced gene expression signature was similar to the original 70-gene signature described by van't Veer and colleagues [10] (Figure 4B).
Patient samples were stratified by estrogen receptor (ER) and lymph node (LN) status, two clinically relevant prognostic markers, to determine whether the RRP1B signature might provide additional clinical stratification. Expression of the RRP1B signature in bulk primary tumor tissue predicted outcome in patients that were both LN negative and LN positive and patients with ER positive tumors (Figure 4C, 4D & 4E, respectively). Patients with ER negative tumors did not show a significant survival benefit (Figure 4F). However, this may be due to the limited sample size and needs to be clarified with additional studies.
To validate a possible role of RRP1B in human cancer, a case-only pilot breast cancer association study was performed to assess the role of a nonsynonymous SNP within the human homolog of Rrp1b (dbSNP ID: rs9306160; 1421G→A, Pro436Leu) in human disease (Table S10). The variant A allele frequency in this Caucasian cohort was 0.362 (n = 269). Univariate analysis revealed a significant difference with respect to disease stage (localized versus nonlocalized p = 0.006; Table 2). Of the 130 patients with localized disease at diagnosis, 85 (65%) were carriers of the variant allele compared with 74 of 139 (53%) of those with advanced regional or metastatic disease. Significant associations were also observed with tumor ER and progesterone receptor (PR) status, the presence of LN disease, and primary tumor grade. The variant allele was more frequent among patients with ER positive and PR positive primary tumors: 122 of 190 individuals (64%) with ER positive tumors had the variant allele versus 25 of 54 patients (46%) with ER negative tumors (p = 0.001), and 104 of 160 individuals (65%) with PR positive tumors versus 41 of 82 subjects (50%) of those with PR negative tumors (p = 0.001). Furthermore, the AG and AA genotypes were more frequent among patients with well to moderately differentiated tumors (76 of 116 individuals (66%) versus 49 of 96 subjects (51%) with poorly differentiated tumors; p = 0.001). The variant allele was also more frequent among LN negative patients when compared with LN positive patients (81 of 125 patients (65%) with no positive LN versus 65 of 123 individuals (53%) with ≥1 LN; p = 0.033). No significant differences were observed with respect to primary tumor size and variant allele status did not influence disease-free survival in this cohort. Multivariate analysis that included age at diagnosis as covariate confirmed these results (Table 2), which again demonstrated that the variant allele was associated with a number of indicators of improved outcome.
RRP1B SNP analysis was performed in a second small pilot cohort consisting of 248 surgical breast cancer patients (58% African-American, 42% Caucasian) from the greater Baltimore area (Table S11) to attempt to replicate the findings of the initial cohort study. Stratification on race/ethnicity and age at disease onset in the Baltimore cohort indicated that neither of these variables was a significant confounding factor. Consistent with the Orange County cohort, the variant A allele was less frequent in patients with a high stage or poor grade tumor, with ER negative or PR negative tumors, and with a LN positive disease (Table 3). Most associations between the A allele and tumor markers were best explained by assuming an additive effect of the variant allele (Tables 2 & 3), however, studies in larger populations are required to better define the relative effect of the variant allele on outcome markers in breast cancer. We also examined the association between the 1421G→A SNP and breast cancer survival by assuming a dominant effect of the variant allele on survival, a model that best reflects our survival data. Carriers of the variant allele had a significantly better breast cancer-specific survival compared to homozygous carriers of the common allele (Figure 5). Multivariate Cox regression analysis with adjustments for age at diagnosis, race, ER status, tumor–node–metastasis (TNM) stage, and chemotherapy, further confirmed this observation. Patients who carried the variant allele had improved survival when compared to patients with the G/G genotype (hazard ratio of death (HR) = 0.46; 95% confidence interval (CI) = 0.21–0.97). This effect was stronger among patients with an ER positive tumor (HR = 0.17; 95% CI = 0.04–0.70) suggesting that RRP1B may have a particular function in the ER pathway.
The Pro436Leu SNP was selected for analysis in the pilot epidemiology analysis due to its potential effect on RRP1B function as the result of the nonsynonymous amino acid substitution. The pilot epidemiology data, while consistent with the possibility of role of RRP1B in breast cancer progression, does not distinguish between causal polymorphisms and polymorphisms in high linkage disequilibrium (LD) with the causal variant. To gain a better understanding of whether the Pro436Leu SNP might be the causal variant effecting RRP1B function, haplotype analysis was performed using the publicly available HapMap data (http://www.hapmap.org/downloads/index.html.en). rs9306160 was genotyped on the 30 CEPH trios (90 samples) used in the HapMap project and the linkage disequilibrium (LD) structure determined. Analysis of the current version of the HapMap revealed that the Pro436Leu SNP was in a large haplotype block spanning 212,658 bp, with 40 SNPs in high LD with rs9306160 (r2 > 0.8). The haplotype block encompasses only two genes, RRP1B and HSF2BP. Among the 673 SNPs in this region rs9306160 is the only missense polymorphism within this haplotype block based on RefSeq annotation (http://www.ncbi.nlm.nih.gov/RefSeq/), and therefore remains an interesting candidate for the causative polymorphism. However, we cannot exclude at this time the possibility that the causative effect is due to a different linked polymorphism. Further analysis at the epidemiology and molecular level will be required to resolve this question.
The diverse techniques employed in this study, including genetical genomics [21], functional genomics, sequence analysis, and molecular epidemiology, have allowed us to identify Rrp1b as a candidate for both a tumor progression and metastasis modifier in mice, and a marker of inherited breast cancer metastasis susceptibility in humans. Furthermore, functional data provides evidence that Rrp1b regulates ECM gene expression, and that a nonsynonymous SNP in RRP1B is associated with tumor progression and disease-specific survival in pilot epidemiology experiments.
Rrp1b was identified through two distinct experimental approaches designed to address two independent questions: (a) what is the molecular mechanism(s) by which Sipa1 modulates metastasis, and (b) what drives ECM dysregulation in metastasis-prone primary tumors? Yeast two-hybrid assays and functional genomic studies addressed the first question, which identified Rrp1b as binding to the polymorphic PDZ domain of Sipa1. The second question probed the origins of metastasis-predictive gene expression signatures. ECM genes are components of all metastasis-predictive gene expression signatures in both humans [9–11,22] and mice [12,13], a finding that may well be explained in part by constitutional variation [13]. To test this question, we examined whether ECM eQTLs and metastasis modifiers might be the same entities by analyzing ECM gene expression in a RI mouse panel. This led to the identification of several eQTLs, with a locus on proximal chromosome 17 displaying the strongest linkage. The peak linkage region of this locus encompasses both Rrp1b and a tumor growth and progression QTL [6], and rather remarkably, when we examined expression of transcripts within the peak eQTL linkage region, Rrp1b was highly correlated with metastasis-predictive ECM gene expression. Taken together, these data suggest that Rrp1b is a potential dual ECM and tumor progression candidate.
Further experimentation demonstrated that ectopic expression of Rrp1b in two highly metastatic mammary tumor cell lines and a mouse fibroblast cell line modulates ECM gene expression, a finding concurrent with our initial hypothesis that Rrp1b is indeed the chromosome 17 ECM eQTL. It should be noted however, that the directionality of the ECM expression changes observed in response to Rrp1b activation in these cell lines are not directly comparable to the changes in ECM expression observed in metastasis-predictive gene expression signatures [9–13]. Primary tumors are composed of a variety of cell and tissue types, and this cellular and microenvironmental heterogeneity is not accurately reflected by in vitro growth conditions of single cell types. Nevertheless, we do argue that our in vitro experimentation demonstrates that Rrp1b modulates the expression of metastasis-predictive ECM genes in a variety of individual cell lines. It is also evident that further work will be necessary to define the complex microenvironmental relationships that modulate ECM gene expression in bulk tumor tissue and their relationship to overall levels of Rrp1b activation.
In addition, it should be noted that eQTL analysis is most commonly associated with expression changes in normal tissues, and not the neoplastic tissues analyzed in this study. However, eQTLs owe their existence to germline polymorphism, and such variation will be present in tumor tissue in addition to normal tissues. With this in mind, it is therefore not unreasonable to assume that the phenotypic effects of eQTLs will be observed in tumors as well as normal tissue. Yet neoplastic tissues possess an inherent genomic instability, and it is therefore possible that variations in tumor gene expression patterns could also arise from somatic mutation. However, we have demonstrated that eQTLs can be genetically mapped in tumors, which suggests either that similar somatic mutations consistently occur in the majority of the tumor tissue in a subset of the RI strains, or that the observed eQTLs result from inherited polymorphisms. Our previous demonstration that the same ECM genes used to define tumor eQTLs are differentially expressed in normal mammary tissues derived from high- and low-metastatic mouse genotypes [13] suggests that such differential expression may be partially regulated by germline polymorphism. At this stage, however, we cannot formally dismiss a role for somatic mutation in ECM gene expression variation within RI mammary tumors.
We also described a number of important functional differences in Rrp1b between the progenitors of the AKXD RI panel, the high metastatic AKR/J, and low metastatic DBA/2J genotype mice. Specifically, the proximal promoter of Rrp1b in AKR/J mice contained polymorphisms that may reduce Rrp1b expression. This difference in promoter activity is one possible explanation for the observed differences in Rrp1b expression in the normal mammary tissue of these mice, with the high metastatic capacity AKR/J mouse having significantly lower levels of Rrp1b activity than the low metastatic DBA/2J mouse. Furthermore, ectopic expression led to reduced metastatic potential and primary tumor growth following tumor cell implantation into mice. In combination, these observations suggest, at least in this mouse model that increased Rrp1b expression correlate with a better outcome.
We have used a dual approach to try to address the importance of RRP1B in human breast cancer progression. The first of these approaches was to use data derived from microarray expression analysis of the Mvt-1/Rrp1b cells to address one of the central goals of our research: the translation of experimental data from mouse models of human breast cancer into potentially clinically relevant observations. Based on the work of Bild et al. [20] we identified an RRP1B gene expression signature and demonstrated it that predicts outcome in a publicly available and well-characterized breast cancer cohort [10]. This gene signature not only held strong prognostic value in the Dutch study cohort [10], but was also able to stratify those patients with ER positive tumors and LN negative disease at presentation into high and low risk categories. There has been significant interest in using gene expression profiles for improved patient stratification [23,24] in the clinic since it raises the possibility of improvements in breast cancer subtype classification, which in turn could enable clinicians to tailor treatment to individual patients. Whether the RRP1B signature proves of clinical value in this respect is at present unclear, and further testing of its prognostic value in different cohorts will be required to address this possibility. The significance of this study, however, is not the identification of yet another prognostic signature, but the fact that the underlying casual element is known. Identification of other genetic elements that drive the predictive gene expression patterns may provide a more robust means of complementing currently available tests used for the assessment of prognosis in breast cancer. Furthermore, this type of study also provides us with potentially novel and important insights into the mechanisms underlying the metastatic process.
Further supportive evidence for the role of RRP1B in human breast cancer progression was evident in the two pilot epidemiological studies, both of which found an inverse relationship between the variant A allele of the 1421G→A RRP1B SNP and poor outcome markers. These consistent findings indicate that 1421G→A is a marker for disease progression, and patients who carry the A allele are less likely to present with advanced disease than homozygous carriers of the more common G allele. These data are consistent with the results of functional analysis of Rrp1b, and associate RRP1B with disease outcome in human breast cancer. It should be noted that the variant 1421G→A allele was associated with improved outcome in those individuals with ER positive tumors, which may permit better stratification of patients who are currently thought to be in a low risk category. A particularly intriguing question is if and how a patient's 1421G→A genotype affects expression of the 33-gene RRP1B expression signature, and whether polymorphisms in the promoter of RRP1B in linkage disequilibrium with the 1421G→A SNP are more important in this respect. These studies are currently ongoing in this laboratory. Indeed, a link between constitutional polymorphism and bulk tumor gene expression would be particularly significant given the technical difficulties associated with tumor gene expression profiling and the relative ease of SNP genotyping. These results, while consistent between the studies and in support of our hypothesis, must be considered only as preliminary. Further investigations in larger epidemiology studies specifically designed to address tumor progression and outcome, rather than tumor incidence, will be necessary to gain further support for the role of germline polymorphism in RRP1B in breast cancer progression.
Several differences were evident between the epidemiology study cohorts, the most notable of which was that the 1421G→A SNP was associated with breast cancer-specific survival only in the Baltimore cohort, a discrepancy that is likely due to several factors. First, both cohorts are relatively small (n < 300), thus some differences might arise from statistical power issues. More importantly, the study population compositions differed: the Orange County cohort was derived from a population-based case-cohort study, including all cancer patients, regardless of stage, with more than 10 years of follow up, whereas the Baltimore cohort is a surgical breast cancer population, and therefore is biased against patients with metastatic disease at presentation. Furthermore, unlike the Orange County cohort, the Baltimore cohort contained African-American and Caucasian women. Differences in race/ethnicity influence allele frequency and disease outcome, and some variability in results is therefore expected. While the participation of patients from different race/ethnicities also strengthens study design, stratification is required. Stratification on race/ethnicity in the Baltimore cohort indicated that race/ethnicity was not a significant confounding factor. It is interesting to note, however, that the variant allele distribution is different in African-Americans and Caucasians in the Greater Baltimore cohort (0.099 in African-Americans versus 0.411 in Caucasians). It is known that African-American women have a poorer prognosis compared to other breast cancer patients [25], and given the protective effect exerted by the variant 1421G→A, it is interesting to speculate that polymorphisms in genes such as RRP1B may be driving these ethnicity-specific differences in outcome. Thus, further characterization of this SNP in larger and ethnically diverse cohorts is required to determine the influence of race/ethnicity upon the association between this SNP and breast cancer survival.
The function of Rrp1b and its exact role in metastasis remain unclear at this time. Sipa1 was originally cloned as a mitogen-inducible protein [26] that was subsequently shown to be a negative regulator of Rap1 by serving as a GAP for Rap1 [27]. Sipa1 has significant effects on cellular adhesion [28], primarily related to its effects on Rap1, which has been implicated in maintaining the integrity of polarized epithelia [29] and intercellular adherens junctions [30], and potentially integrating signaling between cadherins and integrins [31]. Rrp1b may therefore mediate tumor cell adhesion properties by altering intercellular and cell–ECM contacts in a Sipa1-dependent and Rap1-dependent manner. It should be noted, however, that the human polymorphism in RRP1B falls outside of the domain that directly interacts with the PDZ domain of Sipa1. Whether this polymorphism impacts the enzymatic function of Sipa1 or mediates metastatic potential through some other mechanism is unclear and currently under investigation. Similarly, it is unclear whether the amino acid substitution in human RRP1B directly affects function. Based on the mouse model, where increased expression confers protection against malignant progression, the variant leucine in the human ortholog may phenocopy the mouse situation by activating some function of RRP1B. Further in vitro analysis will be required to clarify the different situations in the two species and is currently under investigation in our laboratory.
In addition to the negative regulatory role on Sipa1 function, there is some evidence to suggest that Rrp1b may be involved in RNA metabolism. Protein homology analysis has shown Rrp1b to contain a Nop52 domain, a motif found in proteins critical to 28S rRNA generation [32], and a previous yeast two-hybrid analysis has shown that Rrp1b may also interact with Lsm1, which is a protein involved in regulation of mRNA degradation [33]. Publicly available databases show that RRP1B is ubiquitously expressed at a somewhat low level, although it is expressed at a slightly higher level in lymph nodes in humans (http://smd.stanford.edu/cgi-bin/source/sourceSearch). Differential expression of RRP1B has been reported in fibroblasts from patients with systemic sclerosis, an autoimmune disorder characterized by dysregulation of a variety of ECM genes, including procollagens I, III, and VI [34], consistent with our results. Further research, however, is clearly needed to fully explore the role of Rrp1b and Sipa1 in human breast cancer and other tumor types. Unraveling the mechanisms of action and the molecular pathways that they regulate are likely to provide novel and valuable insights into tumor dissemination and metastasis.
Yeast two-hybrid screens using different regions of the human Sipa1 protein (Entrez Gene ID No: 6494) as bait were performed by ProNet technology (Myriad Genetics, Salt Lake City, UT). Methodology for these commercially performed experiments is provided by Myriad Genetics and is available in Text S2.
The various genes were cloned into pcDNA3. HEK293 (293) cells were transiently transfected with lipofectamine (Invitrogen) according to the manufacturer's instructions. They were cotransfected with pcDNA3 vector or mouse Sipa1-V5 from DBA [7], and Rrp1b-V5, Rrp1b-HA, or AQP2 (from American Type Culture Collection). Two days after transfection, cells were lysed with Golden Lysis Buffer (GLB) containing 20 mM Tris (pH 7.9), 137 mM NaCl, 5 mM EDTA, 1 mM EGTA, 10 mM NaF, 10% Glycerol, 1% Triton X-100, 1 mM sodium pyrophosphate, 1 mM leupeptin, 1 mM PMSF, and aprotinin (10 μg/ml). Cell extracts were immunoprecipitated with anti-Spa-1 (Sipa1) mAb (BD BioSciences), and protein A/G Sepharose (Pierce) were added and rotated overnight at 4 °C. The immune complexes were washed once with GLB, once with high salt HNTG (20 mM Hepes, 500 mM NaCl, 0.1% Triton X-100, 10% Glycerol), and twice with low salt of HNTG (20 mM Hepes, 150 mM NaCl, 0.1% Triton X-100, 10% Glycerol). The immune complexes were then analyzed by immunoblotting with antibodies against AQP2 (Santa Cruz), V5 (Invitrogen), or HA (Convance). For each immunoblot, horseradish peroxidase-conjugated anti-rabbit, anti-mouse, or anti-goat immunoglobulin G was used for the second reaction at a 1:10,000 dilution. Immunoblots were visualized by enhanced chemiluminescence with an ECL Kit (Amersham).
293 cells were transiently cotransfected with pcDNA3 vector or mouse Sipa1 and Rrp1b-V5 or AQP2. Epac-HA was also cotransfected, to elevate the level of Rap1-GTP, and pcDNA3 vector was added as necessary to ensure that equal amounts of DNA were transfected. Transfected cells were processed two days later, using a Rap1 activation Kit (Upstate Biotech) according to the manufacturer's instructions. Equal amounts of the total protein from cell extracts were estimated based on BCA protein assay kit (Pierce). Rap1-GTP protein was pulled down by RalGDS-RBD beads and washed three times, then subjected to gel analysis and immunoblotting using anti-Rap1 antibody (Santa Cruz). Cell extracts from transfectants were analyzed for protein expression by immunoblotting, using anti-AQP2 antibody, anti-Spa-1 mAb, or anti-V5 antibody.
The Mvt-1 and 4T1 cell lines were obtained as a gift from Lalage Wakefield (National Cancer Institute, Bethesda). The cells were cultured in Dulbecco's Modification of Eagle's Medium (DMEM; Cellgro, VA) containing 10% fetal bovine serum (FBS; Cellgro, VA), with culture medium being replaced at three day intervals. When the cells achieved confluency, they were washed once with 5 ml phosphate-buffered saline (PBS), incubated with 2 ml trypsin-EDTA for 5 min, and passaged at a 1:30 dilution into a fresh culture flask. NIH-3T3 cells were maintained in the same manner as Mvt-1, except cells were passaged at a 1:15 ratio when they achieved confluency.
Microarray hybridization methodology and generation of the microarray expression data from AKXD × PyMT primary tumors has been described previously [13]. Affymetrix .CEL files were normalized using the RMA method, averaged for each AKXD RI strain, and loaded into the GeneNetwork web service (http://www.genenetwork.org) [15]. The database was then searched for the 11 probe sets from our previously described metastasis signature profile [13] classification as within an “ECM component” (1437568_at, Mmp16; 1418454_at, Mfap5-pending; 1416168_at, Serpinf1; 1425896_a_at, Fbn1; 1416625_at, Serping1; 1450798_at, Tnxb; 1439364_a_at, Mmp2; 1427884_at, Col3a1; 1416808_at, Nid1; 1423407_a_at, Fbln2; and 1420924_at, Timp2). Additionally, a further five probe sets for the ECM genes represented the human breast carcinoma metastasis gene signature profile described by Ramaswamy et al. [9] were also included (Col1a1: 1423669_at_A, 1455494_at_A and Col1a2: 1423110_at_A, 1446326_at_B, 1450857_a_at_A). eQTLs were defined as described above.
An expression vector encoding the full length Rrp1b cDNA BC016569 in pCMV-SPORT6 was obtained from the Mammalian Gene Collection (MGC:27793, IMAGE ID: 3157173). The control cell line was generated using the vector pCMV-SPORT-β-Galactosidase (Invitrogen). The identity of the vector was sequence verified before transfection. Supercoiled plasmids were transfected into Mvt-1 and 4T1 cells using Superfect Transfection Reagent (Qiagen, Valencia, CA) as per the manufacturer's instructions. Briefly, transfections were performed in 100 mm diameter culture dishes, with 2 ×106 Mvt-1 or 4T1 cells being seeded 24 h prior to transfection. The Rrp1b-pCMV-Sport6 and pCMV-SPORT-β-Galactosidase vectors were cotransfected with the vector pSuper.Retro.Puro (Oligoengine) containing no insert as a selectable marker for transfectants. Cells in each culture vessel were transfected with a total of 20 μg vector DNA using Superfect at a 6:1 lipid to DNA ratio. Twenty-four hours after transfection, the cells were selected in normal growth medium containing 10 μg/ml puromycin (Sigma Aldrich), transferred to 96 well plates, and individual clones were selected by limiting dilution. Colonies were screened by qPCR as described below to identify clones ectopically expressing Rrp1b.
With regards to the NIH-3T3 cell lines, PCR primers were designed to encompass the entire length of the Rrp1b cDNA BC016569. The following primer sequences generated a 2,248 bp product from normal mammary tissue cDNA, with the downstream primer being designed to omit the transcription termination codon and to remain in coding frame: 5′-CCCATACGCAGACGCAGT-3′ and 5′-GAAGAAGTCCGCAGCCCT-3′. Full length Rrp1b cDNA was then amplified using rTth DNA Polymerase, XL (Applied Biosystems) as per the manufacturer's protocol. Following PCR amplification, Rrp1b cDNA was inserted into the reporter vector pcDNA3.1 V5-His using a pcDNA3.1 V5-His TOPO® Cloning Kit (Invitrogen) and transformed into TOP10 competent cells (Invitrogen) as per the manufacturer's protocol. Plasmids were propagated in 100 ml LB Medium containing 100μg/ml ampicillin. Plasmid DNA was extracted using a Qiagen EndoFree Maxi Kit, and insert identity and integrity of the full-length sequence was verified prior to further experimentation. Transfection of this vector into NIH-3T3 cells was performed in the same manner as for Mvt-1 cells except that the Rrp1b construct was cotransfected with the puromycin selection marker pPur (Clontech), and selection of transformants performed using 5 μg/ml puromycin and 700 μg/ml G418 (Sigma Aldrich).
Total RNA samples were isolated from cell culture samples using an RNeasy Mini Kit (Qiagen) with sample homogenization being performed using a 21 gauge needle and syringe as per the manufacturer's protocol. All samples were subjected to on-column DNase digestion, and RNA quality and quantity determined by an Agilent Technologies 2100 Bioanalyzer (Bio Sizing Software version A.02.01, Agilent Technologies). Only those samples containing high-quality total RNA with A260/A280 ratios between 1.8 and 2.1 were used for further analysis.
cDNA was synthesized from RNA isolated from either primary tumor tissues or transfected cell lines using the ThermoScript RT-PCR System (Invitrogen, Carlsbad, CA) by following the manufacturer's protocol. Single RT-PCRs were performed for each Mvt-1 and 4T1 clonal isolates, and in triplicate for the untransfected or Rrp1b-expressing NIH-3T3 total RNAs. SYBR Green qPCR was performed to detect the cDNA levels of Rrp1b and a variety of metastasis predictive ECM genes (see above) using an ABI PRISM 7500 and/or 7900HT Sequence Detection Systems and custom designed primers (Table S12). Reactions were performed using QuantiTect SYBR Green Master Mix (Qiagen, Valencia, CA) as per the manufacturer's protocol. The cDNA level of each gene was normalized to peptidylprolyl isomerase B (Ppib) cDNA levels using custom-designed primers for SYBR green-amplified target genes (Table S12).
Complete sequencing of the exons, intron-exon boundaries, promoters, and the regions immediately upstream of the promoters was performed in two highly metastatic (AKR/J, FVB/NJ) and three low metastatic (DBA/2J, I/LnJ, NZB/B1NJ) strains of mice [35]. The sequences of the primers for Rrp1b are shown in Table S4. PCR products were generated under standard amplification conditions (5 min at 94 °C, 30 s at 57 °C, 30 s at 72 °C, and 5 min at 72 °C), purified with a Qiagen PCR purification kit, and double strand sequencing performed with a Perkin Elmer BigDye Terminator sequence kit. Analysis was performed on a Perkin Elmer 3100 Automated Fluorescent Sequencer. Sequences were compiled and analyzed with the computer software package VectorNTI [36].
PCR primers were designed to encompass the two proximal promoter polymorphisms identified by sequencing Rrp1b. The following primer sequences generated a 1,672 bp product with AKR/J genomic DNA and a 1,670 bp product with DBA/2J: 5′-AACCTCATCGTCCCTTGG-3′ and 5′-GCACTCGCTTCAGCATCC-3′. Proximal promoter sequences were amplified using rTth DNA Polymerase, XL (Applied Biosystems) as per the manufacturer's protocol. Following PCR amplification, proximal promoter sequences were inserted into the reporter vector pBlue TOPO using a pBlue TOPO® TA Expression Kit (Invitrogen) and transformed into TOP10 competent cells (Invitrogen) as per the manufacturer's protocol. Plasmids were propagated in 100 ml of Luria-Bertani Medium containing 100 μg/ml ampicillin. Plasmid DNA was extracted using a Qiagen EndoFree Maxi Kit. AKR/J and DBA/2J promoter constructs were sequence verified prior to further experimentation as described above.
Supercoiled plasmids were transfected into NIH-3T3 cells using Superfect Transfection Reagent (Qiagen, Valencia, CA) as per the manufacturer's instructions. Briefly, transfections were performed in 100 mm diameter culture dishes, with 2 × 106 NIH-3T3 cells seeded 24 h prior to transfection with 10μg vector DNA (8μg Rrp1b promoter construct and 2μg of pGL3-Control (Promega)) using Superfect at a 5:1 lipid to DNA ratio. Transfections were performed in triplicate for each promoter construct. Twenty-four hours after transfection, the cells were washed with PBS, trypsinized, collected, and washed twice with cold PBS. Cell lysis was achieved using 100 μl of lysis buffer per 100 ml plate from a β-Galactosidase Assay Kit (Invitrogen). β-galactosidase activity in each sample was assayed as per the manufacturer's protocol using either 20 μl or 30 μl of the lysate. The remaining lysate was used to determine lysate protein concentration, which was assayed using a BCA Assay Kit (Pierce). β-galactosidase activity was calculated for each sample as per the β-Galactosidase Assay Kit (Invitrogen), but is essentially based upon the concentration and A562 of each lysate. These specific activities were normalized against firefly luciferase activity as driven by the pGL3 control vector to account for differing transfection efficiencies. The activity of firefly luciferase for each sample was assayed using a Dual-Luciferase Reporter Assay Kit (Promega) and quantification performed by VICTOR2 (Perkin-Elmer Life Sciences).
Total RNA extractions from tissue samples were carried out using TRIzol® Reagent (Life Technologies, Inc.) according to the standard protocol. RNA quantity and quality were determined by the Agilent Technologies 2100 Bioanalyzer (Bio Sizing Software version A.02.01, Agilent Technologies) and/or the GeneQuant Pro (Amersham Biosciences). Samples containing high-quality total RNA with A260/A280 ratios between 1.8 and 2.1 were purified with the RNeasy Mini Kit (Qiagen). An on-column genomic DNA digestion was performed as part of this purification step using the RNase-Free DNase Kit (Qiagen). TaqMan qPCR was performed to detect the cDNA levels of Rrp1b using an ABI PRISM 7500 Sequence Detection System (Applied Biosystems, Foster City, CA). Rrp1b expression was quantified using the Applied Biosystems Assay-On-Demand Mm00551206_m1. The housekeeping gene Peptidylprolyl Isomerase B (Ppib), used for normalization of Rrp1b expression between samples, was quantified using the primers described in Table S12 and the following fluorogenic probe: 6FAM-TCTATGGTGAGCGCTTC-MGB. Reactions were performed using TaqMan Universal PCR Mastermix (Applied Biosystems) per the manufacturer's protocol.
Transfected cells proven to be stably expressing Rrp1b were subcutaneously implanted into virgin FVB/NJ mice. Two days before injection, cells were passaged and permitted to grow to 80%–90% confluence. The cells were then washed with PBS and trypsinized, collected, washed twice with cold PBS, counted in hemocytometer, and resuspended at 106 cells/ml. One hundred thousand cells (100 μl) were injected subcutaneously in the vicinity of the fourth mammary gland of 6-wk-old virgin FVB/NJ female mice. The mice were then aged for 4 wk before they were killed by anesthetic overdose. Tumors were dissected and weighed. Lungs were isolated and surface metastases enumerated using a dissecting microscope. Tumor growth and metastasis was compared to mice injected with 105 Mvt-1 cells stably cotransfected with pCMV-Sport-β-Gal and pSuper.Retro.Puro. These experiments were performed in compliance with the National Cancer Institute's Animal Care and Use Committee guidelines.
The Rrp1b 1421G→A polymorphism was characterized using SNP-specific PCR. PCR primers were designed using Primer Express software (Applied Biosystems) according to parameters described elsewhere [37]. Each probe was labeled with a reporter dye (either VIC®, a proprietary fluorescent dye produced by Applied Biosystems, or FAM, 5-(&6)-carboxyfluorescein) specific for the wild-type and variant alleles of the Rrp1b SNP. Sequences of PCR primers are as follows: 5′-TGGACGTGGCCTCTGCAC-3′ and 5′-CACCACCTGCAGCCTGAAA-3′; and the sequences of fluorogenic probes are as follows: 6FAM-AGGGCTTTCAGCCCAGAG and VIC-AGGGCTTTCGGCCCAG. Reaction mixtures consisted of 300 nM of each oligonucleotide primer, 100 nM fluorogenic probes, 8 ng template DNA, and 2× TaqMan Universal PCR Master Mix (Applied Biosystems, Foster City, CA) in a total volume of 10 μl. The amplification reactions were performed in a MJ Research DNA Engine thermocycler (Bio-Rad, Hercules, CA) with two initial hold steps (50 °C for 2 min, followed by 95 °C for 10 min) and 50 cycles of a two-step PCR (92 °C for 15 s, 60 °C for 1 min). The fluorescence intensity of each sample was measured post-PCR in an ABI Prism 7500 sequence detection system (Applied Biosystems, Foster City, CA), and Rrp1b SNP genotypes were determined by the fluorescence ratio of the nucleotide-specific fluorogenic probes.
Total RNA was extracted using TRIzol Reagent (Life Technologies) according to the standard protocol. Total RNA samples were subjected to DNase I treatment, and sample quantity and quality determined as described above. Purified total RNA for each clonal isolate were then pooled to produce a uniform sample containing 8 μg of RNA.
Double stranded cDNA was synthesized from this preparation using the SuperScript Choice System for cDNA Synthesis (Invitrogen) according to the protocol for Affymetrix GeneChip Eukaryotic Target Preparation. The double stranded cDNA was purified using the GeneChip Sample Cleanup Module (Qiagen). Synthesis of biotin-labeled cRNA was obtained by in vitro transcription of the purified template cDNA using the Enzo BioArray High Yield RNA Transcript Labeling Kit (T7) (Enzo Life Sciences). cRNAs were purified using the GeneChip Sample Cleanup Module (Qiagen). Hybridization cocktails from each fragmentation reaction were prepared according to the Affymetrix GeneChip protocol. The hybridization cocktail was applied to the Affymetrix GeneChip Mouse Genome 430 2.0 arrays, processed on the Affymetrix Fluidics Station 400, and analyzed on an Agilent GeneArray Scanner with Affymetrix Microarray Suite version 5.0.0.032 software. Normalization was performed using the BRB-Array Tools software [38,39].
To generate a high confidence human transcriptional signature of Rrp1b expression, 98 probe sets whose differential expression demonstrated p < 10−7 were selected by matching the gene symbols from the mouse dataset to the published Hu25K chip annotation files. Analysis of tumor gene expression from breast cancer datasets was performed using BRB ArrayTools. Expression data were downloaded from the Rosetta Company website (http://www.rii.com/publications/2002/vantveer.html). Expression data were loaded into BRB ArrayTools using the Affymetrix GeneChip Probe Level Data option or the Data Import Wizard. Data were filtered to exclude any probe set that was not a component of the Rrp1b signature, and to eliminate any probe set whose expression variation across the data set was p ≥ 0.01.
Unsupervised clustering of each dataset was performed using the Samples Only clustering option of BRB ArrayTools. Clustering was performed using average linkage, the centered correlation metric and center the genes analytical option. Samples were assigned into two groups based on the first bifurcation of the cluster dendogram, and Kaplan-Meier survival analysis performed using the Survival module of the software package Statistica. Significance of survival analyses was performed using the Cox F-test.
Orange County, California. This patient population is a random sample of 269 probands successfully genotyped for the Rrp1b SNP that were diagnosed between March 1, 1994 and February 28, 1995 with invasive breast cancer. Probands were ascertained through the population-based Hereditary Breast Cancer Study funded by the National Cancer Institute. A description of the study and details of data collection methods have been reported previously [8,40]. Briefly, there were two case groups: cases with localized disease (n = 130), and cases with regional/metastatic disease (n = 139). The average age at diagnosis of this cohort was 55.9 ± 13.4 years, and the average body mass index was 25.7 ± 4.9.
Greater Baltimore, Maryland. Surgical breast cancer cases were recruited at the University of Maryland Medical Center, the Baltimore Veterans Affairs Medical Center, Union Memorial Hospital, Mercy Medical Center, and the Sinai Hospital in Baltimore, Maryland between February 15, 1993 and August 27, 2003. We collected blood, tissue specimens, and survival information from 248 patients. These patients had pathologically confirmed breast cancer, were of African-American or Caucasian descent by self-report, were diagnosed with breast cancer within the last six months prior to recruitment, and had, by self-report, no previous history of the disease. Patients were excluded if they were HIV, hepatitis B virus, or hepatitis C virus carriers, were intravenous drug users, were institutionalized, or were physically or mentally unable to sign consent and complete the questionnaire. Of the eligible patients that were identified through surgery lists, 83% participated in the study. The subjects signed a consent form and completed an interviewer-administered questionnaire. Additional information to determine the ER-α status, disease stage, treatment, and survival was obtained from medical records and pathology reports, the Social Security Death Index, and the National Death Index. Disease staging was performed according to the TNM system of the American Joint Committee on Cancer/ the Union Internationale Contre le Cancer (AJCC/UICC). The Institutional Review Boards at the participating institutions approved the study.
Student's t-tests were used to compare means for continuous variables and Wilcoxon's sum rank test to compare medians. Variables that were not normally distributed, such as tumor size, were log transformed. Chi-square test or Fisher's exact tests were used to test for differences between categorical variables and to test for Hardy-Weinberg equilibrium. Unconditional logistic regression adjusting for multivariate covariates, such as age at diagnosis, was used to estimate the adjusted odds ratios. We used likelihood ratio tests to calculate p-values comparing a model with covariates to a model without them. We used Cox regression models to perform survival analysis. All p-values presented are two-tailed and were considered to be statistically significant if they were below 0.05. The Kaplan-Meier method and the log-rank test were used for univariate survival analysis and Cox regression models were used to perform multivariate survival analysis. For the Baltimore cohort, survival was determined for the period from the date of hospital admission to the date of the last completed search for death entries in the Social Security Index (August 18, 2004) for the 248 case patients. The mean follow-up time for breast cancer survival was 55 months (range: 12 to 140 months). A total of 59 (24%) of these 248 patients died during this period. We obtained death certificates of the deceased case patients and censored all causes of death that were not related to breast cancer, such as accidents, violent crimes, stroke, heart attack, and liver cirrhosis, in our analysis.
The genotype data of the 90 HapMap CEPH trio in the 250 kb flanking region of rs9306160 was downloaded from the HapMap database (http://www.hapmap.org). A total of 304 SNPs with minor allele frequency ≥10% were selected to evaluate the LD structure. The CEPH genotype data we generated for rs9306160 were integrated with the HapMap data in this analysis. Using LDSelect [41] with a cut of r2 ≥ 0.64 we found a large LD block spanning a 210 kb genomic region (chr21:43739961–43952618) that has 104 SNPs in high LD with rs9306160 (Figure 6). In this 210 kb region, there are a total of 673 dbSNP markers and rs9306160 is the only nonsysnonymous SNP based on RefSeq annotation.
Accession numbers for genes mentioned in this paper from the Mammalian Gene Collection (http://mgc.nci.nih.gov/) are Rrp1b cDNA BC016569 (MGC:27793, IMAGE ID: 3157173).
Accession numbers for genes mentioned in this paper from the Entrez Gene database (http://www.ncbi.nlm.nih.gov/sites/entrez?db=gene) are β-galactosidase (957271), Col1a1 (12842), Col3a1 (1281), Col6a2 (12834), Fbln2 (14115), Fbn1 (14118), Mfap5 (50530), Ppib (19035), Rrp1b (72462), RRP1B (23076), Sipa1(20469), Serpinf1 (20317), and Serping1(12258).
Accession numbers for probe sets mentioned in this paper from the Affymetrix NetAffx Analysis Center (http://www.affymetrix.com/analysis/index.affx) are Col3a1 (1427884_at), Col1a1 (1423669_at_A, 1455494_at_A), Col1a2 (1423110_at_A, 1446326_at_B, 1450857_a_at_A), Fbn1(1425896_a_at), Fbln2 (1423407_a_at), Mfap5-pending (1418454_at), Mmp2 (1439364_a_at), Mmp16 (1437568_at), Nid1 (1416808_at), Serpinf1(1416168_at), Serping1 (1416625_at), Timp2(1420924_at), and Tnxb (1450798_at).
Accession number for the polymorphism mentioned in this paper from the Entrez SNP database (http://www.ncbi.nlm.nih.gov/sites/entrez?db=Snp) is Rrp1b (dbSNP ID: rs9306160; 1421G→A, Pro436Leu) |
10.1371/journal.pgen.1000273 | Gamma-Linolenic and Stearidonic Acids Are Required for Basal Immunity in Caenorhabditis elegans through Their Effects on p38 MAP Kinase Activity | Polyunsaturated fatty acids (PUFAs) form a class of essential micronutrients that play a vital role in development, cardiovascular health, and immunity. The influence of lipids on the immune response is both complex and diverse, with multiple studies pointing to the beneficial effects of long-chain fatty acids in immunity. However, the mechanisms through which PUFAs modulate innate immunity and the effects of PUFA deficiencies on innate immune functions remain to be clarified. Using the Caenorhabditis elegans–Pseudomonas aeruginosa host–pathogen system, we present genetic evidence that a Δ6-desaturase FAT-3, through its two 18-carbon products—gamma-linolenic acid (GLA, 18:3n6) and stearidonic acid (SDA, 18:4n3), but not the 20-carbon PUFAs arachidonic acid (AA, 20:4n6) and eicosapentaenoic acid (EPA, 20:5n3)—is required for basal innate immunity in vivo. Deficiencies in GLA and SDA result in increased susceptibility to bacterial infection, which is associated with reduced basal expression of a number of immune-specific genes—including spp-1, lys-7, and lys-2—that encode antimicrobial peptides. GLA and SDA are required to maintain basal activity of the p38 MAP kinase pathway, which plays important roles in protecting metazoan animals from infections and oxidative stress. Transcriptional and functional analyses of fat-3–regulated genes revealed that fat-3 is required in the intestine to regulate the expression of infection- and stress-response genes, and that distinct sets of genes are specifically required for immune function and oxidative stress response. Our study thus uncovers a mechanism by which these 18-carbon PUFAs affect basal innate immune function and, consequently, the ability of an organism to defend itself against bacterial infections. The conservation of p38 MAP kinase signaling in both stress and immune responses further encourages exploring the function of GLA and SDA in humans.
| Polyunsaturated fatty acids are vital for optimal physiological functions, including immunity. Much of these effects are mediated by eicosanoids, which are metabolites of arachidonic acid (AA) and eicosapentaenoic acid (EPA). In mammals, PUFAs cannot be synthesized de novo. They are produced from essential dietary fatty acids, which are first converted to gamma-linolenic acid (GLA) and stearidonic acid (SDA) by a rate-limiting step catalyzed by a Δ6-desaturase, FADS2. Activity of FADS2 is impaired under numerous conditions—including aging, diabetes, stress, and smoking—and could lead to reduced production of GLA and SDA. In this study, we examined the effects of loss-of-function mutations in PUFA biosynthetic genes on the ability of C. elegans to survive infection by the Gram-negative human pathogen P. aeruginosa. We show that the enhanced pathogen susceptibility of the C. elegans Δ6-desaturase mutant fat-3 is associated with decreased basal expression of immunity genes and disrupted activity of the p38 MAP kinase. These defects could be fully restored when both GLA and SDA, but not AA or EPA, were added into the diets of fat-3 mutants, further supporting the conclusion that GLA and SDA are required for basal immunity in C. elegans.
| Polyunsaturated fatty acids (PUFAs) are a class of long chain fatty acids of 18 carbon atoms or more in length that contain two or more double bonds. PUFAs are classified into two groups, the omega-6 (n-6) or the omega-3 (n-3) fatty acids, depending on the position of the double bond (n) closest to the methyl end of the fatty acid chain. In mammals, the 18-carbon and longer omega-6 and omega-3 PUFA families cannot be synthesized de novo. They are produced, instead, from the dietary essential fatty acids linoleic acid (LA, 18:2n6) and alpha-linolenic acid (ALA, 18:3n3) through a series of desaturation and elongation reactions catalyzed by desaturase and elongase enzymes, respectively [1],[2]. Omega-6 PUFAs, such as arachidonic acid (AA, 20:4n6) are converted into eicosanoids, leukotrienes and prostanoids through the actions of lipoxygenase and cyclooxygenase enzymes [3]. In vertebrates, these eicosanoids variously exert stimulatory and inhibitory influences and have profound effects on multiple aspects of organismal physiology, including immunity [4],[5]. For example, prostaglandins and leukotrienes are pro-inflammatory mediators that are vital for the initial containment of an infection and for the recruitment of phagocytes and other immune cells to a site of infection [6]. Omega-3 fatty acids such as eicosapentaenoic acid (EPA, 20:5n3) influence the T cell response to infection and demonstrate strong anti-inflammatory effects [7]. Dietary fatty acids and eicosanoids have also been shown to bind nuclear receptors, such as the Peroxisome Proliferator Activated Receptor γ (PPAR-γ), which modulates activation of dendritic cells, NK cells and T cells [8]–[10]. The influence of PUFAs on immune functions also extends to other organisms that possess only an innate immune response. For example, eicosanoids are crucial mediators and coordinators of insect cellular immune reactions to bacterial, fungal, and parasitoid invaders, specifically microaggregation, nodulation, and encapsulation [11]. In the silkworm Bombyx mori, eicosanoids are involved in the expression of the antibacterial proteins cecropin and lysozyme in the fat body [12]. In Drosophila, a functional coupling between eicosanoid biosynthesis and the IMD pathway for the induction of the antibacterial peptide diptericin by LPS has been reported [13],[14]. An analogous fatty acid-derived signaling pathway has also been shown to be important for defense in plant. However, instead of the 20-carbon AA, which is a minor PUFA in plants, 18-carbon PUFAs serve as major precursors for the synthesis of jasmonates and other oxylipins that play important roles in pathogen defense [15]. Jasmonates regulate the expression of defense genes that are essential for survival against insects and necrotrophic pathogens [16]. Oxylipins, such as crepenynic and dehydrocrepenynic acids are biologically active anti-fungal compounds [17].
Innate immunity forms a common first line of defense for most organisms, providing a highly conserved but generally non-specific response to pathogens and parasites. The innate immune system can be distinguished into two separate but overlapping components, the constitutive or basal branch of innate immune defense, and the pathogen-induced responses [18],[19]. Constitutive or basal immunity involves the constant production of effector molecules such as defensins and other antimicrobial peptides, providing a preventative barrier and allowing the organism to instantaneously respond to an immunological insult [20]–[22]. The inducible branch of the innate immune system, on the other hand, is only activated after the host has encountered a pathogen, and typically includes the induction of additional effector molecules, and where present, the recruitment and activation of phagocytic cells [23].
Both constitutive and inducible innate immunity has been described in the soil nematode C. elegans. For example, a number of antimicrobial peptides are constitutively expressed in healthy worms, including lysozymes [24], the ABF-2 defensin [25] and the SPP-1 saposin [24]. A subset of these constitutively-expressed antimicrobials, and a suite of additional effector molecules, including members of the C-type lectin family, are up-regulated at different time points after infection [26]–[28]. The ability of C. elegans to defend against infections requires several conserved signaling pathways [19],[29],[30]. They include a MAP kinase cascade, resulting in the activation of the p38 MAP kinase homologue PMK-1 [29], an insulin-like defense pathway that activates the FOXO transcription factor homologue DAF-16 [31] and a TGF-β pathway [26]. These pathways are required for the elevated production of a number of effector molecules, including antimicrobial peptides, lysozymes and lectins, to levels above those seen under basal conditions, in healthy worms [24],[28]. The p38 MAP kinase pathway also plays a vital role in maintaining the basal immune response, and mutants in this pathway, such as the p38 MAP kinase mutant pmk-1, show defects in the constitutive expression of lysozymes, lectins and other effector molecules [28].
In addition to the innate immune response, pathways for lipid synthesis and metabolism are also largely conserved in C. elegans [32]–[34], making the worm an ideal model to investigate the effects of lipids on immune function. Unlike mammals, C. elegans is able to synthesize all its required long chain fats from its bacterial food source (Figure 1A), allowing for the manipulation of lipid synthesis and content in the worm [32],[35]. Synthesis of these fatty acids is catalyzed by elongase and desaturase enzymes, which in C. elegans are encoded by elo and fat genes, respectively, and the C. elegans genome contains the full complement of enzymes required for the synthesis of long chain fatty acids (LCFAs) [32],[35]. The absence of obvious mammalian orthologs of cyclooxygenases and lipoxygenases or of prostanoid and leukotriene receptors in the C. elegans genome [36] provides the opportunity to investigate the roles for PUFAs in innate immunity that could otherwise be masked by the dominant influences of the prostaglandin and leukotriene eicosanoids.
Here, we use the infection of C. elegans by a human Gram-negative bacterial pathogen, Pseudomonas aeruginosa as an experimental system to investigate the interplay between PUFAs and innate immunity, in the context of the whole organism. We identify two long chain PUFAs, gamma-linolenic acid (GLA, 18:3n6) and stearidonic acid (SDA, 18:4n3), as vital for C. elegans defense against P. aeruginosa infection. Disrupting the production of these two fatty acids results in increased mortality following exposure to the pathogen. We demonstrate, by deficiency and exogenous supplementation studies, that GLA and SDA are required for both the basal activity of the p38 MAP kinase pathway and the basal expression of immunity genes.
Although lipids are known to play multiple roles in immunity, relatively little evidence exists for the specific manipulation of the lipid metabolism in response to infection. Detailed analysis of a whole genome microarray study for gene expression in P. aeruginosa-infected C. elegans [27] revealed an enrichment for genes required for the synthesis of LCFAs. This modulation of the lipid metabolism in response to infection hinted at potential roles for fatty acids in C. elegans immunity. To confirm and extend the microarray observations, we used quantitative real time PCR (qRT-PCR) to compare mRNA levels of 16 LCFA synthesis genes in age-matched adult animals raised on E. coli OP50-1, the standard laboratory food source, or following infection by P. aeruginosa strain PA14 (Figure 1B). Since the activity of lipid metabolism genes could be greatly influenced by available nutritional sources, and having determined that the fatty acid contents of E. coli and P. aeruginosa were different (Figure S1A), we also quantified the mRNAs of elo and fat genes in worms exposed to PA14ΔgacA, an isogenic strain of P. aeruginosa PA14 in which the global virulence gene gacA, has been deleted [37]. PA14ΔgacA mutants were highly attenuated in their ability to kill C. elegans (Figure S1B) [38], but had a fatty acid composition very similar to the parental PA14 strain (Figure S1A), thus providing a useful control for changes due to nutritional differences between E. coli and P. aeruginosa. Of the nine C. elegans genes known to be involved in the synthesis of the majority of 18- and 20-carbon PUFAs and monounsaturated fatty acids (MUFAs), fat-6, fat-2, fat-3 and fat-4 were expressed at higher levels in worms exposed to P. aeruginosa than to E. coli or PA14ΔgacA (Figure 1B). These represent genes whose expressions were significantly induced under infection conditions, indicating a modulation of PUFA synthesis by P. aeruginosa infection. Of the five elo genes of unknown function, elo-7 and elo-8 were also significantly induced under infection conditions (Figure 1B). Although expression of fat-5, elo-9 and the branched chain fatty acid (BCFA) biosynthetic genes (elo-5 and elo-6) were lower in animals exposed to P. aeruginosa compared to E. coli (Figure 1B), they were also lower in animals exposed to attenuated PA14ΔgacA, suggesting that these changes in expression may be due to differences in fatty acid content between the bacterial species. These results confirmed a specific modulation of host LCFA synthesis in response to P. aeruginosa infection.
We next determined the effect of infection on the abundance of specific fatty acids in the worm. Gas chromatography followed by mass spectrometry (GC-MS) [32] was used to identify and compare the content of individual species of LCFAs in age-matched P. aeruginosa-infected worms and worms grown on E. coli. To rule out changes caused by nutritional differences between the two bacterial species, we also determined LCFA content of worms exposed to PA14ΔgacA mutant bacteria. We conclude that the higher levels of vaccenic acid (VA, 18:1n7) in P. aeruginosa-infected worms is most likely due to nutritional differences between P. aeruginosa and E. coli because the same increase was seen in worms that were exposed to the relatively avirulent PA14ΔgacA (Figure 1C). This is consistent with GC-MS results showing that both the P. aeruginosa and PA14ΔgacA strains had more than twice the VA content compared to E. coli (Figure S1A). Worms infected with P. aeruginosa had significantly lower levels of stearic acid (SA, 18:0), oleic acid (OA, 18:1n9), LA, ALA and GLA compared to worms exposed to E. coli or the attenuated PA14ΔgacA strains (Figure 1C). These changes in PUFA content also corresponded with the previously observed infection-induced changes in gene expression. Three of the genes up-regulated in response to infection, fat-6, fat-2 and fat-3, are involved in the synthesis of fatty acids listed above, potentially indicating a feedback loop, where decreases in LCFA levels during infection could induce increased expression of corresponding biosynthetic genes. The infection-specific decreases in fatty acid levels led us to hypothesize that these LCFAs may be involved in immunity against pathogens.
To determine if specific LCFAs could be important for immune function in vivo, we analyzed a series of mutants that were unable to synthesize specific MUFAs and/or PUFAs for their ability to survive infection by P. aeruginosa (Figure 2A). These strong or complete loss-of-function mutants in the fat and elo genes were also analyzed by GC-MS to confirm that the genetic lesion or RNAi knockdown resulted in the expected alterations in the fatty acid profile (Figure 2A). The ELO-2 elongase is thought to catalyze the elongation of palmitic acid (PA, 16:0) to SA (Figure 1A). Reducing elo-2 expression by RNAi resulted in significant changes in LCFA profile that is consistent with a previous report [39] and a significant increase in susceptibility to killing by P. aeruginosa (Figure 2A). The next step in PUFA synthesis, the conversion of SA to OA is catalyzed by two functionally-redundant desaturases, encoded by fat-6 and fat-7 (Figure 1A) [40],[41]. We verified previous reports that neither the loss of fat-6 nor fat-7 function resulted in any significant alteration in PUFA composition, and showed that neither mutant was susceptible to P. aeruginosa infection (Figure 2A). By contrast, the fat-6(tm331); fat-7(wa36) double mutant, which lacked OA and the 18- and 20-carbon PUFAs derived from OA [41], was highly susceptible to killing by P. aeruginosa. fat-2(wa17) animals that lacked all 18- and most 20-carbon PUFAs were also significantly more susceptible to P. aeruginosa (Figure 2A). Loss of fat-3(wa22) function resulted in animals that lacked two specific 18-carbon PUFAs, GLA and SDA, as well as all the 20-carbon PUFAs. fat-3(wa22) animals were also significantly more susceptible to infection, suggesting that GLA, SDA and/or 20-carbon PUFAs that were missing in these animals could be vital for infection response in C. elegans (Figure 2A). Interestingly, two different elo-1 mutants, elo-1(gk48) and elo-1(wa7), that had decreased levels of all the 20-carbon PUFAs but accumulated the upstream 18-carbon precursors GLA and SDA [32], were significantly more resistant to killing by P. aeruginosa (Figure 2A). These results suggest that a lack of 20-carbon PUFAs does not compromise immune function. Consistent with this interpretation, neither the fat-1(wa9) mutant lacking two 20-carbon PUFAs, dihomo-γ-linolenic acid (DGLA, 20:3n6) and EPA, nor two fat-4 mutants, fat-4(wa14) and fat-4(ok958), that lacked AA and EPA, were significantly different from wild-type animals for susceptibility to P. aeruginosa (Figure 2A). Collectively, resistance of the elo-1 mutants and increased susceptibility of the fat-3(wa22) mutants to P. aeruginosa, suggest that GLA and SDA that accumulated in the elo-1 mutants but were absent in the fat-3(wa22) mutant may be required for immune function.
PUFA levels in fat and elo mutants could be restored through dietary supplementation of the missing fatty acids [42]. To confirm the requirement of GLA and SDA for C. elegans to survive a pathogen challenge, fat-2(wa17), fat-3(wa22) and elo-1(gk48) mutants were raised from embryos to 1-day-old adults in the presence of exogenously supplied PUFAs. A sub-population of these PUFA-supplemented adults was subjected to GC-MS to confirm that the procedure effectively restored the levels of the missing PUFAs (Figure S2) while the remaining population was subjected to survival assays. In parallel, wild-type worms were also supplemented with the respective PUFAs and the PUFA levels following supplementation were determined by GC-MS (Figure S2A). PUFA-supplemented wild-type animals were not significantly different from untreated wild-type for pathogen survival (Table 1, Figures S3A and S3B). Supplementation with ALA completely restored PUFA levels (data not shown) and survival of fat-2(wa17) animals on P. aeruginosa to that of wild-type (Figure 2B). ALA supplementation, however, failed to rescue fat-3(wa22) susceptibility to P. aeruginosa (Figure 2D, Table 1). These results were expected because the fat-3 gene remained functional in the fat-2(wa17) mutant and could convert the exogenously added ALA into the required downstream fatty acids, thus rescuing the immune defects of fat-2(wa17) animals. The fat-3(wa22) mutant, on the other hand, was unable to process the supplied ALA, and consequently remained susceptible to P. aeruginosa. We, therefore, conclude that ALA is not directly required for immune function. Instead, ALA is likely to be modified by the FAT-3 Δ6-desaturase enzyme into functional molecules that affect immune function.
Supplementation with either GLA or SDA that were absent in both the fat-2(wa17) and fat-3(wa22) mutants, resulted in a partial rescue of pathogen susceptibility (Figure 2B, C). GLA supplementation increased the mean survival period of both the fat-2(wa17) and fat-3(wa22) mutants to approximately 90% of wild-type. GLA supplementation limited to only the adult stage was also sufficient to partially rescue the susceptibility of the fat-3(wa22) mutant (Figure S3C), suggesting that the presence of GLA during growth and development is not necessary for its effect on the survival against P. aeruginosa infection. Addition of SDA alone increased the mean survival period of both fat-2(wa17) and fat-3(wa22) mutants on P. aeruginosa to approximately 85% of wild-type (Figure 2B, C). Supplementation with both GLA and SDA, however, completely rescued fat-3(wa22) survival against P. aeruginosa infection (Figure 2C, Table 1), indicating that GLA and SDA together are required for optimal infection response. We also note that, similar to the association between pathogen resistance and accumulation of GLA and SDA seen with elo-1 mutants (Figure 2A), wild-type and fat-3(wa22) animals supplemented with both GLA and SDA were marginally more resistant to P. aeruginosa, although these increases were not statistically significant (Table 1).
The P. aeruginosa-resistant elo-1(gk48) animals were also supplemented with ALA, GLA and SDA. The elo-1(gk48) mutant already accumulated these same fatty acids (Figure 2A), and additional supplementation did not enhance pathogen resistance (Figure S3D). Pathogen resistance of elo-1(gk48), coupled with wild-type phenotypes of the fat-1(wa9), fat-4(wa14) and fat-4(ok958) mutants on P. aeruginosa (Figure 2A), indicate that 20-carbon PUFAs are not necessary for C. elegans immunity. To further support these conclusions, we analyzed the effect of supplementation with AA and EPA on the fat-3(wa22) mutant. Dietary supplementation of either AA or EPA to fat-3(wa22) animals effectively restored the respective PUFAs in these animals (Figure S2C, S2D), but could not rescue pathogen sensitivity of the fat-3(wa22) mutant (Figure 2D), confirming that the 20-carbon PUFAs do not have any detectable roles in immune function. These results further implicate the requirement for both GLA and SDA in C. elegans immunity.
The fat-3 gene is expressed in multiple tissues, including the intestine, pharynx and body wall muscles, as well as some head and tail neurons [42]. To determine the tissue in which the fat-3 gene is required for immune function, we obtained transgenic strains that express a functional fat-3 gene only in the neurons, the muscles or the intestine of a fat-3 mutant using well-established tissue-specific promoters [36] and assayed the ability of these animals to survive P. aeruginosa infection. We first confirmed that the two deletion alleles of fat-3 used to generate the tissue-specific rescue strains, fat-3(lg8101) and fat-3(lg8101/qa1811) [36], demonstrated equivalent susceptibilities as fat-3(wa22) [32] to P. aeruginosa (Figure 3A). The fat-3; [Pfat-3::fat-3] strain, carrying a transgene consisting of the endogenous fat-3 promoter and the fat-3 coding region in the fat-3(lg8101) mutant [36], showed wild-type survival on P. aeruginosa, confirming that pathogen sensitivity is a direct consequence of loss of fat-3 function (Figure 3A, Table 2). fat-3; [Punc-119::fat-3] transgenic animals expressing the fat-3 coding sequence under the control of the neuron-specific unc-119 promoter, however, remained significantly more susceptible to P. aeruginosa (Figure 3B, Table 2). Expression of fat-3 under the control of the muscle-specific myo-3 promoter also failed to rescue the fat-3 mutant sensitivity to P. aeruginosa (Figure 3B, Table 2). Together, these results indicate that fat-3 gene expression and fat-3-dependent PUFA synthesis in the muscles or neurons does not significantly affect immune function. By contrast, fat-3; [Pelt-2::fat-3] transgenic animals expressing the fat-3 gene under the control of the intestine-specific elt-2 promoter showed survival kinetics on P. aeruginosa that were indistinguishable from wild-type (Figure 3B, Table 2). Intestine-specific rescue of fat-3 pathogen sensitivity indicates that fat-3-dependent synthesis of PUFAs in the intestine is sufficient for normal immune function in response to bacterial infection.
Animals that have lost fat-3 gene function display pleiotropic abnormalities, including impaired motility, a weakened cuticle, decreased defecation rate and irregular expulsion [36],[42]. Many of these defects are associated with impaired neurotransmission due the loss of fat-3 function in the neurons [36],[42]. We also found that fat-3(wa22) animals had a marginal but significant decrease in adult lifespan (Table 1), contrary to a previous report [42]. Together, these defects may indicate a general poor health of fat-3 mutants that could indirectly impact their ability to survive P. aeruginosa infection. To determine if these pleiotropies could be dissociated from immune defects, we first analyzed the effects of PUFA supplementation in fat-3(wa22) animals on these defects, in addition to survival on P. aeruginosa. We note that, with the possible exception of AA on adult life span, PUFA supplementations did not have any significant effects on wild-type animals (Table 1). Consistent with a previous report, supplementation with GLA [42] or GLA and SDA combined rescued the defecation and locomotion defects, and lifespan of fat-3(wa22) animals (Table 1). GLA and SDA, however, failed to rescue aldicarb resistance indicating that these PUFAs were not sufficient to restore synaptic transmission, as measured by acetylcholine release [43] (Table 1). By contrast, supplementation with either of the 20-carbon PUFAs, AA or EPA rescued fat-3(wa22) for all the phenotypes tested: adult lifespan, aldicarb resistance, defecation and locomotion defects (Table 1). Yet, fat-3(wa22) animals supplemented with either EPA or AA remained sensitive to P. aeruginosa (Figure 2D, Table 1). Failure to rescue the fat-3(wa22) immune defect was not due to insufficient incorporation of EPA or AA because the levels of these 20-carbon PUFAs in the fat-3(wa22) mutants following supplementation was equivalent to, or higher than, in wild-type (Figure S2C and S2D). Since AA or EPA could rescue fat-3(wa22) neuronal and muscular defects and adult lifespan but not pathogen sensitivity, while GLA and SDA rescued pathogen sensitivity, but not neurotransmission (Figure 2C and 2D, Table 1), we can conclude that pathogen susceptibility of the fat-3(wa22) mutant was not due to neuromuscular defects or a shortened adult lifespan. Instead, pathogen sensitivity of fat-3 mutants is likely to be caused by factors dependent on levels of GLA and SDA.
This conclusion is further supported by tissue-specific rescue experiments using transgenic animals. Intestinal expression of the fat-3 gene only partially rescued the locomotion defects (Table 2) despite completely rescuing the pathogen sensitivity of the fat-3(lg8101) mutant (Figure 3B, Table 2). Full rescue of the defecation defect and the partial rescue of locomotion by intestinal expression of fat-3 are not surprising because most FAT-3 protein is in the intestine [42] and it is therefore likely that PUFAs synthesized in the intestine could be transported to other parts of the body. As with pathogen sensitivity, muscle-specific expression of fat-3 was not sufficient to rescue defecation and movement defects (Table 2). By contrast, neuronal expression of the fat-3 gene that also failed to rescue pathogen susceptibility (Figure 3B), could fully rescue the defecation and locomotion defects of the fat-3(lg8101) mutant (Table 2), indicating neuromuscular and immune functions may be independently regulated by fat-3. Together, the PUFA supplementation and tissue-specific rescue experiments indicate that the susceptibility of fat-3 mutants to P. aeruginosa infection is not associated with neuromuscular and lifespan defects.
To further rule out the possibility that the increased pathogen sensitivity of fat-3(wa22) mutants was a consequence of a general increased sensitivity to any insults, we determined the ability of fat-3(wa22) animals to survive or develop under a number of additional stress conditions. We assayed the sensitivity of fat-3 animals to heavy metal stresses by determining the proportion of embryos that could develop into adults in the presence of toxic concentrations of cadmium or copper metals [44]. Exposure to toxic levels of cadmium results in cell damage and is thought to induce the transcription of a number of defense and repair genes [45]–[47]. Following exposure to 30 µM cadmium chloride, less than 65% of fat-3(wa22) embryos successfully developed into adults. By contrast, approximately 75% of wild-type embryos grew to adults, indicating that fat-3(wa22) animals were more sensitive to cadmium (Table 3). fat-3(wa22) animals were also more sensitive to copper, with significantly fewer fat-3(wa22) adults than wild-type developed from embryos following exposure to 250 µM copper sulfate (Table 3). The fat-3(wa22) mutant was also more susceptible to a 1% solution of the detergent Triton X-100. Approximately 26% of fat-3(wa22) animals survived a 1-hour incubation with the detergent, compared to almost 83% for wild-type (Table 3). This susceptibility may be associated with the compromised cuticle of the fat-3 mutant [42], but may also indicate defects in membrane structure and permeability in fat-3(wa22) animals due to the absence of long chain unsaturated fatty acids. Supplementation with GLA and SDA, as well as AA or EPA fully rescued fat-3(wa22) susceptibility to both heavy metals and to detergent (Table 3). This may indicate that susceptibility to these stresses is not specific to a particular PUFA species but dependent, instead, on the total level of unsaturated fatty acids in the animal. Importantly, that AA or EPA rescued fat-3(wa22) susceptibility to these abiotic stresses but not susceptibility to pathogens further dissociated immune function from general stress resistance in the fat-3(wa22) mutant.
The physiological temperature that supports C. elegans development ranges from 15–25°C [48]. To determine the ability of fat-3(wa22) animals to tolerate extreme temperatures, we assayed for the number of one-day old adults that remained alive following exposure to 36°C and 0°C for a defined period. In contrast to heavy metal and detergent stresses, fat-3(wa22) animals were more resistant than wild-type to extreme temperatures. A significantly higher proportion of fat-3(wa22 ) than wild-type animals survived the 36°C heat stress for 10 hours (Table 3) and 12 hours (data not shown). Similarly, following a 24-hour exposure to 0°C cold stress, significantly more fat-3(wa22) than wild-type adults remained alive (Table 3). The findings that supplementation with GLA and SDA did not restore cold and heat resistance (Table 3), but effectively restored pathogen sensitivity of the fat-3(wa22) mutant to wild-type (Table 1) further disassociates resistance to extreme temperatures from pathogen susceptibility. The observations that fat-3(wa22) animals were not always more sensitive than wild-type to all the abiotic insults tested, and that heavy metal and detergent sensitivity but not immune functions could be rescued by AA or EPA, further support the hypothesis that susceptibility of the fat-3(wa22) animals to P. aeruginosa is likely to be due to specific immune defects.
The requirement for fat-3 in the intestine, the primary site of P. aeruginosa infection in C. elegans, raised the possibility that GLA and SDA could influence immune gene expression. To provide a further link between fat-3 gene function and innate immunity, we compared the expression of 50 infection-response genes [49] by qRT-PCR, in 1-day-old adult wild-type and fat-3(wa22) animals (Table S1). These 50 genes were selected based on one or more of the following criteria: a) genes with known or predicted antimicrobial activity, including spp-1 [50] and abf-2 [25], b) genes required for survival against P. aeruginosa infection [27],[28], and c) genes known to be differentially regulated in response to P. aeruginosa infection [27],[28]. We quantified the expression of these genes under normal growth conditions on E. coli to determine basal or constitutive mRNA levels, and following a 12-hour exposure to P. aeruginosa to compare mRNA levels in infected animals (Table S1). The constitutive expression of 22 genes (44%) was significantly different between age-matched fat-3(wa22) and wild-type animals raised on E. coli. Of these, 12 genes were expressed at significantly lower levels (Figure 4A) and 10 were expressed at significantly higher levels (Table S1) in the fat-3(wa22) mutant compared to wild-type. These results indicated that the missing PUFAs in the fat-3(wa22) mutant are required for proper basal or constitutive expression of a significant subset of infection-response genes tested. Genes expressed at lower levels in fat-3(wa22) included three antimicrobial peptide homologs, spp-1, encoding a saposin-like protein, and two lysozymes, encoded by lys-2 and lys-7 (Figure 4A). The reduced expression of these putative antimicrobial genes in uninfected animals raised the possibility that basal immune function of the fat-3(wa22) mutant may be compromised. To address this hypothesis, we first determined if the 12 constitutively down-regulated genes were required for immunity in wild-type animals. We found that RNAi-mediated knockdown of spp-1, lys-2, lys-7, dct-17 and F08G5.6 resulted in significantly increased sensitivity to P. aeruginosa-mediated killing (Table 4). The remaining genes that did not induce any survival defects on P. aeruginosa following RNAi knockdown were further tested using a colonization assay. Comparing the degree of intestinal colonization by PA14-GFP, a derivative of P. aeruginosa PA14 that expresses the GFP protein [38], provides a more sensitive measure of the infection process, allowing us to detect smaller defects in worm immunity. RNAi-mediated knockdown of lec-11 and F49F1.1 resulted in a significant increase in the rate of colonization by PA14-GFP (Figure S4A), despite a wild-type survival phenotype on the pathogen (Table 4). We henceforth refer to these seven infection-response genes as immunity genes, due to their role in protecting C. elegans from infection. The demonstration that a majority of the genes that were expressed at reduced levels in fat-3(wa22) animals were functionally important for immunity against P. aeruginosa provides a potential molecular basis for the sensitivity of the fat-3(wa22) mutant to infection and suggests that the absence of GLA and SDA can compromise basal immunity.
We next compared expression levels of the 50 infection-response genes following a 12-hour infection with P. aeruginosa. The mRNA levels of only seven genes (14%) were significantly different between P. aeruginosa-infected fat-3(wa22) and wild-type animals (Figure S4B, Table S1), indicating that majority of the genes in the fat-3(wa22) mutant, including a number of genes misregulated under basal conditions, responded to P. aeruginosa infection to reach levels similar to the wild-type worm. With the exception of lys-7, the mRNA levels of all the genes that were constitutively expressed at lower levels in uninfected fat-3(wa22) animals were indistinguishable from wild-type following P. aeruginosa infection (Figure 4B, Table S1), indicating that, at the level of gene expression, the ability of fat-3(wa22) mutants to respond to infection remained largely intact. Taken together, these results indicate that despite displaying a largely normal inducible response to infection, the significant reduction in constitutive expression of immunity genes was sufficient to render fat-3(wa22) animals more susceptible to P. aeruginosa-mediated death. These results underscore the importance of basal or constitutive immunity for protection from pathogens.
Given that fat-3 expression in the intestine is required to protect C. elegans from P. aeruginosa-mediated killing (Figure 3B), we wondered if expressing fat-3 in the intestine would be sufficient to restore the expression of infection-response genes in the fat-3 mutants. Infection-response gene expression was quantified by qRT-PCR in transgenic fat-3 strains that specifically express the fat-3 transgene in the intestine, muscles or neurons. As these transgenic strains were constructed in the fat-3(lg8101) background, we first confirmed that, with the exception of lec-11, the basal gene expression of the 12 genes assayed were similarly misregulated in fat-3(lg8101) and fat-3(wa22) relative to wild-type (Table S2). The reason for this allele-specific effect on lec-11 expression is currently unclear. Excluding lec-11 from the remaining analysis with transgenic animals, we note that the expression of the fat-3 gene under the control of its own promoter was sufficient to rescue the expression of all but one of the 11 infection-response genes tested (Figure S5A), indicating that the requirement for fat-3 in immune function is strongly correlated with the expression of infection-response genes. Intestine-specific expression of fat-3 restored the expression of seven of the eight the down-regulated infection-response genes, including the immune-specific genes spp-1, lys-7 and F08G5.6 (Figure S5B), indicating that fat-3 is required in the intestine to regulate basal gene expression. By contrast, and consistent with the pathogen survival assay (Figure 3B), expression of fat-3 specifically in the muscles failed to restore the expression of any of the genes tested, with the exception of F35E12.8 (Figure S5C). Expression of fat-3 in neuronal tissues was similarly ineffective at restoring infection-response gene expression; expression of only two genes, F35E12.8, and F08G5.6 were restored to wild-type (Figure S5D). Analysis of spp-1 expression in these transgenic animals is revealing, as spp-1 is expressed only in the intestine [24]. Expression of spp-1 was restored to wild-type levels when fat-3 was expressed specifically in the intestine (Figure S5B) but not in the muscles (Figure S5C) or neurons (Figure S5D). Together, these data confirm our hypothesis that fat-3 functions in the intestine to influence the expression of a number of infection-response genes that contribute to the protecting C. elegans from P. aeruginosa infection.
Of the 12 genes that are positively regulated by fat-3 (Figure 4A), the expression of lys-2, dod-19, ZK6.11 and F08G5.6 has been reported to be dependent on the p38 MAP kinase pathway [28]. To determine if altered constitutive gene expression in the fat-3(wa22) mutant correlated with defects in p38 MAP kinase signaling, we compared mRNA levels between fat-3(wa22) and sek-1(km4), a p38 MAP kinase kinase mutant [29], adults raised on E. coli. Using the set of 22 genes that were significantly altered in fat-3(wa22) animals, we found that basal gene expression between fat-3(wa22) and sek-1(km4) animals was highly correlated (Figure 4C), with the expression of 18 out of 22 genes being similarly altered in both strains. A majority of these genes, however, were expressed at lower levels in sek-1(km4) compared to fat-3(wa22) animals (Figure 4A and C). A similarly significant correlation was seen between fat-3(wa22) animals and another MAP kinase pathway mutant, the p38 MAP kinase homolog, pmk-1(km25) (R2 = 0.4409, p = 0.0008). Significant correlations in basal gene expression were also seen between fat-3(wa22) and sek-1(km4) (R2 = 0.332, p = 0.0001), and between fat-3(wa22) and pmk-1(km25) (R2 = 0.259, p = 0.0003) animals when the analysis was extended to the entire 50 gene-set.
In addition to the p38 MAP kinase pathway, the Sma/TGF-beta and Insulin/Insulin growth factor signal transduction pathways also play important roles in C. elegans immunity [19],[51],[52]. However, basal gene expression in fat-3(wa22) animals was not significantly correlated with the null allele of the FOXO transcription factor of the insulin pathway, daf-16(mu86) (Figure 4D), or the null allele of the Sma/TGF-beta receptor, sma-6(wk7) (Figure 4E). This high concordance in altered basal gene expression between fat-3(wa22) and the sek-1(km4) or pmk-1(km25) mutants led us to hypothesize that the basal activity of the p38 MAP kinase pathway may be compromised in fat-3(wa22) animals.
As a direct measure of the effect of the fat-3 mutation on p38 MAP kinase pathway activity, we used immunoblot analyses to determine the levels of activated PMK-1 protein in fat-3(wa22) and wild-type age-matched adults raised on E. coli. The sek-1(km4) mutant, previously shown to have a complete loss of PMK-1 phosphorylation and increased susceptibility to infection [29], was used as a control. Wild-type worms had detectable levels of phosphorylated PMK-1 indicating some basal p38 MAP kinase activity under normal physiological conditions. By contrast, fat-3(wa22) had decreased levels of phosphorylated PMK-1 protein (Figure 4F), indicating a reduction in basal activity of the p38 MAP kinase pathway. This decrease in the basal levels of activated PMK-1 in fat-3(wa22) animals, as opposed to the complete loss of phosphorylated PMK-1 protein in sek-1(km4) animals, is consistent with the trends indicated by the qRT-PCR analysis, showing that the expression levels of infection-response genes in the fat-3(wa22) mutant were not as low as in the sek-1(km4) mutant (Figure 4A). The decrease in PMK-1 phosphorylation and immune gene expression indicate that although the fat-3(wa22) null mutation does not completely abolish PMK-1 activity, it is sufficient to compromise immune function in the worm.
As shown in Figure 4B and Table S1, a majority of the infection-response genes were expressed at wild-type levels in infected fat-3(wa22) animals. Consistent with this gene expression data, immunoblot analysis of PMK-1 activation in fat-3(wa22) and wild-type lysates following a 12-hour exposure to P. aeruginosa revealed that PMK-1 phosphorylation in the infected fat-3(wa22) mutant was restored to 81% of that seen in infected wild-type animals (Figure 4F). By contrast, in sek-1(km4) animals, the level of phosphorylated PMK-1 protein remained at background following infection, (Figure 4F), and immunity genes that were expressed at low levels under basal condition remained low following PA14 infection (Figure 4B). Thus, both the gene expression and PMK-1 phosphorylation analyses support the conclusion that FAT-3 Δ6-desaturase is necessary to maintain basal activation of PMK-1 but appears to be dispensable for PMK-1 activation and the associated immune gene expression during infection.
Since the loss of fat-3 gene function resulted in the reduced phosphorylation of PMK-1, fat-3(wa22) mutants are expected to manifest phenotypes that are associated with a loss or reduction in p38 MAP kinase signaling. A well-characterized defect of the sek-1(km4) mutant that is associated with a complete loss of PMK-1 phosporylation is an increased susceptibility to arsenic-induced oxidative stress [53]. We therefore compared the ability of 1-day-old adult sek-1(km4) and fat-3(wa22) mutants to survive on 3 mM arsenic. As expected, fat-3(wa22) and sek-1(km4) animals had similar survival rate following exposure to arsenic (Figure 4G), further indicating that the p38 MAP kinase pathway is functionally compromised in fat-3(wa22) animals.
The increased sensitivity of fat-3 loss-of-function mutants to pathogen infection and arsenic stress is associated with reduced basal p38 MAP kinase signaling. Among the genes that positively regulated by fat-3 are three antimicrobial peptide homologs: spp-1, which encodes a saposin-like protein, and lys-2 and lys-7 that are predicted to encode for lysozymes (Figure 4A). This raises the possibility that the influence of fat-3 on immune function may be distinct and separable from its arsenic-induced oxidative stress response. To identify fat-3-regulated genes that are specifically require for immunity, we inactivated each of the 12 genes down-regulated in the fat-3(wa22) mutant individually by RNAi and determined the effects of gene knockdown on pathogen and arsenic sensitivity. As shown in Table 4, four groups of genes were identified. Members of the first group, dct-17 and F49F1.1, were like sek-1(km4) and fat-3(wa22) in that they were required to protect C. elegans from both pathogen and arsenic. By contrast, inactivation of the group 2 genes ZK6.11, dod-19 and T01D3.6, had no detectable effect on pathogen or arsenic survival. Of particular interest are members of group 3, spp-1, lys-2 and lys-7, F08G5.6 and lec-11, that are specific to immune functions; they were required to protect C. elegans from P. aeruginosa infection but not from arsenic-induced oxidative stress. F35E12.8 and gst-38 are members of group 4 that were not required for C. elegans survival against infection, but were required for the response to oxidative stress. Interestingly, gst-38 is predicted to encode a glutathione-S-transferase that plays an important role in protection from oxidative stress. We are thus able to distinguish among the fat-3-regulated genes, a set of immune-specific genes that are distinct from those required for protection from arsenic toxicity. These results strongly indicate that fat-3 influences the expression of genes that have specific role in innate immunity.
The specific rescue of fat-3(wa22) survival on P. aeruginosa by GLA and SDA led us to hypothesize that basal PMK-1 phosphorylation and expression of infection-response genes would be restored with the supplementation of these PUFAs. As with the pathogen survival assay, we determined the effect of PUFA supplementations on the basal expression of 10 infection-response genes, 6 that were down-regulated and 4 that were significantly up-regulated in the fat-3(wa22) mutant, by qRT-PCR (Figure 5). As expected, individual addition of GLA or SDA fatty acids resulted in a partial restoration of immune gene expression in uninfected fat-3(wa22) worms, while simultaneous addition of both fatty acids completely restored the majority of basal immune gene expression in the fat-3(wa22) mutant to wild-type levels (Figure 5A), including spp-1, lys-2, lys-7 and lec-11 that are specifically required for immune function (Table 4). By contrast, supplementation with ALA, AA or EPA, the PUFAs that did not rescue fat-3(wa22) pathogen sensitivity (Figure 2D), also had no significant effect on the expression of these misregulated infection-response genes, with the exception of lec-11 (Figure 5B). Two genes that had no effect on survival when knocked down by RNAi; clp-1 and ZK39.6 (data not shown), also did not appear to be affected by the addition of any of the above PUFAs.
The conclusion that both GLA and SDA are specifically required for survival against PA14 infection, through the regulated expression of infection response genes, was further supported by immunoblot assays that determined the effect of fatty acid supplementation on PMK-1 phosphorylation (Figure 5C). Supplementation with both GLA and SDA completely restored the level of phosphoryated PMK-1 in the fat-3(wa22) mutant to wild-type, indicating that both fatty acids are required to maintain basal PMK-1 activation, and thus the basal PMK-1-dependent MAP kinase immune function. Supplementation with ALA, AA or EPA, on the other hand, had no detectable effect on the levels of phosphorylated PMK-1 in the fat-3(wa22) mutant.
Supplementation with GLA and SDA also rescued the fat-3(wa22) response to oxidative stress, again functionally confirming the restoration of p38 MAP kinase activity with the supplementation of the two missing fatty acids (Figure S6). Supplementation with AA and EPA had no effect on the oxidative stress response, as expected from their lack on effect on PMK-1 phosphorylation or gene expression in the fat-3(wa22) mutant (Figure S6). We thus provide strong genetic evidence that two specific 18-carbon PUFAs, GLA and SDA play a vital role in maintaining basal activity of the p38 MAP kinase pathway and consequently influence both immune and stress responses in C. elegans. That the fat-3 gene, through the synthesis of SDA and GLA, influences the basal expression of immune-specific genes, such as spp-1, lys-2 and lys-7, and lec-11, further indicates that fat-3 has a specific role in innate immunity, independent of its influences on oxidative stresses.
Using C. elegans mutants defective in PUFA biosynthesis, and detailed analysis of the Δ6-desaturase mutant fat-3(wa22), we identified two 18-carbon PUFAs, GLA, an omega-6 fat, and SDA, an omega-3 fat, that play critical roles in basal immunity. Depletion of GLA and SDA resulted in disrupted basal activity of the p38 MAP kinase pathway and defective basal immune gene expression, leading to increased susceptibility to infection by P. aeruginosa. We also demonstrated that fat-3 is required in the intestine, the site of P. aeruginosa infection, to protect C. elegans from pathogen-mediated death and to regulate the expression of immunity genes. The p38 MAP kinase pathway is required to protect C. elegans from infection and oxidative stress. Importantly, we showed that fat-3, through the synthesis of GLA and SDA, affects the expression of a subset of genes that are specifically required for immune function but not oxidative stress response. We further showed that loss of fat-3 gene function does not result in a general loss of defense against stresses, and provided evidence that support an independent role for GLA and SDA in innate immunity.
Fatty acid desaturases have previously been shown to have important roles in innate immunity in mammals and plants. In mice, the stearoyl-Coenzyme A desaturase protein SCD1 is required for the production of immune effector molecules. SCD1 catalyses the Δ9 desaturation of 16- and 18-carbon saturated fatty acids into the monounsaturated palmitoleic (PLA, 16:1n9) and OA that are bactericidal against Gram-positive pathogens. Consequently, mice carrying loss of function SCD1 mutations are defective in clearing skin infections by Streptococcus pyogenes and Staphylococcus aureus [54]. Whether these MUFAs are also involved in immune signaling remains to be investigated. In plants, mutants in the Arabidopsis SSI2/FAB2 gene, which encodes a Δ9 desaturase, show enhanced resistance to bacterial and biotrophic oomycete fungal pathogens but increased susceptibility to a necrotrophic fungal pathogen [55]–[57]. The immune phenotypes of the ssi2 mutant are due to low levels of OA, which leads to the constitutive activation of the salicylic acid-dependent immune pathway and repression of the jasmonic acid (JA)-dependent pathway by unknown mechanism(s) [55],[57],[58]. Disruption of another desaturase that catalyzes the conversion of LA to ALA, encoded by the spr2 gene in tomato plants and fad-7 and fad-8 in Arabidopsis, also results in diminished JA signaling and a reduced response to wounding by insects and infection by fungal pathogens [59]–[61]. Similar to the lipid-dependent manipulation of immune signaling in plants, we showed for the first time that GLA and SDA, the products of FAT-3, an animal Δ6 desaturase, are required to maintain basal expression of immunity genes through their effect on the phosphorylation of a C. elegans p38 MAP kinase homolog, PMK-1.
In mammals, the omega-6 and omega-3 18-carbon PUFA families cannot be synthesized de novo. They must be produced from the dietary essential fatty acids, LA and ALA through a series of elongation and desaturation reactions. LA and ALA have relatively little pharmacologic action of their own; their effects derive largely from metabolic processing to more active end products. The human ortholog of the C. elegans FAT-3 enzyme, fatty acid desaturase 2 (FADS2) regulates production of GLA and SDA from their LA and ALA precursors [62]. This reaction is slow and can be further impaired by numerous factors, including aging, nutrient deficiencies, diabetes, hypertension, and life style factors, such as stress, smoking and excessive alcohol consumption [63],[64]. Thus, reduced dietary intake of LA and ALA, coupled with any of these conditions could lead to insufficient production of GLA and SDA in the body, potentially leading to compromised basal immunity analogous to the C. elegans fat-3(wa22) mutant. Reduced activity of FADS2 could also result in the decreased production of down-stream metabolites, such as the inflammatory mediators AA and EPA [64].
As noted above, we have provided several lines of evidence that the decreased ability of fat-3 mutants to survive infection by P. aeruginosa is a consequence of diminished synthesis of GLA and SDA (Figure 2). By contrast, the 20-carbon PUFAs, AA and EPA appear to have minimal effects on the infection response to P. aeruginosa. Mutants deficient in different 20-carbon PUFAs, such as elo-1(gk48), elo-1(wa7), fat-1(wa9), fat-4(ok958) and fat-4(wa14) show no defects in their response to infection (Figure 2A). Although supplementation with AA or EPA was sufficient to restore many of the additional defects displayed by the fat-3(wa22) mutant, neither of these PUFAs had any significant effects on the immune defects of fat-3(wa22), as measured by survival on pathogen, expression of immune-specific genes and phosphorylation of PMK-1. In mammals, AA can be metabolized by cyclooxygenase, lipoxygenase and cytochrome P-450 (CYP) enzymes to produce important signaling molecules [3]. Since the C. elegans genome does not contain obvious orthologs of mammalian cyclooxygenases and lipoxygenases or of prostanoid and leukotriene receptors, a role for prostaglandin and leukotrienes in lipid signaling can be largely excluded. However, C. elegans shares with mammals the capacity to produce CYP-dependent eicosanoids. Recently, it was demonstrated that C. elegans contains microsomal monooxygenase systems, consisting of CYP-29A3 and CYP-33E2 cytochromes and an EMB-8 microsomal NADH-cytochrome c reductase that catalyze the epoxidation and hydroxylation of EPA and AA to specific sets of epoxy- and hydroxy-derivatives [65]. The ability of C. elegans to generate endogenous CYP-dependent eicosanoids could be blocked by inhibitors, such as adamantyl-3-dodecyl urea (ADU) developed against mammalian soluble epoxide hydrolases [65] suggesting that this component of eicosanoid metabolism may be conserved between C. elegans and mammals [66]. CYP-derived eicosanoids have been implicated in a variety of critical biological processes in humans, including homeostasis and inflammation [67]. Although our genetic analysis indicates that AA and EPA have no significant effect on the ability of C. elegans to survive infection by bacterial pathogens, we cannot not rule out other, as yet unidentified, roles for these 20-carbon fatty acids in the immune response.
CYP enzymes also play a role in the synthesis of oxylipins, oxygenated fatty acids synthesized from precursor PUFAs [68]. Oxylipins are typically derived from cis PUFAs, such as LA, ALA or AA [15], and act as signaling and effector molecules. Among the best known oxylipins are jasmonic acid and its derivatives that form vital signaling and effector molecules in plant immune responses [16]. In mammals, eicosanoids form one of the major groups of oxylipins, and are potent modulators of various physiological processes, including the regulation of inflammation [4]–[6]. Many oxylipins also show direct antimicrobial activities against bacteria, fungi and oomycetes [69]–[71]. A recent report indicated that in Cyanobacteria, GLA and SDA can be converted to oxylipins by CYP enzymes, but this process is not well characterized [72]. Little is known of oxylipin synthesis in C. elegans, but the presence of functional cytochrome P-450 enzymes leaves open the possibility that GLA and SDA could be processed into functional signaling molecules or immune effectors that directly influence the immune response. Future work will focus on determining if deficiency in CYP-derived eicosanoids or oxylipins could affect innate immune function in C. elegans.
The disruption of the FAT-3 Δ6-desaturase also resulted in altered immune gene expression and defective basal p38 MAP kinase activity (Figure 4). We demonstrate that this reduction in basal activity of p38 MAP kinase signaling and the concomitant increased susceptibility to both infection and oxidative stress, due to loss of fat-3 function, are associated with GLA and SDA deficiencies. In C. elegans, the p38 MAP kinase is required for both the basal and induced expression of genes in response to infection [28] and functions through the activation of the p38 MAP kinase ortholog, PMK-1. Under normal growth conditions, this pathway is active as low levels of phosphorylated PMK-1 can be detected. We present the first evidence that the maintenance of PMK-1 basal activity requires GLA and SDA. Depletion of GLA and SDA in the fat-3(wa22) mutant significantly reduced the levels of phosphorylated PMK-1, without affecting the PMK-1 protein levels. This disruption further resulted in the altered basal expression of a number of immunity genes, as well as an increased susceptibility to oxidative stress. Despite retaining an intact response to infection in the absence of GLA and SDA, reduction in basal p38 MAP kinase signaling in the fat-3 mutant was sufficient to cause increased susceptibility to both infection and oxidative stress, highlighting the vital importance of basal immunity. Previous research has similarly demonstrated the importance of this constitutive response in mammals and other invertebrates. In mammalian systems, beta-defensins form a major part of the constitutive immune response, and are continuously expressed in many epithelial tissues. Mice deficient in the production of the lung β-defensin-1 (mBD-1) showed defects in their ability to clear H. influenzae infections from the lung [73]. In this case, however, the mBD-1 mutant mice were defective in both the constitutive as well the inducible expression of the single effector molecules. Here, with the fat-3 mutant, we demonstrate the essential requirement of a constitutive immune response pathway for survival against the pathogen, despite the presence of a functional inducible infection response in C. elegans. It would be additionally interesting to determine if GLA and SDA deficiencies in humans are also associated with reduced p38 MAP kinase activity and hypersensitivity to infection, and if these pathophysiological conditions could be restored through dietary supplementation of GLA and SDA.
The mechanism by which GLA and SDA affect the activity of p38 MAP kinase signaling and immune gene expression is currently unknown. GLA and SDA have a range of actions, and future work will be required to determine if their effects are direct or indirect. Lipids form a major constituent of cell membranes, and the effects of GLA and SDA may be associated with their influence on the physical properties of these membranes. The extent of membrane fatty acid unsaturation is known to influence membrane structure, fluidity and permeability [74]. Membrane fluidity is the extent of molecular disorder and molecular motion within the lipid bilayer [75]. This physical state of the membrane lipid can act directly to regulate membrane-bound proteins, such as receptor-associated protein kinases and ion channels, leading to alteration of gene expression [76],[77]. Thus, the effect of GLA and SDA depletion on reduced signaling through the p38 MAP kinase pathway may be linked to their effects on membrane fluidity, perhaps analogous to osmoregulation in yeast. When glucose was added to yeast medium to raise osmolarity, an associated reduction in membrane fluidity was observed [78]. When shifted to high osmolarity, yeast cells rapidly stimulate a MAP kinase cascade, the high-osmolarity glycerol (HOG) pathway, which orchestrates part of the transcriptional response [79]. Alternatively, the levels of SDA and GLA could affect lipid-protein interactions of membrane receptors and thus the intensity of signaling, analogous to the effects of OA on G protein coupled receptor (GPCR)-associated signaling. Addition of OA alters membrane structure and results in reduced G protein receptor activity in 3T3 cell derived membranes [80]. GPCRs are capable of activating MAPKs using an intricate signaling network [81]. It would be interesting to determine if depletion of GLA and SDA results in changes in membrane structure or fluidity that leads to reduced p38 MAP kinase signaling, through their effects on GPCRs or other membrane-associated signaling molecules.
Another important aspect of a cell membrane is its selective permeability, which plays a vital role in maintaining cell integrity and preventing entry of toxins [82]. Given that most pathogens secrete toxins and hydrolytic enzymes that can harm host cells, membrane permeability might affect the outcome of an infection. The fat-3(wa22) mutant is more sensitive to the detergent Triton X-100, potentially pointing to a defect in cell membrane permeability. However, the detergent sensitivity of the fat-3(wa22) mutant could be rescued without affecting its sensitivity to P. aeruginosa infection (Table 3, Figure 2C,D), indicating that the potentially altered cell membrane permeability of the fat-3(wa22) mutant does not impact immune function. Of note is that the fat-3(wa22) mutant also has a defective cuticle [42], which could account for, or partly influence, the detergent sensitivity of the fat-3(wa22) mutant, rather than a defect in membrane permeability.
Lipids perform a multitude of roles in the immune system, influencing both the innate and adaptive immune responses to infection. While their roles as inflammatory precursors is well known, studies have also identified lipid derived ligands that function through the mammalian Peroxisome Proliferators Activated Receptors (PPARs) to modulate the adaptive T cell response [83] and activate NK cells and dendritic cells [9],[10]. PPARs are a subset of nuclear hormone receptors (NHRs), a family of transcription factors activated by small lipophilic ligands that control a number of metabolic and systemic processes. In mammals, GLA is primarily converted to DGLA, a precursor of anti-inflammatory eicosanoids [64]. In keratinocytes, however, GLA treatment also results in the induction of COX-2 expression in a PPAR-γ-dependent manner. Addition of GLA results in the translocation of PPAR-γ to the nucleus and a consequent increase in COX-2 promoter activity and COX-2 protein levels in the cell [84]. This suggests a possible direct signaling role for GLA in regulating expression of the COX-2 gene, through PPAR-γ to mediate inflammatory immune responses. C. elegans has no known inflammatory response but does posses 284 putative NHRs [85], several of which affect the fat content [33] or the lipid metabolism of the worm [34],[86]. A number of these NHRs are differentially regulated in response to infection [27],[28], and reducing expression of nhr-112, by RNAi, results in increased sensitivity to infection by P. aeruginosa [27]. The interaction between lipid ligands, NHRs and the MAP kinase pathways has been explored previously in the context of the PPAR receptors in mammalian systems. In CD4+ T cells, unliganded PPARα suppresses p38 MAP kinase phosphorylation. Activation of PPARα by its lipid ligand relieves this restraint, allowing phosphorylation and activation of the MAP kinase pathway to induce cytokine production in these T cells [87]. Conversely, the p38 MAP kinase pathway has also been implicated in the control of PPARα activation and function. In vitro analysis shows that phosphorylation by p38 MAP kinase enhances activity of PPARα in cardiomyocytes [88], suggesting the possibility for similar complex interactions between the GLA and SDA PUFAs, NHRs and the p38 MAP kinase pathway in C. elegans innate immune function.
In summary, the demonstration that GLA and SDA are required for basal immunity adds to out understanding of the varied roles for lipids in immunity. Disrupting the synthesis of GLA and SDA leads to an increased sensitivity to infection, and the disrupted basal activity of the p38 MAP kinase pathway. Given that numerous conditions, including aging, diabetes, stress and smoking could lead to GLA and SDA deficiencies, it will be of interest to explore the roles for these PUFAs in other organisms, including humans.
The strains fat-2(wa17), fat-3(wa22), elo-1(wa7), elo-1(gk48), fat-1(wa9), fat-4(ok958), fat-4(wa14), daf-16(mu86), sma-6(wk7), sek-1(km4), pmk-1(km25) and pha-1(e2123) were obtained from the Caenorhabditis Genome Center (CGC). The fat-6(tm331) strain was obtained from Dr. Shohei Mitani (National BioResource Project, Japan). The fat-7(wa36) and the fat-6(tm331); fat-7(wa36) double mutant were gifts from Dr. Jennifer Watts (Washington State University). fat-3(lg8101) and the tissue specific rescue strains were gifts from Dr. Giovanni M Lesa (University College London). All strains were grown on nematode growth media (NGM) plates at 25°C and fed with the E. coli strain OP50-1 unless noted otherwise. For the assays described below, unless noted otherwise, all the worms were grown at 25°C on E. coli HT115 expressing the pos-1 RNAi construct to prevent the production of progeny [89]. Bacteria expressing dsRNA directed against pos-1, sek-1, lys-7 and lec-11 were part of a C. elegans RNAi library expressed in E. coli strain HT115 (Geneservice, Cambridge, U.K.). Bacteria expressing dsRNA directed against spp-1, lys-2, dct-17, ZK6.11, dod-19, F49F1.1, F35E12.8, F08G5.6, gst-38, T01D3.6 were part of a C. elegans library expressed in E. coli strain HT115 (Open Biosystems, Huntsville, Alabama). All bacterial strains were cultured under standard conditions at 37°C.
ALA, AA and EPA supplements were obtained from NuChek Prep Inc., GLA was obtained from Sigma-Aldrich Co., while SDA was obtained from Cayman chemicals Co. Fatty acids were dissolved in 95% ethanol, and were added to a final concentration of 4 mM to E. coli HT115 carrying the pos-1 RNAi construct and allowed to dry overnight in the dark. Wild-type animals fed pos-1 RNAi bacteria supplemented with an equivalent amount of ethanol were used as controls. Worms were allowed to grow on the supplements from egg to one-day-old adults at a temperature of 25°C. For adult supplementation assays, young adult animals were placed on supplement plates for 48 hours before transfer. Supplementation, in each case, was verified by GC-MS, and collected worms were used for multiple assays.
Survival assays were performed as described [27]. Plates were scored every 12 hours, and worms that showed no response to touch were scored as dead. Worms that died due to desiccation or by bagging due to live progeny were excluded from the analysis. Statistical analyses were performed using a Kaplan-Meier non-parametric comparison and a Logrank test, using Statview (Version 5.0.1, SAS Institute Inc.). All assays were repeated a minimum of three times, with approximately 120 worms tested per condition in each assay.
Synchronized populations of several thousand adult worms were harvested at appropriate time points and washed with M9 buffer to remove excess bacteria. Worm pellets were treated with 3% H2SO4 in methanol and incubated at 80°C for 2 hours. Fatty acids were extracted as described previously [32] and analyzed by GC-MS using an HP 6890 gas chromatograph equipped with an HP-5MS column (30 m×0.25 mm×25 µm).
One-day-old adult animals were exposed to OP50-1, PA14 or PA14ΔgacA for 12 hours. For supplementation assays, worms were allowed to develop from embryos to young adults in the presence of fatty acids prior to analysis. RNA extraction and qRT-PCR were performed as previously described [27]. 25 µl reactions were performed using the iScript One-Step RT-PCR kit with SYBR green according to the manufacturer's instructions (BioRad Laboratories, Hercules, CA). Cycling threshold (Ct) values were normalized to mRNA levels of three primer pairs, pan actin (act-1,3,4), F44B9.5 and ama-1, which did not change with infection. Values and statistical analyses were calculated from normalized cycle threshold values prior to conversion to relative fold change.
Colonization assays were performed on slow killing plates using a GFP-expressing PA14 strain (PA14-GFP) [38] and the pha-1(e2123) temperature sensitive mutant strain, to avoid the presence of progeny on the assay plates at the restrictive temperature of 25°C. Adult worms were exposed to PA14-GFP for 24 hours and the intestinal bacterial load was determined under a fluorescence microscope. The degree of colonization was determined as follows: worms with a lumen completely packed with PA14-GFP were classified as fully colonized, worms that showed a limited presence of GFP in the intestine were classified as partially colonized and worms with no detectable GFP expression in the intestine were classified as having undetectable levels of colonization. A minimum of 2 independent experiments was performed with a total of 60 worms per sample per time point for each experiment. Statistical analyses were performed using Chi-square tests.
Life span assays were performed on NGM plates containing 0.1 mg/ml FUDR to prevent progeny from hatching [90]. Plates were seeded with concentrated OP50-1 and allowed to dry overnight. A synchronized population of L4 worms was placed onto the plates and scored every 24 hours. All strains were compared to the wild-type Bristol N2 strain and approximately 120 worms were used per strain per experiment. Statistical analyses were performed using a Kaplan-Meier non-parametric comparison and a Logrank test, using Statview (Version 5.0.1, SAS Institute Inc.).
Adult worm samples were washed with M9 and frozen for analysis. Animals were homogenized in M9 buffer and protein content was measured with a BCA Protein Assay Kit (Thermo Fisher Scientific Inc.) before loading. A phospho p38 specific monoclonal antibody (Cell Signaling Technology, Inc.), a p38 specific antibody (Cell Signaling Technology, Inc.) and an anti-actin antibody (Sigma-Aldrich Co.) were used at concentrations of 1∶1000, 1∶250 and 1∶250 respectively.
Slow killing plates were coated with a final concentration of 3 mM sodium arsenite and allowed to dry overnight [53]. Plates were then seeded with E. coli OP50-1 and approximately 30 adult worms were placed on each plate, for a total of 120 worms per strain. Plates were scored every 12 hours and worms that showed no response to touch were counted as dead. Aldicarb (2-methyl-2-[methylthio]- propionaldehyde O-[methylcarbamoyl]oxime; Chem Services, West Chester, PA) stocks were dissolved in acetone and added to a final concentration 0.7 mM onto NGM plates [91]. Plates were allowed to dry overnight in the dark and then seeded with E. coli OP50-1. 30 adult worms were placed on each plate and monitored every 4–6 hours for paralysis, with approximately 120 worms used per strain/treatment for each experiment. Statistical analyses were performed using Kaplan-Meier non-parametric comparisons and Logrank tests, using Statview (Version 5.0.1, SAS Institute Inc.).
Defecation assays were performed as described [42]. Defecation cycles were measured as the time between successive posterior body contractions over a period of five minutes. All assays were conducted in closed Petri dishes seeded with OP50-1, and a minimum of six adult animals was scored for per strain for each fatty acid treatment. Movement assays were performed as described [92] with M9 buffer in 96-well microtiter plates. A minimum of six animals was scored for total number of thrashes within a period of 2 minutes. One ‘thrash’ was defined as a change in the direction of bending at the mid-body.
Metal toxicity assays were performed as described [44]. Briefly, sets of three 1-day-old adults were allowed to lay eggs on plates containing either CdCl2 (30 µM) or CuSO4 (250 µM), for 3 hours. Adult worms were then removed, and the number of eggs on each plate was determined. After incubation at 25°C for 48 hours, the number of surviving adults was counted. The percentage of adults was determined as total number of adults divided by total number of eggs. For PUFA supplementation assays, fatty acids were added to the bacteria before seeding the plates. One-day-old adult animals were incubated in a solution of 1% Triton X-100 for one hour. Following removal from the detergent solution and one-hour recovery on NGM plates at 25°C, the number of survivors was determined. For supplementation assays, worms were grown in the presence of different fatty acids, before being placed in the detergent solution.
Heat and cold stress assays were carried out as described [93]. One-day-old adult animals were exposed to 0°C or 36°C for a period of 24 hours and 10 hours, respectively. Number of survivors was determined following one-hour recovery at 25°C. Significant differences in thermal tolerance were determined using a Student's t-test.
|
10.1371/journal.pgen.1003527 | Strong Purifying Selection at Synonymous Sites in D. melanogaster | Synonymous sites are generally assumed to be subject to weak selective constraint. For this reason, they are often neglected as a possible source of important functional variation. We use site frequency spectra from deep population sequencing data to show that, contrary to this expectation, 22% of four-fold synonymous (4D) sites in Drosophila melanogaster evolve under very strong selective constraint while few, if any, appear to be under weak constraint. Linking polymorphism with divergence data, we further find that the fraction of synonymous sites exposed to strong purifying selection is higher for those positions that show slower evolution on the Drosophila phylogeny. The function underlying the inferred strong constraint appears to be separate from splicing enhancers, nucleosome positioning, and the translational optimization generating canonical codon bias. The fraction of synonymous sites under strong constraint within a gene correlates well with gene expression, particularly in the mid-late embryo, pupae, and adult developmental stages. Genes enriched in strongly constrained synonymous sites tend to be particularly functionally important and are often involved in key developmental pathways. Given that the observed widespread constraint acting on synonymous sites is likely not limited to Drosophila, the role of synonymous sites in genetic disease and adaptation should be reevaluated.
| Synonymous mutations do not alter the sequence of amino acids encoded by the gene in which they occur. These synonymous mutations were thus long thought to have no effect on the function of the ensuing protein or the fitness of the organism. At four-fold degenerate sites, every possible mutation is synonymous. For this reason, they are often neglected as a possible source of important functional changes. Using a deep sampling of the variation within a population of the fruit fly Drosophila melanogaster, we show that, contrary to this expectation, 22% of synonymous mutations at four-fold degenerate sites are strongly deleterious to the point of absence in the Drosophila population. The underlying biological function disrupted by these mutations is unknown, but is not related to the forces generally believed to be the principal actors shaping the evolution of synonymous sites. Genes with many such possible deleterious synonymous mutations tend to be particularly functionally important, highly expressed, and often involved in key developmental pathways. Given that the observed functional importance of synonymous sites is likely not limited to Drosophila, the role of synonymous sites in genetic disease and adaptation should be reevaluated.
| As there are 64 codons and only 20 amino acids, most amino acids can be encoded by more than a single codon. Mutations that alter coding sequences (CDS), but do not alter amino acid sequences are referred to as synonymous mutations. Synonymous sites are then the collection of potential synonymous mutations present in a gene. Predicated on the assumption that the CDS of a gene is simply the recipe for making the protein, synonymous mutations were long thought to have no functional effect, in other words to be “silent” and thus selectively neutral [1], [2]. As a result, synonymous variation is often used as the neutral reference when measuring selection at functionally important, non-synonymous sites [3]–[7].
The observation of codon usage bias in many organisms was the first indication of possible functionality encoded by synonymous sites [8], [9]. Different codons for the same amino acid are often utilized at unequal frequencies across the genome. Highly expressed genes and codons encoding functionally important amino acids generally display particularly biased patterns of codon usage [9]–[11]. This observation led to the theory that selection for translation optimization generates higher levels of codon bias [12]–[15]. In other words, it is thought that the speed and accuracy of mRNA translation is higher for a subset of codons, referred to as “optimal” (“preferred”) codons [14]–[19]. Such codons are translated more accurately and more efficiently because they are recognized by more abundant tRNA molecules with more specific anti-codon binding [14], [20], [21]. While this kind of selection acting on synonymous mutations is widely accepted, it is generally estimated to be weak - nearly, if not quite, neutral [22]–[31]. Synonymous variation is therefore still often thought to lack any major functional or evolutionary importance. In this paper, we further investigate the functionality of synonymous sites through detecting the action of purifying selection. If synonymous sites harbor highly deleterious variants under strong purifying selection, then that must change our view of the functional importance of synonymous sites and their potential role in genetic disease, as a possible source for adaptation, and as the neutral foil in tests for selection.
Previous tests for selection on synonymous sites have often been consistent with the presence of weak purifying selection operating on synonymous variation. Using the rate of divergence between species, the signal of purifying selection comes from a lower number of inferred substitutions on a phylogenetic tree at sites allowing for synonymous mutations, compared to the expectation provided by a neutral reference. Simply comparing the rate of evolution between a test and a neutral reference set can be problematic when weak purifying selection and mutational biases interact [32]. Nonetheless, synonymous sites do indeed appear to evolve slower than expected under neutrality for many organisms in a manner seemingly consistent with weak selection [22], [29]–[31], [33]–[40]. More evidence for weak selection acting at synonymous sites comes from the study of polymorphism within species. Purifying selection reduces the frequency of deleterious alleles in the population. To measure its effect, the site frequency spectrum (SFS) tabulates the fraction of observed SNPs in all frequency classes across the sites of interest. The overabundance of low frequency SNPs relative to the neutral expectation is the signal of purifying selection operating on the test sites. From this signal, one can calculate the strength of the selective force and the proportion of the test sites it affects [41], [42]. Such methods have been applied to studying the effects of selection on synonymous sites in a variety of Drosophila species, and have found evidence of weak selection - often favoring optimal codons [24]–[28].
Studies using divergence and polymorphism to infer selection as described above are, however, unable to detect the action of strong purifying selection. Tests that rely on divergence are limited in power to distinguish strong purifying selection from weak or moderate purifying selection. The problem lies in the efficacy of purifying selection (constraint) over a tree: a small, linear increase in the strength of constant purifying selection causes a large, exponential drop in the rate of evolution [43]–[45]. Weak to moderate constraint is thus capable of conserving sites over even large phylogenetic distances and increasing the number of species/tree length results in only a limited increase in power to distinguish strong from moderate or weak purifying selection [32]. Unlike tests on divergence that have difficulty distinguishing between strong and weak constraint, tests using the SFS of observed polymorphism can miss strong purifying selection entirely. While both weak and strong constraint eliminate variation from the population, strong selection does so far more efficiently. Therefore, at sites of strong constraint there are few SNPs and only at very low frequency in the population. Without sequencing enough members of a population to attain a deep sample, such SNPs will not be in the SFS of observed polymorphism. With no signal in the shape of the SFS from shallow population sequencing data, any strong selection acting on synonymous sites could not be detected via these methods.
While strong selection does not significantly affect the shape of a shallow-sample SFS, the lack of polymorphism can itself be a powerful signal of the action of selection [46], [47]. Knowing how many mutations should be present in the population sample, as compared to the amount actually present, can allow the estimation of the fraction of sites under strong selection. To do this, one needs a large sample of sites as the density of polymorphism is always low - on the order of a few percent. Differentiating between low densities in the test set and the neutral reference thus requires a large number of sites from each.
Note that both weak purifying selection and lower rates of mutation can likewise cause a paucity of SNPs. Ultra-low frequency variants can distinguish the signal of strong constraint from that of a variation in the mutation rate between the neutral reference and the set of sites being tested. While mutational cold spots only lead to a lower number of SNPs, under strong purifying selection some mutations should still be observable at very low frequencies in a deep enough sample of the population. Weak selection, meanwhile, will affect the shape of the spectrum beyond the rare alleles and can be estimated from that. Combined, the lack of polymorphism and the excess of rare variants from a genome-wide, deep sample, could give the necessary power to quantify the intensity of the strong constraint and the fraction of sites it affects. Thus, our dataset needs to include both a wide sample of sites from the genome, as well as a deep sample from the population.
The Drosophila Genetic Reference Panel (DGRP) for D. melanogaster provides such a dataset [48]. With 168 sequenced-inbred lines, this data set represents the whole genome (thus providing us with the widest possible sample of sites from the genome). The data also provides a deep sample of the variation within the D. melanogaster population of North Carolina. Using DGRP polymorphism, we estimate that, contrary to long held expectations, a substantial fraction of the synonymous sites in D. melanogaster is evolving under strong selective constraint. The discovery of strong selection on codon usage in Drosophila should dramatically change our collective perspective on the functional and evolutionary significance of synonymous sites.
To detect the action of selection on DGRP variation in synonymous sites, we need a neutral reference against which to compare the site-frequency spectrum and SNP density of synonymous sites. Short introns in Drosophila have been shown to be evolving neutrally or nearly so [49]–[52]. We therefore use sites from introns shorter than 86 bp as the neutral reference, also removing the edges of these introns, 16 bp away from the intron start and 6 bp away from the intron end, as they may contain splicing elements [52].
For our collection of synonymous sites, to prevent any confusion of synonymous vs. non-synonymous selection acting on a given codon position, we focused on the third codon positions of the four-fold degenerate amino acids (Proline, Alanine, Threonine, Glycine, and Valine). All possible mutations in the third codon position are synonymous for these five amino acids. The third codon positions of these amino acids will from hereon be referred to as 4D sites. So that we could later relate our results from this analysis on polymorphism within D. melanogaster to divergence across Drosophila, we used only those 4D sites from genes with 1–1 orthologs across the twelve sequenced Drosophila species as our test set [53].
To normalize the number of D. melanogaster strains sequenced at each position and any sequencing differences between short introns and 4D sites, we took only those positions which had their base pair called in at least 130 out of 168 strains and further resampled all SNPs to a depth of 130 homozygous strains (see Materials and Methods). The resulting data set consists of 863,972 4D sites, 5.58% of them containing a SNP, and 870,364 sites in short introns with 6.0% of these being polymorphic. By comparing the density and SFS of polymorphism between 4D and short intron sites, we can quantify the strength of selective forces operating on 4D sites and the fraction of such sites they affect.
Before doing so, several potential confounding factors to such an analysis need to be removed. The greatest of these is the difference in GC content between short introns and 4D synonymous sites. The GC content of 4D sites in D. melanogaster is 64%, compared to only 31% for short introns. Mutation is known to be generally biased towards A/T with particularly high rates of mutation from C:G to T:A [30], [51], [54]–[57]. With a higher GC content, 4D sites are thus expected to be subject to a higher mutation rate on average compared to short introns increasing their relative density of polymorphism. This mutational-GC effect could mask any effects of selection on 4D sites, which if present, would reduce the density of polymorphism in 4D sites compared to short introns.
A further complication is that there are spatial variations in the rates of mutation and recombination and in the amount and severity of linked selection across the genome [51], [58]–[61]. Sweeps, strong background selection, and variation in mutation rates may all influence the density of polymorphism in short intron sites relative to 4D sites [62], [63].
Outlined in Figure 1A is our bootstrap procedure to control for GC content and spatial variation in levels of polymorphism. We first pair 4D sites with short intron positions, requiring a short intron and 4D pair to have identical major alleles and be within 1 KB of each other. Such pairs are then sampled with replacement, first drawing a 4D site and then picking at random one of its possible short intron partners, until the number of random pairs drawn equals the population of all 4D sites with short intron pairs. This process matches the GC content of the neutral reference to the test set, ensures the same spatial sampling of the intronic and 4D sites, and as a side bonus, normalizes the total number from each.
Figure 1B shows the SFS for the SNPs in short introns and 4D sites from one bootstrap run. The shapes of the short intron and 4D spectra appear nearly identical. However, this similarity in the shapes of the spectra for 4D and short intron sites belies a large disparity in the density of polymorphism between the two sets. We measured the density of polymorphism in short intron and 4D sites and calculated the standard error of our measurement over 10 bootstrap runs. We find that 4D sites have approximately 22.1% (+/−0.6%) fewer segregating sites as compared to short introns (Figure 1C).
To account for the relative paucity of polymorphism in 4D sites when the spectra of 4D and short introns SNPs are so similar, we combined both facets of information in a maximum-likelihood method allowing for the effects of multiple selective forces and demography on polymorphism (see Materials and Methods). We extended the SFS to include the number of non-polymorphic sites, the “zero”-frequency class, in our 4D and short intron bootstrap samples. Using such “amplitude” information along with the distribution of alleles over the observed frequency classes enables better maximum-likelihood estimation for parameters of strong selection. In this model, selection is parameterized by the effective selection coefficient 4Nes: where s is the selection coefficient and Ne is the effective population size of the organism. In our maximum-likelihood model, we used three categories of selection, neutral: 4Nes = 0, weak purifying: |4Nes|<5, strong purifying: |4Nes|>100. The point estimates for the fraction of sites and the strength of constraint in each selection category can be seen in Table 1. While there is no evidence of extant weak selection acting differentially on 4D sites and short introns, ∼22% of 4D sites are estimated to be under very strong constraint, 4Nes∼−283+/−28.3 (standard error estimate by bootstrap). When a coarse-grained demographic correction was applied to the SFS we obtained results that, though quantitatively are somewhat different (4Nes∼−370.1+/−105), are qualitatively similar in that for both cases 100<< |4Nes| <<700 – the calculable limit of our program (see Text S1).
Exposing the action of the strong constraint on divergence between Drosophila species affirms the functional importance of these 4D sites across evolutionary history and reveals how these constrained synonymous sites evolve. If the strong constraint at 4D sites we identified within D. melanogaster has been constant across Drosophila, we would expect it to result in the complete conservation of the constrained 4D sites. If, on the other hand, the strong constraint is not constant and there is functional turnover at these sites, then we would expect to see substitutions occurring even at constrained sites along the Drosophila species tree. In order to compare the divergence between species to the constraint within a species, we considered only those 4D sites in amino acids conserved across the twelve Drosophila species from D. melanogaster to D. grimshawi. This simplifies the analysis as only the synonymous third position of the codon has been allowed to change over time. Thus, we can focus solely on the evolution of the synonymous site itself rather than consider the evolution of the entire codon. Figure 1C shows that the conservation of the amino acid has no bearing on the fraction of missing polymorphism in 4D sites. As such, the 4D sites of conserved amino acids provide a representative sample with which to study the strong constraint over the evolution of all 4D sites.
The gene orthologs in the other species were obtained from the 12 Drosophila Genome Consortium data realigned by PRANK [53], [68], [69]. We used the established 12 Drosophila species tree and re-estimated the branch lengths on the aligned 4D sites with PhyML (see Materials and Methods) [70]. From these alignments we removed the sequences belonging to D. melanogaster and D. willistoni. The D. melanogaster sequences were removed because the polymorphism data was extracted from this species and we wished to avoid a false concordance between the results from polymorphism and divergence. The D. willistoni sequences were removed, because the branch length leading to D. willistoni is long and the codon bias of D. willistoni is significantly different than from the rest of the twelve Drosophila species [71]. Having removed these species, the expected number of substitutions over the now ten Drosophila species tree for synonymous positions in otherwise conserved four-fold amino acids is estimated by PhyML at 3.1 subs/site [70]. To obtain site-wise estimates of conservation, we then inferred the number of substitutions along this tree for each 4D site independently using GERP (see Materials and Methods) [72], [73].
Figure 2 shows that the percentage of sites under strong constraint declines monotonically as the rate of evolution increases. 40.8% (+/−1.9%) of completely conserved sites (0 substitution class), and only 7.1% (+/−3.0%) of the fastest evolving sites (≥9.3 substitution class) are predicted to be under strong constraint. This difference in the amount of constraint between fast and slow-evolving sites allowed us to carry out a further control for any variation in mutation rate between short introns and 4D sites. We carried out an identical bootstrap procedure but pairing slow-evolving 4D sites with neighboring fast-evolving 4D sites instead of short introns as a neutral reference. We recapitulated our result of strong constraint at 4D sites by using slow- versus fast-evolving 4D sites (see Text S2C).
This correlation between a 4D site's conservation across species and strong constraint within a species underscores the functional importance of these synonymous positions over the evolutionary history of the Drosophila clade. However, over 80% of the sites currently under strong constraint in D. melanogaster fall outside the 0 substitution class, i.e. are not conserved across the ten Drosophila species. Indeed, over 11% of 4D sites under strong constraint in D. melanogaster have each acquired 6.2 or more substitutions over the tree, evolving quickly at more than twice the average rate. As even a moderate amount of selection results in complete conservation if it has been consistent over the tree, this suggests there has been functional turnover at these functionally important synonymous sites.
Codon bias is generally thought to be the product of background substitution biases combined with a weak selective force within genes skewing codon usage towards optimal (preferred) codons to increase translation efficiency and accuracy [19]. In Drosophila, translationally preferred codons are always G- or C-ending (except for in D. willistoni) [71]. The five four-fold degenerate amino acids have the following preferred codons: Alanine - GCC; Glycine - GGC; Proline - CCC; Threonine - ACC; and Valine - GTG [71]. Selection for codon bias is thus likely responsible for driving the GC content of 4D synonymous sites in D. melanogaster to 64% and to over 67% in the 4D sites of amino acids conserved over the 12 Drosophila species. While codon bias increases in conserved amino acids [17], as stated above, the strong selection at synonymous sites inferred in this paper does not (Figure 1C). To explore the relationship between codon bias and the strong constraint, we measured the fraction of sites under strong constraint within each codon, in unpreferred versus preferred codons conserved from D. sechellia-D.grimshawi, and across genes ranked by codon bias.
Table 3 summarizes our analyses of how the extent of strong constraint is influenced by different genic features such as gene length, the location of the synonymous site along the gene, the chromosome on which the gene is located, whether or not the synonymous site falls within splice junctions, and nucleosome binding. Many of the associations below, while suggestive, are marginal in effect. The dominant pattern is that strong constraint at synonymous sites appears to be ubiquitous across different gene classes and functional elements within genes.
To map how strong constraint at synonymous sites varies with gene expression over development, we ranked genes by their expression levels at each developmental time point in the ModEncode data set [84]. We split the genes evenly into three categories of expression - highly, moderately, and lowly expressed - within each developmental stage and ran 10 bootstraps for the 4D sites of the genes within each expression category in each developmental time point. The results are shown in Figure 5.
The overall gene expression level across development correlates well with the fraction of sites under strong constraint with lowly expressed genes tending to have fewer sites under strong constraint and highly expressed genes tending to have more sites under strong constraint. This pattern is strongest for the genes expressed highly in mid-late embryos, pupae, and adult males. The association of strong constraint with these developmental stages is further enhanced when the “high” expression group has been split in half into “high” and “very high” expression level categories (see Figure S2). In contrast to this preference of strong selection for genes highly expressed in embryo, pupal, and adult stages, codon bias is highest for genes whose expression peaks in larval stages [85].
The difference in density of polymorphism between 4D sites and short introns does not allow for precise measurements of constraint on the synonymous sites of single genes. To identify a set of genes that are under particularly strong constraint at synonymous sites, we ranked genes by the fraction of their conserved amino acids that are unpreferred and conserved from D. sechellia to D. grimshawi, in the 0-substitution class (see Materials and Methods). Our method left 4,877 genes capable of being ranked of which we took the top sixth (812 genes, see Dataset S1) as our gene set enriched for strong constraint.
To validate our method of selecting genes under strong constraint, we checked that our 812-gene set is indeed enriched for strong constraint at 4D sites. We performed a bootstrap analysis on the 4D sites of variable amino acids in the genes in and out of this top set. Estimating constraint using 4D sites from variable amino acids provides a measure of the fraction of synonymous sites under constraint independent from our surrogate using conserved amino acids. In the top 812 genes, we find a ∼30% reduction in polymorphism at 4D sites in variable amino acids; in all 4,065 genes not in the top 812 set, we find an average of ∼21% of 4D sites in variable amino acids under strong constraint. As such, our top 812 genes are enriched for almost 50% more 4D sites under strong constraint than the average gene. Note that any individual gene in the 812-set does not necessarily have elevated levels of strong constraint at its synonymous sites, nor does any individual gene of the 4,065 necessarily have a lower fraction of 4D sites under strong constraint.
In order to examine whether genes under strong constraint at synonymous sites tend to be enriched for certain functions, we used DAVID 6.7 [86], [87]. DAVID takes all the genes in the background data set (4,877 genes) and all genes in the test data set (812 genes) and looks for the enrichment of biological terms and gene families in the test set relative to the background. In Table 4, we list a subset of those biological terms found by DAVID's functional annotation clustering run on high stringency (for full information on the top 13 clusters, see Table S3). We find that in genes enriched for strong constraint, we co-enrich for many important functional gene sets. In particular, we co-enrich for genes critical in pupae-to-adult morphogenesis and in late embryogenesis. This finding is consistent with the result that genes expressed highly in late embryos, pupae, and adults have elevated levels of strong constraint at 4D sites. Many other functional classes important to the basic development and functioning of D. melanogaster appear to have a higher fraction of synonymous sites under strong constraint including: transcription factors, ribosomal genes, immunoglobulin genes, genes regulating gamete production – particularly oogenesis, cell-signaling genes – particularly synaptic transmission, and more.
The strong constraint at synonymous sites in D. melanogaster measured in this paper represents a powerful force. We estimate that ∼22% of synonymous sites are experiencing, on average, a selective pressure between 4Nes ∼−250–−500 against deleterious mutations. This strength of selection is as strong or stronger as has been measured via population genetic techniques at any class of sites, including non-synonymous ones [46], [47]. Mutations at strongly constrained synonymous sites should never rise above low frequency in the population and certainly will never fix, barring tight linkage to a very advantageous allele or a shift in the functional properties of the site. While detectable within a population, these mutations are effectively lethal over evolutionary time.
We tested a number of controls to rule out the possibility that our observation of strong purifying selection results from other forces with possibly similar signals: A lower mutation rate, for example, can cause a signal indistinguishable from strong selection in polymorphism if the sample depth of the population is too shallow. To account for this, and at the same time account for any variation in the amount of linked selection between 4D sites and short intron sites, we used a bootstrap to control for GC content and distance between the 4D and short intron sites. We also performed bootstraps controlling for dinucleotide content between 4D and short intron sites and performed bootstraps pairing slow-evolving 4D sites against fast-evolving 4D sites as the neutral reference. Neither revealed a mutational force underlying the ∼22% drop in 4D polymorphism compared to short introns. As revealed by simulations, the finite estimate we obtained of the strength of strong selection is itself evidence against a mutational force being responsible for our signal, as mutational variation would behave like infinite selection on 4D sites. While we do not have the frequency depth from the population necessary to estimate a full distribution of selection coefficients for the strong constraint force, our point estimate of 4Nes ∼−283 for these 22% of sites is statistically significantly different from the value of 4Nes ∼−700 (the computational limit of our program) expected if the signal was due to variation in mutation rate.
We also controlled for deviations from mutation-selection equilibrium affecting both the 4D and short intron site frequency spectra using a frequency-dependent correction. Such deviations include demography, shared (linked) selection between 4D sites and short introns, and our own approximations to the SFS. Controlling for these deviations resulted in a higher estimate of the strength of selection (4Nes∼−370) with larger error bars, but still significantly far from the boundary of 4Nes∼−700.
A constant influx of weakly advantageous alleles in coding sequences, as is expected to occur in D. melanogaster [60], could affect variation at nearby 4D sites more than at short introns. The resulting genetic draft generated by adaptive substitutions in coding sequences would weaken the apparent intensity of purifying selection on 4D sites by bringing strongly deleterious alleles to higher frequencies, making our above estimates of selection intensity conservative [88]. Even so, strong selection, rather than a mutational difference, would still underlie our signal, as genetic draft cannot alter the frequency of synonymous mutations that are simply absent from the population. On the other hand, sweeps of weakly advantageous alleles in coding sequences could eliminate polymorphism in 4D sites more so than in short introns. Narrow selective sweeps in coding sequences reducing variation at otherwise neutral 4D sites is, however, an unlikely explanation for our observations. When comparing 4D sites from different substitution rate classes against each other, we found a signal of strong constraint at conserved 4D sites relative to fast-evolving 4D sites. As sweeps should not affect the overall substitution rate of linked sites, strong purifying selection on synonymous sites is the best explanation for the lack of polymorphism at 4D sites relative to short introns.
Our ability to detect strong selection and differentiate it from other forces critically depends on the availability of deep and genome-wide population data. Previous data sets could only find weak or no constraint, thus always confirming our collective biological intuition that synonymous sites had little functional or evolutionary importance. In a shallower sample of even genome-wide data, the highly deleterious variants would be simply missing from the sample and there would be no power to distinguish strong selection from a variation in the rate of mutation. As an example, we simulated 4D sites evolving under the selective regime inferred from the real data (22% of sites at 4Nes = −283) but with only 60 instead of 130 homozygous strains. Attempting to re-estimate the strength of selection from such a shallow sample results in the observation of seemingly infinite selection operating on 22% of 4D sites. Simulating 60 strains under a scenario where neutral 4D sites have a 22% lower mutation rate than do short introns results in the same inference of infinite selection. Genome-wide, deep population data sets were not available before recently and thus strong constraint could never before be unambiguously detected at synonymous sites.
Interestingly, the strong constraint in D. melanogaster appears to be a largely orthogonal force to canonical codon usage bias, favoring an overlapping, but different set of codons with subtly different gene targets. Codon bias increases as the conservation of amino acids increases, while the strong constraint targets the 4D sites of both conserved and variable amino acids equally. We further identified the codons under strong constraint and, for any given amino acid, the codon(s) with the highest fraction of sites under constraint were not necessarily the optimal codon. Other studies have likewise noted signals of selection favoring non-optimal codons in Drosophila [25], [30], [33], [89], [90]. Overall, preferred 4D sites do have greater amounts of strong constraint acting on them, but the strong selective force targets a substantial fraction of the unpreferred 4D sites as well. There is also a weak anti-correlation between genes with a high fraction of constraint and genes with high codon bias, which extends to various gene features. Long genes are associated with higher levels of strong constraint at 4D sites as opposed to shorter genes, in opposition to codon bias in Drosophila [15], [81]. X-linked genes have a lower fraction of 4D sites under constraint than autosomal genes, wheras codon usage bias is stronger on the X [28], [82]. While both codon bias and the fraction of 4D sites under strong constraint are correlated with highly expressed genes, codon usage bias is strongest in genes with their highest expression in larval stages [85] as opposed to the strong constraint seen most often in genes expressed highly in mid-late embryo, pupal, and male adult stages.
The pattern of conservation over 4D sites supports the existence of weak selection in Drosophila favoring the conservation of preferred 4D sites across the twelve species, but it appears to have been relaxed in D. melanogaster. In our SFS analysis, we were not only able to gauge the intensity of strong selection, but also show a lack of contribution from weak purifying selection to our signal. If any weak selection is still acting differentially on synonymous sites relative to short introns, then it is not powerful enough to be detected by our SFS model or contribute much to our signal of lost polymorphism. These results recapitulate some earlier results on D. melanogaster [24], although see [25]–[28]. While weak selection on 4D sites in D. melanogaster may not have vanished completely, the large influx of mutations and substitutions away from optimal codons corroborates some relaxation of constraint for codon bias in D. melanogaster [25], [30], [31], [33], [38]. Overall, weak selection for codon bias would seem to be less of a force on synonymous sites in D. melanogaster than in its sister species where weak selection for codon bias can be detected with far less ambiguity [24], [30], [31], [33], [34], [40]. Thus, evidence suggests that there are at least two major, orthogonal forces affecting the evolution of 4D sites in Drosophila: the weak selective force driving codon bias that favors optimal codons, present in other Drosophila species, but relaxed in D. melanogaster; and an extant strong selective force targeting both optimal and non-optimal codons in D. melanogaster and across the Drosophila phylogeny. The function engendering the strong constraint appears to be independent of the translation optimization for efficiency and accuracy governing canonical codon usage bias.
The presence of splicing enhancers and nucleosomes do not explain the pattern of strong purifying selection either. However, the function underlying the strong constraint of synonymous sites may yet prove to be acting at the level of gene regulation. Those genes where strong selection on synonymous sites acts most frequently are often highly expressed regulatory proteins, operating in essential, tightly controlled developmental pathways. These are genes where the regulation of gene expression will matter most. Regulation of gene expression may be acting at the level of mRNA structures, mRNA stability, miRNA binding sites, and the modulation of translation rate [91]–[104]. Choice of synonymous codons might affect all of these levels of gene regulation. It should be noted that these various hypotheses are not mutually exclusive and may be intertwined. mRNA structures - as well as their avoidance, especially near the start of ORFs - may be involved in translation initiation/elongation, modulation of mRNA half-life, and accessibility of the mRNA to proteins and miRNAs [98], [99], [103]–[105]. Indeed, signatures of selection have been associated both with mRNA accessibility and mRNA structures and overall folding energy [97], [99], [105], [106]. Our initial analysis found no enrichment of conserved unpreferred codons, a first-pass marker of the action of the strong constraint described in this paper, in either structured or unstructured mRNA as determined by ds/ssRNA sequencing [107] (not shown). This analysis, however, is at best preliminary and a strong possibility remains that the function underlying the strong constraint at synonymous sites is related to mRNA structure. miRNAs also have a host of different functional effects in different species and different genes within a species but are well known in their role of mRNA degradation [108], [109]. The dynamics of translation not only affect the overall rate at which proteins are created, but also affect how these proteins fold and even the mRNA half-life [91]–[94], [110]–[112]. The possibility that strong selection acts at the level of modulating translation rate through the presence of slow/fast sites is interesting as the translation speed of a codon is not necessarily related to codon optimality and tight control has been inferred at the beginning and end of ORFs in some species [96], [100]–[102], [113], [114]. Given the pattern of the strong constraint across the different codons both optimal and non-optimal, the strong selective force may be due to the abundance of wobble vs. Watson-Crick tRNAs available for that codon. Ascertaining the functional mechanism underlying the observed strong constraint acting on synonymous sites could reveal deep insights into the regulation of gene expression.
Regardless of the specific functional mechanism underlying the strong constraint, experimental evidence from a wide range of species substantiates an important functional role for synonymous sites. Directed mutagenesis studies targeting synonymous sites as well as studies of natural polymorphism have found consequential changes in protein levels and functionality due to natural synonymous variation and induced mutations [111]–[113], [115]–[127]. In an experiment done on the Alcohol dehydrogenase (Adh) gene in D. melanogaster, changing 10 wild-type preferred Leucine alleles to unpreferred alleles in the 5′ region of the gene lowers the enzymatic activity of collected Adh by 25% [119]. The authors proposed that disruption of the sites' translational efficiency and accuracy caused the drop in activity, but also noted that the functional effect was far larger than expected given the assumption of only weak selection on synonymous sites [119]. ‘Humanized’ versions of protein coding sequences, with codons replaced with synonymous, putatively optimal codons in humans, show much greater protein expression and function when transfected into mammalian cells than the originals or synthetic versions using a non-mammalian species' set of optimal codons [115]–[118]. Human gene Multidrug Resistance 1 (MDR1) contributes to the drug resistance of cancer cells [122]. Both naturally occurring alleles as well as induced novel mutations at synonymous sites in MDR1 affect the resulting protein's conformation, altering its substrate specificity in human cell lines [122]. In the E. coli gene ompA, exchanging eight frequently-used codons for synonymous infrequently-used codons near the gene start results in a 3-fold reduction in mRNA levels and a 10-fold reduction in synthesis of protein OmpA [112]. Meanwhile exchanging codons with low-abundance tRNAs to synonymous codons with high-abundance tRNAs in E. coli gene sufI - or increasing the abundance of those tRNAs - results in misfolding of the protein in vitro and in vivo [110].
What about the presence of strong constraint in the synonymous sites of other species? In addition to the above functional assays, there are reported to be a significant fraction of synonymous sites under an unknown intensity of constraint in many species [22], [29], [35]–[37], [39], [97], [128]–[130] and there is evidence for strong selection in humans [47]. For example, when compared to “neutral” controls, there is a reduction in polymorphism density and/or a lower rate of divergence at synonymous sites for many tetrapods including chicken, hominids, murids, and mammals in general [22], [29], [35]–[37], [128]. Further, some of these species have undetectable or weak levels of codon bias, presumably commensurate with their small effective population sizes and thus the weakness of selection in favor of optimal codons [36], [131]. Using a similar model to the one described in this paper, Keightley and Halligan (2011) found evidence to support that weak selection alone is unable to explain the pattern of diversity at 4D synonymous sites in humans [47]. While that study lacked the sample depth of polymorphism to be able to gauge the intensity of the strong selection, they estimated that 11% of 4D sites are evolving under a strong selection regime of |4Nes|>40 [47]. Our results from Drosophila with a deeper population sample lend credence to the hypothesis that, in humans too, a force of strong constraint is responsible for the lack of polymorphism at 4D sites rather than a mutational force or other confounding factors. For many species, there has been no conclusion that the constraint on their respective synonymous sites is strong, but many of the signals are consistent with what we find in Drosophila with the fraction of sites under constraint, the amount of missing polymorphism, and the lack of relationship to codon bias. Thus with genome-wide, deep population SNP data becoming available for many of these other species, we may well find strong selection on synonymous sites to be ubiquitous.
As synonymous sites have often been used as the neutral reference in tests for purifying and adaptive selection, many estimates of the fraction of sites under constraint in other classes, such as non-synonymous sites, UTRs, and many others, are likely to be conservative. This result from population genetics supports findings that synonymous sites may harbor many, important causal variants and that studies ignoring the potential contribution of synonymous mutations may be likewise unnecessarily conservative [91]. Turnover at these strongly constrained synonymous sites could also represent a significant source of interspecies functional divergence and adaptation. The potential of synonymous sites to be sources of adaptation and genetic disease merits further investigation. Although the functionality underlying this strong constraint remains unknown, recent studies have uncovered a myriad of different types of functional information encoded into the CDS of genes beyond the protein recipe, including controls for translational efficiency and accuracy, splicing enhancers, micro-RNA binding, nucleosome positioning, and more. With the discovery of a significant fraction of sites under strong constraint in Drosophila, two things become clear: the role of synonymous sites in the biology of genomes is far greater than the neutral, “silent” part they were once assumed to play; and we still have much to learn about the functionality encoded in genes.
The SNP data set from DGRP (http://dgrp.gnets.ncsu.edu/data/) consists of 168 inbred lines from a population of North Carolina D. melanogaster [48]. The SNPs were annotated as synonymous, non-synonymous, and intronic using Flybase release 5.33 (ftp://ftp.flybase.net/genomes/Drosophila_melanogaster/dmel_r5.33_FB2011_01/) [132]. If a position was found in multiple gene annotations, only those sites where the SNP was synonymous in all sites was called synonymous. Short intron sites are defined as those sites falling in introns of less than length 86 bp, 16 bp away from the intron start and 6 bp away from the intron end in order to eliminate any functional sequences at the edges of the introns [52]. Eliminating 16 bp from each side did not change SNP density (not shown). Any remaining purifying selection, especially strong purifying selection, in short introns makes our results more conservative. Four-fold (4D) sites are the collection of 3rd codon positions for the following amino acids: Proline, Alanine, Threonine, Glycine, and Valine.
All sites were resampled to a depth of 130 strains. All sites with sequence information for fewer than 130 strains were excluded. For SNPs at sites with more than 130 strains or which contained heterozygous lines at that position, a 130 allele subset was chosen randomly. If the SNP was no longer polymorphic after this random resampling, that position was moved into the non-polymorphic site class. We also removed any position with more than 2 alleles present.
We restricted our analysis to genes with 1–1 orthologs across the 12 Drosophila species tree [53] and where the longest transcript annotation had remained intact in release 5.33 - even if it is no longer the longest transcript in release 5.33. We used the remaining 5,709 coding sequences aligned with PRANK from Markova-Raina and Petrov (2011) [68], [69].
To determine the distribution of selective effects on a group of sites based on the shape and the amplitude of the SFS, we assume a two-state framework where sites are either monomorphic in the wild-type state or polymorphic with a neutral or deleterious mutation at some observed frequency in the population. Using short introns as a neutral reference, our model aims to capture the fraction of synonymous sites falling into three broad selection categories – those with neutral, weakly deleterious, or strongly deleterious mutations – and estimate the effective selection coefficients acting on those mutations.
Strong constraint can be difficult to capture as strong selection has a greater effect on the amplitude of the SFS, the total number of observed mutations, than on its shape, the frequency distribution of observed mutations. Using a similar expansion to the standard SFS to Keightly and Eyre-Walker (2007) [46], we add the zero-frequency class, the fraction of monomorphic sites, to the SFS. The SNP density provides the additional information necessary to infer the action of strong constraint.
Equal to 4Neμ, θ is mutation rate scaled by the effective population size and determines the neutral SNP density. The short intron SFS, used as neutral reference, anchors our estimate of θ which in turn allows us to estimate the amount of missing synonymous polymorphism in each selection category, c. As purifying selection increases, the overall density of observed polymorphism is reduced in the fraction of 4D sites in that selection class and the expected distribution of mutation is further skewed towards rare frequencies in the population. Each category has a single selection parameter, γc, a point estimate of the effective strength of selection, 4Nes, operating on the 4D sites in that class. For those 4D sites in the neutral category, γc = 0. For those in the weakly deleterious category, 0<|γc|<5. For those in the strongly deleterious category, |γc|>5 or 100 – the choice of boundary did not affect results.
For our sample of n chromosomes from the population, assuming mutation-selection balance, we have the following analytical prediction for the SFS, g(x) – the expected fraction of 4D sites with SNPs at frequency x in the sample [43]:(1)(2)g(x,c) is the contribution of each selection category to the overall SFS. L is the total number of 4D sites while fc is the fraction of 4D sites in each selection category c.(3)(4)g(0) are the zero-frequency class, monomorphic, sites and are what gives the SFS “amplitude” information – the density, rather than just the shape, of the spectrum. While m is the total number observed SNPs in the sample.
The theoretical SFS for intronic sites is the same as above, only all sites are assumed to be neutral. However, any real SFS does not reflect the true frequency distribution of the SNPs in the population, but rather a binomial sampling of those SNPs and frequencies. The above is thus an approximation, as the probability of a site with a SNP at a given frequency in the sample from the population is not quite the same as the probability of a site with a SNP at a given frequency in the population as a whole. However, it is much more computational efficient for both speed and memory to use the approximation.
With this theoretical prediction of the distribution of sites over each frequency class in both the neutral reference (short intron SFS) and test set of sites (4D SFS), we can use maximum-likelihood to fit the parameters of our model to real data sets. Our model has 5 free parameters: θ, (γweak, γstrong), and (fneutral, fweak, fstrong) where fneutral = 1-fweak -fstrong. The total likelihood, λ, of the model's fit to the data, D, is equal to product of the fit the short intron and 4D sites spectra:(5)λ4D and λSI are the likelihood of the observed SFS given the expected SFS as determined by the free parameters and equations (1)–(4). These likelihoods are the multinomial probability of observing a certain number of sites, k, with SNPs in frequency class x in the sample given theoretical expectations. Taking short intron sites as an example (same for both):(6)
Equation (6) is thus the probability that the folded theoretical SFS, g(x), matches the empirical folded SFS, kx. We folded the spectrum to avoid any problems with inferring the ancestral state.
We then maximized the parameters θ, (fneutral, fweak, fstrong), and (γweak, γstrong) in Matlab using fminsearch, an implementation of the Nelder-Mead simplex method [133], on the negative log-likelihood of λfull. The observed spectra were obtained from the bootstrapped 4D and short intron pairs. Where simulations were needed in this study, theoretical spectra were calculated using the above equations (1)–(4) and then the parameters were re-estimated by the outlined maximum-likelihood procedure on those theoretical spectra acting in place of the empirical data.
We used the determined 15 species Insect tree topology from the UCSC genome browser (http://hgdownload.cse.ucsc.edu/goldenPath/dm3/phastCons15way/) and paired it down to the 12 Drosophila species [134]. We then input that tree topology into PhyML v3.0 (http://www.atgc-montpellier.fr/phyml) [70] and allowed it to re-estimate the branch lengths on all 4D sites in conserved amino acids using the HKY85 model [135] without a discrete gamma model and without invariant sites. The nucleotide frequencies and transition-transversion rate ratio were inferred by maximum-likelihood. The resulting tree can be found in Text S4.
GERPcol from GERP++ (http://mendel.stanford.edu/SidowLab/downloads/gerp/) [73] was run on the collection of all 4D sites from all 12 Drosophila species excluding D. melanogaster and D. willistoni, estimating the Rscore (tree length - inferred # of substitutions) for each site independently. We input into GERP the tree and transition-transversion ratio from the PhyML results. As these two programs use different parameterizations of the transition-transversion ratio, we translated one to the other (see Text S4).
Our signal from polymorphism does not afford us a precise measurement of constraint on the 4D sites of a single gene (not enough information). Therefore, we use a surrogate to infer the amount of strong constraint at the 4D sites of individual genes. Looking only at sites without SNPs, we use the percentage of 4D sites in conserved amino acids that are unpreferred and themselves conserved from D. sechellia to D. grimshawi (i.e. in the 0-substitution class) as our measure of how extensive the strong constraint has been on the 4D sites of the gene in question. As unpreferred 4D sites in the 0-substitution class have the highest fraction of sites under strong constraint (53%), the reasoning is that the more such sites exist in a gene, the more likely there has been extensive constraint acting on all 4D sites. Since not all genes have enough conserved amino acids to allow a reasonable calculation of the above surrogate, we used only those genes where at least 20% of the four-fold amino acids were conserved along the tree, leaving 4,877 genes in the analysis. We ranked genes by this surrogate and took the top 812 genes (∼ top sixth of genes). We then used the functional annotation clustering tool from DAVID 6.7 (http://david.abcc.ncifcrf.gov/home.jsp) set on high stringency to look for enrichment of GO category terms in this gene set [86], [87].
|
10.1371/journal.pgen.1006727 | Trans-ethnic predicted expression genome-wide association analysis identifies a gene for estrogen receptor-negative breast cancer | Genome-wide association studies (GWAS) have identified more than 90 susceptibility loci for breast cancer, but the underlying biology of those associations needs to be further elucidated. More genetic factors for breast cancer are yet to be identified but sample size constraints preclude the identification of individual genetic variants with weak effects using traditional GWAS methods. To address this challenge, we utilized a gene-level expression-based method, implemented in the MetaXcan software, to predict gene expression levels for 11,536 genes using expression quantitative trait loci and examine the genetically-predicted expression of specific genes for association with overall breast cancer risk and estrogen receptor (ER)-negative breast cancer risk. Using GWAS datasets from a Challenge launched by National Cancer Institute, we identified TP53INP2 (tumor protein p53-inducible nuclear protein 2) at 20q11.22 to be significantly associated with ER-negative breast cancer (Z = -5.013, p = 5.35×10−7, Bonferroni threshold = 4.33×10−6). The association was consistent across four GWAS datasets, representing European, African and Asian ancestry populations. There are 6 single nucleotide polymorphisms (SNPs) included in the prediction of TP53INP2 expression and five of them were associated with estrogen-receptor negative breast cancer, although none of the SNP-level associations reached genome-wide significance. We conducted a replication study using a dataset outside of the Challenge, and found the association between TP53INP2 and ER-negative breast cancer was significant (p = 5.07x10-3). Expression of HP (16q22.2) showed a suggestive association with ER-negative breast cancer in the discovery phase (Z = 4.30, p = 1.70x10-5) although the association was not significant after Bonferroni adjustment. Of the 249 genes that are 250 kb within known breast cancer susceptibility loci identified from previous GWAS, 20 genes (8.0%) were statistically significant associated with ER-negative breast cancer (p<0.05), compared to 582 (5.2%) of 11,287 genes that are not close to previous GWAS loci. This study demonstrated that expression-based gene mapping is a promising approach for identifying cancer susceptibility genes.
| Although individual genetic variant-based genome-wide association studies have greatly increased our understanding of the genetic susceptibility to breast cancer, the genetic variants identified to date account for a relatively small proportion of the heritability. Shifting the focus of analysis from individual genetic variants to genes or gene sets could lead to the identification of novel genes involved in breast cancer risk. Here, we take advantage of a recently developed gene-level expression-based association method MetaXcan to examine the association of genetically-predicted expression levels for 11,536 genes across the human genome with breast cancer risk. The MetaXcan method uses external information on the effects of genetic variants on gene expression. We show that the TP53INP2 gene on human chromosome 20 is significantly associated with estrogen-receptor negative breast cancer (P = 5.35×10−7, Bonferroni threshold = 4.33×10−6). The association is consistent across analyses of four datasets, representing European, African and Asian ancestry populations. As a downstream gene of p53, TP53INP2 may affect breast cancer risk through p53 signaling pathway. Furthermore, TP53INP2, also known as DOR (Diabetes And Obesity-Regulated Gene), has been linked to obesity and diabetes, suggesting a novel biological pathway for the known association between obesity and breast cancer risk.
| Breast cancer is the most common cancer in women in the United States and in the world [1]. It is a heterogeneous disease and the two main subgroups of breast cancer are estrogen receptor (ER)-positive and ER-negative cancer. Genome-wide association studies (GWAS) have identified more than 90 susceptibility loci for breast cancer [2–20], with only a few loci specific for ER-negative breast cancer [3,15,17]. Susceptibility loci for ER-positive loci are often the same as loci for overall breast cancer risk because most of breast cancers are ER-positive, especially in women of European or Asian ancestry [2,4,19].
Women of African ancestry are more likely to be diagnosed with ER-negative breast cancer compared to women of non-African ancestry [21–23]. To date, breast cancer GWAS have been conducted primarily in populations of European ancestry. The difference in linkage disequilibrium (LD) patterns and allele frequencies across ancestry groups may explain the apparent inconsistencies in GWAS findings from studies of women of European ancestry as compared to studies of women of African ancestry [24,25]. The strength and the direction of the association between causal variants and disease are expected to be consistent across populations, and thus cross-population validation provides further evidence of causation. In addition, trans-ancestry analysis could identify novel breast cancer susceptibility variants [26].
The variants discovered by previous GWAS along with previously known high-penetrance genes explain only a modest proportion of the heritability of breast cancer [2]. More genetic factors for breast cancer are yet to be identified, but power for discovery of new loci is limited by the sample size of existing GWASs. Moreover, the biologic significance of the variants identified by GWAS and the genes on which they act, are often unknown. Single nucleoid polymorphisms (SNPs) associated with disease traits are more likely to be expression quantitative trait loci (eQTLs) [27], and regulatory variants can explain a large proportion of disease heritability [28]. Therefore, genes regulated by eQTLs can be used as an enrichment analysis unit to identify more genetic risk factors for breast cancer. Recently, gene-based approaches using eQTL information, such as PrediXcan, have been proposed, which can reduce the multiple testing burden in genome-wide analyses and have been used to identify novel genes for autoimmune diseases [29]. PrediXcan uses individual-level data to estimate the correlation between genetically predicted levels of gene expression and human traits to prioritize causal genes. MetaXcan computes the same correlation as PrediXcan, but does so using summary statistics from GWAS, which are much more readily accessible than individual level data [30].
To identify novel genes involved in breast cancer susceptibility, we utilized a gene-level expression-based association method, implemented in the MetaXcan software [30], to infer gene expression levels using summary statistics from five GWASs. We used an additive prediction model of gene-expression levels trained in Depression Genes and Network (DGN) data [31] and examined the predicted expression of specific genes for association with overall breast cancer risk and estrogen receptor-negative breast cancer risk. The GWAS datasets were made available in dbGaP (https://www.ncbi.nlm.nih.gov/gap) through “Up For A Challenge (U4C)–Stimulating Innovation in Breast Cancer Genetic Epidemiology” launched by the National Cancer Institute. The DGN data included RNA sequencing data from whole blood of 922 genotyped individuals (463 cases of major depressive disorder and 459 controls), all of European ancestry. These individuals consisted of 274 males and 648 females with ages ranged from 21 to 60.
Using logistic regression, we first conducted SNP-level GWAS analysis for overall breast cancer risk among 8605 breast cancer cases and 8095 controls, and for ER-negative breast cancer risk among 3879 cases and 10213 controls. The analyses were performed for each of the five GWAS datasets separately and summary statistics including log odds ratios and standard errors were generated. These summary statistics for each dataset were input to the software MetaXcan [30] to perform genome-wide gene-level expression association tests for 11,536 genes. Then, we performed meta-analysis of the results from individual MetaXcan analyses. Quantile-quantile plots of P-values from the meta-analysis showed little inflation (Fig 1). For overall breast cancer risk, there was no gene with a P-value that deviated from the null distribution (Fig 1A), but for ER-negative breast cancer risk analysis, there were several genes with P-values smaller than expected, including TP53INP2, HP, and DHODH (Fig 1B).
Table 1 lists the top genes with P-values less than 10−3 in the analyses of association between predicted gene expressions and overall breast cancer risk. The sign of Z score indicates the direction of association between genetically-predicted expression and breast cancer risk. None of the genes reached genome-wide significance when a Bonferroni threshold (α = 4.33x10-6) was used.
Of the 249 genes that are 250 kb within known susceptibility loci identified from previous breast cancer GWAS [2–4,17,32], 12 genes (4.8%) were statistically significant associated with overall breast cancer risk at nominal significance level of 0.05, compared to 497 (4.4%) of 11,287 genes that are not close to previous GWAS loci (P for enrichment = 0.75).
Table 2 lists the genes with P-values less than 10−3 in the ER-negative breast cancer analysis. TP53INP2 was the top gene (P = 5.35x10-7), which surpassed the Bonferroni-corrected p-value threshold (α = 4.33x10-6). The false discovery rate for TP53INP2 was 0.0062. Higher genetically-predicted TP53INP2 expression was associated with lower risk of ER-negative breast cancer. The gene with the second smallest P-value was HP, which had p-value of 1.70x10-5, close to but not significant after Bonferroni correction. The false discovery rate for the HP gene was 0.098. For the HP gene, higher expression was associated with higher risk of ER-negative breast cancer. Both genes are novel and no previous studies have found association between these two genes and breast cancer risk.
Of the 249 genes that are 250 kb within known breast cancer susceptibility loci identified from previous GWAS, 20 genes (8.0%) were statistically significant associated with ER-negative breast cancer (p<0.05), compared to 582 (5.2%) of 11,287 genes that are not close to previous GWAS loci (P for enrichment = 0.044), suggesting a moderate enrichment for genes close to known susceptibility loci.
There were six SNPs included in the prediction of the expression of the TP53INP2 gene, from 367 kb upstream to 159 kb downstream of the gene (Table 3). Five of the six SNPs (except for rs8116198) were associated with overall breast cancer risk and ER-negative breast cancer risk (at the nominal level of α = 0.05), and the effects were consistently across studies (none of the heterogeneity tests were significant). These associations were more significant for ER-negative breast cancer risk (p values ranging from 5.0x10-4 to 1.8x10-6) than for overall breast cancer risk (7.0x10-4 to 1.4x10-4). None of the SNP-level associations reached traditional genome-wide significance, thus they have not been reported in previous GWAS publications. However, our study showed the aggregate effects of these SNPs were significantly associated with ER-negative breast cancer after Bonferroni correction. We noticed that one of the six SNPs, rs8116198, is monomorphic in the SBCGS data. Therefore, when MetaXcan was applied to the SBCGS data, the prediction of TP53INP2 expression was based on only five SNPs. To make our results more robust to missing and low quality genotypes, in the DGN prediction model, we used elastic net with 0.5 as the mixing parameter, which sets the degree of mixing between ridge regression and LASSO. In addition, the SNPs in the prediction were not necessarily causal but could be in LD with the causal SNPs.
Fig 2 shows positions of the 6 eQTL SNPs for TP53INP2 in the cytoband 20q11.22. Interestingly, there are several other genes in this region that were associated with ER-negative breast cancer, including MAP1LC3A, ITCH, and TRPC4AP (Fig 2 and Table 2). The 6 SNPs are located either in enhancer elements or in promotor regions (Table 4). The promotor/enhancer features of 4 SNPs were found in human mammary epithelial cells (HMEC) and breast variant human mammary epithelial cells (HMEC.35), and the enrichment was statistically significant for both cell types (both p<0.03).
There were 20 SNPs included in the prediction of the expression of the HP gene (S1 Table). Thirteen of the 20 SNPs were associated with overall breast cancer risk and 17 were associated with the risk of ER-negative breast cancer (at the nominal level of α = 0.05), quite consistently across populations (none of the heterogeneity tests were significant). The strengths of their associations were stronger for ER-negative breast cancer risk than for overall breast cancer risk. Interestingly, none of the associations for individual SNPs reached genome-wide significance, thus they have not been reported in previous GWAS publications.
We used summary results from GAME-ON GWAS (http://gameon.dfci.harvard.edu) to replicate our study findings from the U4C. All the six eQTLs for the TP53INP2 gene were available in GAME-ON (Table 5). Five of the six SNPs that were associated with ER-negative breast cancer in the discovery phase (using U4C datasets) were all statistically significant in GAME-ON at the nominal 0.05 significance level. Gene-level test of TP53INP2 from MetaXcan gave a Z-score of -2.803 (p = 5.1×10−3) for ER-negative breast cancer in GAME-ON. The gene-level test for overall breast cancer risk was not significant in GAME-ON (Z-score = -1.627, p = 0.10). Because the GAME-ON ER-negative data included the BPC3 dataset, in order to show the independent replication, we tested association in the U4C ER-negative data excluding BPC3, and found the Z-score for the TP53INP2 gene was -4.127 (p = 3.67×10−5).
For the HP gene, the direction of association for 19 SNPs (out of 20) were consistent between U4C and GAME-ON for ER-negative breast cancer risk, but only 2 SNPs were statistically significant at nominal 0.05 level in GAME-ON (S2 Table). None of the SNPs were significantly associated with overall breast cancer risk in GAME-ON. In the gene-based analysis using GAME-ON data, the Z-score for overall breast cancer risk was 1.769 (p = 0.077) and the Z-score for ER-negative breast cancer risk was 2.02 (p = 0.043). In addition, we tested this association in the U4C ER-negative data excluding BPC3, and found the Z-score for the HP gene was 2.81 (p = 5.1×10−3).
In this gene-level expression-based genome-wide association analysis of five breast cancer GWAS datasets composed of individuals of diverse ancestry, we identified TP53INP2 (20q11.22) as gene with genetically-determined expression that is associated with ER-negative breast cancer. The gene-based analysis of aggregated eQTLs for a particular gene as an analysis unit can reduce the burden of multiple testing and provide a direction of association between expression of a specific gene and disease risk. We found that increased expression of TP53INP2 expression in whole blood was associated with a decrease in ER-negative breast cancer risk. In addition, we identified the HP gene in the 16q22.2 regions to have expression levels that are positively associated with ER- negative breast cancer.
The TP53INP2 gene (tumor protein p53-inducible nuclear protein 2) is 9150 base pairs long and codes for a 220 amino acid protein, which is a dual regulator of transcription and autophagy and is required for autophagosome formation and processing. One experimental study showed that overexpression of TP53INP2 severely attenuated proliferative and invasive capacity of melanoma cells, via p53 signaling and lysosomal pathways [34]. This inverse correlation between TP53INP2 expression and cancer proliferation is consistent with our finding that TP53INP2 expression inversely correlated with breast cancer risk. P53 is a transcription factor for TP53INP2, and TP53 plays an important role in development of multiple cancers. Germline TP53 mutations cause Li-Fraumeni syndrome, characterized as a cluster of cancers including breast cancer [35]. Somatic TP53 mutation is a common event in ER-negative breast cancer [36]. As a downstream gene of p53, TP53INP2 may affect breast cancer risk through p53 signaling pathway. Also, known as DOR (diabetes- and obesity-regulated gene), TP53INP2 has been linked to obesity and diabetes [37]. TP53INP2 is also associated with triglycerides and cholesterol level. One experimental study found that dietary fat content influenced the expression of TP53INP2 expression in adipose and muscle tissues of mice [38]. This gene has been proposed to serve as a diagnostic biomarker for papillary thyroid carcinoma [39] but no study has linked its expression to cancer risk. Obesity has been convincingly correlated with breast cancer risk in numerous studies, although the relationship is complex and involves additional modifying factors [40,41]. Obesity has been associated with excess risk for breast cancer among postmenopausal women [42–46], while in pre-menopausal women, obesity was associated with decreased breast cancer risk [40,43,47–49]. However, the underlying mechanisms for this association are still not fully understood. The identification of TP53INP2/DOR as breast cancer-related gene could provide novel insight on the mechanism for obesity-breast cancer relationship.
In the 20q11.22 region, several other genes including MAP1LC3A, ITCH, and TRPC4AP were associated with ER-negative breast cancer risk. MAP1LC3A codes for a protein that is important in the autophagy process, and was found to be expressed at higher level in breast cancer tissues than in normal tissues [50]. E3 ubiquitin ligase ITCH plays a role in erythroid and lymphoid cell differentiation and immune response regulation, and ITCH was found to be important in the cross-talk between the Wnt and Hippo pathways in breast cancer development [51]. TRPC4AP is involved in Ca2+ signaling and is part of the ubiquitin ligase complex [52,53]. It is unclear which of these genes (or their interactions) play a role in breast cancer development, but the 20q11.22 locus is worthy of further investigation. Three of the six SNPs for TP53INP2 (rs6060047, rs11546155, and rs1205339) are also shared by the genes MAP1LC3A and TRPC4AP. It is possible that the associations in these three genes are partly generated by the overlapped SNPs, which contribute to predicted expression levels of the three genes and, possibly, to the enrichment observed at this locus.
The HP gene (16q22.2) is 6,491 base pairs long and codes for a 406 amino acid preprotein, which codes haptoglobin. Haptoglobin binds to hemoglobin to prevent iron loss during hemolysis. There are two allelic forms, Hp1 (83 residues) and Hp2 (142 residues), which determine 3 major phenotypes [54]. Haptoglobin genotype has been linked to cardiocerebral outcomes among diabetic patients [55]. A small study found haptoglobin phenotypic polymorphism was associated with familial breast cancer [56], but no studies have reported on the relationship between SNPs in this gene and breast cancer risk. Further larger studies could investigate the relationship between major HP genotype/phenotype (HP1-1, HP1-2, and HP2-2) and breast cancer risk.
The present study has several strengths, including its large sample size, diverse ancestry groups, a cross-replication approach, and a novel gene expression-based analysis method. The gene-level analysis method can combine eQTL SNPs in a biologically informative way to assess relationships between predicated gene expression and disease risk. Compared to SNP-based analysis, the gene-based analysis can gain power by reducing the multiple testing burden by about 100-fold and using external information on correlation between gene expression and SNPs from reference samples. In addition, this approach enables the detection of individual SNPs with weak effects on disease risk by leveraging combined effects of multiple SNPs on gene expression. For example, none of SNPs for TP53INP2 reached traditional genome-wide significance, but their aggregated effect via TP53INP2 expression was genome-wide significant. The gene-based method (MetaXcan) that we employed is an extension of the gene expression-based method (PrediXcan) [29] and allows the use of SNP-level summary statistics without the need to access individual-level genotype data [30]. The MetaXcan method has been shown to produce PrediXcan results accurately, and it is robust to ancestry mismatches between studies and reference/training populations [30]. With this property, we were able to use summary statistics from the GAME-ON consortium for external replication.
Several limitations should be considered when interpreting the study findings. The gene expression-based association method relies on accurate prediction of gene transcript level from genotypes, i.e. identification of eQTLs, but eQTL identification depends on sample size of eQTL studies as well as tissue types. In the current study, we used the transcriptome prediction model that was developed using 922 RNA-seq samples from whole blood and genotype data [31]. Although it has been shown that models developed in whole blood were still useful for understanding diseases that affect other primary tissues [29], we expect there to be a loss of power when studying non-blood diseases using whole blood eQTL data. As a sensitivity analysis, we performed the MetaXcan analysis using the prediction model from breast tissues of 183 donors of multiple ethnicities (http://www.gtexportal.org). Only 4,308 genes had breast tissue specific eQTLs, and no eQTL was available for TP53INP2, perhaps due to the small sample size. We found that DHODH (P = 3.61×10−5), ITCH (P = 1.23×10−4), and TRPC4AP (P = 7.7x10-4) were among the top genes associated with ER-negative breast cancer risk, and TRPC4AP (P = 1.68×10−5) and DHODH (P = 1.12×10−4) among the top genes associated with the overall breast cancer risk using breast tissue eQTLs. In the enrichment analysis, we found that 7 (8.2%) out of 85 genes that are close to known breast cancer susceptibility loci identified in previous GWAS were associated with ER-negative breast cancer and 6 (7.1%) genes were associated with overall breast cancer risk; by contrast, of the 4223 genes away from previous GWAS loci, 199 (4.7%) genes were associated with ER-negative breast cancer and 212 (5.0%) genes were associated with overall breast cancer risk. Here, we have to consider the balance between tissue relevance and sample size in eQTL studies. Further investigations based on large, reliable eQTL datasets are desirable. In future studies, we will seek out larger samples of multi-ethnic breast tissue as training data to construct improved prediction models of gene expression and further investigate trans-ethnic associations for breast cancer.
In conclusion, our study identified TP53INP2 and several other genes in the 20q11.22 region as potential susceptibility genes for ER-negative breast cancer using a novel gene-based analysis method that incorporates genetically determined gene expression. We demonstrated this gene-based method increases statistical power and may be helpful in searching for causal variants. Future studies need to determine whether the TP53INP2 gene is a true susceptibility gene for breast cancer and what are the underlying mechanisms for its association with ER-negative breast cancer.
The study was approved by the Institutional Review Board of the University of Chicago. The Epidemiology and Genomic Research Program within the National Cancer Institute launched a Challenge at the end of 2015 to inspire novel cross-disciplinary approaches to more fully decipher the genomic basis of breast cancer, called "Up For A Challenge (U4C)–Stimulating Innovation in Breast Cancer Genetic Epidemiology”. Several data sets were gathered and made available for use in dbGap (https://www.ncbi.nlm.nih.gov/gap). Our study has two phases; the discovery phase included five U4C GWAS datasets (Table 6). Here, we refer them collectively as “U4C” data. These data were collected from three distinct ancestry groups. The BPC3 [16,18] and CGEMS study [15,20] were conducted in women of European ancestry. The ROOT [17] and AABC study [57] consisted of women of African ancestry. The SBCGS study was conducted in Chinese population [19]. For the analysis of overall breast cancer risk, we used four GWAS datasets: AABC, CGEMS, ROOT, and SBCGS. For the analysis of ER-negative breast cancer risk, we used datasets from AABC, BPC3, ROOT, and SBCGS. All these dbGap datasets included imputed genotype data that were inferred based on reference haplotypes from the 1000 Genomes project.
In the replication phase, we used the summary results from the meta-analysis of 11 breast cancer GWASs in the GAME-ON consortium (http://gameon.dfci.harvard.edu). All participants were of European ancestry. The overall breast cancer analysis included 16,003 cases and 41,335 controls from 11 GWAS studies; The ER-negative breast cancer analysis included 4939 cases and 13128 controls from 7 GWAS studies. The dataset from one study (BPC3; all ER-negative cases) in GAME-ON consortium was also included the U4C datasets. Because only meta-analysis results were available from GAME-ON, we removed the BPC3 data from “U4C” dataset when we compared replication performance to avoid duplicate counting.
Our gene level expression-based association analysis consists of three main steps. First, we conducted SNP-level genome-wide association tests and calculated summary statistics such log odds ratios and their standard errors. We used logistic regression model adjusting for eigenvectors from the principal component analysis and related covariates such as age. Genotypes were coded by an additive genetic model. Eigenvectors in principal component analysis were calculated using the method smartPCA, which is implemented in the software EIGENSOFT version 6.0.1 [58]. For the ROOT dataset, we adjusted for age, study sites, and the top 4 eigenvectors. For the AABC dataset, we adjusted for age, study sites, and top 10 eigenvectors. For CGEMS and SBCGS, we adjusted for age and the top three or two eigenvectors, respectively. The number of eigenvectors we adjusted for was chosen according to published papers from these GWASs [17,57], as well as their association with case-control status. The logistic regression models were fit using software Mach2dat (http://www.unc.edu/~yunmli/software.html) or SNPtest [59], depending on format of the datasets; the Mach2dat software was used for CGEMS and SBCGS and SNPtest was used for ROOT and AABC. For BPC3, the GWAS summary statistics for ER-negative breast cancer have been pre-calculated in the dbGap release, so we used them directly.
Second, we applied the gene level association method, MetaXcan [30] (https://github.com/hakyimlab/MetaXcan), to each of the datasets listed in Table 6. MetaXcan is an extension of the method PrediXcan [29], which uses an additive genetic model to estimate the component of gene expression determined by an individual’s genetic profile and then identifies likely causal genes by computing the correlations between genetically predicted gene expression levels and disease phenotypes. MetaXcan infers the results of PrediXcan using summary statistics from GWAS, which are much more readily accessible than individual level data. In our study, as input files for MetaXcan, we used summary statistics from SNP-based analysis of each dataset obtained in step one. In addition, we used the whole blood genetic prediction model of transcriptome levels trained in the DGN data [31], which can be downloaded from http://predictdb.hakyimlab.orghttps://s3.amazonaws.com/predictdb/DGN-HapMap-2015/. The DGN data provides a large reference sample of 922 individuals with both genome-wide genotype data and RNA sequencing data. The model trained in the DGN data can be useful in estimating gene expression levels and has been successfully applied to the Wellcome Trust Case Control Consortium (WTCCC) data in identifying genes associated with five complex diseases [29]. The DGN prediction model includes a) weights for predicting gene expression using genotypes and b) covariance of the SNPs that takes into account linkage disequilibrium. We tested the association between predicted expression levels of 11,536 genes for each of the two phenotypes, overall and ER-negative breast cancer risk, using the MetaXcan software. To construct the prediction model of expression levels using the DGN data, MetaXcan used SNPs with minor allele frequencies (MAFs) >0.05. When MetaXcan was applied to the breast cancer GWAS data, only SNPs with MAFs >0.05 were used. We also looked up genes within 250 kb of the 93 breast cancer susceptibility loci identified in previous GWAS [2–4,17,32].
Third, we conducted meta-analysis to combine results from MetaXcan analyses for different datasets. The method described by Willer et al. with sample size as meta-analysis weight [60] was used. We also conducted SNP-level meta-analysis using a fixed effect model, as implemented in the software METAL (http://genome.sph.umich.edu/wiki/METAL). False discovery rates were calculated using the Benjamini and Hochberg method [61].
For genes identified in the discovery phase using the U4C datasets, we conducted replication analysis using GAME-ON summary results using the same methods described above. For each top variant and gene identified in this study, we used HaploReg [33] and USCS Genome Browser to explore functional annotations of noncoding variants. Chromatin states (promoters and enhancers), variant effect on regulatory motifs, and protein binding sites were assessed from available data from the ENCODE [62] and Roadmap Epigenomics Consortium [63]. Data from normal mammary epithelial cells (HMEC, MYO, vMHEC) were emphasized.
|
10.1371/journal.pcbi.1004635 | Melanoma Cell Colony Expansion Parameters Revealed by Approximate Bayesian Computation | In vitro studies and mathematical models are now being widely used to study the underlying mechanisms driving the expansion of cell colonies. This can improve our understanding of cancer formation and progression. Although much progress has been made in terms of developing and analysing mathematical models, far less progress has been made in terms of understanding how to estimate model parameters using experimental in vitro image-based data. To address this issue, a new approximate Bayesian computation (ABC) algorithm is proposed to estimate key parameters governing the expansion of melanoma cell (MM127) colonies, including cell diffusivity, D, cell proliferation rate, λ, and cell-to-cell adhesion, q, in two experimental scenarios, namely with and without a chemical treatment to suppress cell proliferation. Even when little prior biological knowledge about the parameters is assumed, all parameters are precisely inferred with a small posterior coefficient of variation, approximately 2–12%. The ABC analyses reveal that the posterior distributions of D and q depend on the experimental elapsed time, whereas the posterior distribution of λ does not. The posterior mean values of D and q are in the ranges 226–268 µm2h−1, 311–351 µm2h−1 and 0.23–0.39, 0.32–0.61 for the experimental periods of 0–24 h and 24–48 h, respectively. Furthermore, we found that the posterior distribution of q also depends on the initial cell density, whereas the posterior distributions of D and λ do not. The ABC approach also enables information from the two experiments to be combined, resulting in greater precision for all estimates of D and λ.
| Quantifying the underlying parameters that drive the expansion of melanoma cell colonies such as the cell diffusivity, cell proliferation rate and cell-to-cell adhesion strength can improve our understanding of melanoma biology and its response to treatment. We combine a simulation-based model of collective cell spreading with a novel Bayesian computational algorithm to estimate these parameters from carefully chosen summaries of collective cell image data and to quantify their associated uncertainty across different experimental conditions. Our summarisation of the image data leads to precise estimates for all parameters. Our analysis reveals that the cell diffusivity and the cell-to-cell adhesion strength estimates depend on experimental elapsed time. Furthermore, the cell-to-cell adhesion strength estimate appears to depend on the initial cell density, whereas the cell proliferation rate estimate is approximately the same over different experimental conditions.
| Skin cancer consists of two groups: melanoma and non-melanoma. Melanoma is the least common, approximately 5% of all skin cancer occurrences, but it is responsible for most skin cancer deaths [1]. It is estimated that 132,000 new cases of melanoma are reported worldwide each year, with more than 12,500 of these cases reported in Australia [2]. During the early stage of the disease, melanoma colonies grow and spread laterally within the epidermis. Thus, quantifying the underlying mechanisms that drive the expansion of melanoma cell colonies such as motility, proliferation, and cell-to-cell adhesion can improve our understanding of melanoma biology and its response to treatment.
Although much progress has been made in terms of developing and analysing mathematical models of expanding cell colonies, far less progress has been made in terms of understanding how to estimate model parameters including the cell diffusivity, D, the cell proliferation rate, λ, and the cell-to-cell adhesion, q, from experimental in vitro image-based data. Obtaining precise estimates of D, q and λ is important for developing a systematic approach to assessing the effectiveness of a potential treatment [3]. Several studies have investigated the in vitro expansion of cell colonies using partial differential equations [4–7]. These approaches are limited in that they provide point estimates, and the uncertainty in the estimate is not quantified. An alternative modelling approach uses discrete, individual-based models [8–10], which can incorporate several important biological factors such as cell heterogeneity [11]. Discrete models can also produce discrete image-based and video-based information which is ideally suited to collaborative investigations involving applied mathematicians and experimental cell biologists. However, the likelihood functions for these discrete models are generally intractable, so standard statistical inferential methods for these models are not applicable.
To overcome these issues, an approximate Bayesian computation (ABC) approach is developed to jointly infer the values of D, q and λ from a discrete stochastic model describing the expansion of cell colonies. ABC is a well established method that has been successfully applied in a wide range of areas such as population genetics [12], infectious diseases [13, 14], astronomical model analysis [15] and cell biology [16]. Generally, ABC approximates the likelihood function by model simulations, the outcomes of which are compared with the observed data [16, 17]. In this paper, we propose a new ABC algorithm that is shown to be more efficient than state-of-the-art algorithms available in the literature [17–20] by developing a new sequential Monte Carlo approach. ABC requires the specification of a set of summary statistics to compare the observed and simulated data. Each of our experimental datasets is initially summarised using a high dimensional vector of summary statistics (hereafter referred to as the pilot summary statistics). Unfortunately, ABC is not able to handle high dimensional summary statistics in an efficient manner [21], so we adopt a semi-automatic approach [22] to reduce the dimension of the pilot summary statistics. Using a synthetically generated dataset, we demonstrate that combining our new ABC algorithm and the derived set of summary statistics can precisely recover all parameters.
We apply this procedure to the experimental data of human malignant melanoma cells (MM127) in a barrier assay [23] in two different experimental scenarios: (1) Mitomycin-C is applied as a treatment to suppress cell proliferation, and (2) no treatment is applied. We aim to obtain a joint approximate posterior distribution for D, q and λ for different combinations of initial cell densities, C(0), and experimental times, T, in each scenario. Through the ABC analyses, the associated uncertainty in the parameter values is quantified and interpreted in terms of the coefficient of variation (CV) and probability intervals of the posterior distribution. Thus, our work adds significant extra information about model parameters relative to the previous analysis [23], which obtained point estimates of D, q and λ separately. In the previous analysis [23], D and q were estimated only from the experiments with cell proliferation suppressed.
Previous approaches often assume that these parameter values are the same over different experimental conditions [3, 23, 24]. The findings from this study show that the posterior estimate of D appears to depend on experimental time and weakly depend on the initial cell density, which is consistent with the results reported in Vo et al. [16] for 3T3 fibroblast cells. A similar trend of dependency is also found for q; but in contrast the posterior estimates of λ remain similar over time. These results suggest that a more complicated model might be warranted. However, this finding could not have been achieved without first exploring the suitability of the standard model under consideration here.
The experimental data analysed in Vo et al. [16] also consists of two separate scenarios, with and without Mitomycin-C pre-treatment. Vo et al. [16] demonstrate that λ cannot be identified by leading edge data solely, unless prior information about D (obtained from the experiment with the treatment applied) is incorporated via a sequential Bayesian learning approach. In this paper, we show that all parameters (including λ) can be estimated precisely through the inclusion of additional summary statistics (cell densities and percentages of isolated cells) even when only vague prior information is specified for parameter values. Nonetheless, we show that the Bayesian sequential learning approach [16] is still useful here as we are able to obtain greater precision of the parameter values.
The details of the experimental method were described previously [23]. Briefly, monolayers of human malignant melanoma cells (MM127, [25, 26]) were cultured in 24-well tissue culture plates, where each well had a diameter of 15.6 mm. Experiments were conducted in two different experimental scenarios: (1) with Mitomycin-C pre-treatment to suppress cell proliferation, and (2) without Mitomycin-C pre-treatment. Mitomycin-C, an alkylating antibiotic, is used to block DNA and RNA replication and protein synthesis. Thus, given an appropriate concentration, Mitomycin-C inhibits mitosis and proliferation of several cell types [27]. For the melanoma experiments here, 10 µg ml−1 Mitomycin-C was added to the cells one hour prior to transfer to the wells [23].
To initiate each experiment, either 20,000 or 30,000 cells were approximately evenly distributed within a circular barrier, of diameter 6.0 mm, located at the centre of the well. After allowing the cells to attach for 1 h, the barriers were lifted and population-scale images were recorded at either 24 h or 48 h, independently. To extract detailed information about the location of individual cells in the population, high magnification images of a transect across the centre of the cell population were also acquired, where the nuclei were stained with Propidium Iodide (PI). Furthermore, each experimental scenario, for each initial cell density and each termination time, was repeated three times. Thus, a 2 × 2 × 2 balanced experimental design was conducted with three replicates, producing a total of 24 independent experimental images of expanding cell colonies and the corresponding transect images.
From preliminary analysis, we note that cell colonies maintain an approximately circular shape during the experiments. Thus, for each population-scale image, which shows the spatial expansion of the entire melanoma cell colony, we detect the position of the leading edge, then estimate the radius of the colony by converting the area enclosed by the leading edge to the equivalent circular radius, R, using a segmentation algorithm written with the Matlab Image Processing Toolbox [16, 28] (Table S1 in S1 Text). We use the exact same edge detection algorithm for both our experimental data and the images produced by the discrete simulation model described in the next section. Images in Fig 1A–1C show the entire expanding cell colonies for the 30,000 initial cell experiments at time 0 h and 48 h where cells were pre-treated with Mitomycin-C, and 48 h without the treatment, respectively, together with the estimated leading edge superimposed.
To extract cell densities and measure cell clustering, we mapped the position of the cells to a square lattice with spacing Δ = 18 µm (Fig 1H and 1I), which corresponds to an average diameter of the cell nucleus [23]. For each experiment, we analyse six sub-regions along a transect image (Fig 1G). Each sub-region has size 700 × 500 µm or 39 × 28 lattice sites. We then count the number of cells in each sub-region, { c i } i = 1 6, together with the proportion of isolated cells, { p i } i = 1 6. A cell is identified as isolated if all of its nearest neighbours (north, south, east and west) are unoccupied. For each experiment at each initial cell density and termination time, at each sub-region, we average ci and pi over three replicates (Tables S2 and S3 in S1 Text).
Summaries of { c i } i = 1 6 and { p i } i = 1 6 (average over the three replicates) for experiments initialised with 20,000 cells are given in Fig 2. We observe that, for experiments where cells were not pre-treated with Mitomycin-C (Fig 2B), { c i } i = 1 6 increases significantly over time, whereas the differences in { c i } i = 1 6 for the corresponding experiments (Fig 2A), where cell proliferation was suppressed, are minimal. Furthermore, { p i } i = 1 6 (Fig 2C and 2D) appear to decrease over time which suggests that melanoma cells possibly form more clusters as the experiments proceed. These trends are consistent with previous research [23], which shows that cell-to-cell adhesion plays an important role in the melanoma expanding colonies.
To describe the expansion of a single layer of melanoma cell colonies, we employ a discrete lattice based model that incorporates cell migration (unbiased random walk), cell proliferation and cell-to-cell adhesion. The discrete model here is similar to the model used in [9, 16, 23]. We incorporate a volume exclusion process and realistic crowding effects [8, 9, 29], so each lattice site can be occupied by at most one agent.
To simulate the experiments, we use a two-dimensional square lattice of size 867 × 867, with lattice spacing Δ = 18 µm, so that the width of the lattice corresponds to the diameter of the well, 15.6 mm (15600 µm/18 µm = 867). Let C(t) be the number of agents in the discrete model at time t, Pm ∈ [0, 1] be the probability that an isolated agent will attempt to step a distance Δ within a time step of duration τ, and Pp ∈ [0, 1] represent the probability that an agent will attempt to proliferate and deposit a daughter within a time step of duration τ. The strength of cell-to-cell adhesion is represented by q ∈ [0, 1].
Initially, C(0) agents (20,000 or 30,000 agents) are placed randomly inside a circle which has a radius of 177 lattice sites, corresponding to the mean radius of the experimental observations at time t = 0 h. We use an approximate random sequential update (RSU) algorithm [30, 31] to perform the simulations. To step from time t to time t + τ, C(t) agents are sampled, with replacement, and given the opportunity to move with probability Pm × (1 − q)n, where 0 ≤ n ≤ 4 is the number of occupied nearest neighbour sites. If an agent is at position (x, y) and has an opportunity to move, it will attempt to step to either (x ± Δ, y) or (x, y ± Δ), with each target site chosen with equal probability. The higher the value of q, the more difficult it is for an agent to move away from its neighbours.
A similar mechanism is employed for proliferation events. A proliferative agent at position (x, y) will attempt to deposit a daughter agent at (x ± Δ, y) or (x, y ± Δ), with each target site chosen with equal probability. Since the model is an exclusion process, any attempted motility or proliferation event that would place an agent on an occupied site is aborted (S1 Text, Algorithm S1). We do not consider any death mechanism in this model since there was no evidence of any cell death in the experiment [23]. Given the termination time, T (24 h or 48 h), the model requires T/τ time steps.
The cell expanding colonies are governed by three parameters (Pm, q, Pp). These parameters are related to the cell diffusivity, D, and the proliferation rate, λ, by D = Pm Δ2/4τ and λ = Pp/τ, respectively [29], with Δ and τ set fixed. In this work, we apply our new ABC algorithm to obtain joint posterior distributions for (Pm, q, Pp), then use these relationships and the values of Δ and τ, to rescale posterior distributions of Pm and Pp into posterior distributions of D and λ, respectively.
We note that the RSU algorithm is an approximation of the exact, continuous time Gillespie algorithm [32]. The value of the time duration τ is a trade-off between the accuracy of the approximation and the computational time to simulate the experiments. To choose a suitable value for τ, we perform 100 model simulations using the same diffusion coefficient D = 220 µm2h−1, obtained with different pairs of parameters (τ = 0.1 h, Pm = 0.2716), (τ = 0.08 h, Pm = 0.2173), (τ = 0.06 h, Pm = 0.1630), (τ = 0.04h, Pm = 0.1086) and (τ = 0.02 h, Pm = 0.0543). We then compare the plots of the probability density of the resulting radii, percentages of isolated cells and total number of cells in six sub-regions. We found that there is a negligible difference between results from simulations with τ = 0.04 h and τ = 0.02 h. This means that τ = 0.04 h is small enough to produce reasonably accurate simulations. Therefore, for all model simulations hereafter, we use τ = 0.04 h. Snapshots of the discrete stochastic models initialised with 30,000 agents and termination time at 0 h, 48 h in Scenario 1, and 48 h in Scenario 2 are shown in Fig 1D–1F, respectively.
In this paper, we do not have any measurement for the uncertainty in C(0). Thus, all of the simulations from the discrete models use the same initial values of C(0), i.e. 20,000 cells or 30,000 cells. However, if we have this measurement, we can easily incorporate it in the ABC algorithms by drawing the value of C(0) from its distribution before proceeding to simulate a realisation of the model.
The discrete models described above can incorporate realistic cell behaviour. However, their likelihood functions are not available in an analytical form and are not computationally tractable, so standard statistical inferential methods for these models are not applicable. Combining ABC and the discrete stochastic model is a promising approach since ABC bypasses the evaluation of the likelihood by a simulation-based procedure [12, 17]. The aim of the ABC approach is to find the joint approximate posterior distributions, which are the distributions of the unknown parameters given the observed summarisation of the data and the prior information. All inferences about the parameters including point estimates and probability intervals are made from the posterior distributions.
Let yobs and ysim represent the observed and the simulated data, θ = (Pm, q, Pp) represent the vector of unknown parameters and π(θ) be the prior distribution for θ. We define a distance metric ρ which is a function of yobs and ysim, ρ = ρ(yobs, ysim). ABC approaches consist of four major steps: sampling a proposed parameter θ⋆, simulating data as per the observed data structure from the model with θ⋆, comparing ysim with yobs by computing ρ = ρ(yobs, ysim) and accepting the proposed θ⋆ if ρ(yobs, ysim) ≤ ϵ, where ϵ ≥ 0 is a tolerance value. The accepted sample of parameter values forms the approximation of the posterior distribution of the model parameters. The choice of ϵ is a trade-off between accuracy and computational effort. In practice, different ABC algorithms have different approaches to sample the values of θ⋆.
ABC rejection is the simplest ABC algorithm, which generally samples θ⋆ from the prior distribution. This algorithm is easy to implement and is embarrassingly parallel. However, for complex models where the prior distribution is substantially different from the posterior, this approach results in low acceptance rates and is computationally inefficient. Vo et al. [16] employed the ABC rejection algorithm to estimate D and λ, using the leading edge data of 3T3 fibroblast cell populations. This study samples a large number of proposed parameters from the prior, each with a corresponding artificial dataset and a value of discrepancy ρ. These parameters are then sorted by their discrepancies and only a small proportion of parameters with the lowest discrepancy are retained. In the study of Vo et al. [16], a uniform prior was used, suggesting that for a reasonably low ϵ, the proportion of parameters being kept is very small, approximately 0.1%. Thus, this study suggests that it is necessary to generate 106 model simulations to obtain an ABC posterior sample of size 1,000.
Several studies [33–35] proposed a Markov chain Monte Carlo approach to ABC (MCMC-ABC). MCMC-ABC algorithms make local proposals in high (ABC) posterior support regions, thus they can improve the acceptance rates. However, the posterior samples are highly correlated and the algorithms can easily be trapped in regions of low posterior density [35]. Another class of ABC is SMC-ABC which was pioneered by [36] to overcome the problems associated with ABC rejection and MCMC-ABC. SMC-ABC algorithms involve sampling from a sequence of ABC posterior distributions with a non-increasing sequence of tolerances, { ϵ k } k = 1 M. Thus, this last class of ABC only draws proposed parameters in sequentially higher posterior support regions, rather than the entire parameter space. A review of ABC algorithms can be found in [37].
In this paper, we only focus on SMC-ABC algorithms. Instead of drawing a proposed value θ⋆ one at a time, the SMC algorithms work with a large set of parameter values simultaneously and treat each parameter vector as a particle. The particles are moved and filtered at each stage of the algorithm. Initially, a set of N particles, { θ i } i = 1 N, is often sampled from the prior distribution π(θ) and each sampled particle has an equal weight of 1/N. To propagate a particle from iteration k − 1 to iteration k, SMC-ABC algorithms involve three steps: (i) re-sampling: a sampled particle candidate θ⋆ is chosen randomly from the set of particles at k − 1 with probability proportional to their weights, θ ⋆ ∼ { θ i k - 1 , W i k - 1 } i = 1 N; (ii) perturbing: the particle candidate θ⋆ is perturbed by a transition kernel to propose a new particle θ⋆⋆, θ⋆⋆ ∼ Kk(⋅|θ⋆), and (iii) simulating ysim from the model, ysim ∼ f(⋅|θ⋆⋆). To maintain N particles throughout the algorithm, the steps (i-iii) are repeated until a parameter value is found such that the condition ρ(yobs, ysim) ≤ ϵk is satisfied. Different SMC algorithms can be distinguished by the transition kernel, the schedule of the tolerances and how sampling weights are assigned to the particles.
In the literature, there are several versions of SMC-ABC algorithms. For example, SMC-ABC algorithms of [18, 19, 38] use a Gaussian Markov kernel with a covariance matrix as twice the empirical covariance matrix of the current set of particles. These algorithms also assign to each particle θk a weight given by:
W k ∝ π ( θ k ) ∑ j = 1 N W j k - 1 K k ( θ k | θ j k - 1 ) . (1)
These algorithms have the advantage that they require fewer model simulations, although the sequence of tolerances in these algorithms is determined manually. Drovandi et al. [17] and Del Moral et al. [39] proposed an adaptive SMC algorithm that can determine a decreasing set of tolerances dynamically. This can be achieved by sorting the particles by their discrepancies and then dropping a proportion of the particles with the highest discrepancy. However, these algorithms use an MCMC kernel which has a drawback of replications of particles. To reduce this problem, Drovandi et al. [17] suggest to repeat the MCMC step (steps (ii) and (iii) above) a number of times, which also can lead to a large number of unused model simulations.
We take the advantage of fewer model simulations from SMC-ABC algorithms [18, 19, 36] and the advantage of automatically determining tolerance values from [17] (also named the SMC replenishment (RSMC) algorithm) and incorporate these in one algorithm, hereafter referred to as ASMC (S1 Text, Algorithm S2). Our ASMC algorithm is similar to that proposed in [20] (also named adaptive population Monte Carlo (APMC) algorithm) who also determine the sequence of tolerances adaptively and use the re-weighting scheme above. However, in each iteration, the APMC algorithm [20] only performs steps (i-iii) above once and keeps all the N particles, so the particle’s discrepancy value is not enforced to be below a particular tolerance.
In the APMC algorithm, the sequence of tolerances fluctuate, whereas the sequence of tolerances in the RSMC and ASMC algorithms is always non-increasing. Therefore, we cannot use a single indicator to compare the performance of the three algorithms. We suggest comparing the RSMC and ASMC using the final tolerance, and comparing the ASMC and APMC using the same computational effort. Using synthetically generated data, we show that our algorithm requires fewer model simulations than the RSMC algorithm [17], given the same target tolerance ϵfinal. In addition, given the same number of simulations, our algorithm is shown to produce a lower tolerance value (thus higher accuracy) relative to the APMC algorithm [20].
To examine the utility of our new ABC algorithm and to investigate whether the derived set of summary statistics is informative for parameter inferences, we simulated a dataset with biologically relevant parameter values (Pm = 0.1, q = 0.2, Pp = 0.0012), which corresponds to (D = 202.5 µm2h−1, q = 0.2, λ = 0.03h−1). The synthetic dataset has C(0) = 20,000 cells, T = 24 h and is replicated three times. This dataset represents experiments in Scenario 2.
We first summarise the synthetic dataset in terms of the pilot summary statistics, including three radii of the expanding cell colonies for three replicates (order statistics), the numbers of cells and the percentages of isolated cells in six sub-regions along a transect after averaging over three replicates. The ABC posterior distributions resulting from the pilot run with the pilot summary statistics have significant spread. So, a multiple linear regression procedure is performed to generate one summary statistic for each parameter. We then apply the new ABC algorithm with the derived set of summary statistics and uniform priors for all parameters, Pm ∼ U(0,1), q ∼ U(0,1) and Pp ∼ U(0,1). The resulting posterior distributions for (Pm, q, Pp) are presented in Fig 3. These results show well-defined posterior distributions with narrow spread and posterior means close to the true values. The posterior correlation coefficients of (Pm, q), (q, Pp) and (Pm, Pp) are between −0.2 to 0.3. Thus, it is evident that our new ABC algorithm combined with our method for determining summary statistics allows us to recover all parameters rather precisely.
Using the synthetically generated data, we also compare the performance of the three algorithms: RSMC, APMC and ASMC. For all algorithms, we set N = 1000 particles and run each algorithm 10 times to compare the resulting posterior distributions, the total number of model simulations and the generalized variance (GV, or the determinant of the posterior variance-covariance matrix). A comparison of ABC posterior distributions from the three algorithms is shown in Fig 4. For RSMC and ASMC, we set ϵfinal = 0.1. For all cases, the posterior distributions from RSMC (the dashed black curves) and ASMC (the solid red curves) are almost indistinguishable, however, the RSMC requires approximately 2.5 times more model simulations than the ASMC algorithm (Fig 5A).
For the APMC algorithm, we use 62 iterations (giving the total number of model simulations similar to the number of model simulations for ASMC, 62,000). Results in Fig 4 suggest that the posterior distributions from the ASMC algorithm has smaller variance than the results from the APMC algorithm (the blue curves with markers) due to the ability of ASMC in getting to a smaller value of ϵ with a similar computational effort. We then compute the GV of the resulting ABC joint posterior distributions from the ASMC and APMC algorithms from the 10 runs (Fig 5B). We observed that the GVs for the resulting posterior distributions from APMC are approximately three times larger than the corresponding GV from the ASMC algorithm. Thus, for this application, our algorithm performs better than the RSMC and the APMC algorithms. We now apply the ASMC algorithm to the experimental data in the two scenarios and interpret the results in terms of the biologically relevant parameters D, q and λ.
This section presents the results for D and q for all experimental conditions in Scenario 1, where cells were pre-treated with Mitomycin-C to suppress cell proliferation. Uniform priors are placed on all parameters, Pm ∼ U(0,1) and q ∼ U(0,1). From the regression procedure to generate one summary statistic S for each parameter, for all cases, we observe that all pilot summary statistics (R(1), R(2), R(3), { c i } i = 1 6 and { p i } i = 1 6) are informative about D. However, to obtain estimates for q, only R(1), { c i } i = 1 6 and { p i } i = 1 6 were significant in the regression.
The ABC estimate of the posterior expected value of D and q, E[D] and E[q], 90% credible intervals, CI, the coefficient of variation, CV, and the correlation coefficient, r, from all experimental conditions, are given in Table 1. To assess the accuracy of our resulting estimates from the true ABC posteriors, we computed the Monte Carlo standard error, MCSE, for E[D] and E[q] in all experimental conditions, MCSE = σ / ESS [41]. Here, σ is the posterior standard deviation and ESS is the effective sample size. We use Kish’s approximation method [42] to compute the ESS, ESS = 1 / ∑ i = 1 N W i 2, where Wi is the normalised weight for the ith parameter value. For all cases, the ABC posterior consists of 1,000 parameter values, which leads to an ESS usually in the range 700–850. Our posterior sample size leads to a small MCSE for both E[D] and E[q], less than 0.2% and 0.4% of the estimate of their expected values, respectively.
From Table 1, we observe that the CV for D and q are also quite small, approximately 6% and 10%, respectively, which means that we can obtain reasonably precise estimates for D and q using the derived summary statistics. The correlation coefficient between D and q for all combinations is between 0.2 to 0.6. This suggests that multiple combinations of values of D and q can generate similar expanding cell colonies in terms of our pilot summary statistics.
For both initial cell densities (20,000 and 30,000 cells), we observe that the values of E[D] for the experiments terminated after 48 h are higher than those values for experiments terminated after 24 h. This finding suggests that estimates of D appear to depend on the experimental time, T, which is consistent with the results reported in [16] for 3T3 fibroblast cells. It is conjectured that some amount of time could be required for the cells to adjust to their new or modified environments encountered as part of the experimental protocol. The cell motility, therefore, could be reduced during this transition phase. A similar trend of dependency is also found for q. This motivates us to investigate the values of D and q for the period 24–48 h.
Let {D(0–24), q(0–24)}, {D(24–48), q(24–48)} and {D(0–48), q(0–48)} represent the cell motility coefficient and strength of cell-to-cell adhesion for the period 0–24 h, 24–48 h and 0–48 h, respectively. Estimates of posterior distributions for {D(0–24), q(0–24)} and {D(0–48), q(0–48)} have already been obtained from experimental data at 24 h and 48 h, respectively. To obtain estimates for {D(24–48), q(24–48)}, two stages of simulations are required, from 0–24 h and from 24–48 h. In the first stage, model simulations use parameter sets that are drawn from the distributions of {D(0–24), q(0–24)}; whereas, in the second stage, the model simulations update the cell colonies with parameter sets that are drawn from the distributions of {D(24–48), q(24–48)}. We consider two approaches to infer the values of {D(24–48), q(24–48)}.
Approach 1: We jointly infer the values of {D(0–24), q(0–24)} and {D(24–48), q(24−48)} by simultaneously comparing experimental data that are terminated at 24 h and 48 h with the simulated data at the corresponding terminated times. In this approach, we place a uniform prior on both parameter sets {D(0–24), q(0–24)} and {D(24–48), q(24–48)}. We observe that the ABC posterior distributions of {D(0–24), q(0–24)} in this approach are indistinguishable with the estimates previously obtained by using the experiments terminated at 24 h. Approach 2: We make use of the ABC posterior of {D(0–24), q(0–24)} previously obtained from the experiments terminated at 24 h, and only infer the values of {D(24–48), q(24–48)} by matching on the summary statistics at 48 h. To achieve this, for each initial cell density, we fit a bivariate normal distribution to the ABC joint posterior distributions of {D(0–24), q(0–24)}. To perform a model simulation, we draw a parameter set from the bivariate normal distribution for the first stage, and another parameter set from the uniform prior for {D(24–48), q(24–48)} for the second stage.
We use the same uniform prior for {D(24–48), q(24–48)} in the two approaches. The second approach has the advantage that the SMC-ABC algorithm only needs to search over the parameter space of 24–48 h, {D(24–48), q(24–48)}. Thus, we expect the second approach to be faster and more efficient. For each joint posterior distribution of {D(0–24), q(0–24)}, we assess the bivariate normality assumption using a Q-Q plot of chi-square quantiles against the squared Mahalanobis distance [43]. The Q-Q plots suggest that the bivariate normality assumption is reasonable for both initial cell densities.
We found that the ABC posterior distributions of {D(24–48), q(24–48)} in the two approaches are indistinguishable. However, the second approach is more efficient in terms of computational time. Therefore, for all experimental conditions in the two scenarios, we first obtain estimates for periods 0–24 h and 0–48 h then use the second approach to obtain estimates for 24–48 h.
A comparison of D and q for different time periods is shown in Fig 6. Results in Fig 6A–6D correspond to experiments initiated with 20,000 and 30,000 cells, respectively. We observe that the estimated posterior distributions of D(0–24) and D(24–48) are non-overlapping, which implies that estimates of cell diffusivity are significantly different for the two periods of the experiment.
Comparing the posterior estimates of D for different C(0) suggests that values of D for the 30,000 initial cell density experiment is higher than for those in the 20,000 initial cell density experiment during the period 0–24 h. However, the difference is insignificant for the period 24–48 h and for the entire period 0–48 h. These findings indicate that estimates of cell diffusivity depend less on the initial cell density for longer experiments.
In contrast, the posterior estimates of cell-to-cell adhesion strength, q, for different C(0) are substantially different for all three periods. In particular, the estimates of q for the experiments initiated with 30,000 cells are higher than the corresponding values from the experiments initiated with 20,000 cells. This implies that estimates of cell-to-cell adhesiveness depend on initial cell densities. The higher the initial density, the stronger the cell-to-cell adhesion strength. In the literature, several studies have investigated the role of cell-to-cell adhesion in collective cell spreading [44–46] by matching the cell density profiles between the experimental data and the model simulation with several values of q. The previous approach is limited in that it can only give a point estimate of q and provide no insight into the uncertainty in the estimate or the correlation between D and q. Therefore, this study is the first attempt to provide a systematic approach to jointly infer the values of D and q, and compare the distributions of D and q for different experimental conditions.
To analyse the second set of experiments, we consider two approaches: (i) assuming that the values of D and q in the two experimental scenarios are completely unrelated, and thus, inferences of D, q and λ are based solely on the experimental data in Scenario 2, and (ii) assuming that the values of D and q from Scenario 1 are equal to those of D and q in Scenario 2. For the latter approach, we adopt a Bayesian sequential learning approach and use the posterior distribution of D and q from Scenario 1 as the prior for D and q for the corresponding experiments in Scenario 2.
Quantifying the underlying mechanisms that drive the expansion of melanoma cell colonies such as migration, proliferation, and cell-to-cell adhesion is important for developing a systematic approach to assessing the effectiveness of a potential treatment. Typical approaches to parameter estimation often use a deterministic framework [4–7, 23] and only produce point estimates. There is, therefore, a risk that future model projections based on such point estimates could be made with undue confidence.
In this paper, we present a new ABC algorithm to estimate D, q and λ which represent the cell motility, the cell-to-cell adhesion strength and the cell proliferation rate, respectively. To the best of our knowledge, this is the first time that joint inferences have been obtained for all three parameters in a discrete stochastic model describing expanding melanoma cell colonies, using data from a single assay. The new ABC algorithm shows favourable performance relative to state-of-the-art algorithms and together with our derived summary statistics, we can estimate all model parameters precisely across different scenarios, even when a vague prior is used (Tables 1 and 2). This emulates a situation in which virtually no biological knowledge about D, q and λ is assumed. Furthermore, the methodology developed here overcomes the limitation in the previous work [16], which demonstrated that without prior information about D, λ cannot be identified using solely leading edge data.
The methodology proposed here allows us to obtain inferences for D, q and λ in a fully Bayesian framework. The resulting posterior distributions enable us to quantify the associated uncertainty with the parameter estimates which can not be achieved using a deterministic approach. Furthermore, comparing the distributions of D, q and λ (Figs 6, 7 and 9) provides insight into the dependency of the parameter posterior estimates on the experimental elapsed time and on the initial number of cells. Thus, our work adds significant extra information about the parameters relative to the previous analyses [23]. Another advantage of using an ABC approach is the possibility of combining information from the two experiments in a principled way. This approach is shown to be useful in our previous work [16]. Here, it also enables us to gain additional information for D and λ.
We acknowledge that our discrete individual-based model, which is straightforward to implement and computationally cheap, makes an assumption that cell diffusivity is constant. Although the density dependence is less pronounced for experiments terminated at 48 h, it suggests that the underlying assumption of a constant diffusion coefficient D is violated. Thus, it is suggested that the use of a non-linear diffusion coefficient, where D is a function of cell density, D(C), may be more appropriate. In particular, using non-linear diffusion coefficients is shown to provide a better description of the collective behaviour of a cell population in a lattice-free model [47] and a model with complex contact interactions [48]. We expect that implementation of the ABC approach for these models will lead to further research.
It should also be noted that [23] obtained point estimates of D, q and λ separately; D and q from the experiments with cell proliferation suppressed, and λ from experiments with cell proliferation. Thus, this approach may not be applicable if one does not have access to this kind of detailed experimental data sets. Furthermore, results from our analyses also indicate that cell-to-cell adhesion may differ between the two scenarios. In particular, the values of cell-to-cell adhesion is slightly higher for the experiments with cell proliferation occurring, due to the increasing cell population. Thus, we suggest that future studies should consider estimating all parameters simultaneously.
One particular finding from our analysis is that the posterior distributions of D and q consistently depend on the experimental time period, whereas the posterior distribution of λ is approximately time constant. This finding is in agreement with the results of [16] for 3T3 fibroblast cells, however, this feature has not been investigated elsewhere. As demonstrated earlier, this effect is significant and should be included when modelling mechanisms governing the expansion of cell colonies in future research. To achieve this, we suggest that experimental data should be collected at several time points and to optimally do this we leave for future research.
In addition, our ABC algorithm together with the derived summary statistics could also be implemented in a model selection algorithm to distinguish between discrete lattice-based and lattice-free models describing the expansion of cell colonies. In lattice-free models, agents are allowed to migrate and proliferate in a continuous domain, and the direction of movement is a continuous variable [10]. Thus this model is considered to be more realistic than the lattice-based model.
|
10.1371/journal.pcbi.1006232 | Transmission of temporally correlated spike trains through synapses with short-term depression | Short-term synaptic depression, caused by depletion of releasable neurotransmitter, modulates the strength of neuronal connections in a history-dependent manner. Quantifying the statistics of synaptic transmission requires stochastic models that link probabilistic neurotransmitter release with presynaptic spike-train statistics. Common approaches are to model the presynaptic spike train as either regular or a memory-less Poisson process: few analytical results are available that describe depressing synapses when the afferent spike train has more complex, temporally correlated statistics such as bursts. Here we present a series of analytical results—from vesicle release-site occupancy statistics, via neurotransmitter release, to the post-synaptic voltage mean and variance—for depressing synapses driven by correlated presynaptic spike trains. The class of presynaptic drive considered is that fully characterised by the inter-spike-interval distribution and encompasses a broad range of models used for neuronal circuit and network analyses, such as integrate-and-fire models with a complete post-spike reset and receiving sufficiently short-time correlated drive. We further demonstrate that the derived post-synaptic voltage mean and variance allow for a simple and accurate approximation of the firing rate of the post-synaptic neuron, using the exponential integrate-and-fire model as an example. These results extend the level of biological detail included in models of synaptic transmission and will allow for the incorporation of more complex and physiologically relevant firing patterns into future studies of neuronal networks.
| Synapses between neurons transmit signals with strengths that vary with the history of their activity, over scales from milliseconds to decades. Short-term changes in synaptic strength modulate and sculpt ongoing neuronal activity, whereas long-term changes underpin memory formation. Here we focus on changes of strength over timescales of less than a second caused by transitory depletion of the neurotransmitters that carry signals across the synapse. Neurotransmitters are stored in small vesicles that release their contents, with a certain probability, when the presynaptic neuron is active. Once a vesicle has been used it is replenished after a variable delay. There is therefore a complex interaction between the pattern of incoming signals to the synapse and the probablistic release and restock of packaged neurotransmitter. Here we extend existing models to examine how correlated synaptic activity is transmitted through synapses and affects the voltage fluctuations and firing rate of the target neuron. Our results provide a framework that will allow for the inclusion of biophysically realistic synaptic behaviour in studies of neuronal circuits.
| Variability in synaptic function arises from stochasticity in processes ranging in scale from the transitory opening and closing of ion channels to probabilistic neurotransmitter release and vesicle restock [1–4]. The transmission of signals between neurons is therefore inherently stochastic and, moreover, will interact in a history-dependent manner with the patterns in the incoming presynaptic drive [5–9]. A common approach to treating this stochasticity analytically assumes that neuronal firing is uncorrelated, with a Poisson process typically used to model spike times [10–12]. However, non-Poissonian activity is regularly observed in vivo [13–16] and models suggest that, even in the absence of short-term synaptic dynamics, it can have a substantial effect on the propagation of activity [17–21]. It can be expected that the impact of non-Poissonian presynaptic activity will be further complicated when combined with vesicle-depletion depression, in which synaptic transmission becomes weaker and less reliable as stores of available neurotransmitter are depleted and yet to be restocked [22–24].
Transmission through plastic synapses has been shown to decorrelate input spike trains [5, 25, 26], typically increasing the computational power [27, 28] and efficiency [29] of a neuronal network. While average rate effects under the influence of depression have been extensively studied [30–32], a compact set of analytical results for correlated spike trains and stochastic quantal vesicle release from multiple sites has remained elusive (for a full discussion of existing results, see the Discussion). Recently, it has been shown [33] that correlated firing patterns more regular than Poissonian can increase the rate of vesicle release, thereby enhancing the fidelity and efficiency of signal transmission, whilst more irregular spike trains can lead to a decrease in neurotransmitter release [25, 27].
To further analyse how the interaction of correlated presynaptic drive and short-term depression affect quantal synaptic transmission, here we derive a number of analytic results for renewal processes, for which the incoming spike train is fully characterised by the inter-spike-interval (ISI) distribution. This type of presynaptic drive includes that generated by the leaky, quadratic and exponential integrate-and-fire models driven by white-noise [34–36], which are models commonly used to fit experimental data [37–39]. It can be noted that these models will also generate ISIs that are well approximated by renewal processes when the correlations of their incoming synaptic drive are much shorter in time than the typical outgoing ISIs.
We derive equations for the occupancy and the temporal structure of release events when presynaptic cells have spiking patterns fully characterised by their inter-spike-interval distribution. We then show how these can be used to calculate the post-synaptic voltage mean and variance when the presynaptic neurons make multiple independent contacts. These results are illustrated using gamma-distributed ISIs and presynaptic integrate-and-fire neurons. We show that the results allow for a good estimation of post-synaptic firing rates thereby opening the way for the analysis of the role synapses with stochastic short-term depression play in feed-forward and recurrent neuronal networks in which presynaptic spike-trains are typically non-Poissonian.
A quantal model of synaptic dynamics is used where the binary variable x represents the occupancy (x = 1) of a release site by a neurotransmitter-filled vesicle, with x = 0 otherwise. On the arrival of a presynaptic action potential, neurotransmitter is released with probability p if a vesicle is present at the site. For brevity, the probability that a present vesicle is not released, 1 − p, is written as q. Sites that are empty are then restocked at memoryless rate λ (Poisson process). Note that because x is a binary variable taking values of 0 and 1, x2 ≡ x and so Var(x) = 〈x〉(1 − 〈x〉) where the notation 〈X〉 is used as the expectation of any quantity X over all stochastic processes (spike times, release and restock events). Values of parameters used to generate figures, unless otherwise stated, are given in Table 1.
The presynaptic spike train is modelled as a renewal process characterised by the ISI distribution f(t) where the firing rate r is the reciprocal of the mean ISI. Though there are no correlations between successive ISIs, the spike train itself will in general be non-Poissonian and can range between bursting and regular extremes: bursting neurons have positively autocorrelated spike trains at short times whereas more periodically firing neurons generate trains that are negatively autocorrelated over short time frames. Many results in this paper will involve expectations over the ISI distribution of exponential functions
⟨ e - z t ⟩ = ∫ 0 ∞ f ( t ) e - z t d t = L ( z ) (1)
where z can be complex (with restrictions on its real part related to the specific functional form the ISI distribution). This can also be interpreted as the Laplace transform of the ISI distribution, which we will denote as L ( z ). For Laplace transforms of other quantities, say X(t), we will use the notation L X ( t ). Note that the Fourier transform f ^ ( ω ) of the ISI distribution can also be interpreted as an expectation
f ^ ( ω ) = ∫ - ∞ ∞ f ( t ) e - i ω t d t = ⟨ e - i ω t ⟩ = L ( i ω ) (2)
and is directly related to its Laplace transform, remembering that f(t) = 0 for t < 0. This will allow many existing results from the literature on integrate-and-fire ISI distributions in the frequency domain to be used in the subsequent analyses.
Gamma distributed ISIs provide a useful illustration of the results presented here as a single shape parameter α continuously varies the train between bursting for α < 1 and more regular α > 1. The ISI distribution f(t) is given by
f ( t ) = ( α r ) α Γ ( α ) t α - 1 e - α r t (3)
for positive t and zero otherwise, where Γ ( α ) = ∫ 0 ∞ t α - 1 e - t d t is the gamma function. When α = 1 the ISI distribution becomes exponential and the presynaptic spike train is a Poisson process. The expectation of an exponential function (Eq 1) over this class of ISI distribution is
⟨ e - z t ⟩ = ( α r z + α r ) α (4)
where z can be a complex constant as long as the real part of z + αr is greater than zero.
Together with this paper we provide JULIA code in the Jupyter Notebook environment for generating each of the figures shown in the paper. All five scripts are published under the GNU General Public Licence, Version 3 (http://www.gnu.org/copyleft/gpl.html).
After writing down some general statements for arbitrary spike times, we focus on deriving exact results for the case when correlated presynaptic spiking is a renewal process and fully described by the ISI distribution. We present formulae for the steady-state vesicle-release site occupancy averaged over time as well as its mean value just before the arrival of a presynaptic action potential. We then derive an integral equation for the occupancy at some later time given a release event at an earlier time: this will allow us to calculate the autocovariance of release events which in turn leads to an analytical form for the postsynaptic voltage variance. We then generalise this result to a scenario in which each presynaptic cell makes multiple independent contacts. Because of the shared presynaptic drive across these contacts, an additional level of correlation is generated in the input to the post-synaptic cell. We characterise these correlations through the cross-covariance of release events and extend the formula for the post-synaptic voltage to multiple contacts. These results are illustrated using gamma-distributed ISIs for the presynaptic trains. It is then further shown how all results can be derived exactly for presynaptic LIF models or numerically for other classes of integrate-and-fire model, such as the EIF model that are themselves all driven by Gaussian white-noise drive. Finally, we consider a straightforward extension to consider biophysically realistic post-synaptic potentials.
The statistics of a binary occupancy variable x(t), where x = 1 if the site is occupied and is zero otherwise, are first considered. We first write down an obvious steady-state result which links two averages of this quantity: 〈x〉, the occupancy averaged over time, and 〈x〉∞, the mean occupancy just before the arrival of a presynaptic spike. It should be noted that these two quantities are only the same for (memory-less) Poisson processes. Given that the total restock rate must equal the total release rate in the steady state, the following balance equation holds
λ ( 1 - ⟨ x ⟩ ) = p r ⟨ x ⟩ ∞ (5)
where λ is the restock rate given no vesicle is present, p is the probability of release given a vesicle is present and r is the firing rate of the presynaptic neuron. As will be seen later, it is the quantity 〈x〉∞ that is required for analysing the effect on the post-synaptic cell.
To derive 〈x〉∞ it is convenient to first consider a less complete expectation of x, denoted by x ¯, which implies the expectation of x for a fixed pattern of spike times {t1, t2, …tm} but averaged over all possible patterns of restock and release events. Consider the expected occupancy x ¯ m immediately before the mth spike at time tm as a function of the expected occupancy x ¯ m - 1 immediately before the (m − 1)th spike: this obeys the recursion equation
x ¯ m = x ¯ m - 1 q e - λ ( t m - t m - 1 ) + ( 1 - e - λ ( t m - t m - 1 ) ) (6)
where q = 1 − p. This can be solved, with the initial condition x ¯ 1 = 1, to give
x ¯ m = 1 - p q ∑ k = 1 m - 1 q k e - λ ( t m - t m - k ) . (7)
Taking expectations over all realisations of presynaptic spike times, as m → ∞, gives the expected occupancy before the arrival of a presynaptic action potential in the steady-state
⟨ x ⟩ ∞ = 1 - p q ∑ k = 1 ∞ q k ⟨ e - λ T k ⟩ (8)
where Tk is the sum of the last k ISIs. This result is quite general and holds for arbitrarily correlated spike trains; however, we now consider the specific case of spike-times generated by a renewal process. In this case the ISIs will be independent so that the expectation over the exponential term factorises 〈 e - λ T k 〉 = 〈 e - λ t 〉 k and the sum can be evaluated for 〈x〉∞ to give
⟨ x ⟩ ∞ = 1 - ⟨ e - λ t ⟩ 1 - q ⟨ e - λ t ⟩ and ⟨ x ⟩ = 1 - p r ( 1 - ⟨ e - λ t ⟩ ) λ ( 1 - q ⟨ e - λ t ⟩ ) . (9)
The corresponding form for 〈x〉 was found from Eq (5). The expectation 〈e−λt〉 is straightforward to evaluate for many classes of renewal process and is directly related to the ISI Laplace or Fourier transform (Eqs 1 and 2) L ( λ ) = 〈 e - λ t 〉.
We now consider the statistics of the release of neurotransmitter and, in particular, the autocovariance of the release. This quantity will be required to calculate the variance of the postsynaptic voltage. We define the release events at a single site as a series of Dirac-delta events
χ ( t ) = ∑ { t k } δ ( t - t k ) (13)
where here {tk} are the times of the neurotransmitter release events. These occur with probability p only when a presynaptic spike arrives and a vesicle is present, so that the steady-state mean of this quantity is 〈χ〉 = pr〈x〉∞, as already observed in the previous section.
We now consider the steady-state autocovariance that, because of its time-translation invariance, can be written 〈χ(t)χ(0)〉 − 〈χ〉2. For t > 0 we note that 〈χ(t)χ(0)〉 = 〈χ(t|0)〉〈χ〉 where 〈χ(t|0)〉 is understood to be the probability density of a vesicle being released at time t given that one was released previously at time 0. The autocovariance can therefore be written
AutoCov ( χ ) = ⟨ χ ⟩ ( δ ( t ) + ⟨ χ ( t | 0 ) ⟩ - ⟨ χ ⟩ ) (14)
for t ≥ 0 (the function is even in time, thus specifying the t < 0 component) with the Dirac delta function coming from the zero-time contribution.
The quantity 〈χ(t|0)〉 itself is the rate of release, given a release at t = 0 and can be written as pG(t) where G(t) is the probability density of presynaptic spike arriving while a vesicle is present at time t given that initially (at t = 0) a presynaptic spike arrived and immediately after there was no vesicle present. We also introduce a related quantity H(t) which has the same conditionality but is the density of a presynaptic spike arriving while there is no vesicle present at a time t. Note that F(t) = G(t) + H(t) where F(t) is the probability density of an action potential arriving at t given there was one at t = 0 (the conditionality is the same because the arrival of a spike at t does not depend on whether a release site was stocked or not just before an earlier spike). This last quantity can be directly related to the ISI distribution f(t) via a convolution
F ( t ) = f ( t ) + ∫ 0 t d s F ( s ) f ( t - s ) (15)
that can be solved in terms of integral transforms. It is straightforward to derive similar formulae for G(t) and H(t) via the introduction of the quantities
g ( t ) = f ( t ) ( 1 - e - λ t ) and h ( t ) = f ( t ) e - λ t (16)
where g(t) and h(t) are the probability densities that the release site (initially empty at t = 0 following a presynaptic spike) are either restocked or not, respectively, by the time the next spike arrives; note also that f(t) = g(t) + h(t). To derive a self-consistent integral equation for G(t) we need to decompose it into the various histories that start with a vesicle absent at t = 0 following a presynaptic spike and end with a vesicle present at t when a spike arrives. There are four distinct contributions that need to be accounted for. The first is straightforward as it arises from the first spike arriving at t giving a contribution of g(t). The other contributions imply that the penultimate spike arrives at an intermediate time s that needs to be integrated over (as in Eq 15). The second contribution has a vesicle present before the intermediate time s and no release and so contributes G(s)qf(t − s). The third contribution has a vesicle present before the intermediate time s and there is a release and so contributes G(s)pg(t − s). The final contribution has no vesicle present before the intermediate time s and then there is a restock, contributing H(s)g(t − s). Combining these four contributions and simplifying, using H = F − G and h = f − g, results in the following integral equation
G ( t ) = g ( t ) + ∫ 0 t d s F ( s ) g ( t - s ) + q ∫ 0 t d s G ( s ) h ( t - s ) . (17)
The equations for F(t) and G(t) can either be solved numerically using iterative procedures or, alternatively, solved using integral transforms (see the next section). Noting that in the limit t → ∞ the conditional quantity G(t) converges to r〈x〉∞ allows the autocovariance
AutoCov ( χ ( t ) ) = p r ⟨ x ⟩ ∞ ( δ ( t ) + p ( G ( | t | ) - r ⟨ x ⟩ ∞ ) ) (18)
of the neurotransmitter release time series χ(t) to be written in terms of G(t) and the occupancy 〈x〉∞.
In the previous section the pre-spike occupancy of a release site was calculated and the release-train autocovariance derived. We now use these results to derive the voltage mean and variance of the postsynaptic neuron. The subthreshold mean and variance of the postsynaptic neuron are the key quantities necessary to estimate the output firing rate of a neuron [18, 36, 41–45] and hence its resultant effect on the network.
It is assumed that the presynaptic neurons are uncorrelated and each fires at a rate r with the same ISI distribution. For simplicity we assume that each neurotransmitter-release event causes the postsynaptic membrane to increase by a fixed voltage a (this restriction is for simplicity and can be relaxed). In this section we consider presynaptic cells that make contacts with only one vesicle release site, leaving multiple release sites to a later section. The postsynaptic voltage therefore follows the equation
τ d v d t = μ - v + a τ ∑ i = 1 N χ i ( t ) (23)
where τ is the membrane time constant, μ the resting potential in absence of synaptic drive and there are N presynaptic neurons, with the ith having a release train χi(t).
The steady-state mean post-synaptic voltage is found using the result 〈χ〉 = pr〈x〉∞ so that
⟨ v ⟩ = μ + a τ N p r ⟨ x ⟩ ∞ . (24)
This is an increasing function of the prespike occupancy 〈x〉∞ and therefore also increases with the regularity of the presynaptic spike train (see Fig 1D for the dependence of 〈x〉∞ on the burstiness of the presynaptic spike train).
To find the post-synaptic voltage variance we first solve differential Eq (23) formally as an integral over the release trains
v ( t ) = ⟨ v ⟩ + a ∫ - ∞ t d t ′ e - ( t - t ′ ) / τ ∑ i ( χ i ( t ′ ) - ⟨ χ ⟩ ) (25)
where we included the component of the mean stemming from the synaptic input in the summation. The voltage variance can therefore be written as a double integral over the autocovariance of χ(t) as follows
Var ( v ) = N a 2 ⟨ χ ⟩ ∫ 0 ∞d s ∫ 0 ∞d s ′ e - s / τ e - s ′ / τ ( δ ( s - s ′ ) + p ( G ( | s - s ′ | ) - r ⟨ x ⟩ ∞ ) ) (26)
where the fact that the χi(t) from different presynaptic neurons are uncorrelated has been used. The Dirac-delta component is straightforward to evaluate and the other components are symmetric around s − s′ so that
Var ( v ) = τ N a 2 2 ⟨ χ ⟩ + 2 N a 2 ⟨ χ ⟩ p ∫ 0 ∞d s ′ e - s ′ / τ ∫ s ′ ∞d s e - s / τ ( G ( | s -s ′ | ) - r ⟨ x ⟩ ∞ ) (27)
which, on a change of the inner integration variable z = s − s′, allows for the outer integral over s′ to be performed resulting in
Var ( v ) = τ N a 2 2 ⟨ χ ⟩ ( 1 + 2 p ∫ 0 ∞ d z e - z / τ ( G ( z ) - r ⟨ x ⟩ ∞ ) ) . (28)
Part of the integral in the above equation is simply the Laplace transform of G(t), which was already provided in the second of equation pair (20). On substitution, this gives a compact form for the postsynaptic voltage variance in terms of L ( z ) = 〈 e - z t 〉 the Laplace transform of the ISI distribution
Var ( v ) = τ N a 2 2 ⟨ χ ⟩ ( 1 + 2 p ( L ( 1 / τ ) - L ( 1 / τ + λ ) ( 1 - L ( 1 / τ ) ) ( 1 - q L ( 1 / τ + λ ) ) - τ r ⟨ x ⟩ ∞ ) ) . (29)
This is a central result of this paper and is general for neurons that receive synapses with single release sites from presynaptic neurons that fire as a renewal process. We now go on to examine the postsynaptic voltage behaviour for gamma-distributed presynaptic ISIs, and consider the case for presynaptic integrate-and-fire models in a later subsection.
A single presynaptic neuron will typically make contacts with a postsynaptic cell that result in multiple vesicle release sites, with estimates of this parameter varying from 1 to as much as 100 [46] for neocortical layer-5 pyramidal cells. Here we consider a case where each of the N presynaptic cells makes connections with n indepedent release sites, a scenario illustrated in Fig 3A for a case N = 1 and n = 3. We assume that all processes (such as restock and release) are statistically independent at each of these sites, but those sharing the same presynaptic neuron receive the same spike train. The postsynaptic-voltage dynamics now take the form
τ d v d t = μ - v + a τ ∑ i = 1 N ∑ j = 1 n χ i j ( t ) (31)
where here χij is the release time course of the jth contact on the ith neuron. Only the χij that share the same presynaptic neuron will be correlated. The steady-state mean voltage is straightforward to derive using the result 〈χ〉 = pr〈x〉∞ so that
⟨ v ⟩ = μ + a τ N n p r ⟨ x ⟩ ∞ . (32)
However, to calculate the variance correlations between release sites need to be accounted for, which requires the solution of an additional integral equation.
In previous sections analytical forms for the pre-spike occupation 〈x〉∞, autocovariance of the release-train χ and the resulting voltage moments were derived with results illustrated using gamma-distributed ISIs (Eq 3). Integrate-and-fire models, such as the Leaky, Quadratic or Exponential IF models driven by white noise that, following a spike, retain no memory of their previous state will generate spike trains with uncorrelated ISIs. In this case the ISI distribution is identical to the first-passage-time density of the steady-state dynamics. Because of this, all the general results derived thus far are applicable when the presynaptic population is comprised of integrate-and-fire neurons. For the LIF the Fourier transform of the first-passage-time density is available analytically, whereas for non-linear IF models like the QIF or EIF it can be straightforwardly obtained numerically. We now consider examples of these two cases.
It has been shown [44] that the postsynaptic firing rate of EIF neurons is surprisingly insensitive to (positive) temporal correlations when driven by coloured Gaussian noise for a given voltage mean and variance. This allows for a matched-variance approximation to be used in which the firing rate of a coloured-noise driven EIF neuron is approximated by its white noise equivalent but using the same subthreshold mean and variance. Eqs (32) and (42) can be used to find the voltage mean and variance and the EIF rate for white-noise drive is given in [36]. To test whether this approximation has validity, we consider the firing rate of an EIF neuron driven by depressing, stochastic synapses from a presynaptic population of EIF neurons. Following the same approach used in Fig 4 for LIF neurons, we covaried vre and σ to get a range of spiking statistics in the presynaptic EIF population (see Eq 49 for the EIF model definition). In Fig 5A the firing rate of an EIF neuron as a function of the constant drive μ is shown for three different combinations of vT − vre and σ, with example time courses given in Fig 5B each having a rate 10Hz. In Fig 5C–5F the occupancy, voltage statistics and post-synaptic rate are plotted, as vT − vre and σ are simultaneosly linearly varied from their values on the blue and red curves in Fig 5A (with μ adapted so that the presynaptic rate is always 10Hz). A range of connectivity is considered by four combinations of N and n (as marked). As can be seen, the simulations agree well with the post-synaptic firing rate over a range of presynaptic firing patterns and forward connectivity choices. Hence, the matched variance approximation provides a fair account of the firing rate even in cases where the incoming fluctuations are negatively correlated, extending previous results [44].
It is worth noting it is fairly straightforward to generalise the voltage-variance calculations Eqs (29 and 42) to neurons receiving more biophysically shaped excitatory post-synaptic potentials (EPSPs). For example, if an isolated release event generates an EPSP of the form E ( t ) then, under the assumption of additivity,
v = ⟨ v ⟩ + ∫ - ∞ t d t ′ E ( t - t ′ ) ∑ i ( χ i ( t ′ ) - ⟨ χ ⟩ ) (50)
becomes the generalisation of Eq (25) where the voltage mean is
⟨ v ⟩ = μ + N ⟨ χ ⟩ ∫ 0 ∞d t E ( t ) . (51)
Following the same approach that led to Eq (29), the voltage variance in this case can be written
Var ( v ) = N ⟨ χ ⟩ ( ∫ 0 ∞ d t E ( t ) 2 ) ( 1 + 2 p ∫ 0 ∞ d z ( G ( z ) - r ⟨ x ⟩ ∞ ) ∫ 0 ∞ d t E ( t ) E ( t + z ) ∫ 0 ∞ d t E ( t ) 2 ) . (52)
If the form of the EPSP is modelled as a sum of multiple exponentials then the variance can again be expressed by Laplace transforms of G(t). For example, a two-exponential model for an EPSP
E ( t ) = a τ 1 + τ 2 τ 1 - τ 2 ( e - t / τ 1 - e - t / τ 2 ) (53)
has the integrals
∫ 0 ∞d t E ( t ) 2 = a 2 2 ( τ 1 + τ 2 ) and ∫ 0 ∞d t E ( t ) E ( t + z ) ∫ 0 ∞d t E ( t ) 2 = τ 1 e - z / τ 1 - τ 2 e - z / τ 2 τ 1 - τ 2 . (54)
On substitution into the equation for the variance, and after identifying any Laplace transforms of G(t), this gives
Var ( v ) = ( τ 1 + τ 2 ) N a 2 2 ⟨ χ ⟩ ( 1 + 2 p ( τ 1 L G ( 1 / τ 1 ) τ 1 - τ 2 + τ 2 L G ( 1 / τ 2 ) τ 2 - τ 1 - r ⟨ x ⟩ ∞ ( τ 1 + τ 2 ) ) ) . (55)
The forms L G ( z ) are provided in the second equation of the pair (20) in terms of the Laplace transform of the first-passage-time density.
We have presented a series of novel analytical results that extend the level of biophysical detail incorporated in models of synaptic transmission. Non-Poissonian activity is commonly seen in vivo and can have a substantial effect on neuronal activity; relating this to synaptic dynamics allows for a more comprehensive understanding of the typical behaviour of plastic synapses. We have derived the exact spike-triggered mean vesicle occupancy, noting that it takes a particularly compact form when the input spike train is a renewal process, and highlighted the relationship between the spike-triggered mean and the overall level of vesicle occupancy, confirming the numerical results of Matveev and Wang (2000, [25]), de la Rocha and Parga (2005, [8]) and Reich and Rosenbaum (2013, [33]). We have derived the autocorrelations in vesicle release in terms of integral transforms, confirming the numerical results of Goldman et al (2002, [26]) and extending the analytical results of Goldman (2004, [29]) to account for the biophysically important case of probabalistic vesicle release.
The exact subthreshold voltage variance calculated from the neurotransmitter release autocorrelations is a potentially useful result and incorporates many biophysical details of autocorrelated input spikes, quantal effects, stochastic and cross-correlated vesicle release, and short-term plasticity. The relative effects of these biophysically relevant phenomena can now be quantitatively analysed [27, 50], providing insights into how changes in release-site number n arising from long-term plasticity [51] or in release probability p due to developmental changes [52–54] reshape the postsynaptic response to correlated spike trains.
The results for synaptic transmission when the presynaptic integrate-and-fire neuron is driven by short-time correlated noise processes are another important application, expanding the utility of the model to allow for study of synaptic dynamics alongside other known results concerning the firing rate [41, 42] and correlation structure [18, 43] of such neurons.
Many previous studies have examined how plastic, probabilistic and quantal synapses affect the statistics of patterned transmission through the synapse, and we now provide a selective overview. Vere-Jones (1966) [5] examined a model of a quantal, probabilistic synapse, finding that the release of neurotransmitter is more Poissonian than the afferent activity. Maass and Zador (1999) [27] considered the binary output of a single vesicle release site, investigating how triplets of incoming spikes corresponded to different release patterns under varying synaptic parameters. In particular they showed that dynamic synapses transmit bursts of spikes more reliably than static synapses and have enhanced computational power. Matveev and Wang (2000, [25]) numerically studied the effects of naturalistic presynaptic firing patterns on vesicle release, both with bursts generated by a two-state Markov model and long-time correlated trains. They found that spikes within a burst are suppressed by synaptic depression compared to isolated spikes, and, like Vere-Jones, that neurotransmitter release is more Poissonian than the incoming spikes. Goldman et al (2002, [26]) examined the transmission of doubly-stochastic Poissonian spike trains, constructed to reflect experimentally recorded neuronal bursting, finding again that dynamic synapses decorrelate afferent spike trains and so reduce coding redundancy across a broad range of synaptic parameters. As part of this study, the autocovariance of neurotransmitter release in response to temporally structured spike trains was calculated numerically. Goldman (2004, [29]) derived the information transmission efficiency of a depressing synapse analytically, finding the autocovariance of neurotransmitter release under the assumption of a synapse that reliably releases neurotransmitter whenever a vesicle is present. Using a model comprising a single neurotransmitter release site, de la Rocha et al (2002, [55]) showed that dynamic synapses were more effective transmitters of afferent signals only when the input is non-Poisson, analytically describing the distribution of synaptic release events when the input is a renewal process. They later [8] numerically studied the impact of temporal correlations on synapses containing multiple release sites, showing that bursty stimuli elicited fewer releases of neurotransmitter but that there could be a non-monotonic relationship between presynaptic and postsynaptic firing rates in the presence of input correlations and synaptic dynamics. Fuhrmann et al (2002, [56]) developed the stochastic quantal model used in this paper, capturing the same processes as the continuous phenomenological Tsodyks-Markram model of short-term plasticity [30], but focussed the initial analysis on Poisson spike trains. Ly and Tranchina (2009, [57]) considered numerically the transmission of temporally correlated spike trains across stochastic, but not dynamic, synapses and plotted the autocovariances in vesicle release, as well as the postsynaptic firing rate for renewal process inputs. Rosenbaum et al (2012, [12]) studied information transmission for Poissonian inputs, finding that the incorporation of stochastic quantal effects differentially affected information transmitted at different presynaptic rates. This paper approximated the auto- and cross covariances in neurotransmitter release in response to Poisson drive and these results were shown to be exact by Bird and Richardson (2014, [9]). Reich and Rosenbaum (2013, [33]) studied models of presynaptic spiking both more and less regular than a Poisson process, showing numerically that more regular firing patterns can increase the rate of vesicle release, thereby enhancing the fidelity and efficiency of signal transmission, whilst more irregular spike trains can lead to a decrease in neurotransmitter release. Zhang and Peskin (2015, [50]) developed the results of [12] on information transfer with unreliable dynamic synapses using a slightly simpler model of vesicle recovery, analytically studying the effects of a more general model of presynaptic spiking on neurotransmitter release rates and numerically simulating the effects on the postsynaptic membrane.
The matched-variance firing rate approximation in Fig 5 does not constitute a complete framework for treating recurrent networks of neurons with stochastic depressing synapses, because only the rate and not the full ISI statistics of the post-synaptic neuron were derived. However, it does suggest that some approximation scheme that goes beyond the first-order statistics of the ISI distribution might be used to analyse recurrent networks. This is currently a problem of great interest [21, 43, 58] given the strong effects correlated spike trains are acknowledged to have even across static synapses [18]. To include the components of vesicle-release autocorrelation arising from short-term depression, as modelled here, would increase physiological relevance and bring studies of output firing patterns into line with much of the literature on neuronal networks [31, 32, 59].
Another interesting extension is to account for the effect of spike-frequency adaptation currents on presynaptic firing. Adaptation is present across the nervous system [60] and can modulate responses to persistent activity by high-pass filtering and response selectivity [61–63]. These functional roles overlap with those attributed to synaptic depression [64], and there have been a number of recent studies on the interactions of short-term synaptic plasticity with slow adaptation mechanisms [65, 66]. A key feature of adaptation currents is the creation of correlations between interspike intervals [67], generating non-renewal spike trains. These correlations have recently been shown to take the form of a geometric series [68]. Eq (8) presents a way of deriving approximate results for the synaptic transmission for weakly correlated ISIs and suggests a promising avenue of research to go beyond renewal processes and study the effect of these two key short-term adaptive processes in neural circuits.
|
10.1371/journal.pntd.0003833 | Risk Factors Associated with Ebola and Marburg Viruses Seroprevalence in Blood Donors in the Republic of Congo | Ebola and Marburg viruses (family Filoviridae, genera Ebolavirus and Marburgvirus) cause haemorrhagic fevers in humans, often associated with high mortality rates. The presence of antibodies to Ebola virus (EBOV) and Marburg virus (MARV) has been reported in some African countries in individuals without a history of haemorrhagic fever. In this study, we present a MARV and EBOV seroprevalence study conducted amongst blood donors in the Republic of Congo and the analysis of risk factors for contact with EBOV.
In 2011, we conducted a MARV and EBOV seroprevalence study amongst 809 blood donors recruited in rural (75; 9.3%) and urban (734; 90.7%) areas of the Republic of Congo. Serum titres of IgG antibodies to MARV and EBOV were assessed by indirect double-immunofluorescence microscopy. MARV seroprevalence was 0.5% (4 in 809) without any identified risk factors. Prevalence of IgG to EBOV was 2.5%, peaking at 4% in rural areas and in Pointe Noire. Independent risk factors identified by multivariate analysis were contact with bats and exposure to birds.
This MARV and EBOV serological survey performed in the Republic of Congo identifies a probable role for environmental determinants of exposure to EBOV. It highlights the requirement for extending our understanding of the ecological and epidemiological risk of bats (previously identified as a potential ecological reservoir) and birds as vectors of EBOV to humans, and characterising the protection potentially afforded by EBOV-specific antibodies as detected in blood donors.
| Ebola and Marburg viruses cause haemorrhagic fevers often fatal to humans. Here, we looked for antibodies to Ebola and Marburg viruses (i.e., markers of previous contact with these viruses) in Congolese blood donors with no previous history of haemorrhagic fever. We found serologic evidence for contact with Marburg and Ebola viruses in 0.5% and 2.5% of blood donors, respectively. The circulation of Marburg virus occurs at a very low rate without any identified risk factor. In contrast, prevalence to Ebola virus was peaking at 4% in rural areas and in Pointe Noire city. Importantly, we identified that contacts with bats and birds constituted two independent environmental determinants of exposure. This study confirms that contact with Ebola virus is not infrequent in Congo and can occur in the absence of haemorrhagic fever. It highlights the requirement for further investigating the role of bats and birds in the ecological cycle of Ebola, and for determining whether asymptomatic contact with Ebola virus can provide subsequent protection against severe forms of the Ebola disease.
| Marburg and Ebola viruses (family Filoviridae, genera Marburgvirus and Ebolavirus) cause severe Viral Haemorrhagic Fever (VHF) in humans, with a high fatality rate in symptomatic cases [1,2,3]. They appear to infect and persist in some species of fruit bats, that may serve as natural reservoirs for these viruses [4,5,6,7,8,9]. Non-human primates have been a source of human infections however they are not thought to be the reservoir as they develop severe, fatal illness when infected [10].
The genus Marburgvirus currently consists of a single species Marburg marburgvirus, of which the recognised members are Marburg virus (MARV) and Ravn virus [11,12]. The first cases of Marburg haemorrhagic fever (MHF) occurred in Germany and Serbia (in the former Yugoslavia) in 1967 and were linked to laboratory work using tissues dissected from African green monkeys imported from Uganda [13,14]. The first major outbreak of MHF occurred in Democratic Republic of the Congo (DRC, formerly Zaire), from 1998 to 2000 [15,16]. A second, even more devastating outbreak occurred in Angola in 2004–2005 with a reported case fatality rate (CFR) of almost 90% [17,18].
In Nigeria and DRC, seroprevalence studies identified antibodies to MARV in less than 2% of apparently healthy people selected in general population in Nigeria and amongst healthcare workers and general population in DRC [19,20]. In the Central African Republic (CAR), antibodies to MARV were observed in both Pygmy (0.7–5.6%) and non-Pygmy (0.0–3.9%) populations [21]. An African serosurvey of VHF (Crimean-Congo haemorrhagic fever, Rift Valley fever, Lassa, Hantaan, EBOV and MARV), conducted in the 1980s in the Central African general population, reported low prevalence values: 0.3% in N’Djamena (Tchad), 2.6% in Bioco Island (Equatorial Guinea) and, in the Republic of Congo, 3% in Pointe-Noire but no seropositive sera to MARV detected in people in Brazzaville [22]. To date, no case of MHF has been reported in the Republic of Congo.
The genus Ebolavirus includes five species: Zaire ebolavirus (Ebola virus: EBOV), Sudan ebolavirus, Taï Forest ebolavirus, Reston ebolavirus and Bundibugyo ebolavirus [11,12]. The genus Ebolavirus is primarily African in origin, with the exception of the species Reston ebolavirus which is Asian [23]. EBOV was first identified in 1976, in Southern Sudan [24] and in the North of DRC [25,26]. Since then, outbreaks have been described in several other African countries (the Republic of Congo, Ivory Coast, DRC, Gabon, Sudan, Uganda, Guinea, Sierra Leone and Liberia) [1,27,28,29,30,31,32,33,34], with reported (CFR) frequently exceeding 50% amongst symptomatic patients. In the Republic of Congo where the current study took place, several outbreaks of (Zaire) EBOV were reported in the North of the country (2001 in Olloba-Mbomo, 2002 in Kéllé, 2003 in Mbandza-Mbomo), with 75 to 89% reported fatality rates [35,36,37].
In previous seroprevalence studies, amongst 1,517 apparently healthy persons tested in five regions of the Cameroon, a positive rate of 9.7% was found with highest rates amongst Pygmies (14.5%), young adults (11.6%) and rain forest farmers (13%) [38]. In CAR, the seropositivity rate was 5.3% and Pygmies appeared to have a higher seroprevalence than non-Pygmies (7% versus 4.2%) [21]. During the 1995 outbreak of Ebola virus disease in the region of Kikwit (Democratic Republic of Congo), villagers had a greater chance of exposure (9.3%) than forest and city workers (2.2%) [39]. In a large study conducted in 220 villages in Gabon (4,349 individuals enrolled), antibodies against EBOV were detected in 15.3% of those tested, with the highest levels in forested regions (17.6% and 19.4% respectively in forest and deep forest areas), suggesting the occurrence of mild or asymptomatic infections [40,41]. In the Republic of Congo, seroprevalence values reported in the late 1980's were 7.8% in Pointe-Noire and 6.2% in Brazzaville [22].
In Sierra Leone, in 2006–2008, among 253 febrile patients negative for Lassa fever and malaria, antibodies against EBOV and MARV were detected in respectively 8.2% et 3.2% of the samples [42].
In this study, we present an analysis of MARV and EBOV seroprevalence amongst blood donors in the Republic of Congo in 2011 and we report associated risk factors for contact with EBOV.
A MARV and EBOV seroprevalence study was performed in 2011 in the Republic of Congo, using a prospective cohort of blood donors.
Field samples for the study were collected from March to July 2011, in the Republic of Congo (Fig 1) in urban areas (Brazzaville and Pointe-Noire) and in rural locations (Gamboma, Owando, Oyo and Ewo). Ewo is the capital of the Department of Cuvette-Ouest, where all previous EBOV outbreaks occurred.
This study was performed in collaboration with the Centre National de Transfusion Sanguine (CNTS) of Congo; the Virology Laboratory UMR_D 190 "Emergence des Pathologies Virales" (Aix-Marseille University, IRD French Institute of Research for Development, EHESP French School of Public Health), Marseille, France and the Virology Laboratory of Bernhard-Nocht-Institut für Tropenmedizin, Hamburg, Germany.
Blood donors of both genders were included. The criteria for enrollment were eligibility for blood donation and provision of informed consent without specific limiting factors. The age of blood donors ranged from 18 to 65 years.
Serum samples for serological analyses were collected in collaboration with the CNTS. Informed, written consent was obtained from each person enrolled in the study and the consent procedure was approved by the Congolese Research in Health Sciences Ethics Committee (N° 00000065 DGRST/CERSSA).
A structured questionnaire was administered face-to-face, in the official language (French) and/or in national languages (Lingala or Kutumba). All questionnaires were completed by the medical personnel conducting the interviews.
The following data were collected: socio-demographic circumstances, domestic characteristics (age, gender, occupation, residence, size of household, type of house, water resource, usage of mosquito nets), environmental characteristics (animal contacts and/or consumption), travel outside the country during their lifetime, history of haemorrhagic fever (in family or personal).
Venous blood samples were drawn using two 4 mL plain tubes which were immediately centrifuged. Sera were kept at -80°C until use. Aliquots were inactivated at 56°C for 30 min and transferred to the Virology Laboratory of the Bernhard-Nocht-Institut (BNI) in Hamburg for serological assays.
Serum IgG antibodies specific for EBOV and MARV were titrated using indirect double-immunofluorescence microscopy assays which were recorded as positive if reciprocal end point titres were ≥20 [43].
Antigens consisted of acetone-fixed Vero cells infected with Ebola virus (strain ATCC 1978) or Marburg virus (strain Popp 1967). Cultivation of the viruses was carried out in an approved and compliant BLS4 laboratory in BNI.
Serum samples were tested as serial twofold dilutions from 1:20 to ≥ 1:160 and antibodies were detected with a Fluorescein isothiocyanate (FITC) labelled anti-human IgG antibody-conjugate. Cell smears were counterstained with specific anti-Ebola or anti-Marburg nucleocapsid monoclonal antibodies (provided by the Institute of Virology, University of Marburg) using a rhodamine-anti-mouse conjugate as secondary antibody [44]. As positive controls for Ebola virus, we used (i) a human polyclonal antibody for the first IF with a titre at 2,560 (IgG) and (ii) a mouse monoclonal antibody with a titre at 1,280 (IgG). As positive controls for Marburg virus, we used (i) a human polyclonal antibody for the first IF with a titre at 1,280 (IgG) and (ii) a mouse monoclonal antibody with a titre at 640 (IgG).
This “double immunofluorescence” protocol provides a much higher specificity than regular immunofluorescence assays, since only antibodies that detect filoviral antigens in co-localisation with a monoclonal antibody are considered.
Statistical analyses were performed using the IBM SPSS statistic 21 software. Analyses aiming at analysing risk factors for seropositivity included univariate, stratified and multivariate analyses. The Fisher’s exact test was used to compare proportions in univariate analysis and the ANOVA test to compare means. The Pearson’s test was used for stratified analysis.
All statistical analyses were performed at the 95% confidence level. The association between anti-EBOV IgG seropositivity and risk factors was determined by binary logistic regression analysis. Stratified analysis based on sex, age and area were performed. The significant variables in univariate analysis were entered in the multivariate model. The quality of the multivariate model was assessed with Hosmer-Lemeshow’s test.
Sociodemographic characteristics are presented in Table 1. Overall, 809 blood donors provided serum samples; 734 (90.7%) lived in urban areas (62.9% and 27.8% in Brazzaville and Pointe-Noire, respectively), 370 (45.7%) of which were younger than 30 years old (yo) with a median age of 31. The gender ratio (Male/Female) was 3.02. Most blood donors lived in a modern house 723 (89.4%). The two most common occupations were “student” (20.9%) and “unemployed” (26.5%).
Epidemiological analyses of EBOV and MARV infections were performed separately. However, the same potential risk factors for seropositivity (gender, age, household size, occupation, mosquito net, travel risk, exposure to rodents/forests animals, and consumption of forests animals) were assessed for both viruses.
Individual titres of the MARV and EBOV-positives are presented in Table 2.
IgG to MARV was identified in 0.5% of the donors tested (4 in 809). Seropositivity could not be significantly associated with any of the risk factors investigated in the individual questionnaire. Antibodies to MARV were detected exclusively in male blood donors from Brazzaville who had all been in contact (touching and catching) with rodents (mice and rats), but this is a common feature in out cohort (>80% of the donors reported such contact). Three out of four were students (median age 23) and one (55yo) was an office worker. None had been in contact with (or had eaten) forest animals or bats. The titre of IgG in positive donor ranged from 80 to 160.
The overall prevalence of positivity for IgG when tested against EBOV was 2.5% (20 in 809). It was 1.6% (8 in 509) in Brazzaville, 4% (3 in 75) in the rural locations (Gamboma, Owando, Oyo and Ewo) and 4% (9 in 225) in Pointe-Noire.
Table 3 summarises the association between potential risk factors and anti-EBOV IgG seropositivity using univariate analysis. Amongst the populations studied there was no statistically significant relationship between gender (p = 0.79), age (p = 0.96), travel (p = 0.62), household size (p = 0.39), exposure to rodents (p = 0.63) and the presence of IgG to EBOV. However, being a hunter (p = 0.01) was a risk factor, whereas the other occupations showed no statistical significance in univariate analysis. The use of simple mosquito nets had a protective effect. Importantly, significant association with bat contact or eating birds (p values respectively of <0.001 and 0.01) was identified.
Seropositivity in age-groups was as follows: 2.7% in 18–29yo, 2.1% in 30–39yo, 2.3% in 40–49yo, 3.2% in 50–59yo and 0.0% in >60yo, with no evidence for an increase of seroprevalence with age. In the 18–29yo age-group, higher seroprevalence was associated with touching bats (p<0.001). In the second group (30–39yo), higher seroprevalence was associated with the military profession (p = 0.04). No significant association was found in the other age-groups. No significant association was identified in a stratified analysis by occupation, including military profession.
Regarding stratification by areas: in Pointe-Noire, to be a student (p = 0.002) or to have exposure to bats (p<0.001) was statistically associated with anti-EBOV IgG. Exposure to bats (p<0.001) was also found to be a risk factor in Brazzaville. Other variables (sex, age, household size, type of house, travel risk, exposure to rodents, and consumption of forest animals) were not significantly associated with the presence of IgG, regardless of the area in which donors lived.
Concerning stratification by gender (S1 Table), seropositivity was 2.6% in males and 2.0% in females. No significant association was identified in females. Amongst the male subpopulation, being a hunter (p = 0.009), having contact with bats (p<0.001) or monkeys (p = 0.01), or consuming birds (p = 0.002) was statistically associated with EBOV IgG positivity. Other variables (age, household size, type of house, travel risk, exposure to rodents) had no statistically significant relationship with the presence of IgG to EBOV.
In the multivariate model (Table 3), the only variables independently associated with Ebola antibody detection were contact with bats (p<0.001) and bird consumption (p = 0.04).
Whilst the Democratic Republic of the Congo and neighbouring Angola have experienced Marburg disease outbreaks [2,45], no Marburg cases have been reported to date in the Republic of Congo. By contrast, human epidemics due to Ebola virus have been reported in the Republic of Congo in 2001, 2002 and 2003. All outbreaks were located in the Cuvette-Ouest Department [1,29,46].
Against this background, we conducted a seroprevalence study of Ebola and Marburg viruses amongst blood donors in the Republic of Congo, to estimate the seroprevalence of both viruses outside the epidemic period and also to identify possible risk factors. This study has some obvious limitations related to the population studied (e.g., the number of blood donors originating from rural areas was small, all participants were older than 18 years old and 75% were males) and to the capacity to collect epidemiological information without interfering with the process of blood donation (for example, travel inside the country and activities outside the main occupation were not documented). However a body of standardised information was obtained from the questionnaires.
Importantly, our biological analyses identified antibodies specific for Ebola virus (and to a lesser extent to Marburg virus) in healthy blood donors who did not report a history of haemorrhagic fever. This confirms previous studies suggesting that filoviruses can circulate in Africa in the absence of severe clinical presentations (i.e., associated with asymptomatic or mild infections) [47] and that Ebola virus-specific seroconversion can be observed without clinical manifestation [21] or possibly associated with atypical clinical presentation. In the case of Ebola virus infections, asymptomatic seropositives have been identified by different techniques (e.g., immunofluorescence [22] or ELISA [21]) and have been found more frequently in areas where Ebola cases were reported [40]. In addition, Leroy and collaborators detected the Ebola virus genome in white blood cells of asymptomatic seroconverters investigated for up to 3 weeks following exposure to documented Ebola symptomatic patients [41]. Contact with limited amounts of virus, specific infection routes [48], specific characteristics of individual immunity leading to production of specific IgG and early and strong inflammatory responses [47], may explain the presence or absence of symptomatic presentation. Finally, a recent study identified asymptomatic seropositives with antibodies against various linear epitopes located in different and both structural and non-structural EBOV proteins [49]. Altogether, this rules out the hypothesis that antibodies to EBOV detected in asymptomatic individuals in previous studies are massively false positives.
It has been proposed that some populations may have been exposed to yet unidentified virus strains that have relatively low pathogenicity for humans. Monath [50] proposed that pathogenic strains may have independent transmission cycles involving species rarely in contact with humans. In support of this hypothesis, Gonzalez [22] reported atypical serological responses in asymptomatic seropositives, consisting in identical titres against Ebola Mayinga from Zaire and Ebola Boniface from Sudan identified in 89% of cases. This hypothesis still stands but is weakened by (i) the fact that fifteen years of investigations of African wildlife species have failed to identify such low pathogenicity strains in host species in close contact with humans and (ii) the identification of asymptomatic seroconversions following contact with symptomatic patients [41].
Monath [50] further proposed that pathogenic strains may emerge from non-pathogenic strains by mutational events, but this hypothesis is weakened by the argument that the same strain of EBOV has been shown to be responsible for symptomatic cases and secondary asymptomatic cases [41].
In the late 1980's, Gonzalez and collaborators identified low MARV seroprevalence values in Pointe Noire and Brazzaville (3 and 0%, respectively) [22]. In our study, the prevalence of IgG to MARV was also low (0.5%) and could not be associated with specific risk factors. In contrast, the prevalence of antibodies to EBOV was 2.5%, peaking at 4% in rural populations and in the city of Pointe Noire. It was 1.6% in Brazzaville. Gonzalez and collaborators had previously reported 7.8 and 6.2% Ebola seroprevalence values in Pointe Noire and Brazzaville, respectively [22]. The differences observed may be explained by different technical protocols used for detection of antibodies, by the fact that studies were performed more than twenty years apart, and by the investigation of different populations since Gonzalez studied randomly selected clusters (providing representativeness of gender and age distribution). Against the latter hypothesis, it should be noted that, in both studies, no significant difference associated with age-group and gender could be identified.
Nevertheless, two remarkable aspects of our epidemiological observations deserve further discussion. On the one hand, antibodies to Ebola virus were detected in urban populations of Congolese blood donors, which should stimulate future investigations relating to asymptomatic seropositives. On the other hand, new environmental factors associated with seropositivity have been identified. In neighbouring Gabon, the seroprevalence was significantly higher in the forest regions (17.6%) than in savannah regions (10.5%) and Lakeland areas (2.7%) [40,48], but no sociodemographic or behavioural risk factors associated with EBOV seropositivity could be identified. Therefore, the finding in the current study that (i) exposure to bats and (ii) bird consumption are associated independently in multivariate analysis with an increased prevalence of antibodies to EBOV is of particular relevance.
The notion of bird consumption must be interpreted with caution. If birds are consumed, they will have been cooked, but bird consumption also implies contact with birds, either in the markets in Africa or when capturing them or preparing them for cooking. Whilst further studies to clarify the role of birds as a possible source of human infection are clearly justified, we prefer to refer to “exposure to birds” as a potential source of the detected seropositivity.
The possibility of exposure to birds, with supporting evidence, has never been described before as a risk factor for filovirus infection. To the best of our knowledge, birds have not been associated with EBOV or virus antigen in previous investigations [51]. In 1979 and 1980, a total of 1,664 animals (117 species, including 67 birds) were collected in the DRC and Cameroon during the dry season near the site of the 1976 Ebola haemorrhagic fever epidemic. This study failed to identify an animal reservoir of EBOV [52]. An ecologic investigation was also performed to identify the reservoir of EBOV following the 1995 outbreak in Kikwit, DRC. Most of the 3,066 collected specimens were mammals (87%, 2663), but birds (9%, 265) were also collected. All attempts at isolation of EBOV remained negative [53]. During a study in CAR (Ngotto forest) in 1998–99, 662 animals were trapped including 16 birds. Only Zaire ebolavirus nucleotide sequences were detected in six organs rodents [54]. Finally, experimental inoculation of birds with EBOV did not evidence significant acute or chronic replication [55].
In 1994, an outbreak of Ebola was described in a wild chimpanzee community in the Taï National Park, Ivory Coast. Before the beginning of the outbreak, chimpanzees spent a long time in a fig tree (Ficus goliath) that was full of fruits. Many birds (pigeons) were seen feeding on the figs during the day, and rodents and fruit bats were feeding on the tree during the night. It was proposed that the F. goliath could have constituted an epidemiological "hub", putting different species-including birds- in contact with one another [56].
The implication of birds in the epidemiology of filoviral disease was evoked by Monath [50] in a hypothetical transmission cycle, taking into consideration the possibility that filoviruses may be primary arthropod or plant viruses, transmitted to vertebrates including bats or other insectivorous species such as birds. Jeffers et al. [57] reported in 2002 a noticeable biochemical similarity between the glycoprotein of certain birds, oncogenic retroviruses and the EBOV glycoprotein. They suggested a possible common evolutionary origin which could be an indication that avian species have or have had a role in the ecology and evolution of filoviruses. Accordingly, the possible role of birds in the transmission of EBOV should not be ruled out and our intriguing observation suggests that additional studies to elucidate the ecology of filoviruses in birds would be of great interest.
The statistically significant identification of exposure to bats as an independent risk factor for EBOV infection is also remarkable since contact with bats has been reported as a risk factor for MARV infection [58,59,60,61], and fruit bats [6] (and possibly insectivorous bats [62]) are thought to have a significant role in the ecological cycle of EBOV. Therefore, the current study reinforces the suspected importance of bats in the natural cycle of Ebola virus and its possible transmission to humans.
Our results imply that in the Republic of Congo, the circulation of Marburg virus occurs at a very low rate without any identified risk factor, but that human exposure to Ebola virus without consequent disease is not infrequent. Living in Pointe Noire or in a rural area, and having contact with bats and birds is associated with a higher risk of exposure to Ebola virus. Unfortunately, little is known about the natural history and biological properties of EBOV antibody in individuals without haemorrhagic fever. Here, we did not observe an increase of seroprevalence with age, which may suggest that, in some individuals, the antibody titre decreases and becomes undetectable with time. This would in turn imply the absence of iterative contacts with Ebola virus antigens or that of a strong antibody response following such secondary antigenic exposure.
Similarly, the protection afforded by antibodies detected in blood donors against Ebola virus infection remains completely unknown. How individuals without any history of haemorrhagic fever acquire specific antibodies to EBOV and what are the biological properties of such antibodies (in particular what is their seroneutralising capacity) deserve further investigations in African populations.
|
10.1371/journal.pcbi.1003068 | Inference of Gene Regulatory Networks with Sparse Structural Equation Models Exploiting Genetic Perturbations | Integrating genetic perturbations with gene expression data not only improves accuracy of regulatory network topology inference, but also enables learning of causal regulatory relations between genes. Although a number of methods have been developed to integrate both types of data, the desiderata of efficient and powerful algorithms still remains. In this paper, sparse structural equation models (SEMs) are employed to integrate both gene expression data and cis-expression quantitative trait loci (cis-eQTL), for modeling gene regulatory networks in accordance with biological evidence about genes regulating or being regulated by a small number of genes. A systematic inference method named sparsity-aware maximum likelihood (SML) is developed for SEM estimation. Using simulated directed acyclic or cyclic networks, the SML performance is compared with that of two state-of-the-art algorithms: the adaptive Lasso (AL) based scheme, and the QTL-directed dependency graph (QDG) method. Computer simulations demonstrate that the novel SML algorithm offers significantly better performance than the AL-based and QDG algorithms across all sample sizes from 100 to 1,000, in terms of detection power and false discovery rate, in all the cases tested that include acyclic or cyclic networks of 10, 30 and 300 genes. The SML method is further applied to infer a network of 39 human genes that are related to the immune function and are chosen to have a reliable eQTL per gene. The resulting network consists of 9 genes and 13 edges. Most of the edges represent interactions reasonably expected from experimental evidence, while the remaining may just indicate the emergence of new interactions. The sparse SEM and efficient SML algorithm provide an effective means of exploiting both gene expression and perturbation data to infer gene regulatory networks. An open-source computer program implementing the SML algorithm is freely available upon request.
| Deciphering the structure of gene regulatory networks is crucial for understanding gene functions and cellular dynamics, as well as system-level modeling of individual genes and cellular functions. Computational methods exploiting gene expression and other types of data generated from high-throughput experiments provide an efficient and low-cost means of inferring gene networks. Sparse structural equation models are employed to: i) integrate both gene expression and genetic perturbation data for inference of gene networks; and, ii) develop an efficient sparsity-aware inference algorithm. Computer simulations corroborate that the novel algorithm markedly outperforms state-of-the-art alternatives. The algorithm is further applied to infer a real human gene network unveiling possible interactions between several genes. Since gene networks can be perturbed not only by genetic variations but also by other means such as gene copy number changes, gene knockdown or controlled gene over-expression, this paper's method can be applied to a number of practical scenarios.
| Genes in living organisms do not function in isolation, but may interact with each other and act together forming intricate networks [1]. Deciphering the structure of gene regulatory networks is crucial for understanding gene functions and cellular dynamics, as well as for system-level modeling of individual genes and cellular functions. Although physical interactions among individual genes can be experimentally deduced (e.g., by identifying transcription factors and their regulatory target genes or discovering protein-protein interactions), such experimental approach is time-consuming and labor intensive. Given the explosive number of combinations of genes involved in any possible gene interaction, such an approach may not be practically feasible to reconstruct or “reverse engineer” gene networks. On the other hand, technological advances allow for high-throughput measurement of gene expression levels to be carried out efficiently and in a cost-effective manner. These genome-wide expression data reflect the state of the underlying network in a specific condition and provide valuable information that can be fruitfully exploited to infer the network structure.
Indeed, a number of computational methods have been developed to infer gene networks from gene expression data. One class leverages a similarity measure, such as the correlation or mutual information present in pairs of genes, to construct a so-termed co-expression or relevance network [2], [3]. Another approach relies on Gaussian graphical models with edges being present (absent) if the corresponding gene pairs are conditionally dependent (respectively independent), given expression levels of all other genes [4], [5]. While the approach based on Gaussian graphical models entails undirected graphs, directed acyclic graphs (DAGs) or Bayesian networks have also been employed to infer the dependency structure among genes [6], [7]. The fourth approach employs linear regression models and associated inference methods to find the dependency among genes and to infer gene networks [8]–[11]. Finally, while these approaches use gene expression data in the steady-state, several methods exploiting time-series expression data have also been reported; see e.g., [12], [13] and references therein.
Recently, gene expression data from gene-knockout experiments have been combined with time series comprising gene expression data with perturbations to considerably improve the accuracy of network inference [14]. When a gene is knocked out or silenced, expression levels of other genes are perturbed. Different from using gene expression levels of the original network alone, comparing gene expression levels in the perturbed network with those in the original network reveals extra information about the underlying network structure. Gene perturbations can be performed with other experimental approaches such as controlled gene over-expression and treatment of cells with certain chemical compounds [8], [9]. However, these gene perturbation experiments may not be feasible for all genes or organisms. To overcome this hurdle, one can exploit naturally occurring genetic variations that can be viewed as perturbations to gene networks [15]. More importantly, such genetic variations enable inference of the causal relationship between different genes or between genes and certain phenotypes.
Several approaches are available to capitalize on both genetic variations and gene expression data for inference of gene networks. The first approach models a gene network as a Bayesian network, and then infers the network by incorporating prior information about the network obtained from expression quantitative trait loci (eQTLs) [16]–[18]. In the second approach, a likelihood test is employed to search for a casual model that “best” explains the observed gene expression and eQTL data [19]–[23]. The third approach relies on the structural equation model (SEM) to infer gene [24]–[27] or phenotype networks [28]–[34]. While these approaches focus on inference of gene networks incorporating information from eQTL, another approach employs both phenotype and QTL genotype data to jointly decipher the phenotype network and identify eQTLs that are causal for each phenotype [35]. Logsdon and Mezey [26] proposed an adaptive Lasso (AL) [36] based algorithm to infer gene networks modeled with an SEM. They compared the performance of a number of methods using simulated directed acyclic or cyclic networks. Their simulations showed that the AL-based algorithm outperformed all other methods tested. Despite its superiority over other methods, the AL-based algorithm does not fully exploit the structure of the SEM. Therefore, it is expected that a more systematic inference algorithm may significantly improve the performance of the SEM-based approach.
Motivated by the fact that gene networks or more general biochemical networks are sparse [8], [37]–[39], a sparse SEM is advocated in this paper to infer gene networks from both gene expression and eQTL data. Incorporating network sparsity constraints, a sparsity-aware maximum likelihood (SML) algorithm is developed for network topology inference. The core technique used is to maximize the likelihood function regularized by the -norm of the parameter vector determining the network structure. The -norm controls complexity of the SEM, and thus yields a sparse network. The key innovative element of the SML algorithm is a block coordinate ascent method derived to maximize the -regularized likelihood function, which makes the SML algorithm computationally efficient. The simulations provided demonstrate that the novel SML algorithm offers significantly better performance than the two state-of-the-art algorithms: the AL [26], and the QDG algorithm [21]. The SML algorithm is further applied to infer a human network of 39 human genes related to the immune function.
Consider expression levels of genes from individuals measured using e.g., microarray or RNA-seq. Let denote the vector collecting the expression levels of these genes of individual . Suppose that a set of perturbations to these genes has been also observed. These perturbations can be due to naturally occurring genetic variations near or within the genes, gene copy number changes, gene knockdown by RNAi or controlled gene over-expression. In this paper, focus is placed on genetic variations observed at eQTLs, although the network model and the inference method described in the next section are also applicable to cases where other perturbations are available. As in [26], it is assumed that each gene has at least one cis-eQTL so that the structure of the underlying gene network is uniquely identifiable. Let denote the genotype of eQTLs of individual . The goal is to infer the network structure of the genes from the available gene expression measurements , , and eQTL observations , .
As in [25], [26], the gene network is postulated to obey the SEM(1)where matrix contains unknown parameters defining the network structure; matrix captures the effect of each eQTL; vector accounts for possible model bias; and vector captures the residual error, which is modeled as a zero-mean Gaussian vector with covariance , where denotes the identity matrix. It is assumed that no self-loops are present per gene, which implies that the diagonal entries of are zero. As mentioned in [26], lack of self-loops and a diagonal covariance matrix of are commonly assumed in almost all graph-based network inference methods. It is further assumed that the loci of eQTLs have been determined using an existing eQTL method, but the effective size of each eQTL is unknown. Therefore, has unknown entries whose locations are known and remaining zero entries (for instance is a diagonal matrix when ).
The network inference task is to estimate unknown entries of , and as a byproduct, the unknown entries of . Without any knowledge about the network, no restriction is imposed on the structure specified by . Therefore, the network is considered as a general directed graph that can possibly be a directed cyclic graph (DCG) or a DAG. Network inference is challenging since the number of unknowns to be estimated is very large for a moderately large . Note that under the assumption that each gene has at least one cis-eQTL, the “Recovery” Theorem in [26] guarantees that the network is identifiable for both DCGs and DAGs.
As discussed in [8], [37]–[39], gene regulatory networks or more general biochemical networks are sparse meaning that a gene directly regulates or is regulated by a small number of genes relative to the total number of genes in the network. Taking into account sparsity, only a relatively small number of the entries of are nonzero. These nonzero entries determine the network structure and the regulatory effect of one gene on other genes. The SEM in (1) under the aforementioned sparsity assumption will be henceforth referred to as the sparse SEM. Exploiting the sparsity inherent to the network, an efficient and powerful algorithm for network inference will be developed in the ensuing section.
Upon defining , , and , the SEM in (1) can be compactly written as , where 1 is the vector of all-ones. Given and , the log-likelihood function can be written as(2)where denotes matrix determinant, and denotes the Frobenius norm.
As mentioned earlier, is a sparse matrix having most entries equal to zero. In order to obtain a sparse estimate of , the natural approach is to maximize the log likelihood regularized by the weighed term , where denotes the th entry of . In a linear regression model, it is well known that the -regularized least-squares estimation also known as Lasso [40] can yield a sparse estimate of the regression coefficient vector. Similarly, the -regularized maximum likelihood (ML) approach used here is expected to shrink most of the entries of toward zero, thereby yielding a sparse matrix. It is easy to show that maximizing with respect to (w.r.t.) yields , where and . Upon defining , , , , and substituting for in (2), the proposed -penalized ML estimation approach yields(3)where denotes the set of row and column indices of the entries of known to be zero. As assumed earlier, each phenotype has at least one cis-eQTL that has been identified, which implies that the locations of nonzero entries of or equivalently the set is known. However, our sparse SEM and inference method are also applicable to more general cases where some or all phenotypes have cis-eQTLs that have not been identified. In these cases, the locations of nonzero entries of corresponding to the unidentified cis-eQTLs are unknown. We can form a weighted -norm of the entries of excluding those corresponding to the identified cis-eQTL and then add a penalty term involving this -norm to the objective function in (3). This new optimization problem can be solved efficiently using a method modified from the one solving (3), as it is described in the supporting text S1.
Weights in the penalty term are introduced to improve estimation accuracy in line with the AL [36]. They are selected as , where is found using a preliminary estimate of obtained via ridge regression as(4)The sparsity-controlling parameters in (3) and in (4) are selected via cross validation (CV), while is estimated as the sample variance of the error using and . In adaptive Lasso based linear regression [36], Zou suggested using the ordinary least squares (OLS) estimate to determine the weights; if the OLS estimate does not exist due to, e.g., collinearity, Zou suggested the estimate obtained from ridge regression, although it remains to show if the ridge regression estimate is consistent in this case and if the resulting adaptive Lasso yields the desired oracle properties. If OLS is used for estimating and in the SEM, the solution usually does not exist since the number of unknowns is typically larger than the number of samples. However, even in this case the solution can always be obtained from ridge regression as in (4). Moreover, every entry of the solution is typically nonzero, which yields a finite weight for every variable, and thus every variable will be included in the following -penalized ML procedure. An alternative approach is to replace the weighed -norm in (3) with an unweighted -norm to obtain a preliminary estimate of and then calculate the weights from this preliminary estimate, as in [26]. However, the unweighted -penalized ML procedure may shrink many variables to zero and exclude them from the weighted -penalized ML estimator, possibly yielding a biased estimate. For this reason, the inference method in this paper uses ridge regression to determine , with the additional advantage of (4) admitting a closed-form solution.
A block diagram of the novel inference algorithm, abbreviated as the sparsity-aware maximum likelihood (SML) algorithm, is depicted in Figure 1. The first and third blocks in Figure 1 perform cross-validation to select optimal parameters and to be used in (3) and (4), respectively (see the description of the cross-validation procedure in the supporting text S1.) The third block produces weights and error-variance estimate after solving (4). Finally, the fourth block takes data and together with , and and solves (3) to yield , representing the SML estimator for in (1) and revealing the genetic-interaction network. As it will be described in the Methods section, (4) is separable across rows of and , and each row of and becomes available in closed form [cf. (8)–(9)]. The -regularized ML problem (3) is solved efficiently using a novel block coordinate ascent iterative scheme given by (11)–(16) in the Methods section. Precise description of the overall SML algorithm is also presented in the Methods section as Algorithm 1, which was used to yield an executable computer program.
In their simulation studies, Logsdon and Mezey [26] compared the performance of their AL-based algorithm with that of several other algorithms including the PC-algorithm [41], [42], the QDG algorithm [21], the QTLnet algorithm [35], and the NEO algorithm [22]. In two out of four simulation setups, the AL outperformed all other algorithms; and in the other two simulation setups, the AL and QDG algorithms exhibited comparable performance, but consistently outperformed the other two algorithms. Logsdon and Mezey [26] also considered other existing algorithms [25], [43], but these were deemed either computationally too demanding [43] or prohibitively complex [25]. For these reasons, the AL and QDG algorithms are regarded as state-of-the-art in the field. Their performance was compared against this paper's SML algorithm.
Following the setup of Logsdon and Mezey [26], two types of acyclic gene networks were simulated first: one with 10 genes and another with 30 genes. Specifically, a random DAG of 10 or 30 nodes with an expected edges per node was generated by creating directed edges between two randomly picked nodes. Care was taken to avoid any cycle in the simulated graph. If an edge from node to node was emerging, was generated from a random variable uniformly distributed over the interval or ; otherwise, . The genotype per eQTL was simulated from an F2 cross. Values 1 and 3 were assigned to two homozygous genotypes, respectively, and 2 to the heterozygous genotype. Hence, was generated as a ternary random variable taking values with corresponding probabilities . Matrix was the identity matrix, was sampled from a Gaussian distribution with zero mean and variance , and was set to zero. Finally, was calculated from
For each type of gene network, 100 realizations or replicates of the network were generated, and then the SML, the AL and the QDG algorithms were run to infer the network topology. When running the SML algorithm, 10-fold CV was employed to determine the optimal values of parameters and and then use these values to infer the network. An edge from gene to was deemed present if . The AL algorithm also automatically ran using CV to determine the values of its parameters. For 100 replicates of the network, counted the total number of edges, denoted the total number of edges detected by the inference algorithm. Among detected edges, stands for the number of true edges presented in the simulated networks, and for the number of false edges. The power of detection (PD) was then found as , and the false discovery rate (FDR) as . The PD and the FDR of the SML, AL, and QDG algorithms for different sample sizes are depicted in Figure 2. It is seen from Figures 2 (a) and (c) that the PD of the SML algorithm exceeds 0.9 for both networks across all sample sizes, whereas the PD of the AL algorithm is about 0.65 for and 0.35 for . The PD of the QDG algorithm is even lower ranging from 0.22 to 0.33. As shown in Figures 2 (b) and (d), the FDR of the SML algorithm is on the order of for most sample sizes, and is much lower than that of the AL and QDG algorithms, which is about 0.3 for and over the range from 0.31 to 0.6 for .
Two types of cyclic networks were subsequently simulated: one with 10 genes and the other with 30 genes. The average number of edges per gene is again equal to 3. The same procedures used in simulating acyclic networks described earlier were employed, except that DCGs instead of DAGs were simulated. Again, 100 replicates for each type of the networks were randomly generated. The PD and the FDR of three algorithms are depicted in Figure 3. As shown in Figure 3 (a) and (c), the PD of the SML algorithm is between 0.83 and 0.9, whereas the PD of the AL algorithm is about 0.52 for and 0.29 for , and the PD of the QDG algorithm is between 0.16 and 0.28. As shown in Figures 3 (b) and (d), the FDR of the SML algorithm is , which is much smaller than that of the AL and QDG algorithms over the range from 0.33 to 0.68. For the convenience of comparison, the results in Figures 2 and 3 at sample size 500 are summarized in Table 1.
As confirmed by Figures 2 and 3, the SML algorithm offers much better performance in terms of PD and FDR than the AL and QDG algorithms. However, these results were obtained for gene networks of small size. To test performance of the SML algorithm for networks of relatively large size, an acyclic network of 300 genes was simulated with an expected edge per node, and 10 replicates of the network were randomly generated. PD and FDR of the SML and AL algorithms obtained from these replicates are depicted in Figure 4. The PD of SML exceeds across all sample sizes from 100 to 1,000, whereas that of the AL algorithm is about 0.04 for sample sizes from 100 to 500, and gradually increases to 0.42 at the sample size of 1,000. The FDR of SML stays below for sample sizes from 400 to 1,000, whereas the FDR of the AL algorithm is on the order of for the same sample size. When the sample size is relatively small (in the range from 100 to 300), the FDR of SML is higher than that of the AL algorithm, but it is still relatively small (). Note that the AL algorithm essentially does not work for sample sizes , since its power is too small. All simulation results show that the novel SML algorithm significantly outperforms the AL and QDG algorithms in terms of PD and FDR.
An extra set of simulations assessing the stability of SML is described in the section of “Stability of model selection under CV with different folds” in supporting text S1, and in Figures S1 and S2. As an alternative to CV, stability selection (STS) [44] provides a means of selecting an appropriate sparsity level to guarantee that the FDR is less than a theoretical upper bound. The STS procedure was applied to the SML algorithm as described in the supporting text S1, and was used with the selection probability cutoff and an upper bound or target FDR = 0.1 in simulations for the networks in Figures 2[(c) and (d)] and 3 [(c) and (d)]. As shown in Figure S3, the FDR of the STS is indeed much smaller than the target FDR and almost uniform across different sample sizes, but the PD of the STS is smaller than that of CV. In fact, the FDR of the STS is on the same order as that of the CV except at the sample size of 100 for the DAG. As seen from these simulation results, although the STS guarantees a FDR upper bound, this upper bound is loose for the simulation setups tested, which may sacrifice detection power. Nevertheless, the STS procedure can select a set of stable variables as described in [44] and verified by our simulations.
So far, all the simulated data were generated with noise variance . Next, the performance of SML was analyzed for simulated networks of 30 genes, when was increased to 0.05 and was changed from 3 to 1 or 5. Reducing from 3 to 1 improved the performance of SML for most of the sample sizes, as it is depicted in Figure 5, withstanding the increase in the noise variance. Increasing at constant , or increasing at constant degraded the performance, most notably in the later case. Comparing Figure 5 with Figures 2 and 3 [(c) and (d)] demonstrates that in both cases the SML estimates still achieve higher detection power and lower FDR than those estimates obtained with the AL algorithm for and .
Pickrell et al. [45] used RNA-Seq technology to sequence RNA from 69 lymphoblastoid cell lines derived from unrelated Nigerian individuals extensively genotyped by the International HapMap Project [46]. For each gene, they evaluated possible associations between its gene expression level calculated from RNA-Seq reads and all 3.8 million single nucleotide polymorphisms (SNPs) using the genotypes from phases II and III of the HapMap Project. At FDR = 0.1, they identified 929 genes or putative new exons that have eQTLs within 200 kb of the gene or the exon. From these 929 genes, 39 genes that are related to immune functions were selected manually by an expert as mentioned in the Acknowledgements section; expression levels and the genotypes of the eQTLs of these 39 genes in 69 individuals were used to infer the underlying regulatory network.
Pickrell et al. normalized expression values using quantile normalization before performing eQTL mapping. They also provided a data set that contains the number of reads mapped to each of 929 genes. This data set was obtained and the number of reads for each of 39 genes was normalized with the length of the gene to yield expression value. Such kind of values may better reflect the real expression values than the values normalized with quantile normalization, and thus they were used to infer the network. To ensure the quality of the data, the SAS ROBUSTREG procedure was applied to 69 expression values of each of 39 genes to detect outliers. The default M estimation method of the ROBUSTREG procedure was employed and the outliers were detected at a significance level of 0.05. Several gene expression values were identified as outliers since they are much larger than the remaining values that were classified as non-outliers. The outliers were replaced with the largest non-outlier. More sophisticated means of revealing and imputing outliers are possible using robust statistical schemes; see e.g., [47]. The genotypes of the eQTLs of the 39 genes were downloaded from HapMap database using the SNP IDs for the eQTL provided by Pickrell et al.. About 12% genotypes are missing. These missing genotypes were imputed using the program IMPUTE2 [48]. The name and a brief description of each gene were obtained from DAVID [49] using the Ensembl gene IDs provided by Pickrell et al. Information of these 39 genes including their Ensembl gene IDs and names, a brief description of each gene, and HapMap SNP IDs of the associated eQTLs can be found in Table S1 in the supporting information.
The SML algorithm was run with the expression levels and genotypes of eQTLs of these 39 genes. An edge from gene to was detected if . To improve the reliability of the detected edges, the SML algorithm was run with stability selection at an FDR using 100 random subsamples, yielding 13 directional edges as shown in Figure 6. The frequency of each edge detected in 100 runs is given in Table S2. It is interesting to see from Figure 6 that only 9 genes are involved in the network, and the remaining 30 genes are not connected with any other genes and thus not shown in the figure. AL and QDG algorithms were also run with stability selection at an FDR using 100 random subsamples. The edges detected by AL and QDG algorithms and their frequencies are included in Table 4. The AL algorithm detected only one edge that was not detected by the SML algorithm. The QDG yielded 3 edges, one of which was also detected by the SML algorithm. The relatively small number of edges detected by three algorithms was likely due to relatively low signal-to-noise ratio (SNR) in this data set. The estimated noise variance was and the estimated SNR was dB, which was much lower than that (about 25.8 dB) in the case of in Figure 5. However, comparing the results of three algorithms shows that our SML algorithm detected more edges than the other two algorithms at the same FDR due to its higher detection power as confirmed also by the simulations. When the FDR was increased to , the SML algorithm with stability selection yielded a network of 16 genes that have 42 edges as shown in Figure S4 in the supporting information. Since only 39 genes were used to construct the network, an edge between two genes may not necessarily imply a direct regulatory effect, but may reflect the fact that two genes are either directly linked or very close to each other in the real network that consists of all genes. Particularly, if two genes are co-regulated by another gene which is not included in the 39 genes, these two genes may have a unidirectional or bidirectional edge.
Most edges in Figure 6 are between major histocompatibility complex (MHC) genes (HLA-A, HLA-DPA1, HLA-DQA2, HLA-DQB1, HLA-DRB4 and HLA-DRB5), which is expected since these genes may interact with each other and/or be co-regulated. FCRLA is a member of Fc receptor-like family of genes. It is expressed in B cells and interacts with IgG and IgM [50], [51]. IGH, encoding the heavy chain of immunoglobulin, characterizes the B-cell origin of the samples. Hence, it is not surprising to see an edge between FCRLA and IGH. Interleukin-4-induced gene 1 (IL4I1) was first described in the mouse [52] and subsequently characterized in human B cells [53]. Human IL4I1 is expressed by antigen-presenting cells [54], which may allude to the edge between HLA-A and IL4I1, but this may be speculative since there is no edges between IL4I1 and MHC class II genes in the network. The edges between IGH and HLA-A and between IGH and HLA-DRB4 may reflect the coordinated effect of antibody and MHC as a response to antigens. In fact, IGH is connected to most of MCH genes in Figure S4, which may imply the wide coordination between the two classes of molecules.
Integrating genetic perturbations with gene expression data for inference of gene networks not only improves inference accuracy, but also enables learning of causal regulatory relations among genes. Although much progress has been made recently on the development of inference methods that integrate both types of data, a truly efficient algorithm is missing. The SEM provides a systematic framework to integrate both types of data, and offers flexibility to model both directed cyclic as well as acyclic graphs. However, there is no systematically designed inference method for SEMs of relatively high dimension, which is particularly true for gene networks typically including hundreds or thousands of genes. Traditionally, inference for SEMs has relied on the ML or generalized least-squares methods implemented with a numerical optimization algorithm [55], [56]; but recently, Bayesian alternatives [57] have emerged too, based on Markov chain Monte Carlo simulations [58], [59]. These methods not only are computationally intensive, but also may be inaccurate for sparse SEMs of relatively high dimension, since they do not account for sparsity present in the model.
In the context of QTL mapping, Newton's method is employed in [27] to implement the ML method, while the genetic algorithm [60], [61] is used in [24], [25] to maximize the likelihood function, and in conjunction with a model selection method using a test or Occam's window to search for the best network topology. These methods are not scalable to SEMs of relatively high dimension. The AL-based algorithm proposed in [26] is more efficient because it automatically incorporates variable selection into the inference process, and also takes into account the sparsity present in gene networks. However, the AL-based scheme borrows the adaptive Lasso [36] optimally designed for the linear regression model instead of the SEM. In contrast, the SML algorithm proposed in this paper directly maximizes the -regularized likelihood function of the SEM, which fully exploits the information present in the data and therefore improves inference accuracy. Moreover, the novel block coordinate ascent method combined with discarding rules can efficiently maximize the -regularized likelihood function, rendering the SML algorithm applicable to SEMs of high dimension. However, unlike the AL-based algorithm, the SML algorithm maximizes a non-convex objective function as given in (3). Although the “Recovery” Theorem in [26] guarantees the identifiability of the network, the algorithm can converge to a local maximum that may not necessarily be coincident with the global maximum corresponding to the optimal network. A common technique for alleviating this problem is to use multiple random initial values. We tested multiple initial values in our simulations and observed that the algorithm converged to the same solution. In Algorithm 2, we used the pathwise coordinate optimization strategy as used in [62], where the solution of (3) obtained with was used as the initial point for the run with . The pertinence of this strategy is corroborated by simulated numerical tests, showing significant performance gains of the SML algorithm in terms of detection power and FDR when compared to the AL-based algorithm.
Comparisons in the Simulation Studies section, as summarized in Figures 2–5, demonstrated that the SML algorithm markedly outperforms two state-of-the-art algorithms: the AL [26] and QDG [21] algorithms. For three directed acyclic networks with number of genes and 300, respectively, the PD of the SML algorithm exceeds 0.9 for all sample sizes from 100 to 1,000, and is greater than 0.99 for most sample sizes. This is much greater than the PD of the AL and QDG algorithm that ranges from 0.004 to 0.67. In fact, The QDG algorithm was too time-consuming to obtain results for . The FDR of SML is on the order of for most sample sizes, which is much smaller than those of the AL and QDG algorithms, that are between 0.25 and 0.6 for and 30. The FDR of the AL algorithm for is between 0.02 and 0.1. The only case where the FDR of SML exceeds that of the AL algorithm is when , and the sample size . However, the AL algorithm essentially does not work in this case, since its PD is about 0.04. In the case of directed cyclic networks, all algorithms offer slightly degraded performance when compared to that of directed acyclic networks. However, the SML algorithm still considerably outperforms the AL and QDG algorithms.
Using a limited amount of available data [45], 39 genes related to the immune system and having one eQTL per gene were selected to infer a possible network among these genes. At an FDR 10% for the detected edges, a network of 9 out of 39 genes containing 13 edges were obtained. An edge between two genes in the inferred network may be an indication of the direct regulator effect, or indirect interaction or co-regulation mediated by some other genes that are not among the 39 genes. The majority of the edges were reasonably expected from the experimental results in the literature, while the remaining edges may represent new interactions to be elucidated.
Structural equation modeling has a long history of about a century, with well-documented contributions to various fields including biology, psychology, econometrics and other social sciences [55], [56], [63], [64]. The model considered in this paper belongs to a class of SEMs with observed variables [55]. The SML algorithm is the first one that is systematically developed for inferring sparse SEMs with observed variables. It is expected to accelerate the application of high-dimensional SEMs not only in biology, but also in other fields.
The overall SML approach described in the Methods section, including the ridge regression weights, the discarding rules, and the coordinate descent cycle is depicted step-by-step in Algorithm 1. The for-loop starting from line 8 and ending at the last line is the -regularized ML method for computing and in (3), which comprises the block coordinate ascent algorithm and discarding rules. In our computer program, these lines were written as a subroutine. Since the CV on line 7 needs to solve (3), the subroutine is also called on line 3 with varying from to . An additional subroutine implementing ridge regression was written to solve (4), and subsequently called on lines 1 and 2.
In the supporting text S1, three relevant extensions to the SML algorithm are described. First, stability selection [44] is applied to the SML, as an alternative to CV, to select the sparsity level so that the FDR is controlled. Second, the SML is extended to handle heteroscedasticity in the SEM error. Third, the SML is modified to enable inference of unknown eQTLs. In addition, supporting text S1 gives a description of the state-of-the-art AL-based and QDG algorithms that were considered for comparison with SML.
|
10.1371/journal.ppat.1000563 | Nuclear Entry of Hepatitis B Virus Capsids Involves Disintegration to Protein Dimers followed by Nuclear Reassociation to Capsids | Assembly and disassembly of viral capsids are essential steps in the viral life cycle. Studies on their kinetics are mostly performed in vitro, allowing application of biochemical, biophysical and visualizing techniques. In vivo kinetics are poorly understood and the transferability of the in vitro models to the cellular environment remains speculative. We analyzed capsid disassembly of the hepatitis B virus in digitonin-permeabilized cells which support nuclear capsid entry and subsequent genome release. Using gradient centrifugation, size exclusion chromatography and immune fluorescence microscopy of digitonin-permeabilized cells, we showed that capsids open and close reversibly. In the absence of RNA, capsid re-assembly slows down; the capsids remain disintegrated and enter the nucleus as protein dimers or irregular polymers. Upon the presence of cellular RNA, capsids re-assemble in the nucleus. We conclude that reversible genome release from hepatitis B virus capsids is a unique strategy different from that of other viruses, which employs irreversible capsid destruction for genome release. The results allowed us to propose a model of HBV genome release in which the unique environment of the nuclear pore favors HBV capsid disassembly reaction, while both cytoplasm and nucleus favor capsid assembly.
| Viral capsids facilitate protection of the enclosed viral genome and participate in the intracellular transport of the genome. At the site of replication capsids have to release the genome, but after replication new capsids have to be assembled for encapsidation of the progeny genomes. Detailed data on stability of capsids and kinetics of their formation and dissociation are obtained for several viruses in vitro, allowing biophysical or electron microscopical techniques. These approaches, however, do not consider the impact of cellular interaction partners. Using digitonin-permeabilized cells which support hepadnaviral genome release actively, we analyzed the disassembly kinetic of the hepatitis B virus (HBV) capsid. Using different analytical methods we found that HBV capsids disintegrate to protein dimers which reassemble to capsids inside the nucleus. The study provides a link between in vitro and in vivo data showing that HBV uses a unique strategy. We propose a model in which the unique environment of the nuclear pore favors the disassembly reaction, while both cytoplasm and nucleus favor assembly.
| Viral capsids facilitate multiple functions in the viral life cycle. Outside the cell, they protect the enclosed viral genome against nucleases, and in case of non-enveloped viruses they mediate attachment and entry. For both enveloped and non-enveloped viruses, they carry the viral genome to the site of replication where they have to release the genome in order to allow access of transcription and/or replication factors. After replication new capsids have to be assembled for encapsidation of the progeny genomes and subsequent release of mature virions. Capsids are assigned to be metastable: early in infection they have to open, later they have to assemble and close. Most data on stability of capsids and kinetics of their formation and dissociation are obtained in vitro allowing analysis by biophysical or electron microscopical techniques (e.g. [1]–[5])
In vivo data on capsid disintegration are rare despite of their importance for genome release and viability of infection. Two different examples are Adenovirus 5 (Ad-5) capsids and the capsid of Herpes simplex virus-1 (HSV-1). Ad-5 capsids disintegrate to penton and hexon subunits after modification upon endocytosis [6],[7] for genome release. Capsids of HSV-1 in contrast remain stable after release of one penton [8] which allows the injection of the viral DNA from that opening [9].
Intensive studies on capsid assembly were performed in vitro on capsids of the medically important Hepatitis B virus (HBV) [1],[5],[10],[11]. HBV infection is endemic in large parts of the world and ∼350 Mio people are chronically infected, accounting for 1 million deaths per year. HBV is an enveloped virus with an icosahedral capsid that is composed of 240 or 180 copies of a single protein species called core protein [12]. Within the oxidizing environment outside the cell the two core protein subunits of a dimer become linked by three disulfide bonds (Cys 48, 61 and 183 [13],[14]). The capsid encloses the relaxed circular viral DNA (rcDNA) [15], which is covalently attached to the viral polymerase [16].
HBV cell entry is the limiting step that prevents infection of most cell cultures but it can be by-passed by lipofection of hepatoma cells with virion-derived capsids. Using this artificial mode of capsid entry, HBV production reaches in vivo-like efficiency [17]. This suggests that the capsids either do not become modified during natural entry or that lipofection changes their structure in the same way. Complex interactions with the cellular transport machinery mediate capsid translocation to the nuclear periphery, passage through the NPC and liberation of the viral DNA [17]–[19].
Transport and genome release must be highly efficient and well-coordinated, because ∼80% of virions are infection-competent in vivo [20]. Within the nucleus, viral DNA is converted by cellular repair mechanisms to a covalently closed circular form (cccDNA), which is the template for viral mRNA synthesis including the RNA pregenome. Interaction of the RNA pregenome with the viral polymerase facilitates encapsidation into the viral capsid [21]. The polymerase retrotranscribes RNA pregenome into the rcDNA, but this occurs only within the capsids. This genome maturation requires multiple phosphorylation steps within the C terminus of the core protein [22],[23]. Mature capsids (Mat-C) can either be enveloped by the viral surface proteins in order to form virions or they can be transported through the NPC causing amplification of nuclear viral DNA.
Liver histology of HBV-infected individuals shows massive intranuclear capsid accumulation. However, the number of intranuclear viral genomes is low. Thus, it is believed that unassembled core protein pass the NPC and assembles intranuclear [24].
HBV capsids exhibit predominantly a T = 4 symmetry with a diameter of ∼36 nm. Core protein assembly is independent of eukaryotic host proteins as it occurs also upon core protein expression in E. coli resulting in “recombinant” capsids (rC). In contrast to natural capsids, rC are unphosphorylated and contain E. coli RNA [25]. In addition, they exhibit one unnatural disulfide bond linking the C terminal Cys (C183) of one core protein with a C183 of a neighboring dimer [26].
Structure of the first 143 aa of core proteins in rC has been obtained by X ray crystallography with a resolution of 3.3 Å [27]. The core monomer comprises two long α-helices with a hairpin structure. Association of two hairpins from two monomers forms a spike which protrudes from the particle surface. The connecting loop is exposed on the spike tip and comprises the immune dominant c/e1 B cell epitope, so that most antibodies are conformation-dependent. While the first 143 aa of the core protein are well structured [12], the C terminus is flexible: whereas the C terminus is localized within capsid lumen in E. coli-expressed capsids [28], it is exposed to the exterior in viral mature capsids [19]. The latter observation may however indicate that capsids dynamics increase with genome maturation.
In vitro association kinetics, performed on E. coli-expressed, C terminally truncated core proteins (aa 149), showed that capsid formation starts with core protein dimers. It is thought that the dimers trimerize and the resulting hexamer nucleates capsid assembly without accumulating further distinct populations of capsid subassemblies [10],[29]. According to the laws of thermodynamics, disassembly could just be the inverse reaction because the attractive forces between the subunits are weak, allowing them to transiently dissociate and re-associate (capsid breathing) [1],[5] in a way similar to that observed for polio-, flockhouse- and rhinoviruses [2]–[4]. In fact, recent in vitro evaluations showed that chaotropic agents as urea cause disassembly down to core protein dimers without distinguishable capsid subassemblies [30]. Several differences of these in vitro studies with the in vivo situation deserve attention: the C terminus, which interacts with encapsidated nucleic acids [31] and comprises the phosphorylation sites, was deleted in these studies; the capsids contained neither RNA nor DNA nor the polymerase. Moreover, host factors explaining the highly time- and site-coordinated HBV genome release were not present.
Accounting for the poor knowledge on in vivo disassembly and the medical importance of HBV, we evaluated the fate of HBV capsids within the cell. As no efficient infection system exists, we used digitonin-permeabilized cells that promote genome liberation into the nucleus. In order to distinguish reliably between input capsids and products of disassembly and re-assembly, we determined antibody binding properties, density, and size of authentic and recombinant capsids, and compared these particles with the products of nuclear import.
We characterized 32P labelled mature DNA-filled capsids from cell culture and RNA-filled capsids from E. coli as reagents. These capsids exhibited different densities upon Nycodenz gradient centrifugation. DNA-filled capsids undergo a transition to a density that resembled the RNA-filled state that correlates with nuclear entry. We demonstrated that E. coli expressed capsids dissociate and reassociate. Using the assembly/disassembly states we characterized two anti core protein antibodies and found different specificities for assembly states. These characteristics allowed us to analyse the intracellular fates of DNA-filled capsids by immune fluorescence microscopy.
Phosphorylation of both HBV virion-derived (Mat-C) and E. coli-derived (rC) capsid species resulted in specific labelling of the core proteins (Fig. 1A). In order to exclude that contaminant proteins accounted for the radiolabelled 21.5 kDa band we performed an immune precipitation using a polyclonal anti HBV capsid antibody (DAKO Ab). This antibody does not react with denatured core protein or core protein aggregates generated by acidification [18]. Fig. 1B showed that the DAKO Ab completely precipitated the 21.5 kDa band. This finding confirmed the identity of the band as core protein and showed that no other 32P-labelled protein co-migrated. It further indicated that all core proteins exhibited their proper conformation after phosphorylation. To analyze whether the core proteins were assembled to capsids we performed a native agarose gel electrophoresis of (Fig. 1C). No disintegrated capsids or aggregated core proteins exhibiting slower and diffuse migration were observed [18]. Phosphorylated rC (P-rC) caused a minor, slower migrating band. The additional band is characteristic for E. coli-expressed capsids. It is assumed that the minor band represent two capsids linked by RNA, as RNase A-treatment reduces the slower migrating band. The presence of this band thus indicates that the trapped RNA was not degraded upon in vitro phosphorylation. Both capsid species reacted with the DAKO Ab confirming their exposure of core protein epitopes.
We next analyzed possible contaminations with nucleases in the capsid preparations. We used 32P labelled dsDNA, ssDNA and RNA and determined the hydrolysis after 2 h incubation at 37°C (Fig. 1D). We observed reductions of TCA-precipitable 32P up to 10.2%. As the pipetting error in this assay was determined to be 4% we assume that the nuclease activities in the capsid preparations were not significant.
The 32P capsid preparations were analyzed by Nycodenz gradient centrifugation. Nycodenz was chosen as it conserves protein interactions better than any other gradient media, allowing the recovery of functionally active protein complexes [32]. In our hands, Nycodenz allowed the recovery of 95% of E. coli-expressed capsids loaded, while CsCl, which is known to allow separation of genome-containing and empty capsids - caused significant capsid disintegration to 10% (data not shown). As Nycodenz has properties similar to sucrose that allows a separation based on the sedimentation coefficient. In contrast to sucrose however, Nycodenz allows capsids to reach their equilibrium. In order to determine sedimentation of unassembled core proteins, we analyzed urea disintegrated capsids by centrifugation. Due to the unphysiological disulfide bonds of the C terminal Cys upon expression in E. coli, the assay was performed using the C183S P-rC mutant.
Sedimenting the capsids on Nycodenz gradient resulted in a peak of P-rC from 1.257 to 1.283 g/ml, with a maximum at 1.283 g/ml (Fig. 2A, peak (1)). The three major fractions contained 91% of the radioactivity. P-Mat-C banded at a slightly lower density between 1.242 and 1.284 g/ml with a maximum at 1.263 g/ml (Fig. 2B, peak (2)). The four major fractions contained 98% of the radioactivity indicating that capsids had not undergone significant degradation during centrifugation.
Urea dissociated core proteins remained close to the top of the gradient exhibiting a peak at 1.156 g/ml (Fig. 2C, peak (3)). Likely, these core proteins represent dimers as it was reported upon urea disintegration by others [30].
Despite the fact that rcDNA has a ca. 1.8 fold higher molecular mass than RNA, the density of P-Mat-C was slightly lower. In Nycodenz however, RNA has a higher density (1.18 g/ml) than dsDNA (1.13 g/ml) while proteins have 1.31 g/ml in Nycodenz [32]). Therefore, the different density distribution of Mat-C and rC implies that the gradient medium entered capsid lumen. To test this hypothesis we searched for Nycodenz entry by electron microscopy. Nycodenz, an electron-rich solute, can be seen without addition stain. Fig. 2D depicts that Nycodenz caused 20–24 nm large dots which were absent in the negative control. The size of the stain was similar to the lumen of P-rC stained by phospho tungstic acid, which exhibited the external diameter of 35 nm known for HBV capsids.
We used rC capsids in order to separate capsids from core protein dimers and capsid subassemblies. Separation was performed by Superdex 75 size exclusion chromatography using C183S rC capsids.
Separating a stock solution of C183S-rC revealed a single peak in fractions A4 (0.85 ml) (Fig. 3A). Its appearance in the exclusion volume of the column implies that practically all core proteins were assembled to capsids. However, when a 20 fold lower core protein amount was applied three peaks occurred (Fig. 3B): the strongest peak (peak A; 51% of total protein) appeared in fraction A3/4 (0.85 ml), which is within the exclusion limit of the column. A small peak (B, ∼3%) appeared in fraction B9 (1.73 ml) and another stronger peak (peak C, 46%) was observed in fraction C1/2 (1.92 ml).
Re-application of peak A after 1 h at RT resulted in the re-appearance of peaks A, B and C (69% peak A, 10% peak B and 23% peak C; Fig. 3C). Re-injection of peak C revealed the same peaks but with a different distribution (A 11%, B 10%, C 79%) (Fig. 3D). This finding confirms that rC undergoes dissociation, as described previously [1],[5].
For analyzing the immune reactivity of the fractions, we used two anti capsid antibodies: (1) polyclonal DAKO Ab reacts with entire wt capsids but only weakly with denaturated core proteins (<1% [18]), (2) monoclonal FAb3105 which was shown previously to bind to an epitope on core protein dimers involving the immune dominant loop (aa 77–80 and 83–84 [33]).
Determination of the immune reactivity of peaks A, B and C by dot blot is shown on Fig. 4A. The graph shows the reactivity, normalized to the protein amounts. FAb3105 reacted similarly well with all fractions giving evidence that they contained core proteins which were at least assembled to dimers. The polyclonal DAKO Ab reacted best with fraction B and somewhat less with fraction A containing the capsid. It must be considered however that the capsid peak contains the encapsidated heterogeneous E. coli RNA, which leads to an overestimation of the core protein amount. Reactivity with fraction C was faint. Given that peak C contained core proteins already assembled to dimers we conclude that the DAKO Ab recognizes preferably epitopes that are formed on capsid subassemblies with a higher complexity.
The interference of spike insertions with DAKO Ab binding implies binding at or close to the spikes, similar to FAb3105. We thus analyzed the antibody competition for their epitopes on rC. Fig. 4B shows that preincubation with DAKO Ab completely inhibited FAb3105 binding, suggesting overlapping binding sites. Saturation with FAb3105 prior to DAKO Ab incubation reduced DAKO Ab binding but not entirely, potentially due to its polyclonality.
In order to analyze the fate of Mat-C upon nuclear entry we analyzed the sedimentation in Nycodenz after subjecting P-Mat-C to nuclear import. Import reaction was performed using digitonin-permeabilized cells, which is a well-established system for analysis of nucleo-cytoplasmic traveling and HBV genome release at the nuclear envelope [19],[34]. A control reaction was performed by addition of WGA, which blocks active nuclear import by nuclear transport receptors [35]. Following import reaction, nuclei were lysed by the same non-ionic detergents used for capsid purification of Mat-C from secreted virus. A similar protocol was chosen to exclude that nuclear lysis has an impact on the subsequent capsid analysis on Nycodenz gradients.
Upon WGA inhibition, 97% of P-core proteins of the P-Mat-C migrated as purified P-Mat-C without cell exposure, exhibiting a peak between 1.283 and 1,252 g/ml (Fig. 5A, peak (2)) with a maximum at 1.267 g/ml. It has been shown that P-Mat-C from transport reaction is attached to the nuclear import receptors Importin α and β [18]. Thus, one might expect a higher density. However, precipitating with the DAKO Ab capsids that were preincubated with the cytosolic extract showed, by immune blot, that the number of coprecipitated import receptors is rather low (4 molecules per capsid, not shown). This accounts for an undetectable increase of protein mass of <13% only.
Core proteins derived from P-Mat-C following nuclear import peaked between 1.304 and 1.250 g/ml with a maximum at 1.285 g/ml (Fig. 5B, peak (1)). This maximum was identical to the one of the purified RNA-containing P-rC (1.283 g/ml). As 99% of 32P core proteins were found within these fractions we concluded that all core proteins were assembled into particles without exhibiting significant amounts of dimers or subassemblies. A summary of all capsid sedimentation profiles is given in Fig. 5C. The requirement for an active nuclear import suggested however that the transition from “light” to “heavy” capsids occurred inside the nucleus.
The changed density of the DNA-filled P-Mat-C to the density of RNA-filled E. coli-expressed capsids upon nuclear transport implied that the DNA genome in Mat-C was replaced by RNA. Such a replacement likely involves at least partial capsid disintegration, although transitory capsid subassemblies could not be detected in our assay. Reduction of temperature or modification of salt concentration - successfully used in biophysical assays [1] - could not be applied as physiological nuclear import is temperature and salt dependent. We used an alternative approach based on the observation that HBV capsid assembly rate is enhanced by core protein RNA-interaction [36]. We treated digitonin-permeabilized cells by RNase A, which is a small 13.7 kDa protein that is far below the threshold of diffusion into the nucleus (up to 68 kDa [37]). RNA degradation was monitored by ethidium bromide staining after agarose gel electrophoresis of the cell lysate (Fig. 6A). Specificity of degradation was shown by control digestions using DNase I (37 kDa). DNase I treatment however causes collapse of the nuclear structure with diffuse distribution of the nuclear pores (not shown).
RNase A-treated cells were used for nuclear import of P-Mat-C. Applying the nuclear lysate onto Nycodenz gradient showed two peaks, one with densities from 1.241 to 1.300 g/ml, with a maximum at 1.285 g/ml (Fig. 6B, peak (1a)) and a second with densities from 1.141 to 1.185 g/ml, with a maximum at 1.156 g/ml (Fig. 6B, peak (1b)). Quantification showed that 47% of the 32P core protein migrated in the first peak while 53% were found in the lighter fractions. The heavier peak showed the same peak density of P-Mat-C after nuclear import but with a broader distribution. The lighter one showed the same distribution as urea treated core proteins.
A control was performed by adding WGA during the import reaction. Here, core proteins migrated with the density of P-Mat-C, ranging from 1.242 to 1.282 g/ml (maximum 1.266 g/ml; Fig. 4B, peak (2)), showing that RNase A has no impact on P-Mat-C density prior to nuclear import.
To confirm the presence of capsids, dimers and capsid subassemblies, we analyzed capsids derived from nuclear import by size exclusion chromatography. Fig. 7A (upper panel) shows the elution profile at OD280. The peaks were derived from cellular proteins, as the same profiles were obtained from permeabilized cells without P-Mat-C and from RNase A-treated cells. We analyzed the presence of 32P capsids by native agarose gel electrophoresis (Fig. 7A, lower panels). In lysate from cells not treated with RNase A, 32P signals were obtained in fractions A3–5 migrating as integer rC capsids. 32P-core proteins from RNase A-treated cells, in contrast, were observed in fractions A5–7 and exhibited a migration slower and more diffuse than rC. No core proteins were observed in fractions B9 and C2 in which dimers and more complex capsid subassemblies were found. It must be considered however, that in native agarose gel electrophoresis unassembled core proteins migrate diffusely. Given that the signals were already at the detection limit, diffuse bands could have caused false negative results.
We thus analyzed the fractions derived from the RNase A-treated cells by dot blot (Fig. 7B, upper row), revealing 32P label in fractions C1/2 by phosphoimaging. Immune reaction of the dot blot by FAb3105 confirmed the presence of core proteins in all the 32P-positive fractions (Fig. 7B, middle panels). These fractions however were not reactive with DAKO Ab, which showed a faint signal only in fraction A8. As it overlaps with the peaks of cellular proteins, we assume that it results from unspecific interactions, which is in accordance with the absence of any FAb3105 or 32P label in this fraction.
To confirm the immune reactivity of the capsids, dimers and other capsid subassemblies, we investigated the light and the heavy fractions derived from Nycodenz gradient centrifugation by immune precipitation. Heavy fractions (2–5) from P-Mat-C exposed to untreated cells (1.250–1.304 g/ml) and heavy fractions 3–5, derived from RNase A-treated cells in which nuclear import was blocked by WGA (1.242 to 1.282 g/ml) served as controls. From the RNase A-treated sample, we analyzed the heavy fractions 2–5 (1.241 to 1.300 g/ml) and the light fractions of the same gradient (1.141–1.185 g/ml). Fig. 8 depicts that FAb3105 precipitated the 32P-core proteins in all heavy and light fractions from all gradients with similar efficiency. In contrast, DAKO Ab precipitated 32P-core proteins only from heavy fractions of P-Mat-C from untreated cells and from RNase A-treated cells in which nuclear import was blocked by WGA. No precipitate was found in the heavy and light fractions from RNase A-treated cells in which nuclear import occurred.
Gradient analyses and size exclusion chromatography indicated that P-Mat-C is subject to an RNA-dependent intranuclear reassembly process upon nuclear import. For confirmation we analyzed the intranuclear appearance of capsids and their subassemblies in individual cells. Due to the lack of a suitable infection system we used Digitonin-permeabilized cells, which are a wide-spread technique for studying nuclear import including HBV capsids [18],[19],[34]. Digitonin permeabilizes the plasma membrane leaving the nuclear and ER membrane integer. After addition of exogenous cargos and the addition of nuclear import receptors the fate of the cargo can be analyzed using microscopy. We added Mat-C and rabbit reticulocyte lysate, which is common source of transport receptors [34]. The localization of capsids and capsid subassemblies was determined by confocal laser scan microscopy using indirect immune fluorescence. A control stain was performed with propidium iodide (PI) in order to depict cell nuclei.
Mat-C were added to permeabilized cells that were either untreated or RNase A-treated. Controls were performed by inhibiting nuclear import using WGA and by using cells to which no capsids were added. Fig. 9 shows that both DAKO Ab and FAb3105 exhibited nuclear fluorescence after addition of Mat-C to permeabilized cells which had not been treated with RNase A. This staining pattern is in accordance with data reported previously in permeabilized and integer cells [17],[19]. It is also in agreement with the typical findings in liver histology [38]. The signals of both antibodies were specific, as no fluorescences were observed in cells to which no Mat-C was added, or in which nuclear import was blocked by WGA [18]. In RNase A-treated cells however, no immune fluorescence could be observed using DAKO Ab, but strong signals were obtained using FAb3105. These findings are in accordance with the observation that capsid subassemblies present in the lysates from RNase A-treated cells were not recognized by DAKO Ab but recognized by FAb3105.
FAb3105 staining in RNase A-treated cells was significantly stronger than in cells untreated with RNase A. As both antibodies were added together this observation is in agreement with the competition for binding to capsids.
Several studies describe the in vitro assembly of viral capsids, but there are only few investigations targeting their disassembly or their intracellular fate. In vivo investigations can hardly yield capsids in amounts suitable for biochemical analysis. We thus used two anti capsid antibodies and characterized at first their binding specificity for capsid subassemblies. Both antibodies, raised against entire capsids, are conformation-dependent. We used the well-studied FAb3105 in order to comparatively characterize the DAKO Ab for its reactivity against different capsid subassemblies, obtained by separation on size exclusion columns. Comparison was required, as the calibration of the column with globular proteins (see Material and Methods) showed that the migration of the subassemblies did not correspond to the MW of a single core protein (21.5 kDa) neither to a core protein multimer, so that a form-dependent retardation of the core proteins or unspecific interactions with the matrix were assumed. Both subassemblies (B, C) reacted with the FAb3105 and thus represented dimers or larger multimeric association of core proteins. Based on the known kinetics of HBV capsid assembly, we assume that peak C exhibits core protein dimers and peak B corresponds to a larger assembly state. Our experiments do not allow drawing a conclusion on how many dimers form this complex. According to the literature trimers of the dimers could be present in peak B, as this assembly state was shown to be the only distinct capsid subassembly population apart of dimers and capsids [10]. The DAKO Ab reacted with the capsid subassembly in peak B but poorly with the dimer fraction (C). In fact, the faint signal obtained after blotting of this fraction can be well explained by the limited formation of capsid subassemblies larger than dimers occurring between harvest from the column and blotting. We assume that the DAKO Ab either reacts with epitopes at the interface of dimers or, alternatively, the larger subassemblies may exhibit conformational differences compared to free dimers, as it was suggested previously [1].
The antibodies competed for their substrate implying that they both attach at, or near, the capsid spikes. This was not surprising, as the loop comprises the immune dominant c/e1 epitope of the capsids. In summary, the antibody characterizations indicate that DAKO Ab requires core protein assemblies larger than dimers, and that both antibodies exhibited sterical interference. Considering the kinetics of in vitro assembly [10], it must be assumed that the DAKO Ab reacts with core protein hexamers and fully assembled capsids.
Anti capsid antibodies do not detect small amounts of core protein monomers. For following the migration of all states of capsids assembly on gradient we depended on sensitive and quantitative detection. We took advantage of capsid phosphorylation, which allows labelling of the core proteins by radioactivity. Phosphorylation of the capsids did not significantly affect capsid structure, as indicated by their unaltered migration on native agarose gel electrophoresis and their unchanged reactivity with the DAKO Ab (not shown). This result was expected as the number of transferred phosphates was low. In addition, data from others have shown that E. coli-expressed capsids are identical to liver-expressed capsids within a resolution limit of 30 Å [39]. Recently these data were confirmed with better resolution of 16 Å further showing that no gross structural changes are linked with genome maturation and envelopment [40]. Better resolution with 10 Å however showed that a hydrophobic pocket is present on DNA-containing capsids [41].
As these differences do not affect capsids size, the different densities of Mat-C and rC in Nycodenz gradients suggest that the nucleic acid content caused the difference. This assumption is in accordance with the reported densities of RNA and dsDNA in Nycodenz [32]. Entry of Nycodenz in the capsid lumen was proven by electron microscopy, and could have occurred either via the 1.5 nm-measuring holes in the capsid shell [12], or by capsid breathing.
For analyses of the intracellular fate of HBV capsids, we investigated the dissociation reaction of HBV capsids using digitonin-permeabilized cells. Addition of P-Mat-C to permeabilized cells caused a change of migration towards the density of RNA containing capsids, but only upon active nuclear import. Although the differences appear to be minor, the superposition of the different gradients allows a clear differentiation of the sedimentation profiles (Fig. 3B).
Strikingly, P-Mat-C showed a single density peak after nuclear import, implying that virtually all capsids were converted, and that no significant subpopulations failed to deliver their encapsidated DNA into the nucleus. This high efficiency in our system suggests that it may reflect the situation in infected individuals where up to 80% of all HBV particles are infectious [20].
Genome release and subsequent replacement of the genome by RNA could hardly be explained by passage through holes in the threefold or quasi-threefold axis of the capsid shell [12]. It can be concluded from structural data on HBV capsids that even dissociation of one core protein hexamer upon capsid breathing would cause holes of only 4.3 – 5.7 nm. Further dissociation of the capsid is probably necessary for genome release. The required incubation period of the transport assay for obtaining detectable nuclear import (15 min) was however much longer than the short concentration-dependent association times of >50 s observed in vitro [29].
To observe the disassembly of capsids into subassemblies within nuclei we decelerated reassociation by removal of RNA. Both sedimentation and size exclusion chromatography showed two reaction products. The smaller product corresponded in size to urea treated C183S P-rC by sedimentation, and to peak C of the chromatography, thus likely representing core protein dimers. This assumption was confirmed by their reactivity with FAb3105 and their failure to bind to the DAKO Ab following both methods of separation.
The second product migrated similarly to RNA filled capsids, but exhibited a slightly broader peak. Size chromatography confirmed a similar but nevertheless different migration than capsids, as the peak was shifted by two fractions. Further confirmation was obtained by native agarose gel electrophoresis, after chromatography, showing a diffuse and retarded migration compared to capsids. Such a migration was reported for P-rC upon acid denaturation [18] causing protein aggregation. Immune blotting confirmed the absence of DAKO Ab reactivity arguing against capsid or core protein hexamer formation similar to that observed after acid denaturation. The remaining reactivity with FAb3105 indicated however core protein dimer formation, so that we conclude that the oligomeric reaction product consists of core protein dimers which were assembled in a misfolded manner. In fact, misdirected folding under preservation of intact dimer formation was observed recently upon addition of the HBV capsid assembly inhibitor HAP1 [42].
To confirm the intranuclear localization of the dissociation and reassociation events, we analyzed capsids and their subassemblies by indirect immune fluorescence in permeabilized cells. In accordance with gradient centrifugation and size exclusion chromatography, we observed that RNase A-treatment generated intranuclear subassemblies, which were reacting with FAb3105 but not with the DAKO Ab. Despite the absence of capsids, the nuclear presence of core protein dimers indicates that RNase A treatment did not interfere with nuclear import of core protein. This supports the conclusion that HBV capsids become imported into the nuclear basket as entire particles [43] but disassemble in the nuclear environment.
In the presence of RNA, in contrast, DAKO Ab and FAb3105-reactive assembly forms of the core protein appeared. Considering the absence of detectable capsid subassemblies between complete capsids and core protein dimers upon in vitro disassembly [30], this finding implies that the capsids disintegrate to core protein dimers followed by a rapid RNA-dependent reassociation to capsids, which is misdirected in the absence of RNA. As we used unphosphorylated Mat-C in this assay, these results further exclude that phosphorylation of Mat-C has had a significant impact on formation of capsids and on the capsid subassemblies.
In the present paper we showed that the dissociation of the HBV capsid follows the in vitro association reaction in inverse direction. Our observation that core proteins reassemble to capsids inside the nucleus implies that both compartments - the cytoplasm in which initial capsid formation occurs - and the nucleus, support the assembly reaction. The environment in which disassembly occurs should be consequently before capsid entry into the free karyoplasm. A potential candidate compartment would be the nuclear basket at the karyoplasmic side of the NPC, as it provides the unique proteins of the NPC. We thus hypothesize that the capsids assemble from core proteins via dimer and hexamer formation, as it was proposed recently [10],[29] (Fig. 10). The time at which the polymerase-RNA pregenome complex interacts with the capsid subassemblies has to be left open but may enhance the assembly reaction. During genome maturation and transport, the capsid is subject to capsid breathing and remains stable [17],[19],[43]. Apparently, the basket of the nuclear pore comprises the environment promoting capsid disintegration and genome liberation, which occurs after arrest of the capsid [19],[43]. We assume that core protein dimers derived from disassembled capsids are however able to diffuse deeper into the nucleus. When the threshold concentration of nuclear core protein dimers [36] is reached, capsid formation would occur, which is probably enhanced by cellular RNA. As cytoplasmic capsids, they undergo breathing but remain stable, explaining the capsid accumulation observed in liver biopsies.
Mat-C were prepared and purified from virions of the permanently Hepatitis B virion-expressing hepatoma cell line HepG2.2.15 [44] accordingly to Rabe et al [19]. The capsids of these virions, which were shown to be infectious in chimpanzees [45], contain viral DNA in a relaxed circular form. [44]. E. coli-derived capsids (rC) and a mutant in which the C terminal Cys was replaced by Ser (C183S rC) were expressed and prepared as described previously [12]. Electron microscopy did not show any difference between wt capsids and this mutant (not shown). While Mat-C showed strong contamination of proteins of cell culture medium (approx. 30 fold) both E. coli-expressed capsid preparations exhibited high purity of >90% with respect to the total protein: SYPRO staining after SDS PAGE showed a single band of 21.5 kDa when 400 ng were loaded. Disintegration of capsids was achieved by treating C183S rC by 4 M urea for 10 min at 42°C.
All capsid preparations were analyzed for nuclease contaminations by incubating 50 ng capsids with 32P labelled nucleic acids in 50 µl transport buffer for 2 hours at 37°C in siliconized Eppendorf tubes and determining the amount of radioactive TCA precipitable material at 0 and 120 min. Five µl were removed and spotted in duplicates on Whatman 3 M filters, dried, then immersed in a beaker containing 300 ml 5% TCA and 1% PPi for 15 min on ice. Filters were rinsed 3 times 5 minutes with 300 ml 5% TCA 1% PPi, then immerged 1 minute in 70% ethanol, dried and finally counted in a liquid scintillation counter. Radiolabelled DNA was obtained by random priming of a 700 nt PCR product, whereas radiolabelled RNA was obtained by T7 transcription of a linearized plasmid containing a cloned gene downstream a T7 promoter. Radiolabelled DNA and RNA were separated from free unincorporated nucleotide either by spin column or by ethanol precipitation in the presence of ammonium acetate. Single stranded DNA was obtained by 5 min heat denaturation of DNA at 100°C followed by rapid chilling in ice.
Mat-C was labelled by addition of [γ32P]ATP using the activity of the in vivo incorporated cellular protein kinase (P-Mat-C) as described previously [19] and resulted in the transfer of 2–4 P/capsid (T4) which corresponds to 0.008–0.017 P/core protein P. rC and C183S rC were labelled in vitro by PKC according to Kann et al. [26] (P-rC and P-C183S). In brief this phosphorylation requires partial disintegration of capsid structure by low salt treatment, followed by phosphorylation. Capsids from the stock solution were at first diluted 1∶20 in water and incubated for 15 min at 40°C. 5 µg of the diluted capsids were preincubated in PKC buffer (20 mM HEPES pH 7.4; 10 mM MgCl2; 1.7 mM CaCl2; 1 mM DTT) together with 3 µg phosphatidylserine for 10 min at 42°C. Then 15 ng protein kinase C (Promega) and 10 µCi [γ32P]ATP (3000 Ci/mmol) were added (5 µl final volume) and the kinase reaction let to proceed for 30 min at 37°C, after which 0.5 µl PBS 10X was added. In contrast to the previously reported protocol we did not performed an RNase A digest, resulting in a 50fold reduced labelling of about 0.01 P/core protein. After phosphorylation the capsids were reconstituted by adjusting the salt concentration to physiological concentrations. As we have shown previously by EM these capsids exhibit the same structure than original capsids [26].
For separation by native agarose gel electrophoresis, capsids were loaded on the gel using sample buffer without SDS. Electrophoresis was performed on 1% agarose/TAE using a TAE buffer. Blot onto PVDF membranes was performed according to Southern [46]. Quantification of radioactivity was performed by phosphoimaging using either Typhoon 9200 (Amersham Biosciences) or Pharos FX (BioRad). Quantification of immune blots was performed by ECL (PerkinElmer) using ChemiDoc XRS (BioRad).
For analysis of capsid integrity during transport on Nycodenz gradients, or on Superdex 75 PC 3.2/30 columns, 4×106 HuH-7 cells were permeabilized by digitonin as described previously [19], with the modification that the permeabilized cells were harvested after the washing steps prior to the transport reaction. The transport was performed as described, but using a volume of 100 µl comprising 50 ng P-Mat-C and rabbit reticulocyte lysate. If required, 100 µg/ml of wheat germ agglutinin (WGA) was added during the washing steps. After incubation, cells were lysed using 1% NP-40/PBS/5 mM MgCl2 for 1 h at 37°C. After lysis, Triton X-100 was added to a final concentration of 0.2%, followed by 10 min incubation at 37°C and subsequent sonification (6×15 sec). The sample of 200 µl was then subjected to analysis. RNase A-treatment was performed for 15 min at 37°C at a final concentration of 4 µg/µl.
Transport assays for immune detection, immune staining and subsequent confocal laser scan microscopy were done accordingly to Rabe et al. [19]. Immune staining was performed with the polyclonal rabbit anti capsid antibody (1∶200, DAKO Ab) and with the monoclonal mouse anti capsid protein antibody (1∶200, Fab3105, Institute of Immunology CO., LTD, Tokyo, Japan). Cy5-labelled anti rabbit antibody (1∶400, Dianova) and FITC-labelled anti mouse antibody (1∶200, Dianova) served as secondary antibodies. Nuclei were stained with propidium iodide (1∶5000).
Two hundred µl of samples were added onto 3.6 ml continuous Nycodenz/TN gradient (1.11–1.32 g/ml) in a SW60 rotor. Centrifugation was done for 22 h at 10°C and 36,000 rpm. Fractions of 220 µl were harvested. Density d was determined by refractometry (σ) using the formula: d = (σ×3,287) – 3,383. The fractions were vortexed prior to density determination.
Capsid preparation was centrifuged for 5 min at 17 000 g before being processed through the size-exclusion column. The apparent molecular size of the proteins was analyzed by chromatography on a Superdex 75 PC 3.2/30 column (GE Healthcare), which has a fractionation range of 3 to 70 kDa. The column was equilibrated with transport buffer (20 mM HEPES pH 7.3, 2 mM Magnesium acetate, 110 mM Potassium acetate, 5 mM Sodium acetate, 1 mM EGTA and 1 mM DTT). Proteins (50 µl) were eluted with a flow rate of 40 µl/min and recorded by continuously monitoring the absorbance at 280 nm. The column was calibrated with the following standard proteins: ovalbumin (43 kDa; 1.13 ml)), chymotrypsinogen A (25 kDa; 1.26 ml), RNase A (13.7 kDa; 1.36 ml) and the void volume was determined with dextran blue (>2000 kDa).
Quantification of the capsids were done by immune blot accordingly to Rabe et al. (11) using either the polyclonal rabbit anti capsid antibody (1∶5000, DAKO Ab) or the monoclonal mouse anti capsid protein antibody (1∶2500, Fab3105). As secondary antibody, a horse radish peroxidase anti rabbit or anti mouse antibody (1∶10000, Dianova) were used. Detection was performed by ECL (PerkinElmer) using ChemiDoc XRS (BioRad). For immune precipitations 3.5×106 sheep anti rabbit- or anti mouse-conjugated biomagnetic beads (Dynal) were added to 22 µg DAKO Ab or mouse anti capsid protein antibody (Fab3105) and incubated in 0.1% BSA/PBS overnight at 4° C on a rotating wheel. Afterwards unbound antibodies were removed by washing the beads three times with 0.1% BSA/PBS. The antibody saturated beads were subjected to light and heavy capsid fractions from Nycodenz gradients in the presence of 0.1% BSA and incubated overnight at 37°C on a rotating wheel. The precipitate was washed three times in 0.1% BSA/PBS, one time in 0.1% Nonidet P-40/PBS, transferred into a new cup and again washed three times with PBS. The samples were loaded onto a 4–12% Bis-Tris gradient gel SDS-PAGE (NuPAGE) and transferred onto a PVDF (VWR International) by electro blotting. Precipitated capsid proteins were detected by their intrinsic radioactive signals using phosphoimaging.
|
10.1371/journal.pcbi.1002268 | Dominant Glint Based Prey Localization in Horseshoe Bats: A Possible Strategy for Noise Rejection | Rhinolophidae or Horseshoe bats emit long and narrowband calls. Fluttering insect prey generates echoes in which amplitude and frequency shifts are present, i.e. glints. These glints are reliable cues about the presence of prey and also encode certain properties of the prey. In this paper, we propose that these glints, i.e. the dominant glints, are also reliable signals upon which to base prey localization. In contrast to the spectral cues used by many other bats, the localization cues in Rhinolophidae are most likely provided by self-induced amplitude modulations generated by pinnae movement. Amplitude variations in the echo not introduced by the moving pinnae can be considered as noise interfering with the localization process. The amplitude of the dominant glints is very stable. Therefore, these parts of the echoes contain very little noise. However, using only the dominant glints potentially comes at a cost. Depending on the flutter rate of the insect, a limited number of dominant glints will be present in each echo giving the bat a limited number of sample points on which to base localization. We evaluate the feasibility of a strategy under which Rhinolophidae use only dominant glints. We use a computational model of the echolocation task faced by Rhinolophidae. Our model includes the spatial filtering of the echoes by the morphology of the sonar apparatus of Rhinolophus rouxii as well as the amplitude modulations introduced by pinnae movements. Using this model, we evaluate whether the dominant glints provide Rhinolophidae with enough information to perform localization. Our simulations show that Rhinolophidae can use dominant glints in the echoes as carriers for self-induced amplitude modulations serving as localization cues. In particular, it is shown that the reduction in noise achieved by using only the dominant glints outweighs the information loss that occurs by sampling the echo.
| Rhinolophidae are echolocating bats that hunt among vegetation. The foliage returns clutter echoes that potentially mask the echoes of insect prey. However, prey introduces frequency and amplitude shifts, called glints, into the echo to which these bats are highly sensitive. Therefore, these glints are used by Rhinolophidae to detect prey and infer its properties. One of the defining characteristic of consecutive dominant glints is that they have a very stable amplitude. This is, consecutive wing beats of the insect produce dominant glints with more of less the same amplitude. Owing to the strategy Rhinolophidae use to locate prey, the stable amplitude of glints makes these parts of the echoes ideal signals to use to locate prey. In this paper, we demonstrate the feasibility of strategy under which Rhinolophidae use only the dominant glints in the echo for locating prey.
| Rhinolophidae or Horseshoe bats, a family of echolocating bats, hunt for fluttering insects amongst vegetation [1], [2]. This implies that, with each call, they receive a large number of echoes most of which originate from foliage. They have evolved an echolocation system that allows detecting prey under these difficult circumstances by encoding the presence and the properties of prey by frequency and amplitude modulations in the returning echo (reviewed in [3]).
Rhinolophidae emit long narrowband pulses where most energy is contained in a single and well-controlled frequency component. Fluttering prey introduces frequency and amplitude modulations into the returning echo called glints [4]–[7]. Glints reliably signal the presence of prey to the bat. Indeed, Rhinolophidae only pursue insect prey that introduces glints in the echoes [3], [7], [8]. In addition to merely signalling the presence of prey, it has been argued that the glints provide Rhinolophidae with cues about a number of prey properties (reviewed in refs. [3], [7], [9]). The prey property encoded in the glints that is best studied is the wing beat frequency. The wing beat frequency of an insect can readily be inferred from an echo by counting the glints. In experiments, Rhinolophidae were able to discriminate between targets fluttering at different rates, e.g. [5], [7], [8].
For the localization of echoes in azimuth and elevation bats using broadband calls depend on spectral cues created by the transfer function of the outer ears (e.g. [10]–[12]). The use of a narrow frequency band to perform echolocation prevents Rhinolophidae from using spectral cues to localize reflectors in space [13]. To overcome this, they vigorously move their ears while echolocating [13]–[17]. The movement of the pinnae (which they coordinate with the reception of the echo) imposes amplitude modulations upon the incoming echo. The exact modulation patterns depend on the reflector position (azimuth and elevation). As it has been shown that these amplitude cues provide stable localization information [18], [19], it is assumed that this cue is also used by the bat to estimate the origin of the echo [13]. Indeed, when Rhinolophidae are prevented from moving their pinnae, their ability to locate obstacles deteriorates [14], [20].
In the current paper, we present simulations showing that glints carry the most reliable information for prey localization (as has been shown to be the case for prey classification,). Only fluttering insect prey produces glints. Therefore, the amplitude of glints is not influenced by interfering echoes from the background vegetation. In contrast, the carrier frequency band will contain a number of overlapping echoes from foliage resulting in spurious amplitude variations that are not due to pinna movement [18]. By only processing the glints when determining the location of prey, the bat could effectively reduce the influence of clutter on the localization cues. Fluttering prey not only introduces frequency shifts into the echo but also considerable amplitude variations [4], . However, the amplitude of the dominant glint is rather stable. Dominant glints can be defined as the most Doppler shifted parts in the echo and are produced at the instant the insects wings are perpendicular to the impinging sound waves [4]. In ensonification experiments, the amplitude of the dominant glint was found to have a standard deviation of less than [4] while the amplitude of the echo across its entire duration can fluctuate by up to [3]. In sum, for localization, Rhinolophidae could substantially reduce the noise (i.e. unknown amplitude variations) by focussing on the dominant glints. This would reduce both the interference by echoes from foliage and stabilize the glint amplitude.
Focussing on the dominant glint potentially comes at a cost. Depending on the flutter rate of a target, only a limited number of dominant glints will be present in each echo. By only processing these, the bat would effectively use a sampled version of the echo where information is only available at discrete times of dominant glints caused by the wing beat. This process is illustrated in figure 1. Using a sampled version of the echo, potentially reduces the amount of localization information generated by the moving ears that is transferred to higher auditory centres. Indeed, unless the information generated by the amplitude modulations is robust against being sampled at a low rate (given by the dominant glint rate), the clutter rejection mechanism would pose a limit to the echolocation capacity of the animal.
In this paper, we use a computational model of the echolocation task faced by Rhinolophidae to investigate the feasibility of a localization mechanism that is based on processing the dominant glints. We test whether the localization cues introduced by the moving pinnae are robust against sampling. We do this by evaluating the information transfer in R. rouxii for a range of simulated flutter rates. We hypothesize that, the glint based localization mechanism would be feasible only if the information transfer is not hindered considerably when localization is based on the information carried by the dominant glints.
In addition, we compare the information transfer in two qualitatively different frequency channels available to the bat. In a first alternative, we analyse the information content of the response from a cochlear channel sensitive to frequencies close to the resting frequency. Such a channel produces a non-zero response throughout the duration of the echo. However, the expected response pattern, i.e. the one corresponding with the echo strength modulation due to ear movement, is disturbed by an additional unknown amplitude modulation pattern due to the fluttering prey and clutter echoes. In the other alternative, the information content of the response of a second type of channel (a Doppler shifted frequency channel) is analysed. This cochlear channel is only stimulated when large frequency shifts are introduced into the echo, i.e. during the dominant glints. We hypothesise, that using the Doppler shifted frequency channels in locating the prey will only be adaptive if its advantages (i.e. noise reduction) outweighs its potential disadvantages (i.e. information loss due to sampling).
The calls of Rhinolophidae are often preceded by a short upward sweep and/or followed by a short downward sweep. However, we only consider the constant frequency (CF) component of the calls of R. rouxii in our analysis. The limited bandwidth and relatively small energy in the frequency modulated (FM) component of their call has been taken to indicate that Rhinolophidae rely less on the spectral cues that are used for localization by bats emitting broadband calls [21]–[24]. Moreover, R. rouxii has been observed to omit the FM component in 90 percent of its calls while hanging from a perch and scanning the surroundings for flying insect prey [1], 25.
While the hearing directionality of R. rouxii has been measured [13], this is not the case for the emission directionality. However, simulation methods have become available that allow the evaluation of the directionality of the echolocation system of bats at a high resolution [26]–[31]. Among these simulation methods, Boundary Element Methods (BEM) are well suited to simulate both the emission and hearing directionality of bats [29]. Furthermore, BEM is thus far the only simulation method that has been formally validated for the simulation of HRTFs of small mammals (for bats [18], [29] and for gerbils [32]).
Using BEM to simulate the directionality of a bat requires a 3D model of the morphology of the head of the species under study. In our lab, we have developed a method to create such a model from CT data [29]. The 3D model of R. rouxii used in this study is rendered in figure 2. To obtain this model, a single specimen of R. rouxii (origin: Sri Lanka [13]) was scanned using a MicroCT machine with a resolution of m. Using standard biomedical software and the method described in ref. [29], a 3D model of the morphology was extracted from the data (see ref. [33] for more details on the extraction of the current model).
On our current hardware, the software [26], [27] used to simulate the emission beam and the hearing directionality can only handle models consisting of up to 30,000 triangles. Therefore, we constructed a separate model of the noseleaf to ensure capturing all important features of the baroque facial morphology of R. rouxii. The complete head model and the model of the noseleaf are depicted in figure 2. As the resting frequency used by R. rouxii lies typically around 75 kHz (73–79 kHz; [1]) we use the simulated emission pattern and hearing directionality pattern at this frequency in the current paper.
Figure 2 shows the simulated hearing and emission directionality for the 3D model at 75 kHz. The simulated hearing directionality corresponds well with that reported in ref. [13]. Moreover, the match between the simulated hearing directionality of R. rouxii and the measured hearing directionality [13] was quantified in ref. [18] for a range of frequencies. As reported in ref. [18], we found a good agreement between the simulations and the measurements.
R. rouxii typically moves one of its pinnae to the front and the other one backwards during the reception of an echo. In the closely related specimen Rhinolophus ferrumequinum, the motion of the pinnae describe an arc of about 30 degrees at an oblique angle [15]–[17]. This is, while moving to the front (back) the pinnae also move somewhat inwards (outwards). No accurate description of the motion in R. rouxii is available. Therefore, we model the motion of the pinnae based on the reports on Rhinolophus ferrumequinum as moving from −15 degrees in elevation and −(+)15 degrees azimuth to +15 degrees in elevation and +(−)15 degrees azimuth for the right (left) ear. This is, as one ear moves down the other one moves up. In additional simulations, we confirmed that other arcs of movement influenced our results very little (see [18] for details).
Simulating the movement of the pinnae was done by assuming that this could be approximated by rigid rotations of the hearing directionality while keeping the emission directionality in the same position. In cats it has been shown that rigid rotations are a good approximation of changes to the hearing directionality due to pinnae movement [34]. Moreover, the extent over which the pinnae are moved in R. rouxii is rather small (about 30 degrees). Hence, we assume that the effects of the additional deformation of the pinnae on the combined emission-hearing directionality can be neglected in our analysis.
In this section of the paper, we outline our mathematical model of the echolocation task. This model has been adapted from refs. [12], [18] and is based on the Shannon Information Theory [35], [36].
The basic assumption underlying our model is that the localization of a target can be considered as a template matching task [12], [37], [38]. A fluttering insect produces an echo containing typical target-induced Doppler shifted glints that show up as frequency spreading in the spectrogram (see refs. [4], [5], [8] for examples of spectrograms obtained from measurements). The echo is picked up by the bat's moving pinnae. Based on reports in the literature, we assume, that each pinna moves either up or down during the reception of the echo [15]–[17]. Ear movement introduces additional amplitude modulations of the echo at both tympanic membranes. The exact way in which the echo is modulated depends on the augmented head related transfer function (AHRTF), i.e., the combination of the emission directionality and the HRTF, of the bat. Each different azimuth-elevation position of a target with respect to the bat corresponds to a different expected modulation pattern at the left and the right ear. These expected modulation patterns are termed templates in the remainder of the paper. We assume the bat compares any measurement with a set of stored templates to estimate the direction from which the echo originated.
As argued in the introduction, we assume that R. rouxii uses the dominant glints to perform localization. Therefore, the bat has access to a version of the expected modulation patterns that is sampled at the points in time at which the echo contains a dominant glint. We assume that the bat uses a number of samples taken at discrete points in time from the amplitude modulated signal produced by the moving ears. The number of samples depends on the flutter rate of the insect. This models a worst case scenario in which a fluttering insect introduces only one dominant glint per wingbeat and the bat does not use any other glints apart from the dominant glints. Being a worst case scenario implies that the evaluation of the information transfer in Rhinolophidae using this signal results in a lower estimate. Therefore, if our results show that the frequency channels picking up the dominant glints conserve localization information, this indicates that using these channels is certainly possible for Rhinolophidae hunting under realistic circumstances where the amount of information carried by all the glints is even higher (see Discussion).
Under these assumptions, we regularly sampled the expected modulation patterns at frequencies between 20 and 200 Hz. A realistic range of flutter rates for insects as reported by [4], [5] would be about 50 to 100 Hz (see also the Discussion). Extending this range downwards enables us to assess the extent of the information transfer at very low flutter rates. Moreover, we will use the results for a flutter rate of 200 Hz as a baseline to which we compare the results for lower flutter rates. As we assume that R. rouxii uses calls with a duration of 50 ms [1], flutter rates of 20 to 200 Hz correspond to 1 to 10 dominant glints (samples) for each ear. The point in time of the first sample was uniformly distributed between 0 and 0.5 sampling periods. While the flutter rate of insects is very stable [9], some deviation from regularly spaced sampling are likely to occur. To investigate whether our results also hold when we no longer assume regularly spaced glints in the echo, we ran simulations in which the samples were randomly spaced over time.
The sampled versions of the expected amplitude modulation pattern at the left and the right ear are concatenated into a single vector containing all measurements.
Using the same measurement noise model as proposed in [12], the received amplitudes are assumed to be corrupted both by the unknown and varying reflector strength as well as the system noise. Their different effects on the received amplitudes follow naturally if we represent the received echo amplitudes on a logarithmic scale (in ), i.e., apply a compression very similar to the one performed by the hearing system. System noise is additive but, because of the logarithmic compression, its effect on the received amplitudes can be approximated by a maximum operator as,(1)(2)with the template, i.e., the expected amplitude modulation at the different pinna positions (scaled such that ), stored by the bat for reflector position . The noise level, i.e., the lower threshold below which no signal can be detected, is set at . The vector denotes the unknown and varying echo strength modulation due to the fluttering target. The term represents the mean echo strength averaged over the ear positions. As the noise level is set to zero the parameter can be interpreted to specify the signal to noise ratio of the echo.
The term represents normally distributed multivariate noise, i.e. (the meaning of is explained in the next paragraph). This noise term models the unknown amplitude modulations imposed onto the echo due to target movement (e.g., fluttering target).
Following Bayes' theorem, the posterior probability of a received vector of strength to originate from position can be written as given by equation 3(3)Taking into account that the expected value of , i.e., , depends on , the likelihood of a received vector given a reflector position and echo strength is calculated as,(4)with the total number of ear positions in the binaural template and(5)The covariance matrix gives the variances and covariances of the stochastic vector . However, the amplitude of the echo is unknown to the bat. Therefore, it is treated as a nuisance parameter in the model,(6)with the range of values that can occur. Hence, we rewrite equation 3 to arrive at,(7)
Equation 6 is calculated assuming that the bat considers all echo strengths in the interval equally likely and thus maintains a uniform prior across reflector strengths. Hence, we assume that the bat has no priori knowledge about the fraction of the impinging energy reflected by the target. Equation 7 gives the posterior distribution of . Using Shannon entropy, the uncertainty about the true target position when receiving a particular echo from position can be expressed in bits as,(8)
The quantity of direct behavioural relevance though is the average entropy carried by all possible echoes originating from position . To calculate this quantity one should average over all realizations of the reflector ensemble. is approximated using a Monte Carlo simulation. For each position , 20 realizations of the measurement vector are generated. For each of these realizations, equations 3 to 8 are evaluated and the average value is reported. Twenty realizations for each position were found to yield stable results.
Having introduced the model and the methods, we can summarize all relevant assumptions: (i) localization is considered as a template matching task, (ii) we assume that only upward frequency-shifted dominant glints within a single echo are evaluated and (iii) that the relative position prey with respect to the bat does not change appreciably while it is being ensonified, (iv) it is assumed that the HRTF does not change during pinna movement, but is only rigidly rotated, (v) the parts of the echo that were Doppler-shifted by insect wings are assumed to have more stable amplitude than the echoes of non-moving objects (vi) we assume that the head does not move during call emission and echo reception (vii) FM parts of the echoes are not considered (see also discussion).
As described in the previous section, the model only has a single parameter, the covariance matrix ,(9)with and denoting the sample for the left and the right ear.
To obtain estimates of the values of we use the value reported in ref. [4]. The authors report that the standard deviation of the amplitude of the dominant glints produced by fluttering echoes is or less (see figure 5 in ref. [4]). Based on these data, we use as a lower value for the diagonal of .
The variation of the amplitude of the dominant glints in an echo is markedly lower than the variation of the amplitude throughout the echo as the dominant glints are, by construction, synchronized with a specific point in the wingbeat cycle. Previously, we have reported on asynchronous ensonification measurements of fluttering targets from which we calculated the standard deviation of the amplitude in a narrowband frequency channel [18]. We found a value of about . Therefore, we also evaluate the model for . We use this value to model frequency channels that are stimulated for the entire duration of the echo signal (resulting in more noisy modulation pattern measurements) as opposed to the frequency channels that are stimulated by the dominant glints only.
Finally, in addition to and as values for the diagonal of , we also use as an intermediate value to evaluate how the information transfer deteriorates when moving from a noise level of .
The value of , were set to . This reflects the assumption that the noise is uncorrelated across samples. Previously we found that the model is not very sensitive to the values of and [18]. For similar reasons and with were set to as well. Finally, was set to reflecting the assumption that simultaneous amplitude measurements in the left and the right ear are highly (but not perfectly) correlated (see ref. [18]).
It should be noted that only describes the variations in the glint amplitudes within a single echo. Variations between consecutive calls are modelled as changes in the echo strength . Hence, if the insect returns weaker or stronger glints across consecutive calls, this amounts to variations in the signal to noise ratio under which the bat operates.
The ability of the model to match templates and measurements critically depends on the assumed echo strength or signal to noise ratio of the echo. In the lab, fixated R. rouxii were found to call with an amplitude of about (at 10 cm in front of the bat)[39]. R. rouxii hunts mostly for insects with a wing length smaller than 10 mm [40]. Fluttering insects of this size return an echo that is up to weaker than the impinging sound (depending on the frequencies used) [4]. Therefore, we evaluated the localization entropy predicted by the model for echoes ranging from in steps of as this contains all echo strengths likely to result from prey of interest to R. rouxii.
In the current numerical simulations, we use 3252 templates that code for as many azimuth and elevation positions uniformly distributed over the frontal hemisphere. It was found that using more templates increased the computation time but did not alter the results. Changing the number of sample points changes the results quantitatively, but not qualitatively.
The entropy, i.e., a measure of the remaining ambiguity, about the origin of an echo as function of azimuth and elevation for is plotted in figure 3. From this figure, it can be seen that, as the flutter rate increases, entropy quickly reaches a stable level. Increasing the flutter rate beyond 60 Hz does not reduce entropy significantly. It should also be noted that at 20 Hz, the lowest flutter rate simulated, the predicted echolocation entropy is already considerably lower than chance level (i.e. about 11 bits in the current simulations). In a central area, entropy goes down to a level of about 6 bits even for this low sampling frequency. Note that, at a flutter rate of 20 Hz, the model is only provided with 1 sample per ear to perform localization.
Figure 4 further explores the effect of flutter rate on localization entropy. Figure 4 confirms that entropy mostly depends on the echo strength. For every flutter rate the entropy decreases as echo strength increases. In contrast, entropy depends only little on flutter rate as long as the flutter rate is higher than about 50 Hz when (fig. 4b & c). Indeed, for these flutter rates the difference in entropy between the information transfer at a flutter rate of 200 Hz and a lower flutter rate is less than 1 bit. For higher noise levels, the overall entropy increases (fig. 4a, d & g). Moreover, the effect of flutter rate increases with increasing noise levels (fig. 4b, e & h). However, even for , the effect of flutter rate is mainly located in the region of flutter rates lower than 100 Hz.
The third column of plots in figure 4 confirms that the results also hold when the echo is sampled at random intervals (in contrast to fixed intervals). This indicates that our results are not sensitive to a deviation from regular spaced sampling.
Another way of summarizing the information loss due to sampling is given in the performance plots in figure 5. The performance measure plotted in this figure for a given flutter rate is calculated as follows,(10)with the entropy for a given flutter rate and echo strength (averaged across the frontal hemisphere). Flutter rate is the highest flutter rate simulated, i.e. 200 Hz. The parameters and give the number of echo strengths evaluated and the number of templates (i.e. 3252) respectively. Therefore, in plot 5, a performance of 100% is the entropy level for a flutter rate of 200 Hz and the plot shows the normalized average performance as a function of flutter rate for the three noise levels. It can be seen that 90% performance is reached for the three noise levels at flutter rates 37, 65 and 86 Hz respectively.
The results presented so far indicate that, for the noise levels typical for the dominant glints (i.e. about ), the localization entropy does not depend heavily on the flutter rate of the insect that is to be located. The results plotted in figures 4 and 5 indicate that R. rouxii looses little performance by sampling the echo even when the flutter rate is low. In addition, these results indicate that the robustness against sampling is higher for lower noise levels. The dependence on flutter rate increases gradually as the noise level rises.
In figure 6, the entropy for the two types of frequency channels described in the introduction are compared.
As can be seen in figure 6, the information content of the output of the frequency channel responding for the entire duration of the echo signal but suffering from a higher noise level is almost uniformly the lowest. Indeed, for almost every flutter rate and echo strength the entropy about the location of a target is higher for the ‘non-sampling’ frequency channel than for the ‘sampling’ frequency channel. Only for very low flutter rates and very high echo strengths are the roles reversed. This indicates that, although some information is lost due to sampling the echo signal, the noise reduction that is achieved by processing only the most Doppler-shifted parts of the dominant glints yields an almost universal increase in target location information.
In theory, there are two ways in which templates can be robust against sampling. First, templates could show a high degree of variation. By having templates that have a higher dynamic range, the Euclidean distance between templates increases and any loss in fidelity by sampling would cause less increase in localization entropy. Alternatively, templates could have most of their energy in the lower frequency components of the modulation spectrum. In this case, the spectrum of a template would only contain low frequency components. If templates would only vary slowly as the pinnae move through space, any sub-sampling would be less of a problem. It should be noted that these two strategies to design more robust templates are somewhat contradictory. Templates that have a larger dynamic range will usually contain higher frequencies.
In figure 7a & b, we plotted a histogram of the dynamic range of the templates of R. rouxii and an average spectrum of the templates respectively. In figure 7a, the dynamic range of the templates of R. rouxii is compared with those of Micronycteris microtis and Phyllostomus discolor. We have previously reported on the simulated HRTFs and emission patterns of these bats [30]. Moreover, we have provided an analysis of the localization information transfer of M. microtis [12]. In contrast to R. rouxii , both M. microtis and P. discolor emit short broadband calls and use spectral cues as means of localizing echoes in space. In these animals, as in most mammals, the major part of the localization information is provided by notches in the spectra generated by the filtering of the pinnae [12], [41]. Therefore, in contrast to R. rouxii which is assumed to use amplitude modulations of a narrowband signal, these bats mainly code the position of a target in space by means of spectral notches.
In figure 7a, it can be seen that the dynamic range of the templates of R. rouxii is not larger than that of the two other bats. Inspecting some examples of the templates of the three species (plotted in figures 7c–e) it can be seen that the templates of R. rouxii do not show the deep notches found in M. microtis and P. discolor.
The templates of R. rouxii , consist mostly of low frequency components (figure 7b). However, a major part of the energy is contained in frequency components for which the Nyquist criterion is not reached at typical flutter rates of insect prey. For example, targets fluttering at 100 Hz yield 5 glints in an echo of 50 ms. This only allows to faithfully reconstruct frequencies up to 2.5 Hz (see line in figure 7b). Stated differently, reconstructing the templates from the samples provided by a target that flutters at 50 Hz is only possible if the templates contained only frequencies below 1.25 Hz.
In sum, the templates of the echolocation system of R. rouxii do not seem to be particularly suited to be robust against sampling at the rates their prey flutters. Neither the dynamic range nor the spectra of the templates seem optimized for reconstruction from a small number of samples. Hence, we propose that the localization system of R. rouxii is robust against sampling of the templates only because it can effectively limit the noise by processing the dominant glints. Indeed, the results plotted in figure 4 and 5 indicate that good localization for low flutter rates is only attained if the noise level is low.
Our simulation results show that Rhinolophidae could reject unwanted amplitude variations (i.e. noise caused both by clutter and target movement) by processing dominant glints without substantially reducing the localization information transfer. Plots 1a–c show that almost no localization information is lost once the flutter rate is higher than about 50–60 Hz (for a noise level of ). Indeed, the performance curves in figure 5 show that a performance level of 90% is attained at a flutter rate of about 40 Hz.
Although some insects have flutter rates even below 20 Hz (the lowest flutter rate simulated) [42], a flutter rate of 40 Hz seems in the lower range of the flutter rates frequently encountered by these echolocating bats [4]. Since no data exists, as far as we could find, about the distribution of the flutter rates of insects R. rouxii preys on it is unknown what range of flutter rates is of behavioural importance to the bat. However, indirect evidence, i.e. cortical neurons that encode flutter rates show best phase locking for flutter rates between 40–60 Hz in R. ferrumequinum, seems to indicate that flutter rates in the range 40–60 Hz might indeed be relevant to R. rouxii . Moreover, many insects have flutter rates in this range (see [3] for references). Furthermore, by lengthening its emissions, a simple adaptive strategy Rhinolophidae makes use of when faced with a difficult echolocation task [1], R. rouxii could locate insects with lower flutter rates. Our simulations are based on a call duration of 50 ms. Doubling the length of the call would imply that the simulated flutter rates could be halved without altering the results. In addition, we have assumed that insects produce a single dominant glint per wing beat. However, depending on the wing structure, some insects produce more than a single dominant glint per wing beat cycle [4]. Insects that produce multiple dominant glints would provide the bat with more localization information and should therefore be easier to locate at lower flutter rates. The fact that Rhinolophidae can lengthen their call and that some insects produce multiple dominant glints increases the feasibility of using channels sensitive to the Doppler shifted dominant glints.
More important than the absolute information transfer for any channel is the comparison between the two types of channels proposed in the introduction. We compared the information transfer in both types of channels in figure 6. It was found that Doppler shifted frequency channels almost invariably outperform the channels responding to the centre frequency.
While our simulations show that Doppler shifted channels provide Rhinolophidae with better localization performance than the more noisy channels responding to the reference frequency, bats will have access to both types of channels while locating prey. Indeed, bats can support the information in the Doppler shifted frequency channels with information gathered by reference frequency channels. Therefore, our simulations yield a conservative, i.e. lower bound of the information transfer, and real bats likely use both the reference frequency as well as the frequency shifted parts of the echo, as would be expected from their sensory physiology.
Neurophysiology recordings in the cochlear nucleus suggest that Rhinolophidae posses neurons that can support the processing of self-induced modulations of the dominant glints in the echoes. Their cochlear nucleus contains a large proportion of neurons with a high degree of frequency tuning that respond only to the onset of the preferred frequency [43]. About 40% of these neurons were found to be insensitive to variations in intensity. These neurons would be well suited to detect dominant glints in the echoes. A tentative hypothesis about the implementation of the glint based localization proposed in the current paper could be as follows: neurons selective to frequency and with a phasic response continuously monitor Doppler shifted frequency channels. These onset-coding and intensity-insensitive neurons act as a clock pulse selecting samples from the continuous intensity signals coded by other intensity-sensitive neurons.
In addition, it should be noted that, although 40% of the onset-coding neurons in the cochlear nucleus were found to be insensitive to intensity [43] at least some neurons in the cochlear nucleus are capable of detecting both the onset and encoding the intensity (by means of prolonged firing [43]). Also, similar properties of sharp frequency tuning and amplitude modulation selectivity can be found in other auditory nuclei in rhinolophid bats, e.g. [44], [45]. This indicates that the localization mechanism proposed in this paper could be implemented at different levels in the auditory system of Rhinolophidae.
Finding that the localization information transfer in R. rouxii is robust against sampling, we analysed the templates in order to investigate whether these show any adaptations that support this robustness. However, we could find no evidence of the templates of R. rouxii showing adaptations to being sampled at the flutter rates of likely targets. On the contrary, on average the templates show less dynamic range than those used by M. microtis and P. discolor . Also, the sample rate dictated by the flutter rate of insects is not high enough to comply with the Nyquist criterion as the templates contain frequency components that are too high. Rather, it seems that the reduction in echo amplitude variability (i.e. noise) by focusing on dominant glints allows the bat to locate targets without such adaptations.
Analysing the localization information transfer of the FM bat M. microtis it was found that the spectral templates with the largest dynamic range encode peripheral positions. Notches in the spectral templates of this bat are mostly found for peripheral positions. However, these notches, being created by side lobes in the spatial sensitivity pattern of the system, lower the sensitivity of the system at these locations. Indeed, deep notches in a template denote a combination of a location and a frequency for which the system is insensitive. The effect of this on localization is that weak echoes can be located best in a central region where sensitivity is highest. However, for stronger echoes the region with the best localization is actually the periphery at locations coded by deep spectral notches as these templates are more resistant to unknown reflector filtering (noise), see figure 8 and [12]. Therefore, in the face of noise, these bats are confronted with a trade-off: for any given position the bat can either be highly sensitive or very accurate.
The switch in the region where echoes can be best located, is not observed in R. rouxii . Plotting the localization entropy as a function of echo strength (see figure 8) it is found that lowest entropy is always located in the central region and that this region of low entropy simply expands as the echo strength increases. By focussing on the dominant glints, the entropy in the central region is not increased by noise and the bat does not experience a trade-off. Indeed, figure 8 shows that avoiding the trade-off is only possible for low noise levels (). For a noise level of , entropy in the central region is also higher than for the peripheral region in R. rouxii . In this case, R. rouxii show the same trade-off as M. microtis and, because of the overall lower dynamic range of its templates, the trade-off is even more pronounced as can be seen from the larger contrast between its central and peripheral localization entropy (bottom row of figure 8). This figure also contains a visual analogue using the classic Lenna image [46] to further clarify the sensitivity-accuracy trade-off faced by most bats.
Interestingly, in addition to theoretical evidence for this fundamental trade-off [12], direct behavioural evidence was recently found. The bat Rousettus aegyptiacus was shown to point its beam not directly towards a target it needs to localize but slightly to the left and to the right of it. Hence, it receives less energetic but more informative echoes from the object of interest [47], thereby trading sensitivity for accuracy.
Concluding, we propose that the dominant glints, showing little amplitude variations (i.e. noise), are ideal input signals for a system using self-induced amplitude modulations to locate targets. Indeed, the low noise levels attained by only processing dominant glints, outweighs the loss in information due to sampling of the echo.
|
10.1371/journal.pntd.0001743 | Safety of the Recombinant Cholera Toxin B Subunit, Killed Whole-Cell (rBS-WC) Oral Cholera Vaccine in Pregnancy | Mass vaccinations are a main strategy in the deployment of oral cholera vaccines. Campaigns avoid giving vaccine to pregnant women because of the absence of safety data of the killed whole-cell oral cholera (rBS-WC) vaccine. Balancing this concern is the known higher risk of cholera and of complications of pregnancy should cholera occur in these women, as well as the lack of expected adverse events from a killed oral bacterial vaccine.
From January to February 2009, a mass rBS-WC vaccination campaign of persons over two years of age was conducted in an urban and a rural area (population 51,151) in Zanzibar. Pregnant women were advised not to participate in the campaign. More than nine months after the last dose of the vaccine was administered, we visited all women between 15 and 50 years of age living in the study area. The outcome of pregnancies that were inadvertently exposed to at least one oral cholera vaccine dose and those that were not exposed was evaluated. 13,736 (94%) of the target women in the study site were interviewed. 1,151 (79%) of the 1,453 deliveries in 2009 occurred during the period when foetal exposure to the vaccine could have occurred. 955 (83%) out of these 1,151 mothers had not been vaccinated; the remaining 196 (17%) mothers had received at least one dose of the oral cholera vaccine. There were no statistically significant differences in the odds ratios for birth outcomes among the exposed and unexposed pregnancies.
We found no statistically significant evidence of a harmful effect of gestational exposure to the rBS-WC vaccine. These findings, along with the absence of a rational basis for expecting a risk from this killed oral bacterial vaccine, are reassuring but the study had insufficient power to detect infrequent events.
ClinicalTrials.gov NCT00709410
| Pregnant women are more vulnerable to complications of cholera than other people. It would be helpful to include pregnant women in vaccination campaigns against cholera but pregnant women and their unborn children are highly vulnerable to the potential adverse effects of biological products such as vaccines. The safety of oral cholera vaccines in pregnant women has up to now not been evaluated. During a large mass cholera vaccination campaign in Zanzibar in 2009, women were advised not to participate if they thought they may be pregnant. The large majority (955 or 83%) of women residing in the study area who were to be pregnant during the 9 months following the vaccinations did not participate in the campaign. The remaining 196 (17%) women received the vaccine. A comparison between vaccine exposed and unexposed pregnancies did not reveal any significant differences in outcome between the two groups. The small number of miscarriages, infant deaths and ill infants was similarly distributed between the two groups. These findings are reassuring but continued monitoring of this vaccine when given during pregnancy is recommended.
| The recombinant cholera toxin B subunit, killed whole-cell oral cholera (rBS-WC, Dukoral) vaccine, has been found to be safe and protective in a range of settings over the last 30 years [1], [2], [3]. This vaccine is mainly used by tourists visiting endemic areas [4] where the control of cholera has traditionally been based on safe water supply, sanitation and health education [5]. A more affordable oral cholera vaccine which could be used more widely in endemic settings has recently been developed, licensed, and prequalified for purchase by UN agencies [6]. This second generation killed oral cholera vaccine (Shanchol) is composed of a different set of V.cholerae strains than the rBS-WC vaccine, includes not only O1 but also an O139 strain, does not include the recombinant B subunit (rBS), therefore does not require buffer for administration, and has afforded 66% protection during a 3 year trial in Kolkata, India [7]. In early 2010, the Strategic Advisory group of the World Health Organization (WHO) recommended that oral cholera vaccines be used preventively as well as reactively in the management of cholera outbreaks [8].
Since cholera tends to affect all age groups in endemic settings and during outbreaks, mass vaccination is considered an important vaccine deployment strategy. To achieve maximum impact of mass cholera vaccination, it is crucial to immunize the highest possible percentage of the population at risk. This includes women in the reproductive age group, defined here as being between 15 and 50 years old. In endemic and epidemic settings, women are at high risk for cholera and other diarrhoeal diseases, not least because mothers tend to be exposed to infectious children [9]. Without prompt rehydration, cholera during pregnancy can result in abortions, premature childbirth and maternal death [10], [11]. There are good reasons for women in the reproductive age group in endemic areas to participate in interventions that prevent cholera. Excluding potentially pregnant women from mass vaccination campaigns is logistically and ethically challenging. But administering oral cholera vaccines to this highly vulnerable population causes a dilemma since the safety of the vaccine during pregnancy has not been documented. There are several reasons why it is thought that oral cholera vaccines are unlikely to have a harmful effect on foetal development. First, the bacteria in the rBS-WC vaccine are killed and do not replicate. Second, the vaccine antigens act locally on the gastro-intestinal mucosa is not absorbed and does not enter the maternal or foetal circulation. Third rBS-WC vaccines don't trigger systemic reactions (e.g. fever) linked to abortions early in pregnancy. However, no actual safety studies of the rBS-WC vaccine in pregnancy have been carried out [12].
The uncertainty regarding the use of the vaccine during pregnancy has resulted in differing recommendations. The recommendations from the WHO state the following. “The primary targets for cholera vaccination in many endemic areas are preschool-aged and school-aged children. Other groups that are especially vulnerable to severe disease and for which the vaccines are not contraindicated may also be targeted, such as pregnant women and HIV-infected individuals.” [13]. The package insert of Dukoral, states: “The effect of DUKORAL [Oral, Inactivated Travellers' Diarrhoea and Cholera Vaccine] on embryo-foetal development has not been assessed and animal studies on reproductive toxicity have not been conducted. No specific clinical studies have been performed to address this issue. The vaccine is therefore not recommended for use in pregnancy. However, DUKORAL is an inactivated vaccine that does not replicate. DUKORAL is also given orally and acts locally in the intestine. Therefore, in theory, DUKORAL should not pose any risk to the human foetus. Administration of DUKORAL to pregnant women may be considered after careful evaluation ofthe benefits and risks.” The package insert of the second generation vaccine (Shanchol ) uses similarly guarded language: “The vaccine is not recommended for use in pregnancy. However, Shanchol is a killed vaccine that does not replicate, is given orally and acts locally in the intestine. Therefore, in theory, Shanchol should not pose any risk to the human foetus. Administration of Shanchol to pregnant women may be considered after careful evaluation of the benefits and risks in case of a medical emergency or an epidemic.”
A mass oral cholera vaccination was conducted in Zanzibar in 2009. Pregnant women were advised not to participate in the campaign. To assess whether any pregnant women had inadvertently received the vaccine, and to investigate birth outcomes, we visited all women residing in the study area and in the reproductive age group more than nine months after the last dose of the vaccine had been administered. The objective of the study was to determine whether there was any difference between the outcomes of pregnancies exposed and not exposed to the oral rBS-WC cholera vaccine.
The study methods have been described in more detail in the accompanying paper estimating the effectiveness of the vaccine [14].
The study was conducted according to the principles expressed in the Declaration of Helsinki. Individual verbal consent was obtained from each respondent after the purpose of the study was explained. The Institutional Review Board of the Government of Zanzibar (ZAMREC), of the International Vaccine Institute, Seoul, Korea, and the Research Ethics Review Committee of the World Health Organization, Geneva, Switzerland approved this project.
The informed consent process was done in several phases. Community informed consent was obtained through meetings with the local leaders (She has). A multistage community outreach campaign was conducted to disseminate information about the planned study activities. During the census, individual verbal informed consent was obtained prior to the interview of each household head or his or her representative. During the mass vaccination, individual verbal informed consent was obtained from each participant or from his or her guardian, if they were less than 18 years of age. In addition, verbal assent from children 12 to 17 years of age was obtained. The participants received information regarding the vaccine, including advice for children less than 2 years of age and pregnant women not to receive the vaccine. There was no screening for pregnancy prior to vaccine administration.
The interview of pregnant women was closely linked with the census, for which oral consent was provided. Like the census interview, the interview of pregnant women posed minimal risks and oral consent was deemed appropriate. Provision of oral consent by each participant was documented in a logbook. The use of oral consent was approved by the ethics review boards. After the surveillance was completed the three ethics review boards were informed about the conduct and the findings of the birth surveillance.
The archipelago of Zanzibar lies about 50 kilometres east of mainland Tanzania and consists of two main islands, Unguja and Pemba, as well as smaller islets. Zanzibar had a population of about 1.1 million in 2009. In Unguja, we included the shehias of Chumbuni, Karakana, and Mtopepo, which are informal, urbanized areas extending from the capital, Zanzibar City also known as Stonetown. These shehias arose without the corresponding development of adequate water and sanitation facilities. In Pemba, we included the shehias of Mwambe, Kengeja, and Shamiani, located in the mainly rural southeast of the island.
Each dose of the rBS-WC cholera vaccine (Dukoral ™, SBL Vaccine AB, Sweden) consists of ca. 1×1011 vibrios [12]:
The full dose of vaccine was mixed with 75 or 150 ml of buffer solution for participants aged from two to six years and over six years, respectively.
A formal census was conducted from November to December 2008, collecting demographic and socio-economic information. Verbal informed consent was obtained from the head of each household prior to the interviews. The number of household members, ownership of various capital goods and household building materials were recorded. Data was directly entered into handheld computers, also known as personal digital assistants (PDA) [15]. A unique identification number was assigned to each resident in the study sites. After the census was completed, household identification cards were distributed in early January 2009. At the time of card distribution, all healthy, non-pregnant residents of the study sites who were two years of age and older were invited to participate in the mass vaccination campaign. Study residents were requested to bring their household identification cards when coming to a vaccination outpost to facilitate identification. In August 2009, a second census was conducted in the study sites to update the study population database.
The mass vaccination campaign was implemented by the Expanded Program on Immunization of the Zanzibar Ministry of Health and Social Welfare with WHO technical support. The first round of immunizations was conducted from January 11 to 26, 2009, the second round from February 7 to 16, 2009. The vaccine vial was shaken, opened and its contents poured into a cup with buffer solution and stirred. The participants drank the mixture under direct observation and completeness of ingestion was recorded. During the first round, a card was issued to each vaccine recipient to record the subject's name, age, address, household head, date of vaccination, and completeness of ingestion of the dose. At the time of dosing, this information was also recorded in a PDA-based vaccination registry. Only those who had received a first dose (as documented in the vaccination card or the PDA registry) were given a second dose of the vaccine.
The birth surveillance was conducted more than 9 months after the mass vaccination campaign was completed, between January 15 and February 15, 2010. A list of all women between 15 and 50 years of age at the time of the vaccination campaign and living in the study area was prepared based on the study population database. Following training in study procedures fieldworkers visited the listed women and asked whether they had been pregnant in 2009. Women who had been pregnant were asked about the following: the date of delivery, duration and outcome of the pregnancy based on their last menstrual period, number of deliveries, age of the last child born before this delivery, antenatal clinic attendance during this pregnancy and person who attended the delivery. Birth outcomes were described as miscarriage or live births. We further defined a miscarriage as either a spontaneous abortion or a stillbirth. A spontaneous abortion was defined as a termination of a pregnancy within 20 weeks of conception. A stillbirth was defined as a foetus born after 20 weeks of gestation without a pulse. Live births that died later during infancy were described as infant deaths. For live births, the disposition of the baby was recorded. During the visit the field worker asked whether the baby is free from recurring illness, without gross malformations, and is feeding, urinating, defecating, crying, sleeping and growing normally. For the purpose of this surveillance, a recurrent illness was defined as an illness lasting more than two weeks or occurring twice or more often [16]. Only illnesses requiring the attention of medical staff were included. A gross malformation was defined as a physical defect present in a baby at birth. It includes any abnormality visible on a naked baby (e.g. cleft lip or palate, Down syndrome, spina bifida, limb defects, etc.). Whether feeding, urinating, defecating, crying, sleeping and growing was within the normal range was recorded according to the mother's definition. If the field worker considered the infant as sick or abnormal, the infant was seen by a paediatrician. The paediatrician completed a standardized history, physical examination and assessment and provided treatment or referral according to national guidelines. The field workers and paediatricians were blinded regarding the vaccination status of the mother.
The information collected during birth surveillance was linked to the population census and vaccination databases. Receipt of the cholera vaccine during the mass immunization program was ascertained based on the vaccination database. Linkage to the vaccination registry was made blinded to pregnancy outcome. Baseline data on socio-behavioural, economic, and environmental variables were obtained from the census database.
To calculate the date of conception, we subtracted the duration of the pregnancy (as defined by the mother in weeks based on the last menstrual period) from the date of delivery. The pregnancy was considered exposed to the vaccine if the period from conception to delivery included the dates when the woman received at least one vaccine dose. Additionally, because it is difficult to know the exact date of conception, we included pregnancies with calculated conception dates within two weeks before ingestion of the first vaccine dose as potentially exposed. A pregnancy was considered unexposed if the period from two weeks before the calculated conception date to the date of delivery did not include receipt of any oral cholera vaccine dose. We compared the frequency of adverse birth outcomes between exposed and unexposed pregnancies.
The number of miscarriages, live infants and infant deaths (birth outcomes) among the exposed and unexposed pregnancies were initially compared using chi-square or Fisher's exact test, as appropriate. Characteristics of women who had exposed and unexposed pregnancies were compared using chi-square and Student's t-test for binary/categorical and continuous variables, respectively. In the assessment of the risk for negative outcomes (miscarriage and infant sickness, abnormality or death), a stepwise elimination method was used to select variables most closely associated with exposure and non-exposure and to fit them into a logistic regression model. All p values and 95% confidence intervals were interpreted in a two-tailed fashion. Statistical significance was designated as a p value less than 0.05. Stata/SE 8 (Stata Corporation, Texas, USA) was used for statistical analysis.
The population census enumerated 14,564 women between 15 and 50 years of age residing in the study sites. During the birth surveillance, 13,736 (94%) of this population were located and interviewed. Women who participated had a significantly different health care utilization pattern, tended to be from a lower socio-economic background as suggested by the possession of fewer capital items (mobile phone, bicycle etc.), came from larger households and tended to be less well educated (Table S1).
Out of the interviewed women, 1,453 (11%) had a delivery in 2009; and 1,151 (79%) of these deliveries occurred during the period where the foetus could have been exposed to the vaccine. The large majority 955 (83%) out of these 1,151 mothers had not been vaccinated; the remaining 196 (17%) mothers had received at least one dose of the oral cholera vaccine (82 received 1 dose, 114 received 2 doses). The flow of the pregnant women is shown in Figure 1.
We compared the outcomes of pregnancies exposed and unexposed to the cholera vaccine (Table 1). There was no statistically significant difference in the number of miscarriages among the exposed compared to the unexposed pregnancies [10/196 (5%) vs. 27/955 (3%), adjusted odds ratio (AOR) 1.62 (95% confidence interval (95% CI 0.76 to 3.43)]. Similarly, there was no statistically significant difference in the number of infant deaths among the exposed compared to the unexposed non-miscarriage pregnancies [3/186(2%) vs. 13/928 (1%), AOR 1.46 (95% CI 0.41 to 5.29)]. The frequency of infant illness and abnormalities among the live infants verified by a paediatrician was 8/183 (4%) among the exposed versus 46/915 (5%) among the unexposed (AOR 0.79, 95th CI 0.36 to 1.75). Logistic regression models, adjusted for variation in background characteristics, found no significant difference in the frequency of miscarriages, sickness or abnormality, and infant deaths between the exposed and unexposed pregnancies.
We assessed the timing of the exposure to the cholera vaccine in relation to the gestational period of the ten miscarriages (Figure 2). Vaccine exposure occurred during the first trimester in three, during the second trimester in four, and during the third trimester in three pregnancies.
We compared individual and household characteristics of the mothers who had exposed and unexposed pregnancies (Table 2). Pregnant women who participated in the mass vaccination campaign differed in several aspects from pregnant women who didn't participate in the vaccinations. The women who received the vaccine were significantly older, had had more deliveries, attended antenatal care less frequently, had more frequently lived in the same household during the past 5 years and lived in a larger household with lower socio-economic status as suggested by the ownership of capital items and household construction materials.
This is the first report of the safety of the rBS-WC oral cholera vaccine administered during pregnancy. We found no significant differences in birth outcomes among pregnancies exposed and unexposed to the rBS-WC oral cholera vaccine. Among the 196 pregnancies with gestational exposure to the vaccine, there was no evidence of a statistically significant increase in the number of foetal losses or infant deaths compared to unexposed pregnancies. There was a slightly higher percentage of miscarriages in pregnancies exposed to the oral cholera vaccine than in pregnancies not exposed. This trend did not reach statistical significance and is likely explained by chance.
The study has several limitations. First, foetal losses were probably under-reported since pregnancy is often denied until late into gestation for complex cultural reasons. Second, and more importantly, exposure or non-exposure to the vaccine was not randomized. Instead prior to vaccination women were advised not to participate in the vaccinations if they were pregnant and there was no mandatory pregnancy testing prior to vaccination. This element of self-selection may have led to bias. Third, an element of recall bias can't be ruled out, namely women may have recalled adverse outcomes more frequently when they had been vaccinated than unvaccinated women. Fourth, the study detected 196 pregnancies. Even though the study is the largest to date, the overall number of pregnancies (196) is small and has limited power to detect infrequent adverse events. Fifth our sampling method does not detect maternal deaths. Finally, 6% of the eligible women did not participate in this birth surveillance study. Considering that the large majority (94%) of eligible women participated in the study it seems unlikely that this finding has introduced bias.
To help ensure the validity of the results, we performed the following procedures: To ensure the complete detection of all pregnancies in the study site, we visited and interviewed all women in reproductive age enumerated in the census. Extensive information about potentially confounding variables was available since data on baseline characteristics of individuals and households were collected during the census and during the interview of the mothers, which were controlled for in the analyses. Birth outcomes were linked in a blinded fashion to vaccination status in the database in order to avoid potential observer bias.
Pregnant women who participated in the mass vaccination campaign were older, had had more deliveries, came from bigger households and had lived in the same household for a longer period than pregnant women who did not participate. Younger pregnant women from smaller, better-off households may well perceive themselves at a lower risk for gastro-enteric infections than more experienced, older women from a lower socio economic background. Similar observations of an inverse relationship between participation in free mass vaccination campaigns and socio-economic status have been reported from Kolkata, India [17] and Hue, Vietnam [18]. Alternatively younger women from a higher socio-economic background have a better understanding of the potential risks of vaccination during pregnancy than women from a lower socio-economic background.
This study found no significant increase in adverse events involving the foetus or newborn among pregnant women who inadvertently received killed oral cholera vaccine. Because the sample size was small, our findings cannot rule out the possibility that rBS-WC vaccine could cause adverse events during pregnancy, but the study provides reassurance that such events are not common. The findings from this study support the current recommendation that killed oral cholera vaccine is not contraindicated during pregnancy, but the decision to administer the vaccine should depend on the epidemiological context and after weighing the potential benefits and risks [12]. Randomized, controlled studies of the rBS-WC vaccine in pregnant women are ethically not justifiable, but future, larger mass vaccinations may allow further evaluation of birth outcomes after inadvertent exposure of pregnant women to the vaccine.
|
10.1371/journal.ppat.1005147 | Flavodoxin-Like Proteins Protect Candida albicans from Oxidative Stress and Promote Virulence | The fungal pathogen Candida albicans causes lethal systemic infections in humans. To better define how pathogens resist oxidative attack by the immune system, we examined a family of four Flavodoxin-Like Proteins (FLPs) in C. albicans. In agreement with previous studies showing that FLPs in bacteria and plants act as NAD(P)H quinone oxidoreductases, a C. albicans quadruple mutant lacking all four FLPs (pst1Δ, pst2Δ, pst3Δ, ycp4Δ) was more sensitive to benzoquinone. Interestingly, the quadruple mutant was also more sensitive to a variety of oxidants. Quinone reductase activity confers important antioxidant effects because resistance to oxidation was restored in the quadruple mutant by expressing either Escherichia coli wrbA or mammalian NQO1, two distinct types of quinone reductases. FLPs were detected at the plasma membrane in C. albicans, and the quadruple mutant was more sensitive to linolenic acid, a polyunsaturated fatty acid that can auto-oxidize and promote lipid peroxidation. These observations suggested that FLPs reduce ubiquinone (coenzyme Q), enabling it to serve as an antioxidant in the membrane. In support of this, a C. albicans coq3Δ mutant that fails to synthesize ubiquinone was also highly sensitive to oxidative stress. FLPs are critical for survival in the host, as the quadruple mutant was avirulent in a mouse model of systemic candidiasis under conditions where infection with wild type C. albicans was lethal. The quadruple mutant cells initially grew well in kidneys, the major site of C. albicans growth in mice, but then declined after the influx of neutrophils and by day 4 post-infection 33% of the mice cleared the infection. Thus, FLPs and ubiquinone are important new antioxidant mechanisms that are critical for fungal virulence. The potential of FLPs as novel targets for antifungal therapy is further underscored by their absence in mammalian cells.
| Oxidative damage is a fundamental problem for cells and a particular challenge for microbial pathogens, which require special mechanisms to resist the oxidative attack by the host immune system. We identified four proteins in the human fungal pathogen Candida albicans that belong to a large family of enzymes in bacteria and plants that reduce quinone molecules to detoxify them. Interestingly, mutational studies in C. albicans showed that these enzymes also confer resistance to a wide range of oxidants, suggesting they may have broader impact by reducing the major quinone present in cells (ubiquinone or coenzyme Q). In support of this, we found that mutating the COQ3 gene to block ubiquinone synthesis rendered cells highly sensitive to oxidative stress, revealing that it plays a very important antioxidant function in addition to its well known role in energy metabolism. These quinone reductases play a critical role in vivo, as they were required for virulence in mouse infections studies. Since mammalian cells lack this type of quinone reductase, this difference could be exploited to develop much needed novel therapeutic approaches for fungal and bacterial pathogens.
| Oxidative stress poses a great threat to cells. Unchecked oxidative damage to DNA, proteins, and lipids causes disruption of physiological processes, harmful mutations, and cell death [1]. To prevent these destructive effects, cells utilize a variety of mechanisms to protect against oxidation. These antioxidant mechanisms are especially important for pathogens to resist the oxidative attack by the immune system [2]. As a result, the human fungal pathogen Candida albicans relies on several different mechanisms, such as extracellular, cytoplasmic, and mitochondrial forms of superoxide dismutases to break down superoxide radicals [3–5]. Other intracellular mechanisms include catalase to detoxify H2O2 and glutathione to promote a reducing environment [6].
Cellular membranes require special protection from oxidation. The plasma membrane is particularly vulnerable because it directly faces oxidative attack by macrophages and neutrophils. Protecting the plasma membrane is critical for survival. In addition to forming a protective barrier around the cell, it functions in a wide range of essential processes including nutrient uptake, ion homeostasis, pH regulation, cell wall synthesis, and morphogenesis. This membrane is also vulnerable because it contains polyunsaturated fatty acids (PUFAs). Approximately 30% of the C. albicans fatty acids are polyunsaturated linoleic (18:2) or linolenic (18:3) acids [7, 8]. PUFAs are very sensitive to peroxidation due to the ease with which the hydrogens can be abstracted from the methylene bridges (-CH2-) that lie in between the double bonds [9, 10]. This leaves an unpaired electron on the carbon that can react with O2 to form a peroxyl radical, which can in turn abstract the hydrogen from another PUFA to continue the cycle. Thus, lipid peroxidation starts a chain reaction that propagates to other lipids. The resulting oxidative damage can also spread to other cellular constituents, including proteins and DNA.
Several lines of evidence suggested that a family of four uncharacterized Flavodoxin-Like Proteins (FLPs) present in C. albicans could play a novel antioxidant role at the plasma membrane. The FLPs, which are encoded by PST1, PST2, PST3, and YCP4, are induced by oxidative stress [11]. The FLP genes contain consensus sites in their promoter regions for the binding of Cap1, a transcription factor that is induced by oxidative stress, and for a subset of these genes Cap1 has been shown to bind to the promoter and regulate expression [12, 13]. The S. cerevisiae FLPs (Pst2, Rfs1, Ycp4) have been suggested to promote resistance to oxidative stress [14–16], although their physiological role is not known [17]. It is also interesting that the C. albicans FLPs are likely to act at the plasma membrane, since their orthologs in S. cerevisiae are associated with the plasma membrane [18].
The FLPs are highly conserved in bacteria, plants, and fungi, but surprisingly not in mammalian cells [19]. Biochemically, the most well studied FLP is the E. coli WrbA protein. It uses flavin mononucleotide (FMN) as a cofactor and acts as a NAD(P)H quinone oxidoreductase [20–22]. FLPs from fungi, plants and other bacteria have also been shown to act as NAD(P)H quinone oxidoreductases, indicating that this is a conserved property of this family [15, 23–27]. A special feature of FLPs is that they carry out a two-electron reduction of a quinone to quinol (see structures in Fig 1A). This converts both carbonyl groups on the benzoquinone ring to hydroxyl groups. In contrast, other pathways that promote a one-electron reduction of quinone form a semiquinone intermediate that is a hazardous reactive oxygen species [9, 10]. Although the physiological role of WrbA is not known, there is suggestive evidence that it promotes resistance to oxidative stress [19, 21, 27].
Quinone reductases could promote resistance to oxidative stress in several ways. One is that they can reduce and detoxify small molecule quinones that are produced by some organisms for defense or created as benzene metabolites [28, 29]. In addition, they could act on endogenously produced quinones, such as ubiquinone (coenzyme Q), an isoprenylated benzoquinone. Ubiquinone is well known for its role in the mitochondrial electron transport chain, but it is also present in other cellular membranes, where it can undergo redox cycling to act as an antioxidant [30–34]. Mammalian cells use the enzyme Nqo1 (NAD(P)H quinone oxidoreductase), formerly known as DT-diaphorase, to safely carry out a two-electron reduction of ubiquinone and avoid semiquinone formation [35, 36]. Nqo1 is analogous to FLPs in that it uses NAD(P)H for reducing potential, but it differs in overall amino acid sequence and the active site is distinct from the FLPs, in part due to the fact that the active site of Nqo1 binds FAD as a cofactor rather than FMN [19]. However, it is not known how fungal cells, including C. albicans, carry out this function since they lack an obvious ortholog of NQO1. Therefore, in this study we examined a quadruple mutant lacking all four FLP genes (PST1, PST2, PST3 and YCP4). The results demonstrate that these proteins represent a new mechanism for protecting C. albicans against oxidative stress that is required for virulence in a mouse model of systemic candidiasis.
Four FLPs were identified in C. albicans based on their high sequence identity (45–50%) and similarity (~65%) to the well-studied E. coli WrbA (S1 Fig). This type of enzyme is advantageous because it uses NAD(P)H to carry out a two-electron reduction of toxic quinones that avoids creation of the semiquinone radical (Fig 1A) [21]. The conserved residues are concentrated in the active site near the location of the FMN co-factor. The four C. albicans FLPs share a similar structure, although Ycp4 contains C-terminal extension of about 90 amino acids that ends in a CAAX box, indicating it is likely to be lipid modified (S1 Fig). To examine their role in the diploid C. albicans, a quadruple mutant strain was constructed that lacks both copies of all four FLP genes. Fortuitously, PST3 and YCP4 are adjacent in the genome and were deleted simultaneously using the HIS1 and LEU2 selectable markers. Subsequent deletion of the PST1 and PST2 genes was carried out by successive use of the SAT Flipper that employs a recyclable SAT1 selectable marker [37]. For brevity, this pst1Δ pst2Δ pst3Δ ycp4Δ strain will be referred to as the Δ/Δ/Δ/Δ mutant. The sensitivity of this strain to quinones was tested by spotting dilutions of cells onto agar medium containing p-benzoquinone (BZQ) or menadione (MND), a heterocyclic napthoquinone (Fig 1B). The growth of the Δ/Δ/Δ/Δ strain was clearly inhibited by these small molecule quinones, indicating it is more sensitive to quinones than either the wild type control or a complemented strain in which one copy of each of the FLP genes was reintroduced.
FLPs in bacteria and plants have also been suggested to have a role in fighting oxidative stress, but their physiological role is not known [19, 21, 25, 27]. Therefore, given the importance of antioxidant enzymes for microbial pathogens, we spotted the cells on medium containing H2O2 and found that the Δ/Δ/Δ/Δ mutant was more sensitive to this oxidant (Fig 1B). Since the FLPs are associated with the plasma membrane in S. cerevisiae [18], we further tested two other peroxides that are more hydrophobic. Interestingly, the Δ/Δ/Δ/Δ mutant was also very sensitive to tert-butyl hydroperoxide (TBHP) and cumene hydroperoxide (CHP), which are more hydrophobic than H2O2 and more likely to preferentially oxidize membranes.
The Δ/Δ/Δ/Δ mutant was next assayed for sensitivity to polyunsaturated fatty acids (PUFAs), which can auto-oxidize and initiate a chain reaction of lipid peroxidation [10, 33]. PUFAs are more readily oxidized because the presence of double bonds flanking a methylene group (-CH2-) weakens the methylene C-H bond, making it much easier to abstract a hydrogen [9]. This leaves a carbon with an unpaired electron that readily reacts with oxygen to form a peroxyl radical (LOO•). For example, linolenic acid, which has three unsaturated double bonds, is much more likely to auto-oxidize to form a peroxyl radical than is monounsaturated oleic acid. The peroxyl radical can then abstract a hydrogen from another PUFA to form a lipid peroxide (LOOH) and a new lipid radical that can further extend a chain reaction of lipid peroxidation [9, 10]. Linolenic acid was also used for this analysis because previous studies showed that it efficiently induced lipid peroxidation and cell death in S. cerevisiae [33]. Interestingly, growth of the Δ/Δ/Δ/Δ mutant was strongly inhibited by the polyunsaturated linolenic acid (LNA; Fig 1B). In contrast, the Δ/Δ/Δ/Δ mutant grew as well as the control cells in the presence of the monounsaturated oleic acid (OA). Taken together, these results indicate that the FLPs are needed for C. albicans to combat a variety of oxidative stresses.
The effects of linolenic acid on C. albicans were analyzed further in quantitative assays. A time course of cell death was assayed by incubating cells for different times in the presence of 0.5 mM linolenic acid followed by plating dilutions on agar medium to determine the viable colony forming units (CFUs). The results confirmed the spotting assays. The Δ/Δ/Δ/Δ mutant showed a significant trend toward decreased viability by 6–8 h that was not observed for the wild-type control or complemented strains (Fig 2A). Analysis of the dose-response to incubation with linolenic acid for 6 h revealed a loss of viability starting at 0.25 mM that became more significant at 0.5 and 1.0 mM (Fig 2C). In contrast, the cells remained viable after incubation in the monounsaturated oleic acid (S2 Fig).
To determine whether linolenic acid caused an increase in lipid peroxidation, cells were assayed for thiobarbituric acid reactive substances (TBARS) [33, 38]. This assay detects malondialdehyde (MDA), a common byproduct of lipid peroxidation. As expected, both the Δ/Δ/Δ/Δ mutant and the control cells showed elevated TBARS after incubation for different times with linolenic acid (Fig 2B). However, the Δ/Δ/Δ/Δ mutant showed a significantly higher level of TBARS than the control cells at 4 and 6 h. By 8 h, the results of the TBARS assays were quite variable. This may have been due to difficulties in dealing with the high fraction of dead cells during the analysis. Dose-response assays showed that the TBARS in the Δ/Δ/Δ/Δ mutant started trending upward at 0.25 mM and was significantly higher than control cells at 0.5 mM and 1.0 mM linolenic acid (Fig 2D). These results demonstrate that linolenic acid stimulated higher levels of lipid peroxidation in the Δ/Δ/Δ/Δ mutant.
For comparison, mutants lacking a single FLP gene (pst1Δ, pst2Δ, pst3Δ or ycp4Δ), two genes (pst3Δ ycp4Δ), or three genes (pst2Δ, pst3Δ ycp4Δ) were also tested for their sensitivity to 0.5 mM linolenic acid (Fig 2E and 2F). However, no significant changes in either CFU or lipid peroxidation level were detected compared to the wild type control. As will be described further below, this is consistent with redundancy of the different FLP genes in C. albicans.
To gain additional evidence that the effects of linolenic acid were due to oxidation, cells were incubated with α-tocopherol (vitamin E), a hydrophobic reducing agent that localizes to membranes and has been shown to prevent lipid peroxidation in other organisms [9, 33]. Treatment of cells with α-tocopherol alone had no detectable effects on CFUs or lipid peroxidation. In contrast, the addition of α-tocopherol significantly decreased the killing activity of linolenic acid in both WT and the Δ/Δ/Δ/Δ mutant (Fig 3A). Similarly, α-tocopherol reduced the levels of lipid peroxidation to below the limit of detection, as determined by the TBARS assay (Fig 3B).
To confirm whether quinone reductase activity is important to promote resistance to oxidative stress in C. albicans, the Δ/Δ/Δ/Δ mutant was engineered to express two distinct types of NAD(P)H quinone oxidoreductases: rat NQO1 and E. coli wrbA. NQO1 and wrbA were selected because their proteins have been well-studied biochemically [21, 39, 40]. These genes were expressed under the control of the strong ADH1 promoter. As a control, cells were also engineered to express GFP in a similar manner. Incubation of the cells in the presence of 0.5 mM linolenic acid for 6 h showed that expression of either wrbA or NQO1 rescued the viability of the Δ/Δ/Δ/Δ mutant (Fig 4A). In contrast, the Δ/Δ/Δ/Δ mutant or the Δ/Δ/Δ/Δ mutant that expressed only GFP showed a significant drop in CFUs. Similarly, expression of wrbA or NQO1, but not GFP, diminished lipid peroxidation in cells that were exposed to linolenic acid (Fig 4B). Growth assays on agar plates also showed that wrbA and NQO1 could complement the increased sensitivity of the Δ/Δ/Δ/Δ mutant to H2O2, tert-butyl hydroperoxide, cumene hydroperoxide and menadione (Fig 4C). The ability of two distinct quinone reductases to complement the Δ/Δ/Δ/Δ mutant phenotype demonstrates that this activity plays a key antioxidant role in C. albicans.
The properties of the different quinone reductase homologues were examined by expressing individual genes in the Δ/Δ/Δ/Δ mutant. The C. albicans genes were reintroduced under control of their native promoters, whereas wrbA and NQO1 were controlled by the ADH1 promoter. Growth assays were performed to test the ability of cells carrying only one quinone reductase gene to resist different quinones and oxidants. All of the different quinone reductases were able to promote resistance to H2O2, tert-butyl hydroperoxide, and linolenic acid (Fig 5A). However, some of the strains had differential ability to resist cumene hydroperoxide and the small molecule quinones: p-benzoquinone and menadione (Fig 5A and summarized in Fig 5B).
The strain expressing only PST3 was very interesting in that it showed the strongest resistance to p-benzoquinone and menadione (Fig 5A). Although several strains displayed obvious resistance to 75 μM p-benzoquinone, only the PST3-expressing strain was resistant to 100 μM p-benzoquinone. It grew remarkably better than the other strains, and nearly as well as the complemented strain that carries one copy of each FLP gene. Similarly, it also grew better than the other strains on medium containing menadione. In contrast, the PST3-expressing strain did not show significant resistance to cumene hydroperoxide and was more weakly resistant to linolenic acid, which are considered to be good inducers of lipid peroxidation. This strain was, however, more resistant than the Δ/Δ/Δ/Δ mutant to H2O2 and tert-butyl hydroperoxide, indicating that it can provide protection against some oxidants. Thus, it appears that Pst3 can preferentially act on small molecule quinones. In agreement with this, a pst3Δ strain was sensitive to the inhibitory effects of p-benzoquinone and menadione, whereas the pst1Δ, pst2Δ and ycp4Δ mutants were not (Fig 5C). The increased sensitivity of the pst3Δ mutant to p-benzoquinone and menadione were the only phenotypes we detected for the single mutants as we did not detect increased sensitivity to oxidizing conditions (Fig 2).
Some of the other strains expressing a single quinone reductase showed the opposite phenotype of being more resistant to oxidants than to the small molecule quinones. For example, the strains expressing PST2, YCP4, wrbA, or NQO1 all showed improved resistance to cumene hydroperoxide and linolenic acid compared to the Δ/Δ/Δ/Δ mutant, but were not significantly more resistant or were more weakly resistant to the small molecule quinones under the conditions tested (Fig 5A). The different phenotypes indicate that there are functional differences between the various quinone reductases.
The major quinone found in cells, ubiquinone (coenzyme Q), is known to have two key functions. It plays a central role in the mitochondrial electron transport chain, and it is also present in other cellular membranes where it can function as an antioxidant [30–34]. To investigate the relationship between ubiquinone and oxidative stress, both copies of COQ3 were deleted from C. albicans to prevent ubiquinone synthesis. As expected, a C. albicans coq3Δ mutant was not able to grow on glycerol, a carbon source that requires respiration to be utilized (Fig 6A). In contrast, the Δ/Δ/Δ/Δ mutant readily grew on glycerol (Fig 6A). Interestingly, the coq3Δ mutant was very sensitive to H2O2, even more so than the Δ/Δ/Δ/Δ mutant (Fig 6A). Spot assays also showed that the coq3Δ mutant was more sensitive to linolenic acid than the Δ/Δ/Δ/Δ mutant. For comparison, two previously constructed mitochondrial mutants were examined that lack components of Complex I of the electron transport chain [41]. Both orf19.2570Δ and orf19.6607Δ failed to grow on glycerol medium, as expected (Fig 6A). However, they were not more sensitive to linolenic acid and showed perhaps only a minor increase in sensitivity to H2O2. This indicates that a mitochondrial defect does not account for the increased sensitivity to oxidation of the coq3Δ mutant, consistent with ubiquinone also playing a major role as an antioxidant.
Analysis of cell viability after incubation with 0.5 mM linolenic acid for 6 h revealed a larger drop in CFUs for the coq3Δ mutant than for the Δ/Δ/Δ/Δ mutant (Fig 6B). The coq3Δ mutant also displayed significantly higher levels of TBARS under these conditions (Fig 6C). Similar results have been observed in S. cerevisiae, as a coq3Δ mutant in this yeast is also sensitive to oxidation and lipid peroxidation [33]. These results demonstrate that ubiquinone plays an important role as an antioxidant to prevent lipid peroxidation and oxidative stress in C. albicans.
It is noteworthy that the coq3Δ mutant was not significantly more sensitive to p-benzoquinone and menadione, even though it was very sensitive to H2O2 and linolenic acid (Fig 6A). This suggests that the FLPs in C. albicans can detoxify these small molecule quinones in the absence of ubiquinone, thereby prevent them from causing oxidative damage.
FLPs were fused to GFP to examine their subcellular localization. Pst1-GFP and Pst3-GFP were detected at the plasma membrane by fluorescence microscopy (Fig 7A). To improve detection for the other two FLPs, the strong ADH1 promoter was used to express GFP fusions to the PST2 and YCP4 genes. These GFP-Pst2 and GFP-Ycp4 fusion proteins gave a strong plasma membrane signal (Fig 7B). The GFP-tagged FLPs all showed a slightly patchy distribution in the plasma membrane, suggesting that they localize in part to the eisosome subdomains, as do their S. cerevisiae orthologs [18, 42]. Cytoplasmic GFP signal was also detected in cells. However, this could be due to proteolytic cleavage of the FLP proteins resulting in the presence of free cytoplasmic GFP, as Western blot analysis detected a strong signal at the expected size of GFP (~30 kD) (S3 Fig).
The role of the FLPs in virulence was examined using a mouse model of hematogenously disseminated candidiasis [43]. After injection via the tail vein with 2.5 x 105 C. albicans cells, BALB/c mice infected with the wild type control strain succumbed to infection with a median time of 8 days (Fig 8A). Similar results were observed for the complemented version of the Δ/Δ/Δ/Δ strain. In contrast, all mice infected with the Δ/Δ/Δ/Δ mutant survived to the end of the experiment (Day 28). No CFUs were detected in the kidneys from these mice, indicating that they had cleared the infection (Fig 8B).
To determine whether the Δ/Δ/Δ/Δ mutant failed to initiate an infection, or if it was cleared more rapidly, kidneys were examined at early times post infection. The kidney is a sensitive organ to test the ability of C. albicans to initiate an infection, since this fungus grows rapidly in the kidneys during the first two days after infection [44, 45]. At day 2 post infection, the wild type and Δ/Δ/Δ/Δ mutant were both present at similarly high levels of CFU/g kidney, indicating they grew well initially (Fig 8B). Histological analysis showed that foci of C. albicans growth in the kidney overlapped with clusters of leukocytes (Fig 8C). However, by the 4th day post infection, the median CFU/g kidney was 100-fold lower for mice infected with the Δ/Δ/Δ/Δ mutant than the wild type. Furthermore, 33% of the mice (3/9) had no detectable CFU/g kidney at day 4, indicating that they had cleared the infection. Thus, the FLPs are required for the persistence of C. albicans systemic infection.
Previous studies have shown that oxidation sensitive mutants, including those with defects in catalase, thioredoxin, or superoxide dismutatase, show normal or only slightly increased sensitivity to killing by neutrophils [46]. Similar results were obtained when the Δ/Δ/Δ/Δ mutant was examined for sensitivity to killing by macrophages derived from mouse bone marrow cells. Although the Δ/Δ/Δ/Δ mutant showed a slight increase in killing by macrophages, the difference was not statistically significant (S4 Fig).
Analysis of the pst3Δ ycp4Δ double mutant and the pst2Δ pst3Δ ycp4Δ triple mutant showed that they did not display a significant virulence defect in mice (S5 Fig). Mice infected with the triple mutant appeared to show slightly increased survival (median 12.5 days) compared to the wild type control strain (median 8 days), but this difference was not statistically significant using a log rank test (Mantel-Haenszel). These results are consistent with the general redundancy of the FLP genes seen in the in vitro studies. In addition, since both the double and triple mutant lack PST3, this indicates that the special role this FLP plays in detoxifying small quinones does not appear to be important in systemic candidiasis.
Cells utilize a variety of pathways to protect against oxidation [1, 3, 5, 6]. Cytoplasmic mechanisms include superoxide dismutase, catalase, thioredoxin, and glutathione. In addition, pathogens have also evolved extracellular mechanisms. For example, C. albicans produces three superoxide dismutases that are GPI-anchored and therefore on the cell surface or built into the cell wall (Sod4-6) [5, 6]. One of these, Sod5, was recently shown to have unique properties in that it uses copper as a co-factor, but not zinc [47]. This appears to be designed to take advantage of the fact that copper is pumped into phagosomes but zinc is restricted as part of the antimicrobial attack by leukocytes. However, it is not as well understood how cellular membranes are protected from oxidation, particularly the fungal plasma membrane that is directly exposed to the oxidative attack by neutrophils and macrophages [2].
To better understand how the plasma membrane is protected against oxidation we examined four FLPs in C. albicans that are associated with the plasma membrane (Fig 7). In agreement with their predicted role as NAD(P)H quinone oxidoreductases, a C. albicans Δ/Δ/Δ/Δ quadruple mutant lacking all four FLP genes (PST1, PST2, PST3, and YCP4) displayed increased sensitivity to p-benzoquinone and menadione, a napthoquinone (Fig 1). Interestingly, the mutant cells were also more sensitive to a wide range of oxidants, indicating that they have a broader antioxidant function.
Consistent with the membrane localization of the FLPs, the Δ/Δ/Δ/Δ mutant was very sensitive to hydrophobic oxidants, including linolenic acid (Figs 1 and 5). The increased sensitivity to linolenic acid was particularly significant, since previous studies demonstrated that this PUFA auto-oxidizes and initiates a chain reaction of lipid peroxidation [33]. In agreement with this, the Δ/Δ/Δ/Δ mutant showed elevated levels of TBARS (Fig 2), a hallmark of lipid peroxidation [9, 10]. Furthermore, the effects of linolenic acid could be reversed by the hydrophobic antioxidant α-tocopherol (Vitamin E) (Fig 3). Lipid peroxidation is likely to be a more serious problem for C. albicans than for S. cerevisiae, which lacks significant levels of PUFAs [33]. About 30% of the fatty acids in C. albicans are polyunsaturated [7, 8], which predisposes them to forming lipid peroxides [9, 10]. These PUFAs are typically found in more complex lipids, such as phospholipids, in addition to existing as free fatty acids. Taken together, the results identify FLPs as an important new set of antioxidant enzymes in C. albicans. These results also have broad significance for other pathogens, given that FLPs are induced by oxidative stress in diverse fungi [11, 14, 48–50] and there is suggestive evidence that they play an antioxidant role in bacteria [17, 19, 21, 27].
Biochemical studies of FLPs from bacteria, fungi, and plants have shown that they use NAD(P)H to reduce quinones in a manner that avoids creation of hazardous semiquinone intermediates [23–27]. The Δ/Δ/Δ/Δ mutant was rescued by expression of E. coli wrbA (Fig 4), confirming that NAD(P)H quinone oxidoreductase activity plays an important antioxidant role in C. albicans. Furthermore, heterologous expression of mammalian NQO1 in the Δ/Δ/Δ/Δ mutant also rescued its sensitivity to oxidation and lipid peroxidation. Nqo1 does not share obvious sequence similarity with FLPs even though it carries out a similar enzymatic activity. Although there are some underlying structural similarities between Nqo1 and FLPs, they are quite distinct [19]. For example, Nqo1 binds FAD as a cofactor instead of FMN, and it forms dimers rather than tetramers as seen for wrbA. These observations provide strong support that the key function of the C. albicans FLPs is to act as quinone reductases.
Analysis of Δ/Δ/Δ/Δ cells engineered to express a single FLP gene indicated that they have overlapping but distinct functions. Pst3 provided the best protection against the small molecule quinones p-benzoquinone and menadione (Fig 5). In agreement with this, a pst3Δ mutant was the only single FLP deletion mutant that was more sensitive to the small molecule quinones (p-benzoquinone and menadione) (Fig 5). In contrast, cells expressing only PST3 were less able to resist other oxidants, such as linolenic acid or cumene hydroperoxide. These phenotypes are consistent with different functional properties. However, it is also possible that some of these differences are due to differential expression of the various FLP genes under the different conditions that were tested.
The most likely target for the quinone reductase activity of FLPs in C. albicans is ubiquinone (coenzyme Q). Ubiquinone has a benzoquinone head group and a hydrophobic isoprenylated tail that localizes it to membranes [32, 51]. Analogous to its well-known role as an electron carrier in the mitochondria, ubiquinone is present in other cellular membranes where its reduced form (ubiquinol) can act as an antioxidant [30–34]. In particular, ubiquinol is thought to be able to reduce lipid radicals that would otherwise propagate a chain reaction of lipid peroxidation to cause more extensive damage [9, 10]. To determine if ubiquinol plays an important antioxidant role in C. albicans, COQ3 was deleted to block its synthesis. The coq3Δ mutant was found to be very sensitive to oxidative stress and also displayed increased levels of lipid peroxidation in response to linolenic acid (Fig 6). In further support of the conclusion that FLPs act on ubiquinone, rat NQO1, which is known to reduce ubiquinone [35, 36], can rescue the defects of the Δ/Δ/Δ/Δ mutant (Figs 4 and 5).
There were interesting differences between the Δ/Δ/Δ/Δ mutant and the coq3Δ mutant that reveal insights into their roles. Whereas the coq3Δ mutant was highly sensitive to oxidizing conditions promoted by peroxides or PUFAs, it was not significantly altered in sensitivity to p-benzoquinone and menadione (Fig 6). This indicates that the FLPs can reduce quinones in the absence of ubiquinol. The coq3Δ mutant was also much more sensitive than the Δ/Δ/Δ/Δ mutant to H2O2 and linolenic acid. One possibility is that other reductases can contribute to reduction of ubiquinone in the absence of the FLPs. However, if these enzymes use a one-electron mechanism for reduction of quinones, they will generate deleterious semiquinone radicals that would contribute to the phenotype of the Δ/Δ/Δ/Δ mutant.
The FLPs were required for virulence in a mouse model of hematogenously disseminated candidiasis (Fig 8A). Whereas the median survival time was 8 days for mice injected with 2.5 x 105 wild type C. albicans, all of the mice infected with the Δ/Δ/Δ/Δ mutant survived to the end of the experiment (day 28). Thus, the Δ/Δ/Δ/Δ mutant appears to have a stronger virulence defect than was reported for other C. albicans oxidation sensitive mutants including a cat1Δ catalase mutant [52], a sod1Δ or sod5Δ superoxide dismutase mutant [53, 54], a grx2Δ glutathione reductase mutant [53], or a tsa1Δ thioredoxin peroxidase mutant [55].
Interestingly, the Δ/Δ/Δ/Δ mutant could initially grow in the kidney essentially as well as a wild type strain (Fig 8B). However, by day 4 there was about a 100-fold decrease in median CFUs and 3 out of 9 mice cleared the infection. This decline in CFUs for the Δ/Δ/Δ/Δ mutant correlates with the influx of neutrophils (Fig 8C) that typically peaks about day 2 of infection [44, 45]. By day 28, all of the mice infected with the Δ/Δ/Δ/Δ mutant lacked detectable CFU and appear to have cleared the infection. Generally similar results were reported for a C. albicans cat1Δ catalase mutant that also grew well initially, but then CFUs declined in most infected mice [52]. In this regard it is also significant that a wrbAΔ mutant of the bacterial pathogen Yersinia tuberculosis can initiate an infection but is defective in establishing a persistent infection [56].
This key role in virulence for the FLPs indicates they have strong potential to serve as novel targets for antifungal therapy. New therapeutic approaches are needed; ~40% of patients with systemic candidiasis succumb to the infection even with current antifungal therapy [57, 58]. This outcome is likely to worsen, as drug resistance is a growing problem for two of the three most commonly used antifungal drugs [59, 60]. An important advantage of targeting FLPs is that they do not have orthologs in humans. The analogous NAD(P)H quinone oxidoreductases in mammals, Nqo1 and Nqo2, are very different [19].
Pharmacological studies on Nqo1 have identified multiple ways that quinone reductases can be targeted. One approach is to identify inhibitors, such as dicoumarol that blocks the Nqo1 activity [61]. In addition, the ability of Nqo1 to reduce small molecule quinones has been studied as a basis for cancer chemotherapy. The fact that many cancer cells overexpress NQO1 has been exploited to develop novel therapies in which quinone compounds are reduced by Nqo1 to convert them into a toxic form that preferentially kills cancer cells [36, 62]. Similarly, Nqo1 has also been shown to reduce benzoquinone-containing ansamycin drugs, which makes these compounds more potent inhibitors of the Hsp90 chaperone [63]. This suggests yet another way drugs targeting FLPs could be useful, since Hsp90 inhibitors can prevent the emergence of drug resistance in C. albicans [64]. Thus, the important roles of FLPs in oxidative stress response and virulence, combined with their absence in mammalian cells, identifies them as important new targets for therapeutic strategies aimed at combating fungal and bacterial pathogens.
All procedures were approved by the Stony Brook University IACUC Committee (#1686). Mice were considered to be moribund if food and water could no longer be accessed and then humane euthanasia was performed by carbon dioxide inhalation as per instructions from the Department of Laboratory Animals at Stony Brook University.
Oleic acid, linoleic acid, linolenic acid, α-tocopherol (vitamin E), hydrogen peroxide, tert-butyl hydroperoxide, cumene hydroperoxide, menadione, p-benzoquinone, thiobarbituric acid (TBA), hydrochloric acid, and 1,1,3,3-tetramethoxypropane were purchased from Sigma-Aldrich Corp. Trichloroacetic acid was from the Alfa Aesar Company, and nourseothricin from Werner BioAgents.
The C. albicans strains used in this study are described in Table 1. Cells were grown in SD medium (yeast nitrogen base synthetic medium with dextrose) [65]. C. albicans deletion mutants were created in strain SN152 (arg4Δ his1Δ leu2Δ) by homologous recombination, as described previously [66]. Mutant strains that carry homozygous deletion of PST1, PST2, PST3, YCP4, or COQ3 were constructed with strain SN152 by the sequential deletion of both copies of the targeted gene. Gene deletion cassettes were generated by PCR amplification of the LEU2 or HIS1 selectable marker gene [66], using primers that also included ~80 bp of DNA sequence homologous to the upstream or downstream region of the targeted open reading frame (ORF). Cells that had undergone homologous recombination to delete the targeted gene were identified by PCR analysis. A pst3Δ ycp4Δ double mutant strain was constructed by simultaneous deletion of both genes, taking advantage of the fact that they are adjacent in the genome. Homozygous triple and quadruple deletion mutation strains were then constructed by sequential deletion of both copies of the targeted gene using the SAT1 flipper method to recycle the selectable marker [37]. Similar phenotypes were observed for independent isolates. Deletion strains were then made prototrophic by transforming with the ARG4 gene to correct the remaining auxotrophy.
Complemented strains, in which the wild-type FLP gene was reintroduced into the corresponding deletion mutant, were constructed by first using PCR to amplify the corresponding FLP gene plus 2000 base pairs (bp) upstream and 501 bp downstream of the PST1 open reading frame (ORF), 811 bp upstream and 427 bp downstream of the PST2 ORF, 1681 bp upstream and 427 bp downstream of the PST3 ORF, or 1526 bp upstream and 310 bp downstream of the YCP4 ORF. The DNA fragments were then inserted between the SacI and SacII sites in a derivative of plasmid pDDB57 [67] in which the URA3 gene was replaced with ARG4. The resulting plasmids were then linearized by restriction digestion in the promoter region, and then transformed into the corresponding homozygous deletion mutant strains using ARG4 for selection. The complementing plasmids were also transformed into the Δ/Δ/Δ/Δ mutant to create strains that express only a single FLP gene. A fully complemented quadruple mutant strain was constructed essentially as described above, except that both PST1 and PST2 genes were cloned in tandem into the ARG4 plasmid. The plasmid was digested with BspEI to linearize it in the promoter region of the PST1 gene, and then it was transformed into the quadruple mutant strain LLF054. The PST3 and YCP4 genes were cloned between the SacI and ApaI restriction sites of a derivative of plasmid pDDB57 in which the URA3 selectable marker was changed to the SAT1 gene to confer nourseothricin resistance. Note that the PST3 and YCP4 genes are adjacent in the genome in a head to head manner, such that the corresponding open reading frames are transcribed in a divergent manner. A PCR fragment that contains sequences between 1157 bp downstream of the YCP4 ORF and 466 bp downstream of the PST3 ORF was used to create the PST3-YCP4 complementing plasmid. The resulting plasmid was digested with SnaBI to linearize it about 400 bp downstream of the YCP4 open reading frame, and then the DNA was transformed into strain LLF078, a version of the Δ/Δ/Δ/Δ quadruple mutant in which the PST1 and PST2 genes were already introduced as described above.
The open reading frames for E. coli wrbA and rat NQO1 were synthesized by GeneWiz Corp. so that the codons could be optimized and to avoid CUG codons that are translated differently in C. albicans. The wrbA and NQO1 open reading frames were amplified by PCR and introduced downstream of the ADH1 promoter and GFP in plasmid pND391 that carries the ARG4 selectable marker. The resulting plasmids were then transformed into the Δ/Δ/Δ/Δ mutant strain LLF054 to create strains expressing wrbA (LLF074), NQO1 (LLF076), or GFP (LLF080) as a control.
Spot assays to test growth in the presence of oxidizing agents were carried out essentially as described previously [68, 69]. C. albicans mutant or wild type strains were grown overnight and then adjusted to 107 cells/ml. Serial 10-fold dilutions of cells were prepared, and three μl of each dilution was then spotted onto solid agar SD medium containing the indicated chemical. The plates were incubated at 37°C for 2 days and then photographed. Each assay was done at least three independent times.
Cells were also tested in liquid culture for sensitivity to oxidizing agents by assaying colony-forming units (CFUs). C. albicans cells were grown in synthetic medium with 2% dextrose and without amino acids at 37°C overnight. Cells were harvested at about 6–10 x 107 cells per ml, washed once, and resuspended in phosphate buffer (0.1M sodium phosphate, pH 6.2, 0.2% dextrose) at 107 cells per ml. Three ml was transferred to a 15 ml glass test tube and then fatty acids were added. Cells were then incubated at 37°C on a tube rotator for the designated period of time. At the end of each treatment, cells were harvested by centrifugation and samples were plated to determine the number of viable CFUs. Results represent the average of 3–6 independent assays.
The level of TBARS in yeast whole cell lysates was determined by a modification of a previously described procedure [70]. At the end of the fatty acid treatment described above, 1.5 x 107 cells were harvested by centrifugation at 17,000 x g for 5 min, washed once with distilled water, and resuspended in 100 μl distilled water in a screw cap tube. 1ml of a freshly prepared solution of 0.375% thiobarbituric acid dissolved in 12% trichloroacetic acid and 0.5 M hydrochloric acid was added to each tube. After a 20-minute incubation at 90°C, samples were allowed to cool down, and then the insoluble material was sedimented by centrifugation at 17,000 x g for 5 min. The absorbance of the supernatant was measured at 535 nm, and corrected by subtracting the nonspecific absorbance at 600 nm. The corrected absorbance was then compared with a standard curve created using 1,1,3,3-tetramethoxypropane treated under the same conditions, which generates malondialdehyde (MDA). Results represent the average of 3–4 independent experiments.
The GFPγ variant was fused to the 3’ ends of the open reading frames for PST1 and PST3 using HIS1 selection, in LLF018, as described previously [71]. Strains were verified by PCR analysis and microscopic examination of GFP fluorescence. To add GFP at the 5’ end of the open reading frames to create N-terminal fusions, GFPγ was introduced downstream of the ADH1 promoter followed by the FLP gene and then the ADH1 terminator in pND397, which carries an URA3 selectable marker gene [72]. The plasmids of pADH1-GFPγ-PST1, pADH1-GFPγ-PST2, or pADH1-GFPγ-PST3 were also linearized with Not I, before being individually transformed into LLF089 using URA3 as the selectable marker to create the strains LLF091, LLF092, and LLF093. The plasmid pADH1-GFPγ-YCP4 was linearized with NotI and transformed into LLF018 using URA3 as the selectable marker to construct the strain LLF071. GFP fluorescence was analyzed directly in live cells without further processing using a Zeiss Axiovert 200M microscope equipped with an AxioCam HRm camera and Zeiss AxioVision software for deconvolving images.
The survival of C. albicans cells in the presence of macrophages was assayed essentially as described previously [73]. Bone marrow was isolated from femurs of 6- to 8-week-old female C57BL/6 mice (Jackson Laboratories) and then macrophages were derived from them as previously described [74]. At 18 h prior to infection, bone marrow derived macrophages were seeded into multiwell trays in Dulbecco’s modified Eagle medium (Invitrogen) supplemented with 10% fetal bovine serum (HyClone), 15% L-cell-conditioned medium, 1 mM sodium pyruvate, 2 mM glutamate, and 100 ng/ml E. coli LPS (Sigma). Dilutions of C. albicans cells were then plated in multiwell trays in the presence or absence of the macrophages and incubated for 48 h [73]. Microcolonies of growth in each well were then counted to determine the reduction in C. albicans viability due to the presence of macrophages. The results represent the average of three different experiments in which different batches of bone marrow-derived macrophages were used.
C. albicans strains were tested for virulence in a mouse model of hematogenously disseminated candidiasis similar to previous studies [45, 75]. C. albicans cells were cultured by growing overnight at 30°C in YPD medium with 80 μg/ml uridine, reinoculating into fresh medium, and incubating again overnight at 30°C. Cells were prepared for infection assays by washing twice in phosphate-buffered saline (PBS), counting in a hemocytometer, and then diluting to 1.25 x 106 cells/ml with PBS. Female BALB/c mice were injected via the lateral tail vein with 2.5 x 105 cells, and then monitored at least twice a day for 28 days. Statistical analyses of the results for the survival studies were carried out using a log rank test (Mantel-Haenszel) with the Prism 6 software program (GraphPad Software, Inc., La Jolla, CA). To assess fungal burden, kidneys were excised, weighed, and then homogenized in 5 ml PBS for 30 s with a tissue homogenizer (Pro Scientific Inc.). The CFU per gram of kidney was determined by plating dilutions of the homogenates on YPD agar medium plates, and incubating for 2 days at 30°C. Statistical analysis of the CFU data was carried out with Prism 6 software using one-way analysis of variance with a nonparametric Kruskal-Wallis test and Dunn’s post-hoc test. For histological analysis, kidneys were excised from mice 2 d post infection, fixed with formaldehyde, and then stained with Hematoxylin and Eosin (H&E) to detect leukocytes or with Gomori-Methenamine Silver (GMS) to detect fungal cells by McClain Laboratories, Smithtown, NY.
The accession numbers for the C. albicans genes used in this study are as follows:
Standard NameSystematic NameOrf19NameGenBank DesignationPST1C2_06870Corf19.2241XP_714771.1PST2C2_08640Corf19.3612XP_714456.1PST3CR_05390Worf19.5285XP_710366.1YCP4CR_05380Corf19.5286XP_710367.1COQ3C6_01840Corf19.3400XP_716710.1
|
10.1371/journal.pbio.0050114 | From Parasite to Mutualist: Rapid Evolution of Wolbachia in Natural Populations of Drosophila | Wolbachia are maternally inherited bacteria that commonly spread through host populations by causing cytoplasmic incompatibility, often expressed as reduced egg hatch when uninfected females mate with infected males. Infected females are frequently less fecund as a consequence of Wolbachia infection. However, theory predicts that because of maternal transmission, these “parasites” will tend to evolve towards a more mutualistic association with their hosts. Drosophila simulans in California provided the classic case of a Wolbachia infection spreading in nature. Cytoplasmic incompatibility allowed the infection to spread through individual populations within a few years and from southern to northern California (more than 700 km) within a decade, despite reducing the fecundity of infected females by 15%–20% under laboratory conditions. Here we show that the Wolbachia in California D. simulans have changed over the last 20 y so that infected females now exhibit an average 10% fecundity advantage over uninfected females in the laboratory. Our data suggest smaller but qualitatively similar changes in relative fecundity in nature and demonstrate that fecundity-increasing Wolbachia variants are currently polymorphic in natural populations.
| Wolbachia are endosymbiotic bacteria that live inside the cells of their invertebrate hosts. They are transmitted directly from mother to offspring, and spread through populations by manipulating the reproduction of their hosts. The most common reproductive manipulation responsible for the spread of these bacteria, called “cytoplasmic incompatibility,” arises when infected males mate with uninfected females, resulting in fewer offspring than normal. There are fitness costs for the hosts associated with Wolbachia infections, most commonly involving a reduction in egg production. Theory predicts that this detrimental effect of Wolbachia on its host should result in selection for the bacteria to evolve a more benign lifestyle, changing the bacterium from being parasitic to more mutualistic. We document such a shift in a Wolbachia infection of fruit flies (Drosophila simulans) from California. The shift occurred extremely rapidly, over 20 years. Consequently, Wolbachia-infected hosts now have higher rates of egg production than their uninfected counterparts. Changes in the genome of Wolbachia seem to be responsible for this, rather than changes in the host genome. Our study reveals that bacteria and their hosts represent components of a dynamic interacting system that can evolve rapidly over time.
| When microbes that live within animal cells are transmitted only maternally, their reproductive success is directly tied to that of the matrilines they inhabit. Both intuition and mathematics suggest that such endosymbionts will be selected towards mutualism, if possible, increasing the fecundity of their female hosts [1]. The expectation that vertical transmission favours evolution towards mutualism is supported by both laboratory co-evolution experiments between viruses and bacteria and comparative data from a wide range of natural associations [1,2]. Mutualisms generally have long evolutionary histories, but given the potentially explosive rate of bacterial evolution [3], rapid evolution of mutualisms in nature might also be expected. Here we report such evolution by bacteria (Wolbachia) associated with a dipteran host (Drosophila simulans) in natural California populations. In less than 20 y, the Wolbachia in California D. simulans have changed so that infected females now produce more eggs than uninfected females under laboratory conditions, whereas infected females previously suffered a significant fecundity deficit.
Cytoplasmic incompatibility (CI) in insects is normally caused when Wolbachia-infected males mate with uninfected females or females that carry a different Wolbachia strain [4]. Because CI causes embryo lethality, infected females, who are protected from CI, often have a reproductive advantage over uninfected females. This results in the rapid spread of the infection through host populations when the Wolbachia are faithfully transmitted from mother to offspring and produce relatively minor fitness costs.
Hoffmann et al. [5] discovered Wolbachia-induced CI in California populations of D. simulans. Initially, the California D. simulans Wolbachia (wRi) infection was found only south of the Tehachapi transverse range in the southern quarter of the state. From 1985 to 1994, we monitored the infection's spread north [6,7]. We showed that the dynamics and equilibrium infection frequencies in nature could be described accurately by a discrete-generation model with only three parameters: μ, the average frequency of uninfected ova produced by an infected female; H, the relative hatch rate from “incompatible” fertilisations of uninfected eggs by sperm from infected males (the other three possible fertilisations produce indistinguishable hatch rates); and F, the relative fecundity of infected females [8–10]. Using replicated field assays in the early 1990s, we found the following: μ ≈ 0.045, H ≈ 0.55, and F ≈ 1.0. In contrast, in laboratory populations, the infection showed perfect maternal transmission (μ = 0), and infected females were 10%–20% less fecund than uninfected females (F ≈ 0.8–0.9). Our field-based parameter estimates produce a predicted equilibrium infection frequency,
p̂ ≈ 0.94, consistent with data from several locations, including three populations where we monitored the infection's introduction and spread over about 2 y, with dynamics that roughly matched our simple predictions [7]. The wRi infection quickly spread northward through California and is now pervasive throughout most North American populations of D. simulans.
Models suggest that both Wolbachia infections and the host nuclear background should evolve to reduce deleterious effects associated with the infection and to increase the transmission fidelity of the microbe [11,12]. Despite the fact that CI allows Wolbachia to spread within populations, intrapopulation selection of Wolbachia is not expected to directly affect the level of CI [11,12], unless host populations are structured so that kin selection favours more intense CI [13] (e.g., when infected males can reduce the larval competition experienced by the progeny of their female siblings, a condition that is likely to be rare for D. simulans). In contrast, host evolution is expected to reduce the intensity of CI (i.e., increase embryo viability from incompatible crosses) [12]. Both Wolbachia and host evolution affecting transmission (μ), level of incompatibility (H), and fecundity of infected hosts (F) are plausible, because both Wolbachia- and host-related effects influence CI, transmission, and fitness [14–16]. Indeed, some Wolbachia infections do not induce detectable levels of CI and have no known deleterious effects on host fitness [17,18]. Positive fitness effects from Wolbachia infections have been suggested [19–22], and indirect comparative evidence from different Wolbachia infections in Drosophila indicates that Wolbachia–host interactions may become more benign and potentially mutualistic over time [4,23].
Southern California populations of D. simulans are natural candidates for observing such evolution, because they have been stably infected for at least two decades, the host populations produce on the order of 10–15 generations per year [7], and the parameter values describing transmission, fecundity, and CI level are such that both the infection and its host should experience significant pressure to evolve [12]. We have informally monitored these populations for evolutionary changes in CI and other Wolbachia effects for about a decade. Here we provide our accumulated evidence that the Wolbachia infection in D. simulans has changed to become more mutualistic, while no evolution by either the Wolbachia or D. simulans has been detected affecting CI levels or the fidelity of maternal transmission.
Apart from the viability effects associated with CI, female fecundity is the only fitness component known to be affected by the wRi infection in California D. simulans [9]. To test whether fecundity effects associated with this Wolbachia infection have evolved, we re-examined female fecundity in California D. simulans after 20 y [5]. In the late 1980s, fecundity costs were evident in the laboratory when lines treated with the antibiotic tetracycline, which cures Wolbachia infections, were compared to untreated lines, or when naturally uninfected and infected lines were compared [9]. Flies were collected in 2002 and uninfected lines were generated by treatment with tetracycline. Infected and uninfected lines were then scored for fecundity every day over 5 d. Overall, there was a fecundity advantage associated with the infection; in some lines no fecundity advantage was detected, while in other lines the total egg output was significantly greater in infected individuals (Figure 1A). In contrast, the wRi line collected from Riverside in 1988 and maintained in the laboratory still produced a fecundity deficit, comparable to the deficits found in infected lines previously (see Table 10 of [8] and Table 7 of [9]). We repeated this fecundity assay, again curing many of the same lines after 10 mo of laboratory culture. To test the sensitivity of the fecundity effect to culture conditions, we restricted access to yeast. Overall, the infected lines still showed greater fecundity than the uninfected lines derived from them. Although one line showed a decrease in fecundity when infected, there was a significant increase overall, and no significant interaction effect (Figure 1B). When individual lines were considered, there appeared to be a shift in fecundity for line R3 and R24 (Figure 1A and 1B). However, when we computed confidence intervals on these data for the difference in fecundity relative to the uninfected lines, the bootstrapped confidence intervals (bCIs) overlapped between, for example, the yeasted R3 line (mean difference 42%, 95% bCI 3.7% to 85%) and non-yeasted R3 line assays (mean difference 0.4%, 95% bCI −6.8% to 20.4%). Hence, the lack of repeatability for the statistical significance of individual lines is likely to reflect primarily the large inherent stochasticity of fecundity data. In contrast, the positive effect of the Wolbachia infection on mean fecundity across lines is evident in both of the experimental treatments reported in Figure 1A and 1B (the mean effect is 23% in Figure 1A and 10% in Figure 1B).
Because Wolbachia effects on fitness may be temporarily or permanently induced by tetracycline treatment [14,24], we re-collected isofemale lines of D. simulans from Irvine, California, in 2004. We then generated uninfected lines from these and reciprocally crossed infected and uninfected sublines two generations after tetracycline treatment (using old males in the incompatible cross) to homogenise nuclear backgrounds and remove the effects of the antibiotic treatment. Fecundity was then assayed each day over 10 d, controlling for body size. Again, a significant overall fecundity advantage of approximately 10% was associated with Wolbachia infection (F1,185 = 8.03, p = 0.005), with three lines out of eleven (IR1, IR2, and IR28) showing a significant fecundity advantage (Figure 1C). Therefore, the fecundity effect associated with the wRi infection in the laboratory has changed from negative to positive. The apparent lack of consistency of this fecundity advantage across individual lines suggests that there may be polymorphism among the Wolbachia strains infecting the populations near Riverside and Irvine. These data do not demonstrate polymorphism because there is no significant line-by-infection interaction in our fecundity assays. However, polymorphism is expected from our theoretical analyses (see “Mathematical Analyses”) and is directly demonstrated from additional data presented below.
Our data indicate that the fecundity deficit initially associated with the wRi infection in laboratory assays [7,9] has disappeared. We computed 95% CIs of the mean difference between infected and uninfected lines following Sokal and Rohlf (p. 444 of [25]) for the 2004 Irvine data and earlier published data (Table 10 of [9]; Table 7 of [7]). The mean difference for infected minus uninfected lines from the 2004 data was 10.4% (95% bCI 2.6% to 18.2%) and for the earlier data was −19.4% (95% bCI −24.5% to −13.6%). Thus, the infection has changed from causing a significant fecundity deficit to causing a significant fecundity advantage in the laboratory (an overall shift of 30%).
Evolutionary changes may also influence CI levels and maternal transmission. To test for changes in CI, we reciprocally mated infected and uninfected individuals derived from isofemale lines of D. simulans collected in 1999 and 2002 and compared these with the old wRi line collected at Riverside in 1988. We also mated females to males that were 5, 10, and 15 d post-eclosion, as male age can affect CI [7,9,26]. When males were 5 d old, levels of CI were high, with an average of 92% egg mortality for all lines in crosses between uninfected females and infected males (Figure 4). CI levels did drop off when males were 10 and 15 d old, as described previously [7,9]. However, there was no difference in the level of CI between the Riverside collections in 1999 and 2002 and the old 1988 wRi line, and the power of our CI tests indicated that we could have detected a difference of 8% or greater. Hence, the level of CI has certainly evolved significantly less than the fecundity effects observed in our laboratory assays.
The levels of CI for matings involving males in the three age classes are also similar to those found previously [7,9]. We assume that the lack of change of CI in our laboratory assays implies relative constancy of CI levels in nature. Hence, we expect that H, the average relative hatch rate from an incompatible fertilisation in nature (sperm from an infected male fertilising an uninfected ovum), remains approximately 0.55, as estimated previously [7].
We did not directly re-measure field maternal transmission rates of the Wolbachia infection. Instead, we used field infection frequencies and our assumption that H remains approximately 0.55 to indirectly estimate transmission efficiency, as any significant increase or decrease in maternal Wolbachia transmission would lead to an observable change in the equilibrium infection frequency. For instance, with H = 0.55, explaining an equilibrium frequency of 0.94 requires a transmission efficiency of 0.964 to 0.953 if the relative fecundity of infected females, F, is between 0.9 and 1.1. Our previous estimates of transmission frequency in nature, denoted 1 − μ, averaged 0.96, consistent with the observed infection frequency and field estimates of CI intensity and relative fecundity [7]. Because selection among mutually compatible Wolbachia variants acts to increase the parameter combination F(1 − μ) [12], we expect μ to decrease. If the failure rate of maternal transmission, μ, was halved to 0.02 (which would have roughly the same fitness impact on Wolbachia as increasing F by only 2%), the expected equilibrium infection frequencies would rise to 0.97, with H = 0.55 and F between 1.0 and 1.1 (see “Mathematical Analyses”).
We collected 654 D. simulans females from four locations in California (Irvine, Ivanhoe, Riverside, and Winters), and screened F1 females from each field-collected female for Wolbachia using PCR. We found that infection frequencies did not differ significantly between the locations, with an overall infection frequency of 93% (Table 1; G test for homogeneity, G = 3.09, df = 3, p = 0.38). These infection frequencies do not differ from those found in earlier studies [7,9], suggesting that transmission of Wolbachia from mother to offspring (1 − μ) has not changed appreciably.
To determine if the observed fecundity advantage of wRi infection is associated with evolution of the wRi strain and not merely a new strain of Wolbachia that has invaded California, we tested for compatibility between the old wRi line collected in 1988 and the IR2 line collected from Irvine in 2004. If Wolbachia strains are incompatible (i.e., cause CI in matings between lines), then it is likely that a new strain has invaded California populations of D. simulans. However, we found no difference between the hatch rates of crosses between and within lines (F3,72 = 0.08, p = 0.97), indicating complete compatibility between the Wolbachia strains. In addition, we sequenced part of the Wolbachia wsp gene (611 bp) for the old wRi strain collected in 1988 and strains from 25 isofemale lines collected at Irvine in 2004 and found no differences at the nucleotide level. We also sequenced part of the Wolbachia ftsZ gene (718 bp) for five of these strains and found no differences in nucleotide sequence compared to the 1988 wRi strain. Therefore, it is likely that the change in fecundity has involved evolution of the wRi strain rather than invasion by a new strain.
Essentially all studies of Wolbachia effects have been limited to laboratory populations. The wRi infection of D. simulans is one of very few whose effects have been studied in nature (cf. [27–29]). We have shown that wRi has apparently evolved in nature over the past 15 y so that laboratory isofemale lines with the infection tend to show a fecundity advantage on the order of 10% rather than a fecundity disadvantage on the order of 20%, as was the case when the wRi infection was new in California D. simulans. We do not yet have new fecundity data from wild-collected females to directly compare fecundities of infected versus uninfected females in nature.
However, we previously reported such data while we tracked the northward spread of wRi through California. During this spread, we were able to study populations in which wRi was at intermediate frequencies, allowing us to make comparisons between the fecundities of wild-collected infected versus uninfected females. We performed nine comparisons, four in the Tehachapi mountains of southern California in 1988–1989 (Table 4 of [9]), then five more near Davis in northern California in 1992–1993 (Tables 6 and 7 of [7]) (all of these data are summarised in Table S1). Collectively these studies assayed the fecundity of more than 1,000 females from nature, and each of our nine comparisons was based on 45–203 females. Only the first of these nine studies (104 females) found a statistically significant fecundity deficit for infected females (F = 0.82, p < 0.05). More relevant to our new laboratory fecundity data is that the four 1988–1989 studies produced an average relative fecundity of F = 0.92, whereas the five 1992–1993 studies produced an average of F = 1.03 (with three of the last five point estimates for F above one). Although the difference between these two sets of estimates is marginally non-significant in a one-tailed test (t-test, t = 1.76, df = 7, p = 0.06), it suggests that the changes in relative fecundity we have documented in the laboratory reflect similar changes in nature. This is discussed further below in light of our theoretical analyses.
In less than 20 y, the wRi strain that invaded and spread throughout California in the 1980s has evolved from inducing a fecundity deficit in the laboratory to providing a fecundity benefit to its host, as theory predicts [12]. There has been no detectable change in the level of CI, indicating that the genes controlling fecundity have at most minor pleiotropic effects on CI. Rapid evolutionary change within this system has resulted in a parasitic Wolbachia evolving towards a more mutualistic interaction with its host. Interspecific comparisons (e.g., [30–32]) and laboratory experimental evolution systems (e.g., [33]) provide many examples supporting the theoretical prediction that vertical transmission, as opposed to horizontal transmission or a mixed mode of transmission, tends to promote mutualism [34–36]. There are well-known examples in which viruses have rapidly evolved to become more benign in nature [37,38]. However, these are best interpreted as evolution towards an “optimal” level of virulence, rather than evolution towards mutualism [34]. We know of no previous examples in which an evolutionary shift towards mutualism has been observed over a period of decades in nature.
To understand the evolutionary dynamics in nature that have so rapidly produced the new fecundity effects, we assume—consistent with our laboratory CI data—that the relevant Wolbachia variants are fully compatible with each other. This implies that within the population of infected individuals, the frequencies of alternative Wolbachia types follow haploid selection dynamics with fitness determined by the parameter combination F(1 − μ), irrespective of whether the variants cause different levels of CI with uninfected individuals [12]. Between 1988 and 2002, the California populations of D. simulans have produced about 200 generations [7]. The observed fecundity variation produced by different Wolbachia on a common genetic background (Figure 3) demonstrates polymorphism for the fecundity-increasing Wolbachia variant(s). Irrespective of within-host dynamics, we can use discrete-generation haploid selection theory to explore the selective pressures responsible for the spread of fecundity-enhancing variants among hosts and their likely evolutionary trajectory (see “Mathematical Analyses”).
Assuming that the observed changes are attributable to increased frequency of variants initially present, but extremely rare, in the 1988 southern California wRi population, our analysis suggests that selective advantages in the field are likely to be on the order of 5% (whereas 1% or 15% are unlikely). Theory also indicates that the current polymorphism should be transient and that the fecundity-enhancing variants should reach very high frequencies in these populations over the next 5–10 y. Hence, we predict that the continuing evolution of these Wolbachia populations will be easily documented.
Our data on compatibility of the “new” versus “old” Wolbachia variants and their DNA sequence similarity indicate that Wolbachia effects on its host evolve readily in natural populations by selection among closely related Wolbachia variants. Such rapid evolution helps to explain the diversity of effects of Wolbachia on host fitness noted in the literature: these effects range from negative [9,15,39] to positive [19,21,22] to the extreme where Wolbachia becomes essential for host survival [20] or host fertility [40,41]. It also helps to explain the inconsistent effects of Wolbachia on host fitness detected in previous experiments [22,42]; changes in the apparent host effects of Wolbachia over time or between experiments may well reflect selection among Wolbachia variants rather than residual effects of antibiotics or changes in Wolbachia density. The rapid evolution of wRi, as well as rapid evolution of Wolbachia hosts [43], suggests a dynamic interaction between parasitic and mutualistic life modes and rapidly changing effects of endosymbionts in host insect evolution.
The CI assays used D. simulans collected from Riverside, California, in 1998 and 2002 and maintained in the lab as isofemale lines until testing. Fecundity assays included the 2002 isofemale lines and those established from females collected at Irvine, California, in 2004. A California wRi-infected line from 1988 was included in some assays. To determine the infection frequencies in California populations, approximately 200 female D. simulans were collected at each of four localities (Irvine, Ivanhoe, Riverside, and Winters) in 2004, and F1 individuals scored for infection status by PCR assay (described below).
To produce uninfected sublines for each line, larvae were treated with 0.03% tetracycline [5] for two generations. Lines were reared for at least two generations without tetracycline before the CI and fecundity experiments.
Level of CI was determined by mating virgin 5-, 10-, and 15-d-old Wolbachia-infected males to uninfected virgin females (>5 d old) from the same 1998 and 2002 collected lines. Reciprocal crosses acted as controls. Males were mated once, and females were placed after mating in a vial with a spoon containing 5 ml of agar-treacle-yeast medium and left for 24 h at 25 °C. The number of unhatched eggs was counted >24 h later.
CI data (egg hatch rates) were angular transformed prior to analysis. Model I ANOVA (analysis of variance) and t-tests were used to compare CI levels between the Riverside collections from 1998 and 2002 and the wRi line from 1988.
Five fecundity experiments were done. In the first two (Figure 1A and 1B), lines from the 2002 Riverside collection were cured, and infected and uninfected females from each line were mated to uninfected males from the same line. In the first experiment, the 1988 wRi line was included to re-test the previously described fecundity deficit [9]. Flies were reared at low densities by placing 20 eggs per vial on 15 ml of medium. To measure fecundity of emerging flies, pairs of 1-d-old virgin females and males were placed in vials with spoons as for the CI tests. Spoons were replaced every 24 h for 5 d and eggs counted. Between ten and 15 females were assayed for each line. Yeast paste was added to the medium surface in the first experiment, but not in the second experiment, to see if the same fecundity-enhancing Wolbachia effect could be detected when egg output was suppressed due to the absence of live yeast.
In the third experiment (Figure 1C), lines from the 2004 Irvine collection were cured as above. To control for nuclear background, we crossed uninfected and infected flies from the same line reciprocally and scored F1 offspring for fecundity (with live yeast) after they had been reared and set up as above. Fecundity scoring was extended from 5 to 10 d to increase the likelihood of detecting small differences. Between 15 and 20 replicates were assayed per infected/uninfected treatment of each isofemale line. Wing size (measured as centroid size based on landmarks [44]) was also measured for each female and used as a covariate in analyses to control for body size.
To assign the effects on fecundity to either Wolbachia or a host–Wolbachia interaction, we backcrossed the nuclear background of one Irvine line showing the greatest fecundity advantage (Figure 1C; IR2) into the 1988 wRi strain, and the 1988 wRi line nuclear background into the IR2 strain, both for five generations (Figure 2). Ten-day fecundity was measured on 20 replicate pairs of males and females per backcross line as above. Wolbachia strain and nuclear background were treated as fixed effects in the ANOVA for fecundity.
Finally, to determine whether the Wolbachia fecundity effect was polymorphic within the 2004 Irvine lines, we backcrossed the 1988 wRi line nuclear background into 20 strains (isofemale lines) from the 2004 Irvine collection for two generations (Figure 3). Ten-day fecundity was measured on 20 replicate pairs of males and females as above. Model I ANOVA was used to determine Wolbachia strain differences for fecundity. We also determined the coefficient of variation [25] for the lines in this experiment and the infected and uninfected lines from the third fecundity experiment (2004 Irvine lines) to see if they fitted the patterns expected (fecundity of infected > fecundity of uninfected > fecundity of infected with a homogenised background).
We determined the infection status of all lines collected from the field or after treatment with tetracycline using extracted DNA from females in a PCR with the Wolbachia-specific primers 76–99 forward and 1012–994 reverse which amplify a ~ 950-bp fragment of bacterial 16S rDNA [45]. The D. melanogaster primers su(s) forward 724–753 and su(s) reverse 1113–1092 were included in each reaction as a control [7].
To determine the infection frequency in the four populations collected from California in 2004, we first assayed DNA as above from a single F1 female from each field-collected female. In addition, another PCR with the Wolbachia-specific primers wsp81F and wsp619R [46] was performed with the same DNA to minimise the chance of false positives. If either or both of these assays were negative, DNA was extracted from a second F1 fly from the same isofemale line and subjected to the two PCRs. This second fly was used to control for PCR artefacts and imperfect maternal transmission of Wolbachia [7,9].
To determine compatibility between the 1988 wRi line and the IR2 line collected at Irvine in 2004, we crossed virgin males and females (>5 d old) between and within each line in a reciprocal design. Males were mated once, and females were placed after mating in a vial with a spoon containing medium as above for the CI assays. The number of unhatched eggs was counted after 48 h. The analysis was as above for the CI assays.
To compare the similarity between the 1988 wRi strain and the Irvine strains collected in 2004, we sequenced 611 bp of the highly variable Wolbachia wsp cell-surface gene [46] and 718 bp of the Wolbachia ftsZ cell-cycle gene. DNA was extracted from a single female from each of 25 isofemale lines from the 2004 Irvine collection and the laboratory line of the 1988 wRi strain. The partial wsp gene fragment was amplified from all lines using the primers and protocol found in [46]; the partial ftsZ gene was amplified from only five isofemale lines from the 2004 Irvine collection and the 1988 wRi line as in [47]. Amplified fragments were sequenced using the BigDye Terminator cycle sequencing kit (v3.1, Applied Biosystems, http://www.appliedbiosystems.com). Sequences were aligned using the CLUSTAL W algorithm [48]. We also included in the analysis the original wsp and ftsZ sequences of the wRi strain found in GenBank (http://www.ncbi.nlm.nih.gov/Genbank; accession numbers AF020070 and U28178, respectively).
We analysed various mathematical models to address three issues discussed in the text: (1) inferences concerning transmission-rate evolution based on the dependence of equilibrium infection frequencies on the three parameters that are sufficient to explain dynamics and equilibria in nature [7], (2) the intensity of selection responsible for the observed evolution, and (3) predicted future frequency changes in the fecundity-enhancing Wolbachia variant(s). Our methods and analyses leading to our conclusions are described below.
Regarding dependence of equilibria on parameter values, to make inferences concerning whether Wolbachia's maternal transmission rate has evolved, we first considered how the stable equilibrium infection frequency, denoted
p̂, changes with the parameters H, the relative hatch rate from incompatible crosses, F, the relative fecundity of infected versus uninfected females, and μ, the fraction of uninfected ova produced by an infected female. Based on field estimates of infection frequencies, we concluded that the equilibrium frequency throughout central and southern California in 1992 was approximately
p̂ ≈ 0.94 (with 95% confidence interval 0.92 to 0.96) [7]. This was consistent with our theoretical prediction for
p̂ from field-based parameter estimates [7]. Our new laboratory data suggest that F has evolved significantly, while H has remained relatively constant. Given the change in F, it may seem surprising that our new estimate of the infection frequency in central and southern California, approximately 93% (with 95% confidence interval 0.90 to 0.94), does not differ significantly from the frequency estimated previously. Our formula for
p̂ allows us to examine the consistency of these observations.
Evolutionary theory suggests that if Wolbachia variants remain fully compatible, F(1 − μ) should tend to increase [12]. Thus, we are particularly interested in determining whether μ has decreased. However, because fitness is proportional to F(1 − μ), changing F from 0.9 to 1.0 or from 1.0 to 1.1 involves a selection coefficient on the order of s = 0.1 (which should produce significant changes in polymorphic Wolbachia variant frequencies over tens of generations). In contrast, halving μ from 0.04 to 0.02 involves much weaker selection, on the order of s = 0.02, so that hundreds of generations would be required for significant evolution. Hence, we expect that detectable evolutionary changes since the mid-1980s in μ are much less likely than detectable changes in F.
Figure 5A and 5B explore how varying F changes the values of H and μ necessary to explain equilibrium population frequencies of 0.90 or 0.94. Figure 5 assumes μ = 0.04 and shows the values of H needed to produce
p̂ = 0.90 (solid line) versus 0.94 (dashed line) as F varies from 0.9 to 1.1. As shown, varying F over this range requires very little change in H to preserve
p̂. Similarly, Figure 5B assumes H = 0.55 and shows the values of μ needed to produce
p̂ = 0.90 (solid line) versus 0.94 (dashed line) as F varies from 0.9 to 1.1. Again, changing F has little effect. Both graphs indicate that changes in F are likely to have little impact on
p̂. This is shown directly in Figure 5C, which assumes H = 0.55 and μ = 0.045 (or μ = 0.0225) and plots
p̂ as F varies from 0.9 to 1.1. Clearly, changes in F over the range suggested by our laboratory and field data have little impact on
p̂. In contrast, a change in μ that would have a much smaller impact on Wolbachia fitness would produce changes in
p̂ that our samples would have detected.
Regarding selection intensity, within the population of infected individuals, the frequencies of mutually compatible Wolbachia variants follow haploid selection dynamics with the fitness of each variant proportional to F(1 − μ), irrespective of the level of CI they produce with uninfected females [12]. All else being equal, two conclusions follow: (1) the level of CI is not subject to direct selection based on between-host frequency dynamics, and (2) for values of F near one and μ near zero (as suggested by our data), selection for modifying F is much stronger than selection for modifying μ. These inferences are consistent with our data, which suggest that H and μ have remained relatively constant, while F has increased.
To make quantitative inferences, we assumed discrete generations. If we consider two Wolbachia variants such that F1(1 − μ1)/[F2(1 − μ2)] = 1 + s, the frequency of variant 1 in generation t, denoted pt, changes according to
Our inferences about plausible selection intensities follow from equation 1, assuming that the observed changes have occurred over roughly 200 generations.
Figure 6 illustrates the selection intensity needed to explain a transient polymorphism, assuming that the fecundity-enhancing variant was rare in the population 200 generations ago. It shows that for low initial frequencies (say, between 10−3 and 10−6), selection coefficients, s, on the order of 0.04–0.08 would produce polymorphic frequencies (between 0.1 and 0.8, for instance) that are consistent with our data. In contrast, for s on the order of 0.01, a significantly longer time would be necessary to produce the polymorphism observed (see Figure 7), whereas if s were as large as 0.15, polymorphism for the fecundity-enhancing variant would tend to be very short-lived (on the order of 2 y).
Regarding future evolution, our analysis suggests that the currently inferred polymorphism for the fecundity-enhancing variant(s) is likely to be transient. We can use equation 1 to understand the time scale over which near-fixation is expected. Figure 7 plots the time required for a favoured variant to increase from an initial frequency of 0.001 up to a frequency of between 0.1 and 0.9. The difference between the highest and lowest lines indicates the time required for the frequency to increase from 0.1 to 0.9. Note that for s as large as 0.15, this time is only about 31 generations, or approximately 2 y in these populations. Given that samples collected in 2002 and 2004 both showed an apparent polymorphism, it seems unlikely that selection was this intense. This inference is consistent with our conjecture that the current Wolbachia-induced fecundity advantage in nature is likely to be less than the 10% effect observed in the laboratory, just as the fecundity deficit of roughly 15% found in the laboratory in 1989 [9] corresponded to a fecundity deficit for infected field-collected females that was probably less than 10% in 1989 [9] and less than 8% in 1992 [7]. Conversely, if the fecundity advantage was as small as 1% (corresponding to s = 0.01), as Figure 7 shows, the inferred polymorphism would be unlikely to arise in only 15 y (about 200 generations). Hence, for plausible levels of selection, we are likely to be able to observe significant frequency increases of the fecundity-enhancing Wolbachia variant(s) in nature over the next few years.
Sequences for the wsp and ftsZ genes sequenced in this study have been deposited in GenBank (http://www.ncbi.nlm.nih.gov/Genbank) under the accession numbers EF423730–EF423735 for ftsZ and EF423736–EF423761 for wsp.
|
10.1371/journal.pcbi.1006605 | Binding of the general anesthetic sevoflurane to ion channels | The direct-site hypothesis assumes general anesthetics bind ion channels to impact protein equilibrium and function, inducing anesthesia. Despite advancements in the field, a first principle all-atom demonstration of this structure-function premise is still missing. We focus on the clinically used sevoflurane interaction to anesthetic-sensitive Kv1.2 mammalian channel to resolve if sevoflurane binds protein’s well-characterized open and closed structures in a conformation-dependent manner to shift channel equilibrium. We employ an innovative approach relying on extensive docking calculations and free-energy perturbation of all potential binding sites revealed by the latter, and find sevoflurane binds open and closed structures at multiple sites under complex saturation and concentration effects. Results point to a non-trivial interplay of site and conformation-dependent modes of action involving distinct binding sites that increase channel open-probability at diluted ligand concentrations. Given the challenge in exploring more complex processes potentially impacting channel-anesthetic interaction, the result is revealing as it demonstrates the process of multiple anesthetic binding events alone may account for open-probability shifts recorded in measurements.
| General anesthetics are central to modern medicine, yet their microscopic mechanism of action is still unknown. Here, we demonstrate that a clinically used anesthetic, sevoflurane, binds the mammalian voltage-gated potassium channel Kv1.2 effecting a shift in its open probability, even at low concentrations. The results, supported by recent experimental measurements, are promising as they demonstrate that the molecular process of direct binding of anesthetic to ion channels play a relevant role in anesthesia.
| Volatile and injected general anesthetics encompass a diverse array of small and uncharged chemotypes including haloalkanes, haloethers and alkylphenols. Despite efforts reaching back over a century, clarification of their microscopic mechanism in general anesthesia has proven difficult and wanting. A favored hypothesis proposes that ion channels in the brain are implicated, among which members of ionotropic neurotransmitter receptors, voltage-gated and non-gated ion channels are best-known players [1–3]. Primary exemplars are the Cys-loop nicotinic acetylcholine and γ-aminobutyric acid class A receptors, the voltage-gated sodium and potassium channels, and the tandem pore potassium channels. An extensive series of electrophysiological studies corroborate the hypothesis by demonstrating a range of effects, from inhibition to potentiation, of general anesthetics on the various receptor targets. Beyond these electrophysiological studies of reductionist systems, the current view has gained additional support from gene knockout experiments demonstrating for some of these channels the in vivo role on a clinically-relevant anesthetic outcome. For instance, the knockout of the non-gated tandem pore potassium channel trek-1 produces an animal model (Trek1-/-) resistant to anesthesia by inhalational anesthetics [4].
How general anesthetics modulate ion channels to account for endpoints of anesthesia must at some point build on understanding electrophysiological data in the context of ligand binding, a reasoning that has driven mounting efforts in the field. Currently, though not refuting other molecular processes likely to contribute to anesthetic action [5–7], crystallography and molecular dynamics studies support that anesthetics bind ion channels at clinical concentrations [8–16]. Binding interactions have been evidenced in anesthetic containing systems of mammalian voltage and ligand-gated channels, as well as bacterial channel analogs. Specifically, partitioning of anesthetics in the membrane core allows it to access and bind multiple transmembrane (TM) protein sites, featuring single or multiple occupancy states–a process that might depend further on chemotypes, channel types and conformations. Although some progress has been made in validating one or more aspects of the direct-site hypothesis, a first-principle demonstration that anesthetics bind ion channels to affect protein equilibrium and function as recorded in experiments is still unaccounted for.
Here, we focus our efforts on the haloether sevoflurane and its molecular interaction to Kv1.2, a mammalian voltage-gated potassium channel. Experimental work demonstrates that sevoflurane potentiates the channel in a dose-dependent manner [3,17,18]. Effects on current tracings include a leftward shift in the channel’s conductance-voltage relationship and an increased maximum conductance. As extensively discussed in these past publications, at least two molecular mechanisms are expected to be involved in Kv channels potentiation by sevoflurane. One mechanism (i) might involve sites allosterically coupled to the electromechanical transduction directly responsible for controlling voltage-dependent gating. The other (ii) might involve distinct sites, which could modulate the channel’s pore region and influence the stability of the conductive state and/or the unitary conductance. Here, we are interested in the investigation of mechanism (i) and its underlying structural hypothesis that sevoflurane binds the channel’s open and closed states to impact protein equilibrium and therefore its voltage dependence. Among all other aspects that might impact channel-anesthetic interactions in general, our specific goal is to determine if sevoflurane binds the well-characterized open-conductive (O) and resting-closed (C) structures of Kv1.2 [19,20] in a conformation-dependent manner to impact its voltage-dependent open probability as recorded experimentally. Very recently, we have put forth an innovative structure-based study [21] dealing with the concentration-dependent binding of small ligands to multiple saturable sites in proteins to show that sevoflurane binds the open-pore structure of Kv1.2 at the S4S5 linker and the S6P-helix interface–a result largely supported by independent photolabeling experiments [22,23]. To our current goal, we aim therefore at extending these calculations to investigate sevoflurane interactions with the entire channel TM-domain and, more importantly, to resolve any conformational dependence in its binding process to channel structures C and O. In the following sections, we first provide the theoretical framework to study sevoflurane binding to a specific channel conformation under equilibrium conditions. A state-dependent strategy is put forward to describe anesthetic binding in terms of occupancy states of all identified potential channel binding sites, embodying both concentration and multiple sites saturation effects. The strategy is then generalized to account for ligand impact in the C-O equilibrium, allowing for reconstruction of voltage-dependent open probabilities of the channel at various ligand concentrations. Anticipating our results, we find that sevoflurane binds Kv1.2 structures at multiple sites under saturation and concentration effects. Despite a similar pattern of molecular interactions, binding of sevoflurane is primarily driven towards the open-conductive state shifting leftward the open probability of the channel at diluted ligand concentrations.
We applied large-scale and flexible docking calculations to solve sevoflurane interactions to Kv1.2 structures X ≡ {C, O} (Fig 1). A total of ~ 6,000 docking solutions was generated per channel conformation and clustered into 21 ligand interaction sites. The interaction sites spread over the transmembrane region of the channel at the S4S5 linker, S6P-helix interface and at the extracellular face, next to the selectivity filter. Further docking sites were resolved within the voltage-sensor, at the S4Pore interface and within the channel central cavity. Re-docking of sevoflurane generated in turn a total of ~ 13,000 solutions per channel conformation, solving the interaction of two ligands for all sites but the extracellular face.
From the docking ensembles, there are up to 2 × 321 channel occupancy states that might contribute to sevoflurane binding and functional effects. To quantitatively evaluate this, we performed an extensive series of decoupling FEP calculations to estimate the per-site binding affinity for one and two bound ligands against the channel structures (cf. Computational Methods for details). Here, FEP calculations started from equilibrium ligand-bound channel structures, embedded in an explicit water-membrane environment (cf. RMSD analysis in S1 Fig). For the purpose of improving statistics, FEP estimates and the associated statistical errors were determined from at least two independent decoupling runs. Calculations were performed over ~ 7.0 ns per replica, per site, per conformation, to converge FEP estimates; in a total MD simulation time of ~ 2.0 μs. S2 Fig shows the effectively sampled configuration space in FEP calculations for each of the channel structures. Systematic errors related to lack of site rehydration or relipidation during ligand decoupling were ruled out in S3 Fig showing equilibrium-like lipid or water coordination numbers of the channel structure at the final stages of FEP. Under these technical details, per-site equilibrium binding constants were quantified relative to a homogeneous and diluted aqueous solution occupied by ligands, with an excess chemical potential of μ¯=0.10±0.09kcal.mol−1. As shown in S1 and S2 Tables, per-site binding constants are heterogeneous and take place over a diverse range, i.e. 10−8 (mM-1) -10+2 (mM-2). There is however a decreasing trend of affinities involving sites respectively at the S4S5 linker, S4Pore and S6P-helix interfaces, voltage sensor, central cavity and extracellular face.
To determine if sevoflurane binds channel structures X ≡ {C, O} at clinically relevant concentrations, we computed binding probabilities ρX(n1,…,ns) for dilute concentrations of the ligand in solution, i.e. 1mM, 10mM and 100mM. Equilibrium constants KX(n1,…,ns) for every occupancy state of the channel were then reconstructed from the per-site affinities to determine state probabilities via eq (2). Here, estimates of KX(n1,…,ns) were determined for the condition of independent binding sites, as the minimum site-to-site distances of ~15 Å demonstrated their non-overlap distributions in each of the channel structures (cf. Computational Methods for details). At low 1mM concentration, ρX(n1,…,ns) is largely dominated by the empty state probability ρX(01,…,0s) indicating only a small fraction of bound states with non-negligible occurrences (S4 Fig). Within this fraction, the most likely states involve single occupancy of the S4S5 linker or the S4Pore interface as shown by the marginal probabilities ρX(nj) of individual sites (Fig 2). At higher concentrations, there is a clear shift of ρX(n1,…,ns) towards channel occupancy states that significantly enhance the average number of bound ligands. Careful inspection of ρX(nj) confirms the major relevance of sites at the S4S5 linker and S4Pore interface over the entire concentration range, accompanied by an increasing importance of binding regions at the S6P-helix interface. In contrast, ρX(nj) for sites within the voltage-sensor, central cavity and nearby the extracellular face of the channel remain negligible over all concentrations. For completeness, note in S1 Table that equilibrium constants for doubly-occupied sites are comparable to or even higher than estimates for one-bound molecule thus revealing important saturation effects in which one or two sevoflurane molecules can stably bind the channel structures at individual sites. The result is especially true for spots at the S4S5 linker and S4Pore interface.
The complex distributions of the multiple occupied states of structures X ≡ {C, O} were described in three dimensions by mapping ρX(n1,…,ns) into the position-dependent density ρXj(R) of sevoflurane in each binding site j (cf. Computational Methods for details). As shown in Fig 3 and supplementary S1 and S2 Movies, the density of sevoflurane better convey the results by showing the spatially-mapped concentration dependent population of bound ligands. Projection of ρXj(R) along the transmembrane direction z of the system, ρXj(z), stresses further the results. Note from ρXj(R) that sevoflurane binds channel structures in a concentration dependent manner, binding preferentially the S4S5 linker and the interfaces S4Pore and S6P-helix over a range of concentrations.
So far, our calculations demonstrate that sevoflurane binds Kv1.2 structures over a spectrum of concentrations, preferentially at the linker S4S5 and at the segment interfaces S4Pore and S6P-helix. From a physical-chemical point of view, spots at these channel regions are primarily dehydrated, lipid accessible, amphiphilic pockets providing with favorable interaction sites for the polar lipophilic sevoflurane molecule (S5 Fig). It is worth mentioning that these findings recapitulate recent photolabeling experiments demonstrating that photoactive analogs of sevoflurane do interact to the S4S5 linker and at the S6P-helix interface of the open-conductive Kv1.2 channel [22,23]. In detail, Leu317 and Thr384 were found to be protected from photoactive analogs, with the former being more protected than the latter. As shown in S6 Fig, atomic distances of bound sevoflurane to these amino-acid side chains are found here to be respectively 7.28±2.5 Å and 10.44±3.66 Å, in average more or less standard deviation. Such intermolecular distances are consistent with direct molecular interactions and therefore consistent with the measured protective reactions–similar conclusions hold for the closed channel as well. Besides that, our calculations recapitulate the stronger protection of Leu317 in the sense that, relative to sites at S6P-helix, the affinity of sevoflurane is found to be higher at the S4S5 linker considering its stable occupancy either by one or two ligands. The stable occupancy of the linker by one or two ligands as computed here, is also consistent with recent flooding-MD simulations of the homologous sodium channel NaChBac [14,24] and more importantly, with previous Ala/Val-scanning mutagenesis showing a significant impact of S4S5 mutations on the effect of general anesthetics on members of the K+ channel family [10]. In particular, a single residue (Gly329) at a critical pivot point between the S4S5 linker and the S5 segment underlies potentiation of Kv1.2 by sevoflurane [18]. Sevoflurane is found to be close to that amino acid when bound to the S4S5 linker.
In contrast to the aforementioned spots, sites within the voltage-sensor, within the main pore and nearby the extracellular face of the Kv1.2 structures are primarily hydrated, lipid-inaccessible, amphiphilic pockets (S5 Fig) that weaken sevoflurane interaction as reflected in the state- and space-dependent densities shown in Figs 2 and 3. The binding probabilities at these sites thus support that a non-negligible fraction of poses determined from docking (Fig 1D) corresponds to low affinity or false positives. In particular, because sevoflurane induces potentiation rather than blocking of Kv1.2 [17,18], we read the negligible or absent density of the ligand in the channel central-cavity as a self-consistent result of the study–especially for the open-conductive state. Supporting that conclusion, note that binding constants as computed here are upper bounds for the affinity of sevoflurane under ionic flux conditions in which potentiation takes place. Accordingly, as shown in S7 Fig, the binding affinity of a potassium ion at the central cavity overcomes that of sevoflurane due its binding and excess free-energies under applied voltages. Once bound, the ion destabilizes sevoflurane interactions and the molecule is not expected to bind the channel cavity at low concentrations. As also shown in S7 Fig and supplementary S3 Movie, even under the occurrence of rare binding events, sevoflurane appears unable to block the instantaneous conduction of potassium which is also consistent with its potentiating action.
Weak interactions at the main pore and nearby the selectivity filter of Kv1.2 contrasts with sevoflurane binding at analogous regions of NaChBac [14,24], likely due major structural differences between Na+ and K+ channels. Specifically, the pore of potassium channels lacks lipid-accessible open-fenestrations of the sodium relatives and K+-selective filters are sharply distinct from Na+-selective ones.
Despite a comparable pattern of molecular interactions, careful inspection of ρX(nj) or ρXj(R) reveals that for most sites there is an obvious differential affinity of sevoflurane across Kv1.2 structures (Figs 2 and 3). The overall consequence for sevoflurane binding is then clear: the average number of ligands bound to the open-conductive channel systematically exceeds that for the resting-closed channel over the entire concentration range. There is therefore a remarkable conformational dependence for the anesthetic interaction, such that sevoflurane preferentially binds the open-conductive structure.
Implications for Kv1.2 energetics were then investigated by quantifying changes to the channel open probability ρO(V) induced by sevoflurane at concentrations of 1mM– 100mM (Fig 4). Specifically, from the partition functions ZC(n1,…,ns) and ZO(n1,…,ns) across the entire ensemble of occupancy states of the channel, solution of eqs (5) and (9) show that sevoflurane shifts leftwards the open probability of Kv1.2 in a concentration-dependent manner–voltage shifts amount from -1.0 mV to -30.0 mV with concentration increase of the ligand in solution. The result is particularly interesting, supporting that the approximately 3 mV probability shift recorded experimentally at 1 mM sevoflurane concentration can be explained by our structure-based probability predictions in the concentration range of 1–10 mM. The latter thus provides a theoretical basis to predict sevoflurane impact on channel energetics in a larger, not yet experimentally probed, range of concentrations. Additionally, for a fixed ligand concentration (100 mM), decomposition analysis reveals further that ratio values for the partition functions at individual sites j can be smaller, equal or larger than unity, implying a non-trivial interplay of conformation-dependent modes of action involving distinct sites (cf. Computational Methods for details). In detail, binding of sevoflurane at low affinity sites within the voltage-sensor, central cavity and next to the extracellular face of the channel are mostly conformation-independent and do not impact open probability (ratio ≈ 1). On the other hand, conformation-dependent binding of sevoflurane to sites at the S4S5 linker and the S4Pore interface accounts for the overall stabilization of the open channel (ratio < 1). That effect contrasts with the mild stabilization of the closed conformation of Kv1.2 induced by binding of sevoflurane at S6P-helix and reflected in rightward shifts of ρO(V) (ratio > 1). The overall conformation-dependent binding process is therefore differentially encoded across distinct channel regions.
As extensively discussed in past publications [3,17,18], potentiation of Kv1.2 by sevoflurane has been attributed to stabilization of the open-conductive state of the channel via at least two molecular mechanisms. One mechanism (i) likely involving sites allosterically coupled to the electromechanical transduction responsible for controlling voltage-dependent gating; and another (ii) implicating distinct sites, which could influence the pore conductive state stability and/or unitary conductance. Here, our structural calculations based on the open-activated and resting-closed states of Kv1.2 were consistently designed to investigate the first (i) of these mechanisms. Given the critical role of S4 and S4S5 linker on the channel gating mechanism [19], it is reasonable that sevoflurane interactions with these segments, as found here, are at the origins of the experimentally measured voltage-dependent component of anesthetic action. While restricted to sevoflurane interactions with the resting-closed and open-conductive structures, the presented two-state binding model only embodies left or rightward shifts in the open probability of the channel, therefore it cannot clarify any molecular process accounting for the maximum conductance increase recorded experimentally. As supported by a recent kinetic modeling study [17], generalization of eqs (9) to include a third non-conducting open state yet structurally unknown is needed to account for such conductance effects and for that reason, the investigation of mechanism (ii) is beyond the scope of our study. We speculate however that binding of sevoflurane at the S4Pore and S6P-helix interfaces could allosterically interfere with pore domain operation, thus affecting channel’s maximum conductance. A working hypothesis also raised in the context of anesthetic action on bacterial sodium channels [12,14], assumes indeed that non-conducting states of the selectivity filter are implicated. Corroboration of such an assumption from a molecular perspective is however not trivial and will necessarily involve further structural studies to demonstrate how ligand binding might impact non-conducting open states of the channel to affect maximum conductance.
Here, we carried out extensive structure-based calculations to study conformation-dependent binding of sevoflurane to multiple saturable sites of Kv1.2 structures X ≡{C, O} under equilibrium conditions–the total MD simulation time was ~2.0 μs. Binding of sevoflurane was studied for ligand concentrations in the range of 1mM–100mM and saturation conditions up to njmax=2. Our study relied on the assumption that molecular docking calculations performed in vacuum can faithfully describe ligand interactions at protein sites. Specifically related to that assumption, we have considered the generated ensemble of docking solutions to estimate the location of binding sites δVj and the local distribution of the ligand ρX(R|nj). The generation of false positive hits is however a well documented drawback of docking algorithms as a result of limitations of the scoring function in describing ligand solvation energies and protein flexibility [25]. Given the same limitations of the scoring function, it is also not guaranteed that neither all binding hits, nor that ρX(R|nj) can be accurately known from docking. In this regard, although not considered here, it might be important to integrate docking results from different algorithms involving different scoring functions in order to characterize the bound ensemble. Still, thanks to the generality of the presented formulation, extension of the current investigation to sampling techniques other than docking, including all-atom flooding-MD simulations [9,11,12,14,16], might also be an important refinement in that direction.
Despite these sampling improvements that may eventually be obtained, the presented combination of extensive docking calculations against an ensemble of equilibrium receptor structures fit to handle protein flexibility, and FEP calculations based on fine force-fields to accurately estimate solvation energies are critical technical aspects of the applied methodology devised to minimize such drawbacks. Whereas docking was performed in vacuum, FEP calculations were carried out in presence of explicit all-atom lipids and water, therefore taking into consideration environmental effects in the estimated ligand binding free energies. In this regard, standard binding free-energies estimated from FEP (S1 and S2 Tables) are comparable or significantly smaller than the standard free energy for transferring the ligand from water to a pure lipid bilayer–supporting that sevoflurane is expected to partition preferentially into channel sites rather than the membrane. The partition coefficient (log K) and the related transfer free energy of sevoflurane between water and the lipid bilayer (POPC) amount respectively to 2.64±0.96 and –3.12±0.32 kcal.mol-1 as recently estimated by Tajkhorshid and coworkers [26]. Besides that, it is also important to note that the configuration space in FEP calculations overlap between channel structures at individual sites, i.e. sampling and binding affinities were evenly resolved between states (S2 Fig and Fig 4B)–meaning eventual biases or systematic errors were mitigated when comparing similar calculations between channel states according to the main goal here.
Under these technical considerations, we conclude that most of the identified binding sites are located nearby flexible protein regions for which the root-mean-square deviation between channel structures is larger than 4.0 Å. Then for the purpose of quantifying any direct ligand effect on channel energetics, the determined conformational dependence of sevoflurane binding to these gating-implicated protein regions appears robust and likely to impact function. Structural knowledge allied to solid electrophysiological data available for Kv1.2 make this channel an interesting model system for molecular-level studies of anesthetic action thereby justifying our choice. In detail, the atomistic structures account for most of the available experimental data characterizing closed and open conformations of the channel in the native membrane environment [20]. Previous findings support further that sevoflurane binds Kv1.2 to shift leftward its voltage-dependence and to increase its maximum conductance in a dose-dependent manner [18]. Despite a similar pattern of interactions, we found here a clear conformational dependence for sevoflurane binding at multiple channel sites. The ligand binds preferentially the open-conductive structure to impact the C-O energetics in a dose-dependent manner as dictated by the classical equilibrium theory for chemical reactions embodied in eq (9). Front of the difficulty in conceiving and characterizing other, still more complex molecular processes that might impact channel energetics under applied anesthetics [5–7], the result is revealing by showing that in principle the isolated process of sevoflurane binding to Kv1.2 accounts for open-probability shifts as recorded in experiments. Within this scenario, the calculations reveal contrasting per-site contributions to the overall open probability of the channel. For instance, at 100mM concentration, binding of sevoflurane at the S4S5 linker and S4Pore interface significantly stabilizes the open structure of the channel overcoming the mild stabilization of the closed structure by ligand binding at the S6P-helix interface. By showing this non-trivial interplay of conformation-dependent modes of action involving distinct binding sites, the result is particularly insightful and should guide us to design novel site-specific mutagenesis and photolabeling experiments for further molecular characterization of anesthetic action.
Although not addressing the paucity of in vivo experimental evidences that a binding process to a specific molecular target as presented here is related to any clinically-relevant anesthetic outcome, our study adds support to the direct-site hypothesis by linking binding free-energy and protein energetics. As such, our study treats and reveals a new layer of complexity in the anesthetic problem that brings us novel paradigms to think their molecular action and to design/interpret research accordingly. To the best of our knowledge, the main-text Figs 3 and 4 represent in the context of structural studies, a deeper and first revealed view on the intricate mode of interactions that might take place between general anesthetics and ion channels to impact function in general.
A procedure was designed to solve the molecular binding of sevoflurane to the open-conductive (O) and resting-closed (C) structures of Kv1.2 for saturation conditions up to njmax=2, under equilibrium conditions. For both channel structures, the procedure consisted of (i) an extensive production of docking solutions for the ligand-receptor interaction, (ii) clustering of docking solutions into binding sites along the receptor structure and (iii) estimation of binding affinities using the free-energy perturbation (FEP) method. First completion of steps (i) through (iii) solved the ligand channel interaction for singly-occupied binding sites. Double occupancy of the receptor sites was investigated by inputting the first generated ensemble of docked structures into another round of (i) through (iii) calculations. In detail, step (i) was accomplished by docking sevoflurane as a flexible ligand molecule against an MD-generated ensemble of membrane-equilibrated structures of the protein receptor. Docking calculations included the transmembrane domain of the channel, free from the membrane surroundings. Step (ii) provided the location of δVj volumes lodging docking solutions for the ligand along the channel structures. Each of these volumes was treated as binding site regions in FEP calculations. Here, FEP calculations started from equilibrium structures of the ligand-bound channel embedded in a explicit water-membrane environment. For the purpose of improving statistics, FEP estimates and the associated statistical errors were determined from at least two independent decoupling runs. Calculations were performed over ~ 7.0 ns per site, per conformation and per replica to converge FEP estimates, in a total MD simulation time of ~ 2.0 μs. Following this procedure, binding constants KX(n1,…,ns) for channel structures X ≡ {C, O} were solved by inputting FEP estimates into eq (1), allowing for direct solution of state-dependent probability distributions ρX(n1,…,ns) via eq (2). Here, binding constants KX(n1,…,ns) and the related standard binding free energies ΔGXo(n1,…,ns) were solved for the condition of independent binding sites and relative to a homogeneous and diluted aqueous solution occupied by ligands at constant density ρ¯ and excess chemical potential μ¯. Ligand-free and ligand-bound open probability curves ρX(V) were respectively computed from eqs (5) and (9) by taking into consideration the previously determined mid-point voltage and steepness of the open probability curve of Kv1.2 free of ligands, i.e. Vm = −21.9 mV and ΔQ = 3.85eo as determined from best two-state Boltzmann fit of the measured conductance-voltage data of the channel [18]. Note that any ligand-induced shift in eq (9) is determined by the partition function ratio between open and closed structures and not by the choice of these reference parameters. Probabilities ρX(n1,…,ns) and ρX(V) were determined for sevoflurane concentrations in the range of 1mM–100mM (or in density units, 6.02x10-7Å-3–6.02x10-5Å-3).
|
10.1371/journal.pgen.1005849 | Accounting for Population Structure in Gene-by-Environment Interactions in Genome-Wide Association Studies Using Mixed Models | Although genome-wide association studies (GWASs) have discovered numerous novel genetic variants associated with many complex traits and diseases, those genetic variants typically explain only a small fraction of phenotypic variance. Factors that account for phenotypic variance include environmental factors and gene-by-environment interactions (GEIs). Recently, several studies have conducted genome-wide gene-by-environment association analyses and demonstrated important roles of GEIs in complex traits. One of the main challenges in these association studies is to control effects of population structure that may cause spurious associations. Many studies have analyzed how population structure influences statistics of genetic variants and developed several statistical approaches to correct for population structure. However, the impact of population structure on GEI statistics in GWASs has not been extensively studied and nor have there been methods designed to correct for population structure on GEI statistics. In this paper, we show both analytically and empirically that population structure may cause spurious GEIs and use both simulation and two GWAS datasets to support our finding. We propose a statistical approach based on mixed models to account for population structure on GEI statistics. We find that our approach effectively controls population structure on statistics for GEIs as well as for genetic variants.
| Although genome-wide association studies (GWASs) have discovered numerous novel genetic variants associated with many complex traits and diseases, those genetic variants typically explain only a small fraction of phenotypic variance. Factors that account for phenotypic variance include environmental factors and gene-by-environment interactions (GEIs). Recently, several studies have conducted genome-wide gene-by-environment association analyses and demonstrated important roles of GEIs in complex traits. One of the main challenges in these association studies is to control effects of population structure that may cause spurious associations. In this paper, we show both analytically and empirically that population structure may cause spurious GEIs and use both simulation and two GWAS datasets to support our finding. We propose a statistical approach based on mixed models that can effectively correct for population structure when searching for GEIs.
| Over the past decade, genome-wide association studies (GWASs) have been a predominant approach to identify genetic variants involved in many complex traits and diseases.[1–3] While GWASs have discovered associations of many genetic variants, a large proportion of phenotypic variance for most traits is not explained by these variants.[4] Among several possible factors that explain this phenotypic variance such as effects of rare variants and epistasis, gene-by-environment interactions (GEIs) have drawn significant attention because of their important effect in many traits and diseases.[5–8] Discovering GEIs involved in diseases is of major interest in genetic research because they can provide insight into disease pathways, an understanding of the effect of environmental factors in disease, better risk prediction and personalized therapies. Similar to traditional GWASs that attempt to detect associations of genetic variants, researchers have recently performed gene-by-environment genome-wide association studies (GxE GWASs) to identify GEIs associated with diseases.[9–11]
One major difficulty in association studies is that population structure can easily confound the studies.[12] Association studies assume that individuals are unrelated, and if they are not, inflation of test statistics and possibly spurious associations may arise if genetic relatedness within individuals is imprecisely modeled. Several statistical approaches have been proposed to address this problem including genomic control [13], principal components analysis [14], and linear mixed models.[15] In particular, methods based on linear mixed models which incorporate pairwise relatedness between individuals has been shown to capture complex sample structure more effectively than other methods.[15, 16] It is important to note that all of these methods are designed to correct for population structure on statistics for genetic variants.
In contrast to numerous studies that have analyzed effect of population structure on association statistics in real GWAS datasets, few studies have investigated its effect on GEI statistics empirically. There have been, however, a few studies that evaluated bias caused by population structure on GEI statistics through simulations. Wang et al.[17] showed that population structure may have small effect on GEI statistics when genetic variants and environments have small correlations while Cheng and Lee [18] showed that it may introduce unacceptable bias to the estimation of GEIs in the presence of selection bias. Wang and Lee [19] also demonstrated that population structure may cause serious bias on estimated GEI effects in case-only studies. Recently, Dudbridge and Fletcher [20] showed that confounding due to population structure may cause dependence between gene and environment, and spurious GEIs can arise under this dependence. Although these studies provide useful information on theoretical impacts of population structure on GEI statistics, its influence in actual GxE GWASs has not been investigated comprehensively.
In this paper, we first show analytically that for the same reason that population structure causes spurious associations of genetic variants, it also causes spurious GEI associations based on the polygenic model. We show that disregarding sample structure can easily inflate test statistics for GEIs, leading to false positives. We then simulate a GxE GWAS using the 1000 Genomes Project dataset.[21] This simulation demonstrates the impact of population structure on GEI statistics more accurately than previous simulations because it is based on actual genotype data that resemble traditional GWAS datasets whereas previous simulations are not. We show that test statistics for GEIs as well as those for genetic variants are inflated due to population structure.
In addition to the simulation, we utilize two GxE GWAS datasets to show that population structure may cause serious effects on GEIs. One dataset is an expression quantitative trait loci (eQTL) study of the human aortic endothelia cell collected by Romanoski et al.[22] and Erbilgin et al.[23] Gene expression was collected with and without a certain treatment, which corresponds to an environmental exposure. The other dataset is a GWAS dataset of inbred mouse strains termed Hybrid Mouse Diversity Panel (HMDP) that consists of 100 classical inbred and recombinant inbred strains.[24] We analyze their lipid phenotypes, and the environment exposure is a thioglycollate injection to recruit macrophages. Both datasets are ideal for evaluating effect of population structure on GEI statistics for following reasons. First, it is known that population structure exists in both datasets; individuals in the human eQTL dataset are from multiple ethnicities, and mouse strains in the HMDP dataset have very diverse genetic backgrounds. Second, both datasets have many quantitative phenotypes to test effect of GEIs; the human eQTL dataset has gene expression measured at more than 18,000 probes, and the HMDP dataset has more than 20 different quantitative phenotypes. This variety of phenotypes allows us to comprehensively determine the impact of population structure on GEI statistics.
We also propose a statistical approach based on a linear mixed model to correct for population structure on GEI statistics. We show that the traditional mixed model approach [15] that incorporates genetic relatedness between individuals only corrects for population structure on effects of genetic variants and does not correctly control inflation of test statistics for GEIs. To solve this problem, we consider two types of pairwise similarities between individuals. One is the traditional genetic similarity that causes a pair of individuals who are genetically similar to have correlated phenotypes, and this causes inflation of test statistics on genetic effects. The other type of similarity is that individuals who are related and have the same environment or exposure status have similar phenotypes, which causes spurious GEIs. We extend the linear mixed model to take into account both types of similarities and show that our approach effectively removes inflation of test statistics for both GEIs and genetic variants in our simulation and the two GxE GWAS datasets.
We generate a simulated GxE GWAS using two populations (GBR and TSI) of 1000 Genomes Project dataset.[21] Each population has 1,000 individuals whose genotypes are generated using only common variants found in a standard SNP chip. In this simulation, we consider a dichotomous environmental exposure and two scenarios; (1) each population has the same number of exposed and unexposed individuals and (2) one population has more exposed individuals than the other population. We generate the genetic kinship matrix (K) from genotype data and the GxE kinship matrix (KD) from K and the environmental exposure. Phenotypes are generated such that the genetic kinship (K) explains 40% of phenotypic variance while the GxE kinship (KD) explains 20% (See Materials and Methods). There is no causal variant in the simulation, meaning that the genomic control inflation factor (λGC) should be close to one for both SNP and GEI statistics. We generate 100 replicates of simulation and measure inflation factors on SNP and GEI statistics of three different approaches. The first approach is one with no population structure correction on both SNP and GEI statistics (“OLS”), and another approach is a linear mixed model approach that incorporates the genetic kinship and accounts for population structure only on SNP statistics (“one RE”). The last approach is our proposed mixed model approach that uses both genetic and GxE kinship to correct for population structure on both SNP and GEI statistics (“two RE”).
Fig 1 shows that population structure may cause spurious GEI associations because inflation factors on GEI statistics are on average greater than one. When the number of exposed and unexposed individuals is the same for both populations, the median λGC of the OLS approach is 1.032 and as high as λGC = 1.363. The results are similar when the ratio of exposed and unexposed individuals is different between the two populations. Population structure in the presence of GEIs may also cause inflation of SNP statistics, and S1 Fig shows that test statistics for SNPs are inflated. Also, λGC on SNP statistics tend to be higher than that on GEI statistics; the median inflation factor on SNP statistics is about 1.12. One of the reasons is that the genetic kinship (K) captures more phenotypic variance than the GxE kinship (KD) does in this simulation. The result demonstrates that both SNP and GEI effects are susceptible to false associations due to population structure.
The result of the simulation also indicates that we need to incorporate both genetic and GxE kinship matrices into the linear mixed model to correct for population structure on SNP and GEI statistics. While the one RE approach that uses only genetic kinship reduces inflation of test statistics on SNPs (S1 Fig), it has almost the same or slightly worse inflation factors on GxE statistics than OLS (Fig 1). With our approach, λGC becomes very close to one; the median λGC values on GEI statistics are 0.9969 and 0.9982 when the ratio of exposed and unexposed individuals between the two populations is the same and different, respectively. The maximum λGC values are also 1.026 and 1.0158, respectively. Interestingly, inflation factors on SNP statistics after applying our approach are even better than those after applying one RE; the median λGC with two RE is 0.9926 while that with one RE is about 1.02 when the ratio of exposed and unexposed individuals is the same. Hence, this shows that incorporating both kinship matrices also reduces inflation of test statistics on SNPs.
To assess the influence of population structure on a real GxE GWAS, we first analyze the eQTL study of human aortic endothelial cell (HAEC).[22, 23] Erbilgin et al. measured gene expression levels of 147 individuals with and without the oxidized phospholipid species, oxidized 1-palmitoyl-2-arachidonoyl-snglycero-3-phosphatidylcholine (Ox-PAPC) treatment. In order to have independent samples, to perform a GxE GWAS, we randomly selected 74 samples where we only used the treated samples and 73 samples where we only used the untreated samples. Due to the normality assumption of the linear regression model, we filter out probes of gene expression that do not follow the normal distribution and choose 8,720 probes for our analysis (See Materials and Methods). We also perform the same quality control as in the original paper for the genotype data, and about 575,000 SNPs are included in our analysis. We compute λGC for each probe on SNP and GEI statistics of the three methods as in the previous simulation.
Fig 2 shows the distribution of inflation factors on GEI statistics with (Fig 2A) and without (Fig 2B) outliers. The results show that population structure indeed causes inflation of test statistics for GEIs, and our method can effectively correct for population structure in a real GxE GWAS. Although all three approaches have very similar median inflation factors for GEI statistics (0.98), OLS and one RE approaches have many more probes whose λGC values are greater than one than our approach. There are 2,687 (31% of total probes) and 2,509 (29%) probes with λGC > 1.02 according to OLS and one RE approaches, respectively, and the maximum λGC values are 1.492 and 1.498, respectively. After applying our approach, there are only 950 probes (11% of total probes) with λGC > 1.02 and the maximum is 1.096. Fig 2B shows that even after removing outliers from the plot, our method has a narrower range of inflation factors than OLS and one RE approaches do. S2 Fig shows that our method also reduces inflation of test statistics on SNPs. Most of probes whose λGC values on SNP statistics are around or greater than 1.4 in the OLS approach have λGC < 1.4 after applying our method although the median λGC of our method is greater than one (1.0365).
We then compare the correlation between λGC and the variance of phenotype explained by the GxE kinship, denoted as σ ^ d 2. We estimate variance components (σ g 2 , σ d 2 , σ e 2 in Eq (8)) using GCTA software [25], and we obtain the ratio of each variance component to the total phenotypic variance. We focus on only probes whose σ ^ d 2 > 10 % because they are the probes in which the GxE kinship explains a certain amount of phenotypic variance. We find that about 24% (2,065) of probes have σ ^ d 2 > 10 %. Fig 3A shows that inflation factors of OLS on GEI statistics tend to increase as the variance of phenotype explained by the GxE kinship increases; r2 between λGC and σ ^ d 2 is 0.4631. This is expected because when σ d 2 is higher, GEI effects become more susceptible to false positives due to population structure. This is similar to higher λGC on SNP effects for phenotypes with higher σ g 2. [15] r2 between σ ^ d 2 and λGC of the one RE approach (0.4620) is similar to that of the OLS approach (Fig 3B), meaning that it does not correct for population structure on GEI statistics. However, after applying our approach, r2 becomes 0.0058 (Fig 3C). This means that even when the GxE kinship explains high phenotypic variance and hence population structure can easily confound GEI associations, our method can successfully correct for population structure.
Next, we utilize the HMDP GxE GWAS dataset [24] that consists of many inbred mouse strains with very different genetic backgrounds. This diversity creates severe population structure, which was shown to easily cause spurious associations of SNP effects. [26] Hence, this dataset allows us to measure the impact of strong population structure on GEI statistics. We analyze 23 lipid phenotypes measured in more than 700 samples, and we test associations of about 74,000 SNPs after QC with these phenotypes. Macrophage recruitment was simulated in mice by injecting thioglycollate solution, which corresponds to environmental exposure in a GxE GWAS. The percentage of samples that received the injection varies between 30% to 42% for different phenotypes. We apply the three methods to each phenotype and measure the inflation factors on SNP and GEI statistics.
Fig 4A shows that population structure causes serious inflation of test statistics for GEIs; the median inflation factor of the OLS approach is 1.77. Inflation factors of the HMDP dataset are generally much greater than those of the human eQTL dataset, and this is expected because the HMDP dataset has much stronger population structure effect than the human eQTL dataset does. The results also show that λGC value becomes close to one and more stable after applying our approach. The median λGC of two RE is 1.092, and especially the maximum λGC is 1.19, which is much smaller than 6.27 of the OLS approach. Interestingly, the one RE approach has a worse distribution of inflation factors than the OLS approach as both median and maximum λGC values of one RE are much greater than those of OLS. This result is to a certain degree consistent with results of the previous 1000 Genome simulation; one RE tends to have higher λGC than OLS in the 1000 Genomes simulation. The one RE model performs similarly to the OLS model which demonstrates that traditional mixed model methods do not correct for GxE interactions. In fact, the one RE model performed slightly worse than the OLS model which is likely because it is attempting to fix a statistical model which doesn’t fit the data. Fig 4B is a QQ plot of one of the phenotypes, free fatty acids (ffa), and it shows that test statistics for GEIs from our method follow the expected distribution while those from the two other methods clearly have inflation of test statistics. S3 Fig shows λGC on SNP statistics, and the results are similar to those of the human eQTL dataset; both one RE and two RE approaches successfully removes inflation of test statistics on SNPs.
Table 1 lists the variance of phenotype explained by the genetic kinship matrix (σ ^ g 2), one by the GxE kinship matrix (σ ^ d 2) and inflation factors on GEI statistics for each phenotype. The genetic kinship matrix accounts for more phenotypic variance than the GxE kinship matrix for all phenotypes; the average σ ^ g 2 is 50% while the average σ ^ d 2 is 12%. However, for certain phenotypes, the GxE kinship explains more than 20% of phenotypic variance, and inflation factors on GEI statistics are greater for these phenotypes than for phenotypes with lower σ ^ d 2. Fig 5 shows the correlation between λGC and σ ^ d 2, and the OLS (Fig 5A) and one RE (Fig 5B) approaches have high correlations, which is similar to the results of the human eQTL dataset. However, our approach significantly reduces the correlation between σ ^ d 2 and λGC (Fig 5C) meaning that our approach effectively removes effect of population structure.
In Fig 4, we observe an anomalous result in HMDP GWAS data; the OLS results significantly outperform the 1RE method. This is somewhat surprising which computes a single variance component to account for the genetic heritability of the trait.
To determine why this is the case, we compare the amount of variance explained by KG in the HMDP data with and without the use of a GxE study. To explore this, we simulated a population using the population structure derived from the HMDP data, as well as a population with a population with structure derived from our 1000G data. In this simulated data, we implanted several GxE associations and used GCTA to estimate the variance components of the model. These numbers are demonstrated in S1 Table. In this simulated data, we implanted several GxE associations and used GCTA to estimate the variance components of the model. We show that in the 1RE model, the variance components are substantially overestimates. We suspect that this is the cause of the higher observed inflation in the p-values compared to the OLS method.
To simulate a gene-by-environment (GxE) GWAS with population structure, we utilize HAPGEN2 software [27] to generate genotype data of two populations in 1000 Genomes Project [21]; GBR (British in England and Scotland) and TSI (Toscani in Italy). We use only common variants in chromosomes 11, 12, 13, and 14 whose minor allele frequency is greater than 5% in both populations. We also use variants present in the Illumina OmniExpress 730K genotyping chip to simulate a typical GWAS. The number of SNPs after this filtering is 99,612, and we generate 1,000 individuals for each population for a total of 2,000 individuals.
To generate phenotype values, we sample them from the following multivariate normal distribution.
y ∼ N 0 , σ g 2 K + σ d 2 K D + σ e 2 I
We create a genetic kinship matrix (or genetic relationship matrix) K using the GCTA software from the simulated genotype data. To create a GxE kinship matrix KD, we first need to assign an environmental exposure to each individual. We assume a dichotomous variable where a half of individuals (1,000) are exposed and the rest (1,000) are unexposed. When assigning an environmental exposure to each sample, we consider two possible cases. One is that each population has the same number of exposed and unexposed individuals. In other words, each population has exactly 500 exposed and 500 unexposed individuals. The other case is that one population has more exposed than unexposed individuals. For example, the ratio between exposed and unexposed in one population is 0.6 while it is 0.4 in other population. This is possible in actual GxE GWASs when individuals in one population are more easily exposed to the environment than those in other population. We vary this ratio from 0.54 to 0.6 in one population. We consider both cases in our simulation to determine how they influence results. Once we decide on the number of exposed and unexposed for each population, we randomly assign the environmental exposure to each individual and create the GxE kinship matrix. The phenotype values are generated such that the genetic kinship matrix and GxE kinship matrix explain 40% and 20% of phenotypic variance, respectively. In other words, σ g 2 is 0.4, σ d 2 is 0.2 and σ e 2 is 0.4. We generate 100 replicates of this simulation.
We also simulated our method at values of σ d 2 (i.e. low GxE kinship) between 0 and 0.2, while holding the ratio between σ g 2 and σ e 2 constant; results are shown in S4 Fig. There appears to be an approximately linear relationship between the amount of inflation (λGC) and the size of σ d 2. As less of the total variance is explained by the GxE kinship, the 2RE method begins to deflate p-values slightly, while both the 1RE and OLS methods improve. However, the over-correction of 2RE methods is small, and near σ d 2 = 0, 1RE and 2RE methods perform similarly.
Because the sample sizes in the HMDP and HAEC datasets are rather low for GWAS, we wanted to demonstrate that our method can accurately detect the amount of GxE kinship.
In order to do this, we performed simulations of populations with the same genetic kinship matrix as was computed from the HMDP dataset, and a phenotype distributed with an implanted GxE kinship component. The results, which demonstrate accurate estimation of this GxE component are in S6 Fig.
A alternate approach for correcting for population structure in gene-by-environmental interaction studies is to include principal components of the genetic relatedness matrix as covariates similar to the approach of EIGENSTRAT. We can extend such approaches to the scenario of gene-by-environment interactions by adding additional covariates of the form of the principal component times the environmental covariate (PCs × environment). We find that using PCs × environment does reduce inflation, but at a lesser extent than including a specifically GxE-based kinship matrix. This is consistent with comparisons of mixed models and principal components in traditional association studies [15]. We demonstrate these results in S7 Fig.
Many of the phenotypes explored in the HMDP and HAEC datasets are close to normally-distributed but are not completely normal. We examine if this is a source of inflation by quantile normalizing the data. In S5 Fig, we observe that quantile normalization of the phenotypes does improve the performance of 1RE and OLS methods, but only slightly.
We demonstrated that population structure may cause spurious associations of gene-by-environment interactions. Using the same argument that population structure can inflate test statistics for genetic variants in the polygenic model, we were able to derive analytically that the same phenomenon may occur for GEI statistics. We then used the 1000 Genomes simulation and two GxE GWAS datasets to observe the impact of population structure on GEI statistics. When the severe population structure exists as in the mouse GxE GWAS dataset, we observed very high inflation factors for GEI statistics. When the influence of population structure is relatively moderate as in the 1000 Genome simulation and the human eQTL GxE GWAS dataset, we found that test statistics for GEIs are nonetheless inflated, which may cause spurious associations. While Wang et al.[17] showed through simulations that population structure may cause small biases to estimated GEI effects when there exists small correlation among environments and genetic variants, their results are not based on actual GxE GWAS datasets. Hence, our results that make use of current GxE GWASs may more accurately represent the impact of population structure on GEI statistics, and our results indicate that even moderate population structure may cause unacceptable inflation of test statistics for GEIs.
To correct for population structure on GEI statistics, we proposed a linear mixed model approach that includes two random effects to take into account two types of similarities between individuals. One is the genetic similarity, and the other is the similarity caused by both genetic and environment. By incorporating two kinship matrices corresponding to the two similarities into linear mixed models, we were able to correct for population structure on GEI statistics successfully. We showed that accounting for only the genetic similarity controls the inflation of test statistics for SNPs, but not for GEIs. This is important because GWASs typically include only the genetic kinship matrix to correct for population structure.[15] Sul and Eskin [16] proposed the idea of including two random effects in linear mixed models to account for two types of population structure; one caused by SNPs under selection and the other by rest of SNPs. As demonstrated in their and this papers, this approach is effective in removing inflation caused by two different types of population structure or confounding.
Recently, Zheng et al.[11] studied the roles of GEIs on type 2 diabetes (T2D) related traits and observed inflation of test statistics on GEIs. They collected information regarding to several dietary and lifestyle factors that may influence the T2D-related traits. These factors were considered as environmental exposure in their GxE analysis, and they measured the variance of T2D-related phenotype explained by GEIs for the different environmental factors. They also performed a GxE GWAS and observed that test statistics for GEIs were inflated for environmental factors that explained a significant amount of the phenotypic variance while they observed no inflation for factors that did not contribute to the phenotypic variance. One of the possible reasons for this inflation is population structure because they did not correct for population structure on GEI statistics. Their result is also consistent with our finding that the inflation factor is higher for a phenotype with higher σ ^ d 2, the phenotypic variance explained by the GxE kinship. Hence, as more GxE GWASs are conducted to discover GEIs associated with traits, correcting for population structure will become important to reduce inflation of test statistics and to remove possible false positive associations.
The linear mixed model in our two RE approach is based on the GCTA GEI model [25], and we use GCTA software to estimate variance components. While previous GxE studies utilized GCTA software to estimate phenotypic variance explained by GEIs, to the best of our knowledge, they did not attempt to measure effect of population on GEIs and to correct for population structure. Our approach is the first method to use the linear mixed model with two kinship matrices to correct for population structure on both SNP and GEI effects.
In this paper, we mainly focused on inflation of test statistics for measuring GEI effects of the three different approaches. We showed that only two RE approach achieves correct false positive rate while the two other methods do not. When comparing performance of different statistical tests, it is also important to compare power of tests in addition to measuring false positive rates. However, power comparison can only be made when all tests achieve the correct false positive rates. In our simulation and real data, it is not possible to compare power among the three different methods because OLS and one RE have incorrect false positive rate.
We note that in our results, we are including the marker that we are testing in the kinship matrix. Recently, several studies have pointed out that including the tested marker in the kinship matrix effectively includes the marker twice in the statistical model and this is what causes the inflation factor to be one even when there are many genetic effects throughout the genome.[28] Our results are orthogonal to these approaches and the recommendation of those studies should apply to testing for GEI as well. Nevertheless, in our experiments, we decided to include the tested marker in the kinship because this more easily exposes inflation of test statistics since we expect to observe an inflation factor of one if there is no inflation.
Before we discuss how population structure may cause spurious GEI associations, we first review how it influences associations of genetic variants because the two concepts are closely related. We assume that genetic effects are additive and there are M variants. Then, the standard genotype-phenotype model is
y k = μ + ∑ i = 1 M β i X i k + ε k (1)
where yk is individual k’s phenotype value, μ is the mean of the phenotype, Xik is the genotype of individual k at variant i, βi is the effect of genetic variant i, and εk is the residual. The polygenic model assumes that there are many variants with small effects, which means that many βi’s are non-zero. The traditional association study considers each genetic variant individually and tests the effect of each genetic variant on the basis of the following model.
y k = μ + β r X r k + η r ¯ k (2)
The goal of association studies is to identify the set of genetic variants with β ≠ 0 since these are variants that putatively affect the phenotype. Note that we use different notations for the residual terms (εk in Eq (1) and η r ¯ k in Eq (2)) to emphasize the difference between the two residuals. The residual in Eq (2) in relation to Eq (1) is exactly
η r ¯ k = ∑ M : i ≠ r β i X i k + ε k (3)
According to Eq (3), people who are related would have similar residual terms (η r ¯ k) because they share the same genotypes (Xik). This violates the assumption of the traditional linear regression model in Eq (2) that residuals are independent and hence causes bias in the estimation of βr. Therefore, sample structure in GWAS datasets such as population structure or cryptic relatedness may cause inflation of test statistics (βr) for genetic variants.[12, 29] For clarity, we refer to the statistics testing for the effect of a genetic variant as SNP statistics to distinguish from statistics testing for the presence of GEI which we refer to as GEI statistics.
One approach to account for the sample structure is through the use of a linear mixed model.[15, 26, 30, 31] This approach introduces a random effect into the linear model in Eq (2) to account for the global genetic relatedness resulting in the following model
y = μ + β r X r + u + e (4)
where y = [y1, y2, …, yn]T and Xr = [Xr1,Xr2, … ,Xrn]T where n is the number of individuals. u is the random effect in the mixed model that captures effect of population structure, and var ( u ) = σ g 2 K and var ( e ) = σ e 2 I where K is an n × n kinship matrix and I is an identify matrix of size n. Then, the total variance of phenotype is given as var ( y ) = σ g 2 K + σ e 2 I. It has been shown that this linear mixed model approach that incorporates the pairwise genetic relatedness into the linear model effectively controls inflation of test statistics for genetic variants due to sample structure.[15, 26]
We extend the standard model to consider an environmental factor D. For simplicity, we assume it is a dichotomous variable. The exposure of an individual k to the environmental factor is denoted as Dk; Dk = 0 for the unexposed and Dk = 1 for the exposed. The model corresponding to Eq (1) is now
y k = μ + ∑ i = 1 M β i X i k + δ D k + ∑ j = 1 M γ j D k X j k + ε k (5)
where δ represents the fixed effect of environmental factor D and each γj is the gene-by-environment interaction effect of variant j and environmental factor D. The goal of GxE association studies is to discover genetic variants or SNPs whose γj ≠ 0 because they have effects on the phenotype in the presence of the environmental factor. While it appears from Eq (5) that interaction effects only affect the phenotype when Dk = 1, this is not always the case. For example, when βi = −γj and Dk = 1, a SNP does not influence the phenotype because SNP effects and interaction effects cancel each out. In this case, the SNP has effects on phenotype only for unexposed individuals (Dk = 0). Similar to the association study of genetic variants, we test the effect of each genetic variant and its GxE effect individually, which corresponds to fitting the following model.
y k = μ + β r X r k + δ D k + γ k D k X j k + τ r ¯ k (6)
where τ r ¯ k is the residual. This residual is precisely
τ r ¯ k = ∑ M : i ≠ r β i X i k + ∑ M : j ≠ r γ j D k X j k + ε k (7)
These residuals (τ r ¯ k) are not independent if individuals are related. For people who are genetically similar, they would have similar value for the first sum (∑M: i≠r βi Xik), and people who are genetically similar and are in the same environment, they would have similar value for the second sum (∑M: j≠r γj Dk Xjk). Hence, this equation shows that for the same reason population structure causes spurious genetic associations as shown in Eq (2), it may inflate test statistics of GEIs and cause false positive associations due to correlated residuals.
We extend the linear mixed model approach to correct for population structure on GEI statistics by introducing an additional random effect that captures the similarity of individuals due to GEI effects. Given the kinship matrix (K), we define the matrix KD where each entry K i j D = K i j if Di = Dj and K i j D = 0 otherwise.[25] This matrix KD describes how individuals are related both genetically and environmentally because a pair of individuals who are genetically related and share the same environment exposure have a non-zero kinship coefficient. We name KD “GxE kinship” and K “genetic kinship” to distinguish two kinship matrices. We propose the linear mixed model that incorporates both kinship matrices as following
y = μ + β r X r + δ D + γ r D · X r + u + v + e (8)
where D = [D1,D2, … ,Dn]T is a column vector of environmental exposures and D⋅Xr is the element-wise product. The random effect v accounts for the relatedness of individuals due to GEI effects and var ( v ) = σ d 2 K D. The total variance of y is then given as var ( y ) = σ g 2 K + σ d 2 K D + σ e 2 I. We call this approach “two RE” because it uses two random effects to correct for population structure on both SNP and GEI statistics.
We compare our approach to other approaches that do not consider effect of population structure on GEI statistics. One such approach is a simple linear regression without any random effect. We name this approach “OLS” from ordinary least squares, and it is defined as
y = μ + β r X r + δ D + γ r D · X r + e (9)
Note that this does not correct for population structure either on SNP statistics or on GEI statistics. Another approach is to correct for population structure only on SNP statistics by including one random effect that accounts for genetic relatedness. Its model is
y = μ + β r X r + δ D + γ r D · X r + u + e (10)
This approach would account for the similarity due to genetic effects (the first sum in Eq (7)), but would not correct for the similarity due to GEI effects (the second sum in Eq (7)). This is because it is likely that values for βi and γj are different for each variant, and the random effect u would not capture GEI effects which is the second sum in Eq (7). We name this approach “one RE” because it uses only one random effect.
P-values of all three approaches can be estimated using a standard F-test. Let Σ = I for OLS, Σ = σ ^ g 2 K + σ ^ e 2 I for one RE, and Σ = σ ^ g 2 K + σ ^ d 2 K D + σ ^ e 2 I for two RE where σ ^ g 2 , σ ^ d 2 , σ ^ e 2 are estimated variance components. We utilize GCTA software [25] to estimate these variance components (σ g 2 , σ d 2 , σ e 2), and we estimate them for each phenotype once and apply them for all SNPs. This is the same as the EMMAX approach [15], and this approach markedly reduces the computational time while maintaining the similar power to that of an approach that estimates variance components for each SNP.
Let β include effects of all covariates in the linear regression model that includes a SNP effect (βr) and a GEI effect (γr) and let X include all covariates including a SNP (Xr) and a GEI (D⋅Xr). Then, the estimated β is
β ^ = X ′ Σ - 1 X - 1 X ′ Σ - 1 y (11)
We perform a standard F-test for the null hypothesis βr = 0 and γr = 0 to obtain p-values for SNP and GEI effects, respectively. We provide a software package that implements the two RE approach at http://genetics.cs.ucla.edu/pylmm/. Our approach can efficiently be applied to standard GWAS datasets that contain thousands of individuals and hundreds of thousands of SNPs, similar to the linear mixed models for GWAS [15].
It is very attractive to extend this model to the case of continuous covariates; however, it is not straightforward to create KD from K. Under a binary enviornmental exposure, setting K i , j D = K i , j × δ D i , D j where δDi,Dj is 1 if the environmental exposures of indivdiuals i and j are the same and 0 otherwise is intuitive. For continuous covariates, we can extend this formalism to: K i , j D = K i , j × f ( D i , D j ) for some function f. A good general guideline for a choice of function is to use f = 1−d(Di,Dj) where d is a metric with range [0, 1]. Two of the most natural choices that satisfy the above recommendation are: f ( i , j ) = 1 - | D i - D j R |, where R is the range of the environmental exposures, or f ( i , j ) = 1 - | Φ ( D i - μ D σ D ) - Φ ( D j - μ D σ D ) |, where μD and σD are the mean and standard deviation of the environmental exposures, and Φ is the standard normal cumulative distribution function. However, the best choice of f will depend on the scale of the environmental exposure.
Erbilgin et al.[23] performed the expression quantitative trait loci (eQTL) study of human aortic endothelial cell (HAEC). They collected HAEC cultures from 147 unrelated heart transplant donors, and the oxidized phospholipid species, oxidized 1-palmitoyl-2-arachidonoyl-snglycero-3-phosphatidylcholine (Ox-PAPC) treatment was applied to the cells. It has been known that Ox-PAPC promotes vascular inflammation and regulates more than 1,000 transcripts in this cell type.[22, 32] Gene expression was collected both with and without the Ox-PAPC treatment. In order to have the two conditions have independent samples, we randomly chose a subset of 74 individuals where we only used the treated samples and a different 73 individuals where we only used the untreated samples, which represents two exposure statuses of environment.
The gene expression on 18,630 probes is collected using Affymetrix HT HG-U133A microarrays. The COMBAT software was utilized to correct for batch effects in the expression data [33]. Since the linear regression model assumes that expression values follow the normal distribution, we filtered out probes whose Shapiro-Wilk test p-values are less than 0.05. Additionally, we computed the number of outliers for each probe whose expression values are two standard deviations apart from the mean, and excluded probes containing five or more outliers (5% of total samples). These filters removed 9,910 probes, leaving 8,720 probes for the subsequent analysis. To verify that our results are not affected by the fact that the data still deviates from the normal distribution, we reanalyzed the data after performing quantile normalization both within each exposure group and over the entire dataset. In these additional experiments we observed equivalent results as shown in S5 Fig.
SNPs are genotyped using Affymetrix Genome-Wide Human SNP Array 6.0. We used the same QC filters as in the original paper [23]: MAF of 10%, HWE p-value of 10−4, and genotype completeness of 5%, and 575,042 SNPs in autosomes are tested for associations. Erbilgin et al. performed a principal component analysis to identify population structure among the 147 individuals with 11 HapMap3 populations and found that there are groups of individuals with different ethnicities.
Hybrid Mouse Diversity Panel (HDMP)[24] consists of 100 inbred strains including 29 classic inbred strains and three sets of recombinant inbred strains. The 23 lipid phenotypes (and their abbreviations) that were analyzed are: body weight (bw), fat mass by NMR (fat_mass), free fatty acids (ffa), Femoral fat pad (ffp), Femoral fat pad/total bw (ffp_percentage), water weight by NMR (free_fluid), gonadal fat pad (gfp), gonadal fat pad/total bw (gfp_percentage), glucose at time for sac (glucose), glucose by lipid core (glucose_lc), HDL (hdl), LDL and VLDL (ldl_and_vldl), lean mass by NMR (lean_mass), mesenteric fat pad (mfp), mesenteric fat pad/total bw (mfp_percentage), Body fat percentage determined by NMR (nmr_bf_percentage), total mass by NMR (nmr_total_mass), retroperitoneal fat pad weight (rfp), retroperitoneal fat pad weight/total bw (rfp_percentage), spleen weight (spleen_wt), total cholesterol (tc), triglycerides (tg), unesterified cholesterol (uc). The number of samples from the 100 inbred strains that are phenotyped varies between 735 and 894 among these phenotypes. These strains are genotyped at more than 130,000 SNPs, and we applied following QC; genotype completeness of 98% for both SNPs and individuals and minor allele frequency threshold of 10%. After the QC, we have about 74,000 SNPs for performing association studies. The environment that we are interested in is thioglycollate injection to recruit macrophages. Macrophages play an important role in inflammatory component of many common diseases.[34] The percentage of exposed samples is between 30% and 42% depending on phenotypes. Note that individuals from the same strain can be both exposed and unexposed.
|
10.1371/journal.pgen.1001296 | Parallel Evolution of a Type IV Secretion System in Radiating Lineages of the Host-Restricted Bacterial Pathogen Bartonella | Adaptive radiation is the rapid origination of multiple species from a single ancestor as the result of concurrent adaptation to disparate environments. This fundamental evolutionary process is considered to be responsible for the genesis of a great portion of the diversity of life. Bacteria have evolved enormous biological diversity by exploiting an exceptional range of environments, yet diversification of bacteria via adaptive radiation has been documented in a few cases only and the underlying molecular mechanisms are largely unknown. Here we show a compelling example of adaptive radiation in pathogenic bacteria and reveal their genetic basis. Our evolutionary genomic analyses of the α-proteobacterial genus Bartonella uncover two parallel adaptive radiations within these host-restricted mammalian pathogens. We identify a horizontally-acquired protein secretion system, which has evolved to target specific bacterial effector proteins into host cells as the evolutionary key innovation triggering these parallel adaptive radiations. We show that the functional versatility and adaptive potential of the VirB type IV secretion system (T4SS), and thereby translocated Bartonella effector proteins (Beps), evolved in parallel in the two lineages prior to their radiations. Independent chromosomal fixation of the virB operon and consecutive rounds of lineage-specific bep gene duplications followed by their functional diversification characterize these parallel evolutionary trajectories. Whereas most Beps maintained their ancestral domain constitution, strikingly, a novel type of effector protein emerged convergently in both lineages. This resulted in similar arrays of host cell-targeted effector proteins in the two lineages of Bartonella as the basis of their independent radiation. The parallel molecular evolution of the VirB/Bep system displays a striking example of a key innovation involved in independent adaptive processes and the emergence of bacterial pathogens. Furthermore, our study highlights the remarkable evolvability of T4SSs and their effector proteins, explaining their broad application in bacterial interactions with the environment.
| Adaptive radiation is the rapid origination of an array of species by the divergent colonization of disparate ecological niches. In the case of pathogenic bacteria, radiations can lead to the emergence of novel human pathogens. Being divergently adapted to a range of different mammalian hosts, including humans as reservoir or incidental hosts, the genus Bartonella represents a suitable model to study genomic mechanisms underpinning divergent adaptation of pathogens. Here we show that two distinct lineages of Bartonella have radiated in parallel, resulting in two arrays of evolutionary distinct species adapted to overlapping sets of mammalian hosts. Such parallelisms display excellent models to reveal insights into the genetic mechanisms underlying these independent evolutionary processes. Our genome-wide analysis identifies a striking evolutionary parallelism in a horizontally-acquired protein secretion system in the two lineages. The parallel evolutionary trajectory of this system in the two lineages is characterized by the convergent origination of a wide array of adaptive functions dedicated to the cellular interaction within the mammalian hosts. The parallel evolution of the two radiating lineages on the ecological as well as on the molecular level suggests that the horizontal acquisition and the functional diversification of the secretion system display an evolutionary key innovation underlying adaptive evolution.
| Adaptation to different ecological niches can lead to rapid diversification of a single ancestor into an array of distinct species or ecotypes. This process, called adaptive radiation, typically occurs after the arrival of a founding population in a novel environment with unoccupied ecological niches (‘ecological opportunity’) and/or by the acquisition of a novel trait (‘evolutionary key innovation’) allowing the exploitation of so far unapproachable niches [1]. Spectacular examples of adaptive radiation come from different metazoan lineages with the cichlid fishes of the East African Great Lakes and the Darwin finches on Galapagos Islands representing the most prominent examples [2], [3]. Although known from a few cases only, bacterial lineages also underwent adaptive radiation - as documented in natural settings as well as in evolution experiments [4]–[6]. It remains a fundamental problem to biology to understand why and how certain lineages diversified; adaptive radiations, and in particular the genetic and genomic basis thereof, provide an ideal set-up to address this question [3], [7].
One of the most fascinating aspects of adaptive radiation is the frequent occurrence of evolutionary parallelism resulting in independent adaptation to same ecological niches [8]–[10]. Such evolutionary parallelisms are excellent examples for the action of similar, yet independent, selective forces and, hence, for the key role of natural selection in evolution [11]. Furthermore, parallel adaptive radiations in a single group indicate the existence of traits conferring a high degree of adaptability allowing the group members to efficiently occupy distinct environments. Therefore, lineages that radiate in parallel are of great value to study the molecular basis of adaptation and their independent evolutionary trajectories [1], [9]. This is of particular interest in case of host-adapted bacteria differentiating into divergent ecological niches and potentially resulting in the emergence of new pathogens. Species of the α-proteobacterial genus Bartonella are specifically adapted to distinct mammalian reservoir hosts where they cause intra-erythrocytic infections [12]. Different animal models revealed that Bartonella upon reservoir host infection colonizes a primary cellular niche from where the bacteria get seeded into the bloodstream adhering to and invading erythrocytes [13]–[15]. In most cases, infections of the reservoir host do not lead to disease symptoms suggesting a highly specific adaptation to the corresponding host niche. The transmission between host individuals is mediated by blood sucking arthropods.
An integrative genome-wide analysis showed that most factors essential for Bartonella to colonize their mammalian reservoir hosts are found within the core genome of this genus [16]. This is not surprising as it reflects the common strategy used by divergently adapted species to colonize their hosts. However, this study also revealed that two type IV secretion systems (T4SS), Trw and VirB, which are essential for host interaction at different stages of the infection cycle represent the few colonization factors exclusively found in the most species rich sub-lineages of bartonellae. It was assumed that the horizontal acquisition of these T4SS substantially refined the infection strategy of Bartonella facilitating concurrent adaptation to a wide range of different hosts [16]. The VirB T4SS translocates a cocktail of evolutionarily related effector proteins into host cells of the primary infection niche where they modulate various cellular processes [17]–[21]. The Trw T4SS is involved in the erythrocyte invasion by binding to the erythrocytic surface with its manifold variants of pilus subunits [22]–[24].
Here, we study the evolutionary relationship of Bartonella species adapted to distinct reservoir hosts and investigate the genetic mechanisms underlying adaptive radiation in different lineages. We uncover two parallel adaptive radiations in the genus Bartonella. Our genome-wide analysis revealed a remarkable evolutionary parallelism in the horizontally acquired VirB T4SS in the two radiating lineages. This parallelism is characterized at the molecular level by the lineage-specific chromosomal integration of the virB loci and the independent origination of versatile sets of effector proteins for the interaction with host cells. Providing an arsenal of host-subverting functions that can be efficiently modulated, the VirB T4SS thus seems to represent an evolutionary key innovation triggering the independent radiations of the two lineages. Our study provides detailed insights into the molecular mechanisms underlying parallel adaptive radiations in a bacterial pathogen. Furthermore, many of the diversified T4SS effector proteins carry a FIC domain recently shown to mediate ‘AMPylation’, a lately recognized post-translational modification [25], [26]. FIC domains are highly conserved in evolution and the diversified variants of the Bartonella effector proteins may display a suitable model to study their activity spectrum in the future.
To study the adaptive evolution of Bartonella on a genomic level, we aimed for a set of genome sequences from species adapted to distinct mammalian reservoir hosts. To this end, we included in our analysis the published genome sequences of Bartonella bacilliformis (Bb), Bartonella grahamii (Bg), Bartonella henselae (Bh), Bartonella quintana (Bq), and Bartonella tribocorum (Bt) [16], [27], [28]. These five species are adapted to human (Bb and Bq), cat (Bh), mouse (Bg), and rat (Bt). Further, we sequenced the complete genome of Bartonella clarridgeiae (Bc) and generated draft sequences of Bartonella schoenbuchensis (Bs), Bartonella rochalimae (Br), Bartonella sp. AR 15-3 (BAR15), and Bartonella sp. 1-1C (B1-1C). Bs was selected as representative of a solely ruminant-infecting clade [16]. Bc, Br, BAR15, and B1-1C were previously shown to be closely related [29]–[31]. However, they were isolated from different mammalian reservoir hosts and therefore display a suitable set of species to study adaptive processes on the genomic level. BAR15 and B1-1C were recently isolated from American red squirrel and rat, respectively [30], [31], whereas Br was predominantly recovered from canidae like dogs or foxes, and Bc from cats [32]–[34].
Genome sequencing by 454-pyrosequencing resulted in an average sequence coverage of >35x. The single chromosome of the completely assembled genome of Bc was found to be 1,522,743 bp in size and thus belongs to the smaller genomes of Bartonella (Table 1). The draft genomes of Br, BAR15, B1-1C, and Bs consist of 13 to 19 contigs with total genome sizes similar to the one of Bc. On average, 99% of all 454-sequencing reads were assembled into the analyzed 13 to 19 contigs indicating that our draft genomes did not miss essential sequence data for subsequent analysis. Genomic features of the strains used in this study are summarized in Table 1.
We inferred a robust species tree of the genus Bartonella based on 478 core genome genes of the ten available Bartonella genomes sequenced (Figure S1). To exclude that recombination or horizontal gene transfer within this set of core genome genes was affecting our phylogenetic analysis, we reconstructed single gene trees of the entire data set and performed a recombination analysis using the GARD algorithm [35]. 471 of the 478 genes revealed the same overall topology as our genome-wide phylogeny with the two monophyletic clades of lineage 3 and lineage 4 (see Figure 1). Further, the GARD analysis detected significant recombination breakpoints with a p-value<0.01 in only two out of the 478 core genome genes. Together these analyses show that our genomic data set is suitable for inferring a consistent species tree.
Based on available sequence information for the housekeeping genes rpoB, gltA, ribC, and groEL, we included most other Bartonella species in the analysis resulting in a so-called supertree phylogeny (Figure 1) [36]. Just as the analysis based on the 478 core genome genes of the ten sequences species alone, this supertree revealed four major clades in the monophyletic bartonellae: ancestral lineage 1 represented by the highly virulent human pathogen Bb [37]; lineage 2 comprising of Bs and three other ruminant-infecting species; lineage 3 consisting of the closely related Bc, Br, BAR15, and B1-1C; and the most species-rich lineage 4 with 13 species including Bg, Bh, Bq, and Bt (Figure 1). A phylogeny based on only the four housekeeping genes resulted in the same clustering of these taxa into the four different Bartonella lineages (Figure S1).
In contrast to the ancestral lineage 1, lineages 2, 3, and 4 are ramifying to different degrees comprising species isolated from various hosts. While the species of lineage 2 are limited to infect ruminants and have overlapping host range [38], the diversification of lineage 3 and 4 seems to result from the specific adaptation to distinct mammalian hosts [12]. To substantiate the ecological divergence within these two lineages, we analyzed the genotype-host correlation of Bartonella isolates sampled from diverse mammals. Based on gltA sequences, this analysis revealed clustering of strains isolated from same or similar hosts in lineage 3 and 4 (Figure S2). Further support for the host specific adaptation of different Bartonella species comes from recently published laboratory infections [39], [40] and from our own rat infection experiments with the strains of lineage 3 (Figure S3). It is to mention that some Bartonella species can incidentally be transmitted to other hosts like humans [12]. These so-called zoonotic Bartonella species do not cause intraerythrocytic bacteremia in the accidental human host reflecting the lack of specific adaptation. However, such accidental transmissions might facilitate the emergence of new specificity resulting in host switches and the origination of new species. In particular, several species of lineage 3 and 4 are known to display such zoonotic pathogens, whereas for lineage 1 or 2 to our knowledge no such case has been reported so far [12], [29].
In summary, our genome-wide phylogenetic analysis shows that the sister lineages 3 and 4 have evolved by adaptive radiations into same or similar ecological niches (i.e. hosts). Long internal branches separating the two lineages from each other and preceding the radiations are evidence for their independent occurrence (Figure 1). Due to the lack of calibration time points, the exact timing of these independent radiations cannot be deduced. However, the phylogenetic tree in Figure 1 might suggest that lineage 3 diversified more recently compared to lineage 4. This is supported by the mean p-distances inferred for the sequenced taxa of these two lineages: lineage 3 = 0.07±0.0002, lineage 4 = 0.12±0.0003 (see also Table S1). Two alternative explanations for the observed differences in lineage diversification could be (i) a sampling bias, i.e. the full diversity of lineage 3 was not captured or (ii) smaller population sizes for species of lineage 4 over lineage 3 leading to faster evolution at purifying sites. Significant differences in population size might be rather unlikely as Bartonella species are thought to share a common life style in their respective reservoir host. Whether sampling of Bartonella species in animal populations was exhaustive enough is difficult to assess. However, a newly discovered species would only change the coalescent point of a lineage if it would hold a more ancestral position than the already known species of this lineage. In contrast to these alternative hypothesis, epidemiological studies rather seem to support the scenario of a more ancient onset of the radiation in lineage 4: (i) lineage 4 comprises a much wider range of divergently adapted species and (ii) they represent the most frequently found Bartonella species in natural host populations [30], [41]. In contrast, except for Bc, taxa of lineage 3 were only recently detected and so far only sampled at low prevalence [29]–[32].
The evolutionary parallelism of the radiating Bartonella lineages provides an ideal setting to study independent evolutionary processes linked to the adaptation to divergent niches. An important driving force for these adaptive radiations might be the presence of ecological opportunities. The niche of Bartonella in the mammalian reservoir host, the bloodstream, displays a privileged environment in which other resource-competing microbes are typically absent. Thus, the adoption of the characteristic intra-erythrocytic infection strategy together with the vector-borne transmission route might have enabled adaptive radiations of Bartonella by the specialization to different hosts. However, not all Bartonella lineages appear to have diversified to the same extent (see Figure 1) suggesting that the availability of such an ecological opportunity alone is not sufficient to explain the pronounced radiation of lineages 3 and 4. Supposedly, key innovations, i.e. lineage-specific traits underlying the adaptation to the mammalian host niches are responsible for the adaptive radiations. Potential adaptive traits would have to be involved in species-environment interactions, such as molecular factors responsible for causing bacteremia. Further, in analogy to the modulation of the adaptive traits in metazoan radiations [2], [3], any molecular factor used to exploit distinct environments in a specific manner should be divergent among niche-specialized species [1]. Molecular evolutionary analyses provide the means to identify divergent adaptive traits as the genes encoding them are expected to show signs of adaptive evolution, i.e. an excess of non-synonymous (dn) over synonymous (ds) substitutions as the result of positive selection. We performed a genome-wide natural selection analysis in the two radiating lineages to detect genes (and therefore traits) with divergent evolution. To this end, we analyzed all orthologous genes from the available genomes of the radiating lineage 3 (Bc, Br, BAR15, and B1-1C) and lineage 4 (Bh, Bg, Bq, and Bt) for signs of adaptive sequence evolution by inferring the natural selection of orthologs by estimation of ω, the ratio of non-synonymous (dn, amino acid change) to synonymous (ds, amino acid conservation) substitution rates (ω = dn/ds). Generally, ω<1, ω = 1, ω>1 represent purifying, neutral, and positive selection (adaptive evolution), respectively [42].
We first calculated “gene-wide” dn/ds for all orthologous genes of the two lineages (lineage 3: 1,097 genes, lineage 4: 1,091 genes). We excluded one gene from this analysis for each of the two lineages, because the GARD analysis detected statistically significant recombination breakpoints. As adaptive evolution is typically affecting only a few sites of a gene rather than the entire gene sequence [42], we first looked for genes exhibiting an elevated value of ω≥0.25 over the entire gene length. This analysis revealed 133 (12%) and 86 (8%) genes in lineage 3 and lineage 4, respectively, under relaxed purifying selection indicating signs of adaptive evolution (Figure 2). To have an additional measurement for adaptive evolution, we subjected our genomic data sets to a maximum likelihood analysis for the detection of site-specific positive selection. To this end, we used the CodeML module implemented in the PAML package. CodeML compares the likelihoods of different evolutionary models for each analyzed gene alignment. When comparing model M2a (PositiveSelection) vs. model M1a (NearlyNeutral), we detected 62 or 34 genes for lineage 3 and 26 or 14 genes for lineage 4 harboring sites under positive selection with a p-value of <0.05 or <0.01, respectively. A large fraction of these genes (29 for lineage 3 and 12 for lineage 4) exhibited also gene-wide dn/ds values ≥0.25 indicating them as good candidates for encoding adaptive traits (Table 2). Comprehensive lists of the genes identified to have dn/ds values ≥0.25 and/or exhibiting site-specific positive selection in the CodeML analysis are provided in Table S2 and Table S3 for lineage 3 and lineage 4, respectively. As non-synonymous mutations accumulate over time, the higher number of genes identified for lineage 3 could be a further indication for its more recent radiation, or alternatively, the effect of larger population sizes compared to lineage 4. Irrespectively, these findings render the dataset derived from lineage 3 to be more sensitive to the detection of adaptive sequence evolution (Figure 2, Table 2).
Interestingly, there was a marked number of genes with ω≥0.25 over the entire gene length that overlapped in both lineages (Table 3). Among those were genes encoding autotransporters, hemin-binding proteins, and different components of the VirB T4SS. They all constitute important host colonization factors [43]–[45] and thus are likely to display adaptive traits of Bartonella. For many of these genes, also our analysis of site-specific natural selection detected positive selection (Table 3). Most remarkably, all analyzed Bartonella effector protein (bep) genes of the VirB T4SS were among the genes with ω≥0.25. Particularly in lineage 3, they showed strong signs of adaptive evolution by exhibiting ω>0.4 over the entire gene length. Further, in eight out of nine analyzed bep genes of lineage 3, we detected site-specific positive selection (Table 3). Being exclusively found in the radiating lineages and showing strong signs of adaptive evolution, the VirB system and its effector proteins, thus, fulfill the criteria of an evolutionary key innovation likely contributing to the parallel adaptive radiations of Bartonella. Autotransporters and hemin-binding proteins could represent further adaptive traits important for radiations in Bartonella. They exhibited strong positive selection in our analysis and are known to be important factors for host colonization. Further, their conservation throughout the genus Bartonella indicates an important role for the life style and infection strategy of this pathogen (Table S2, Table S3). However, factors conserved in radiating and non-radiating lineages appear unlikely to represent specific key innovations, unless other factors, such as the absence of ecological opportunities or ecological separation, prevented certain lineages to radiate and to colonize more divergent niches.
Importantly, our analysis revealed some lineage-specific colonization factors to carry signs of adaptive evolution. Among others, surface-exposed pilus-components of the Trw T4SS exclusively present in lineage 4 were found to exhibit elevated dn/ds values and sites under positive selection (Table S3). This is in agreement with previous studies and appears to reflect the adaptation of this putative adhesion factor to the erythrocytic surface of different host species [22], [24]. As the Trw T4SS is only present in lineage 4, it might have specifically contributed to the radiation of this most species-rich clade of Bartonella. Factors known to be important for colonization and exclusively present in lineage 3 were not identified by our analysis.
We cannot exclude that the selective pressure imposed by the immune-system might have contributed to the adaptive evolution detected in our genome-wide analysis. It was previously reported that the arms race between host and pathogen can drive the diversification of secretion system- and effector protein-encoding genes [46], [47]. However, in case of the Trw T4SS, recently published in vitro infections with erythrocytes isolated from different mammals demonstrated that the Trw-dependent binding and invasion of Bartonella is host-specific [22]. Although experimental data is not yet available, our data suggest that the VirB T4SS and its effector proteins evolved by similar mechanisms. Together with the previous finding that the VirB T4SSs belong to the few colonization factors specific to the radiating lineages [16], this analysis reveals these horizontally acquired host interacting systems as potential key innovations facilitating adaptation to new hosts and therefore driving the radiations of Bartonella.
To further assess the role of the VirB T4SS for the independent adaptive radiations of Bartonella, we compared the chromosomal organization of the VirB and effector protein-encoding genes of the two lineages. Remarkably, our analysis uncovered independent evolutionary scenarios for the chromosomal incorporation of this horizontally acquired trait. In the genomes of lineage 4 (Bg, Bh, Bq, and Bt), the virB T4SS genes, virB2-virB11 and the coupling protein gene virD4, are encoded at the same chromosomal location (Figure 3, Figure S4). Also, the bep genes are encoded in this region. In contrast, the genome sequences of lineage 3 (Bc, Br, BAR15, and B1-1C) revealed marked differences in organization, copy number, and chromosomal localization of the genes encoding the VirB T4SS. In the completely assembled genome of Bc, we found three copies of the virB2-virB10 genes encoded at two different chromosomal locations (Figure 3, Figure S4). Two copies are encoded at the same locus and belong to inverted repeats of ∼10kb. They are separated by several bep genes and the gene virD4. A third copy of the virB2-virB10 cluster including an additional bep gene is encoded in another genomic region highly conserved across different Bartonella lineages (Figure 3).
The same chromosomal integration and amplification of the virB T4SS genes was found in the other three genomes of lineage 3. However, in a common ancestor of Br and B1-1C, one of the three copies must have been partially deleted, as only virB2, virB3, and a remnant of the virB4 gene were found in the corresponding region of these two genomes (Figure 3, Figure S4). Interestingly, the different copies of virB2-virB10 are identical to each other within one species, but divergent across different species indicating the presence of an intra-chromosomal homogenization process. The fact that duplicated components of another T4SS, Trw, also evolved in concert, and the finding of several other identical genes or gene clusters in different Bartonella genomes [24] suggests that sequence homogenization is a common mechanism in Bartonella to conserve paralogous gene copies. The inverted organization of the two virB T4SS gene clusters seems to result from a duplication event subsequent to the integration of a first copy. Evidence comes from a remnant of the glutamine synthetase I gene (glnA) flanking the entire locus at its upstream end. The full-length copy of this vertically inherited housekeeping gene is located directly downstream of the integration site (Figure 3, Figure S4).
In addition to the effector genes adjacently located to the virB genes (as in lineage 4), we found six additional loci encoding bep genes in lineage 3 (Figure S4). These effector genes are not entirely conserved throughout lineage 3, and the existence of gene remnants provides evidence of their deterioration in certain species. Altogether, we identified 12 to 16 bep genes in lineage 3, whereas only five to seven bep genes are present in lineage 4.
Incomplete synteny in the corresponding regions may hinder comparison between the two different lineages, however, no gene remnants could be found at the different integration sites across the two lineages. We cannot fully exclude that massive genomic recombination events resulted in the different chromosomal locations and the lineage-specific dissemination of the virB and bep genes. Yet, such a scenario appears unlikely, as the overall genomic backbone is largely conserved (Figure 3) and the flanking regions of the virB T4SS integration sites do not encode vertically-inherited orthologs across the two lineages. Furthermore, the absence of mobile elements adjacent to the virB T4SS genes such as recombinases, transposases, or integrases is not supportive of an intra-chromosomal mobilization of this genomic locus. T4SS are ancestrally related to conjugation machineries [48]. Thus, the virB genes might have been transferred from a conjugative plasmid into the chromosome by independent events after the divergence of lineage 3 and lineage 4. In Bg, a closely related T4SS, the Vbh, is encoded on a plasmid in addition to a chromosomally integrated copy [28]. This indicates that these horizontally acquired elements can be maintained on extra-chromosomal replicons within Bartonella from where they are integrated into the chromosome. Similarly, pathogenic Escherichia coli strains from different phylogenetic clades were shown to have evolved in parallel by the independent incorporation of virulence traits from mobile genetic elements [49].
As the chromosomal organization implicates different evolutionary histories of the VirB T4SS in the two radiating lineages, we investigated the relation among the effector proteins translocated by this secretion system. It was previously shown that the bep genes have evolved from a single ancestor by duplication, diversification, and reshuffling of domains resulting in modular gene architectures [17]. The C-terminal BID (Bartonella intracellular delivery) domain is shared by all Beps as it constitutes the secretion signal for the transport via the VirB T4SS. In their N-terminal part, Beps either harbor a FIC (filamentation-induced by cAMP) domain or repeats of additional BID domains (Figure S5). We assessed the evolutionary relationship among the bep genes by inferring phylogenetic trees on the basis of either the BID or the FIC domain, or the entire gene sequence. This revealed that the bep genes of lineage 3 and lineage 4 form two separate clades (Figure 4, Figure S6). Apparently, consecutive rounds of lineage-specific duplications of an ancestral effector gene resulted in the parallel emergence of two distinct arsenals of bep genes. These duplication events preceded the adaptive radiation in both lineages as phylogenetic clusters of effector genes (Bep clades in Figure 4) comprise positional orthologs present in all or a subset of the analyzed genomes of the corresponding lineage (Figure S4). Gene duplications frequently display the primary adaptive response after the acquisition of beneficial factors, because they occur at much higher frequency than other adaptive mutations [50]. This might have been the initial selective pressure for the independent amplification processes. However, in both lineages, the duplicated bep genes subsequently diversified by accumulating mutations as indicated by different branch lengths separating bep genes in Figure 4.
To analyze the sequence evolution during the parallel amplification and diversification processes, we used a branch test for positive selection. Since positive selection is not continuously acting during evolution, this analysis allows the detection of episodic adaptive evolution on single phylogenetic branches. We detected positive selection on many of the internal branches suggesting that subsequent to their duplication different Bep clades have undergone adaptive sequence evolution in both lineages (Figure 4, Figure S6). For Bartonella, experimental studies showed that effector proteins exhibit distinct phenotypic properties on host cells indicating that the evolutionary diversification of the duplicated effectors was substantially driven by the acquisition of novel functions [18]–[21]. Not all branches exhibit dn/ds values >1, though, suggesting episodic changes in the selection pressure acting on different effector gene copies. For example, functional redundancy of paralogous effector copies could have resulted in neutral drift, whereas conservation of an advantageous function might have led to purifying selection on certain branches.
The basis for the functional versatility seems to lie in the adaptability of the domains encoded by bep genes. In case of the FIC domain, recently published work showed that this domain mediates a new post-translational modification by transferring an AMP moiety onto a target protein [25]. Proteins ‘AMPylated’ by FIC domains belong to the family of GTPases. The diversity of these targets and their numerous functions in cellular processes might allow the diversified Beps to target and subvert a variety of host cell functions by target-specific ‘AMPylation’. The high degree of conservation of the FIC domain in different kingdoms of life provides further evidence for the remarkable versatility of this domain [51]. Interestingly, also the BID domain, constituting part of the translocation signal of the effector proteins, seems to be capable of adopting various functions in the host cell [17]. In case of BepA, it was shown that the BID domain is sufficient to mediate the anti-apoptotic property of this effector protein [18]. BID domains of other Beps with the same domain constitution as BepA do not exhibit this phenotype indicating specific adaptive modulation of this domain for BepA. This functional adaptability might also explain why certain effector genes carry more than one BID domain.
At last, tandem-repeated tyrosine-phosphorylation motifs found in a subset of effector proteins confer another multifaceted molecular mechanism to modulate cellular processes. Phosphorylated effector proteins are thought to recruit cellular binding partners resulting in the formation of signaling scaffolds that interfere with specific host cell signaling pathways [21]. For several effector proteins of Bh (lineage 4), tyrosine phosphorylation by host cells has been reported and the targeted host interaction partners studied [17], [21]. Beside Bartonella, a number of other pathogens, as E. coli (EPEC), Helicobacter pylori, or Chlamydia trachomatis are using tyrosine-phosphorylation of effector proteins to modulate their hosts in very distinct ways demonstrating the versatility of this type of host subversion [21]. In Bartonella, the tyrosine-phosphorylated effector proteins seem to display an important functionality of the VirB-mediated host modulation as we found effector proteins of this type in both radiating lineages (see below).
Our analyses suggest that the domain structure of the ancestral effector gene consisted of an N-terminal FIC and a C-terminal BID domain (FIC-BID). In both lineages, the FIC-BID domain structure displays the most abundant effector protein type. In lineage 3, only effector genes of Bep clade 9 consist of domain architectures different than FIC-BID (Figure S5). The gene tree in Figure 4 shows that bep genes with the shortest evolutionary distance across the two lineages are the ones harboring the FIC-BID structure (BepA clade and Bep clade 1). bep genes with different domain architecture constitute more distantly related clades across the two lineages indicating that they derived by independent recombination from the ancestral domain structure. Furthermore, the distantly related Vbh T4SS of Bg and Bs encodes an effector protein consisting of the FIC-BID domain structure [28].
As mentioned above, it was shown for lineage 4 that some of the derived bep genes become phosphorylated by host cell kinases at conserved tandem-repeated tyrosine-phosphorylation motifs leading to the interference with specific host cell pathways [21]. Strikingly, we found that bep genes with derived domain architecture in lineage 3 also harbour regions with tandem-repeated tyrosine motifs (Figure 5, Figure S5). In silico predictions of tyrosine-phosphorylation sites with three different programs [52]–[54] consistently revealed a high number of potentially phosphorylated motifs within these repetitive regions (Figure 5, Table S4). We ectopically expressed these effector proteins in HEK293T cells and showed that they are indeed phosphorylated within eukaryotic cells by tyrosine kinases implicating their functional importance for host interaction (Figure 5). Interestingly, the motifs found in lineage 3 are clearly different from the ones present in lineage 4 and are also less conserved as depicted by their consensus sequences in Figure 5. This suggests that the motifs in lineage 3 may generally be under weaker purifying selection than in lineage 4, because they target either less conserved or even different pathways in their hosts. Further, the lower degree of conservation and the higher number of motifs per effector found in lineage 3 could also indicate that these proteins and particularly their motifs are under positive selection and evolved more recently than their equivalents in lineage 4.
Together with the fact that tandem-repeated phosphorylation motifs are only found in bep genes with derived domain architecture, our findings, thus, suggest parallel evolution of this class of effector proteins within the two radiating lineages. Whether similar pathways are targeted by these effectors in the two lineages remains unknown. Yet, the striking parallelism in the molecular evolution of this class of effector proteins indicates their central role in the VirB T4SS mediated host modulation by Bartonella.
Emerging infectious diseases are frequently caused by zoonotic pathogens which are incidentally transmitted to humans from their reservoir niche (e.g. other animal hosts). Therefore, the understanding of the mechanisms driving diversification of host-adapted bacteria in nature is of relevance for human health. In the present study, we explored the adaptive diversification of host-restricted bartonellae. Our genome-wide phylogeny revealed that two sister clades of this α-proteobacterial pathogen have evolved by parallel adaptive radiations (lineage 3 and lineage 4 in Figure 1). Both lineages comprise species adapted to same or similar reservoir hosts including zoonotic (e.g. Bh and Br) or human specific (Bq) pathogens. The more recent diversification of lineage 3 including the recently recognized incidental human pathogen Br [29] underlines the importance to study the molecular basis of such lineage diversifications.
In line with the ‘ecological’ parallelism of their radiations, our comparative genomic analyses between lineage 3 and lineage 4 uncover striking evolutionary parallelisms at the molecular level of a likely key innovation - the VirB T4SS – essentially involved in the infection of the mammalian hosts. Chromosomal fixation of this horizontally transferred trait occurred by independent evolutionary events. In both lineages, the arsenal of effector proteins translocated via the VirB T4SS was shaped independently by gene duplications and positive selection of diversified gene copies. This amplification process mostly occurred before the onset of the radiations. Strikingly, beside the diversification of effector proteins encoding the evolutionary conserved ‘AMPylase’ domain (FIC), both lineages have convergently evolved a novel effector class with derived domain structure and tandem-repeated tyrosine-phosphorylation motifs. By these evolutionary processes, large reservoirs of distinct biological functions were invented from a single ancestral effector gene. This functional versatility provides the framework for the adaptive potential of the VirB T4SS. Apparently, the plasticity of the underlying genomic loci seems to have favored the parallel occurrence of these adaptive processes in two distinct lineages, thereby essentially contributing to the parallel radiations of Bartonella.
Animals were handled in strict accordance with good animal practice as defined by the relevant European (European standards of welfare for animals in research), national (Information and guidelines for animal experiments and alternative methods, Federal Veterinary Office of Switzerland) and/or local animal welfare bodies. Animal work was approved by the Veterinary Office of the Canton Basel City on June 2003 (licence no. 1741).
Bc strain 73 [34], B1-1C [31], and Br strain ATCC BAA-1498 [29] were grown routinely for 3–5 days on tryptic soy agar containing 5% defibrinated sheep-blood in a water-saturated atmosphere with 5% CO2 at 35°C. BAR15 [30] and Bs strain R1 [55] were grown under the same conditions on heart infusion agar and Colombia base agar, respectively.
Using the QIAGEN Genomic DNA Isolation kit (Qiagen), DNA was isolated from bacteria grown from single colonies. For 454-sequencing, the DNA was prepared with an appropriate kit supplied by Roche Applied Science and sequenced on a Roche GS-FLX [56]. To assemble the reads, Newbler standard running parameters with ace file output were used. Newbler assemblies were considerably improved by linking overlapping contigs on the basis of the “_to” and “_from” information appended to the read name in the ace files. For the assemblies of Bc, BAR15, B1-1C, Br, and Bs, we obtained a 454-sequence coverage of 35x, 37x, 39x, 39x, and 29x, respectively (for details on 454-sequencing see Table S5). Repeats were identified by analyzing the coverage of each Newbler contig. If the link between two contigs was ambiguous, PCR and long-range PCR were used to confirm contig joins. For the complete assembly of the Bc genome, a library of 35 kb inserts was generated using the CopyControl Fosmid Kit (Epicentre). By end-sequencing of library clones with Sanger technology, 983 high-quality reads were obtained and mapped onto the 454-sequencing-based assembly. Remaining sequence gaps were closed by PCR. The final singular contig was fully covered by staggered fosmid clones indicating a correct assembly of the circular chromosome of Bc. Gene predictions of the genome of Bc and the draft genomes of Bs, Br, BAR15, and B1-1 were performed using AMIGene software [57]. Automated functional gene annotation was conducted with the genome annotation system MaGe [58]. For orthologous genes, the annotation was adopted from the manually annotated genome of Bt [16]. Manual validation of the annotation was performed for the virB and bep genes. By using the “FusionFission” tool of MaGe [58] fragmented genes were identified and the corresponding sequences subsequently examined for 454-sequencing errors. After correcting these errors, the updated sequences were re-annotated as described above. The sequence data of the genome of Bc and the contigs of the draft genomes of Br, BAR15, B1-1C, and Bs is stored on the web-based interface MaGe (Bartonella2Scope, https://www.genoscope.cns.fr/agc/mage/bartonella2Scope) and has been deposited in the EMBL Nucleotide Sequence Database under accession numbers FN645454–FN645524.
Phylogenetic trees were based on nucleotide sequence data. Alignments were generated on protein sequences with ClustalW [59] and back-translated into aligned DNA sequences using MEGA4 [60]. Tree topologies were calculated with maximum likelihood and Bayesian inference methods as implemented in the programs PAUP* [61] and MrBayes [62], respectively. The genome-wide phylogeny of Bartonella was calculated on the basis of 478 orthologous genes of the ten sequenced Bartonella genomes and the genome of Brucella abortus (bv. 1 str. 9-941). Orthologs were determined by using the “PhyloProfile Synteny” tool of MaGe [58] with a threshold of 60% protein identity over at least 80% of the length of proteins being directional best hits of each other. The alignments of the 478 identified genes were concatenated resulting in a total of 515,751 aligned nucleotide sites. Tree topology and branch lengths were obtained by maximum likelihood analysis using the HKY85 model. Bootstrap support values were calculated for 100 replicates. For Bayesian inference, the program MrBayes [62] was run for one million iterations with standard parameters (two runs with four heated Monte-Carlo Markov chains in parallel; number of substitutions = 6; burnin = 25%). For the Bartonella ingroup, single gene trees were calculated with maximum likelihood and tree topology congruency assessed with PAUP*. 471 of the 478 single gene trees revealed the same monophyletic clustering of the eight taxa into lineage 3 and lineage 4 as the genome-wide phylogeny. Further, we performed a recombination analysis for each of the 478 single gene alignments using the GARD algorithm as implemented in the HYPHY package [35]. The GARD analysis was run with the GTR model using a general discrete distribution with three rate classes. To identify statistical significant recombination breakpoints in our alignments, we used the Kishino-Hasegawa test as implemented in the GARDProcess.bf algorithm of the HYPHY package. To include non-sequenced Bartonella species in the genome-wide phylogeny, we used available sequence data from the gltA, groEL, ribC, and rpoB genes (7731 aligned sites). Trees were obtained as described above. MrBayes [62] was run for five million iterations. Branch lengths for tip branches of non-sequenced taxa are calculated on the basis of the four housekeeping genes. Branch lengths for tip branches of sequenced taxa and internal branches separating sequenced and non-sequenced taxa are based on the genomic data set. The maximum likelihood tree only based on the gltA, groEL, ribC, and rpoB genes was inferred as described for the genome-wide phylogeny. Bep gene trees were inferred from nucleotide alignments of either the most C-terminal BID domain including the C-terminus (948 sites), the FIC domain including the N-terminal extension (1,305 sites), or the entire bep sequence of genes harboring FIC domains (3,972 sites). To select an appropriate substitution model, the Akaike information criterion of Modeltest 3.7 [63] and MrModeltest 2.0 [64] was used for the maximum likelihood and Bayesian inference analysis, respectively. For the alignments based on the BID domain or the entire bep gene sequence, we obtained the GTR+G+I model with both programs. For the alignments based on the FIC domain, the TVM+I+G model (Modeltest 3.7) and GTR+G+I model (MrModeltest 2.0) were selected. Trees were inferred with the parameters provided by these models as described above. MrBayes [62] was run for one million iterations. The Neighbor-joining phylogeny of different Bartonella isolates in Figure S2 was inferred from a 242 nt segment of the gltA gene with the program MEGA4 [60]. Bootstrap values were calculated for 1,000 replicates.
Based on the four available genomes, orthologous genes for each of the two lineages 3 and 4 were determined by using the “PhyloProfile Synteny” tool of MaGe [58]. The threshold was set to 30% protein identity over at least 60% of the length of proteins being directional best hits of each other. The same tool was used to detect genes without orthologs. By comparing these automatically identified orthologs and non-orthologs, genes present in neither of the two lists were detected and manually assigned to one of the two lists. Alignments were generated and a GARD recombination analysis conducted as described above. To obtain the average dn/ds value (ω) of each ortholog, the arithmetic mean of pair-wise dn/ds values (calculated by the method of Yang and Nielsen implemented in PAML 4.1 [65]) was used. Site tests of positive selection were performed with PAML 4.1 using the CodeML module [65]. To detect positive selection model M1a (NearlyNeutral) vs. model M2a (PositiveSelection) and model M7 (beta) vs. model M8 (beta+ω) were analyzed. PAUP* [61] was used to infer maximum likelihood trees for each set of orthologs. For the CodeML control file, standard parameters were used. The relative significance of model M2a (PositiveSelection) vs. model M1a (NearlyNeutral) and model M8 (beta+ω) vs. model M7 (beta) was assessed using likelihood-ratio-tests (two degrees of freedom). Genes for which significant positive selection was detected were inspected for alignment errors potentially affecting the results of this analysis. If necessary, the alignments were manually modified and the CodeML analysis repeated. Phylogenetic branches were tested for positive selection by using the TestBranchDNDS.bf module implemented as standard analysis tool in HyPhy [66].
Ten weeks old female WISTAR rats obtained from RCC-Füllinsdorf were housed in an BSL2-animal facility for two weeks prior to infection allowing acclimatization. For inoculation, bacterial strains were grown as described above, harvested in phosphate-buffered saline (PBS), and diluted to OD595 = 1. Rats were anesthetized with a 2–3% Isuflurane/O2 mixture and infected with 10 µl of the bacterial suspension in the dermis of the right ear. Blood samples were taken at the tail vein and immediately mixed with PBS containing 3.8% sodium-citrate to avoid coagulation. After freezing to −70°C and subsequent thawing, undiluted and diluted blood samples were plated on tryptic soy agar and heart infusion agar containing 5% defibrinated sheep-blood. CFUs were counted after 8–12 days of growth.
Nucleotide distances were calculated with the program MEGA4 [60] for the alignments based on the genome-wide dataset and the four housekeeping genes. The numbers of base substitutions per site from averaging over all sequence pairs within and between groups were calculated. Codon positions included were 1st, 2nd, and 3rd. All positions containing gaps and missing data were eliminated from the dataset (Complete deletion option).
To construct the plasmids pPE2002 and pPE2004, bep genes BARCL_1034 (Bc) and BARRO_80017 (Br) were amplified from genomic DNA with primer pairs containing flanking BamHI/NotI sites: prPE453 (ATAAGAATGCGGCCGCGATGAAAAC-CCATAACACTCCTG)/prPE454 (CGGG-ATCCTTAATGTGTTATAACCATCGTTC) and prPE455 (ATAAGAATGCGGCCGCG-ATGAATTTTGGAGAAAAGAAAAAAATG)/prPE456 (CGGGATCCTTAAATAGC-TACAGCTAACGATTTTTTC), respectively. PCR products were digested with the enzymes BamHI and NotI and ligated into the BamHI/NotI sites of the backbone of plasmid pAP013 (kindly provided by Arto Pulliainen). The resulting constructs pPE2001 (BARCL_1034) and pPE2003 (BARRO_80017) were cut with NotI and ligated with a GFP fragment obtained from NotI digested pAP013. The plasmid pPE2007 was constructed by cutting bepE of B. henselae from plasmid pRO1100 (kindly provided by Rusudan Okujava) with NotI and BamHI and ligating it into pAP013. All plasmid DNA isolations and PCR purifications were performed with Macherey-Nagel and Promega columns according to manufacturer's instructions.
The protocol for growth and transfection of HEK293T was performed as described previously [18]. 36 h after transfection, cells were incubated for 10 minutes with 10 ml Pervanadate medium (5 ml PBS containing 100 mM orthovanadate and 200 mM H2O2, incubated for 10 min with 500 µl Catalase [2 mg/ml in PBS] before 45 ml M199 medium were added). After washing three times with 7 ml of PBS at room temperature, cells were scraped off and resuspended in 1 ml of ice-cold PBS containing 1 mM EDTA, 0.5 mM phenylmethylsulfonyl fluoride (PMSF), 1 mM orthovanadate, 1 mM leupeptin, and 1 mM pepstatin and collected by centrifugation (3,000g at 4°C for 60 sec). The resulting pellet was lysed in 300 µl of ice cold modified RIPA buffer (50 mM Tris-HCl [pH 7.4], 75 mM NaCl, 1 mM EDTA, 1 mM orthovanadate, 1 mM leupeptin, 1 mM pepstatin) for 1 hour at 4°C. The lysate was centrifuged (16,000g at 4°C for 15 min) and 12 µl of anti-HA-agarose (Sigma) added to the supernatant. After 150 min of incubation at 4°C on a slowly turning rotation shaker, the agarose was washed three times with 300 µl of modified RIPA buffer (3,000g for 10 sec). The affinity-gel pellet was then resuspended in 20 µl of modified RIPA buffer, 20 µl of SDS-sample buffer (2×) were added, and the sample was heated for 5 min at 95°C. Proteins were separated on a 10% SDS-polyacrylamide gel, blotted on a nitrocellulose membrane (Hybond-C Extra, Amersham Pharmacia), and examined for tyrosine phosphorylation by using monoclonal antibody 4G10 (Millipore) and anti-mouse IgG-horseradish peroxidase (HRP) afterwards. The HRP-conjugated antibody was visualized by enhanced chemiluminescence (PerkinElmer). For visualization of the signal from GFP-fusion proteins, the membrane was subsequently incubated in 4% PBS-Tween containing 0.02% NaN3 and anti-GFP antibody (Invitrogen), followed by incubation with anti-mouse IgG-HRP and visualization by enhanced chemiluminescence.
|
10.1371/journal.pgen.1006593 | Direct Keap1-Nrf2 disruption as a potential therapeutic target for Alzheimer’s disease | Nrf2, a transcriptional activator of cell protection genes, is an attractive therapeutic target for the prevention of neurodegenerative diseases, including Alzheimer’s disease (AD). Current Nrf2 activators, however, may exert toxicity and pathway over-activation can induce detrimental effects. An understanding of the mechanisms mediating Nrf2 inhibition in neurodegenerative conditions may therefore direct the design of drugs targeted for the prevention of these diseases with minimal side-effects. Our study provides the first in vivo evidence that specific inhibition of Keap1, a negative regulator of Nrf2, can prevent neuronal toxicity in response to the AD-initiating Aβ42 peptide, in correlation with Nrf2 activation. Comparatively, lithium, an inhibitor of the Nrf2 suppressor GSK-3, prevented Aβ42 toxicity by mechanisms independent of Nrf2. A new direct inhibitor of the Keap1-Nrf2 binding domain also prevented synaptotoxicity mediated by naturally-derived Aβ oligomers in mouse cortical neurons. Overall, our findings highlight Keap1 specifically as an efficient target for the re-activation of Nrf2 in AD, and support the further investigation of direct Keap1 inhibitors for the prevention of neurodegeneration in vivo.
| As our population ages the incidence of neurodegenerative diseases, including Alzheimer’s disease (AD), is predicted to increase dramatically. Despite providing important symptomatic relief, existing treatments for such conditions do not slow-down disease progression, and this will cause an overwhelming future burden on our healthcare system and immense suffering for many more patients and their families. Nrf2 is a gene that normally protects cells from stressful conditions. Although we don’t know why, Nrf2 is reduced in the brains of AD patients and this may explain the increased susceptibility of neurons to damage in neurodegenerative diseases. Our research, using a fruit fly model, identifies Keap1, a negative regulator of Nrf2, as a valid target for the rescue of AD-related Nrf2 defects and the subsequent prevention of neuronal degeneration. Moreover, we show that a new compound, which directly blocks the binding between Nrf2 and Keap1, can prevent toxicity of the AD-initiating Aβ peptide in mouse neurons. Hence, our study provides strong evidence that direct Keap1-Nrf2 disruptors can specifically target the defects in Nrf2 activity observed in neurodegenerative diseases, and supports the further development of such compounds as potential new drugs to prevent neuronal decline AD and other neurodegenerative conditions.
| The transcription factor Nrf2 (nuclear factor E2-related factor 2) targets cellular defence genes containing antioxidant response elements (ARE), which include antioxidant enzymes (glutamate cysteine ligase; GCL), drug metabolising enzymes (cytochrome P450s, glutathione S-transferases; GSTs), molecular chaperones, DNA repair enzymes and proteasome subunits[1]. Activation of these protective genes in response to Nrf2 enables the cell to maintain redox balance and to remove damaged proteins under conditions of oxidative and xenobiotic stress.
Such cellular stress is a key feature of several neurodegenerative diseases. Markers of oxidative damage are increased in the brains of Alzheimer’s disease (AD)[2,3], Parkinson’s disease (PD)[4–6], Huntington’s disease (HD)[7] and, in CSF, of Amyotrophic Lateral Sclerosis (ALS) patients[8]. Some evidence also suggests that AD, PD and ALS patients have reduced xenobiotic metabolism [9] and that the AD-causing Aβ42 peptide may act as a xenobiotic[10]. As Nrf2 is inhibited in several neurodegenerative diseases[1, 11], including AD[11] and ALS[12], as well as in APP/PS1 mutant mouse models of AD[13], deficits in this important cellular protection pathway may, in part, explain the cellular damage associated with these conditions. Conversely, Nrf2 over-expression protects against toxicity induced by the Aβ42 peptide in mammalian cells[13,14], and prevents neuronal pathology in mouse models of ALS[15], PD[16] and AD[17]. Activation of Nrf2 is, therefore, increasingly implicated as an attractive target for the prevention of several neurodegenerative conditions.
Potent activators of Nrf2 have been developed[18], and confer protection against chemically-induced neurotoxic insults[19] and improve memory deficits in mouse models of AD[20]. Many of these Nrf2 activators, however, have been reported to exert toxicity due to off-target effects[21][22]. Additionally, unregulated activation of Nrf2 can have detrimental consequences, with prolonged, ubiquitous activation shortening lifespan in Drosophila[23], and mutations in the Nrf2 inhibitor Keap1 (kelch-like ECH-associating protein 1) causing cancer in humans[24]. Hence, a better understanding of the mechanisms by which Nrf2 function is inhibited in neurodegenerative disease may enable the design of drugs targeted for the prevention of these conditions with minimal side-effects.
Nrf2 activity is tightly regulated by two main inhibitors, Keap1 and GSK-3 (glycogen synthase kinase-3)[25] (S1 Fig). GSK-3 plays a well-established role in the pathogenesis of AD[26], and GSK-3 inhibitors, including lithium, prevent pathology in animal models of AD[26,27], ALS[28] and PD[29]. Emerging evidence also suggests that inhibiting Keap1 can ameliorate neuronal degeneration, with Keap1 RNA interference (RNAi) protecting against Aβ42[30] and MPTP[31] toxicity in cells, and heterozygous loss of Keap1 protecting against neuronal pathology in Drosophila models of PD[32,33]. Both GSK-3 and Keap1 may, therefore, serve as valid candidates for mediating the inhibition of Nrf2 in neurodegenerative diseases, and their inhibition may enable the prevention of Nrf2 deficits, neuronal stress and degeneration specifically in these conditions. A comparative analysis of the efficiency of their inhibition in rescuing Nrf2 deficits in neurodegenerative diseases specifically, however, is required.
We aimed therefore to investigate the role of Nrf2 in mediating the protective effects of GSK-3 and Keap1 inhibition against Aβ42 toxicity. Using an inducible Drosophila model of AD[34], we confirmed that Aβ42 inhibits activity of the fly homolog of Nrf2 (cap-n-collar isoform C, cncC[35]) in neurons, consistent with previous findings in mice[13]. Both inhibition of GSK-3, using lithium, and loss-of-function mutations in Keap1 protected against Aβ42 toxicity in this model. We found, however, that neuronal protection in response to Keap1 inhibition correlates with the rescue of Aβ42-induced Nrf2 defects, whereas lithium treatment appears to exert neuro-protection independently of Nrf2.
Consistent with Nrf2 activation, Keap1 inhibition prevented the enhanced sensitivity of Aβ42-expressing flies to xenobiotic stress, but exerted minimal protection against oxidative damage in comparison to lithium treatment. Combined modulation of Keap1 and lithium additively protected against Aβ42 toxicity, in comparison with either treatment alone, but did not improve their respective effects on xenobiotic and oxidative damage. This further supports the divergent beneficial effects of these manipulations. Down-regulation of Keap1 alone additionally protected against Aβ42 toxicity by mechanisms correlating with enhanced degradation of Aβ42 peptide.
Overall our data highlight Keap1 as an efficient target for the amelioration of Nrf2 deficits and protection against neuronal damage in AD. Finally, we show for the first time that a newly-described direct inhibitor of the Keap1-Nrf2 protein-protein interaction[22] can indeed protect against the synapto-toxicity of naturally-derived Aβ oligomers in primary mouse neuronal cultures. As current Nrf2 activators may exert toxicity due to off-target effects, our data suggest that blocking the specific interaction of Nrf2 with Keap1 may provide an exciting new avenue for the discovery of disease-modifying treatments for AD, and potentially other neurodegenerative conditions.
Over-expression of APP in mice inhibits expression of Nrf2 target genes[13]. Comparing expression, at equivalent levels[36], of various Aβ species in adult fly neurons, we have shown that the presence of aggregating Aβ42 peptides may specifically mediate this effect. Using Nrf2/cncC reporter flies (gstD1-GFP; [35]), RU486 induction of a single, site-directed, copy of Arctic mutant Aβ42 (ArcAβ42), but not WT Aβ40 or Aβ42 peptides, significantly reduced Nrf2/cncC activity compared to un-induced gstD1-GFP-expressing controls (Fig 1A and 1B). Only ArcAβ42 flies develop pathological phenotypes under these expression conditions. Moreover, inducing high levels of WT Aβ42 using an independent random-insertion line, which does cause neuronal decline [37], also significantly reduced Nrf2/cncC activity (Fig 1C). This suggests that the inhibition of Nrf2 may be related to the concentration-dependent ability of Aβ42 to aggregate and exert toxicity.
Consistent with a specific inhibition of neuronal Nrf2/cncC activity in response to Aβ42 expression, ArcAβ42 reduced GFP levels in the heads, but not bodies (S2A Fig), of RU486-induced flies. Microarray analyses further revealed several pathways and processes that might mediate toxicity in response to ArcAβ42 expression in fly neurons (Fig 1D, Arc Aβ42). Comparing this data-set to a previously published microarray analysis using flies ubiquitously over-expressing Nrf2/cncC (Fig 1D, cncC[38]), we found that Aβ42 expression induced reciprocal effects on gene ontology (GO) categories normally regulated by the Nrf2 pathway. Aβ42 activated processes that are down-regulated, and inhibited processes that are up-regulated, by cncC (Fig 1D & S2B Fig), further confirming its suppressive effect on Nrf2 signalling.
Overall these findings suggest that the inhibition of Nrf-2 in AD is robustly conserved in Drosophila, and that this deficiency is caused by the Aβ42 peptide directly. Moreover, as we also observed a suppression of Nrf2/cncC in flies over-expressing human 0N3R tau or the ALS-related C9orf72 (GR) 100 di-peptide repeat protein (DPR) [40] in adult neurons (S3A and S3B Fig), our findings suggest that the accumulation of toxic proteins may lead to the generalised defect in Nrf2 signalling observed in neurodegenerative diseases.
Drosophila thus provide an excellent context for further analysis of the mechanisms by which Aβ42 regulates Nrf2 in vivo. As cncC, MafS and Keap1 mRNA transcripts were unaltered by Aβ42 in our flies (S3C Fig), we hypothesised that Aβ42 modulates Nrf2 activity by altering its biochemical interaction with upstream regulators.
Keap1 is a well-known negative regulator of Nrf2 activity in response to oxidative and xenobiotic stressors[41], but its role in mediating damage in response to neurotoxic proteins has not been widely studied in vivo. Genetically reducing Keap1, alone, extended lifespan (Fig 2A) and rescued neuronal-specific motor defects in both WT (S4 Fig) and ArcAβ42-expressing flies (Fig 2B) using two independent alleles, Keap1del [42] and Keap1EY5[35]. GSK-3 has more recently been shown to inhibit Nrf-2[43], independently of Keap1, but has long been implicated in the pathogenesis of several neurodegenerative diseases including AD[26]. We have previously shown that lithium, a well-described GSK-3 inhibitor[44–46], prevents Aβ42 toxicity in our fly model [34]. Thus both GSK-3 and Keap1 may serve as effective targets for the prevention of neurodegeneration in AD.
We next compared the effects of manipulating both Keap1 and GSK-3 regulatory pathways on Aβ42 toxicity and Nrf2/cncC activity in a parallel study. Using a dose of lithium (25 mM) that prevented Aβ42-induced lifespan-shortening to a similar extent as heterozygous loss-of-function mutations in Keap1 (Fig 2C, P = 0.739 comparing +RU, + LiCl 25 mM to +RU, Keap1 del), this comparative analysis revealed that both manipulations rescued lifespan-shortening in Aβ42-expressing flies in an additive manner (Fig 2C, P<0.001 comparing +RU + LiCl 25 mM to +RU + LiCl 25 mM, Keap1 del), consistent with an independent mechanism of action.
At these therapeutic levels, genetically reducing Keap1, alone, significantly rescued the decline in Nrf2/cncC reporter expression observed in Aβ42-expressing flies (Fig 2D, p<0.05 comparing +RU vs +RU, Keap1 del). By contrast, although 25 mM lithium was sufficient to rescue Aβ42 toxicity, a dose of 50–100 mM lithium, which may exert toxic effects [47], was required to rescue Nrf2/cncC activity as measured using both microarray (S5A Fig) and gstD2 expression (S5B Fig) analyses. A heterozygous loss-of-function mutation in cncC (cncCK6[48]; S5C Fig) did not alter the ability of lithium to protect against Aβ42-mediated lifespan-shortening, further suggesting that Nrf2 is not required for lithium to exert its protective effects in vivo. Consistent with our previous finding that lithium reduces Aβ42 levels, and prevents toxicity, by blocking translation[47], early intervention concurrent with RU486 induction was required for lithium to protect against Aβ42-induced climbing defects (S5D Fig) and was correlated with reduced Aβ42 levels from the point of induction (S5E Fig). This suggests that, at therapeutically active concentrations, the protective effect of lithium against Aβ42 toxicity is not predominantly mediated by activating Nrf2, but rather by blocking Aβ42 peptide accumulation through inhibition of translation.
Our data provide the first evidence that specific Keap1 inhibition can protect against Aβ42 toxicity in vivo, and they highlight Keap1 as a more efficient target for the prevention of Aβ42 peptide-induced Nrf2 inhibition in comparison with lithium treatment.
As Nrf2 plays an important role in protecting against oxidative and xenobiotic stress, we investigated the nature of the Aβ42-induced molecular damage that was ameliorated by Keap1 inhibition (Fig 3). UAS-ArcAβ42>elavGS control and heterozygous Keap1 mutant flies were induced with RU486 for 14 days before measuring sensitivity to xenobiotic (DDT; dichlorodiphenyltrichloroethane) or oxidative (hyperoxia or paraquat) stressors (see methods).
Reducing Keap1 alone protected against Aβ42-induced sensitivity to DDT (Fig 3A) and paraquat (Fig 3C), but not hyperoxia (Fig 3B; +RU vs +RU, Keap1 del). Lithium treatment, on the other hand, protected against sensitivity of Aβ42 flies to DDT to a similar extent as Keap1 inhibition (Fig 3A), but more significantly protected against sensitivity to both hyperoxia (Fig 3B) and paraquat (Fig 3C)-induced oxidative damage. Moreover, when combined, lithium did not add significantly to the ability of reduced Keap1 to protect against the sensitivity of Aβ42 flies to DDT (Fig 3A, +RU, +LiCl 25 mM Keap1 del vs +RU, Keap1 del) and, conversely, reduced Keap1 did not add to the ability of lithium to protect against paraquat (Fig 3B) or hyperoxia (Fig 3C, +RU, +LiCl 25 mM Keap1 del vs +RU, +LiCl 25 mM). Consistent with the apparently different mechanisms of neuronal protection by Keap1 and lithium, this finding suggests that Keap1 inhibition acts mainly through protection against xenobiotic stress and lithium predominantly by limiting oxidative damage.
In addition to increasing Nrf2/cncC activity, the protective effect of reducing Keap1 on Aβ42 toxicity correlated with a reduction in Aβ42 peptide levels (Fig 4A). Although Aβ42 mRNA was also slightly reduced by down-regulation of Keap1 (Fig 4B), the peptide levels were not altered until 14 days of age (Fig 4C), suggesting that reducing Keap1 may clear Aβ42 peptide by activating protein degradation mechanisms.
We have shown previously that Aβ42 peptide is stable for several weeks following induction in our inducible fly model[49]. To assess whether the peptide is indeed degraded as a consequence of reducing Keap1, we induced Aβ42 expression in Keap1del and control flies for one week with RU486 then measured Aβ42 levels by ELISA at several time-points following transfer to non-RU486-containing medium (Fig 4D). Aβ42 levels were equivalent between heterozygous Keap1del and control flies at the end of the induction period (0d since switch-off; age 7d), but then declined only in flies with reduced Keap1, starting from 3 days following switch-off (age 10d). This finding confirms that inhibition of Keap1 does indeed lead to degradation of the Aβ42 peptide. Moreover, analysis of insoluble proteins one week following RU486 induction for 10 days (Fig 4E), revealed that inhibition of Keap1 reduced the level of aggregated Aβ42.
Keap1-Nrf-2 signalling has been implicated in both autophagy and proteasomal degradation. p62, a selective autophagy substrate, competes with Nrf-2 for binding to Keap1[50][51] and Keap1 may activate autophagy directly by binding to p62[52]. However, we did not observe any alteration in autophagy activity upon reduction of Keap1 in the Aβ42-expressing flies, as measured by western blotting using an antibody specific for Drosophila ATG8 (Fig 5A). Proteasome subunits are transcriptional targets of Nrf2/cncC[53] and over-expression of cncC in flies, either directly or by reducing Keap1 levels, increases proteasome expression and activity[23]. Accordingly, we detected an increase in proteasome activity in the heads of Aβ42-expressing flies upon reduction of Keap1 (Fig 5B), at a time-point coinciding with reduced Aβ42 levels (age 14d). Pharmacological inhibition of this enhanced proteasome activity (S6A Fig), however, did not prevent the degradation of Aβ42 in response to Keap1 inhibition (S6B Fig), suggesting that this may not be the mechanism by which loss of Keap1 induces Aβ42 degradation. Further studies are therefore required to elucidate the precise clearance mechanisms mediating Aβ42 degradation following Keap1 inhibition.
Our data highlight Keap1 as a valid in vivo therapeutic target for AD. Current activators of Nrf2 are electrophilic compounds, which commonly act by modifying cysteine residues on Keap1 and thus disrupt its interaction with, and inhibition of, Nrf2[54]. Of these, the synthetic triterpenoids, such as Bardoxolone methyl (CDDO-Me), are potent Nrf2 activators and have exhibited therapeutic potential against several diseases, including neurodegeneration, in animal models and clinical trials[55]. One problem with the mechanism of action of these electrophilic Nrf2 activators, however, is the modification of cysteine residues on other targets[21,22], which may lead to side-effects in humans[22,55]. To improve specificity, compounds have more recently been designed to directly disrupt Keap1-Nrf2 binding, and these enhance Nrf2 activity in vitro [56] and in cells[22,57]. Activity of these compounds against disease models, however, remains to be examined.
We tested the ability of a recently published direct Keap1-Nrf2 inhibitor, 22h (Fig 6A[22]), to protect against exogenous amyloid toxicity in cultured cells. At their EC50 concentrations for Nrf2 activation (S7 Fig), we first performed a comparative analysis of 22h, CDDO-Me and the GSK-3 inhibitor, TDZD-8, for protection against natural Aβ oligomer-induced toxicity[58] in SH-SY5Y cells (Fig 6B). Interestingly, Nrf2 and Keap1 protein levels were increased following 24 h Aβ oligomer treatment (S8 Fig), suggesting that the Keap1-Nrf2 pathway is indeed dysregulated by acute Aβ exposure. TDZD-8 was a poor activator of Nrf2, inducing its target gene NQO1 only at a single concentration of 1 μM (S7C Fig), and exerted toxicity in control SH-SY5Y cells (Fig 6B). Although CDDO-Me was a more potent activator of Nrf2 (S7D Fig), 22h significantly protected against Aβ toxicity, in comparison with both CDDO and TDZD-8, in this experimental paradigm (Fig 6B). This comparative analysis supports our suggestion that Keap1 may serve as a more efficient target, than inhibition of GSK-3, for the rescue of Nrf2-dependent effects of Aβ42 toxicity. Moreover, as TDZD-8 has previously been shown to protect against Aβ toxicity in primary neuronal cultures, our data suggest that the threshold for this protective effect of GSK-3 inhibition lies below that for its effects on Nrf2 activity. These findings indicate that direct Keap1-Nrf2 disruptors may protect against Aβ42 toxicity more effectively than established Nrf2 activators.
To further test the effects of Keap1-Nrf2 disruption on neuronal function in response to Aβ oligomers, we measured the effects of 22h in primary mouse neurons (Fig 6C). Conditioned medium obtained from Tg2576 mouse neurons (Tg2576CM) has previously been shown to contain oligomeric Aβ species at concentrations similar to those in human CSF, and to reduce spine density of GFP-transfected WT mouse cortical neurons[59]. We pre-treated GFP-transfected WT mouse neurons of 12 d in vitro (DIV), with 1 or 10 μM 22h for 16 h prior to administration of either wt or Tg2576 conditioned media (wtCM or Tg2576CM), with continued 22h treatment, for a further 24 h before examining spine density (see Materials & Methods). As previously reported[59], Tg2576CM reduced total spine density of cortical neurons compared to wtCM (Fig 6C and 6D). Strikingly, spine density was rescued by treatment with compound 22h (Tg2576CM, 0.01% DMSO vs Tg2576CM, 10 μM 22h) at non-toxic doses (wtCM, 0.01% DMSO vs wtCM, 10 μM 22h). Aβ42 levels in conditioned media were unchanged following treatment with 10 μM 22h (Fig 6E) which further correlated with increased expression of Nrf2 target genes[60] (Fig 6F).
These findings suggest that directly blocking the Keap1-Nrf2 interaction can protect neurons downstream of extracellular amyloid toxicity.
As the prevalence of ageing-related neurodegenerative diseases, such as Alzheimer’s, is predicted to rise dramatically[61], and with a lack of disease-modifying therapies currently available, there is an urgent need to find new therapeutic targets to slow down neuronal degeneration in these conditions. Accumulating evidence suggests that down-regulation of the cell protective transcription factor Nrf2 may enhance neuronal susceptibility to molecular damage [1, 62], and that its activation may confer neuronal protection[63]. Nrf2 is, therefore, a promising novel target for the prevention of several neurodegenerative diseases, but its activation must be tightly regulated to have beneficial effects[23,24]. Moreover, although current activators of Nrf2, including electrophilic compounds, such as sulforaphane, and the more potent bardoxolone methyl (CDDO-Me), can exert neuroprotective effects[20,64], all of these compounds are indirect activators and likely to exert multiple off-target effects[55], potentially leading to toxicity. A better understanding of the precise mechanisms regulating Nrf2 activity (S1 Fig) under particular pathological conditions would, therefore, enable the design of drugs targeted to the prevention of specific diseases with minimal side-effects. Our work aimed to investigate the mechanisms regulating Nrf2 inhibition specifically in response to the AD-related Aβ42 peptide, with a view to identifying novel targets for the prevention of AD and other ageing-related neurodegenerative conditions.
Our study firstly confirms, in Drosophila, previous reports that Nrf2 target genes are down-regulated in AD brain[11], and in mouse models of AD[13,65,66]. Some studies suggest that this inhibition may be mediated by Aβ42 directly, with exogenous, synthetic, Aβ42 peptide reducing Nrf2 target gene expression in primary mouse neurons[13], and blocking Nrf2 nuclear translocation following injection into the hippocampus of rat[65] and mouse[66] brain. As Aβ42 peptide is expressed in our fly model independently of APP processing, our study further confirms that this direct effect on Nrf2 activity occurs in response to naturally-derived Aβ42 peptide conformations in vivo (Fig 7B). Importantly, other disease-related proteins, including tau and C9orf72 DPRs, also suppressed Nrf2/cncC signalling, but the non-toxic Aβ40 peptide did not. This suggests that inhibition of the Nrf2 pathway is a generalised response to the accumulation of aberrant proteotoxic proteins.
The mechanisms mediating the effect of Aβ42 on Nrf2 activity remains to be established. Although a recent report, using sweAPP-expressing cells, has suggested that Nrf2 transcripts can be replenished by altering DNA methylation [67], Aβ42 did not alter mRNA expression of Nrf2/cncC, or of its co-transcription factor MafS, in our Drosophila model. This suggests that in vivo Nrf2 de-regulation in AD may be post-transcriptional. We have shown that Aβ42 oligomers increased Nrf2 and Keap1 proteins in SH-SY5Y cells after 24 h treatment, suggesting that Nrf2 is initially stabilised in response to acute amyloid exposure. This may represent a protective response to the initial toxic insult, as similar effects on Nrf2 have been observed at early time-points following transient focal ischaemia and correlate with preservation of peri-infarct regions of the brain under these conditions [68]. Keap1 is also an Nrf2 target gene [35], however, and may subsequently be upregulated to control Nrf2 activity following the initial exposure to Aβ42 in our study. If sustained, this increase in Keap1 levels could provide a potential mechanism for the inhibition of Nrf2 observed in AD brain and other chronic neurodegenerative conditions [11]. This hypothesis is supported by observations that Nrf2 protein is downregulated and Keap1 upregulated in mouse brain following 15 days exposure to synthetic Aβ42 [66]. Further work is required, however, to investigate the detailed timing of these events following chronic in vivo exposure to natural Aβ oligomers.
Since Aβ42 did not directly affect Nrf2/cncC gene expression, we investigated the role of its upstream inhibitors, GSK-3 and Keap1, on toxicity. GSK-3 has a well-documented role in Alzheimer’s, and is suggested to provide a pivotal connection between the characteristic amyloid plaque and neurofibrillary tangle pathologies of this disease[26]. Activation of GSK-3 has been proposed to exert neuronal toxicity by many mechanisms, including increasing apoptosis and inflammation, and impairing axonal transport, synaptic function, cell cycle regulation and adult neurogenesis [26]. Conversely, several GSK-3 inhibitors, including lithium, protect against AD-pathology in mice[69–71] and some have been tested in clinical trials for AD[72]. Although GSK-3 inhibition increases Nrf2 activity in AD models[73], however, only one study has shown epistatically that Nrf2 mediates the neuroprotective effect of lithium treatment against paraquat toxicity in cells[74]. The role of Nrf2 in mediating the protective effect of GSK-3 inhibition in AD has also not been empirically investigated.
Our current study confirms that lithium treatment increases transcription of Nrf2 target genes in a dose-dependent manner. Concentrations of lithium required to activate Nrf2 (≥50 mM-100 mM LiCl), however, have previously been shown to exert toxicity in Drosophila[42,47]. Lower doses of lithium (25 mM) were sufficient to protect against Aβ42 toxicity, to a level comparable with reducing Keap1, but did not significantly activate Nrf2/cncC. Moreover, genetically reducing Nrf2/cncC function did not prevent the lifespan-extending effects of lithium, in Aβ42 flies, even at a maximizing concentration (50 mM). Our data, therefore, suggest that lithium mediates neuroprotection against Aβ42 independently of its effects on Nrf2 (Fig 7B). Consistent with our previous observations that lithium (10–100 mM) reduces Aβ42 peptide levels by inhibiting translation[47], early lithium administration reduced Aβ42 peptide, from the point of induction, in our current study and this was required for lithium to exert its protective effects.
These findings are further supported by our observation that the specific GSK-3 inhibitor TDZD-8 has a narrow window for Nrf2 activation in cultured cells, inducing its target gene NQO1 only at a single concentration of 1 μM (Fig 6B). At concentrations ≤ 1 μM TDZD-8 has been described to prevent Aβ-induced reductions in spine density in mouse neurons[59], but doses ≥ 1 μM exerted toxicity (Fig 6B and [59]). Together these data suggest that the threshold for protection against Aβ toxicity by GSK-3 inhibitors lies below the concentration required for activation of Nrf2 in neuron. This supports the hypothesis that blocking GSK-3 exerts its protective effects independently of Nrf2.
Contrary to our findings with lithium, we provide the first in vivo evidence that Keap1 inhibition can exert neuroprotective effects in response to Aβ42 by ameliorating deficits in Nrf2 activity (Fig 7C). As with Drosophila models of PD[32,33], heterozygous loss of Keap1 protected against Aβ42-induced lifespan-shortening and climbing defects in the fly. Moreover, we show that this protective effect of reducing Keap1 correlates significantly with a rescue of Nrf2/cncC activity in response to Aβ42.
Only one previous study has addressed the role of Keap1 inhibition, specifically, in AD. Keap1 RNA interference increased expression of Nrf2 target genes, protected against synthetic Aβ42-induced cytotoxicity and oxidative damage to proteins and lipids, and enhanced autophagy in cultured cells[30]. Our study adds to these in vitro findings, demonstrating that Keap1 inhibition rescues Aβ42-induced Nrf2 deficits, and protects against neuronal decline in vivo. These effects occur in correlation with prevention of xenobiotic and, to a lesser extent, oxidative damage as well as the degradation of endogenous, aggregated, Aβ42 peptide.
Using primary cortical neurons, our findings further confirm that direct pharmacological inhibition of Keap1-Nrf2 binding can protect against neuronal damage downstream of extracellular Aβ42 oligomers. The mechanisms by which reducing Keap1 enhances degradation of Aβ in vivo, however, requires further investigation. Autophagy, as measured by cleavage of ATG8, was unaltered in response to Keap1 inhibition in our Aβ42 expressing flies. Consistent with the established effect of Nrf2 on transcription of proteasome subunits[23], loss of Keap1 enhanced proteasome activity in Aβ42 flies, but blocking this increase, using the proteasome inhibitor Bortezomib, did not prevent the reduction in Aβ42 levels. This suggests that the reduction of aggregated Aβ42 in response to Keap1 inhibition may not be directly mediated via enhanced autophagy or proteasomal degradation. Nrf2 has also been implicated to play a role in the unfolded protein response (UPR), serving as a target for PERK[75] and activating transcription of several components of the ER associated degradation (ERAD) pathway[76], including chaperones and ubiquitin-conjugating enzymes, in addition to proteasome subunits and autophagy. Future studies will therefore be required to investigate the functional role of such protein quality control pathways, potentially upstream of the proteasome and autophagy, in the clearance of Aβ42 following Keap1 inhibition in vivo.
Overall, our data point to Keap1, in comparison with GSK-3, as an efficient target for Nrf2 activation in response to Aβ42 toxicity in vivo. Recent advances in the development of direct inhibitors of the Keap1-Nrf2 binding domain may, therefore, enable the prevention of Nrf2 deficits in neurodegenerative diseases with minimal side-effects. Our study shows for the first time that a direct, small molecule inhibitor of Keap1-Nrf2 binding, 22h[22], can ameliorate synaptotoxicity in response to naturally-derived Aβ oligomers in mouse cortical neurons. As synaptic loss correlates well with cognitive decline[77], our work suggests that these compounds show promise as therapeutic agents for AD. Moreover, as this is the first demonstration that these compounds can prevent neuronal toxicity, our findings warrant their investigation in other neurodegenerative conditions.
Both GSK-3 inhibitors[71] and Nrf2 activators[55] can exert toxicity due to off-target effects and thus it is important that modifiers of these pathways achieve therapeutic benefits at low doses. More specific inhibitors of each of these targets are currently being developed[22,71]. It has also been postulated that combined inhibition of both GSK-3 and Keap1 may activate Nrf2, and confer protection, at lower doses than either intervention alone, hence limiting their detrimental effects[25].
We present the first in vivo study showing that combined Keap1 and GSK-3 inhibition, by lithium, confers additive protective effects against Aβ42 toxicity. Keap1 deletion and lithium treatment extended lifespan and improved climbing ability of Aβ42-expressing flies to a greater extent than either manipulation alone. Additionally, reducing Keap1 limits the dose of lithium required to reach maximal levels of protection since heterozygous Keap1del combined with low dose (25 mM) lithium extended lifespan of Aβ42 flies to a similar extent as high dose (50 mM) lithium alone. Our data, however, do not predict that these additive protective effects are due to mechanisms converging on Nrf2. Rather, low dose lithium treatment exerts protective effects independently of Nrf2 activation, whereas the neuroprotective effects of Keap1 inhibition correlate strongly with increasing Nrf2. Hence, the additive nature of combined lithium and Keap1 inhibition appears to be mediated through divergent, complementary protective mechanisms.
Oxidative[2,3] and xenobiotic damage[9] are key features of AD brain and may potentially be explained by the down-regulation of Nrf2, which is also observed in several neurodegenerative diseases. Consistent with this idea, Keap1 inhibition and, therefore, Nrf2/cncC activation, correlated strongly with protection against Aβ42-induced sensitivity to the xenobiotic DDT in our fly model. Conversely, Keap1 inhibition exerted minimal protection against oxidative damage in comparison with lithium treatment, which strongly protected against Aβ42-induced sensitivity to both paraquat and hyperoxia-induced damage. As therapeutic concentrations of lithium, which can prevent oxidative damage, did not strongly activate Nrf2/cncC in our flies, this suggests that rescue of Nrf2 activity is not required to protect against oxidative damage in response to Aβ42. Although pharmacological activation of Nrf2, has been shown to protect against oxidative damage, in response to Aβ42 peptide, in cells[78,79] and in animal models of AD[65,66], conflicting reports have also described protective effects against such damage that are mediated independently of Nrf2[80]. Moreover, studies showing protection against Aβ42-induced oxidative damage in response to GSK-3 inhibition, using antisense oligonucleotides[73] and in response to lithium[81], have not demonstrated a causal role of increasing Nrf2 activity. Our direct comparison of Keap1 and GSK-3 pathways, however, reveals that Nrf2 activity and prevention of Aβ42-induced oxidative damage do not strongly correlate.
Finally, combining lithium treatment with manipulation of Keap1 did not confer additional protection against oxidative and xenobiotic damage respectively. Overall this suggests that Keap1 and lithium treatment combine to maximise protection against AD-phenotypes, through divergent effects on Nrf2 and by additively protecting against Aβ42-induced oxidative and xenobiotic damage (Fig 7D).
Our findings provide compelling support for the use of direct Keap1-Nrf2 inhibitors for the treatment of neurodegenerative diseases, particularly AD. Future work is warranted to develop these compounds further for in vivo use, and to investigate their effects in combination with other established therapeutic targets for AD, such as specific GSK-3 inhibitors.
Stocks were maintained at 25°C on a 12:12-h light:dark cycle at constant humidity on a standard sugar-yeast (SY) medium (15gl-1 agar, 50 gl-1 sugar, 100 gl-1 autolysed yeast, 100gl-1 nipagin and 3ml l-1 propionic acid). Adult-onset neuronal-specific expression of Arctic mutant Aβ42 peptide was achieved by using the elav GeneSwitch (elavGS)-UAS system (GAL4-dependant upstream activator sequence) and treatment with 200 μM mifepristone (RU486), as previously described[34]. ElavGS was derived from the original elavGS 301.2 line and obtained as a generous gift from Dr H. Tricoire (CNRS, France). Aβ lines used in Fig 1 were: UAS-attP Aβ lines, as previously published [36], and UAS-WT Aβ42x2, obtained from Prof. P. Fernandez-Funez (University of Florida, USA)[37]. Random insertion UAS-ArcAβ42 was obtained from Dr D. Crowther (University of Cambridge, UK) and was used in all other experiments. UAS-0N3R tau line was obtained from Guy Tear (Kings College London, UK). UAS-C9orf72 (GR)100 flies are published[40]. Keap1del was generated by P-element mediated male recombination using the P-element insertion line, Keap1[EY02632][42], Keap1EY5 and gstD1-GFP lines were obtained from Dr D. Bohmann (University of Rochester, USA), and cncCK6 mutant flies, originally generated in the laboratory of William McGinnis (University of California, San Diego, USA), from Dr A. Whitworth (University of Cambridge, UK). All fly lines were backcrossed six times into the w1118 genetic background.
LiCl (Sigma) was dissolved in ddH20 at a concentration of 5 M before diluting to the indicated final concentrations in SYA medium.
Bortezomib (New England Biolabs) was dissolved in ethanol at a stock concentration of 10 mM, and stored frozen at– 20°C, before diluting to the indicated final concentrations in SYA medium.
Flies were raised at a standard density on SY medium in 200 mL bottles. Two days after eclosion once-mated females were transferred to experimental vials containing SY medium with or without RU486 at a density of 10 flies per vial. Deaths were scored and flies were transferred to fresh food 3 times per week. Data are presented as cumulative survival curves, and survival rates were compared using log-rank tests.
Climbing assays were performed using methods adapted from Sofola O et al., 2010 [34]. Briefly, 15 adult flies were placed in a vertical glass column (SciLabware), tapped to the bottom, and their ability to climb subsequently analysed. Flies reaching the top (above 10 cm) and flies remaining at the bottom (below 3 cm) of the column after a 30 sec period were counted. Scores recorded, from three trials per biological repeat, were the mean number of flies at the top (ntop), the mean number of flies at the bottom (nbottom) and the total number of flies assessed (ntot). A performance index (PI) defined as ½ (ntot + ntop—nbottom)/ ntot) was calculated. Data are presented as the mean PI ± SEM obtained in three independent repeats for each group.
cncC activity was measured by crossing UAS-ArcAβ42;elavGS flies to gstD1-GFP reporter flies, expressing green fluorescent protein (GFP) under the control of a 2.7kb genomic sequence upstream of the cncC target gene gstD1 [35], and analyzing GFP levels by western blotting.
Fly heads were homogenized in 2× laemmli sample buffer containing 100mM DTT. Proteins were separated by SDS-PAGE at 150V for 1h using 10% Bis-Tris gels and Mes-SDS running buffer (Invitrogen). Gels were then transferred to nitrocellulose membrane, incubated in a blocking solution containing 5% milk proteins in TBST for 1h at room temperature, then probed with GFP (Cell Signaling, 2955S, 1:1000), ATG8 (custom made anti-rabbit polyclonal against peptide EP113385 (Eurogentec) [42], 1:1000,) or actin (Santa Cruz, sc-47778, 1:5,000) primary antibodies overnight at 4°C. Anti-horseradish peroxidase (HRP)-conjugated secondary antibodies (Abcam, 1:12,000) were used and blots were developed using the enhanced chemiluminescence method (ECL) according to the manufacturers’ instructions (Luminata Crescendo; Millipore). Proteins were visualized using a luminescent image analyzer (ImageQuant LAS 4000; GE Healthcare) and relative intensities measured using ImageQuant software. Proteins of interest were expressed as a ratio relative to actin levels in each sample.
RNA extraction, cDNA synthesis and quantitative PCR (qPCR) reactions were performed as previously published[34,49]. For gene expression in Drosophila, primers, 5’ to 3’, were as follow: Aβ forward GATCCTTCTCCTGCTAACC and reverse CACCATCAAGCCAATAATCG; cncC forward GAGGTGGAAATCGGAGATGA and reverse CTGCTTGTAGAGCACCTCAGC; MafS forward AGATCGTTCGGATGAAGCAG and reverse GTCTCCAGCTCGTCCTTCTG; gstD2 forward CATCGCCGTCTATCTGGTGGA and reverse GGCATTGTCGTACCACCTGG; and eIF-1A reference gene forward ATCAGCTCCGAGGATGACGC and reverse GCCGAGACAGACGTTCCAGA[82]. Genes of interest were expressed as a ratio relative to eIF-1A.
For gene expression in primary mouse neuronal cultures, primers 5’ to 3’ were: Hmox1 forward AGCACAGGGTGACAGAAGAG and reverse GGAGCGGTGTCTGGGATG, Srnx1 forward GACGTCCTCTGGATCAAAG and reverse GCAGGAATGGTCTCTCTCTG, and xCT forward ATACTCCAGAACACGGGCAG and reverse AGTTCCACCCAGACTCGAAC, as previously published [60]. Reference gene primers were to mouse actin forward AACCGTGAAAAGATGACCCAGA and reverse CACAGCCTGGATGGCTACGTA. Genes of interest were expressed as a ratio relative to actin.
Total Aβ42 peptide, from fly heads, was extracted into guanidinium HCl (GnHCl) buffer based as previously described[34,83]. Alternatively, soluble and insoluble Aβ pools were isolated by differential centrifugation followed by formic acid extraction, as previously described [49]. Aβ42 levels were then measured using the High Sensitivity Human Amyloid Aβ42 ELISA kit (Millipore). Samples were diluted 1:100, for total Aβ42, or 1:10, for insoluble Aβ42, in dilution buffer and ELISA performed according to the manufacturers’ instructions. Protein extracts were quantified using the Bradford protein assay (Bio-Rad laboratories Ltd, UK) and the amount of Aβ42 in each sample expressed as a ratio of the total protein content (pmoles/g total protein).
For cell culture, conditioned media was removed from cells, following compound treatment, and diluted 1:2 in dilution buffer before measurement of Aβ42 levels by ELISA as described above.
Fly heads were homogenized in 25 mM Tris, pH 7.5 and protein content determined by Bradford assay. Chymotrypsin-like peptidase activity of the proteasome was assayed using the fluorogenic peptide substrate Succinyl-Leu-Leu-Val-Tyr-amidomethylcoumarin (LLVY-AMC), as previously described[49]. Proteasome activity was determined as the slope of AMC accumulation over time per mg of total protein (pmoles/min/mg).
Flies were prepared as for lifespan experiments then aged to the indicated time-points before measuring stress resistance. For xenobiotic stress, flies were exposed to DDT vapour, 1 mg/mL diluted in acetone, in glass vials in the absence of food. For oxidative stress, flies were subjected to hyperoxia (95% oxygen; PROOX model 110, Biospherix, USA) or injected with paraquat (75 nLs of 1 mg/ml in Ringers buffer) and maintained on SYA media containing RU486 ± LiCl as indicated. As survival times were short in these experiments flies were not transferred to fresh food. Survival under each stress condition was then monitored by recording the number of deaths every 2 hours from the start of decline. Data are presented as cumulative survival curves, and survival rates were compared using log-rank tests.
The raw microarray data generated in this study are deposited in ArrayExpress (http://www.ebi.ac.uk/arrayexpress) with identifier E-MTAB-4611. UAS-ArcAβ42/+;elavGS/+ flies were treated, for 17 days, with standard SY medium alone (-RU) or medium containing RU486 in the absence (+RU) or presence of 100 mM Lithium Chloride (+RU, +LiCl). Flies used for microarray analyses were snap-frozen in liquid nitrogen and, for each array, RNA extracted from 200 heads using RLT buffer + 0.01% β-mercaptoethanol and purified with RNeasy columns (Qiagen, Valencia, CA, USA) following the manufacturer's instructions. The quality and concentration of RNA was confirmed using an Agilent Bioanalyzer 2100 (Agilent Technologies, Santa Clara, CA, USA), and further procedures followed the standard Affymetrix protocol. All samples were hybridized to the Drosophila Genome 2.0 Genechip. In total, 4–5 biological replicates of each condition (-RU, +RU and +RU, +LiCl) were performed.
Raw data (cel files) were processed to correct for probe-sequence biases, by using bioconductor's package gcrma (http://www.bioconductor.org) in R (http://www.r-project.org). Affymetrix's MicroArray Suite 5.0 (bioconductor's package affy) was used to determine present target transcripts[84]. Transcripts were deemed present if the p-value was <0.111 and absent otherwise. The raw data were summarized and normalized by using Robust Multichip Average (rma function, part of bioconductor's package affy [85]. In order to identify differentially expressed genes a linear model was fitted and differential expression was assessed using the empirical Bayes moderated t-statistic as implemented in R's limma package [86]. P-values were adjusted for multiple hypothesis testing by applying the Benjamini and Hochberg correction for false discovery rate. Summarized probe-sets were mapped to transcripts using R's package "drosophila2.db". Transcripts not mapping to any known or predicted genes were excluded from further analysis. The following freely available gene expression microarray datasets were used: control vs. hsp70-CncC (E-GEOD-30087).
The Wilcoxon rank sum test, as implemented in Catmap[87], was used to perform functional analysis, that is significant enrichment of Gene Ontology categories. FlyBase (http://flybase.org) gene identifiers were mapped to Gene Ontology identifiers (FlyBase version FB2014_01). Ranks of genes were based on the p-value derived from the Bayes t-statistic for differential expression. To account for multiple hypothesis testing, an enrichment of GO terms was deemed statistically significant if the p-value derived from the wilcoxon rank sum test was ≤1.0x10-05. Gene lists were sorted by log-fold change and p-value. For all microarray experiments two sets of lists were derived; a gene list comprising most differentially up-regulated (log-fold change > 0) genes at the top of the list and most differentially down-regulated genes (log-fold change < 0) at the bottom of the list (termed up-to-down) and vice versa (termed down-to-up). If a GO category was found to be statistically significant in the up-to-down list, this GO was referred to as up-regulated, meaning that a large enough proportion of the genes in this GO category were found to be up-regulated or at the top of the list. If a GO category was found to be statistically significant in the down-to-up list, this GO was referred to as down-regulated, meaning that a large enough proportion of the genes in this GO category were found to be down-regulated or at the top of the list.
Statistical significance of overlaps of GOs in two microarray experiments was determined using fisher's exact test. To account for multiple hypothesis testing a p-value cut-off of ≤1.0x10-05 was used.
SH-SY5Y cells were incubated in a humidified atmosphere at 37°C, 5% CO2 in Dulbecco’s modified medium (DMEM) supplemented with 4.5g/L glucose, 10% FBS, 1% penicillin- streptomycin, and differentiated with 10 μM retinoic acid. Cells were seeded at an appropriate density in 96-well plates prior to subsequent analyses.
SH-SY5Y cells were seeded at a density of 2 x 104 cells per well in a 96-well plate and cultured for 2 days before treatment for 24 hours with compound or vehicle (0.1% DMSO final concentration). Cells were then lysed in 50 μl/well lysis buffer (0.1% Tween20 in 2 mM EDTA [pH 7.5]) for 15 mins at room temperature. 200 μl enzyme reaction mixture (25 mM Tris buffer [pH 7.5] containing BSA [0.067%], Tween20 (0.01%), FAD (5 μM), Glucose 6 Phosphate (G6P) (1 mM), NADP (30 μM), G6P dehydrogenase (40 units), MTT (0.03%), and menadione (50 μM)) was added to each well. After 5 min at room temperature, 40 μl/well stop solution (10% SDS) was added and the absorbance at 595 nm measured. The background optical density was measured using wells containing lysis buffer, enzyme and stop solutions without SH-SY5Y cells. The optical density values at 595 nm were averaged and the background corrected ratio of optical densities (compound treated/control) was calculated.
SH-SY5Y cells were seeded at a density of 2 x 104 cells per well in a 96-well plate. Cells were pre-treated with compound or vehicle overnight before either Aβ-conditioned Chinese Hamster Ovary (CHO) cell media (7PA2CM) or WT-conditioned CHO media (wtCM) was added to wells at a dilution of 50% for 24 hours. Resazurin (final concentration 20 μM) was added to wells and incubated for 4 hours at 37°C, 5% CO2. The resulting fluorescence intensity was measured at 590 nm. The relative fluorescence values were averaged and normalised to wtCM, DMSO-treated control intensities.
SH-SY5Y cells were seeded at a density of 6,000 cells per well, in a 96-well plate then, after 24 hours, treated with 50% 7PA2CM or wtCM for a further 24 hours. Plates were fixed with 4% PFA (v/v) for 15 minutes, washed three times with PBS, permeabilized with 0.5% Triton-X before staining with primary antibodies against Keap1 (1:50; ab150654, Abcam) and Nrf2 (1:100; ab62352, Abcam) in blocking solution containing 10% goat serum and 3% BSA overnight at 4 C. After 3 washes in PBS secondary antibodies (α rabbit Alexa 594, ab150080, α mouse FITC, ab6785, Abcam) were applied. Nuclei were stained in a PBS solution of 10 μM Hoechst 33342 before high-resolution digital imaging using a Zeiss LSM700 confocal microscope.
For quantification of Nrf2 and Keap1 staining, and cell morphology, 250 cells per well were imaged using the IN Cell Analyzer 6000 automated laser-based imaging platform with confocal modality (GE Healthcare) at 40X magnification. Automated image analysis was conducted with the IN Cell Developer software using custom-developed analysis protocols. Cell nuclei were identified by acquiring images in the DAPI channel (405 nm excitation, 455/50 emission). Whole cells were identified using images of Keap1 from the FITC channel (488 nm excitation, 524/48 emission). Images were also acquired of Nrf2 in the dsRed channel (561 nm excitation, 605/52 nm emission). The whole-cell intensity of Keap1 FITC and Nrf2 Alexa 594 was measured and expressed as average fluorescence units per cell for each well. N = 5 wells per condition.
To obtain transgenic conditioned medium (TgCM) enriched in Aβ, neurons from Tg2576 mice were maintained for 14 days in vitro (DIV) without changing the medium[88]. Medium from wild type neurons from littermates, wild type conditioned medium (wtCM), was used as a control. The genotype of the animals was determined by polymerase chain reaction on DNA obtained from fibroblasts.
Wild type primary neurons were obtained from cerebral cortex of CD1 mouse embryos, at embryonic day 15, as previously described [89]. Neurons were plated to a density 6×105 viable cells/35-mm2 on glass-bottomed dishes (MatTek Corporation, Ashland, MA, USA) previously coated with poly-D-lysine (10 μg/ml) for at least 1 h at 37°C. Cultures were maintained at 37°C with 5% CO2, supplemented with neurobasal medium with 2% B27 nutrient, 2 mM L-glutamine, penicillin (100 units/ml) and streptomycin (100 μg/ml). At 12 days in vitro (DIV) neurons, transfected on day 7 with the plasmid peGFP-N1 (Clontech, Mountain View, CA) using lipofectamine 2000 (Invitrogen), were pre-treated with the Keap1 inhibitor, 22h (10 μM[22]), or 0.1% DMSO, for 16 h. Conditioned medium (CM) from Tg2576 transgenic, Aβ-enriched, or wild type mouse neurons, with and without 22h, was then added for a further 24h before analysis.
Neuronal morphology was assessed at 14 DIV by high-resolution digital imaging using a Zeiss LSM700 confocal microscope and analysis using NeuronStudio software (CNIC, Mount Sinai School of Medicine). Spine density was defined as the number of spines per micrometer of dendrite length according to previously published protocols[59]. Dendritic spine densities were calculated from 6–17 neurons per condition.
Data are presented as means ± SEM obtained in at least three independent biological samples. Log-rank, analysis of variances (ANOVA) and Tukey’s HSD (honestly significant difference) post-hoc analyses were performed using JMP (version 11.0) software (SAS Institute, Cary, NC, USA).
Animals were maintained and treated in accordance with the Animals (Scientific Procedures) Act, 1986, following approval by the Animal Welfare and Ethical Review Body (AWERB), KCL, and the Home Office Inspectorate, and performed in accordance with the European Communities Council Directive of 24 November 1986 (86/609/EEC).
|
10.1371/journal.pntd.0001330 | The Long Term Effect of Current and New Interventions on the New Case Detection of Leprosy: A Modeling Study | Although the number of newly detected leprosy cases has decreased globally, a quarter of a million new cases are detected annually and eradication remains far away. Current options for leprosy prevention are contact tracing and BCG vaccination of infants. Future options may include chemoprophylaxis and early diagnosis of subclinical infections. This study compared the predicted trends in leprosy case detection of future intervention strategies.
Seven leprosy intervention scenarios were investigated with a microsimulation model (SIMCOLEP) to predict future leprosy trends. The baseline scenario consisted of passive case detection, multidrug therapy, contact tracing, and BCG vaccination of infants. The other six scenarios were modifications of the baseline, as follows: no contact tracing; with chemoprophylaxis; with early diagnosis of subclinical infections; replacement of the BCG vaccine with a new tuberculosis vaccine ineffective against Mycobacterium leprae (“no BCG”); no BCG with chemoprophylaxis; and no BCG with early diagnosis.
Without contact tracing, the model predicted an initial drop in the new case detection rate due to a delay in detecting clinical cases among contacts. Eventually, this scenario would lead to new case detection rates higher than the baseline program. Both chemoprophylaxis and early diagnosis would prevent new cases due to a reduction of the infectious period of subclinical cases by detection and cure of these cases. Also, replacing BCG would increase the new case detection rate of leprosy, but this effect could be offset with either chemoprophylaxis or early diagnosis.
This study showed that the leprosy incidence would be reduced substantially by good BCG vaccine coverage and the combined strategies of contact tracing, early diagnosis, and treatment of infection and/or chemoprophylaxis among household contacts. To effectively interrupt the transmission of M. leprae, it is crucial to continue developing immuno- and chemoprophylaxis strategies and an effective test for diagnosing subclinical infections.
| Leprosy is a contagious disease that will remain prevalent, despite the declining number of patients worldwide over the last century. With approximately 250,000 new cases detected annually, leprosy is far from being eradicated. Leprosy can be treated with drugs after disease detection.
Some cases can be prevented with a tuberculosis vaccine (BCG) that cross-reacts with the bacterium responsible for leprosy, but this vaccine might be replaced in the future. Furthermore, preventive drugs can reduce the number of new cases among people in contact with infectious patients, but this strategy has not yet become established in common practice. Also, a new test is under development for the detection of infections before the appearance of symptoms.
In this study, we used a computer model to assess the effectiveness of seven possible leprosy control activities. Our results showed that the decline in incidence of leprosy would slow down or halt with the introduction of a new tuberculosis vaccine that is ineffective against leprosy. However, this effect could be offset by the implementation of effective tests for early diagnosis or the routine administration of preventative drugs to contacts of patients.
| The global new case detection rate of leprosy has dropped considerably during last century, but with approximately 250,000 new cases detected annually, leprosy is far from being eradicated [1]. Currently, the primary strategy for controlling leprosy is case detection and treatment with multidrug therapy (MDT). Although new interventions are under development, their potential impact on disease control is unknown. Recent clinical trials have indicated that a single chemoprophylactic dose of rifampicin given to individuals in contact with newly diagnosed leprosy patients could protect these contacts against leprosy disease [2]. The results with a single dose of rifampicin are very comparable to trials with dapsone that were conducted in the pre-MDT era. A meta-analysis showed that the combined results from the randomized controlled trials favored chemoprophylaxis to placebo with 2–4 years of follow-up (relative risk 0.59, 95% (CI) 0.50–0.70) [3]. The advantage of a single dose of rifampicin is that it is only given once, while dapsone prophylaxis is given for at least 2 years. Furthermore, new tests are under development for identifying subclinical infections [4].
Other recent developments however, are cause for concern. For example, the integration of leprosy control activities into general health care programs has in many countries led to the cessation of active case finding and contact tracing. Consequently, diagnosis is delayed and patients are therefore infectious for a longer period causing more people in contact with patients to become infected.
Another concern is that a new vaccine may replace the current Bacillus Calmette-Guérin (BCG) tuberculosis vaccine, which is given to infants to prevent tuberculosis (TB), but which also protects against leprosy [5]. An update on progress describing new TB vaccine candidates that are currently entering clinical trials has recently been published [6]. Most are pre-exposure vaccines and will most likely prevent TB disease. Such vaccines are intended either to replace BCG (recombinant live vaccines) or to be given after BCG prime as boosters (protein adjuvant formulations or recombinant viral carriers). New and more specific TB vaccines may not induce cross-immunity to the bacterium responsible for leprosy, Mycobacterium leprae [6], [7]. It was recently pleaded that new candidate vaccines must be developed taking both diseases into account, and that the current TB candidate vaccines should be assessed for their potential to protect against leprosy as well as against TB [8]. Therefore, the effect of new leprosy interventions strategies should be tested in the context of other related developments, such as possible changes to BCG.
Although the short-term effectiveness of new interventions can be assessed in trials, extrapolation to long-term effectiveness in the general population is difficult, due to the complex impact on transmission dynamics. Hence, dynamic simulation models are necessary to assess the possible impact of different intervention strategies on future trends in the new case detection rate of leprosy.
We have developed a microsimulation model that simulates the transmission and control of leprosy (the SIMCOLEP model), taking into account the population structure of households [9]. The model has been quantified by data from northwest Bangladesh in 2003. Very detailed data were available for that year from a large randomised controlled trial of chemoprophylaxis with single dose rifampicin (the COLEP study) that was being conducted at the time [2], [10]–[13]. This is an area with a well-organized leprosy control program and with a decreasing trend in new case detection since the mid-1990′s. Regardless, the current case detection rate remains one of the highest in Bangladesh, 2–3 per 10,000 population. We applied our model to the situation in this area as a starting point for exploring the potential impact of seven different intervention strategies on the detection of new cases of leprosy over a 50-year period.
For the COLEP trial (ISRCTN 61223447), on which the data of this modeling study is largely based, ethical clearance was obtained from the Ethical Review Committee of the Bangladesh Medical Research Council in Dhaka (ref. no. BMRC/ERC/2001–2004/799). All subjects were informed verbally in their own language and invited to participate. Written consent was requested from each adult. For children consent from a parent or guardian was given.
The microsimulation model simulates the life history of fictitious individuals [9]. These individuals are members of a household that is formed, changes, and dissolves during the simulation. Individual household movement occurs during adolescence and after marriage. Some married couples start living in the household of the parents-in-law, and will form their own separate household after on average 12 years. The life span of individuals is drawn from a life-table at birth; the number of newborn individuals maintains the simulated population growth rate equal to the observed population growth rate; newly born individuals are placed into the household of their mothers; and mothers are drawn from the population of married women and weighted with an age-dependent fertility function.
An individual that is susceptible to leprosy is defined as an individual that developed leprosy sometime during their lifetime, after acquiring the infection. The large majority (say 80–95%) of the population is assumed not to be susceptible to leprosy [14]–[16].The remaining 5–20% of the population is susceptible. For these individuals, it is assumed that 80% undergoes a self-healing infection and is never infectious to other individuals, that is 20% will become chronically infected and infectious [16].
The mechanisms underlying leprosy susceptibility are currently unknown [9]. Therefore, the model used six hypothetical mechanisms: Random (no mechanism, but each individual has a fixed probability of being susceptible); Household susceptibility (all susceptibles live in a fraction of households, within these susceptible households a fraction of inhabitants is susceptible); Dominant (susceptibility is inherited by a dominant gene); Recessive (susceptibility is inherited by a recessive gene); Household & dominant (50% of susceptibility is determined by the Household and 50% by a dominant gene); Household & recessive inheritance (50% of susceptibility is determined by the Household and 50% by a recessive).
As described in a previous paper [9], the model was unable to identify one single mechanism that could best explain the observed data. However, for Random it turned out that 20% susceptibles provided the best fit, whereas this was 10% for the other mechanisms. For Household this 10% was established by assuming on average 25% of the households contain on average 40% susceptible individuals.
The quantification of the model is based on the leprosy situation in 2003 and the control program of the last decades in the Nilphamari and Rangpur districts of Bangladesh [9]. This control program consisted of passive case detection, with in 2003 an average detection delay of 2 years, treatment with MDT, and active tracing of people in contact with patients. Contacts are examined annually for three consecutive years. In this area, BCG vaccination was routinely given to newborn infants. Since the introduction of the BCG vaccination in 1974, the coverage had gradually expanded to 80% in 1990 and remained at that level in 2003 [17]. BCG had a protective effect of 60% [18].
For a full and detailed description of the model, we refer to our previous paper [9].
In the study we considered seven potential intervention scenarios for the future control of leprosy. The baseline scenario was the current leprosy control program in the Bangladesh study area, as described above. The other scenarios were modifications of the baseline control program. These other six scenarios were: 1) no contact tracing; 2) with a single chemoprophylactic dose of rifampicin, which cured 50% of subclinical cases, for each individual in contact with a leprosy patient [2]; 3) with diagnosis of subclinical cases with a sensitivity of 70% [19] followed by effective treatment; 4) with all newly born infants in the population receiving a new (hypothetical) tuberculosis vaccine that is ineffective against leprosy instead of BCG (no BCG); 5) with the combination of no BCG and chemoprophylaxis; and 6) with the combination of no BCG and early diagnosis with effective treatment.
In our intervention scenarios, contact tracing, chemoprophylactic treatment and early diagnosis were performed only on household members. Contact tracing was repeated three times in three consecutive years with a 10% probability of loss to follow-up and a 90% of symptomatic cases being detected. Early diagnosis was performed in the same schedule as the contact tracing with three consecutive visits to the household. Chemoprophylactic treatment was given only once after examination, in which 90% of symptomatic cases will be detected.
The simulation of interventions was started based on the quantification of 2003, because a detailed data set [16] was available from the COLEP study conducted during that period. The Bangladesh districts at the time when the COLEP study took place can be seen as fairly representative for other areas in the Indian subcontinent with regard to demography, socio-economic condition, cultural tradition and the organization of the health system, including the leprosy control program. The prevalence rate of leprosy at the time was well above the WHO elimination target of 1 per 10,000 population, which was also the case in many areas in India around the year 2000.
Table 1 shows the predicted new case detection rates at 25 years after the initiation of the interventions. Under the baseline control program, the different mechanisms that determined susceptibility showed up to three-fold differences in the predicted number of cases per 100,000 people. In Figure 1, the trends in the new case detection rates over 50 years are shown for all seven interventions. All susceptibility mechanisms give qualitatively comparable trends. When the intervention scenarios were ordered after 50 years by the amount of reduction in new case detection rates, the order was as good as identical for all mechanisms; i.e. early diagnosis lowest; then no BCG & early diagnosis; then chemoprophylaxis; then baseline; then no BCG & chemoprophylaxis together with no contact tracing; and finally no BCG had the highest new case detection rate.
Both the cessation of contact tracing and the replacement of BCG vaccine by a tuberculosis vaccine ineffective for leprosy (no BCG) would have detrimental effects on the rate of decline in leprosy (Figure 2). Twenty-five years after introduction of the ineffective vaccine (no BCG), the new case detection of leprosy was approximately 1.5 times higher than the baseline (Table 1). The cessation of contact tracing was predicted to have a smaller impact, with a marked drop in detection of new leprosy cases during the first few years. This sudden drop was due to the reduced number of examinations of people in contact with patients; thus, these cases would not be detected until later, through passive detection (self-reporting).
Both chemoprophylaxis and early diagnosis were predicted to have substantial effects on the new case detection of leprosy (Figure 2). With no BCG, chemoprophylaxis would partially compensate for the predicted increase in new case detection rates. Furthermore, early diagnosis was predicted to more than compensate for the adverse effects of a leprosy-ineffective tuberculosis vaccine, and reduce the rate of new case detection compared to the baseline. The effects were more promising with the ongoing presence of the BCG vaccine. Under those conditions, at 25 years after the introduction of chemoprophylaxis, the new case detection rate was predicted to be 25% lower than baseline control. Moreover, with the introduction of early diagnosis, the new case detection rate was predicted to halve the baseline incidence after 25 years (Table 1).
Early diagnosis of infection allows the detection of subclinical cases, of which part would be detected later or never at all. These subclinical cases are added to the number of detected cases. This is seen in the results of this intervention. The introduction of early diagnosis would increase the total number of detected cases in the first 18 years, simply because of the detection of previously undetectable subclinical cases. Over time however, the total number of new cases (subclinical and clinical) would finally drop below the number detected in the baseline control program (Figure 3). In Figure 3, we show that the new cases detected under the chemoprophylaxis intervention strategy drop immediately below the level of the baseline control program. The additional effect of chemoprophylaxis is that additional new infections are prevented on top of the cure of subclinical infections. These additional prevented infections are due to a shorter infectious period of the cured subclinical infections. To illustrate this effect we show in Figure 3 the clinical and the subclinical cases that were cured by the chemoprophylactic intervention. During the first 10 years, this total number of newly detected cases plus cured cases is equal to the number of newly detected cases under the baseline control program, but afterwards the number of cases plus cured subclinical cases in the chemoprophylaxis intervention group drops under the baseline control program, indicating the prevention of new infections.
This study used a microsimulation model to compare the future outcomes of different leprosy intervention programs. The baseline program consists of passive case detection, treatment with MDT, contact tracing, and infant BCG vaccination. The predicted rate of decline in new case detection depends on the intervention scenario chosen over the next 50 years. Early diagnosis and/or chemoprophylaxis added to the baseline program can result in a considerable reduction in the new case detection rate. Furthermore, these interventions were predicted to compensate for the adverse effect of replacing BCG by a leprosy-ineffective tuberculosis vaccine.
Our microsimulation modeling approach was able to capture individual (stochastic) processes. Complex infection dynamics could thus be simulated on an individual basis. Aggregating the model outcomes enabled the analysis of trends at the population level. The quantification of the model was based on an area, Nilphamari and Rangpur districts in Bangladesh, where leprosy is highly endemic and which has a well-organized control program [10]. The chemoprophylaxis intervention parameters were based on the COLEP trial conducted in this population [10]. Because of the COLEP study, a large amount of very detailed information was available for this population directly prior to the trial that started in 2003 [11]. We did not use data after 2003, because the impact of the chemoprophylactic intervention (a randomised controlled trial) is difficult if not impossible to mimic in our model.
The use of our microsimulation model is limited to interventions on a household basis. A future challenge would be to extend the modeling to interactions between households. This requires data from molecular epidemiology studies, for which techniques only recently have been developed and field tested [20], [21]. Our objective in this study was however to compare interventions that are most feasible, namely targeting household members. The main uncertainty in our modeling is the mechanism that causes susceptibility to leprosy. The transmission parameters as well as future trends differ greatly between these mechanisms [9]. Other parameters will have less influence on the outcomes than these large differences due to these mechanisms. Our conclusions remain the same for all mechanisms.
The absence of data on the households in the years prior to the COLEP trial makes a quantitative validation of the results impossible. Therefore, the conclusions of our study are only qualitative, i.e. ranking of the intervention strategies. However, these can be used to prioritize implementation and fundamental research into chemoprophylactic treatment and early diagnosis. We are confident that our results will apply to Bangladesh and many other regions in the world.
The current leprosy control program in the Nilphamari and Rangpur districts of Bangladesh is more extensive than usual. The primary advantage of this program is the active tracing of individuals that had been in contact with newly diagnosed leprosy patients. Contact tracing is not common in leprosy control programs. Our modelling showed that contact tracing and subsequent treatment of newly found patients could, in itself, contribute to a reduction in the transmission of M. leprae in the population. Nevertheless, we argue that the qualitative results, i.e. the ranking of the intervention strategies, will not differ when implemented in a currrently less intensive control program.
The primary concern of this study was to estimate the relative, not the absolute, impact of the various interventions and take into account alternative hypotheses for mechanisms of susceptibility to leprosy. We compared the results based on different hypothesized mechanisms for susceptibility, because each of these mechanisms could be valid [9]. The quantitative results were sensitive to the mechanism chosen. Nevertheless, when the different interventions were ordered by the magnitude of effect, that order was identical for all the mechanisms of susceptibility. Thus, the qualitative results were robust, and suggested this order of effectiveness for the different interventions can be generalized.
The new interventions chemoprophylaxis and early diagnosis (which necessarily include contact tracing) were predicted to have a clear added impact for leprosy control. We assumed that the effect of chemoprophylaxis with a single dose rifampicin (SDR) could prevent 50% of subclinical infections to develop leprosy. This assumption was based on the outcome of the COLEP trial and represented the overall effect of SDR in the contacts [2]. In the trial, this effect of SDR was a 56% reduction in new leprosy cases after two years for all contacts. The effect of SDR, however, varied among the different types of contacts, with a 49% prevention in neighbors, 54% prevention in household contacts, and 76% prevention in social contacts [2]. Thus, the choice of contacts to be included in contact tracing and subsequent chemoprophylactic treatment is very important. Ideally, it should go beyond the immediate household of the index patient. The choice of the contact ‘ring’ will likely depend on the acceptance of contacts to be involved and the feasibility of running an extended program. Moreover, rather than providing chemoprophylaxis to all, one would prefer to first test for a subclinical infection and then treat individuals appropriately.
Our modeling showed that identification and treatment of subclinical infections among household contacts had the largest effect in reducing transmission of M. leprae in the population. Part of the better performance of early diagnosis compared to chemoprophylaxis was that the early diagnosis strategy comprised three consecutive annual tests with 70% sensitivity, compared to a single round of rifampicin with a cure rate of 50%. Thus, more subclinical cases could be cured after the early diagnosis than with chemoprophylaxis. Meima et al. [4] showed that a short detection delay is key to the success of the current MDT-based leprosy control strategy. Detection of subclinical cases would be a major improvement because it provides an even shorter detection delay. As shown in Figure 3, the detection of subclinical cases also reduced transmission, and the total number of new cases detected (clinical and subclinical) was predicted to eventually drop below the number of new cases detected under the baseline control program.
Our study shows that BCG may have an important effect on the reduction of the case detection of leprosy. Previously, Meima et al. [4] showed, just as in this study, that BCG vaccination may have a large impact on the expected incidence of leprosy in the population. The current knowledge about the effect of the BCG vaccination on leprosy strongly supports maintaining the current BCG vaccination practice [5], [22]. Alternatively, a leprosy-specific compound should be added to an improved tuberculosis vaccine in leprosy endemic areas.
We showed that the leprosy incidence would be reduced substantially by good BCG vaccine coverage and the combined strategies of contact tracing, early diagnosis, and treatment of infection and/or chemoprophylaxis among household contacts. To effectively interrupt the transmission of M. leprae, it is crucial to continue developing immuno- and chemoprophylaxis strategies and an effective test for diagnosing subclinical infections.
|
10.1371/journal.pbio.2006810 | Sex-specific dominance reversal of genetic variation for fitness | The maintenance of genetic variance in fitness represents one of the most longstanding enigmas in evolutionary biology. Sexually antagonistic (SA) selection may contribute substantially to maintaining genetic variance in fitness by maintaining alternative alleles with opposite fitness effects in the two sexes. This is especially likely if such SA loci exhibit sex-specific dominance reversal (SSDR)—wherein the allele that benefits a given sex is also dominant in that sex—which would generate balancing selection and maintain stable SA polymorphisms for fitness. However, direct empirical tests of SSDR for fitness are currently lacking. Here, we performed a full diallel cross among isogenic strains derived from a natural population of the seed beetle Callosobruchus maculatus that is known to exhibit SA genetic variance in fitness. We measured sex-specific competitive lifetime reproductive success (i.e., fitness) in >500 sex-by-genotype F1 combinations and found that segregating genetic variation in fitness exhibited pronounced contributions from dominance variance and sex-specific dominance variance. A closer inspection of the nature of dominance variance revealed that the fixed allelic variation captured within each strain tended to be dominant in one sex but recessive in the other, revealing genome-wide SSDR for SA polymorphisms underlying fitness. Our findings suggest that SA balancing selection could play an underappreciated role in maintaining fitness variance in natural populations.
| Evolution requires genetic variation, but selection will tend to fix whichever alleles confer the highest fitness, depleting the genetic variation upon which it acts. Sexually antagonistic (SA) genetic variation—in which alternative alleles have opposite fitness effects in the sexes—can generate balancing selection that maintains genetic variation for fitness if the alleles that benefit a given sex are also dominant in that sex. Here, we show that the SA genetic variation underlying fitness in a well-known seed beetle population exhibits these beneficial reversals of dominance, suggesting SA selection may commonly maintain heritable genetic variation for fitness throughout the genome.
| One of the most longstanding challenges for evolutionary biologists has been to explain the maintenance of genetic variance in fitness [1–7]. Selection should erode genetic variation as it eliminates deleterious alleles and fixes beneficial ones. Yet natural populations harbor abundant heritable variation for fitness and life history traits [8–9]. The two general hypotheses for explaining this are mutation-selection balance and balancing selection [2–4]. Under the former, many polymorphisms throughout the genome are maintained at low allele frequencies because of a constant influx of deleterious mutations [10–14], yet this process cannot single-handedly explain the extent and pattern of the genetic variance observed in nature [4–7] (explained below). Thus, some form of balancing selection—including scenarios in which alternative alleles offer fitness benefits in different contexts (e.g., environments, genotypes, seasons, or sexes)—must contribute to the maintenance of polymorphisms for fitness throughout the genome [4–7].
Sexually antagonistic (SA) selection can cause alternative alleles to have opposite fitness effects in males and females [15–18] and has the potential to be the most widespread source of balancing selection among eukaryotes. Sex is a nearly ubiquitous feature of eukaryotic life [19], SA genetic variation is an inevitable outcome of two sexes sharing the same genome whilst having different fitness optima [20–22], and antagonistic forms of balancing selection should generate more stable (less-transient) polymorphisms for fitness than nonantagonistic forms of balancing selection [23]. Further, well-adapted populations should exhibit an overabundance of SA genetic variance in fitness relative to sexually concordant (SC) genetic variance (i.e., that which affects the sexes similarly; see Fig 1) because (1) purifying selection should remove SC genetic variation relatively efficiently [24], and (2) the rate at which SA polymorphisms can be resolved should be low relative to the rate at which novel SA mutations occur [15,17–18,25–29]. A growing body of evidence for standing SA genetic variation in natural and laboratory populations largely supports these predictions (e.g., [26,30–40]).
The capacity for SA selection to generate balancing selection that results in stable polymorphisms for fitness drastically increases with sex-specific dominance reversal (SSDR) [41–42], in which the alleles that benefit a given sex’s fitness are also dominant in that sex, generating a net heterozygote advantage in the population. Such beneficial reversals of dominance, in the more general case of antagonistic pleiotropy [43], were met with early skepticism [44–46], but more recent theory is changing that view. The average locus underlying two polygenic homologous phenotypes with antagonistic fitness effects (e.g., male and female fitness) is actually expected to exhibit at least partial dominance reversal for fitness under the nonrestrictive assumption that the fitness functions are concave in the vicinity of their overlap [47]. Further, SA genetic variation should in turn select for modifier loci that enable heterozygous males and females to exhibit favorable dominance relationships between SA alleles [48], making SSDR of SA genetic variation for fitness a predicted outcome of adaptive evolution.
The ability of SSDR to promote the maintenance of SA polymorphisms for fitness has been known for decades [41], yet there are currently no direct empirical tests of SSDR for fitness. Barson and colleagues [39] recently demonstrated SSDR for a single major-effect locus affecting age at maturity in salmon. Fitness, however, is a highly polygenic trait, and investigating the properties of many loci typically requires a quantitative genetic approach [49–50].
The full diallel cross [50] is the premier quantitative genetic breeding design, capable of partitioning phenotypic variance into that attributable to additive genetic effects, parental effects, dominance, and epistasis. This enables tests of several of the hypotheses for the maintenance of genetic variance in fitness. Because the many weakly deleterious alleles maintained by mutation-selection balance should necessarily exhibit partial dominance [4,7,10–14,51–52], it should generate pronounced additive genetic variance relative to dominance variance if it is the predominant contributor to fitness variance. By contrast, balancing selection acting to maintain relatively fewer polymorphisms of larger effect size, which may involve dominance coefficients up to and including complete dominance, would generate pronounced dominance variance relative to additive genetic variance if it predominates [4,7–8,50]. As for distinguishing among the many forms of balancing selection, diallel data offer the possibility to distinguish SA balancing selection from other forms of balancing selection by testing for its hallmark SSDR (see above). To this end, variance partitioning is not a sufficient test. Rather, full diallel data offer the possibility to quantify the relative amount of fixed recessive allelic variation among a set of inbred strains, and if the data are sex-specific, then it enables one to do this for males and females independently. A positive correlation among strains for their relative amounts of recessive allelic variation when measured in males versus females would demonstrate that strains tend to exhibit a relatively high or low number of fixed recessive alleles (regardless of which sex they are expressed in), whereas a negative correlation would demonstrate that strains tend to be fixed for allelic variation that is recessive in one sex but dominant in the other (i.e., SSDR). It is conceivable that SSDRs of this nature could evolve in the context of SC allelic variants that merely exhibit sex differences in the relative strength of selection acting on each of them. Thus, the most conservative and explicit test of SSDR for the SA genetic variation underlying fitness would be to analyze this correlation after statistically removing SC additive genetic effects from the data.
Here, we used a full diallel cross among 16 isogenic strains to partition genetic variance in sex-specific competitive lifetime reproductive success (hereafter “fitness”) in a wild-caught population of the seed beetle C. maculatus that is known to exhibit pronounced SA genetic variance in fitness [37–38,53]. This species exhibits a polyandrous mating system, X/Y sex determination, and pronounced sexual dimorphism and sex-biased gene expression [33,38,54–56]. In total, 3,278 individual fitness assays (1,731 male and 1,547 female) were conducted over the 256 possible cross types (hereafter “families”)—i.e., the 240 outcrossed (heterozygous) families and 16 parental-self (homozygous) families of a full 16 × 16 diallel. Considering that these inbred strains originate from a population whose genetic variance in fitness is predominantly SA [37], the present findings of pronounced dominance variance and sex-specific dominance variance relative to additive genetic variance (Fig 2) suggest that this population’s fitness variance is largely underlain by relatively few, large-effect polymorphisms under SA selection, as opposed to many small-effect polymorphisms in mutation-selection balance [4,7–8,50]. Our analyses revealed that strains exhibited significantly negative rank correlation for their relative amount of fixed dominant alleles for fitness when measured in males versus females (Fig 3). Thus, whether the average allele underlying fitness in this population is dominant or recessive in a heterozygote depends on whether it is being expressed in a male or a female. As mentioned above, this could still be the case for some SC allelic variation as well, but this relationship remained strong when the SC additive genetic effects were statistically removed from the data beforehand (S1 Fig), explicitly demonstrating SSDR for the SA genetic variation underlying fitness. Our findings are consistent with genome-wide SSDR maintaining balanced SA polymorphisms for fitness, which has important implications for the capacity of SA selection to explain fitness variance in natural populations.
We performed a full diallel cross among 16 isogenic strains, generating an F1 generation of 240 possible outcrossed families and 16 parental selfs. We assayed F1 male and female competitive lifetime reproductive success (i.e., fitness) separately to obtain sex-specific measures of fitness for each outcrossed family and selfed strain (S2 Fig). Phenotypic variance in fitness was partitioned into that attributable to the overall fixed effects of replicate block (x), inbreeding (b1), sex (S), and sex-specific inbreeding (S×b1), as well as the following strain- and cross-specific random effects (i.e., variance components) using restricted maximum likelihood (REML) estimation: additive genetic variance (a), dominance (b2), parental effects (c), symmetric epistasis (b3), asymmetric epistasis (d), and sex-specific versions thereof (i.e., their interaction with S). For this REML approach, we used the FDIALLEL procedure [57] for GenStat (v.18.2.0.18409; [58]) to fit a custom version of Hayman’s [59] model that we modified to accommodate sex-specific effects (hereafter the “full sexed” model). We also applied Lenarcic and colleagues’ [60] full sexed hierarchical Bayesian model fit to our data using the “BayesDiallel” MCMC Gibbs sampler [61] for R (v.3.2.1; [62]). The two approaches yielded qualitatively similar results (the Bayesian approach is reported in the S3–S5 Figs, S1 and S2 Tables, and S1 Text).
To aid interpretation, separate male and female models were also performed (on separate male and female data sets), which provided the sex-specific variance components (qM and qF of Fig 1) that were used in the following geometric interpretation of “sexed” and “unsexed” variance components from the full sexed model (see below) for all inheritance classes (see S3 Table for separate sex-specific variance component estimates). For a given inheritance class Q (e.g., additivity, dominance, epistasis, etc.), the full sexed model provides sexed (S×q) and unsexed (q) variance components. Strains’ best linear unbiased predictions (BLUPs) for a given variance component represent their estimated values along those axes of variation. All unsexed axes of variation (q) from the full sexed model were highly correlated to (i.e., ≈) the derived SC dimension of their respective inheritance class (qSC; all Pearson’s correlation coefficients rq,qSC were > 0.98, all P values < 0.0001; see Materials and methods and Fig 1). Thus, all unsexed variance components from the full sexed model were found to represent SC effects. This was true even in cases of technically improper correlations in which the standard error of q, qM, or qF overlapped zero (e.g., ra,aSC = 0.99, P < 0.0001, despite a being approximately 0; see below). Further, the sexed variance components for additivity (S×a), dominance (S×b2), and epistasis (S×b3) were found to represent SA effects: the inner product between these unsexed (i.e., SC, see above) and sexed effects revealed them as describing approximately orthogonal axes of variation (rendering the latter SA; see Fig 1). That is, their BLUPs fell along axes of variation that were at approximately 90° angular displacements, θ, from one another (θa,S×a = 91.50°, θb2,S×b2 = 90.00°, θb3,S×b3 = 91.49°). Negative variance component estimates for parental effects (c) and sex-specific asymmetric epistasis (S×d) rendered their BLUPs invalid and excluded their inheritance classes from this geometric interpretation. Again, all unsexed variance components describe SC effects, whereas the sex-specific additive (S×a), dominance (S×b2), and epistatic (S×b3) variance components (being orthogonal to their SC counterparts) describe SA effects (see Fig 1). We therefore refer to them accordingly (e.g., SA additive genetic variance). Further information on the terminology and meaning of variance components is available in the S1 Text.
There were no overall differences between the sexes (S) in mean fitness (Table 1). The overall effect of inbreeding (b1) was large and statistically significant (Table 1). The sexes differed in this regard, as revealed by a significant sex-specific inbreeding effect (S×b1; Table 1). Separate male and female models revealed that the effect of inbreeding was stronger in males than females (S3 Table), a result confirmed by other studies (e.g., [63]). Note that these sizeable inbreeding and sex-specific inbreeding effects have been accounted for as overall (fixed) effects prior to the estimation of random effects variance components and therefore do not inflate or disrupt estimates of dominance (b2), sex-specific dominance (S×b2), or any other variance component.
In general, inheritance for fitness was characterized by pronounced SC dominance (b2) and SA dominance (S×b2; Table 1 and Fig 2). There was also a relatively small but nonzero contribution from SC epistatic variance (b3; Table 1 and Fig 2). Of those, b2 and S×b2, respectively, contributed approximately 17 times and approximately 8 times the genetic variance to fitness relative to b3 (Table 1).
Although the standard error of SA additive genetic variance (S×a) did barely overlap zero, we note that its estimate was 10 times that of SC additive genetic variance (a) (Table 1 and Fig 2). Parental effects (c), sex-specific parental effects (S×c), SA epistasis (S×b3), asymmetric epistasis (d), and sex-specific asymmetric epistasis (S×d) were all estimated as near zero with standard errors overlapping zero (Table 1 and Fig 2).
Separate male and female models revealed about 5 times the residual variance (ε) for males than for females, complicating precise quantitative comparison between these models (S3 Table). Qualitatively, however, nonzero variance was estimated for dominance (b2) and epistasis (b3) in both male and female models, and these were the only inheritance classes with nonzero variance in either model (S3 Table). In both models, b2 was estimated at about 13 times that of b3 (S3 Table).
A full diallel cross among isogenic strains offers additional valuable insight regarding the dominance relationships between the fixed allelic variants among strains [50,59]. In particular, an estimate of the relative proportion of dominant:recessive alleles among the homozygous parental strains is given by the array covariances between strains’ outcrossed (heterozygous) family means and the means of the inbred (homozygous) parental selfs of the strains to which they were crossed (S6 Fig [50,59]). Strains whose outcrossed (heterozygous) family means are determined by (i.e., covary with) the inbred (homozygous) means of the strains to which they are crossed harbor alleles that are apparently recessive to those of other strains. By contrast, strains whose outcrossed (heterozygous) family means are instead independent of (i.e., do not covary with) the inbred (homozygous) means of the strains to which they are crossed harbor alleles that are apparently dominant to those of other strains. After removing environmental and epistatic variance from the data, σP,r, each strain’s covariance between its outcrossed family means (r) and the means of the inbred parental selfs (P) of the strains to which they were crossed [50,59] was used as an estimate of the relative amount of recessive alleles fixed within each strain (see Materials and methods and S6 Fig). This was done for male (σPM,rM) and female (σPF,rF) fitness separately, and these estimates were significantly negatively correlated (Pearson’s rσPM,rM,σPF,rF = −0.779 (95% CI −0.92 to −0.46), P = 0.0004; Fig 3A). This effect was not driven by any particular strain(s), as evidenced by a significantly negative nonparametric rank correlation, indicating strains’ relative “dominance” over one another (Spearman’s rσPM,rM,σPF,rF: −0.738, P = 0.0016; Fig 3B). These tests remained strongly significant after first having statistically removed SC additive genetic variance (Pearson’s rσPM,rM′,σPF,rF′ = −0.665 (95% CI −0.87 to −0.25), P = 0.005; Spearman’s rσPM,rM′,σPF,rF′: −0.635, P = 0.0098; S1 Fig), explicitly demonstrating SSDR of the SA genetic variation for fitness.
Progress in our general understanding of the maintenance of genetic variance in fitness requires the identification of broadly applicable mechanisms. Knowing that mutation-selection balance cannot explain everything [4–7], the question for many evolutionary biologists becomes: which form(s) of balancing selection account for the remaining genetic variance? There are many mechanisms of balancing selection with theoretical and empirical support that are capable of maintaining stable polymorphisms for fitness under certain conditions. However, SA selection is potentially one of the most widespread among eukaryotes—pending the prevalence of SSDR—offering a general solution to this classic evolutionary question. Unfortunately, SA selection often goes unmentioned in reviews on the topic (e.g., [5–6]), leaving it relatively poorly understood from a modern genomics perspective [64–65]. We will therefore briefly review the inevitability of SA genetic variation, for which the present findings and other recent studies substantially strengthen the case for its role in maintaining genetic variance in fitness.
Antagonistic mutations (showing any form of antagonistic pleiotropy) sweep to fixation more slowly than nonantagonistic mutations, meaning that they will generate a weaker genomic signature of selection but will actually have a greater, more sustained contribution to genetic variance in fitness [23]. In accordance, experimental evolution in microorganisms often reveals that adaptation in a given context or environment comes at the price of reduced fitness in other contexts (reviewed in [66]), indicating substantial standing genetic variation with antagonistic pleiotropic effects. Likewise, Qian and colleagues [67] demonstrated widespread antagonistic pleiotropy in hundreds of genes across the yeast genome. The prevalence of these genetic trade-offs is expected to increase with increasing organismal complexity such as specialized tissues, developmental stages, and sexes [67–69]. Further, the resolution of such genetic trade-offs evolves more slowly under smaller effective population sizes and with longer generation times [67], again implying an even greater likelihood for antagonistic pleiotropy to maintain genetic variance in fitness in complex multicellular organisms (i.e., eukaryotes). Thus, considering that sexual reproduction is a nearly ubiquitous feature of eukaryotic life [19], the sexes likely represent the most consistent and widespread set of contexts over which antagonistic pleiotropy could ensue.
Although the sexes share largely the same genome, they accrue fitness in distinct ways, meaning that their fitness optima for a range of life history traits commonly differ [16–18,22]. Thus, just as Fisher’s [1] geometric model predicts the large majority of mutations to be deleterious, a rare mutation with fitness benefits in one sex will tend to pose fitness detriments in the other—the shared genome thus making SA genetic variation inevitable [20–22]. Such antagonistic mutations will then tend to reach intermediate equilibrium allele frequencies [23], while purifying selection will tend to eliminate mutations that generate SC genetic variance, leaving behind mostly SA genetic variance in fitness [24] (see Fig 1). In addition, the constraint that a shared genome poses to sex-specific adaptive evolution implies that there is only a limited extent to which SA polymorphisms can be resolved relative to the rate at which novel SA mutations occur [17–18,25–29], consistent with a growing body of empirical evidence for standing SA genetic variation (e.g., [26,30–40]). Lastly, even sex differences in the strength of selection for alternative allelic variants underlying SC forms of antagonistic pleiotropy between different components of fitness (e.g., survival, fecundity, fertility, mate competition, etc.) can render allelic trade-offs to have SA effects on overall fitness [70], suggesting that (1) the likelihood of SA genetic variation for fitness is even greater than previous theory would suggest and (2) studies that lack evidence of SA genetic variance for a given component of fitness do not speak to the presence or absence of SA genetic variation for fitness.
Though SA genetic variation for fitness may be theoretically inevitable and empirically common (see above), the extent to which it may explain fitness variance would be substantially broadened if SA polymorphisms commonly exhibited SSDR [41–42]. This would cause a net heterozygote advantage and generate stable balancing selection on those polymorphisms [41–42]. Dominance reversal between allelic variants at loci exhibiting antagonistic pleiotropy [43] was met with early skepticism (e.g., [44–46]) but is actually expected for the average SA polymorphism [47]. Further, theory predicts that SA polymorphisms favor the evolution of mechanisms that enable SSDR [48]. In support of this notion, one of the most convincing cases of a specific SA locus, albeit not for fitness per se, does indeed exhibit SSDR [39].
Our results are consistent with polygenic SSDR for the SA allelic variation underlying fitness. Using a diallel cross among isogenic strains from a well-characterized population of C. maculatus known to exhibit predominantly SA genetic variance in fitness [37–38,53,71–72], we show that the dominant–recessive relationship between alternative alleles at the loci underlying fitness was reversed in heterozygous males versus females (Fig 3). Specifically, strains whose outcrossed male fitness values tended to covary with the inbred male fitness values of the strains they were crossed with tended not to exhibit this covariance with regard to female fitness and vice versa (see Materials and methods and S6 Fig). In other words, strains whose outcrossed male fitness was determined by the strains they were crossed with (indicating those other strains’ allelic variation was dominant to their own with regard to male fitness) tended to be the determinant of female outcrossed fitness (indicating their own allelic variation was dominant to other strains’ with regard to female fitness). Thus, whether the fixed allelic variation for fitness in a given genotype tended to be dominant or recessive to that of the other genotypes (in the heterozygous progeny of crosses among them) depended upon the sex in which it was being expressed. Such SSDR will facilitate the stable maintenance of SA polymorphisms for fitness via balancing selection, enhancing their contribution to a population’s genetic variance. Indeed, this is likely a major contributor to the fact that this population exhibits predominantly SA genetic variance in fitness [37]. We repeated our test of SSDR after accounting for SC effects in the data to provide a more explicit test of SSDR for the remaining SA allelic effects, per se. The result remained highly significant but was slightly weaker (S1 Fig), which may suggest some SSDR for allelic variation with SC fitness effects (see Introduction).
The prediction that dominance deviations (the BLUPs for dominance variance) should be negatively correlated between the traits/sexes under antagonistic pleiotropy [4] was not upheld by our data (not reported). Other approaches using the dominance deviations (e.g., correlating additive breeding values and dominance deviations) were likewise ineffective at revealing the apparently strong signature of SSDR in our data (not reported). Lastly, even our extensive geometric interpretation of variance components—revealing that sex-specific dominance variance stemmed from SA dominance deviations—provided an inadequate picture of the true extent of SSDR exhibited by this population. Thus, variance partitioning and correlating dominance deviations are likely not sufficient to document dominance reversal. Future studies aiming to investigate dominance reversal via quantitative genetic methods should aim to perform full diallel crosses and assess the relative amount of dominant alleles within each strain via Hayman’s method [50,59] (see Materials and methods) so as to correlate those measures between traits, niches, and sexes that are hypothesized to be under antagonistic selection.
In addition to dominance reversal, theory has shown that epistatic interactions between SA loci can promote the maintenance of SA polymorphisms for fitness [73]. Epistasis is an expectation for polygenic traits that is derived from Fisher’s geometric model [1,73–75]. That is, because of the diminishing returns of additional beneficial alleles in increasingly beneficial backgrounds, the effect of a given allele underlying a continuous trait depends on which alleles are present at the remaining or background loci underlying that trait [73–75]. Despite it being difficult to detect empirically [74] and sometimes analyzed inappropriately [76], there is evidence of diminishing returns epistasis (e.g., [77]). However, the role of epistasis in contributing to the maintenance of SA polymorphisms for fitness has received little if any empirical attention (e.g., [78]). We found that the variance component explaining the next most phenotypic variance in fitness after dominance and SA dominance was SC epistasis. Its sex-specific counterpart was identified as describing SA epistatic effects (though its standard error overlapped zero). This represented one of the few discrepancies between our REML and Bayesian analyses, the latter identifying a sizeable contribution to fitness variance from both SC and SA epistatic effects (S3 Fig and S1 Table). However, an interpretation of the SA epistatic deviations is less straightforward than that for SA dominance deviations (see above). First, (sex-specific) epistatic deviations are a property of crosses, not strains, but the geometric interpretation of this inheritance class was based on strain means for (sex-specific) epistatic deviations (see Materials and methods). Thus, it is the crosses of some strains that tended to exhibit epistatic deviations (i.e., deviations from the expectation based on the additive contribution from each parent strain) in one sex but not the other or to different degrees in the sexes. Second, (SA) epistatic variance can be generated by a variety of interactions among loci and may not merely be due to the diminishing returns of additive loci. Thus, whether the SA epistatic effects detected are explicitly the type that could maintain genetic variance in fitness remains unclear. At most, therefore, our findings can only provide mixed evidence of a putative role for SA diminishing returns epistasis among the loci underlying fitness in this population.
As with dominance reversal, parental effects such as sex-linkage, cytoplasmic, or epigenetic effects could also partially resolve SA polymorphisms and contribute to the maintenance of SA genetic variance in fitness [25,79–83]. There is some empirical support for this (e.g., [32]), but we detected little or no variance in fitness attributable to any form of parental effects or asymmetric epistasis (i.e., parental-effects epistasis), despite the unrivaled explicit exposure of parental effects variance via the reciprocal crosses of a full diallel [50]. Although this is not to say that parental effects are nonexistent in this population, it does suggest that such effects have a relatively minor role in generating fitness variance.
Our study provides novel evidence for polygenic SSDR for the SA genetic variation underlying fitness, adding significant insights to our understanding of SA genetic variation and the maintenance of genetic variance in fitness. We hope that our findings will stimulate further efforts along these lines, which we suspect will add to the growing consensus that SA selection is a widespread phenomenon among sexually reproducing species that commonly acts to maintain genetic variance in fitness.
C. maculatus (Coleoptera: Bruchidae) is a pest of leguminous crops that has colonized most of the tropical and subtropical regions of the world [84]. Thus, laboratory conditions (see below) closely resemble the grain storage facilities and crop fields they have inhabited since the early Holocene. Females lay eggs on the surface of dry beans and hatched larvae bore into the beans, where they complete their life cycle, emerging from the beans as reproductively mature adults [84]. This species is facultatively aphagous (requiring neither food nor water to reproduce successfully) and exhibits a polyandrous mating system [54], X/Y sex determination [55], and pronounced sexual dimorphism [33,38] and sex-biased gene expression [56].
The origin of our study population has been described by Berger and colleagues [37] and Grieshop and colleagues [53]. Briefly, the population was isolated from Vigna unguiculata seed pods collected at a small-scale agricultural field close to Lomé, Togo (06°10′N 01°13′E) during October and November 2010. Seed pods were stripped in the laboratory and beans isolated individually. Virgin males and females hatching out of these beans were paired randomly, and each pair founded an isofemale line (n = 41), each of which was thus derived from a single maternal and a single paternal genome. These isofemale lines were expanded and cultured at population sizes of 200–400 adults on 150 ml of V. unguiculata seeds at 29°C, 55% RH and a 12L:12D light regime for about 12 generations prior to the sex-specific fitness assays conducted by Berger and colleagues [37]. The development of isogenic (inbred) strains from these isofemale lines is described by Grieshop and colleagues [53] and in the S1 Text. The 16 inbred strains used in the present study were reasonably evenly distributed about the population’s original intersexual genetic correlation for fitness (S7 Fig) and apparently captured an unbiased representation of the standing SA genetic variation for fitness exhibited by the wild population from which it was derived (see Results and S1 Text).
We performed a full diallel cross [50] among 16 inbred strains, generating an F1 generation of 240 possible outcrossed combinations and 16 parental selfs. The phenotype we measured in F1 individuals was sex-specific competitive lifetime reproductive success (i.e., fitness). F1 male (N = 1,731) and female (N = 1,547) fitness was assayed separately by placing a single focal individual in a container with approximately 25 g of V. unguiculata seeds, a sterile same-sex reference competitor (from an outbred base population established at the same time and from the same population as the isofemale lines; see above), and two opposite-sex reference beetles (a 1:1 sex ratio; see S2 Fig). Assays were placed in incubators at 29°C, 55% RH and a 12L:12D light regime until all F2 offspring emerged from the beans. The number of F2 offspring produced by an assay represents the focal (nonsterile) individual’s fitness. The sperm of sterilized males still function and fertilize eggs (but the zygotes are inviable), such that male fitness assays included pre- and postcopulatory selection (see [72] for details). Female assays also included mating competition, as well as competition for oviposition substrate (beans), and females’ ability to endure harmful repeated mating attempts by competing males in order to survive and oviposit [38,85–86]. These fitness assays not only include many aspects of the natural ecological setting for these beetles but also represent complex physical environments (i.e., the geometry of the beans), which may play an important role in enabling laboratory fitness assays to reflect complex natural environments and enable mating interactions, sexual selection, and sexual conflict to ensue more naturally [87–88].
The diallel experiment was performed twice, in two “blocks” (with cells replicated within and between blocks). In total, we performed 3,278 fitness assays (1,731 male and 1,547 female), with only moderate imbalance over the 256 families and 2 replicate blocks. Imbalance over sex, cross, and/or block categories was, however, unavoidable: different crosses (including inbred selfs) produced different numbers of F1 offspring in different sex ratios and had different probabilities of producing zero F1 offspring, providing variable opportunity to assay F1 fitness throughout the diallel. However, because of the large sampling effort, only 3 out of 240 outbred crosses (and 0 parental selfs) were missing from the total data set.
We fit a custom version of Hayman’s [59] model (modified to accommodate sex-specific data) using the FDIALLEL procedure [57] in GenStat (v.18.2.0.18409; [58]):
y=μ+x+b1+a+b2+c+b3+d+S+S×b1+S×a+S×b2+S×c+S×b3+S×d+ε,
where μ is the intercept for the total phenotypic variance in fitness y, which is partitioned into that attributable to residual error variance ε, the overall fixed effects of replicate block x, inbreeding b1, sex S, and sex-specific inbreeding S×b1, as well as the following strain- and cross-specific random effects (i.e., variance components): additive genetic variance a, dominance b2, parental effects c, symmetric epistasis b3, asymmetric epistasis d, and the interaction of each of those random effects with S. FDIALLEL forms the factors and matrices necessary to fit a diallel model using GenStat’s REML directive [57,89]—the REML directive accepting other fixed (e.g., S) and/or random effects (e.g., S×b2). The model was performed on log-transformed data, as this provided a superior model fit. This customization of Hayman’s [59] approach did not alter any of the underlying modeling of the variance components as defined in GenStat’s [58] FDIALLEL procedure [57]. We report F statistics and P values for the overall (fixed) effects and variances (σ2) with standard errors for the variance components (random effects). A variance component with a positive variance and standard error that excludes zero is interpreted as evidence for that mode of inheritance contributing to the observed phenotypic variance in fitness. It is possible, by this approach, to attain negative variance component estimates, which should, of course, be interpreted as not differing from zero and should not disrupt the estimation of other variance components in the model.
Replicate block (x) was included as an additional fixed effect because it has only two levels—i.e., too few levels to be modeled as a random effect [90]. Block had a significant effect, likely stemming from imbalance described above—the two replicate blocks differed in overall sample size and in the patterns and degree of sampling imbalance among crosses.
To aid interpretation, we also performed separate male and female models:
yM=μ+x+b1M+aM+b2M+cM+b3M+dM+ε,
and
yF=μ+x+b1F+aF+b2F+cF+b3F+dF+ε,
respectively.
The separate male and female models enabled a geometric validation of the meaning of variance components by relating the predicted values for each strain (i.e., their BLUPs) between sexed and unsexed models as follows. The BLUPs for a given inheritance class Q from separate male and female models (qM and qF) can be set as variance-standardized y and x axes, respectively (Fig 1). That coordinate system can be rotated 45° to derive BLUPs for SC (qSC) and SA (qSA) additive genetic variance (Fig 1) as done by Berger and colleagues [37] and Grieshop and colleagues [53], like so:
qSCBLUPs=qFBLUPssin(45°)+qMBLUPscos(45°),
and
qSABLUPs=qFBLUPscos(45°)−qMBLUPssin(45°).
Note that no variance is lost during this rotation.
For many inheritance classes, the BLUPs for qSC and qSA (derived from separate male and female models as described above; see Fig 1) were highly correlated with the BLUPs of q and S×q (estimated by the full sexed model), respectively. For example, in the case of additivity, a was correlated to aSC (ra,aSC = 0.98, P < 0.0001), and S×a was correlated to aSA (rS×a,aSA = 0.99, P < 0.0001). Thus, staying with the specific example of additivity,
a≈aSC,
S×a≈aSA,
and
a⊥S×a,
or in words, the additive genetic variance a from a full sexed model represents SC additive genetic variance, the sex-specific additive genetic variance S×a represents SA additive genetic variance, and they are orthogonal to one another (see Fig 1).
Having verified that all unsexed variance components q from the full sexed model were modeling SC effects (see Results), a more straightforward measure of the angular relationship between two axes of variation q and S×q for a given inheritance class Q (in order to assess whether the latter represents SA variance) is their inner product q ∙ S×q, defined as:
q∙S×q=q1∙S×q1+q2∙S×q2+…+qn∙S×qn,
where q1…qn and S×q1…S×qn (nonitalicized) are the BLUPs of each strain for variance components q and S×q, respectively. Note that the inner products between sexed and unsexed variance components for symmetric and asymmetric epistasis were calculated based on strain means of BLUPs, since those variance components are based on 120 and 240 unique strain–strain combinations (in a 16 × 16 diallel), respectively.
The inner product between two axes of variation equals zero when they are orthogonal to one another (or in this case, when two variance components are describing orthogonal axes of variation). For normalized axes of variation (ours being variance-standardized prior to coordinate system rotation; see above), the inner product between q and S×q can be converted to the more intuitive angular displacement, θ, like so:
θq,S×q=arccos(q∙S×q).
We can thus calculate and verify the angular displacement of all sexed variance components from their respective unsexed/SC counterparts via their BLUPs. Note that BLUPs from variance components with negative estimates (e.g., c and S×d; see Results) are not valid and can therefore not be used in this geometric interpretation. The major insight gained by this exercise is that all unsexed variance components are modeling SC effects and that some sexed variance components (see Results) are modeling SA effects in this population—symmetric orthogonal deviations from the mean of their respective unsexed/SC counterparts (see Fig 1).
Environmental and epistatic variance was removed from the data by taking the residuals from the following model fit, again using the FDIALLEL procedure [57] in GenStat [58] (see above):
y=μ+x+b3+ε,
where μ is the intercept, x is the fixed effect of replicate block, b3 is a random effect modeling symmetric epistasis, and ε is the residual error variance. Thus, all other effects (but namely, sex-specific additive and dominance effects) remain as underlying contributions to variance in the residual data. These residuals were not variance-standardized.
Family means were tabulated from these residuals and, σP,r, each strains’ covariance between its outcrossed family means (r) and the means of the respective inbred parental selfs (P) that correspond to each of those outcrossed families (i.e., Wr of Fig 1 in Hayman [59] and σP2,r of Fig 20.4 in Lynch and Walsh [50]) was used as an estimate of the relative amount of recessive alleles within each strain. Each strain’s sire- and dam-specific covariances are averaged, and this was done separately for male (σPM,rM) and female (σPF,rF) fitness. If we denote the family mean z¯ for a given sex of a given dam–sire combination as z¯dam,sire, then σP,r for strain 1 of that sex would be the covariance of the elements in these two vectors:
rdam:(z¯1,2,z¯1,3,z¯1,4,…z¯1,16)
P:(z¯2,2,z¯3,3,z¯4,4,…z¯16,16),
averaged with the covariance of the elements in these two vectors:
rsire:(z¯2,1,z¯3,1,z¯4,1,…z¯16,1)
P:(z¯2,2,z¯3,3,z¯4,4,…z¯16,16).
Again, this was done for each strain and for male and female fitness separately, for a total of 32 independent covariances (i.e., 16 σPM,rM, and 16 σPF,rF; see S6 Fig).
Strains whose outcrossed (heterozygous) means (r) are determined by (i.e., covary with) the inbred (homozygous) means (P) of the strains to which they are crossed have alleles that are apparently recessive to those of other strains, whereas strains whose heterozygous means (r) are independent of (i.e., do not covary with) the homozygous means (P) of the strains to which they are crossed have alleles that are apparently dominant to those of other strains. Thus, the correlation between σPM,rM and σPF,rF provides an indication of whether the allelic variation among strains tends to exhibit the same dominant–recessive relationship in both sexes (given by a positive correlation) or whether the dominant–recessive relationship of the allelic variation among strains is reversed between the sexes (given by a negative correlation).
Note that Hayman [59] and Lynch and Walsh [50] point out that σP,r and the variance among outcrossed family means r (i.e. Vr of Fig 1 in Hayman [59] and σr2 of Fig 20.4 in Lynch and Walsh [50]) should scale perfectly with a regression coefficient of 1, where the intercept of that slope indicates the degree of dominance exhibited by the underlying loci (e.g., partial dominance, complete dominance, or overdominance). This is, of course, in the absence of epistatic variance, environmental variance, and substantial remaining heterozygosity in the inbred strains. We removed epistatic and environmental variance (see above), and our inbred strains appear to harbor little remaining heterozygosity (as indicated by the large inbreeding effect in our data; see Table 1 and S1 Table). However, we still found no relationship between σP,r and σr2 (see above) with regard to male or female fitness, perhaps indicating various degrees of dominance among the underlying loci and/or unexplained environmental or epistatic variance in the residuals.
This test was repeated after removing the SC additive genetic effects (a) from the data by taking the residuals from the following model:
y=μ+x+a+b3+ε,
and applying the same procedure described above, which provides a more explicit test of SSDR for the SA genetic variation, per se.
|
10.1371/journal.pgen.1002329 | Transcriptome-Wide Binding Sites for Components of the Saccharomyces cerevisiae Non-Poly(A) Termination Pathway: Nrd1, Nab3, and Sen1 | RNA polymerase II synthesizes a diverse set of transcripts including both protein-coding and non-coding RNAs. One major difference between these two classes of transcripts is the mechanism of termination. Messenger RNA transcripts terminate downstream of the coding region in a process that is coupled to cleavage and polyadenylation reactions. Non-coding transcripts like Saccharomyces cerevisiae snoRNAs terminate in a process that requires the RNA–binding proteins Nrd1, Nab3, and Sen1. We report here the transcriptome-wide distribution of these termination factors. These data sets derived from in vivo protein–RNA cross-linking provide high-resolution definition of non-poly(A) terminators, identify novel genes regulated by attenuation of nascent transcripts close to the promoter, and demonstrate the widespread occurrence of Nrd1-bound 3′ antisense transcripts on genes that are poorly expressed. In addition, we show that Sen1 does not cross-link efficiently to many expected non-coding RNAs but does cross-link to the 3′ end of most pre–mRNA transcripts, suggesting an extensive role in mRNA 3′ end formation and/or termination.
| Transcription in eukaryotes is widespread including both protein-coding transcripts and an increasing number of non-coding RNAs. Here we present the results of transcriptome-wide mapping of a set of yeast RNA–binding proteins that control expression of some protein-coding genes and a number of novel non-coding RNAs. The yeast Nrd1-Nab3-Sen1 pathway is required for termination and exosome-mediated processing of non-coding RNA polymerase II transcripts. Our data show that these components bind unexpected targets including a large number of antisense transcripts originating from the 3′ end of genes that are poorly expressed in the sense direction. We also show that Sen1 helicase, involved in termination of non-coding RNAs, is also present at the 3′ end of mRNAs, suggesting a more fundamental role in transcription termination. Mis-regulation of transcription is the underlying cause of many disease states. For example, mutation of the human Sen1 gene, senataxin, causes a range of neurodegenerative disorders. Understanding the roles of yeast RNA–binding proteins in controlling termination of coding and non-coding RNAs will be useful in deciphering the mechanism of these proteins in human cells.
| Early in each transcription cycle RNA polymerase II (Pol II) can follow one of two paths; terminate early through the Nrd1-Nab3-Sen1 pathway or continue on to form longer, potentially coding transcripts [1], [2]. Yeast Nrd1, Nab3 and Sen1 are part of a complex [3] that interacts both with the phosphorylated Pol II C-terminal domain (CTD) [4]–[6], and with specific sequences in the nascent transcript [7], [8]. If this set of protein-protein and protein-RNA contacts is sufficient, the Nrd1-Nab3-Sen1 complex directs Pol II termination and couples this to processing of the nascent transcript by TRAMP and the nuclear exosome [9], [10]. In addition to directing termination of snoRNAs and cryptic unstable transcripts (CUTs) the Nrd1-Nab3-Sen1 pathway directs premature termination of several pre-mRNA transcripts including NRD1, IMD2, URA2, URA8, and ADE12 [11]–[15]. The mechanism by which the Nrd1-Nab3-Sen1 complex leads to Pol II termination is unknown but involves recognition of specific terminator sequences by Nrd1 and Nab3 [7], [8], [16] and the putative helicase activity of Sen1 [16], [17].
In this study we have used a recently developed in vivo cross-linking approach [18] to derive high-resolution transcriptome-wide maps of binding sites for Nrd1, Nab3, Sen1, and the Pol II subunit Rpb2. This approach yields a more precise picture of known Nrd1 and Nab3 binding sites on snoRNA, CUT and mRNA targets and reveals a set of previously unknown Nrd1 binding sites both on the 5′ ends of mRNAs and on 3′ antisense transcripts. Surprisingly, Sen1 does not cross-link at many of these Nrd1-Nab3 binding sites. Instead, we observe Sen1 cross-linking on mRNA transcripts, particularly at the 3′ end suggesting a potential role for Sen1 in mRNA 3′ end formation through the cleavage and polyadenylation pathway.
The Photoactivatable-Ribonucleoside-Enhanced Crosslinking and Immunoprecipitation (PAR-CLIP) technique [18] was adapted to yeast (Materials and Methods). Briefly, cells are grown in the presence of 4-thiouracil which is incorporated into RNA. UV irradiation at 365 nm results in high-efficiency covalent cross-linking of RNA to protein [18]. We then purify the tagged protein under denaturing conditions [19], prepare cDNA libraries to the covalently attached RNAs and sequence the libraries. As previously reported, reverse transcription of the cross-linked 4-thiouracil leads to a characteristic T to C transition in the cDNA sequence [18]. Among all single-base changes observed in the sequences we derived from our initial Nrd1 and Nab3 cross-linking experiments approximately 90% were T to C transitions; about 10-fold greater than the number expected for random errors. Data analysis was limited to sequences containing a single base change from the reference sequence thus pinpointing the site of cross-linking in these sequences.
In all, we report sequence data for five libraries (Table S1) including data for Nrd1, Nab3, Sen1 and the RNA Pol II subunit Rpb2. Similar datasets have recently been reported by the Tollervey lab [20] who crosslinked Nrd1, Nab3 and Trf4 using an alternative UV cross-linking procedure (CRAC) [21]. The Nrd1 and Nab3 cross-linking datasets obtained by PAR-CLIP and CRAC are very similar. In both cases the majority of cross-links occur on Pol II transcripts with snoRNA transcripts the most abundant. Pol III transcripts are also cross-linked to Nrd1 and Nab3 and Wlotzka et al. provide evidence for a role for Nrd1 in processing pre-tRNAs [20]. A minor amount of Nrd1, Nab3 and Sen1 cross-links to ribosomal RNA and we have not pursued this further. Our cross-linking of Nrd1, Nab3, Sen1 and Rpb2 to snoRNA, snRNA, and tRNA will be discussed in another publication [22].
To validate the PAR-CLIP procedure in yeast we identified Nrd1 and Nab3 binding sites downstream of the SNR13 gene (Figure S1) with a peak of cross-linking within a region containing termination elements identified through genetic analysis [7], [8]. In addition, Figure S1 shows multiple Nrd1 and Nab3 binding sites in the upstream region of the Nrd1 gene, an observation consistent with our previous work showing that multiple mutations were required to inactivate autoregulation of Nrd1 expression [11]. While we cannot say with certainty whether the extent of cross-linking is exhaustive, we observe cross-linking at all previously observed Nrd1 and Nab3 binding sites including those on snoRNAs, snRNAs, attenuated mRNA transcripts and CUTs.
Previous studies have shown that Nrd1 and Nab3 are present on chromatin and localized to genes that are regulated by Nrd1 and Nab3 [4], [6], [9], [11], [23]. Because the Nrd1 complex interacts both with the nascent transcript and with Pol II, it is not clear which interaction is primarily responsible for chromatin binding. We have analyzed Nrd1 and Nab3 interactions with both chromatin and RNA. Figure 1 shows a comparison of ChIP-chip of Nrd1 compared to Nrd1 and Nab3 PAR-CLIP. Two tracks displaying Pol II subunits Rpb2 and Rpb3 are shown for reference, as they are both part of the catalytic core and thus are expected to co-localize. All five data sets were obtained from cells growing in log phase. Nrd1 and Pol II co-localize by ChIP at both the RPS5 and SNR3 genes (Figure 1). Nab3 ChIPs similarly to Nrd1 (not shown). In contrast, Nrd1 and Nab3 cross-link to RNA in a subset of the major Nrd1 ChIP peaks. For example, in Figure 1 we observe efficient cross-linking of Nrd1 and Nab3 to SNR3 transcripts but not to RPS5 transcripts. Genomewide, the Nab3 RNA cross-linking pattern is nearly identical to Nrd1 (Wilcoxon rank sum p<10−8). Similar specificity of Nrd1 RNA cross-linking is seen at other highly expressed genes. Among the top 100 Nrd1 ChIP peaks determined using CisGenome [24] are ten other ribosomal protein genes, none of which are among the top 100 Nrd1 PAR-CLIP peaks (not shown). We conclude that Nrd1, Nab3, and presumably Sen1 are able to enter the early Pol II elongation complex independent of RNA binding and monitor the nascent RNA for appropriate binding sites.
Figure 2A shows the logos [25] of motifs in a representative MEME [26] trial using sequences of 40 nt centered on the most prominent RNA cross-link (T to C) sites. Nrd1 cross-links reveal a consensus sequence UGUAG with an E-value of 1.8e−25 where the underlined U is the most frequently cross-linked residue. The top-scoring motif in the Nab3 pool is a consensus sequence GNUCUUGU. The E-value under the same conditions as the Nrd1 analysis is much higher (6.7e3) indicating that this motif is not nearly as constrained. These motifs are nearly identical to the GUA[A/G] and UCUU sequences that we previously identified using genetic and biochemical approaches [7], [8] and contain many of the over-represented 4 mer sequences in sequences that cross-link to Nrd1 and Nab3 by the CRAC protocol with 254 nm UV irradiation [20]. Our Nab3 motif is very similar to the UUCUUGUW motif identified by microarray analysis of RNA co-purified with Nrd1 in the absence of cross-linking and under non-denaturing conditions [27]. No conserved motifs were observed for Sen1 and Rpb2, consistent with their presumed roles as enzymes that interact non-specifically with RNA.
About 50% of the Nrd1 and Nab3 cross-linked regions overlap. This is not surprising given that Nrd1 and Nab3 are known to dimerize in vivo and in vitro [7], [28]. The second most significant motif, as determined in MEME, in the Nrd1 set corresponded to the top rated motif in the Nab3 analysis and vice versa, again suggesting that Nrd1 and Nab3 binding sites are located close to each other.
In Figure 2B we show that Nrd1 cross-linking sites are clustered. The 50 Nrd1 cross-linked sites with the most reads (as determined by MochiView [29]) were used as an anchor in this plot. The number of cross-links observed as a function of the distance from the top 50 cross-linked sites is increased in a window approximately 30 nt upstream and downstream of the central Nrd1 cross-link. No similar clustering of Nab3 cross-linking sites was observed indicating that most Nab3 binding regions do not contain multiple motifs. Together, these results suggest that these strong Nrd1 binding sites have been selected to bind multiple Nrd1 proteins, a result consistent with our in vitro binding experiments suggesting that the Nrd1-Nab3 complex binds cooperatively to RNA targets [7].
For both Nrd1 and Nab3 data sets binding sites on snoRNA transcripts were among the most extensively cross-linked (Tables S2 and S3). At most snoRNAs we observed a peak of Nrd1 binding downstream of the mature 3′ end of the RNA consistent with the previously demonstrated role in termination [16]. Figure 3A and 3B shows the distribution of cross-linked sequences downstream of SNR3 and SNR13. Nrd1 binds predominantly downstream of the mature RNA, while Rpb2 cross-links across the transcript. Surprisingly, Sen1 cross-links efficiently to some snoRNA transcripts like snR3, but not to others, like snR13. This is an unexpected result considering previous experiments showing that at both of these snoRNA genes Pol II reads through the terminator in a sen1 mutant at the non-permissive temperature [16] and both genes have Sen1 present on chromatin as determined by ChIP [30], [31].
We have previously shown that Nrd1 regulates its own expression by binding to sites in the 5′ end of its mRNA and directing premature termination. This attenuation mechanism requires the function of Nrd1, Nab3 and Sen1 [16]. The Nrd1-Nab3-Sen1 pathway is also required for the formation of attenuated transcripts from the IMD2 promoter [12]. We were therefore surprised that Sen1 does not crosslink as efficiently to NRD1 or IMD2 mRNA as it does to other mRNAs, for example CCW12 (Figure 3C and 3D and Figure S2).
Of the top 300 Nrd1 and Nab3 peaks, 93 and 54, respectively, overlap with CUTs (out of 925 annotated CUTS [32], [33]). By contrast, only 23 (Nrd1) and 8 (Nab3) peaks overlap with SUTs (stable unannotated transcripts, out of 847 annotated SUTs [32], [33]). This preference for binding unstable transcripts is consistent with Nrd1-Nab3 binding being a key feature that distinguishes between CUTs and SUTs [20]. In Figure S3 we show that Nrd1 binds to several known CUTs. The Nrd1 cross-linking sites on CUT transcripts tend to be located toward the 5′ end, a position in which Nrd1 and Nab3 may direct the observed range of termination sites downstream of these binding sites [9], [10], [34]. The observation that some CUTs are not bound by Nrd1 is somewhat surprising. In Figure S3 we observe that some CUTs and SUTs are not expressed under our experimental conditions as shown by the lack of cross-linking to Rpb2.
Nrd1, Nab3, and Sen1 also bind to RNA sequences derived from protein-coding genes. Cross-linked sequence reads were sorted into ten bins covering each coding region in order to display genes of different length on the same scale. Figure 4A–4C shows the distribution of Nrd1 plus-strand and minus-strand reads on genes ranked by expression level. Figure 4D–4F shows a similar distribution of Sen1 reads. Highly expressed genes have a peak of Nrd1 sense-strand reads derived from the 5′ end. Genes with the lowest level of expression show a peak of Nrd1 cross-linked antisense reads at the 3′ end. For Sen1 we observe the largest number of reads on the 3′ end of the most abundantly transcribed genes.
Nrd1 has been implicated in several regulatory mechanisms involving binding to sequences in the 5′ end of transcripts. Autoregulation of Nrd1 expression by attenuation is one example. Another form of regulation involving Nrd1 and Nab3 binding near the 5′ end of transcripts is alternative transcription start site (TSS) selection on genes involved in nucleotide biosynthesis. IMD2, URA2, URA8 and ADE12 have been shown to use alternative transcription start sites (TSSs) in response to nucleotide availability. For IMD2 and URA2 the upstream starts used in the presence of sufficient NTP result in the elongation complex passing through a Nrd1-Nab3 terminator thus reducing transcription of mRNA [12], [13], [15]. We observe peaks of Nrd1 binding on all of these transcripts (not shown). In addition, we see similar binding on other genes involved in nucleotide metabolism including HPT1, GUA1 and ADE17 (Figure S4). For each of these genes multiple TSSs have been reported and are located upstream and downstream of the Nrd1-Nab3 binding region, consistent with regulation by TSS selection.
Figure 5 shows two additional genes, PCF11 and RPB10 that have 5′ Nrd1 binding peaks. In each case the peak of binding is downstream of at least one mRNA 5′ end suggesting that the gene may be regulated by premature termination. Increased levels of RNA from cells in which Nrd1 levels were depleted confirm that PCF11 and RPB10 are negatively regulated by Nrd1 (Figure 5C and 5D). In the case of PCF11 this regulation is particularly interesting because PCF11 encodes a protein with a similar termination function to Nrd1. While Pcf11 has primarily been associated with termination of mRNAs it also plays a role in termination of non-poly(A) transcripts 30,35. Nrd1 and Pcf11 have been proposed to compete for recruitment to the transcription complex [36], [37] and the ability of Nrd1 to regulate Pcf11 expression may play a role in balancing this competition. For example, conditions that reduce Nrd1 protein levels would be expected to lead to compensatory increases in both NRD1 and PCF11 expression. Nrd1 binding also regulates expression of RPB10, a gene encoding an RNA polymerase subunit that is common to all three nuclear RNA polymerases. This observation suggests a possible role for the Nrd1 pathway in regulating expression of the transcription machinery.
In addition to sense-strand binding, we observe a large number of Nrd1 reads derived from RNAs that are anti-sense to annotated genes. Figure 4 shows that these reads are concentrated at the 3′ end of genes that are not heavily transcribed in the sense direction. Figure 6 shows Nrd1 and Rpb2 binding to antisense transcripts at the 3′ ends of the YKL151C and USA1 genes. We do not observe abundant Sen1 cross-linking to the antisense transcript despite the fact that it cross-links efficiently to the sense transcript of the downstream gene. In Figure S5 we show that YKL151C encodes a 3′ antisense CUT. We think it is quite likely that these antisense transcripts originate from divergent transcription from the downstream promoters [32], [33] and may be involved in suppressing transcription in the sense direction of YKL151C and USA1.
Our cross-linking experiments show that Sen1 cross-links to abundantly transcribed mRNAs. In Figure 7E we show examples of Sen1 binding to transcripts derived from the heavily transcribed RPL28, RPS13, RPL30 and PMA1 genes. Several arguments suggest that this Sen1 cross-linking occurs on nascent transcripts. First, Sen1 and Rpb2 both cross-link in clusters that are spaced along the transcript in coincident peaks. A similar clustering of Pol II-associated RNA has recently been attributed to pausing of Pol II as nucleosomes are removed from the path of the elongating polymerase [38]. Such clustering would not be expected on mature transcripts. Second, although upstream cross-links are less abundant on ribosomal gene transcripts, there is some cross-linking to intron sequences that are enriched on nascent transcripts. Taken together, these results suggest preferential binding to nascent RNA but we cannot rule out binding to unprocessed precursors that have been terminated and released from the template.
We note that Sen1 peaks are stronger toward the 3′ end, especially on the ribosomal protein transcripts. This is confirmed in Figure 7D by plotting all Sen1 and Rpb2 reads with respect to the 3′ end of transcripts derived from the most heavily transcribed genes [39]. Interestingly, the peak of Sen1 is broadly distributed over the 75 nt before the polyadenylation site (pA) with few if any reads extending beyond this cleavage site. In Figure 8 we show that this downstream Sen1 cross-linking peak does not correspond to Nrd1 or Nab3 cross-linking sites. A distinct peak of Pol II is observed between 25 and 50 nt downstream of the pA site. Clearly, Pol II continues to transcribe beyond the pA site but we see little evidence for Pol II more than a few hundred bases further downstream (Figure 7E).
In Figure 8 we have also compared our Rpb2 cross-linking data set with the Net-seq data set derived from RNA non-covalently associated with affinity purified Pol II [38]. While we see a very similar pattern of reads within coding regions, we notice a difference in the pattern of Pol II downstream of the pA site. The downstream peak of Pol II that we observe in our PAR-CLIP data is often missing in the Net-seq data (Figure 8).
Nrd1, Nab3 and Sen1 have been shown to form a complex with Pol II [3] and to direct the termination and subsequent processing of nascent transcripts [16], [40]. We have used an in vivo cross-linking approach [18] to map the positions of these yeast RNA-binding proteins in living cells. Our protocol maps Nrd1 and Nab3 binding sites to downstream snoRNA sequences that have previously been shown to direct termination in vivo [7], [8], [10], [16], [41] thus validating our procedure. These downstream sequences are rapidly processed by the nuclear exosome [42], [43] indicating that cross-linking of Nrd1 and Nab3 occurs preferentially on nascent transcripts.
The observation that Nrd1 ChIPs to chromatin at highly expressed genes but only cross-links to a sub-set of these transcripts (Figure 1) is consistent with a model in which Nrd1, Nab3, and Sen1 enter the early Pol II elongation complex by association with the Ser5 phosphorylated CTD [4], [6]. This association places Nrd1 and Nab3 in position to monitor the nascent transcript for suitable RNA binding sites. The presence of Spt5 in the Nrd1-Nab3-Sen1 complex [3] suggests a possible role for Nrd1 and Nab3 in an early transcription elongation checkpoint that controls the transition between early inefficient elongation and the processive elongation that characterizes many Pol II genes [1], [2]. Nrd1 and Nab3 binding to the nascent transcript could help slow elongation until all components of the elongation complex are assembled.
This early elongation checkpoint may serve as a regulatory step and several yeast genes including NRD1, IMD2, URA2, and SER3 contain Nrd1-bound RNAs derived from early elongation complexes. We provide evidence here that several additional genes can be added to this list including PCF11 and RPB10 (Figure 5); HPT1, GUA1 and ADE17 (Figure S4); and DIP5, LSR1, RSF1, SWI5, and MEP1 (not shown). Data about the transcription start site(s) of these genes is limited leaving open the question of whether these genes are regulated by premature termination [11], alternative start site selection [12], [13], [15], or promoter occlusion by an upstream ncRNA [44].
Another unexpected observation is that many poorly expressed protein-coding genes express Nrd1-bound 3′ antisense sequences. Previous studies in yeast have identified regulatory antisense transcripts for IME4, PHO5, PHO84 and the Ty1 retrotransposon [45]–[49]. In the case of PHO84 and Ty1, antisense transcripts appear to act in trans [47], [49], although in the case of PHO84 cis-suppression is also observed [48]. In Figure 6 we show several genes with Nrd1 cross-linked anti-sense RNA peaks localized to the 3′ coding region. In each case the gene exhibiting 3′ antisense transcripts is poorly transcribed in the sense direction as determined by RNA sequencing [39] and the downstream gene is highly expressed in glucose-containing media [50]. These antisense transcripts likely result from bi-directional transcription from the downstream promoter [32], [33]. The Rpd3s histone deacetylase complex has been shown to repress antisense initiation at many promoters [38] and it is possible that the Nrd1-Nab3 non-poly(A) termination pathway prevents elongation of those antisense transcripts that lack or escape Rpd3s control. The question of whether these antisense transcripts are regulatory remains to be answered, but we note that each of these Nrd1 antisense peaks correlates with Rpb2 cross-linking sites but does not show efficient Sen1 binding. We propose that Nrd1 pauses Pol II at these sites, preventing sense strand transcription either through transcription interference or by establishment of chromatin marks that are inappropriate for transcription of the sense strand.
The distribution of Sen1 is surprising on two counts. First, we failed to observe efficient cross-linking on some transcripts that had previously been shown to depend on Sen1 function for proper termination. SNR13 transcripts normally terminate just downstream of the Nrd1 binding site but in a sen1 mutant Pol II reads through into the downstream TRS31 gene [16] and Sen1 ChIPs to this downstream region [31]. Similarly, Nrd1 autoregulation is disrupted in a sen1 mutant with steady-state levels of Nrd1 mRNA increasing about 10-fold [11], [16]. The failure of Sen1 to efficiently crosslink to these transcripts can be explained in several ways. First, the RNA at these sites may interact with Sen1 in a manner that prevents close apposition of Sen1 amino acid side chains with 4SU residues in the bound RNA. A second possibility is that Sen1 may play a structural role, stabilizing the Nrd1-Nab3-Sen1 complex in a manner that does not require the helicase activity. In this model the sen1 mutation may alter the structure of the complex leading to disruption of Nrd1 and/or Nab3 termination function. A third possibility is that Sen1 may be required for expression of another factor that is required for termination of non-poly(A) transcripts. Finally, the low amount of Sen1 cross-linking may indicate that low levels of Sen1 are sufficient for proper termination at some genes. Future experiments must be directed at understanding the role of Sen1 in termination of snoRNA and attenuated Pol II transcipts.
A second unexpected result from the Sen1 cross-linking data set is the widespread distribution of Sen1 on mRNA transcripts. A role for Sen1 in mRNA 3′ end formation has been suggested by previous experiments. Sen1 interacts with Glc7, a protein phosphatase that is part of the CPF complex [51], [52]. In addition, sen1 mutants display a weak read through phenotype on some pA terminators [53], [54], [55] but not others [30]. Although a genome-wide survey of Pol II distribution in a sen1-E1597K mutant did not detect widespread read through of pA terminators [31] some read through was observed and our data shows that several of those genes including RPL43B, RPS28A, RPL36B, and SOD1 display prominent Sen1 binding near their pA site (not shown).
Our data shows that Sen1 cross-linking occurs all along mRNA transcripts, peaking at the 3′ end (Figure 7). Sen1 interacts with the Pol II CTD [5] but whether this interaction is direct is not known. Based on the increased cross-linking toward the 3′ end of genes it is possible that Sen1 interacts with the Ser2 phosphorylated form of Pol II or another protein that binds to this phosphorylated form of the CTD.
The failure of Sen1 to cross-link downstream of the pA site would seem to rule out a rho-like termination model in which Sen1 translocates along the transcript facilitating termination upon reaching the paused downstream Pol II. Recently Mischo et al. [54] have shown that Sen1 helicase activity is required to remove R loops that form at the 3′ end of some transcripts. This proposal is based in part on the identification of DNA∶RNA hybrids downstream of the pA site [54]. Sen1 helicase activity could act to remove RNA from the DNA∶RNA hybrid and expose RNA downstream of the pA site for degradation by the 5′-3′ exonuclease Rat1. Our data argue, however, that Sen1 acts upstream but not downstream of the pA site. We can clearly observe cross-linked Pol II downstream of the pA site but there is no corresponding Sen1 cross-linking in this region. Sen1 could act upstream of the cleavage site to remove RNA from R loops and allow access of the cleavage/polyadenylation machinery and subsequently the 5′-3′ exonuclease Rat1. This model is consistent with previous experiments showing a synergistic effect of sen1 and rat1 mutations [53], [55].
Our data also suggest that the downstream peak of Pol II represents the Pol II termination complex. Kinetic modeling of yeast Pol II transcription suggests a termination time of about one minute [56]. This is greater than the amount of time needed to transcribe many yeast genes, suggesting that termination is the rate-limiting step for formation of some mRNAs. We suggest that the peak of Rpb2 cross-linking we observe downstream of the pA site of heavily transcribed genes (Figure 7) represents this rate-limiting Pol II elongation complex that is in the act of terminating. We note that this downstream peak of Pol II is not as prominent in the NET-seq data (Figure 8; [38]). The NET-seq data is obtained by affinity purifying Pol II elongation complexes from chromatin after digestion with DNase I. We propose that the downstream termination complex is sensitive to DNase I digestion because of an allosteric change that takes place as the elongation complex passes through the pA site [57]. Thus, the RNA is lost from these complexes and is under-represented in the NET-seq data.
The genomic NRD1, NAB3, SEN1 and RPB2 genes in BY4733 were tagged with both 6His and biotin tags (HTB) [19]. The resulting strains expressed proteins with a slightly higher molecular weight that could be enriched on streptavidin beads (Figure S6). These strains displayed no abnormal growth phenotypes. A TAP-tagged NRD1 yeast strain was used for chromatin immunoprecipitation.
For the first two data sets, yeast cells expressing HTB tagged Nrd1 or Nab3 were grown at 30°C in 2 L of synthetic complete (SC-URA) medium supplemented with 2% dextrose, 120 µM Uracil, 0.01 µM Biotin from OD600∼0.1 to mid-exponential phase (OD600∼0.5). 4SU was added to a final concentration of 300 µM and growth continued at 30°C until OD600∼1.5. Addition of 4SU had no effect on the growth rate of yeast during the time course of the experiment. Cells were harvested by centrifugation and the cell pellet was re-suspended in 60 ml of ice-cold water, separated into two 30 ml aliquots and placed in 145×20 mm sterile tissue culture dishes kept on ice. Cells were irradiated on ice with 365 nm UV light (0.15 J/cm2) in a Stratalinker 2400 (Stratagene), three times for 10 minutes each with shaking between irradiations. Cells were pooled, centrifuged, and the cell pellet was re-suspended in 5 ml of buffer-1 (8 M urea, 300 mM NaCl, 0.5% Nonidet P-40, 50 mM sodium phosphate, 50 mM Tris-HCl, pH 8.0, and EDTA-free protease inhibitor mix for His-Tag sequences (RPI)) then frozen in droplets in liquid nitrogen. Cell droplets were kept in −80°C until processed as described below.
In the analysis of the initial Nrd1 and Nab3 data sets we observed binding to a number of RNAs derived from stress-induced genes. The RNA in these early experiments was obtained from cells that were irradiated after centrifugation and re-suspension in ice-cold water [21], a procedure that is likely to induce a stress response. To eliminate this possibility we developed a second technique to irradiate cells growing in liquid media at 30°C. Two liters of growing cells were placed in a two-liter beaker on a magnetic stirrer and irradiated from a distance of 10 cm for 10 min with a UV Power–Shot 1100 Lamp. This lamp delivers 1 W/cm2 primarily in the in the 300–400 nM range with a peak at 365 nm. To eliminate shorter wavelength light the beaker was covered with a Pyrex baking dish. Cells irradiated in this manner were processed as described below.
Protein purification was based on a previously published protocol [58]. Cell droplets were lysed in liquid nitrogen using a Spex SamplePrep 6870 freezer mill with 10 cycles of one minute of breakage and two minutes of cooling, at a frequency setting of 15 cps. Lysates were thawed at room temperature, resuspended in 5 ml of buffer-1 then sonicated using a 1/8″ microprobe tip of Branson sonifer cell disruptor Model 250/450. Sonication was performed three times at 50% power for 5 seconds with 30 second intervals at room temperature. Cell lysates were cleared by centrifugation at 40,000 rpm in Beckman L-80 ultracentrifuge at room temperature for 30 minutes using Ti 70.1 rotor. Cleared lysates were incubated with Ni-NTA agarose (QIAGEN, 500 µl slurry pre-equilibrated in buffer-1) for 3 hours at room temperature. Ni-NTA agarose was then washed in 5 ml of buffer-1; 5 ml of buffer-1, pH 6.3; and 5 ml of buffer-1, pH 6.3, +10 mM imidazole. Proteins were eluted in 8 ml of buffer-2 (8 M urea, 200 mM NaCl, 2% SDS, 50 mM sodium phosphate, 10 mM EDTA, 100 mM Tris-HCl, pH 4.3, and EDTA-free protease inhibitor mix for His-Tag sequences (RPI)). The pH of the eluate was neutralized and loaded onto streptavidin magnetic beads (New England Biolabs). A 200 µl slurry of beads was pre-equilibrated in buffer-3 (8 M urea, 200 mM NaCl, 0.2% SDS, 100 mM Tris-HCl, pH 8.0, and EDTA-free protease inhibitor mix for His-Tag sequences). After incubation overnight at room temperature the streptavidin magnetic beads were washed in 3×500 µl of buffer-3, 3×500 µl of buffer-3 with 2% SDS, 3×500 µl of buffer-3 without SDS, and then 3×500 µl of T1 ribonuclease buffer (150 mM KCl, 2 mM EDTA, 0.5 mM DTT, 50 mM Tris-HCl, pH 7.4, and EDTA-free protease inhibitor mix for His-Tag sequences (RPI Crop.)). The streptavidin magnetic beads were resuspended in 0.5 ml of T1 buffer before RNase T1 (Fermentas) was added to obtain a final concentration of 40 U/ml and the bead suspension was incubated at room temperature for 15 minutes. Beads were washed three times with 500 µl of T1 wash buffer (500 mM KCl, 0.05% NP40, 0.5 mM DTT, 50 mM Tris-HCl, pH 7.8, and EDTA-free protease inhibitor mix for His-Tag sequences (RPI Corp.)) and three times with 500 µl of polynucleotide kinase (PNK) buffer (50 mM NaCl, 10 mM MgCl2, 5 mM DTT, 50 mM Tris-HCl, pH 7.4, and EDTA-free protease inhibitor mix for His-Tag sequences). Beads were resuspended in 160 µl of PNK buffer before Thermosensitive Alkaline Phosphate (TSAP) (Promega) was added to obtain a final concentration of 0.15 U/µl, and SuperRNase inhibitor (Ambion) was added to obtain a final concentration of 1 U/µl. The bead suspension was incubated for 30 minutes at 37°C then was washed once with 500 µl of buffer-3 without SDS, and 3 times with 500 µl of PNK buffer.
Streptavidin magnetic beads from the previous step were resuspended in 200 µl of PNK buffer then γ-32P-ATP (MP Biomedicals) was added to obtain a final concentration of 0.5 µCi/µl and T4 PNK (New England Biolabs) was added to obtain a final concentration of 1 U/µl. The bead suspension was incubated at 37°C for 30 minutes before non-radioactive ATP was added to obtain a final concentration of 100 µM, the incubation was continued for 10 minutes at 37°C. The bead suspension was washed 4 times with 650 µl of T4 RNA Ligase2 truncated (Rnl2 (1–249)) buffer (2 mM MgCl2, 1 mM DTT, 50 mM Tris-HCl, pH 7.5) (New England Biolabs).
Streptavidin magnetic beads from the previous step were resuspended in 19 µl of Rnl2 (1–249) buffer combined with 19 µl of 50% Polyethylene Glycol 8000 (PEG 8000) (Promega). To the bead suspension was added 2 µl of 100 µM adenylated 3′ adapter oligodeoxynucleotide (AppATCTCGTATGCCGTCTTCTGCTTGTC; IDT), 0.4 µl of 1 M MgCl2, 2.5 µl of Rnl2 (1–249) (200 U/µl) (New England Biolabs), 1.25 µl of RNase inhibitor (40 U/µl) (Invitrogen). The bead suspension was incubated at room temperature for 4 hours then washed once with 500 µl of buffer-3 without SDS, and 3 times with 500 µl of PNK buffer. The washed streptavidin magnetic beads were re-suspended in 38 µl of PNK buffer, then to the bead suspension was added 2 µl 100 µM 5′ adapter oligonucleotide (GUUCAGAGUUCUACAGUCCGACGAUC), 1.25 µl of RNase inhibitor (40 U/µl), 2.5 µl of T4 RNA ligase (10 U/µl) (Fermentas), 3 µl of 10 mM ATP (New England Biolabs). The bead suspension was incubated overnight at 16°C then washed once with 500 µl of buffer-3 without SDS, and 3 times with 500 µl of Protease K buffer (75 mM NaCl, 6.25 mM EDTA, 1% SDS, 50 mM Tris-HCl, pH 7.5). The streptavidin magnetic beads were resuspended in 200 µl of protease K buffer follow by the addition of Protease K (Ambion) to the final concentration of 1.2 mg/ml. After incubation at 55°C for 30 minutes, the supernatant was transferred to a new tube and another 200 µl of protease K buffer was added to the streptavidin magnetic beads. The incubation was continued for 5 minutes at 55°C before the supernatant was combined and the RNA was recovered by acidic phenol/chloroform extraction followed by a chloroform extraction and an ethanol precipitation. The pellet was dried and then dissolved in 12 µl of DPEC water. The recovered RNA was divided into 3×4 µl aliquots and kept in −80°C until further preparation.
The recovered RNA was used to synthesize a cDNA library. Time course PCR amplification was performed in order to determine the optimum number of cycles for amplifying the cDNA library. One aliquot of recovered RNA was taken out of −80°C to thaw on ice, then reverse transcription oligonucleotide (CAAGCAGAAGACGGCATACGA) was added to the final concentration of 5 µM. The mixture was briefly centrifuged, heated at 70°C on a preset thermal cycler for 2 minutes and the tube was then placed on ice. To the iced tube containing the primer-annealed template RNA was added 4 µl of the premixed reverse transcription reaction mixture containing 2 µl of 5 X first strand buffer (375 mM KCl, 15 mM MgCl2, 250 mM Tris-HCL, pH 8.3) (Invitrogen), 0.5 µl of 12.5 mM dNTP mix (Fermentas), 1 µl of 0.1 M DTT (Invitrogen), and 0.5 µl of RNase inhibitor (40 U/µl) (Invitrogen). The tube was heated on the preset thermal cycler at 48°C for 3 minutes then Superscript II reverse transcriptase (Invitrogen) was added to the final concentration of 20 U/µl, the incubation was continued for 1 hour on the preset thermal cycler at 44°C. Time course PCR was carried out on a 50 µl scale using cDNA from the previous step and Phusion DNA polymerase (Finnzymes) (10 µl of the cDNA, 0.5 µl of 25 mM dNTP Mix, 10 µl of 5 x Phusion HF buffer (Finnzymes), 1 µl of 100 µM primers, 0.5 µl of 2 U/µl Phusion DNA polymerase). PCR cycle conditions of 10 s at 98°C, 30 s at 60°C, 15 s at 72°C were used. Aliquots of 6 µl were removed every other cycle starting with cycle number 14 by temporarily pausing the PCR cycle at the end of the 72°C step. The maximum cycle of the PCR was set at 30 cycles. PCR aliquots were analyzed on a 6% Novex TBE PAGE gel (Invitrogen) and the optimal cycle number for cDNA amplification was chosen as five cycles prior to reaching the saturation level of PCR amplification. The optimal cycle number PCR was performed on a 50 µl scale using cDNA prepared from another aliquot of recovered RNA (4 µl). The amplified cDNA product was separated on a 6% Novex TBE PAGE gel (Invitrogen) and the band of interest was excised and eluted using 1 x gel elution buffer (Illumina). RNA fragments with both adapters produce PCR fragments that span a range of 100–150 nt. The 5′-adapter-3′-adapter products without inserts may be detected after amplification of the cDNA as an additional PCR band at around 75 nt in length. The eluted DNA from gel extraction was ethanol precipitated followed by DNA analysis using Agilent 2100 Bioanalyzer. DNA was sequenced using an Illumina GAII sequencer (University of California, Riverside).
Sequences were trimmed using the function trimLRPatterns from the ShortRead package in Bioconductor [59]. Reads were aligned to the Sacchromyces cerevisiae genome using the short read aligner Bowtie [60]. The bowtie alignment allowed the read to have up to 1 mismatch and align once to the genome. Reads that aligned more than once to the genome will be discussed elsewhere [22]. All figures were made using the Mochiview genome browser [29]. The processed wig files for each dataset, along with fasta file of the yeast genome for alignment, can be downloaded from Gene Expression Omnibus [61] under the series number GSE31764 (http://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE31764).
Cells containing a TAP-tagged NRD1 gene were grown in yeast extract peptone dextrose (YPD) to an absorbance of about 0.8 and cross-linked with formaldehyde for 20 min at room temperature and processed for chromatin immunoprecipitation according to standard protocols [62]. ChIP and Input DNA were first amplified using linker adapted random primer and sequenase 2 and further amplified and labeled using PCR and Cy3- and Cy5- CTP, respectively. Cy dye labeled ChIP and Input DNA was combined with Agilent aCGH/ChIP hybridization buffer containing 2x Hi-RPM Hyb Buffer, Agilent blocking agent and Human cot-1 DNA and hybridized to Agilent 244K yeast tiling array (G4491A) for 40 hours on MAUI hybridization system with constant mixing. The hybridized array was washed using Agilent aCGH/ChIP-on-chip array washing buffer kit (5188–5266) and scanned on Axon GenePix Scanner (GenePix A4300) and the raw data extracted using GenePix Pro 6.0. Preliminary analysis of the ChIP data was carried out using CisGenome to identify ChIP enriched regions.
Total RNA was extracted from yeast with hot acid phenol and run on a 1% denaturing formaldehyde MOPS agarose gel, visualized by ethidium bromide and Northern blotted as previously described [11]. Samples with clear rRNA bands and no visible degradation were analyzed by quantitative real-time RT-PCR. RNA was treated with turbo-DNA-free (Ambion) according to the manufacturer's instructions for the most stringent treatment. Reverse transcription was performed using the iScript cDNA Synthesis Kit (BioRad). Real-time PCR was performed in triplicate 20 µl reactions on a CFX96 Real-time PCR detection system (BioRad) using iQ SYBR green supermix (BioRad) according to the manufacturer's instructions. Data from at least two replicate experiments were pooled using the Gene Study feature of the CFX96 real-time software, which normalizes for fluorescence intensity differences between plates. Expression was normalized to both ACT1 and 18S ribosomal RNA and the ratios were graphed relative to the wild-type control sample, which was set to 1 for each gene. Error bars represent the positive and negative range of the standard error of the mean.
Tet-promoter yeast strains (Open Biosystems) and a control strain lacking a tet-responsive promoter were seeded to an initial A600 of 0.06 in YEPD. Cultures were grown at 30° for 2.5 hrs prior to the addition of doxycycline at a final concentration of 10 µg/ml and were grown an additional five hrs before collection.
|
10.1371/journal.pntd.0006465 | Prevalence and risk factors for Taenia solium cysticercosis in school-aged children: A school based study in western Sichuan, People’s Republic of China | Taenia solium cysticercosis affects millions of impoverished people worldwide and can cause neurocysticercosis, an infection of the central nervous system which is potentially fatal. Children may represent an especially vulnerable population to neurocysticercosis, due to the risk of cognitive impairment during formative school years. While previous epidemiologic studies have suggested high prevalence in rural China, the prevalence in children as well as risk factors and impact of disease in low-resource areas remain poorly characterized.
Utilizing school based sampling, we conducted a cross-sectional study, administering a questionnaire and collecting blood for T. solium cysticercosis antibodies in 2867 fifth and sixth grade students across 27 schools in west Sichuan. We used mixed-effects logistic regression models controlling for school-level clustering to study associations between risk factors and to characterize factors influencing the administration of deworming medication. Overall prevalence of cysticercosis antibodies was 6%, but prevalence was significantly higher in three schools which all had prevalences of 15% or higher. Students from households owning pigs (adjusted odds ratio [OR] 1.81, 95% CI 1.08–3.03), from households reporting feeding their pigs human feces (adjusted OR 1.49, 95% CI 1.03–2.16), and self-reporting worms in their feces (adjusted OR 1.85, 95% CI 1.18–2.91) were more likely to have cysticercosis IgG antibodies. Students attending high prevalence schools were more likely to come from households allowing pigs to freely forage for food (OR 2.26, 95% CI 1.72–2.98) and lacking a toilet (OR 1.84, 95% CI 1.38–2.46). Children who were boarding at school were less likely to have received treatment for gastrointestinal worms (adjusted OR 0.58, 95% CI 0.42–0.80).
Our study indicates high prevalences of cysticercosis antibodies in young school aged children in rural China. While further studies to assess potential for school-based transmission are needed, school-based disease control may be an important intervention to ensure the health of vulnerable pediatric populations in T. solium endemic areas.
| The zoonotic tapeworm, Taenia solium, affects millions of impoverished people worldwide and can cause neurocysticercosis (NCC), an infection of the central nervous system which is potentially fatal. Hypothetically, children may be a vulnerable population to infection as neurological problems and cognitive impairment caused by NCC during formative school years may lead to poor academic performance, contributing to drop-out rates and, eventually, propagating cycles of poverty. We carried out a school-based study of T. solium cysticerosis in primary school-aged children in rural western Sichuan. Our results indicate high levels of T. solium exposure in young school-aged children in rural China. While further studies to assess disease transmission within schools are needed, school-based disease control may be an important intervention to ensure the health of pediatric populations at risk for infection.
| Infection with the zoonotic tapeworm Taenia solium affects millions of people living in poverty throughout Asia, Africa, and Latin America [1]. Considered a neglected tropical disease, infection is linked to inadequate sanitation and hygiene, presence of free roaming pigs, and poverty [2]. Infection in humans has two manifestations: intestinal taeniasis where humans serve as the definitive hosts for the adult tapeworm which inhabits the gastrointestinal tract, and cysticercosis, a tissue infection where humans are the accidental dead-end host for the cystic larvae (cysticercus).
Intestinal infestation with the adult tapeworm develops when humans consume improperly cooked pork containing cysticerci. The consumed cyst is released in the small intestines where the adult worm develops, attaches to the intestinal wall, and liberates thousands of eggs, which are shed, along with gravid proglottids, in human feces. T. solium eggs contaminate the environment and are consumed by pigs that ingest human feces directly or indirectly through contaminated agricultural products. Once consumed, the larval forms encyst in porcine muscle completing the cycle [3]. Human cysticercosis develops when T. solium eggs are consumed by humans through auto-infection, consumption of contaminated food or water, or close contact with a tapeworm carrier [4]. Upon ingestion of mature eggs, the hatched parasite migrates to tissues throughout the body including muscle, sub-cutaneous tissues, orbits, and the central nervous system (CNS) [5–7].
Neurocysticercosis (NCC) develops when T. solium larva establishes itself in the CNS and may lead to morbidity which can be fatal [5,8,9]. NCC causes a range of symptoms depending on number, stage of involution, volume, and location of the lesions [10], including seizures, chronic headaches, focal neurological deficits, psychiatric disturbances, and cognitive impairment [3,11–15].
The full prevalence of NCC is difficult to establish, it is estimated to be responsible for 29% of acquired epilepsy in endemic areas [16], and has been identified as a leading cause of death from foodborne disease resulting in 2.8 million disability-adjusted life years lost in 2010 [2]. T. solium has been reported throughout China, with hyperendemic foci mainly in southwest regions [17]. Overall, NCC may affect up to an estimated 7 million people in China [1]. However, the risk factors for infection and impact of disease in rural areas remain poorly characterized [17,18].
Children may represent an especially vulnerable population to NCC, with resulting neurological problems and cognitive impairment during formative school years possibly leading to poor academic performance, contributing to high drop-out rates and, eventually, propagating cycles of poverty. Schools may represent centers of transmission combining poor hygienic standards and close contact in a large and vulnerable population. Despite this hypothetical risk, the prevalence of cysticercosis within schools has not been well evaluated.
Here we report the prevalence of cysticercosis antibodies and associated risk factors in school-aged children using school-based sampling in western Sichuan, People’s Republic of China.
The school principals, who are the children’s legal guardians while they are boarding at school, provided initial written consent for student participation, and each student provided verbal assent prior to participating. Consent and assent status were documented by field staff, and students who did not assent did not participate in the questionnaire or blood collection. Field staff and school staff described the purpose of the study using pre-written scripts and were available on hand to answer questions that students might have either about the study or questionnaires. Information on the study and a consent form were part of the take-home questionnaire, requesting written consent from adult caretakers in regard to their and their family member’s involvement. All participants were allowed to keep a copy of the consent/assent document. The study and all methodology were approved by the institutional review board of Stanford University (study ID 35415) and the ethical review board of West China School of Public Health, Sichuan University (K2015031).
The data used in this analysis were collected as part of an epidemiological field study conducted in November 2015 investigating T. solium cysticercosis and NCC prevalence, associations between NCC infection and academic performance, and risk factors for infection in school-aged children. Results of the cognitive and academic assessments, prevalences of NCC cases, and frequency of neurologic problems will be presented in another manuscript and are not further discussed here.
The study was carried out in rural mountainous areas of western Sichuan. Located at the eastern extremity of the Tibetan Plateau and with an average altitude of 2700 meters, these areas are largely characterized by smallholder farmers who raise pigs and partake in small-scale agriculture. These areas were selected as they had been identified as having high endemicity of Taenia species based on smaller scale studies [18]. In these remote areas, there is no routine mass antihelminthic drug administration for pediatric or adult populations.
To collect data which could be generalizable to school-aged children living in farming and pig raising communities in western Sichuan with known T. solium risk factors we used a school-based sampling technique. The study was conducted in three counties, each from one of three prefectures in western Sichuan. In selecting the three study counties (S1 Fig), we considered all counties within Aba (13 counties total), Ganzi (18 counties total), and Liangshan (16 counties and 1 autonomous county total) prefectures for the study. All counties in these prefectures are known to have smallholder pig raising activity and risk factors for human cysticercosis including open defecation and free range pigs based on reports from local public health workers to Sichuan Centers for Disease Control and Prevention (Sichuan CDC). Because we wanted to enroll students at risk for T. solium cysticercosis, we only considered counties where previous cases of human cysticercosis and NCC had previously been detected by Sichuan CDC or where there were previous reports to suggest T. solium taeniasis [18]. We further favored counties with the largest populations of fifth and sixth grade students as estimated by demographic data collected previously by colleagues at Sichuan CDC. Because these mountainous communities are often difficult to access due to poor roads and long distances, we focused on counties where all schools could be accessed by field teams and where local public health services were supportive of the field work.
All schools enrolling fifth and sixth graders were sampled within the three selected counties and all students in both grades were sampled at each school. We selected fifth and sixth grade students for the cognitive study because standardized tests can be administered to this age group easily and these students are old enough to have repeat and significant exposures.
At each school, field teams administered a student questionnaire, aseptically collected approximately 5 ml of blood by venous puncture for T. solium cysticercosis IgG testing by ELISA, and provided a take-home questionnaire to be completed by the child’s parent or guardian.
Student questionnaires (S2 Document) covered basic demographic data, home environment and family asset ownership, animal ownership, pork consumption, and student toileting behavior. If the family owned pigs, students were asked if they ever saw their family’s pigs eating human feces or if the pigs ever went to the areas where people defecated. While data on specific drugs could not reliably be collected, use of antihelminthics was assessed by asking students if they had taken any medication for “intestinal worms” in the past year. Finally, students were asked about symptoms and perceptions of intestinal worms. To assess if tapeworms or proglottids might be present in the child’s feces, students were asked if they had seen “worms” or “pieces of worms” in their feces.
Children were also provided with a take-home adult questionnaire which was completed by the head of household in their home. After completion, adult questionnaire forms were returned to the school by the child. On the adult questionnaire, head of households were asked about pig ownership and pig husbandry methods over the last year and slaughtering and meat preparation practices. Adults were also asked about agricultural practices in the five years preceding the questionnaire and about their knowledge and attitudes towards intestinal worms and the administration of medications for intestinal worms.
Serum was tested using an enzyme-linked immunosorbent assay (ELISA) based on low-molecular-weight antigens (LMWAgs) of T. solium cysticerci collected from pigs in Chinese endemic areas. LMWAgs based assays have been shown to be highly sensitive and specific [19,20], have been used in previous field studies [21], and are especially attractive—given their low cost, quantifiable result, and simplicity—for use in low resource areas [20]. Detailed assay methodology has been published previously [20]. Antigen for the LMWAgs ELISA was obtained from cyst fluids of T. solium metacestodes collected from infected pigs in endemic areas of China. ELISA plates (Nunc-ImmunoTM plate, Maxisorp TM Surface, Thermo Fisher Scientific, Denmark) were coated with 100 μl of diluted LMWAgs at 1 μg/ml in PBS overnight at 4°C. Plates were rinsed twice with PBS containing 0.1% Tween 20 (PBST) and then blocked with 300 μl of blocking solution (10 mM Maleic acid pH 7.5, 150 mM NaCl, 1.0% casein, 0.1% Tween 20) at 37°C for 1 hour. Serum samples were diluted in blocking solution at 1:100. Plates with 100 μl of diluted sera in duplicate wells were incubated at 37°C for 1 hour. The wells were washed 3 times with PBST, incubated with 100μl of recombinant protein G conjugated with peroxidase (Invitrogen) at 1:4000 in blocking solution at 37°C for 1 hour. After washing 3 times with PBST, plates were incubated with 100 μl of substrate (0.4 mM 2,2’- azino-di-[3-ethyl-benzhiazoline sulfonate] in 0.1 M citric acid buffer, pH 5.3) for 30 minutes at 37°C. Color reaction was stopped by application of 1% SDS in each well. The optical density at 405 nm was evaluated with an ELISA reader. The cut-off point was determined as the mean optical density (OD) plus 3 times the standard deviation for a panel of serum samples obtained from healthy Chinese donors (n = 30).
We first characterized prevalence of cysticercosis antibodies, student demographics, reported toileting behaviors, and reported pig husbandry and agricultural practices using basic descriptive statistics. We constructed mixed-effects logistic regression models to better characterize differences in T. solium antibody seroprevalence at the school-level, to study associations between T. solium cysticercosis exposure and demographic, environmental, and behavioral factors, and to investigate what factors affected the likelihood of children receiving deworming medications in the year preceding the study.
We identified schools with higher student populations with T. solium cysticercosis antibodies by building mixed-effects logistic models with the serologic result as the dependent variable, school as the independent variable, and county as a random effect. A school with a cysticercosis antibody prevalence closest to the mean value in the dataset was used as the model reference.
To build models to assess the associations between risk factors and the presence of T. solium cysticercosis IgG antibodies or deworming medication administration, we first framed a causal diagram to identify associations between variables of interest (Fig 1) [22]. In the model assessing associations between risk factors and the presence of T. solium cysticercosis IgG antibodies we used the child’s serologic result for human cysticercosis IgG as the dependent variable. In the model assessing deworming medication administration we used the report of the child receiving medication for intestinal worms in the year preceding the study as the dependent variable.
Missing values in independent variables were imputed assuming an ignorable missingness mechanism using multivariate imputation by chained equations [23]. Numeric variables were imputed using predictive mean matching, logical variables using logistic regression, and categorical variables with more than two levels with multinomial logit models [23]. Predictors for multiple imputation included the independent variable in question, all dependent variables outlined in the causal diagram, demographics (age, sex, ethnicity, and asset score), and geographic location (school and county). Fifty imputed datasets were generated. Conditional rules were applied to ensure that imputation did not create impossible combinations (for example, a non-crop growing household reporting using human waste as crop fertilizer). Finally, results obtained using multiple imputation were compared both to available-case and complete-case analyses (shown in supplemental tables). All results reported within the manuscript are from multiple imputed analyses. Odds ratios were calculated by pooling results across all imputed datasets.
For each of the two dependent variables, we first constructed mixed-effects models consisting of a single independent variable controlling for school clustering as a random effect. To select variables to include in a best-fit multivariable mixed-effects logistic regression model, we used a two-step approach. Five imputed datasets were used for variable selection, and the final selected models were run on all fifty imputed datasets [24]. We first used information-theoretic model selection to generate all possible combinations of independent variables which had resulted in p-values less than 0.1 in our initial analysis and selected the model with the lowest corrected Akaike information criterion (AIC) for each of the five imputed dataset [25]. In the second variable selection step, variables selected in at least 50% of the five imputed datasets were assessed by backwards selection and retained if the Wald test resulted in a p value of less than 0.05 [24, 26]. To add a measure of wealth to our models we used principal components analysis (PCA) to aggregate asset ownership variables into one standardized asset score [27].
To identify if exposures and reported behaviors differed between students attending schools with the highest seroprevalences of T. solium IgG antibiodies compared to students attending the lowest prevalence schools, we compared the proportion of students reporting selected behaviors and exposures in the highest prevalence schools to students in all remaining lower prevalence schools using Fisher’s Exact Test.
Independent continuous variables included in mixed-effects models were confirmed to be linearly related to the log odds. We assessed multicollinearity between independent variables in multivariable models using variance inflation factors.
Analysis was conducted in R [28] utilizing the lme4 package [29] for logistic mixed-effects models, the MuMIn package for information-theoretic model selection [30], and the MICE package to perform multivariate imputation by chained equations [23].
A total of 3036 fifth and sixth grade students completed the student questionnaire. A total of 169 students refused to give blood, resulting in a final total of 2867 students with both serologic and exposure data included in the analysis. The highest refusal proportion was in Yajiang County where 8% (104/1242) of students refused to give blood, this was followed by Muli County where 5% refused (55/1071) and Ruoergai County where 2% (12/728) refused. The higher refusal proportion in Yajiang was driven by a single school were several teachers reportedly suggested their students refuse blood draws and 30% (35/118) of students in the targeted study population within the school refused to participate. Refusal proportions were 5% in the fifth-grade population (74/1634) and 7% (97/1407) in the sixth-grade population. The final study sample consisted of 1016 students in 12 schools in Muli County, 713 students in 9 schools in Ruoergai County, and 1138 students in 6 schools in Yajiang County. Of the 2867 students included in the analysis, 211 failed to return the take-home household questionnaire, resulting in a dataset including both student and parental reported data for 2656 students.
Enrolled students (Table 1) had a mean and median age of 13; 52% were female (1474/2847); 84% were of Tibetan ethnicity (2398/2866); and the majority boarded in school dormitories (65%, 1852/2861), sleeping and eating most meals during the week at school and returning to their village at regular intervals on weekends.
Pig ownership was commonly reported, with 74% of households (2107/2843) owning pigs at home. Families reported owning a median of 3 pigs. Seventy percent of household heads (1220/1735) in pig-owning households reported that their family allowed their pigs to freely forage in the surrounding village and mountains. Sixty-one percent (1060/1724) of household heads in pig-owning households reported that they had observed their pigs consuming human feces in the environment while foraging, while 21% (513/2425) reported their pigs consumed human feces from their own household.
Pork consumption was almost universal, with only 3% of children reporting that they never consumed pork (79/2865). Sixty-one percent (1737/2825) of children reported that pork they consumed at home was raised in their household. Fifteen percent (431/2781) of children reported consuming undercooked pork in the past month. In the take-home household questionnaire, 21% (304/1435) of head of households in pig-owning households that slaughtered pigs reported seeing “cysts” in freshly butchered pork in the five years preceding the study.
Adults in the majority of households (89%, 2202/2470) reported growing crops. Crops were generally for household animal and human consumption and less commonly used as a source of cash income: 96% (2106/2184) consumed their own crops, 73% (1493/2052) fed their crops to their pigs, and 30% (635/2148) sold their crops for profit. Human feces were used as a fertilizer in 33% (696/2101) of households that grew crops. Human feces were often not treated—for example by composting or fermentation—before applying to crops. Of the households that reported using human feces as a fertilizer, 31% (217/687) reported never treating and only 16% (112/687) always treated prior to using.
Thirty-eight percent (1095/2861) of children reported having no toilet in their home. Forty-six percent (1330/2870) of children reported defecating someplace other than a toilet, with the two most common locations for outdoor defecation reported as within the village limit but outside the courtyard of their home (29%, 391/1330) and in the fields surrounding the village (55%, 740/1330).
Overall, 11% (283/2606, 95% confidence interval [CI] 10–12%) of all fifth and sixth grade children self-reported that they had seen what appeared to be worms or worm segments in their feces (Table 2). The highest prevalence was in Muli County, where 14% (141/1013, 95% CI 12–16%) reported worms or worm segments in their feces.
The overall prevalence of serum T. solium cysticercosis IgG antibodies in fifth and sixth grade students in the study area was 6% (180/2867, 95% CI 5–7%) (Table 2). The county prevalences in all enrolled fifth and sixth grade students ranged from 8% (78/1016, 95% CI 6–10%) in Muli County to 6% (72/1138, 95% CI 5–8%) in Yajiang County, and 4% (30/713, 95% CI 3–6%) in Ruoergai County.
Three schools had significantly higher prevalences of students with serum T. solium cysticercosis IgG antibodies when compared to a school with the mean prevalence in the study area (Fig 2): a school in Muli County with a prevalence of 15% (16/105, odds ratio [OR] 2.3, 95% CI 1.2–4.6), a school in Yajiang County with a prevalence of 22% (19/88, OR 3.6, 95% CI 1.9–6.9), and a school in Ruoergai County with a prevalence of 20% (22/111, OR 3.2, 95% CI 1.7–5.9). Four schools, three in Ruoergai County and one in Muli County, had no students with antibodies, although these results were not found to be statistically different than the mean prevalence in the study area.
In mixed-effects logistic models consisting of a single independent variable (Table 3, see S1 Table for comparison of available-case, complete-case, and multiple imputed analyses) and controlling for school-level clustering, children were more likely to have IgG antibodies to T. solium cysticerci if they lived in households that owned pigs (4% vs 7%, OR 1.81, 95% CI 1.09–3.01) and if the family reported feeding household human feces to pigs (5% vs 11%, OR 1.54, 95% CI 1.07–2.24). While the odds ratio crossed one, suggesting no clear benefit, there was a trend suggesting treatment of human feces prior to use as a crop fertilizer resulted in less cysticercosis exposure: 6% of children had cysticerosis antibodies in households reporting never treating feces before use compared to 2% if human feces were always treated. There was also a trend towards more children with cysticercosis antibodies in households that did not own toilets (5% compared to 8%). Children who reported worms or worm segments in their feces were more likely to have serologic evidence of T. solium cysticercosis (10% compared to 6%, OR 1.60, 95% CI 1.03–2.5).
The factors most associated with the presence of IgG antibodies to cysticercosis were pig ownership, feeding human feces to pigs, the presence of worms or worm segments in the child’s feces, and the child having been given medication for gastrointestinal worms (see S2 Table for variable selection results). In the multivariate model (Table 3), children who came from households that owned pigs and reported feeding the household’s human feces to their pigs were more likely to have serologic evidence of cysticercosis antibodies (adjusted OR 1.81, 95% CI 1.08–3.03 and adjusted OR 1.49, 95% CI 1.03–2.16, respectively). In the multivariate model, children who reported worms or worm segments in their feces were more likely to have antibodies (adjusted OR 1.85, 95% CI 1.18–2.91), children who reported receiving medication for gastrointestinal worms in the year preceding the study were less likely to have cysticercosis antibodies (adjusted OR 0.52, 95% CI 0.31–0.90).
Students attending the three schools with the highest seroprevalences of T. solium IgG antibiodies differed from those attending schools with lower prevalences in reported demographics, behaviors, and exposures (Fig 3). Students attending the highest prevalence schools were more likely to be Tibetan (OR 7.15, 95% CI 3.66–13.99), report boarding at school (OR 3.69, 95% CI 2.64–5.14), and come from households without toilets (OR 1.84, 95% CI 1.38–2.46). Students attending the highest prevalence schools were less likely to have received medication for gastrointestinal worms in the year preceding the study (OR 0.36, 95% CI 0.20–0.66) (Fig 3A).
Students attending the highest prevalence schools were more likely to come from pig owning households (OR 4.75, 95% CI 3.08–7.33), report that their household’s human feces were fed to pigs (OR 1.55, 95% CI 1.18–2.05), and report that they had seen pigs consuming human feces in the environment (OR 2.66, 95% CI 1.93–3.67) (Fig 3B). Students attending the highest prevalence schools were more likely to come from households reporting always allowing their pigs to freely forage (OR 2.26, 95% CI 1.72–2.98) and less likely to come from households not allowing their pigs to forage (OR 0.30, 95% CI 0.20–0.44). While the frequency of reported pork consumption (Fig 3C) was the same in both student populations, students attending the highest prevalence schools were more likely to report that their pork came from home raised pigs (OR 2.62, 95% CI 1.94–3.55).
Differences in agricultural practices were less pronounced (Fig 3D), although children attending the highest prevalence schools were slightly more likely to come from households growing crops (OR 2.82, 95% CI 1.59–5.00), using human feces as a fertilizer (OR 1.34, 95% CI 1.06–1.70), and attempting to treat human feces through fermentation or composting before use as a fertilizer (OR 1.35, 95% CI 1.04–1.77).
Thirty-two percent of children reported having “intestinal worms” in the year preceding the study (818/2579). Most children reported that they had realized they were infected due to abdominal pain (55%, 448/818), while a smaller number reported seeing worms or worm pieces in their feces (20%, 163/817) or had been told that they had intestinal worms by a doctor (14%, 114/818).
When asked about their impression of gastrointestinal worms in their children, 35% (866/2469) of adult household respondents felt that intestinal worms had no adverse effects, 30% felt that worms could stunt children’s growth (765/2469), and 3% (69/2468) felt that worms could have a positive effect on children. When asked to identify the best treatments for gastrointestinal worms, 75% of adults (1832/2431) identified medication provided by a doctor, 10% (232/2431) suggested a combination of reducing outdoor activity and drinking hot water, and 9% (210/2431) suggested that spicy food should be consumed.
Fourteen percent (379/2613) of children reported that they had taken a medication for gastrointestinal worms in the year preceding the study. In mixed-effects logistic models consisting of a single independent variable (Table 4, see S3 Table for comparison of available-case, complete-case, and multiple imputed analyses) and controlling for school-level clustering, children who reported their ethnicity as “other” (OR 2.43, 95% CI 1.36–4.32) and who reported seeing worms or worm segments in their feces in the preceding year (OR 4.41, 95% CI 3.29–5.91) were more likely to have received treatment. Children who were older (OR 0.89, 95% CI 0.82–0.97), were of the Yi ethnicity (OR 0.48, 95% CI 0.23–0.99), were boarding at school (OR 0.52, 95% CI 0.38–0.70), and were wealthier as classified by household asset score (wealthiest quartile: OR 0.60, 95% CI 0.40–0.91) were less likely to have received medication for gastrointestinal worms in the year preceding the study.
Parental educational level and impressions of gastrointestinal worm influenced if children received treatment. In families where adults reported worms having no adverse effects, fewer children received medication (OR 0.72, 95% CI 0.54–0.95). Higher levels of education were associated with increasing medication administration, with parents achieving a junior high education or higher more likely to provide medication (junior high school: OR 1.60, 95% CI 1.08–2.38; high school or higher: OR 2.37, 95% CI 1.48–3.79).
The factors most associated with children receiving medication for gastrointestinal worms were age, school boarding status, level of parental education, parental understanding of adverse events caused by gastrointestinal worms, and the child reporting worms or worm segments in their feces in the last year (see S4 Table for variable selection results). In the multivariate model (Table 4), children of more highly educated parents (high school or higher: adjusted OR 1.81, 95% CI 1.11–2.98) and with worms or worm segments in their feces (adjusted OR 4.43, 95% CI 3.29–5.98) were more likely to receive treatment for gastrointestinal worms. Children who were older (adjusted OR 0.91, 95% CI 0.83–0.99), boarding at school (adjusted OR 0.58, 95% CI 0.42–0.80), and who had parents who felt intestinal worm infestation had no adverse effects (adjusted OR 0.68, 95% CI 0.51–0.92) were less likely to receive medication.
Our study demonstrates high prevalence of T. solium cysticercosis antibodies in school-aged children in poor, pig-raising areas in western Sichuan. The use of schools as a unit, rather than the typical village, is a unique approach and reveals variation in T. solium antibody seroprevalence across schools in close geographic proximity.
We identified three schools with significantly higher prevalences of T. solium cysticercosis antibodies than surrounding schools. Schools with the highest prevalence of T. solium cysticercosis antibodies had differences in reported behaviors and exposures compared to lower prevalence schools, with higher proportions of students in the highest prevalence schools reporting the consumption of home raised pigs, living in households without toilets, and coming from households were the family’s pigs are allowed to freely forage and fed human feces.
The seroprevalence of cysticercosis in children varies widely in the literature, from approximately 20% of 10–19 years olds found to be antigen positive in a village based study in the Democratic Republic of Congo [31], to approximately 12% of 11–20 year olds having antibodies in a hyperendemic area of Peru [32], and a study in three provinces of Burkina Faso showing T. solium antigen prevalence ranging from 2.3% to 0.7% in the youngest cohorts [33]. Although it is difficult to compare results given differing laboratory methodology and our school centered approach, the prevalence of T. solium cysticercosis antibodies in the highest prevalence schools in our study seem similar to levels reported in children in high endemic areas.
The risk factors most associated with cysticercosis antibodies in children identified in our best-fit multivariate analysis included pig ownership, the child self-identifying worms or worm segments in their feces, and households allowing their pigs to consume human feces. The link between the presence of antibodies and a recent history of young children possibly passing proglottids is consistent with previously published literature [32]. It is unclear from our study how many children are tapeworm carriers and are auto-infecting themselves, although given the poor handwashing practices among young children [34], there is likely substantial risk for auto-infection. Another potential explanation for this finding is that children who are passing proglottids are, along with their family members, consuming undercooked pork and therefore are likely surrounded by multiple members of their household who are harboring intestinal tapeworms.
The consumption of human feces by pigs results in infected pork and likely results in higher proportions of human intestinal infestation with the adult T. solium tapeworm. Children and adults in the study area often report either pigs eating the household’s human feces or report seeing their pigs eating human feces while foraging in the environment. Free range pigs’ access to human feces—made easier by lack of latrines and open defecation—has been frequently identified as a risk factor for cysticercosis [33, 35–37]. The use and acceptance of human feces as pig feed has been recognized as affecting household practices and preferences, for example respondents in a qualitative study on latrine use in Zambia voiced concern that building latrines would result in less available pig feed [38].
Risk factors identified in the school comparison are also consistent with factors previously identified in the literature. Pig ownership has been identified as a risk factor for cysticercosis and taeniasis in studies in Africa and South America [33,39], and higher risk of seropositivity has been seen in households that consume home raised pigs [39].
While not always meeting criteria for statistical significance, our analysis did suggest trends between agricultural techniques and cysticercosis exposure. Our comparison of the highest with lower prevalence schools suggested that children who attend the highest prevalence schools are more likely to come from households that grow their own crops. The use of human feces as fertilizer for crops has been recognized as a potential risk factor in previous studies [40,41]. While some households reported treating human feces prior to use as fertilizer, this practice is not common and the effectiveness of household techniques is unclear, especially given that consistently achieving the required levels of temperature, pH, and dryness to deactivate T. solium eggs may be difficult in household latrines [42]. The role of treating human feces prior to use as fertilizer and other agricultural practices in reducing infection risk deserves further characterization and study.
Unlike previous published studies, our study did not show any association between serologic status and pork consumption levels [33,39]. Our failure to detect this is perhaps due to the high pork consumption in the area and the fact that children, who do not prepare their own meals at home or at school, are likely poor judges of their pork intake. Additionally, we did not find an association between serologic status and poverty level [33]. This may be related to economic status being similar across the entire study area or to little variation in risk factors with increasing wealth given the agricultural and remote nature of our study communities.
Given that cyticercosis cases cluster around tapeworm carriers [32, 35, 43], it is possible that schools are acting as centers for transmission in pediatric populations. Schools represent large congregations of children, and risk for fecal-oral transmission and passage of eggs from tapeworm carriers is likely high. If this is the case, efforts to reduce school fecal-oral transmission may serve as a tool to interrupt disease transmission.
Treating human tapeworm carriers with antihelminthic medication eliminates the adult tapeworm, destroying the source of infection and preventing human and porcine cysticercosis [44]. While we did not collect data on specific antihelminthic use, our study, which did include questions regarding administration of medication to treat gastrointestinal worms, suggests that few children receive therapy. More concerning, our analysis showed that children who are boarding at school were less likely to receive medication than students living at home. Treatment is most likely to be administered by more educated parents who are aware of the potential adverse effects of tapeworm infestation. If tapeworm carriers are present in schools, the distribution of antihelminthic medications at schools could decrease possible school-based transmission between students.
Our study has some weaknesses that limit our scientific inference. We used a LMWAgs ELISA to detect antibodies to human T. solium cysticercosis. While historically less sensitive and specific than enzyme-linked immunotransfer blot (EITB) assays [45], ELISA performance for serodiagnosis of human cysticercosis has improved with the development of more sophisticated methods for producing antigenic proteins [46]. Use of LMWAgs rather than crude cyst fluid results in improved performance and less cross-reactivity with other pathogens [20]. However, evaluation of LMWAgs based ELISAs continues to suggest some weak cross-reaction with alveolar and cystic echinococcosis [20]. Because echinococcosis is endemic throughout regions of northwestern Sichuan [47], cross-reactivity may be causing us to over-estimate the prevalence of human T. solium cysticercosis antibodies. In this case, we suspect that misclassification caused by cross-reactivity is minimal given that echinococcosis in northwestern Sichuan is more common in pastoral herding communities than the farming communities which inhabit the regions included in this work [47].
Because the overall larger study was designed to assess the relationship between NCC and cognitive outcomes, we selected counties and schools to maximize disease based in small initial studies suggesting presence of NCC and human cysticercosis in the study areas. This means that our results may not be fully representative over a larger geographical area, as prevalences of disease may be higher in our area of study.
Some of the risk factors that failed to achieve significance in the best-fit model are risk factors for gastrointestinal taeniasis and likely failed to achieve significance because our selected laboratory outcome was a serologic test for cysticercosis antibodies. Our measure of taeniasis was based on students self-reporting worms or worm segments in their feces. We were not able to confirm if these reported gastrointestinal infestations were caused by T. solium nor were we able to conduct large scale stool testing to detect cases of taeniasis. T. saginata and T. asiatica are known to be present in the region [18], so some of these self-reported cases may represent other Taenia species or soil transmitted helminths. Our findings do suggest that students reporting worms or worm segments in their feces are more likely to have T. solium antibodies. Given the likely inclusion of gastrointestinal worms other than T. solium in our data collection, we may be underestimating the risk for cysticercosis associated with T. solium taeniasis. Stool testing and laboratory confirmation will be required to better characterize the prevalence of taeniasis and better clarify the associated risk for cysticercosis in the study area.
Because we do not have infection data on pigs raised in the study area, we cannot correlate prevalence with presence and density of infected pigs. Because of the mountainous terrain, long distances, and presumed limited movement of both villagers and pig populations, it is very possible that prevalence of T. solium cysticerci in pigs may vary widely in areas that are geographically proximal but isolated due to terrain features, and this may explain some of the variation in human seroprevalence. Given the complex biology of T. solium, the addition of measures for gastrointestinal taeniasis in humans and prevalence of cysticercosis in pigs would provide a more complete picture of the disease ecology.
Finally, because our study is questionnaire based, children failing to answer questions and adults failing to return take home questionnaires may have limited our ability to make school specific characterizations. In this case, however, overall participation in the study across the entire geographic area was high.
We have shown a high prevalence of T. solium cysticercosis antibodies in school-aged children with school-based clustering. These findings raise concerns for NCC in school-aged children and possible cognitive deficits caused by CNS infection, which could result in long-term negative health, economic, and social effects. T. solium is an eradicable disease. Combined approaches addressing community education, improvements in hygiene and sanitation, improved pig management and meat handling, treatment of tapeworm carriers with antihelminthics, and porcine treatment through vaccination and chemotherapy have shown success in reducing transmission [48–50]. While further work identifying tapeworm carriers and potential routes of transmission within schools is needed, our work raises the hypothesis that schools may be sites of T. solium cysticercosis transmission and that school based interventions may, therefore, be an important addition to reduce disease among vulnerable pediatric populations in T. solium endemic areas.
|
10.1371/journal.pgen.1000336 | Pervasive Hitchhiking at Coding and Regulatory Sites in Humans | Much effort and interest have focused on assessing the importance of natural selection, particularly positive natural selection, in shaping the human genome. Although scans for positive selection have identified candidate loci that may be associated with positive selection in humans, such scans do not indicate whether adaptation is frequent in general in humans. Studies based on the reasoning of the MacDonald–Kreitman test, which, in principle, can be used to evaluate the extent of positive selection, suggested that adaptation is detectable in the human genome but that it is less common than in Drosophila or Escherichia coli. Both positive and purifying natural selection at functional sites should affect levels and patterns of polymorphism at linked nonfunctional sites. Here, we search for these effects by analyzing patterns of neutral polymorphism in humans in relation to the rates of recombination, functional density, and functional divergence with chimpanzees. We find that the levels of neutral polymorphism are lower in the regions of lower recombination and in the regions of higher functional density or divergence. These correlations persist after controlling for the variation in GC content, density of simple repeats, selective constraint, mutation rate, and depth of sequencing coverage. We argue that these results are most plausibly explained by the effects of natural selection at functional sites—either recurrent selective sweeps or background selection—on the levels of linked neutral polymorphism. Natural selection at both coding and regulatory sites appears to affect linked neutral polymorphism, reducing neutral polymorphism by 6% genome-wide and by 11% in the gene-rich half of the human genome. These findings suggest that the effects of natural selection at linked sites cannot be ignored in the study of neutral human polymorphism.
| There is much reported evidence for positive selection at specific loci in the human genome. Additional papers based on comparisons between the genomes of humans and chimpanzees have also suggested that adaptive evolution may be quite common. At the same time, it has been surprisingly hard to find unambiguous evidence that either positive or negative (background) selection is affecting genome-wide patterns of variation at neutral sites. Here, we evaluate the prevalence of positive or background selection by using two genome-wide datasets of human polymorphism. We document that levels of neutral polymorphism are substantially lower in the regions of (i) higher density of genes and/or regulatory regions, (ii) higher protein or regulatory divergence, and (iii) lower recombination. These patterns are robust to a number of possible confounding factors and suggest that effects of selection at linked sites cannot be ignored in the study of the human genome.
| The neutral theory of molecular evolution [1] postulates that adaptive substitutions occur so rarely that they can be safely ignored in most studies in population genetics or molecular evolution. This view has dominated the field of molecular evolution for the past 40 years. However, the past 4–6 years have seen a strong challenge to this view. This challenge comes not only from numerous studies detailing specific cases of molecular adaptation in a number of organisms (for example, see [2]–[8]) but also, and most compellingly, from a number of studies that indicate that adaptation might be common on the genomic scale [9]–[17].
High rates of adaptation on the genomic scale have been inferred from the excess of substitutions in functional regions relative to neutral expectations. The neutral expectations are derived from the polymorphism data at functional and putatively neutral sites and the divergence at the neutral sites using the reasoning of the McDonald-Kreitman (MK) test [18]. The excess in the number of substitutions at functional sites over this expectation can be used to estimate the number of adaptive substitutions [10],[19]. McDonald-Kreitman approaches can be modified to account for the presence of deleterious polymorphisms in the sample and the effects of demographic processes on polymorphism [10],[20],[21]. The approach can also be extended to estimate rates of adaptation in regulatory regions [12],[22].
McDonald-Kreitman analysis indicates that adaptive evolution in functional regions might be common in a range of organisms. In Drosophila, it has been estimated that from 30 to 60% of amino acid substitutions and ∼20% of substitutions in non-coding regions are adaptive [10], [11], [16], [23]–[25]. The rate appears similarly high in E. coli (>56% of amino acid substitutions are adaptive) [26] but not in Arabidopsis (0–5% of amino acid substitutions are adaptive) [27] and yeast [28].
In humans, McDonald-Kreitman-based estimates have varied from zero to ∼35% of all amino acid substitutions being adaptive [15], [29]–[33]. A recent estimate by Boyko et al [21] used information from the allele spectra of nonsynonymous and synonymous SNPs in human genes and the divergence with chimpanzee orthologs to estimate that ∼10% of amino acid substitutions between humans and chimpanzees have been fixed by positive selection. Thus, some of these studies suggest that adaptation might be fairly common in humans, although probably substantially less common than in Drosophila or E. coli.
McDonald-Kreitman approaches are very powerful at detecting positive selection, however, they can be misleading for a variety of reasons [15],[34],[35]. For example, if the strength of purifying selection over the evolutionary period separating two species has been different than it is in the present, McDonald-Kreitman-based approaches can either over- or underestimate the rate of adaptive evolution. As these estimates do not provide consistent answers about the prevalence of adaptation in humans and because they can be misleading under plausible demographic scenarios, reaching more reliable conclusions about the importance of adaptations in humans requires the investigation of other signatures of positive selection.
An adaptive substitution reduces the level of polymorphism at neutral sites in its vicinity in a phenomenon known as a selective sweep [36]. The width of the region in which the polymorphism is reduced is inversely proportional to the local recombination rate and directly proportional to the selection coefficient associated with the adaptive substitution [37]–[39]. The reduction of polymorphism is transient and the levels of polymorphism are expected to recover within roughly Ne generations [40]. In addition to the reduction of the level of polymorphism, recurrent selective sweeps may also generate other signatures such as (i) an overabundance of low-frequency alleles [41],[42], (ii) a greater proportion of high-frequency derived alleles [43],[44], (iii) unusual haplotype structures [45],[46].
A number of these expectations have been used to define signatures of positive selection for genome-wide scans for recent adaptation in humans: i.e., the detection of candidate regions that are likely to be experiencing a selective sweep at present or that have experienced one recently. For example, Nielsen et al. [30] and Kelley et al [47] used the deviation of the allele frequency spectrum from its background characteristics to detect candidate regions that may have experienced a sweep; several other methods have used summaries of haplotype structure and their deviation from the background to detect candidate regions that are undergoing a selective sweep [45],[46].
Genomic scans for positive selection are primarily used to choose candidate regions for future investigation, but their application to the quantification of positive selection or even the establishment of its prevalence is problematic. To quantify the extent of positive selection based on the deviations of these signatures from the background requires a prior expectation about the likelihood of observing them under neutrality. These expectations, however, may be sensitive to the effects of non-equilibrium demography [44], [48]–[50]. As a result, it is difficult to generate robust a priori expectations for these statistics under neutrality. Therefore, scans for positive selection do not, by themselves, provide reliable quantification of the extent of positive selection in humans or establish that positive selection is prevalent in humans.
To evaluate whether selective sweeps are common in the human genome, we require signatures that are unlikely to be generated by demography alone. The effects of recurrent selective sweeps (RSS) should be stronger in the regions of lower recombination and in regions of more frequent and selectively consequential adaptation. In Drosophila, for example, the level of neutral polymorphism is positively correlated with the recombination rate [51]–[53] and negatively with the rate and number of nonsynonymous substitutions in a region [9],[13],[14]. These correlations, which are expected under models of RSS but should not be generated by demography alone, support the notion of high rates of adaptation in these taxa.
Despite several compelling examples of adaptations, clear genome-wide signatures of RSS have been difficult to detect in humans. A relationship of diversity and recombination has been reported, but was attributed primarily to an association between recombination and mutation processes rather than to the effects of selected at linked sites [54]–[57], with the possible exception of telomeric and centromeric regions [58]. In turn, the relationships between levels of polymorphism and functional divergence have not yet been examined.
If the recent MK estimates of the rate of adaptive evolution are correct and approximately 10% of amino acid substitutions are adaptive [21], we should expect to see a substantial number of recent selective sweeps in the polymorphism data. Indeed, ∼7×104 amino acid differences between human and chimpanzee proteins [31] have accumulated over the past ∼14 million years. If 10% of these have been adaptive, then we can estimate that ∼7×103 adaptive amino acid substitutions have taken place over ∼14 million years. Assuming a constant rate of adaptation, this translates into ∼100 adaptive amino acid substitutions that occurred during the past Ne generations (Ne = ∼2×105 years) [59]. Moreover, if regulatory adaptations are common as well, then hundreds of recent selective sweeps should be detectable in the human polymorphism data.
With these considerations in mind, we analyze genomic patterns of nucleotide polymorphism, recombination, functional density and functional divergence in humans using two independent, genome-wide SNP datasets. Consistent with the expectations of positive selection, we detect a positive correlation between levels of neutral polymorphism and recombination rate and a negative correlation between levels of nucleotide polymorphism and both functional density and functional divergence. These correlations remain intact after controlling for a number of possible covariates. The evidence is consistent with positive selection in both regulatory and protein-coding regions. We consider alternative explanations for these findings and argue that, in addition to recurrent selective sweep, only background selection (BS) (loss of neutral variants due to hitchhiking with linked deleterious mutations) can possibly generate most of these patterns. Hitchhiking of neutral polymorphisms with linked selected variants—either due to recurrent positive selection or background selection or possibly both—appears to be a substantial force determining levels of neutral polymorphism in the human genome.
To study the effects of RSS, we separate the genomic sequences into two mutually exclusive sets of sequences: “functional” (genic and regulatory) and “nonfunctional”. Both sets of sequences are taken only from the internal parts of autosomes; specifically, we remove all sequences located within 10 Mbp of a telomere or a centromere. We further remove all sequences that cannot be aligned with the chimpanzee genome [31]. The functional set is composed of several types of sequences (see Material and Methods). First, it contains all the genic regions, specifically those that (i) encode exons or are located within 1 kb of any predicted exon and (ii) are located within 5 kb from the starting and ending position of transcripts of protein-coding genes. Because many functional, noncoding sequences are located far from genes in the human genome [60]–[62], we also take all the sequences that can be aligned between primates and zebrafish; sequences that can be aligned over such large evolutionary distances are very unlikely to be unconstrained [63] (see Materials and Methods). The nonfunctional set contains all other sequences except for the repetitive sequences that are filtered out using RepeatMasker [64]. We remove repetitive regions because both alignment and SNP discovery are more problematic in such regions [65]. Hereafter, we will refer to the sequences in the primarily nonfunctional set (totaling ∼1,080 Mbp) as “neutral” sequences for brevity.
We use two SNP datasets: (i) ∼1.2 million Perlegen [66] “A” SNPs discovered using Perlegen chip technology [67] in a panel of 71 individuals of mixed ancestry [68] and (ii) ∼2.0 million SNPs discovered in the diploid sequence of James Watson [69] (see Materials and Methods). In the remainder of the paper, we show the results derived from the analysis of the Perlegen dataset. The results derived from the analysis of the Watson SNPs are shown in the Supplementary Materials. All of the conclusions in the paper are supported by the analysis of either dataset.
We measure the level of neutral nucleotide variation in a genomic window using the number of SNPs within the neutral regions divided by the total number of neutral sites (θneu) in a window (see Materials and Methods). This measure is proportional to the conventional Watterson's θ [70]. In the remainder of the paper, all measurements are carried out over 400 kb windows. We have also carried out all of the analyses with two other window sizes, 200 and 600 kb; none of the conclusions change depending on the window size (Table S1, S2 and S3, Figures S7, S8, S9, and S10).
The level of neutral polymorphism (θneu) depends both on the average time to coalescence within a particular genomic region and on the local constraint and mutation rate. For the purposes of detecting signatures of RSS, variation in constraint and mutation rate generates noise. We assess variability in constraint and mutation rate by measuring divergence per neutral site (dneu) within the neutral regions between the human and chimpanzee genomes (see Materials and Methods). We detect a positive correlation between dneu and θneu (Table 1), confirming that, as expected, constraint and/or mutation rate vary across the human genome. We control for the variation in neutral mutation rate either by carrying out partial correlations with dneu or by using a normalized measure of neutral variation, , where #SNPneu stands for the number of SNPs found in the neutral regions and Dneu stands for the number of divergent sites within neutral regions between humans and chimpanzee genomes. Pneu and θneu also correlate significantly with repeat density (RD) and GC content (GC) (Table 1, Table S1). Finally, in the case of the Watson data, we further carry out controls for the depth of sequence coverage (Table S4).
The overall effect of RSS on the regional levels of neutral polymorphism should depend on (i) the regional rate of recombination, (ii) the number of recent sweeps (the rate of RSS), and (iii) the strength of positive natural selection associated with a typical adaptive substitution (the strength of RSS). The levels of neutral polymorphism across the genome should correlate positively with the rate of recombination and negatively with the rate and the strength of RSS.
We take estimates of recombination rate from Myers et al. [71], who used a statistical approach to infer recombination rates from linkage disequilibrium data in humans; these rates have been shown to be highly reliable by comparison to pedigree data [72]. The levels of neutral polymorphism measured by both θneu and Pneu increase with the recombination rate (Figures 1, S1, and S2). The correlation remains when we control for possible confounders such GC content (GC), repeat density (RD), and divergence at neutral sites (dneu) separately (Table 2S) or together (Pearson r (θneu, RR|GC, RD, dneu) = 0.254, Pearson r (Pneu, RR|GC, RD) = 0.209, P<0.001 in both cases).
Under a model of RSS regions experiencing more frequent or stronger selective sweeps should show lower levels of neutral polymorphism. Because positive selection should be more prevalent in regions of greater functional density, RSS is expected to generate a negative correlation between the degree of functional density and the level of neutral polymorphism. We measure functional density in two complementary ways. First, in each 400 kb window, we count the number of protein-coding codons (FDn) as a proxy of protein-coding density. In addition, we count the number of nongenic sites that can be aligned between primates and zebrafish (FDx) as a proxy of the number of conserved noncoding sites (CNRs) (see Materials and Methods for details).
Consistent with the predictions under RSS, there are strongly negative correlations between either measure of functional density (FDn,FDx) and measures of neutral variability (Figures 2, S3, S4, and Tables 2, S3). After controlling for GC content (GC), recombination rate (RR), repeat density (RD), and divergence at putatively neutral sites (dneu) (in the case of θneu) the correlations become substantially weaker but do remain statistically significant (Tables 2, 3S). The correlations between FDn and both θneu and Pneu remain significant after we control for FDx; and similarly, the correlations between FDx and both θneu and Pneu are still significant when we control for FDn (Tables 2, 3S).
The number of differences between humans and chimpanzee genomes at functional regions is likely to be a more direct proxy of the rate of positive selection than the functional density. Consistent with the expectations of RSS, we detect lower levels of θneu (Pneu) in regions of higher Dn (the count of divergent amino acid coding sites) or Dx (the count of divergent sites within conserved noncoding regions) (Tables 3, S3, Figures 3, S5, S6). These correlations remain significant when we control for GC content (GC), recombination rate (RR), repeat density (RD), and functional density (FDn, FDx, or both) (Tables 3, S4). The correlations between either Dn or Dx and either of the two measures of neutral variation (θneu or Pneu) remain statistically significant when we control/correct for the other measure of functional divergence (i.e. control for Dn in the case of correlations of neutral diversity with Dx and, similarly, control for Dx in the case of correlations of neutral diversity with Dn) (Tables 3, S4).
The genome-wide patterns of nucleotide polymorphism in the human genome contain much information about the historical patterns of mutation, recombination, natural selection and population histories of modern humans. Here we search for traces of recurrent positive selection in the patterns of diversity at (mostly) neutral sites across the human genome. A number of studies argued that positive selection is reasonably common in humans [15],[30],[31],[33], although substantially less common than in Drosophila [10], [11], [16], [23]–[25],[29] and E. coli [26]. A recent study estimated that ∼10% of all amino acid substitutions between humans and chimpanzees have been driven by positive selection [21]. If true, then signatures of hundreds of recent selective sweeps should still be detectable in the pattern of neutral variation in the human genome.
Because recurrent adaptive substitutions leave local (on the order of 0.1 s/ρ) and transient (on the order of Ne generations) dips in neutral polymorphism, persistent adaptation should lead to lower levels of neutral polymorphism in regions of lower recombination and regions where selective sweeps are more frequent and/or stronger on average. Here we have confirmed these predictions by showing that levels of SNP density are lower in the regions of lower recombination and in the regions of higher functional density and functional divergence.
In addition to RSS, a number of other evolutionary forces can generate heterogeneous patterns of polymorphism: (i) variation in mutation rates and selective constraint, (ii) demographic events such as population structure, bottlenecks, and fast recent population growth, and (iii) hitchhiking of neutral variants with recurrent deleterious mutations (background selection (BS)). In addition, uneven ascertainment of SNPs across the genome could generate spurious variability in SNP density. Below we discuss the evidence in relation to these alternative possibilities and argue that hitchhiking—due to selective sweeps or background selection—needs to be invoked to explain the detected patterns.
All SNP datasets suffer from ascertainment biases during the SNP discovery phase that can systematically under- or overestimate numbers of SNPs in particular genomic regions or at particular types of sites. We address this concern by using two very different SNP datasets that are likely to have different ascertainment biases: (i) the high quality (type A) SNPs from the Perlegen dataset [66] and (ii) SNPs discovered in the sequenced diploid genome of James Watson [69]. The type A SNPs were discovered using Perlegen oligo hybridization chip technology in a panel of 71 individuals of mixed ancestry [66]. This set is biased against SNPs located in repetitive regions, given that it is difficult to design uniquely hybridizing oligonucleotides in such repetitive regions [66]. The diploid genome of James Watson was sequenced using the 454 technology and does not suffer from the same technological problems as the Perlegen oligonucleotide chip hybridization technology.
We obtain very similar results using both datasets, which argues that it is unlikely that specific ascertainment biases are responsible for the observed patterns. In addition, we also used the density of the repeats, GC content and functional density as variables in our statistical analyses and showed that all of the signatures of genetic hitchhiking in our data are robust to statistical controls for these variables. The depth of coverage in the Watson sequencing data also does not noticeably affect any of the detected correlations (Table S4).
The demographic history of human populations in general, and specifically of the populations that have been used for SNP discovery and SNP typing in the Perlegen data, is very complex. Bottlenecks, quick population growth and complex patterns of admixture (for example in the African–American population) are expected to perturb levels of neutral polymorphism across the genome. Collectively, we will denote these forces as “demography”.
The effects of recent demography undoubtedly generate much variation in neutral polymorphism; however, the correlations that we observe are likely to be weakened and unlikely to be generated by the demographic processes alone. For instance, the lower levels of neutral polymorphism in the regions that have large numbers of the protein-coding (Dn) and functional noncoding (Dx) differences are hard to explain by demography; demographic events cannot easily affect the longer-term rates of functional divergence that have been accumulating for ∼10–14 million years between chimpanzees and humans [73]. On the other hand, it is clear that demography needs to be taken into account in order to use the detected signatures to evaluate the strength of hitchhiking in the human genome.
Some of the variation in levels of polymorphism in the sequences that we use to measure levels of neutral polymorphism could be due to the variability in the rates of mutation and levels of selective constraint. We measure levels of neutral variation in the sequences that are less likely to be under selective constraint: they are noncoding, located far from exons, and cannot be aligned with distantly related species such as zebrafish. Nevertheless, some residual variation in constraint is likely to remain. Indeed, the positive correlation between our measures of the levels of neutral polymorphism (θneu) and divergence (dneu) (Table 1 and 1S) suggests that mutation rates and/or levels of constraint vary systematically in these regions. It is therefore important to control for the variation in the levels of selective constraint and mutation rate; we do so by using the levels of divergence (dneu) as a variable in partial correlation analyses or by using the measure Pneu ().The levels of neutral variation correlate strongly with recombination rate, functional density and functional divergence after controlling for neutral divergence suggesting that these correlations are not due to the variation in mutation rate or constraint
Partial correlations may not remove all of the effects of the variation in mutation rate and constraint, however. The variation in selective constraint among neutral regions should have a stronger effect on the levels of neutral divergence (dneu) than on the levels of neutral polymorphism (θneu) because deleterious mutations have a greater chance of segregating in the population than to become fixed. This implies that if the negative correlation between θneu and levels of functional density were entirely due to the variation in selective constraint (specifically higher remaining constraint in regions of higher functional density), then controlling for divergence (dneu) should make the partial correlation between neutral polymorphism (θneu) and functional density positive. Yet we see the opposite: the correlations between Pneu and functional density and the partial correlation between θneu and functional density with respect to dneu both remain strongly negative. This suggests that the variation in selective constraint is unlikely to generate the correlations between levels of neutral variation and recombination rate, functional density and functional divergence that we see in this study.
On the other hand, variability in mutation rates might contribute to some of the observed patterns. Specifically, the positive correlation between neutral diversity and rates of recombination could be due to the mutagenic effects of recombination. Because rates of recombination at local scales (although not necessarily at the 200–600 kb scales relevant to this study) evolve fast [55], [56], [74]–[77], mutagenic effects of recombination should have more pronounced effects on the levels of polymorphism than on the levels of divergence. If so, controlling for neutral divergence (dneu) may not entirely account for the higher mutation rates produced by recent recombination [55].
Mutagenic effects of recombination are expected to affect levels of polymorphism proportionately to the rate of recombination in the area, whereas hitchhiking (RSS or BS) is expected to affect levels polymorphism in regions of very low recombination much more substantially [78]. We observe a mostly linear effect of recombination on divergence (dneu) suggestive of the mutagenic effect of recombination and further arguing that the regional recombination rates at the level of our analysis (200 to 600 kb) do not evolve as fast as the location of recombination hotspots. In contrast, the effect of recombination on the levels of polymorphism (θneu and Pneu) is curvilinear, with most of the effect limited to the regions of the lowest recombination rates (Figure 1 and S1). Indeed, when we split the data by the median value of recombination rate (RR = 1.040 cM/MBp), the correlation between the levels of neutral divergence (dneu) and recombination rate (RR) for the two halves of the data are of similar strength (r (dneu, RR|RR<1.040) = 0.197 and r (dneu, RR|RR>1.040) = 0.220). However, the correlations between recombination rate and levels of polymorphism (θneu or Pneu) are much stronger in the low recombination regions than in the high recombination regions ((r (θneu, RR|RR<1.040) = 0.249 versus r (θneu, RR|RR>1.040) = 0.045; r (Pneu, RR|RR<1.040) = 0.194 versus r (Pneu, RR|RR>1.040) = −0.0241). These considerations suggest that most of the positive correlation between recombination rates and levels of neutral polymorphism, and especially the reduction at lower recombination rates, is caused by some form of hitchhiking. These results are consistent with the findings of Hellman et al [58] who detected lower levels of polymorphism in the areas of low recombination close to centromeres and telomeres. Note that in our study we explicitly excluded telomeric and centromeric regions (see Materials and Methods), making our findings complementary to those of Hellman et al [58].
Background selection (BS) is the process of hitchhiking of neutral or weakly deleterious polymorphism with linked strongly deleterious polymorphisms [79]–[82]. BS should be more efficacious and lead to lower levels of neutral polymorphism in regions of lower recombination. It is thus quite possible that the positive correlation between neutral polymorphism and recombination rate is due in part to BS. In addition, BS should be stronger in the more constrained genomic regions because such regions should experience higher rates of deleterious mutation (e.g. [58]). Therefore BS is likely to contribute to the negative correlation between levels of neutral polymorphism and functional density as well. Because regions of higher functional density also exhibit higher rates of functional divergence (Tables 1 and S1), BS could contribute to the negative correlation between levels of neutral polymorphism and functional divergence as well.
It is less clear whether BS could generate the negative correlation between the levels of neutral polymorphism and functional divergence after controlling for levels of functional density (Tables 3, S4, Figure S6). Two regions of equal functional density can differ in the strength of BS if they differ in the rate of deleterious mutations in the functional sequences. The higher level of deleterious mutations should lead to stronger BS and therefore lower levels of polymorphism in the linked neutral sequences. At the same time, the higher rate of deleterious mutations is likely to come at the expense of neutral mutations at functional sites and thus should lead to lower levels of protein and regulatory divergence. The reduction of neutral mutation rate in the regions of higher deleterious mutation should lead to a positive correlation between levels of neutral polymorphism and functional divergence after controlling for functional density—the opposite of what is seen. On the other hand, the increase in the rate of fixation of weakly deleterious mutations, also expected in the regions of stronger BS, counteracts the reduction of the rate of functional divergence due to the reduction of neutral mutation rate. The combined effect is difficult to estimate given that we do not have information about the distribution of the rates of mutations of different selective effects along the genome.
There is another pattern we observed that is not naturally predicted by BS. The correlations between functional density (FDn or FDx) and neutral polymorphism weaken very substantially and in some cases become nonsignificant when we control for functional divergence at replacement (Dn) and conserved noncoding sites (Dx) (Table 2). Functional density is likely to be a better of proxy of regional constraint than functional divergence. If BS is indeed the dominant force in the generation of the observed patterns, we might have expected correlations between neutral polymorphisms with FDn and/or FDx to be the most robust.
Without a better understanding of the distribution of selective effects and rates of new mutations, we cannot reject the possibility that BS contributes substantially to all of the detected patterns. It appears, however, that only specific distributions of selective effects of new mutations would generate all of the observed patterns. Whether such a distribution exists in principle and whether the distribution of selective effects of human mutation satisfies these requirements in fact remains to be determined.
The arguments above suggest strongly that some form of hitchhiking, either BS or RSS, needs to be invoked to explain the results presented in this paper. These results also suggest that natural selection at both coding and regulatory sites affect linked neutral polymorphism. This is because the measures of the rate of functional evolution at coding and regulatory sites appear to influence levels of neutral polymorphism independently of each other. Specifically, divergence at coding sites and divergence at regulatory sites correlate negatively with the levels of neutral polymorphism after controlling for each other and for the variation in levels of functional divergence (Table 3, S3). To the extent that this is due to recurrent adaptation selection at both coding and regulatory sites, this would echo results of McDonald-Kreitman analyses of adaptation in Drosophila [12].
Levels of neutral polymorphism correlate stronger with divergence at coding than at non-coding regions, possibly implying that either a higher proportion of nonsynonymous changes are adaptive compared to changes in regulatory regions or that the nonsynonymous adaptations have higher selective coefficients. It is also possible and even likely that Dx is a noisier measure than Dn due to greater difficulties in identification of regulatory regions and the noise in estimating Dx due to misalignments. This pattern may also be due to different rates or distributions of the selective effects of deleterious mutations located in coding and regulatory regions, leading to varying effects of BS on linked neutral polymorphism and functional divergence.
These results can also be used to assess the importance of hitchhiking (either RSS or BS) in affecting patterns of neutral polymorphism. The levels of neutral polymorphism appear to be ∼50% lower in the regions of high Dn or Dx (Figures 3, S5) relative to the regions of zero functional divergence (Dn or Dx = 0). If we assume that this effect is entirely due to hitchhiking, then by using the observed correlation between θneu and Dn, we estimate that the levels of polymorphism genome-wide are reduced by 6% genome-wide (Materials and Methods). This reduction is much more pronounced in the more gene-rich regions. For instance, in the 50% of the most gene-rich regions (regions that have greater than the median density of codons (FDn)), the neutral polymorphism is reduced by 11%, while in the regions that contain 50% of the genes (regions that have greater than the mean density of codons (FDn)), the neutral polymorphism is reduced by 13%.
It is clear that hitchhiking has left a significant imprint on the patterns and levels of neutral variability in the human genome and that the effects of natural selection at linked sites cannot be ignored in the analysis of polymorphism data in humans. The challenge for the future is to use these signatures to answer a number of outstanding questions. What are the selective effects and genomic distributions of adaptive and deleterious changes responsible for RSS and BS? What is the biological nature of these changes? What is the relative importance of RSS and BS? Can we estimate parameters of adaptive evolution in the presence of BS? The availability of whole genome sequences in a large number of humans may provide the necessary data to answer these questions. What is needed now are the models and tools to harness these data to provide a cogent picture of the effects of natural selection on human genome and human evolution.
All analyses have been carried out using two SNP data sets—Perlegen data [66] and Watson data [69]. Perlegen data were downloaded from http://genome.perlegen.com. These data were annotated based on the NCBI build 35 of the human genome sequence. We updated all the genomic positions of the SNPs to match the latest NCBI build 36, according to the rs number of SNPs in the dbSNP build 127. During the processing, 1,361 SNPs were discarded because they could not be uniquely mapped to the human genome. Perlegen data contain three classes of SNPs: (A) array-based genomic resequencing, (B) reliable external SNP collections, and (C) unvalidated, lower confidence sources (see Supplementary text of [66]). We excluded class B and C SNPs and retained 1,235,057 class A SNPs located on autosomes for our analysis. The Watson data were downloaded from http://jimwatsonsequence.cshl.edu/. The genome of James Watson was sequenced at 6× coverage using 454 Life Sciences Technology [69] and matched to the human genome project's published reference sequence [83]. In the Watson DNA sequence, heterozygous sites, in which each site was sequenced multiple times and both forms of the base were found in the diploid genome, were ascertained as SNPs. Homozygous sites of Watson's DNA sequence that have been sequenced multiple times and that differ from the reference sequence of the human genome were also ascertained as SNPs. In total our Watson dataset consisted of 2,020,767 SNPs.
Whole-genome alignments of human (H), chimpanzee (C), and zebrafish (Z) sequences were obtained from the Ensembl compara database [84] through the Ensembl Application Program Interfaces (APIs). We defined the “neutral” genomic regions of the human genome if the regions were: (1) H-C aligned, (2) not H-C-Z aligned, (3) located at least 5 kb away from the starting and ending position of transcripts of protein-coding genes and at least 1 kb away from any exons, (4) located on autosomes at least 10 Mbp away from the boundaries of centromeres and the ends of telomeres, (5) not located in the simple repetitive regions of the human genome. The chromosomal coordinates of exons, transcripts and simple repeats were obtained from the finished and annotated human chromosome sequence from the Ensembl database (build 36).
Neutral divergence was assessed from H-C alignments. The accuracy of estimation of neutral divergence may be influenced by the misaligned sequences. Indeed, we discovered some short (2 kb on average) neutral genomic regions having extremely high levels of divergent sites, which may result from misalignments (data not shown). To minimize the possible influence of misalignments, we only counted “isolated” substitutions that are flanked by two monomorphic positions on each side (i.e. no substitutions or SNPs were mapped to these sites). We denoted the number of isolated substitutions between human and chimpanzee sequences as Dneu, and the number of isolated substitutions per neutral site, dneu. To measure neutral polymorphism, we counted the number of SNPs in neutral regions and denoted the number of SNPs per site as θneu. Alternatively, we measured neutral polymorphism with . Data manipulation was done using Matlab functions based on PGEToolbox [85] and MBEToolbox [86].
We used four metrics as proxies of the rate of adaptive evolution for a given region in the human genome. Functional density was measured using FDn, the number of codons, and FDx, the number of aligned bases in the H-C-Z three-way alignments. Functional divergence was measured using Dn, the number of codons involved in nonsynonymous substitutions between H-C orthologous gene pairs, and Dx, the number of H-C substitutions in H-C-Z alignments that are located in noncoding human genomic regions. For each pair of genes, the amino-acid sequences were extracted and aligned using CLUSTALW [87] with the default parameters. The corresponding nucleotide sequence alignments were derived by substituting the respective coding sequences from the protein sequences. The synonymous substitution rate (Ks) was then estimated by the maximum-likelihood method implemented in the CODEML program of PAML [88]. Insertions and deletions within alignments were discarded. Poorly aligned orthologous pairs, as indicated by Ks>5, were excluded. The codons containing nonsynonymous substitutions were mapped back onto the human genome and positions were recorded. For simplicity we counted the numbers of codons causing amino-acid changes instead of the numbers of single nucleotide replacement substitutions. In calculation of Dx, we excluded “tri-allelic” sites where the bases of H, C and Z all differ from each other.
We used 400 kb (as well as 200 and 600 kb) sliding window with a step of 100 kb to scan along the human genome. For each window, two measures of neutral polymorphism (θneu and Pneu) and four proxies of the rate adaptive evolution (FDn, FDx, Dn, and Dx) were estimated. To reduce noise arising from small sample size, we also discarded the windows with Dneu<500 and the ones with the total amount of “neutral” sequence less than 2 kb. 22,553 400 kb windows have been used for the correlation analysis. Spearman rank correlation or Kendall's correlation coefficients have been calculated in all cases. To visualize correlations between variables, we used scatter plots with regression lines superimposed. We also pooled the data points of neutral polymorphism by the values of the proxy of adaptation under consideration (e.g. Dn). To do this, we ranked all the data points of the neutral polymorphism by the values of the proxy and then pooled them into 100 bins such that each bin had equal size (i.e., 1%) of the data points. We then computed average values of the proxies of adaptation and the average value of neutral polymorphism for each bin, and superimposed them onto the scatter plots.
To control for confounding variables, we calculated Spearman partial correlation coefficients between variables X and Y controlling for Z, using the function partialcorr in the Matlab statistic toolbox. Recombination rate estimated by using the coalescent method of [89] were downloaded from http://hapmap.org/downloads/recombination/. The density of simple repeats was computed as the proportion of bases of simple repeats in the given region. Chromosomal coordinates of simple repeats in the human genome, identified by RepeatMasker [64], were obtained from the UCSC genome browser [90].
We also calculated the partial correlation coefficients between variables X and Y by calculating the correlation between the two sets of residuals formed by two linear models X∼Z and Y∼Z (see also [91]) where Z stands for either one or a series of variables. The distribution of Dn, Dx, FDn, and FDx values is approximately exponential, which is a problem in a least squares linear model framework in controlling for a third variable, Z. The linear model used to regress out Z is sensitive to the highly non-normal distribution of variables, and the residuals will be highly non-normal, making the results difficult to interpret. Therefore, we quantile-normalized values, replacing the original estimates with their theoretical quantiles based on a normal distribution. Then, we fitted linear models, using as the response variable quantile-normalized Dn, Dx, FDn, or FDx, and using as the predictor variables various combinations of recombination rate, GC content, and the density of simple repeats.
Estimation of the effect of hitchhiking on the level of neutral polymorphism was calculated using the regression between θneu on Dn, using the formulawhere q is the reduction of polymorphism due to hitchhiking, i is a window count for the subsets of windows used in the analysis (e.g. FDn>median (FDn)), b is the intercept of the regression of θneu on Dn.
|
10.1371/journal.pgen.1005967 | Competition between Jagged-Notch and Endothelin1 Signaling Selectively Restricts Cartilage Formation in the Zebrafish Upper Face | The intricate shaping of the facial skeleton is essential for function of the vertebrate jaw and middle ear. While much has been learned about the signaling pathways and transcription factors that control facial patterning, the downstream cellular mechanisms dictating skeletal shapes have remained unclear. Here we present genetic evidence in zebrafish that three major signaling pathways − Jagged-Notch, Endothelin1 (Edn1), and Bmp − regulate the pattern of facial cartilage and bone formation by controlling the timing of cartilage differentiation along the dorsoventral axis of the pharyngeal arches. A genomic analysis of purified facial skeletal precursors in mutant and overexpression embryos revealed a core set of differentiation genes that were commonly repressed by Jagged-Notch and induced by Edn1. Further analysis of the pre-cartilage condensation gene barx1, as well as in vivo imaging of cartilage differentiation, revealed that cartilage forms first in regions of high Edn1 and low Jagged-Notch activity. Consistent with a role of Jagged-Notch signaling in restricting cartilage differentiation, loss of Notch pathway components resulted in expanded barx1 expression in the dorsal arches, with mutation of barx1 rescuing some aspects of dorsal skeletal patterning in jag1b mutants. We also identified prrx1a and prrx1b as negative Edn1 and positive Bmp targets that function in parallel to Jagged-Notch signaling to restrict the formation of dorsal barx1+ pre-cartilage condensations. Simultaneous loss of jag1b and prrx1a/b better rescued lower facial defects of edn1 mutants than loss of either pathway alone, showing that combined overactivation of Jagged-Notch and Bmp/Prrx1 pathways contribute to the absence of cartilage differentiation in the edn1 mutant lower face. These findings support a model in which Notch-mediated restriction of cartilage differentiation, particularly in the second pharyngeal arch, helps to establish a distinct skeletal pattern in the upper face.
| The exquisite functions of the vertebrate face require the precise formation of its underlying bones. Remarkably, many of the genes required to shape the facial skeleton are the same from fish to man. In this study, we use the powerful zebrafish system to understand how the skeletal components of the face acquire different shapes during development. To do so, we analyze a series of mutants that disrupt patterning of the facial skeleton, and then assess how the genes affected in these mutants control cell fate in skeletal progenitor cells. From these genetic studies, we found that several pathways converge to control when and where progenitor cells commit to a cartilage fate, thus controlling the size and shape of cartilage templates for the later-arising bones. Our work thus reveals how regulating the timing of when progenitor cells make skeleton helps to shape the bones of the zebrafish face. As mutations in many of the genes studied are implicated in human craniofacial defects, differences in the timing of progenitor cell differentiation may also explain the wonderful diversity of human faces.
| Morphogenesis of the facial skeleton in zebrafish is tightly linked with the early differentiation of pharyngeal arch neural crest-derived cells (NCCs) into cartilage. Shortly after migration into the pharyngeal arches, NCCs form a series of pre-cartilage condensations that prefigure the distinct shapes of the later cartilage-replacement bones. As near-isometric growth of these cartilages during the later larval period largely preserves these initial shapes [1], early patterning, not later growth, is the major determinant of facial skeletal shaping. Identifying the local signals that sculpt and arrange early condensations in specific regions of the developing arches is therefore critical to understanding how the facial skeletal bauplan is established.
Genetic studies in a wide range of vertebrates has revealed that patterning of arch NCCs along the dorsoventral axis is an important early step in regionalization of the face, with ventral (distal) cells generating the lower jaw and hyoid bone, maxillary cells forming the upper jaw, and more posteriorly located dorsal (proximal) cells making the lateral upper face. These dorsoventral domains are established in large part by interactions between dorsal Jagged-Notch, ventral/intermediate Endothelin1 (Edn1), and ventral Bmp signaling. Mutation of Edn1 signaling components and key downstream targets (e.g. Dlx5/6) in mice and zebrafish results in homeotic transformations and/or losses of skeletal elements derived from the ventral and intermediate domains of the arches, such that the lower jaw adopts an ectopic upper jaw morphology [2–12]. Downregulation of Bmp signaling results in a similar loss of ventral arch-derived structures in mice and zebrafish [13–16], whereas loss of the Notch ligand jag1b in zebrafish conversely affects bones and cartilages of the upper/dorsal face, particularly those from the second arch and the dorsal-posterior region of the first arch [17]. These pathways are actively antagonistic: Edn1 and Bmp signaling prevent jag1b expression in ventral/intermediate cells, Notch signaling blocks the expression of Edn1 target genes dorsally (e.g. dlx3b/5a/6a, msxe, nkx3.2) [17], and Jagged-Notch and Edn1 signaling limit Bmp activity to the most ventral arches in part through upregulation of the Bmp antagonist Gremlin2 in the intermediate domain [13, 14]. The end result of these interactions is the establishment of a distinct dorsal domain (excluding the anterior/maxillary region of the first arch, which is not patterned by Notch [17]) and the subdivision of an initial ventral arch region into distinct ventral/lower and intermediate regions [14]. How this dorsoventral patterning is translated into region-specific cartilage shapes has, however, remained unresolved.
Previous microarray studies of dissected arches in mice lacking the key Edn1 target genes Dlx5/6 [18] or overexpressing Bmp4 [16] revealed a number of misregulated ventral- and dorsal-specific genes. However, an overarching logic by which the Edn1 and Bmp pathways impart region-specific skeletal shapes remained elusive, with the role of Notch signaling in this process even less clear. In the present study, we perform genome-wide expression analyses of purified arch NCCs to correlate how gene expression patterns change over time in wild-type zebrafish with how gene expression is affected by reduction or elevation of Edn1 or Jagged-Notch signaling. In so doing, we find a prominent role for Jagged-Notch signaling in repressing, and Edn1 in activating, the expression of a set of genes that are strongly induced as arch progenitors mature and begin to acquire cartilage fates, implying that Notch and Edn1 signaling exert opposite effects on cartilage differentiation within the arches.
Two such downstream effectors identified in our genomic analysis are the pre-cartilage condensation marker barx1 (inhibited by Notch, activated by Edn1) and the early progenitor markers prrx1a and prrx1b (inhibited by Edn1). In mouse and chick, the homologs of prrx1a and prrx1b (Prrx1/PRRX1, previously called Prx1 or mHox) are expressed in uncondensed preskeletogenic mesenchyme [19–22], whereas Barx1/BARX1 is found in cells of nascent pre-cartilage condensations that have not or are just beginning to upregulate Sox9 [23–25]. Studies using a Prrx1 proximal promoter to drive lacZ expression [22, 26] or Cre recombinase [27] revealed that the cells that make up the limb skeleton and associated connective tissues all pass through a Prrx1+ state at some point during their differentiation program. Though no similarly definitive lineage-tracing studies exist for Barx1, corollary evidence suggests that most populations of Barx1+ cells mature into Sox9+ chondrocytes [25, 28–31]. In mammals, PRRX1 is required to repress cartilage differentiation in certain parts of the face: Prrx1 mouse mutants develop a large ectopic cartilage in place of the dermal squamosal bone on the side of the skull (derived from the dorsal first arch) as well as an aberrant sigmoidal process off of a shortened Meckel’s cartilage; these mutants also show chondrification of the stylohyoid ligament between the styloid process and Reichert’s cartilage (second arch), among numerous other craniofacial and limb skeletal defects [21, 22]. By contrast, impaired cartilage development is observed in barx1 mutant zebrafish, particularly in the ventral/lower face [32]. In edn1 mutant zebrafish and mice mutant for the Edn1 receptor (Ednra2), defects in ventral and intermediate facial structures are preceded by a loss and shift of Barx1/barx1 expression, particularly in the second arch [10, 33]. These studies indicate that although cartilage differentiation does not strictly require Barx1, chondrogenesis in the ventral and intermediate arches is most sensitive to its loss. Here we demonstrate that early arch patterning pathways compete to drive (Edn1) or restrict (Jagged-Notch, Bmp) the commitment of NCCs to chondrogenic differentiation, in part through antagonistic regulation of barx1 and prrx1a/b. These region-specific differences in the timing and extent of cartilage formation thus establish the template for the later formation of uniquely shaped bones in the upper and lower face.
In an unbiased approach towards understanding facial skeletal patterning, we first performed a global gene expression analysis of pharyngeal arch NCCs at three time-points in wild-type embryos. To purify arch NCCs, we conducted fluorescence-activated cell sorting (FACS) on dissociated cells doubly positive for sox10:DsRed and fli1a:EGFP transgenes (and single-positive and double-negative cells for comparison) (Fig 1A and 1B). sox10:DsRed labels all NCCs and the ear, and fli1a:EGFP labels arch NCCs, blood vessels, and macrophages. These transgenes uniquely intersect within the arch NCC population, allowing us to selectively enrich for these cells shortly after NCC migration into the arches (20 hours post-fertilization, hpf) and during the initiation of pre-cartilage condensation formation (28 and 36 hpf). cDNA libraries were then constructed for each cell population and subjected to next-generation sequencing. To remove genes with low expression in the arches, we excluded genes with RPKM values ≤ 3 in the wild-type 36 hpf sample. However, a number of genes with known expression in the erthyroid lineage (e.g. hemoglobin genes hbae1/3, hbbe1/3), macrophages (e.g. mfap4 [34]), and the ear (e.g. mvp [35] and oc90 [36]) were found in this filtered list, suggesting some degree of contamination of the GFP/DsRed double-positive population by single-positive fli1a:EGFP or sox10:DsRed cells. We therefore further filtered for genes with expression ratios of 1.5-fold or higher in the double-positive cells relative to both single-positive populations. This left 536, 668, and 741 arch-enriched genes in the 20, 28, and 36 hpf samples, respectively, with the latter group comprising the “total” arch gene list in Fig 1 (also see S1 Table).
In order to understand how Edn1 and Notch signaling control the expression of these arch NCC-enriched genes, we next performed FACS purification and next-generation cDNA sequencing of GFP/DsRed double-positive cells from 36 hpf embryos with gain or loss of each signaling pathway. Specifically, we compared fold-change differences between edn1 mutants and stage-matched controls, jag1b mutants and wild-type siblings, and hsp70I:Gal4; UAS:Edn1 or hsp70I:Gal4; UAS:NICD (Notch1 intracellular domain) versus hsp70I:Gal4 controls (subjected to a 20–24 hpf heat-shock to overactivate Edn1 or Notch signaling) (see Methods). The top 20 genes up- and down-regulated in the Edn1 and Notch mutant and overexpression datasets (prior to filtering for arch NCC-enriched genes) are presented in S2 Table. Known targets of Notch (e.g. jag1b, hey2 and her2/4/15 genes) and Edn1 (e.g. dlx3b/4a/4b/6a and Evf1/2) are highly represented in these lists. All subsequent analyses were performed using the filtered list of 741 genes with arch-enriched expression in the 36 hpf wild-type sample. To identify those genes most strongly regulated by the Edn1 or Notch pathway, we divided the fold-change of the overexpression (OE) sample by the fold-change of the corresponding mutant (mut) sample. Genes considered ‘activated’ by the Edn1 or Notch pathways had an OE-fold-change/mut-fold-change ratio ≥ 1.5 as well as an OE-RPKM/control-RPKM ratio ≥ 1. Genes considered ‘inhibited’ by Edn1 or Notch had an OE-fold-change/mut-fold-change ratio ≤ 0.667 and a mutant-RPKM/control-RPKM ratio ≥ 1. Lastly, we performed one further refinement for the Notch lists by analyzing genes affected by treatment of embryos with the γ-secretase inhibitor dibenzazepine (DBZ), which blocks processing of the Notch receptor into its active intracellular form [37], starting at 24 hpf. After FACS-purification and next-generation sequencing of double-positive cells from 36 hpf embryos, we calculated the fold-change between DBZ-treated and control samples. Eleven of the top 20 genes downregulated in DBZ-treated embryos belong to the Her/Hes/Hey family of Notch targets [38, 39] (7 of which were shared with the jag1b mutant list) (S2 Table), showing that DBZ is primarily affecting Notch signaling in this experiment. However, as γ-secretase inhibitors such as DBZ are also known to affect other signaling pathways [40], we only used the DBZ dataset to further refine the lists generated from the jag1b and NICD analyses. Specifically, we excluded genes from the ‘Notch activated’ list that were not also elevated in NICD versus DBZ (fold-change ratio ≥ 1.25) and from the ‘Notch inhibited’ list those not also decreased in NICD versus DBZ (fold-change ratio ≤ 0.8). These filtered gene lists (Fig 1E, S3–S6 Tables) were then used for the global analyses described below.
Consistent with our previous data that the Edn1 and Jagged-Notch signaling pathways antagonize one another during facial development [17], we observed a disproportionately high number of genes oppositely regulated by these pathways. Of the 67 ‘Notch-inhibited’ genes and 107 ‘Edn1-activated’ genes, 22 were in common (Fig 1E). Conversely, 29 genes were in common between the 93 ‘Notch-activated’ and 137 ‘Edn1-inhibited’ genes (Fig 1E). These commonly regulated genes include many known positive Edn1 targets (e.g. dlx3b, dlx4a, dlx4b, epha4b, Evf1/2, msxe, and notch2) [5, 17, 18, 33, 41, 42] and negative Edn1 targets (e.g. jag1b and pou3f3a/b) [17, 18]. Smaller groups of genes were co-regulated by Notch and Edn1 in the same direction (positive, n = 9; negative, n = 6; S3–S6 Tables), as we previously observed for the BMP antagonist grem2 [14].
We next examined whether genes activated or inhibited by Notch or Edn1 presented any common temporal signatures during arch development in wild-type embryos. To do so, we first determined the fold changes in wild-type RPKM values for the 741 total arch genes from 20 to 28 hpf and from 28 to 36 hpf (Fig 1C, 1D and 1F; S1 Table). Total arch genes increased by a median of 1.46-fold between 20 and 28 hpf, and 1.18-fold between 28 and 36 hpf. In contrast, we found that the subset of genes that we had annotated as ‘Notch inhibited’ increased 2.62-fold from 20 to 28 hpf in wild types, with many of these upregulated more than 10-fold (p < 0.001). These strongly upregulated genes presented a range of expression levels at 20 hpf, showing that the stronger upregulation of ‘Notch inhibited’ genes is likely not an artifact of these having very low initial expression levels. ‘Edn1 inhibited’ genes also displayed a modest but significantly higher upregulation than total arch genes (median 1.95-fold increase; p < 0.001), though no significant differences were observed for ‘Notch activated’ genes (median = 1.51). Although the ‘Edn1 activated’ genes were not more highly upregulated than total arch genes (median = 1.68), the subset of ‘Edn1 activated’ genes in common with ‘Notch inhibited’ genes were the most strongly induced (median 4.01 fold increase; p < 0.001). Of these 22 common genes, 8 were induced more than 10-fold between 20 and 28 hpf, out of only 48 total >10-fold-upregulated arch genes. At later stages (28–36 hpf; Fig 1D and 1F), only ‘Edn1 inhibited’ genes (median 1.52-fold increase) showed a small but significant difference (p < 0.001) relative to all arch genes (median 1.18-fold increase). In summary, these data show that genes commonly inhibited by Notch and activated by Edn1 are some of the most highly induced during early arch differentiation, consistent with a global role for Notch repression in limiting the differentiation of NCCs in the dorsal arches.
Given the genome-wide role of dorsal Jagged-Notch signaling in repressing strongly induced genes during arch maturation, we investigated whether this might reflect a delay in cartilage differentiation in the dorsal domain versus the rest of the arches. From our genomic analysis, we observed that barx1, which marks early pre-cartilage condensations [23, 29], was negatively regulated by Notch signaling (~5-fold lower in NICD versus jag1b mutant; S3 Table) and 12.5-fold upregulated between 20 and 28 hpf in wild-type NCCs (S1 Table). By examining a time-course of barx1 expression (also see [29, 32]), we find barx1 to be confined to the intermediate/ventral portions of the first and second arches at 26–28 hpf, with maxillary first arch and dorsal second arch expression not initiating until 30–32 hpf (Fig 2B). To determine whether this delay reflects later cartilage differentiation in the dorsal second arch, we made time-lapse recordings of fish expressing sox10:DsRed (which shows biphasic expression–first in all NCCs and later in differentiating chondrocytes) along with the arch NCC transgene fli1a:EGFP or the chondrocyte transgene col2a1aBAC:GFP (Fig 2C and 2D and S1 and S2 Movies). In both cases, chondrocyte differentiation was first evident in the palatoquadrate cartilage (Pq, primarily an intermediate first arch element with a small amount of dorsal contribution at its posterior end), the symplectic cartilage (Sy, intermediate second arch), at either end of the ceratohyal cartilage (Ch, ventral-intermediate second arch), and the proximal portion of Meckel’s cartilage (M, ventral-intermediate first arch). Chondrocyte transgene expression then spread into the center of the Ch and more ventral portions of the M cartilage. The last elements to differentiate were the hyomandibular cartilage (Hm, dorsal second arch) and the pterygoid process cartilage (Ptp, maxillary) (schematized in Fig 2A and 2E). We also observed that sox9a, an early marker of pre-chondrocytes [43–46] that is positively regulated by Edn1 signaling (S6 Table), was expressed only in ventral-intermediate arch NCCs at 36 hpf, with expression spreading to dorsal arch NCCs by 48 hpf (Fig 3A and 3K). Our findings point to cartilage differentiation occurring first in discrete zones, primarily within the intermediate arches, then spreading to other ventral regions and lastly to dorsal regions, consistent with previous studies based on Alcian Blue staining of sulfated proteoglycans typical of cartilage [47].
The earlier chondrogenic differentiation in intermediate/ventral arch cells relative to dorsal cells led us to hypothesize that antagonism between dorsal Jagged-Notch and ventral Edn1 signaling may serve to establish barx1+ condensations earlier and/or more extensively in the lower face. As reported previously [10], we find that barx1 expression is lost from the ventral second but not first arch of edn1 mutants at 36 hpf (Fig 3E). Conversely, elevation of Edn1 signaling (via 20–24 hpf heat-shock induction of hsp70I:Gal4; UAS:Edn1 fish) resulted in an expansion of barx1 expression throughout the arches (Fig 3F). In contrast, we find Jagged-Notch signaling to be required to restrict dorsal barx1 expression, consistent with our RNAseq data (S3 Table) and the mutually exclusive expression of barx1 and jag1b at 36 and 48 hpf (Fig 3B and 3C). In jag1b mutants, barx1 expression expands into the dorsal-posterior regions of both the first and second arch (Fig 3G; similar to Edn1 overexpression (Fig 3F)), domains that correlate precisely with jag1b expression at this stage (Fig 3B). Conversely, jag1b expression is unaltered in barx1 mutants (S1 Fig), indicating that Jagged-Notch signaling functions largely upstream of barx1 and not vice versa. This ectopic dorsal barx1 expression was also observed in notch2; notch3 double mutants (Fig 3H), which display similar facial cartilage defects to jag1b mutants (consistent with notch2 and notch3, but not notch1a or notch1b, being expressed in arch NCCs; S2 Fig). Reciprocally, forced activation of Notch signaling in heat-shock-treated hsp70I:Gal4; UAS:NICD fish eliminated nearly all barx1 expression in the arches (Fig 3I). Finally, we find that the positive effect of Edn1 on ventral second arch barx1 expression can be explained at least in part by the previously reported role of Edn1 in blocking jag1b expression [17], as mutation of jag1b partially restored ventral second arch barx1 expression in edn1 mutants (Fig 3J).
We next examined whether the ectopic dorsal expression of barx1 persisted in jag1b mutants, as well as the consequences of this for cartilage differentiation. At 36 hpf, sox9a expression marks the nascent cartilages in the ventral-intermediate arches that are the first to differentiate, with barx1 expression in a partially overlapping set of cells nearer to the poles of each arch (Fig 3A). By 48 hpf, sox9a expression has spread into the nascent dorsal cartilages yet remains only minimally overlapping with barx1 (Fig 3K, also see S3 Fig). These results are consistent with previous literature showing that Barx1 is expressed in nascent pre-cartilage condensations that have not or are just beginning to upregulate Sox9 [23–25]. In jag1b mutants at 48 hpf, we observe an expansion of barx1 but not sox9a expression in the dorsal first and second arch (Fig 3L), suggesting that a subset of dorsal arch NCCs may be trapped in an early barx1+ condensation state in the absence of Jagged-Notch signaling. This failure of ectopic dorsal barx1+ cells to transition to a more mature sox9a+ state may help explain why the dorsal cartilages of jag1b mutants are truncated rather than expanded (Fig 4B).
Because ectopic dorsal barx1 expression correlated with dorsal cartilage defects in jag1b and notch2; notch3 mutants, we next investigated whether this reflected a common early requirement for Jagged-Notch signaling for both processes. To temporally inhibit Notch signaling, we treated embryos at different stages with 10 μM DBZ, and evaluated the effects on barx1 expression and skeletal patterning. Although DBZ may also affect other signaling pathways [40], our RNAseq analysis showed that the majority of the most strongly downregulated genes were canonical Notch targets (S2 Table). This analysis focused on the first arch phenotypes, which have proved the most penetrant and consistent across all of our Notch loss-of-function models. Compared with DMSO-treated controls, treatment of embryos with DBZ starting at 8 hpf resulted in a highly penetrant expansion of barx1 expression into the posterior dorsal first arch (12/12), as well as dorsal cartilage defects similar to jag1b mutants (19/20 with Pq malformations; Fig 3M and 3N). DBZ treatment starting at 24 hpf resulted in a weaker and less penetrant barx1 expansion (10/13 embryos with ectopic first arch barx1) and milder dorsal cartilage defects (41/44 with moderate Pq malformations). In contrast, treatment at 28 hpf only mildly affected barx1 expression in 7/13 embryos, with only 16/39 embryos displaying weak dorsal cartilage malformations (Fig 3M and 3N). Treatments initiated at 32 hpf did not affect barx1 expression or skeletal patterning. Inhibition of Notch signaling at these stages also had other effects on embryo development, including cardiac edema, which likely contributed to the general reductions in cartilage size. In summary, we observe a tight correlation between barx1 expression changes and subsequent malformations of dorsal cartilages in Notch-deficient embryos, with the requirement for Notch inhibition by approximately 24 hpf being consistent with the predicted global effects of Notch in repressing arch gene induction between 20–28 hpf (Fig 1F).
We next investigated the extent to which the ectopic dorsal expression of barx1 in Notch pathway mutants contributes to the dorsal cartilage malformations. In particular, jag1b mutants display several characteristic changes in cartilages of the upper face, including truncation of the posterior end of Pq (i.e. the portion deriving from dorsal first arch NCCs; Fig 2A) and a variable reduction of the anterior part of Hm (Fig 4A and 4B). jag1b mutants also display a highly penetrant posterior shift of Hm such that it sits closer to the ventral Ch cartilage [17]. In barx1 mutants, the dorsal cartilages are largely unaffected, with there instead being conspicuous reductions of the ventral M and Ch cartilages (Fig 4C) [32]. In jag1b; barx1 mutants, we observed an incompletely penetrant rescue of posterior Pq (truncation in 11/27 double mutants versus 16/16 jag1b mutants, p < 0.0001) and the position of Hm (posterior shift in 13/27 double mutants versus 14/16 jag1b mutants, p < 0.0001) (Fig 4D and 4E). However, ventral M and Ch defects were not restored, and the anterior Hm was more prominently diminished (loss in 18/27 double mutants versus 5/16 jag1b mutants, p < 0.0001). Of note, the two regions of skeletal rescue (posterior Pq and posterior Hm) correlate precisely with the earlier ectopic expression of barx1 in dorsal-posterior first and second arch domains of jag1b mutants (Fig 3G), suggesting that the ectopic barx1 may account in part for these phenotypes. On the other hand, the incompletely penetrant rescue of these elements, in addition to exacerbated phenotypes in other regions (e.g. anterior Hm), indicates the presence of other causative changes in jag1b mutants beyond barx1 misexpression.
The finding that posterior-dorsal cells ectopically express barx1 but fail to turn on sox9a in jag1b mutants, as well as the fact that cartilage defects were only modestly rescued in jag1b; barx1 double mutants, suggest that Notch signaling has additional roles in dorsal cartilage development. In order to better understand the reductions of dorsal cartilage in jag1b mutants, we used photoconversion of the kikGR protein to follow the fate of dorsal second arch NCCs in wild types versus mutants (Fig 5). When wild-type cells were converted at 36 hpf and then re-imaged at 6 days post fertilization (dpf), we found that anterior dorsal second arch NCCs contributed to the anterior portion of the Hm cartilage, central dorsal second arch NCCs to the posterior portion of Hm, and posterior dorsal second arch NCCs to a small amount of Hm and the opercle bone to which it attaches (Fig 5A–5C). In jag1b mutants, all three domains contributed to similar portions of the malformed Hm cartilage as in wild types (Fig 5D–5F), indicating no major shift in the fate map of skeletal precursors in mutants. However, whereas cells from all three regions spread along the dorsoventral axis in wild types, cells from comparable domains in jag1b mutants gave rise to much smaller domains of cartilage (Fig 5G). These results suggest that Jagged-Notch signaling is also required for the expansion of the dorsal second arch NCCs that generate cartilage in the upper face.
While loss of Jagged-Notch signaling can rescue barx1 expression and ventral cartilage development in edn1 mutants (Fig 3J; [17]), the partial and largely second-arch nature of this rescue implies the presence of other important pathways downstream of Edn1. Our expression analysis of sorted arch NCCs identified two genes implicated in early skeletogenic mesenchyme identity, prrx1a and prrx1b, which, like jag1b, were upregulated in edn1 mutants and downregulated in Edn1-overexpressing embryos (S5 Table). Loss of the homologous Prrx1 gene in mice results in ectopic dorsal facial cartilage [21, 48], implying that Prrx1 genes may also restrict cartilage formation in the upper face. We thus reasoned that Edn1-mediated repression of prrx1a and prrx1b could help to explain the observed acceleration of cartilage differentiation in the intermediate domain (Fig 2B–2E). Consistently, we observed that expression of prrx1a and prrx1b was largely excluded from NCCs in the intermediate domain, instead being confined to the dorsal-most and ventral-most poles of the first two arches in 36 hpf wild types (Fig 6A and 6B). However, as predicted by our RNAseq analysis, prrx1a and prrx1b expression was upregulated along the ventral border and expanded into the intermediate arches of edn1 mutants, and lost in Edn1-overexpressing embryos (Fig 6C–6F), in accord with the elevated ventral Prrx1 expression observed in Dlx5/6 mutant mice [6]. Conversely, overactivation of Bmp4 signaling (via 20–24 hpf heat-shock induction of hsp70I:Gal4; UAS:Bmp4 fish) upregulated prrx1a and prrx1b expression throughout the arches (Fig 6G–6J), in accord with previous findings that Bmp signaling promotes genes associated with progenitor status and self-renewal in arch NCCs [16]. Positive regulation by the ventral Bmp signal, combined with negative regulation by intermediate Edn1, could help explain the restriction of prrx1a/b expression to the ventral pole of the arches.
We next examined whether the Bmp4 induction of prrx1a/b is mediated by Hand2, a strong Bmp target that is specifically expressed in NCCs at the ventral border of the arches in both mice and fish [5, 49], domains that closely overlap with ventral prrx1a/b expression. While Hand2/hand2 expression requires positive input from both the Edn1/Dlx and Bmp pathways [5, 6, 13, 49, 50], overexpression of Bmp4 –but not Edn1 –induces its widespread ectopic expression [13, 14], similar to the patterns observed here for prrx1a/b (Fig 6G–6J). However, consistent with previous results in mice [51], prrx1a and prrx1b were expressed largely normally in hand2 mutants, with a limited expansion of prrx1a in the ventral domain (S4 Fig). Thus, Bmp signaling appears to positively regulate prrx1a/b expression largely independently of Hand2 function.
In order to interrogate Prrx1 function in zebrafish, we used TALENs to generate prrx1ael558 and prrx1bel491 mutant alleles resulting in early truncation of the Prrx1a and Prrx1b proteins upstream of the conserved DNA-binding homeobox domains (S5 Fig). Whereas prrx1a and prrx1b single mutants did not show craniofacial defects during larval stages (consistent with their identical expression patterns), double homozygous mutants exhibited highly penetrant abnormalities affecting dorsal skeletal elements of the first two arches (Fig 7A–7F). Identical dorsal skeletal phenotypes were seen in double mutants carrying prrx1ab1246 and prrx1bb1247 alleles independently generated by CRISPR-mediated mutagenesis (S5 Fig). In the first arch of double mutant larvae, ectopic cartilage develops along the dorsomedial surface of the Pq cartilage in place of the dermal entopterygoid bone. This extra cartilage is occasionally fused with the trabecular cartilages of the neurocranium. In approximately 40% of double mutant embryos, Pq also extended dorsal-posteriorly to fuse with the otic capsule. In the second arch, the top of the Hm cartilage is malformed, with two highly penetrant cartilaginous fusions to the anterior and middle parts of the otic capsule. The foramen of the Hm, a channel for the VIIth cranial nerve and the anterior lateral line nerve [52], is absent, and the opercle bone is reduced. Despite the expression of prrx1a/b at both the dorsal and ventral poles of the arches, double mutants had no detectable defects in ventral cartilages.
Consistent with the ectopic dorsal cartilage, we also found that double mutants displayed ectopic barx1 expression at earlier stages (36 hpf) in dorsal arch regions that generate the parts of Pq, Hm, and otic cartilages affected in mutants (Fig 7G and 7H). This upregulation of barx1 in double mutants is consistent with the near mutually exclusive expression of prrx1a/b and barx1 in 36 hpf wild-type embryos (Fig 6G, 6H and 6L). In contrast to jag1b mutants (Fig 3L), these ectopic barx1 expression domains largely disappeared by 48 hpf (Fig 7I and 7J), perhaps accounting for the ectopic formation of cartilage in Prrx1 but not Notch pathway mutants. Interestingly, despite hand2 being expressed in a similar domain to prrx1a/b in the ventral arches, barx1 has been reported to be lost in hand2 mutants [32], opposite to the barx1 expansion we observe in prrx1a/b mutants. However, hand2 expression was unaffected in prrx1a; prrx1b mutants (S4 Fig), similar to previous observations in Prrx1-/- mice [53], suggesting that Prrx1 and Hand2 act antagonistically and independently to regulate barx1 expression and chondrogenesis in the ventral second arch.
Despite both prrx1a/b and jag1b expression being mutually exclusive to barx1, we found only limited overlap between these genes (Fig 6K and 6L). We therefore hypothesized that these pathways function independently to limit barx1 expression and cartilage formation in distinct domains of the arches. Consistently, we observed no defect in prrx1a or prrx1b expression in 36 hpf jag1b mutants, although forced activation of Notch signaling expanded prrx1a and prrx1b expression ventrally and decreased it dorsally (Fig 8A–8F). Loss of jag1b also partially rescued the ventral expansion of prrx1b observed in edn1 mutants, especially in the second arch (Fig 8G–8I). These findings indicate that, although jag1b is not required for prrx1a/b expression, high levels of Notch signaling (either artificially or by loss of Edn1) can induce prrx1b expression ventrally. Reciprocally, a subset of prrx1a; prrx1b mutants showed a modest reduction of jag1b expression limited to the dorsal posterior second arch (Fig 8J and 8K). To further clarify the genetic interaction between these genes, we analyzed jag1b; prrx1a; prrx1b triple mutants (Fig 8L–8O). In 9/13 triple mutant sides examined, we observed the ectopic posterior extension and fusion of the Pq cartilage to the ear (and not the Pq truncations seen in Notch pathway mutants), indicating that prrx1a/b are largely epistatic to jag1b with respect to the ectopic Pq phenotype. However, in addition to this extra cartilage, skeletons of the triple mutants (but not prrx1a; prrx1b double mutants) showed irregular gaps within the body of Pq, reminiscent of abnormalities seen in Notch pathway mutants. These observations support the triple mutant phenotype being largely additive, in line with Prrx1a/b and Jagged-Notch signaling having distinct roles in regulating condensation and cartilage formation in the upper face.
Because ventral prrx1a/b expression increases in the ventral arches of edn1 mutants (Fig 6C and 6D), we speculated that increased repression of cartilage differentiation by Prrx1 proteins might contribute to the ventral skeletal losses seen in edn1 mutants. Indeed, we found that homozygous loss of both prrx1a and prrx1b resulted in a modest rescue of ventral cartilage formation in edn1 mutants, particularly in the second arch (Fig 9B and 9C), as well as rescue of ventral barx1 in the second arch and dlx5a expression in both the first and second arch (Fig 9G–9J). The partial rescue of ventral cartilage in prrx1a; prrx1b; edn1 mutants, as well as the earlier recovery of barx1 and dlx5a expression, are qualitatively similar to the phenotypes seen in jag1b; edn1 mutants (Fig 9D and [17]). By contrast, there was no rescue of ventral hand2 expression, consistent with our finding that Prrx1a/b do not regulate hand2 (S4 Fig). As Hand2 normally restricts Dlx expression into the ventral-most arches [54, 55], the lack of hand2 recovery in the triple mutants may explain the ectopic ventral expansion of dlx5a in the prrx1a; prrx1b; edn1 mutants.
In jag1b; edn1 mutants, the partial recovery of ventral barx1 expression correlated with zones where prrx1b expression was reduced to control levels (Fig 8I). We therefore asked whether the remaining areas of elevated Prrx1 expression in jag1b; edn1 mutants might account for the incomplete rescue. Consistently, we found that progressive reduction of prrx1a/b gene dosage in jag1b; edn1 mutants resulted in a progressively better rescue of ventral cartilages, with 5/6 quadruple homozygous prrx1a; prrx1b; jag1b; edn1 mutants showing a prominent rescue of the ventral second arch-derived Ch cartilage and improved elongation of the first arch-derived M cartilage (Fig 9E and 9F). However, even in these quadruple mutants, the ‘rescued’ ventral cartilages are still smaller than in wild types, and the dorsal skeletal phenotypes associated with jag1b and prrx1a; prrx1b mutants are still present. These findings reveal important parallel contributions of ectopic Prrx1 and Jagged-Notch activity to the ventral craniofacial defects of edn1 mutants, yet indicate that Edn1 has additional functions beyond inhibiting Prrx1 and Notch activity.
RNAseq analyses of facial NCCs confirmed our previous findings that Notch acts oppositely to Edn1 during pharyngeal arch development [17]. At a mechanistic level, this global analysis revealed that a major function of Notch signaling is to repress the expression of some of the most strongly upregulated genes in early arch development. These include the homologs of a number of genes implicated in mesenchymal condensation, chondrogenesis, and general skeletogenesis in mammals: e.g. barx1 [23, 31, 32], ctgfb [56], col6a1 and col6a6 [57, 58], and tbx22 [59, 60]. As we only profiled global gene expression patterns in mutants and overexpression embryos at 36 hpf, we cannot conclude whether Notch represses these highly-induced genes only at this later stage, or whether it also restrains their initial upregulation.
The concept of Notch limiting differentiation is becoming a common theme in many developmental and regeneration contexts. For example, sustained Notch signaling in preskeletogenic mesenchyme in vivo or mesenchymal progenitors in vitro severely abrogates cartilage formation, with cells inappropriately maintained in a precursor state [61–64]. Likewise, Notch signaling promotes regeneration of the caudal fin of zebrafish by maintaining the blastema in a proliferative, undifferentiated state [65, 66]. Though Notch can also promote differentiation in other contexts (e.g. stimulating maturation and hypertrophy in committed chondrocytes [reviewed by [67]]), our findings are consistent with the large body of literature describing roles for Notch in resisting differentiation of progenitor cell populations, in this case specifically in the dorsal arches.
Our genomic analysis also identified two Prrx1 homologs (prrx1a and prrx1b) as negative targets of Edn1 that function in parallel to Jagged-Notch signaling to restrain cartilage differentiation, yet these pathways appear to do so in different ways (Fig 10). jag1b and prrx1a/b are expressed in largely non-overlapping domains and are generally not required for the other’s expression. The skeletal phenotypes of Notch and Prrx1 mutants also differ in critical ways. Mutants in both pathways develop ectopic barx1+ condensations and malformed cartilages in the dorsal arches, but only prrx1a; prrx1b mutants form ectopic dorsal cartilage. One potential explanation is that ectopic barx1 expression persists in dorsal NCCs at later stages in jag1b but not prrx1a; prrx1b mutants. Perhaps, Jagged-Notch signaling is also required for ectopic barx1+ cells to progress to a sox9a+ chondrogenic state. Further, our fate-mapping studies show that dorsal arch NCCs expand less in jag1b mutants compared to wild types, which could be due to persistent barx1 expression restricting the proliferation of chondrogenic cells. However, loss of barx1 improved only a subset of skeletal defects in jag1b mutants, suggesting that skeletal changes in Notch-deficient embryos result from more than just ectopic barx1 expression. In prrx1a; prrx1b mutants, the release of dorsal cells from a transient barx1+ state may instead allow these cells to reach a critical threshold for making ectopic cartilage (as proposed for the Prrx1 mouse mutant [68]).
The finding that prrx1a; prrx1b double mutants presented skeletal defects only in dorsal elements was somewhat unexpected, given the expression of prrx1a and prrx1b in both dorsal and ventral arch regions. While homozygous or dominant-negative mutations in PRRX1 have been associated with loss of the lower jaw in humans [69–73], Prrx1 mutant mice are similar to zebrafish mutants in displaying ectopic dorsal cartilage, but dissimilar in showing minor abnormalities of the lower jaw [21]; Prrx1; Prrx2 double mutants display much more pronounced jaw reductions [22, 48, 53]. As zebrafish lack a Prrx2 homolog [74, 75], the lack of lower jaw defects in prrx1a; prrx1b double mutants could reflect redundancy with other pathways, or, alternatively, the evolution of different requirements for Prrx1 genes between fish and mammals. At a molecular level, the ectopic barx1 expression we observe in the dorsal arches of prrx1a; prrx1b mutant fish is reminiscent of the medial expansion of Barx1 seen in the ventral first arch of Prrx1; Prrx2 mutant mice [53].
While our data implicate Prrx1 genes and Jagged-Notch as two important negative targets of Edn1 in the ventral arches of zebrafish, the fact that edn1 mutant phenotypes are only partially rescued by the combined loss of Prrx1a/b and Jag1b suggests other yet to be identified key targets of Edn1. Indeed, our RNAseq analysis revealed two different classes of genes activated by Edn1: (1) those that are highly upregulated during early arch development and also inhibited by Notch (including many well-known Edn1 targets such as dlx3b/4a/4b/6a, hand2, epha4b, Evf1/2, and msxe) and (2) those that are Notch-independent and only modestly upregulated during early arch development. Given that jag1b itself is negatively regulated by Edn1 signaling [17], many of the genes on the first list may in fact be Notch targets that are only indirectly stimulated by Edn1. Functional interrogation of these two classes of target genes should help uncover additional functions of Edn1 in arch development.
Our findings in zebrafish also support a greater role for the Jagged-Notch and Prrx1 pathways in patterning the second arch and posterior half of the first arch compared with the anterior portion of the first arch, which generates the bulk of the lower and upper jaw skeleton (Fig 10). For example, jag1b is expressed in only a limited posterior dorsal domain of the first arch and not in the maxillary or mandibular prominences [17], and first arch-derived skeletal structures are less affected than second arch-derived structures in jag1b and prrx1a; prrx1b mutants. barx1 expression is also primarily lost in ventral NCCs of the second but not first arch in Edn1 pathway mutants [10, 33], consistent with the more pronounced upregulation of jag1b and prrx1a/b in this domain. This second arch bias is also reflected by greater rescue of second versus first arch ventral cartilages upon loss of Prrx1 and Notch signaling in edn1 mutants. Given the very different arrangements of cartilage and bone in the second versus first arch, it is not surprising that programs that restrict cartilage formation have distinct roles in each arch. In the future, it will be interesting to explore how Hoxa2 and Hoxb2, which confer second arch identity [76–79], impact the Notch- and Prrx1-based cartilage restriction programs we have identified.
We also note that other major signaling pathways, such as Bmp, Fgf, Tgfβ, Shh, and Wnt, also influence the spatiotemporal patterns of differentiation within the arches–including control of prrx1a/b and barx1 expression [13, 14, 25, 29, 80–86]. For example, our work indicates that Bmp signaling likely helps to establish prrx1a/b expression at the ventral poles of the zebrafish arches. Whether Bmp signaling regulates Prrx1 genes in other vertebrates remains unclear, as previous studies did not detect changes in Prrx1 expression in conditional Bmp4 deletion mice [15] or chicken mandibular explants exposed to exogenous Bmp ligands and antagonists [80]. Future work will need to integrate these other key patterning programs into the model to more fully explain how the timing and extent of chondrogenesis is precisely controlled in the developing face.
Heterochrony in skeletal differentiation is an important mechanism contributing to the evolution of morphological differences between species [reviewed by [87], also see [88, 89] and references therein]. This concept of variation in developmental timing of homologous structures between species has been proposed to explain, for example, differences in beak length and morphology, as well as the shape of Meckel’s cartilage between quail and duck [90–92]. Our work supports the idea that differential developmental timing can also be a critical driving force for varying skeletal structure within an individual. In the arches of zebrafish, chondrocyte differentiation invariably occurs first in intermediate/ventral before dorsal regions [47]. We have found that these events are presaged by an earlier initiation of barx1 and sox9a expression in intermediate/ventral relative to dorsal arch NCCs, with Jagged-Notch and Prrx1a/b circumscribing the size of the later-forming dorsal condensations.
In a previously proposed ‘hinge-and-caps’ model [93, 94], arch polarity is established by differential signaling in the intermediate regions of the arches (i.e., ‘hinges’) versus the dorsal and ventral poles of the arches (i.e., ‘caps’). Our work provides potential cellular correlates to these hinges and caps in zebrafish, particularly in the second arch and posterior portion of the first arch [95]. We propose that the poles of the arches, or caps, represent progenitor domains, consistent with their expression of the mesenchyme progenitor marker Prrx1 in many species [22, 80, 96]. In contrast, the intermediate arches, or hinges, reflect the sites of initial chondrogenesis, as evidenced by their earlier expression of sox9a. Whereas this model predicts that Prrx1 expression at both the ventral and dorsal poles would restrict chondrogenesis relative to the intermediate hinges, dorsal-specific Jagged-Notch signaling would further restrict chondrogenesis in dorsal relative to ventral regions. This model would explain our observations that cartilages generally form first in the intermediate regions, then spread next to the ventral pole, and lastly to the dorsal pole due to combined repressive effects of Prrx1 and Jagged-Notch. However, the timing of cartilage differentiation is clearly more complex. For example, our time-lapse imaging revealed that the Ch cartilage first undergoes chondrogenesis at its tips and then later in its center, potentially correlating with expression of the Bmp target gene msxe in a subset of ventral second arch Ch precursors [14]. Hence, layering of additional signaling pathways, such as Bmp, may further refine the timing of cartilage differentiation within the arches.
Given the expression of Prrx1 homologs at the ventral and dorsal poles of the arches from sharks through mammals [96], and conserved expression of Jag1 in the dorsal arches of mice [97, 98], it appears likely that a similarly regulated intermediate—ventral—dorsal gradient of chondrogenesis may be conserved across vertebrates. For example, in human embryos, Meckel’s cartilage (ventral) differentiates before those elements that form in more proximal/dorsal positions (i.e. the malleus, sphenoid, and styloid process) [99]. On the other hand, differences in the timing and extent of cartilage differentiation might account for the striking differences in facial form between species. In larval zebrafish, the majority of the bony visceral skeleton arises through cartilage templates in the first two arches. In contrast, much of the mammalian facial skeleton forms through direct ossification, with exceptions including Meckel’s cartilage in the lower jaw and the ossicles of the middle ear. These differences might be reflected in the fact that loss of the pre-cartilage marker Barx1/barx1 has more profound effects on the facial skeleton of zebrafish than mice [32, 100, 101], and, reciprocally, that loss of Prrx1 genes impacts lower jaw development in mammals but not fish [21, 22, 48]. It will therefore be interesting to examine whether differences in the requirements and/or regulation of Prrx1 and Barx1 genes underlie differences in the extent and timing of chondrogenesis between species.
An unanswered question is how heterochrony in cartilage differentiation might translate to the distinct shapes of skeletal elements along the dorsoventral axis. One possibility is that dorsoventral differences in the timing at which progenitors commit to a cartilage fate influences the duration and types of signals they encounter from the surrounding endoderm and ectoderm. For example, Jagged-Notch signaling in the dorsal posterior second arch would protect progenitors from early chondrogenesis, thus allowing these cells to receive later osteogenic cues that direct them to form the large, fan-shaped opercle bone. Such an interpretation is consistent with the reciprocal expansion of barx1+ pre-cartilage condensations and loss of opercle bone in jag1b mutants, and the formation of an ectopic opercle bone upon forced expression of JAG1 in ventral regions [17]. Another possibility is that the timing of condensation formation and subsequent chondrogenesis influences the degree of proliferative expansion of elements in different arch domains [1]. In conclusion, our study revisits heterochrony, a fundamental concept of evolutionary biology, from a developmental perspective, showing that the timing and extent of cartilage differentiation within specific arch regions contributes to the diversity of skeletal shapes within the skull.
All zebrafish (Danio rerio) were maintained and handled in strict accordance with good animal practices as defined by the relevant national and local animal welfare bodies. Zebrafish embryos were anesthetized for time-lapse imaging or prior to fixation by adding tricaine to their water. All animal experiments performed in this study were approved by the Institutional Animal Care and Use Committee of the University of Southern California (No. 10885, 20193).
Zebrafish (Danio rerio) embryos were reared at 28.5°C and staged as previously described [102]. The following transgenic lines were maintained as heterozygotes: Tg(fli1a:EGFP)y1 [103], Tg(sox10:DsRed-Express)el10 [104], Tg(col2a1aBAC:GFP) [105], Tg(hsp70I:Gal4)kca4/+ and Tg(UAS:myc-Notch1a-intra)kca3 (hereafter UAS:NICD) [106], Tg(UAS:Edn1;α-crystallin:Cerulean)el249 and Tg(UAS:Bmp4;cmlc2:GFP)el49 (hereafter UAS:Edn1 and UAS:Bmp4, respectively) [14]. The hsp70I:Gal4 and UAS:NICD lines do not contain selectable markers and were genotyped using primers for Gal4 (F: 5′-CTCCCAAAACCAAAAGGTCTCC-3′; R: 5′-TGAAGCCAATCTATCTGTGACGG-3′) and UAS:NICD (F: 5’-CATCGCGTCTCAGCCTCAC-3’; R: 5’-CGGAATCGTTTATTGGTGTCG-3’). For the UAS:Edn1 line, in cases where it was not possible to ascertain α-crystallin:Cerulean expression in living animals, individuals carrying the transgene were identified by genotyping for the lens marker (F: 5’-TGGTGCAGATGAACTTCAGG-3’ and R: 5’- GCATGCAGACAGCAGCAATA-3’). Gal4 expression was induced in hsp70I:Gal4; UAS:NICD, hsp70I:Gal4; UAS:Edn1, and hsp70I:Gal4; UAS:Bmp4 embryos by heat-shocking from 20–24 hpf in a 40°C incubator. The sucker/edn1tf216 [5], jag1bb1105 [17], barx1fh331 [32], notch3fh332 [107], and Df(Chr1)hand2s6 [108, 109] mutant lines were described previously and genotyped by PCR using GoTaq (Promega, Madison, WI) with published primer sequences followed by digestion with the appropriate restriction enzymes.
Three new mutant lines (notch2el515, prrx1ael558, prrx1bel491) were generated for this study via TALEN-mediated mutagenesis. The notch2el515 allele was generated with the same TALEN pair used for the previously reported notch2el517 allele [110]. Exon 2 (of 4) of prrx1a was targeted with TALENs that recognize the following sequences: Left: 5’-CGTTGAGCTGCTCGTCTGGA-3’; Right: 5’-TGTTTCGCCTCTGTTTACGC-3’, and exon 1 (of 5) of prrx1b was targeted with TALENs that recognize the following sequences: Left: 5’-TGGCGAAACGGGCAGGACTA-3’; Right: 5’-TGTATCACTGCCACTCGTTA-3’. TALEN constructs were produced using a PCR-based platform [111]. The TALEN plasmids were linearized by StuI digestion (New England Biolabs, Ipswich, MA), and RNAs were synthesized with the mMessage mMachine T7 Ultra kit (Ambion/Life Technologies, Carlsbad, CA, USA). TALEN RNAs (100 ng/μl) were injected into 1‐cell-stage embryos. Germline founders were identified among the injected individuals by screening outcrossed progeny by PCR followed by restriction digestion. The primers used to identify mutations in each gene are listed in S7 Table. Stable mutant alleles predicted to result in immediate stop codons or frameshifts followed by stop codons were identified by sequencing PCR products in the F1 generation. The notch2el515 allele consists of a 2-bp deletion and a single nucleotide polymorphism (SNP) that destroy a ClaI site in the target region and result in an immediate stop after aa 208 (of 2471), within the extracellular EGF-like domains. The prrx1ael558 allele is an 8-bp deletion that destroys a BseRI site and produces a frameshift after aa 90 (of 245; upstream of the homeodomain at aa 101–155), causing the addition of one incorrect amino acid followed by a stop codon. The prrx1bel491 allele is a 2-bp insertion that abolishes a HinfI site and causes a frameshift after aa 68 (of 245; upstream of the homeodomain at aa 87–165), resulting in the addition of four incorrect amino acids followed by a stop codon. Two additional alleles, prrx1ab1246 and prrx1bb1247, were independently generated via CRISPR-mediated mutagenesis. CRISPR gRNA templates were produced via PCR following a published protocol [112], and gRNAs were synthesized with the MEGAScript T7 transcription kit (Ambion) and column-purified with the mirVana miRNA isolation kit (Ambion). Cas9 RNA was transcribed from pT3TS-nCas9n with the T3 mMessage kit (Ambion) and purified with an RNeasy Mini Kit (Qiagen, Hilden, Germany) [112]. gRNAs (25 ng/μl) plus Cas9 RNA (50 ng/μl) were injected into 1‐cell-stage embryos, and stable lines were identified by sequencing as described above. The prrx1ab1246 allele is an 11-bp deletion that causes a frameshift after aa 62, which results in the incorporation of 28 additional amino acids followed by a stop codon. The prrx1bb1247 allele consists of an 8-bp deletion that causes a frameshift after aa 24 and the inclusion of 29 incorrect amino acids before termination. All animal experiments performed in this study were approved by the Institutional Animal Care and Use Committee of the University of Southern California.
For RNA sequencing experiments, fli1a:EGFP fish were crossed to the sox10:DsRed line, and doubly transgenic fli1a:EGFP; sox10:DsRed fish were further crossed to the edn1, jag1b, hsp70I:Gal4, UAS:Edn1, and UAS:NICD lines. Each of these lines were then separately incrossed to generate embryos for FACS sorting. Wild-type fli1a:EGFP; sox10:DsRed (20, 28, and 36 hpf) embryos were sorted for co-expression of GFP and DsRed expression under a fluorescent dissecting stereomicroscope (Leica M165 FC, Wetzlar, Germany) prior to dissociation. Single-positive and double-negative embryos were also saved as controls for FACS. Mutant edn1; fli1a:EGFP; sox10:DsRed embryos were selected under the fluorescent microscope at approximately 34 hpf based on the reduced distance between the bottom of the first pharyngeal pouch and the ventral border of the arches. To identify jag1b mutants and doubly-transgenic hsp70I:Gal4; UAS:Edn1 or hsp70I:Gal4; UAS:NICD individuals, we genotyped cell lysates of tail biopsies collected from anesthetized individual 24-hpf fli1a:EGFP; sox10:DsRed double-positive embryos. To induce Edn1 or NICD overexpression in the hsp70I:Gal4; UAS:Edn1 and hsp70I:Gal4; UAS:NICD lines, embryos were heat-shocked from 20–24 hpf in an incubator set at 40°C. As another means of inhibiting Notch signaling, fli1a:EGFP; sox10:DsRed embryos were treated with dibenzazepine (DBZ; Tocris, Bristol, UK; final concentration of 10 μM in embryo medium) from 24–36 hpf. The number of embryos used for each sort and the number of cells obtained are presented in S9 Table.
To facilitate FACS analyses at the 36 hpf time point, embryos were moved at 27 hpf to an incubator set at 22°C to delay their development such that they reached an approximation of the 36 hpf stage the following morning. fli1a:EGFP; sox10:DsRed double-positive embryos were dissociated following [113], with minor modifications. Briefly, 30–40 dechorionated embryos were incubated in fresh Ringer’s solution for 5–10 minutes and agitated by pipetting to remove the yolk. The deyolked embryos were then mixed with a protease solution containing 0.25% trypsin (Life Technologies), 1 mM EDTA, and 2 mg/ml Collagenase P (Roche Life Science, Indianapolis, IN) in PBS and incubated at 28.5°C for 15 min, pipetting up and down every 5 min to aid the dissociation. The reaction was stopped by the addition of a 6x stop solution consisting of 6 mM CaCl2 and 30% fetal bovine serum (FBS) in PBS. The cells were pelleted via centrifugation at 2000 rpm for 5 min at 4°C, resuspended in suspension medium (1% FBS, 0.8 mM CaCl2, 50 U/ml penicillin, and 0.05 mg/ml streptomycin (Sigma-Aldrich, St. Louis, MO) in phenol red-free Leibovitz’s L15 medium (Life Technologies)), pelleted again as above, and then resuspended in 500 μl suspension medium and placed on ice. Cells were sorted by FACS for GFP and DsRed expression on a MoFlo Astrios instrument (Beckman-Coulter, Brea, CA, USA). GFP/DsRed double-positive, double-negative, and single-positive populations were collected directly into RLT lysis buffer (Qiagen). Total RNA was immediately extracted using the RNeasy Micro kit (Qiagen) following the manufacturer’s protocol and quantified on a NanoDrop 2000 spectrophotometer (NanoDrop Products, Wilmington, DE, USA).
The quality and quantity of extracted RNA were assessed on a Bioanalyzer Pico RNA chip (Agilent, Santa Clara, CA). cDNA was then made from the extracted RNA using the SMARTer V3 kit (Clontech, Mountain View, CA), according to the manufacturer’s instructions. The number of amplification cycles for cDNA synthesis was estimated based on input amounts of RNA. The size and amount of the resulting cDNA were then confirmed by Bioanalyzer. Sonication was performed on a S2 ultrasonicator (Covaris, Woburn, MA) according to Clontech’s recommended conditions. DNA libraries were constructed using the Kapa Hyper prep kit (Kapa Biosystems, Wilmington, MA) and NextFlex adapters (Bioo Scientific, Austin, TX). Libraries were visualized by Bioanalyzer analysis and quantified by qPCR (Kapa library quantification kit). Sequencing was performed on Illumina HiSeq 2000 (50-bp paired end reads) and NextSeq 500 (75-bp paired end reads) machines (Illumina, San Diego, CA). DNA libraries were constructed and sequencing was performed at the Norris Cancer Center Molecular Genomics Next Gen Sequencing Core at USC.
Raw sequencing data in Fastq format was imported into the Partek Flow interface for alignment and quantification. Pre-alignment QC showed that the reads from all samples had generally high quality, with the average Phred quality score for each sample being above 30. Reads were then trimmed from both ends based on Phred quality score with a minimum end quality level of 20 and a minimum acceptable read length of 25. The TopHat 2 algorithm was used to align the trimmed reads to the zebrafish GRCz10 genome assembly (Ensembl_v80). Aligned reads were then quantified using the Partek E/M algorithm with default parameters to yield the RPKM values. RNAseq files have been deposited in NCBI’s Gene Expression Omnibus and are accessible through the GEO Series accession number GSE72985. Filtered gene lists were derived in MS Excel as described in the Results section. Six genes on the list of arch NCC-enriched genes had passed the ≥ 3 RPKM threshold at 36 hpf but had RPKM values of 0 in the 20 hpf sample, leading to a division error that would have precluded their inclusion in the temporal expression analysis; we thus set the 20 hpf RPKM value for these genes to 0.01 based on the lowest positive RPKM values in the dataset.
Alcian Blue and Alizarin Red staining to detect cartilage and bone, respectively, was performed on 4–6 dpf larvae as previously described [114]. Two-color fluorescent in situ hybridizations were carried out as previously reported [17]. Published probes used in this study include dlx2a [115], dlx5a [10], notch2, jag1b [17], and sox9a [44]. Partial cDNAs for barx1, notch1a, notch1b, notch3, prrx1a, and prrx1b were cloned into the pCR-Blunt II-TOPO vector (Life Technologies) and sequence-verified prior to plasmid linearization and in vitro transcription with Sp6 or T7 polymerase (Roche) (S8 Table).
To determine when Notch signaling affects skeletal patterning, we treated embryos with the γ-secretase inhibitor DBZ. DBZ dissolved in dimethyl sulfoxide (DMSO; 10 mM stock) was added to embryo medium to a final concentration of 10 μM. Embryos (n = 30–50 per treatment) were incubated in this solution starting at 8, 24, or 28 hpf until fixation at 36 (8 hpf group) or 42 hpf (other groups) for in situs or at 4 dpf for Alcian and Alizarin staining. In the groups used for skeletal staining, the DBZ solution was refreshed at 48 hpf and thoroughly washed out at 56 hpf. Clutch-mate controls were exposed to the same concentration of DMSO. Embryos were dechorionated at 24 hpf to improve drug accessibility.
Fate-mapping was performed with the green-to-red photoconvertible kikGR protein [116]. kikGR RNA was injected into embryos from a jag1b; fli1a:EGFP cross at the one-cell stage. At 36 hpf, embryos were anesthetized in Tricaine (Sigma-Aldrich) and mounted in 0.2% agarose for confocal imaging. Small groups of cells in the dorsal arches were selected using the region of interest tool in the Zeiss LSM software and exposed to the UV 405 laser until the red photoconverted protein became apparent (typically ~10 seconds using 50% laser power). The same animals were reimaged at 6 dpf to determine the destination of the photoconverted cells and then genotyped for the jag1bb1105 mutation.
Confocal images of in situ hybridizations (~30 μM z-stacks) were captured on a Zeiss LSM5 microscope using ZEN software. Time-lapse imaging of doubly transgenic fli1a:EGFP; sox10:DsRed and col2a1aBAC:GFP; sox10:DsRed larvae followed [117], with ~130 μM of z-stacks collected every 10 or 12 minutes starting at 48 hpf. Skeletal preparations were photographed using a Leica DM2500 microscope. Image levels were adjusted in Adobe Photoshop CS6, with care taken to apply identical adjustments to images from the same data set and to avoid removing information from the image.
To analyze changes in gene expression between 20 and 28 hpf and 28 and 36 hpf for the total arch and Edn1- and Notch-regulated gene lists, we first calculated the median and quartile values for each list. The full lists were then collectively compared first by a Kruskal-Wallis test and then pairwise by Mann-Whitney U tests, with the Bonferroni correction applied to an α-value of 0.05 to account for multiple comparisons. Chi-square was used to compare the proportions of embryos showing different skeletal phenotypes in jag1b, barx1, and jag1b; barx1 mutants, with p < 0.05 considered significant. JMP 7.0 software (SAS) was used for statistical analysis. Numbers for each experiment are presented in S10 Table.
|
10.1371/journal.pntd.0003581 | Beriberi (Thiamine Deficiency) and High Infant Mortality in Northern Laos | Infantile beriberi (thiamine deficiency) occurs mainly in infants breastfed by mothers with inadequate intake of thiamine, typically among vulnerable populations. We describe possible and probable cases of infantile thiamine deficiency in northern Laos.
Three surveys were conducted in Luang Namtha Province. First, we performed a retrospective survey of all infants with a diagnosis of thiamine deficiency admitted to the 5 hospitals in the province (2007–2009). Second, we prospectively recorded all infants with cardiac failure at Luang Namtha Hospital. Third, we further investigated all mothers with infants (1–6 months) living in 22 villages of the thiamine deficiency patients’ origin. We performed a cross-sectional survey of all mothers and infants using a pre-tested questionnaire, physical examination and squat test. Infant mortality was estimated by verbal autopsy. From March to June 2010, four suspected infants with thiamine deficiency were admitted to Luang Namtha Provincial hospital. All recovered after parenteral thiamine injection. Between 2007 and 2009, 54 infants with possible/probable thiamine deficiency were diagnosed with acute severe cardiac failure, 49 (90.2%) were cured after parenteral thiamine; three died (5.6%). In the 22 villages, of 468 live born infants, 50 (10.6%, 95% CI: 8.0–13.8) died during the first year. A peak of mortality (36 deaths) was reported between 1 and 3 months. Verbal autopsy suggested that 17 deaths (3.6%) were due to suspected infantile thiamine deficiency. Of 127 mothers, 60 (47.2%) reported edema and paresthesia as well as a positive squat test during pregnancy; 125 (98.4%) respected post-partum food avoidance and all ate polished rice. Of 127 infants, 2 (1.6%) had probable thiamine deficiency, and 8 (6.8%) possible thiamine deficiency.
Thiamine deficiency may be a major cause of infant mortality among ethnic groups in northern Laos. Mothers’ and children’s symptoms are compatible with thiamine deficiency. The severity of this nutritional situation requires urgent attention in Laos.
| Infantile thiamine deficiency (beriberi), is rarely seen today after decades of strong public health attention. Infantile beriberi occurs mainly in infants breastfed by mothers with inadequate intake of thiamine. There is evidence of the persistence of infantile thiamine deficiency in Vientiane, the capital of Laos, but insufficient data from outside Vientiane to justify a policy recommendation. We describe possible and probable cases of infantile thiamine deficiency in northern Laos using retrospective and prospective hospital data. In addition we conducted a cross sectional survey in 22 villages where the infants originated. Infantile thiamine deficiency was quite common in retrospective and prospective (hospitals: 54, villages: 17) and cross-sectional surveys (hospital: 4, villages: 10). A second peak of infantile mortality was observed before 6 months and was associated with a high infant mortality in the villages, 106 per 1000 live births (95%CI: 86–128). A total of 60 pregnant mothers and 70 lactating women showed signs of thiamine deficiency. This situation requires urgent attention in Laos.
| Thiamine (vitamin B1) acts as an important cofactor in metabolism and energy production. It is required for the biosynthesis of neurotransmitters and the production of substances used in defence against oxidant stress [1]. Thiamine deficiency occurs predominantly in populations, in which the diet consists mainly of very poor sources of thiamine such as milled white cereals, including polished rice (the rich thiamin envelop is removed by polishing) and wheat flour, and where other key sources of thiamine (meat, fish, and vegetables) are infrequently consumed [2]. It is also related to diets that are rich in thiaminase, the natural thiamine-degrading enzyme, which is abundantly present in raw and fermented fish sauce (a common Asian delicacy) certain vegetables and roasted insects consumed primarily in Africa and Asia [3]. Thiamine deficiency can develop within 2–3 months from a deficient intake and can cause illness and death [4].
Clinically apparent thiamine deficiency, also known as Beriberi, has historically been described in vulnerable populations such as refugees, prisoners, during times of war [2,5] and in developed countries with alcohol abuse or parenteral nutrition with insufficient thiamine intake in adults [6,7].
More recently, thiamine deficiency outbreaks were described among young healthy Thai construction workers in Singapore in the 1980s [8,9] and among commercial fishermen in Thailand in 2005 [10,11]. Worldwide, outbreaks of thiamine deficiency have recently been reported in Ivory Cost jails [10,12], in the Gambia [13,14], among African Union troops in Mogadishu, Somalia [15], in Brazil [15–18] and the French island of Mayotte [19]. Adult thiamine deficiency was also recently described in Cambodia, China, India, Thailand, Laos and various countries with unbalanced thiamine/thiaminase diets [8,19–24].
Cardiac failure associated with thiamine deficiency has also been described in Japanese teenagers consuming excessive sweet carbonated soft drinks, instant noodles and polished rice [25].
Thiamine deficiency, is rarely seen today in infants after decades of strong public health attention [26]. It is an acute disease mainly affecting infants that are breastfed by women with deficient thiamine levels [7]. The onset of symptoms is often very rapid and the fatality rate is very high with death often occurring within a few days from the onset of symptoms. Historically, the clinical features have been categorized into three main types; the pure cardiac form or wet thiamine deficiency, the aphonic form, and the neurologic or dry form [27]. The more severe form is called Shoshin beriberi and presents as cardiac failure and lactic acidosis [28]. Beriberi poses difficult diagnostic issues and can be a missed diagnosis, as the dry or wet forms can mimic critical illness or polyneuropathies. In addition, clinical manifestations such as tachypnea, chest indrawing, tachycardia and cardiomegaly can suggest other diagnoses [29–31] [32].
The content of thiamine in breastmilk is related to the mother’s thiamine status [7]. Post partum thiamine deficiency in refugee mothers was associated with high infant mortality in Karen refugees [2,5]. The overall infant mortality rate declined from 183 before the recognition of thiamine deficiency to 78 per 1000 live births afterwards. Thiamine deficiency has also been described in breastfeeding Cambodian and Lao mothers [20,33], and pregnant mothers in China [24]. Recent surveys using whole blood thiamine diphosphate (TDP) revealed that thiamine deficiency was associated with cardiac dysfunction and tachypnea in Cambodian infants [32,34].
In developed countries infantile thiamine deficiency outbreaks have recently been described. For example in Israel an outbreak was due to thiamine deficient soya formula, with a high fatality rate and long term sequelae [35]. In the French island of Mayotte, deficiency was related to inadequate nutrition [28,36]. Infantile thiamine deficiency is periodically reported in intensive care units in babies receiving parenteral nutrition without thiamine or babies with malabsorption receiving prolonged but inadequate vitamin supplements [6,28,37,38].
Infantile thiamine deficiency was described in Laos in the sixties [39]. More recently, cases of young infants with cardiac failure in Mahosot Hospital, Vientiane, suggested the persistence of thiamine deficiency as a cause of infantile mortality [40]. Traditional food avoidance during the post-partum period, nutritional habits, and the high rate of childhood stunting (40%) may all be related to thiamine deficiency [33]. A recent publication revealed that clinically unapparent thiamine deficiency was common among sick infants without overt clinical thiamine deficiency admitted in 2003–2004 [41]. Alarming reports have been received from physicians about the possibility of thiamine deficiency in infants with cardiac failure in northern Laos in recent years [23,41].
However, there is insufficient data from outside Vientiane to provide evidence for discussions about thiamine supplementation in the Lao national nutrition strategy. To help fill this gap, we describe possible and probable cases of infant and maternal thiamine deficiency in Luang Namtha province.
Luang Namtha province is located in the northwest of the country, bordering Myanmar and China. It is one of the country’s poorest areas with a population of 148,797 people, many of whom live in remote mountain villages. Approximately 23.4% of women had at least one ante natal care visit during their most recent pregnancy in the province [32,42]. The infant mortality in the province (112 per 1000 live births) is one of the highest in the country (national rate 70 per 1000 live births) according to the 2005 national census [43]. At the time of this study (2008) the national infant mortality rate was estimated at 48 per 1000 births [44]. The provincial hospital in Namtha district is the referral centre for all 5 districts and the military hospital. It is a 50-bed hospital with 98 medical and non-medical staff in 2007. Approximately 1,400 outpatients are seen each month and 330 are admitted. It was the only health facility in the province with X-ray and surgery. The 4 other district hospitals are Long, Sing, ViengPhoukha and Nalea.
We conducted various surveys in Luang Namtha province as shown in the flow chart (Fig. 1). First, in a retrospective survey we recorded all infant inpatients with recorded thiamine deficiency in the 5 hospitals in the province between 2007 and 2009. Second, in a prospective survey we recorded all infants admitted with cardiac failure at the emergency ward of Luang Namtha hospital from March to June 2010. Third, from these two surveys we identified 22 villages from where the patients originated and then investigated all mothers with infants (1–6 months) living in these villages. We conducted i) a verbal autopsy of all infants’ deaths and estimated infantile mortality; ii) a cross sectional survey of all mothers and infants (1–6 months) using a pre-tested questionnaire, physical examination and squat test (defined by “the inability of the individual to rise from a squatting position, due to weakness then flaccid paralysis of the lower limbs, without assistance”) [7,12].
Characteristics of infants with a discharge diagnosis of thiamine deficiency admitted at the hospitals were retrieved from hospital records. A standardized form was used that included age of infant, main symptoms, treatment received, and response to treatment. An 80 item questionnaire was used in the villages. It included general information on the population (8 questions), presence of a rice-mill, type of rice consumed, socio-economic characteristics (32 questions), maternal food avoidance behaviour, food given to the child in the previous 15 days and information on the children, age by day, sex, mode of birth and detailed information regarding the causes of infants’ death (40 questions).
We used definitions of possible or probable thiamine deficiency for mothers and for infants with sudden cardiac failure or death, based on symptoms and response to thiamine treatment [36].
Possible adult thiamine deficiency was defined in a pregnant women or a mother with a child less than 6 months if she presented or had presented during her pregnancy with at least two of the following signs: motor deficits, paresthesia of the limbs (peripheral numbness, tingling or plantar pain), loss of reflexes, signs of heart failure (jugular swing, cardiac gallop rhythm on auscultation, hepatomegaly), associated with a positive squat test (unable to rise after squatting) [2].
Possible thiamine deficiency in infants was defined as acute symptoms in previously healthy breastfeeding infants associated with cardiac failure (tachypnea> 50/min, tachycardia> 170/min, gallop, hepatomegaly> 3 finger's breadth) or loss of voice. Probable thiamine deficiency was defined if symptoms recovered after thiamine treatment.
Death was diagnosed as due to thiamine deficiency for previously healthy breastfed infants with less than two days of illness fulfilling the possible or probable thiamine deficiency definitions above, and as probable if associated with mother’s symptoms of thiamine deficiency. Due to possible misdiagnosis with acute pneumonia (though infection can precipitate thiamine deficiency [40]) children with signs of pneumonia (cough, fever, +- dyspnea) were excluded if no signs of thiamine deficiency were present in the mother.
The final verbal autopsy diagnosis was proposed during a review meeting of all cases by a committee including one pediatrician, one public health advisor, and two physicians. Only consensual diagnoses were retained.
Mothers suspected of thiamine deficiency were treated with vitamin B1 tablets 100mg, twice daily for 20 days and infants with suspected thiamine deficiency were treated with vitamin B1 tablets, 30 mg per day for 20 days. Patients with acute symptomatic thiamine deficiency received an intramuscular or slow intravenous injection of thiamine (100mg IM for mothers and 50mg for infants). Treatment for associated infection, if any suspected, was provided appropriately. Treatment was provided free of charge. Moreover, all families and village populations received information regarding thiamine deficiency prevention.
Data was entered into EpiData freeware. All records were crosschecked with the original data sheets. Analysis was carried out with STATA, Version 8 (Stata Corporation, College Station, TX, USA). Chi-squared, Fisher’s exact tests and Student’s t-test were used to compare categorical variables and continuous data, respectively. 95% confidence intervals were calculated for continuous and categorical data. We considered p < 0.05 as statistically significant. Infant mortality rates were calculated and compared to the national rate in Laos at the time of survey [44].
The study was authorized by the Lao health authorities. Information about the study was provided in Lao language and translated into the local ethnic language by one volunteer from each village. All participants gave informed oral consent in the presence of one village witness as the majority could not read. The procedure of the study was granted ethical approval by the Lao Medical Ethics Committee.
Between 2007 and 2009, 54 infants with sudden onset of cardiac failure were admitted to the 5 hospitals of Luang Namtha province. This number increased from 9 to 24 per year. The infant clinical characteristics and treatment evolution are presented in Table 1. Among them, 20 (37%) probably had an associated infection which may have triggered cardiac failure. Of the 54 infants with cardiac failure, 49 (90.7%) were cured after thiamine administration, three died (5.6%) and two had an unknown status (3.7%). Time of cure was not recorded in patients’ files.
Four infants with clinical thiamine deficiency were observed during the prospective study at the provincial hospital from March to June 2010. All recovered after thiamine administration and were visited in their own village later. We present a typical description of one patient (Box 1).
Of 22 villages visited, 18 (81.8%) had an electric rice mill. Of a total of 167 mothers with an infant less than 6 months, 127 mothers and their infants were present and gave consent to be interviewed and undergo a physical examination (Table 2). All mothers consumed polished white rice, 36 (28.3%) had at least one antenatal visit and 28 (22%) reported they received some information on nutrition from health staff during antenatal care. Less than half of the children had received some immunizations (60, 47.2%) (Table 3). Nearly all mothers (125, 98.4%) respected food avoidance after delivery with a median of 30 days.
A third of the mothers (45, 35.4%) reported to have had at least one of their children die. Of 468 live born infants, 50 (10.6%, 95%CI: 8.0–13.8) infants died during the first year. Based on this survey, the infant mortality rate was 106 per 1000 live births (95% CI: 86–128).
Thirty-six infants (7.6%) died below the age of 6 months. According to mothers, 22.8% of infant deaths occurred during the neonatal period while 29.5% and 23.8% of deaths occurred during the second and third month respectively, and dropped to 11.8% during the fourth and fifth months which suggested a plateau of infantile mortality during the first 3 months of life (Fig. 2). Twenty (10.6%) children presented with sudden death compatible with thiamine deficiency. The verbal autopsy suggested that 17 (3.6%) infants died of thiamine deficiency, 13 (2.7%) as probable and 4 (0.8%) as possible thiamine deficiency. Loss of voice was reported in 10/17 (58.8%). A typical patient is presented in Box 2. For the remaining deaths the verbal autopsy suggested other pathologies (meningitis, laryngitis, convulsions, neonatal infection) as probable causes.
This survey suggests the presence of infantile thiamine deficiency and thiamine deficiency in pregnant women and breastfeeding mothers from diverse ethnic groups in Luang Namtha province. Thiamine deficiency was either diagnosed or suspected alone or in association with an infection at the hospital level and in the community. A dramatic therapeutic response to thiamine supplementation was observed in nearly all hospital cases (90.7%). The survey also shows an excessive infant mortality in the 22 villages compared to the national rate [45]. Thiamine deficiency remains a poorly recognized but readily treatable cause of infant death, which strongly advocates for public health prevention and education. We previously recommended using antenatal visits to provide such prevention in Laos [33]. However the rate of antenatal care which varies widely between rural (29%) and urban areas (71%) may limit this recommendation [46]. Extensive public education needs to be conducted with a particular emphasis on mothers belonging to ethnic groups, having a low dietary diversity and those performing hard physical labor or being farmers; three risk factors that were previously pointed out and that also apply to our study population [40]. This survey of highland Lao complements recent work conducted in Vientiane capital where both thiamine deficiency and clinically unapparent thiamin deficiency were described among a majority of lowland Lao (Lao Loum) [39–41]. Another survey shows that 12% of malaria patients (including 165 children less than 15 years of age) had severe biochemical thiamine deficiency in Savannakhet region, southern Laos, without clinical features of thiamine deficiency [23].
The number of thiamine deficiency inpatients increased over the 3 years of the retrospective survey. Over the last years, electricity and electric rice mills were installed in remote villages (H Barennes, personal observations, 2005 to 2012). Traditional foot operated mills which protect the vitamin content of rice, were abandoned and most of the women reported eating mill polished rice which has lost its vitamin B rich envelops. The general practice (98.4% of mothers) of strict food avoidance (eating mostly milled glutinous rice, soaked for hours in water and tea, rich in thiaminases) during the postpartum period further contributes to thiamine deficiency [7,40]. The practice of postpartum food avoidance is very common in Laos and has been associated with a possible reason for infant micro nutrient deficiency and stunting in a prospective survey in 41 randomly selected villages on the outskirts of Vientiane [33]. Nearly all mothers (96.6%) had insufficient thiamine intake [33]. Other possible reasons for the increased diagnosis of thiamine deficiency at the hospitals during the 3 year survey were the arrival of a trained and dedicated pediatrician at the provincial hospital, and the improvement of hospital services which attracted more patients from ethnic groups living outside of the main city (Hubert Barennes, Leila Srour, Gunther Slesak, personal observations). Nevertheless, the increasing rate of thiamine deficiency cases observed at the Luang Namtha hospital might be attributable to a local outbreak related to economic progress and the appearance of electric mills in rural communities of the region. These villages may thus be more at risk than remote communities still relying on foot pounding mills [39].
Before 1991, fatality of infants with congestive heart failure was high at Mahosot Hospital, Vientiane [40]. When infantile beriberi was recognized and measures were taken, clinical response to intravenous thiamine was obtained within 2 hours (15 minutes up to 2 hours) and the in-hospital death rate attributed to thiamine deficiency declined. Nowadays, physicians in Laos are familiar with thiamine deficiency and often treat infants with congestive cardiac failure systematically with parenteral thiamine. However, unapparent clinical thiamine deficiency is still highly prevalent among sick children attending the Vientiane hospital [41] and up to 50–90 children with suspected thiamine deficiency are seen at Mahosot hospital each year. Altogether, this suggests a failure in prevention of this easily curable deficiency.
Detecting thiamine deficiency cases retrospectively from hospital records can be questioned. Hospital records in resource limited settings are usually restricted to basic information. Indeed, the information on treatment efficacy was usually available but not the exact timing of cure. A rapid improvement is usually recognized after parenteral thiamine administration in the wet form of thiamine deficiency [5,40]. This treatment response is different in the neurological forms of thiamine deficiency which usually occur at an older age and were not addressed in this survey [7,35,47,48].
In 2010, in the 22 villages affected by thiamine deficiency, the infant mortality rate was 106 per 1000 live births (95% CI: 86–128), 2.2 fold the 2008 estimated national rate [44]. After neonatal death, instead of decreasing, the mortality remains stable showing a second peak of mortality. A second infant mortality peak before 6 months of age has historically been proposed as a significant feature of a potential thiamine deficiency health problem in the population [4,5,49]. It is important to highlight that public health intervention could reduce the infant mortality as was evidenced with Karen refugees in Thailand [5].
Mothers living in the 22 villages had combinations of all risk factors described for thiamine deficiency [5,40,41]. They were farmers, predominantly illiterate, had low family income, poor dietary diversity and almost all respected food avoidances during the postpartum period.
Typical symptoms of thiamine deficiency were common among women. Though they appeared to be quite common during pregnancy and after delivery, they should be addressed in the context of poor dietary diversity and for women living in similar situations as the one described in rural areas of Cambodia, Thailand, Laos and Vietnam [2,5,20]. During the outbreak in Mayotte among Comorian breastfeeding women, it was similarly reported that mothers complained of paresthesias of lower extremities, pain and some association with walking difficulty, called “lalavy” which resolved after thiamine administration. Interestingly, Comorian mothers had a post partum 40 days regimen which includes a large quantity of a specific ad hoc rice cooked with a lot of water [36].
Women in Laos should be educated about the importance of a diverse diet before and after delivery and how to maintain a sufficient thiamine intake. Pregnant and lactating mothers must be encouraged to eat unpolished rice, prepare their rice avoiding loss of micronutrients by avoiding unnecessarily long soaking, avoid fermented fish paste, and betel nut chewing during pregnancy and breastfeeding periods. Culturally acceptable ways need to be identified to limit postpartum food avoidance. These measures might be challenging in these populations; hence daily thiamin supplements which are affordable could be considered.
Health professionals should provide nutritional advices during the precious time of the antenatal visits and should be trained to recognize, prevent and treat early symptoms of thiamine deficiency and to offer thiamine supplements to mothers who are not able to systematically comply with dietary advice. For many years, Xayaboury province (north-western Laos) has included thiamine supplementation in the prenatal care program [50]. The midwife in charge of the program reported that thiamine deficiency cases are very uncommon in the province (Leila Srour, personal communication).
Educational campaigns which are now focusing on implementing exclusive breastfeeding in Laos must include thiamine deficiency prevention and detection, as an important component of these educational campaigns. The story of a mother in Savannakhet province helps to understand the context. Attending the district hospital in 2004 with a severely malnourished infant (6 kg at 1 year) she was asked why she fed the infant with coffee creamer. She reported that her first four infants had died before 6 months of age. Suspecting her breastmilk being the problem she decided to feed her fifth child with coffee creamer, as she could not afford expensive infant formula (Leila Srour, personal communication).
Further research is needed to evaluate which preventive strategies are the most effective to reach mothers living in remote villages. Determination of thiamine concentration in breastmilk and the infants’ thiamine status are still needed [40]. Further research is also needed to assess and prevent the hidden consequences of infant thiamine deficiency, especially neurological development and epilepsy [51,52].
A concomitant concern is the need for more accessible and inexpensive tests to evaluate thiamine deficiency, as the current basal erythrocyte transketolase activity (ETK) assays remains unavailable in settings where they are most needed [39]. Another concern is to clarify which tests are most useful. Recently the less conventional whole blood thiamine diphosphate (TDP) concentrations have been used in the field to assess thiamine deficiency [20,34].
This survey has several limitations, including the use of retrospective hospital data and the confounding factor of associated antibiotics together with thiamine administration, the possible recall bias while interviewing mothers regarding the subjective nature of signs such as paresthesia, the lack of laboratory testing for thiamine deficiency, and the use of clinical criteria only to assess thiamine deficiency [2]. Due to time and budget constraints the team could not follow treatment efficacy in the villages and could not validate cases.
We adapted our definitions of thiamine deficiency cases from 2 surveys: one hospital-based survey and one epidemiological outbreak survey [36,40]. Our hospital case definition did not exclude the presence of fever or suspected sepsis since there is evidence that these conditions contribute to precipitating thiamine deficiency [7,41,47]. This may have overestimated the thiamine deficiency frequency but the response to thiamine treatment, an important criterion for diagnosis, was positive in all but three. Conversely, retrospective case review did not include the presence of fever or signs of sepsis and we may have underestimated the number of true thiamine deficiency cases in these high risk populations.
Finally, we screened 22 villages with suspected thiamine deficiency cases but this strategy cannot provide a representative overview of the situation of thiamine deficiency in the region.
This survey suggests that thiamine deficiency is a major cause of high infant mortality among ethnic groups in northern Laos. Prevention of thiamine deficiency and nutritional education should be addressed on a larger population scale, particularly for pregnant and breastfeeding women, their offspring and their families. It should also focus on at risk Asian populations reporting similar low diversity diets, low thiamine intake, thiaminase rich diets and food avoidance during and after pregnancy.
|
10.1371/journal.ppat.1000923 | Analysis of Virion Structural Components Reveals Vestiges of the Ancestral Ichnovirus Genome | Many thousands of endoparasitic wasp species are known to inject polydnavirus (PDV) particles into their caterpillar host during oviposition, causing immune and developmental dysfunctions that benefit the wasp larva. PDVs associated with braconid and ichneumonid wasps, bracoviruses and ichnoviruses respectively, both deliver multiple circular dsDNA molecules to the caterpillar. These molecules contain virulence genes but lack core genes typically involved in particle production. This is not completely unexpected given that no PDV replication takes place in the caterpillar. Particle production is confined to the wasp ovary where viral DNAs are generated from proviral copies maintained within the wasp genome. We recently showed that the genes involved in bracovirus particle production reside within the wasp genome and are related to nudiviruses. In the present work we characterized genes involved in ichnovirus particle production by analyzing the components of purified Hyposoter didymator Ichnovirus particles by LC-MS/MS and studying their organization in the wasp genome. Their products are conserved among ichnovirus-associated wasps and constitute a specific set of proteins in the virosphere. Strikingly, these genes are clustered in specialized regions of the wasp genome which are amplified along with proviral DNA during virus particle replication, but are not packaged in the particles. Clearly our results show that ichnoviruses and bracoviruses particles originated from different viral entities, thus providing an example of convergent evolution where two groups of wasps have independently domesticated viruses to deliver genes into their hosts.
| The polydnaviruses (PDVs) are a unique virus type used by an organism (a parasitic wasp) to manipulate the physiology of another organism (a lepidopteran host) in order to ensure successful parasitism. The evolutionary origin of these unusual viruses, found in ∼17,500 braconid wasps (Bracoviruses) and ∼15,000 ichneumonid wasps (Ichnoviruses), has been a major question for the last decade. We thus undertook an exclusive work aiming at investigating this origin via the characterization of genes encoding structural components for both types of PDVs. The present paper constitutes the first report on the identity and genome organisation of the viral machinery producing Ichnovirus virions. Our results strongly suggest that Ichnoviruses originated from a virus belonging to a group as yet uncharacterized that integrated its genome into that of an ichneumonid wasp ancestor. More importantly, our results demonstrate that the ancestor of Ichnoviruses differs from that of Bracoviruses, which originated from a nudivirus. We have now identified, for the two types of PDVs, the non packaged viral genes and their products involved in producing particles injected into the host during oviposition. Together, these data provide an example of convergent evolution where different groups of wasps have independently domesticated viruses to deliver genes into their hosts.
| Polydnaviruses (PDVs) are unique viruses symbiotically associated with endoparasitic wasps belonging to the families Braconidae and Ichneumonidae. Virus particles produced in the ovaries [1] are injected into lepidopteran hosts during wasp oviposition. The PDV genomes packaged in the particles are composed of circular dsDNA molecules that harbor from 60 to 200 genes [2]–[5]. These genes are expressed in infected caterpillar tissues, and their products ensure successful parasitism by abolishing host immune responses and/or altering host larval development [6]–[9].
Viruses belonging to a given family typically share a set of conserved genes (core genes) involved in DNA replication, transcription of viral genes and particle morphogenesis. Strikingly, PDV genomes packaged in the particles lack such typical virus genes. This is not completely unexpected given that no PDV replication takes place in the caterpillar; rather, replication is confined to the wasp ovary where viral DNAs destined for packaging are generated from proviral copies maintained within the wasp genome [10], [11]. Thus the genes involved in particle replication are not required within the particles. We recently identified genes encoding structural components of PDVs associated with braconid wasps (Bracoviruses or BVs). These genes derive from an ancestral nudivirus (nudiviruses are a sister group of baculoviruses), but instead of being packaged in BV virions they are transcribed from the wasp genome [12], [13]. Overall the data support the hypothesis that a nudivirus integrated its own genome into that of the ancestor of bracovirus-associated wasps, which lived ∼100 million years ago, according to a recent estimation based on the age of fossils in amber [14]. Since their integration into the wasp genome, the original nudivirus genes that were not essential to the parasitoid host interaction appear to have been replaced, in the packaged BV genome, by genes contributing to the success of parasitism. The nudivirus-like genes are specifically expressed in the calyx region of the wasp ovaries, where BV virions are produced, and they have been maintained in the ancestor wasp lineage and selected for their contribution to the success of parasitism as gene transfer agents for ∼100 million years.
PDVs associated with ichneumonid wasps (Ichnoviruses or IVs) share many features with BVs: they deliver genes in the parasitized host that are necessary for the successful development of the parasite, their packaged genome is composed of dsDNA circles and their virions are specifically produced in the calyx. However, IVs do not appear to derive from a nudivirus and their origin remains unclear. Indeed, the analysis of cDNAs from the wasp Hyposoter didymator did not lead to the identification of nudivirus genes expressed in the ovaries [12]. Moreover no set of genes having significant similarities with known viral genes could be identified. The recent discovery of a new lineage of insect viruses [15], [16], sharing few genes with other viruses, indicates that the extent of diversity of insect viruses is not completely known. This could explain our inability to identify core genes of viral origin in ichneumonid genomes. To overcome this problem, we searched for genes involved in IV particle production by analyzing the protein components of purified H. didymator ichnovirus (HdIV) particles and we studied the protein-coding DNA organization in the wasp genome.
IV particles have an ovocylindrical shape and a large nucleocapsid (330×85 nm) surrounded by two envelopes [17], [18]. Whereas the inner envelope is acquired de novo in the nucleus of calyx cells, where the particles are produced, the outer envelope is acquired from the cell membrane, during particle exocytosis into the oviduct lumen [1]. As expected from this complex structure, purified particles have a protein profile comprising several dozens of components [19], [20]. So far, only two structural proteins associated with the IV virions of the wasp Campoletis sonorensis (CsIV) have been characterized: p12 (gi|4101554) and p44 (p53 gene product, gi|4101552) [21], [22]. They display no overall resemblance to known proteins, although some similarities based on secondary structure analyses between domains of p44 and of an ascovirus protein have recently been described [23].
Here we report the characterization of 40 genes involved in HdIV virus particle production. Strikingly they are densely clustered in specialized regions of the wasp genome that we named “Ichnovirus Structural Protein Encoding Regions (IVSPERs)”. These genes are specifically expressed in the calyx of H. didymator and at least 19 of them encode components of virus particles, including homologues of the p12 and p53 genes originally identified in C. sonorensis. Furthermore, we showed that 11 homologues of HdIV IVSPER genes are expressed in the ovaries of the ichneumonid wasp Tranosema rostrale, indicating that the set of structural genes is conserved among wasps associated with IVs.
Unexpectedly, IVSPERs are amplified during virus replication in H. didymator ovaries despite not being packaged in the particles. Altogether, IVSPER genomic structure, replication properties and involvement in particle production suggest they originated from a common set of genes, which could correspond to the genome of an ancestral virus.
In order to identify genes expressed during HdIV virus particle production, 5636 clones from H. didymator ovary cDNA libraries were sequenced, resulting in the identification of 1956 non redundant coding regions. No significant similarities were detected with genes of conventional viruses, but two genes similar to CsIV p12 (named H. didymator p12-1 and p12-2) and two similar to CsIV p53 (named p53-1 and p53-2) were identified. Quantitative RT-PCR (qPCR) analyses indicated that transcript levels of these four H. didymator genes were at least 25 times higher in the calyx, where HdIV particles are produced, than in the ovarioles of female pupae (Table S1). This strongly suggested that these genes encode structural particle components, as shown for the CsIV p12 and p53 proteins, but also that sequences of additional IV structural genes were likely present in the libraries.
Several nudivirus-like genes involved in BV particle production are clustered in the wasp genome [13]. We hypothesized previously that this cluster corresponds to a remnant of the nudivirus genome acquired by the braconid ancestor wasp. To determine whether HdIV structural genes were similarly clustered, we isolated H. didymator genomic DNA clones containing the p12-2, p53-1 and p53-2 genes by screening a wasp bacterial artificial chromosome (BAC) library using gene-specific probes (BAC clones BQ, BR and BT; Figure 1).
Sequencing of these clones revealed that the p12-2, p53-1 and p53-2 genes reside in genomic regions characterized by a high density of coding sequences (exon density: 62.2%), making them atypical compared to the rest of the wasp genome (exon density: 21%). There was a significant difference in the mean length of intergenic sequences between these atypical regions (638 bp) and other portions of the wasp genome (1669 bp). Moreover the 40 genes in these regions consist of a single exon while a large majority of wasp genes are predicted to contain multiple exons. The atypical regions seemed to harbor virus structural genes since, in addition to p12-2, p53-1 and p53-2, they also contained p12-1 and another p12 homolog, designated p12-3. We therefore named them “IchnoVirus Structural Proteins Encoding Regions” (IVSPER, Figure 1). Strikingly, two IVSPERs were located respectively 3 kb upstream and 4 kb downstream the chromosomal form of an HdIV genome segment that is packaged in the particles (Figure 1). These two HdIV sequences (SH-BQ and SH-BR) did not show significant similarity to each other when compared at the nucleotide level but both contained a member of the N-gene family that is conserved among IVs [5]. Interestingly, an N-gene was also present in each of the three IVSPERs (N-1, N-2 and N-3; Figure 1).
As H. didymator IVSPERs contain genes (p12 and N) that are related to coding sequences known to be present in packaged CsIV or HdIV DNA, we examined the possibility that IVSPERs may be part of the packaged genome as well. PCR experiments were thus conducted using specific primers and template consisting of either wasp genomic DNA or DNA extracted from purified HdIV particles. Using HdIV particle DNA, no amplification could be obtained with several primer pairs corresponding to sequences scattered along the IVSPERs, whereas amplification products could be obtained with primers specific for the SH-BQ viral segment (Table S2). This showed that the IVSPERs are not packaged in the particles but are expressed in calyx cells at the time of virus production.
To confirm that the identified p12 and p53 genes encoded structural components of HdIV particles and to assess the possibility that IVSPERs contained other structural genes, proteins extracted from purified HdIV particles were analyzed by mass spectrometry (LC MS/MS). After separation of HdIV proteins by SDS-PAGE, more than 70 bands were detected, ranging from 10 to 250 kDa (Figure 2). Among them, the 16 most intense bands were selected and trypsin digested to produce peptides. Strikingly, comparison of peptides identified by LC MS/MS with translated coding sequences showed that 19 IVSPER predicted gene products were components of virus particles (Figure 2; Table S3). They included the p53-2 and p12-1 proteins and the product of the N-2 gene. Products of p53-1 and of other p12 genes were not detected, but could be present in the less intense bands (not analyzed by LC MS/MS) as other IVSPER proteins. Altogether the results obtained indicate that at least half of the IVSPER genes (19/40; Figure 3) encode virion structural proteins and that the IVSPERs constitute clusters of HdIV structural genes.
The 19 genes shown to encode components of the particles are expected to be transcribed in the tissue producing the particles, i.e., the calyx. To verify this prediction and to determine whether the other 21 IVSPER genes might also be involved in the production of virus particles we analyzed the expression of these genes by qRT-PCR. All the 25 IVSPER genes examined were found to be specifically transcribed in calyx cells at levels at least 13 times higher than in the ovarioles (Figure 3). In accordance with these results, blast similarity searches against sequence database generated from H. didymator ovarian cDNA libraries, using IVSPER gene sequences as queries, identified 26 different IVSPER-derived cDNAs (Table S1), thus verifying our initial prediction that cDNAs from genes involved in IV viriogenesis were present in the libraries. Altogether these results suggest that all IVSPER genes are likely to be involved in virus particle production, either directly by encoding structural proteins or indirectly by promoting their production.
The IVSPER gene products display no significant similarity to protein sequences deposited in public databases, and only one conserved domain has been identified in IVSPER proteins: a cyclin domain present in the U12 protein (Tables S3 and S4; Text S1). In addition, the U22 product shows weak similarity with a baculovirus P74 envelope protein (gi|48843584|). The presence of the P74 domain was confirmed when conserved structural signatures in IVSPER products were searched for using HHPred (Table S3; Text S1). The P74 protein is an envelope protein involved in the entry of baculovirus virions into midgut cells and is conserved among nudiviruses and bracoviruses. However the presence of a single gene is not sufficient to draw conclusions as to the nature of the IV ancestor; rather, the H. didymator IVSPER gene products appear to constitute a set of proteins specific to IVs.
In addition to the p12, p53 and N-gene families found in the IVSPERs, we identified members of four new gene families, named IVSP1 to IVSP4 (for “IchnoVirus Structural Protein”; Figure 1). Altogether members of these seven gene families represent 40% (16/40) of the IVSPER genes, and proteins within a given family display >60% sequence similarity (Table S5). The observation that IVSPERs share a combination of related genes suggests they may have originated from a common ancestor having this set of genes.
IVs are associated with species from the Campopleginae and Banchinae subfamilies of ichneumonid wasps. The different features of the virions and the fact that PDVs have not been recorded in species from several groups separating Campopleginae and Banchinae [24], suggest that two distinct ancestral wasp-virus associations may have arisen during the diversification of ichneumonid wasps. In this context, if one assumes that the associations in Campopleginae have a common origin, the genes encoding structural proteins expressed in H. didymator are predicted to be conserved in wasps from this subfamily. We thus searched for IVSPER homologs by sequencing cDNAs (4992 clones) generated from the ovaries of Tranosema rostrale (Campopleginae), which carries the ichnovirus TrIV. As observed for H. didymator, no significant similarities were found with known virus genes, except with those described in IVs [4]. Strikingly, a similarity search allowed the identification of 11 genes expressed in T. rostrale ovaries whose products display significant similarity (60 to 93% similarity) to those of H. didymator IVSPERs (Table S1): seven were homologs of genes shown to encode HdIV structural proteins (U1, U3, IVSP4-1 and 2, p12-1, U23, N-2) and four to other IVSPER genes (N-1, U10, U16, U19). Interestingly these genes were not identified in the packaged genome of TrIV [4], indicating that, like HdIV IVSPER genes, they reside in the wasp genome. These results strongly suggest that HdIV IVSPER genes are conserved among campoplegine wasps and point to a common origin of the set of IV structural genes.
Unlike the cluster of nudivirus-like genes involved in BV particle production, two IVSPERs are located in the vicinity of the integrated form of a viral DNA sequence packaged in the particles (Figure 1). This linkage could have a role in the coordinated expression of genes involved in IV virion production. To assess whether IVSPER DNA could be amplified with the packaged DNA we studied the level of IVSPER DNA during particle production using qPCR. The levels of nine genes chosen in the three IVSPERs and of packaged DNA (SH-BQ, Vinnexin gene) were measured in calyx cells from wasps just after their emergence, when particle production is highest and in adult wasp (24h hours after emergence). As shown in Figure 4, the results indicated that the nine IVSPER genes examined are amplified in calyx cells at a level comparable to that of the viral DNA packaged in the particles. It is noteworthy that the IVSPER-3 genes, which do not appear linked to a packaged DNA sequence, are also amplified. Relative to the levels measured in 2 h-old females, there was a coordinated drop in the amplification of both HdIV segment and IVSPER DNA in females one day after emergence (Figure 4), further confirming the existence of a direct correlation between the level of amplification of these two groups of genes in the calyx. Altogether these results indicate that IVSPERs have retained an important property of virus DNA: they are amplified during virus particle production.
Another link between IVSPERs and packaged viral DNAs is the phylogenetic relationship between IVSPER-2 and CsIV viral segment SH-C (Figure 5). The comparison of nucleotide sequences revealed important similarities which encompass 7014 nt in H. didymator IVSPER-2 and 5328 nt in CsIV SH-C. They consist in a succession of comparable (65 to 77% identity) and more divergent sequences (less than 10% similarity). The highest similarities concern regions containing coding sequences, and 5 homologs of the HdIV structural genes (including p12 gene) are encoded, in CsIV, by a viral segment. This suggests that CsIV SH-C and H. didymator IVSPER-2 have a common ancestor sequence and that during evolution, the CsIV segment has retained the ability to be encapsidated whereas the HdIV segment has lost this ability and is now expressed in the calyx but not packaged.
Because PDV packaged genomes lack typical viral genes, their relationship to conventional viruses has been a subject of debate. We recently identified genes encoding structural components of PDVs associated with braconid wasps, based on their mRNA expression in the ovaries. Present in the wasp genome and expressed specifically in the calyx, these structural protein genes resemble protein-coding genes of nudiviruses, a sister group of baculoviruses [12]. These data strongly suggest that PDVs from braconid wasps originated from a nudivirus. The same approach performed using the ovaries of H. didymator did not lead to the identification of coding sequences showing significant similarity to the core genes of a known virus. To overcome this problem, we conducted mass spectrometry analyses of purified virion proteins to identify genes encoding HdIV particle components.
We discovered that the proteins associated with HdIV particles are encoded by genes located in specialized regions of the wasp genome, the IVSPERs. A subset of 19 IVSPER gene products were identified as components of viral particles and the other IVSPER genes were shown to be highly expressed in the tissue where HdIV particles are produced, suggesting that IVSPER proteins contribute directly (as structural proteins) or indirectly to HdIV particle production. Thus, the IVSPERs clearly encode the protein machinery involved in HdIV viriogenesis. Consistent with this key role and the hypothesis that wasp-IV associations in this group have a common origin, IVSPER genes are conserved among IV-associated campoplegine wasps: in addition to the p12 and p53 genes first described in CsIV, 11 homologs of H. didymator IVSPER genes were found to be expressed in T. rostrale ovaries and four H. didymator IVSPER-2 genes have homologues in CsIV segment SH-C.
Analysis of the gene content of IVSPERs points to a relationship between some of the genes they contain and those packaged in virus particles, a situation that differs from that described for BVs where the packaged genome does not contain genes that are similar to those involved in particle production. More specifically, we found that IVSPERs contain members of the N-gene family, also present on HdIV segments and previously described in the packaged DNA of CsIV [5], Hyposoter fugitivus IV and TrIV [4]. The presence of related genes in IVSPER and packaged DNA, along with the absence of some CsIV genes including the p12 gene in the packaged HdIV genome may reflect the fact that different IV genomes are at different stages of their evolution.
Except for the eight proteins encoded by the p53, p12 and N-gene families, U12, which contains a cyclin domain, and U22, which displays a weak similarity with a baculovirus P74 protein, the other IVSPER gene products do not resemble any previously described protein. In particular, we did not find similarity with ascovirus sequences or structures, a finding that does not support the hypothesis that IVs have an ascovirus origin, as previously suggested [23]. However, the absence of conserved proteins among IV structural protein genes is not completely surprising since several sequencing programs focusing on viral genomes have led to similar findings. For example, the Mimivirus genome consists of 1262 putative open reading frames, among which only 10% exhibit significant similarity to proteins of known functions [25]. Similarly, in a comparison of the herpes virus infecting oysters and those infecting vertebrates, only the structure of the genome was found to be conserved [26].
Although they are not packaged in virus particles, IVSPERs are physically, functionally, and phylogenetically related to the packaged IV DNA and could thus be considered as an integral part of the IV genome. First, we have shown that IVSPERs are amplified in calyx cells during virus production at a level comparable to that measured for packaged segments. The genomic proximity and comparable amplification of IVSPERs and packaged segments strongly suggest they belong to common viral replication units, whereas the IVSPER-3, not in the close vicinity of an HdIV segment and flanked by wasp genes, may constitute an independent unit. A second source of evidence for a close relationship between IVSPERs and packaged IV DNA is the synteny between IVSPER-2 and CsIV segment SH-C, suggesting a common origin of these DNA regions. A simple explanation could be that during evolution of the H. didymator lineage, IVSPER-2 (but not the corresponding region of CsIV) may have lost the ability to be packaged. Conceptually, IVSPERs could thus be considered as elements of the IV genome that no longer require encapsidation. Due to the exclusive vertical transmission of the IV chromosomally integrated genomes, structural protein genes are not required on the viral segments injected into the host, but their amplification may have been selected for to allow production of high levels of virion structural components in the calyx. This appears to differ from the situation described for braconid wasps where the high production of structural proteins is presumed to be effected by a nudiviral RNA polymerase expressed in the calyx [13].
In addition to their functional role in particle production, IVSPERs display other notable features, including (i) their high exon density relative to regions of the wasp genome containing cellular genes, and (ii) the simple structure of their genes (made of a single exon), which is more typical of virus genes than of wasp genes, which more often consist of multiple exons. Strikingly, this organization resembles that of the “nudivirus cluster” in the genome of the wasp Cotesia congregata, which is thought to constitute a remnant of the ancestral nudivirus genome integrated into the genome of the ancestor of BV-associated wasps. This cluster contains 10 genes made of a single exon, is densely packed (exon density: 50%) and the products of five of its genes display similarities to conserved proteins of nudiviruses. The similar organization of IVSPERs suggests that they constitute, like the nudivirus cluster, remnants of foreign DNA integrated into the wasp genome. Altogether, IVSPER genomic structure, gene content, replication properties and involvement in particle production suggest they originated from a virus, belonging to an uncharacterized or extinct group. The nature of the ancestral virus genome could not be established using viral sequences currently available in public databases: sequences of the ancestor group are missing or IV sequences have diverged to such an extent that a relationship is undetectable. However it is interesting to note that IVSPERs contain a combination of related genes that are members of seven families. We hypothesize (Figure 6) that the IV ancestor possessed a member of each gene family. After duplications, different copies of this ancestral genome may have diversified, leading to the current IVSPERs, containing both common and specific genes that cooperate to produce HdIV particles.
Clearly IVs associated to campoplegine wasps originate from an entity that differs from that of the nudiviral BV ancestor, demonstrating that the association between wasps and viruses arose at least twice during the evolution of parasitic wasps. The use of PDVs by two groups of wasps to deliver genes into the host thus represents an example of convergent evolution. Recently a PDV from a banchine wasp has been described and was proposed to belong to a third group, based on its unusual features, in particular the morphology of the particles and the content of its packaged genome [3]. It will be of interest to determine whether this association constitutes a third event of viral capture by parasitoid wasps. Given that these associations between viruses and eukaryotic organisms have only been described for parasitic wasps, one may ask whether they are specific to these insects because of their unusual life-style, i.e. larvae living within the body of a caterpillar, or whether they occur more commonly. One might predict that virus domestication allowing gene transfer has arisen several times in the course of evolution in situations where interactions between organisms are both intimate and antagonistic.
Hyposoter didymator wasps were reared in laboratory and Tranosema rostrale wasps were obtained from the field as described [1], [6].
The libraries were constructed as described [12]. Briefly, ovaries were dissected from H. didymator pupae of different developmental stages and total RNA was extracted using the Qiagen RNeasy Mini Kit. The cDNA synthesis was performed using the Creator SMART cDNA Library Construction Kit (Clontech) from 2 µg of total RNA. The cDNAs were cloned into the pDNR-LIB vector (Clontech). A total of 5636 clones were sequenced from the 5′-end. The sequences cleaned from vector stretches were subjected to clustering using the TIGR software TGI Clustering tool (TGICL), as described [27]. They corresponded to 597 clusters (containing more than one sequence) and 1359 singletons, and thus to 1956 non redundant sequences. To identify similarities with known proteins, the sequences were searched using the Blastx algorithm against a local non-redundant protein database (NCBI, release july 15, 2008) with no cut-off for the E-value.
Ovaries were dissected from adult wasps shortly after emergence, and total RNA was extracted using the RNeasy Mini Kit (Qiagen). 250 ng of RNA was treated with amplification-grade DNAse I (Invitrogen) [28] and reverse transcribed using an oligo dT primer, followed by a second strand synthesis and ligation of an adapter. Using a distal adapter primer and the oligo dT primer, the cDNAs were amplified and then ligated into the pGEM-T-Easy vector (Promega). A total of 4992 colonies were selected and sequenced from both ends at the Genome Sciences Centre, BC Cancer Agency (Vancouver, Canada).
To obtain a H. didymator BAC library, high molecular weight DNA was extracted from larval nuclei and partially digested with HindIII. The fragments thus obtained were ligated into the pBeloBAC11 vector. High-density filters (18,432 clones spotted twice on nylon membranes) were screened using specific 35-mer oligonucleotides. Positive BAC clones were analyzed by fingerprint. One genomic clone was selected for each probe and sequenced by a shotgun method. Coding sequences were predicted using Kaikogas (http://kaikogaas.dna.affrc.go.jp/). A Blastn similarity search against the ovary EST libraries was performed with no cut-off for the E-value. The sizes of the intergenic regions within the IVSPER and other available genomic regions (over 1.40 Mb of wasp genome) were compared using a Student T-test (t = 4.552, df = 49.238, p-value = 3.497e-05).
Total RNA from H. didymator calyx and ovariole fractions was extracted using the Qiagen RNeasy Mini Kit and treated with the Turbo DNAse kit (Ambion). First strand cDNA was synthesized from 3 to 5 µg of RNA using the Invitrogen Superscript III Reverse Transcriptase. Absence of DNA contamination and first-strand cDNA synthesis were verified by PCR with primers specific to Elongation Factor EF1-α (Table S6). The qPCR was performed using the Applied Biosystem 7000 sequence detection system in 96-wells PCR plates (ABgene) that comprised triplicates of 2 or 3 biological replicates. Primer pairs (Table S6) were designed using the Primer ExpressTM software (Applied Biosystems) to generate 51 bp amplicons. The final qPCR reaction volume of 25 µl contained an amount of cDNA equivalent to 20 ng of total RNA, 0.4 µM of primer pairs, and the Platinum SYBR Green qPCR SuperMix-UDG with ROX (Invitrogen). The dissociation curve method was applied to ensure the presence of a single specific PCR product.
The data were analyzed either with the classical CT method or with an alternative assumption-free method [29]. The latter gives the relative N0 values corresponding to the initial transcript levels of each gene in a given tissue. Four endogenous reference genes (EF1-α, ribosomal L55, cytochrome VIIC and histone H1) were used for normalization.
A first purification was performed by filtration from 300 dissected ovaries as described [30], and the viral particles were further purified on a sucrose gradient (20–50%). Centrifugation was performed at 154,324 g during 1.5 h at 4°C in a Beckman L7 ultracentrifuge, using a SW-41 swing-out rotor. Viral fractions were collected, diluted in saline buffer (PBS) and submitted to a second centrifugation (154,324 g during 1 h at 4°C) in order to pellet the viral particles. The resulting pellet was re-suspended in PBS and submitted to dialysis during two days at 4°C. The presence of viral particles was verified by TEM followed by SDS-PAGE.
Gel electrophoresis was carried out as described [31] on a 12% acrylamide gel. After gel staining with colloidal blue (Fermentas), gel slices were cut out, washed with 50% acetonitrile, 50 mM NH4HCO3 and incubated overnight at 25°C (with shaking) with 15 ng/µl trypsin (Gold) in 100 mM NH4HCO3. The tryptic fragments were extracted with 1.4% (v/v) formic acid. Samples were analyzed online using a nanoESI LTQ-OrbitrapXL mass spectrometer (Thermo Fisher Scientific) coupled with an Ultimate 3000 HPLC (Dionex). Details are given in Text S1. Data were acquired using Xcalibur software (v 2.0.7, Thermo Fisher Scientific). Identification of proteins was performed using the Mascot v 2.2 algorithm (Matrix Science Inc.), by searching against the entries of H. didymator sequences. The data submission was performed using ProteomeDiscoverer v 1.0 (Thermo Fisher Scientific). Peptides with scores greater than the identity score (p<0.05) were considered as significant. All spectra were manually validated for proteins identified with less than three different peptides.
Presence of selected genes in the HdIV packaged genome was verified by PCR using gene-specific primers (Table S6). Templates consisted of either 20 ng of viral DNA or 100 ng genomic H. didymator DNA. HdIV DNA was extracted from viral particles purified on a sucrose gradient (see above). Genomic wasp DNA was extracted with the Promega Wizard Genomic DNA Purification System. The 50 µl reactions were conducted using the GoTaq Flexi DNA Polymerase (Promega) following standard PCR protocol.
|
10.1371/journal.ppat.1000080 | Varicellovirus UL49.5 Proteins Differentially Affect the Function of the Transporter Associated with Antigen Processing, TAP | Cytotoxic T-lymphocytes play an important role in the protection against viral infections, which they detect through the recognition of virus-derived peptides, presented in the context of MHC class I molecules at the surface of the infected cell. The transporter associated with antigen processing (TAP) plays an essential role in MHC class I–restricted antigen presentation, as TAP imports peptides into the ER, where peptide loading of MHC class I molecules takes place. In this study, the UL49.5 proteins of the varicelloviruses bovine herpesvirus 1 (BHV-1), pseudorabies virus (PRV), and equine herpesvirus 1 and 4 (EHV-1 and EHV-4) are characterized as members of a novel class of viral immune evasion proteins. These UL49.5 proteins interfere with MHC class I antigen presentation by blocking the supply of antigenic peptides through inhibition of TAP. BHV-1, PRV, and EHV-1 recombinant viruses lacking UL49.5 no longer interfere with peptide transport. Combined with the observation that the individually expressed UL49.5 proteins block TAP as well, these data indicate that UL49.5 is the viral factor that is both necessary and sufficient to abolish TAP function during productive infection by these viruses. The mechanisms through which the UL49.5 proteins of BHV-1, PRV, EHV-1, and EHV-4 block TAP exhibit surprising diversity. BHV-1 UL49.5 targets TAP for proteasomal degradation, whereas EHV-1 and EHV-4 UL49.5 interfere with the binding of ATP to TAP. In contrast, TAP stability and ATP recruitment are not affected by PRV UL49.5, although it has the capacity to arrest the peptide transporter in a translocation-incompetent state, a property shared with the BHV-1 and EHV-1 UL49.5. Taken together, these results classify the UL49.5 gene products of BHV-1, PRV, EHV-1, and EHV-4 as members of a novel family of viral immune evasion proteins, inhibiting TAP through a variety of mechanisms.
| Herpesviruses have the conspicuous property that they persist for life in the infected host. This is also the case for varicelloviruses, a large subfamily of herpesviruses with representatives in humans (varicella zoster virus or VZV), cattle (bovine herpesvirus 1 or BHV-1), pigs (pseudorabies virus or PRV), and horses (equine herpesvirus or EHV type 1 and 4), among many others. Cytotoxic T-lymphocytes play an important role in the protection against viral infections, which they detect through the recognition of virus-derived peptides, presented in the context of MHC class I molecules at the surface of the infected cell. The transporter associated with antigen processing (TAP) plays an essential role in this process, as TAP imports peptides into the compartment where peptide loading of the MHC class I molecules takes place. In this study, we show that the UL49.5 proteins of BHV-1, PRV, EHV-1, and EHV-4 all block the supply of peptides through the inhibition of TAP, but that the mechanisms employed by these proteins to inhibit TAP function exhibit surprising diversity. VZV UL49.5, on the other hand, binds to TAP, but does not interfere with peptide transport. Our study classifies the UL49.5 proteins of BHV-1, PRV, EHV-1, and EHV-4 as members of a novel family of viral immune evasion proteins, inhibiting TAP through a variety of mechanisms.
| Evolving under the selective pressure of the host immune system, herpesviruses have developed countermeasures to prevent recognition of infected cells by cytotoxic CD8+ T lymphocytes (CTLs). CTLs recognize viral antigens presented as peptides bound to major histocompatibility complex (MHC) class I molecules at the surface of infected cells. Herpesviruses in particular have acquired diverse mechanisms to inhibit antigen presentation in the context of MHC class I molecules, thereby escaping from elimination by CTLs [1]–[4].
Most peptides presented by MHC class I molecules are transported into the endoplasmic reticulum (ER) lumen by the transporter associated with antigen processing, TAP. TAP is a heterodimer composed of TAP1 and TAP2 subunits and belongs to the ATP-binding cassette family of transporters [5],[6]. TAP translocates peptides across the ER membrane via a conformational transition that is energized by the hydrolysis of ATP. TAP is part of the MHC class I peptide-loading complex that also contains tapasin, MHC class I heavy and light chains, and several auxiliary proteins including calreticulin and ERp57 [5], [7]–[10].
Several herpesviruses have acquired mechanisms to interfere with TAP function. Interestingly, inhibition of TAP transport is achieved through different strategies, exerted by unique gene products. Although the varicellovirus bovine herpesvirus 1 (BHV-1) and the simplexviruses herpes simplex virus type 1 and 2 (HSV-1 and -2) all belong to the subfamily of alphaherpesviruses, they block TAP through proteins that have an entirely different structure and mode of action. The inhibition of TAP by BHV-1 relies on the UL49.5 (Unique Long 49.5) gene product, a type I transmembrane protein of 75 amino acids [11]. Inactivation of TAP by UL49.5 involves two events: the arrest of the peptide transporter in a translocation-incompetent state and the proteasomal degradation of both subunits of TAP [11]. In contrast, the ICP47 proteins of HSV-1 and -2 are soluble cytosolic proteins acting as high-affinity competitors for peptide binding to TAP [12]–[18]. Within the subfamily of betaherpesviruses, human cytomegalovirus (HCMV) was found to encode a protein, US6, that inhibits TAP function by reducing the interaction of ATP with TAP [19]–[24]. The murine gammaherpesvirus-68 (MHV-68) encodes the mK3 protein that acts as a ubiquitin ligase linking MHC class I molecules and TAP to the ubiquitin/proteasome degradation pathway [25]–[35]. Recently, the BNLF2a protein of Epstein-Barr virus (EBV) and of related primate gamma-1 herpesviruses has been characterized as a potent TAP inhibitor, preventing the binding of both peptides and ATP to TAP [36].
Homologs of UL49.5 (commonly known as glycoprotein N; gN) are encoded by every alpha-, beta- and gammaherpesvirus sequenced to date [37]–[39]. The UL49.5 genes are all predicted to encode a type I membrane protein with a putative cleavable signal sequence. The UL49.5 proteins interact with another herpesvirus protein, glycoprotein M (gM), with which they form a disulfide-linked heterodimer through a conserved cysteine residue within their ER-luminal/extracellular domain [37], [40]–[44]. Nevertheless, the amino acid sequences of UL49.5 proteins demonstrate considerable heterogeneity, even among varicellovirus UL49.5 proteins (Fig. 1). The only exceptions are EHV-1 and EHV-4 UL49.5, which differ by only seven amino acid residues. Thus, at this moment, it is impossible to predict on the basis of amino acid sequence whether any of these proteins have the same capacity to inhibit TAP that was found for BHV-1 UL49.5. The UL49.5 gene products of HSV-1, HSV-2, HCMV, and EBV fail to block TAP, indicating that not all UL49.5 molecules act as inhibitors of TAP [11],[37].
In this study, the effects on TAP function were assessed in more detail for UL49.5 encoded by various members of the genus Varicellovirus. The UL49.5 proteins of BHV-1, PRV, EHV-1, and EHV-4 were found to down-regulate MHC class I cell surface expression through TAP inhibition. Their ability to block TAP was observed in cells of the relevant host species, as well as in human cells. Using UL49.5 deletion mutants of BHV-1, PRV and EHV-1, it was shown that the UL49.5 proteins of these viruses are responsible for the inhibition of TAP-dependent peptide transport. The UL49.5 homologs of canine herpesvirus (CHV) and VZV did not affect MHC class I surface expression. BHV-1 UL49.5 strongly reduces the steady state protein levels of TAP in both bovine and human cells, whereas the UL49.5 proteins of EHV-1, EHV-4 or PRV were not observed to have this capacity. Interestingly, the EHV-1 and EHV-4 UL49.5 homologs interfere with the binding of ATP to TAP, a function that is not influenced by BHV-1 or PRV UL49.5. The UL49.5 proteins of PRV and EHV-1 arrest TAP in a translocation-incompetent state, a property that is shared with BHV-1 UL49.5. Thus, the BHV-1, PRV, EHV-1 and EHV-4-encoded UL49.5 proteins all induce a similar phenotype, i.e. inhibition of peptide transport, but their modes of action demonstrate a surprising diversity.
To evaluate the TAP-inhibiting capacity of the UL49.5 proteins encoded by the varicelloviruses PRV, EHV-1, EHV-4, CHV and VZV, cell lines of the relevant host species were transduced using a retrovirus-based gene delivery system to express the corresponding UL49.5 proteins. Down-regulation of MHC class I expression by the UL49.5 gene products was evaluated using flow cytometry. In cells expressing UL49.5 of BHV-1, PRV, EHV-1 and EHV-4, MHC class I surface expression was reduced (Fig. 2A). The UL49.5 proteins of CHV and VZV failed to down-regulate MHC class I surface expression. These results indicate that UL49.5 of BHV-1, PRV, EHV-1 and EHV-4 interfere with MHC class I-restricted antigen presentation.
To investigate whether the observed down-regulation of MHC class I cell surface expression relies on the inhibition of TAP, species-specific cell lines stably expressing these UL49.5 homologs were evaluated for TAP-dependent peptide transport. The UL49.5 proteins of BHV-1, PRV, EHV-1 and EHV-4 strongly inhibited TAP activity in the corresponding natural host cell lines (Fig. 2B). Despite the absence of a detectable reduction in cell surface MHC class I levels (Fig. 2A), some inhibition of TAP-dependent peptide transport was observed in canine cells expressing the CHV UL49.5 protein (Fig. 2B). Apparently, the inhibition of TAP by CHV UL49.5 was insufficient to observe MHC class I downregulation at the cell surface. VZV UL49.5 had no significant effect on TAP activity. Thus, although the amino acid sequences of the UL49.5 proteins of BHV-1, PRV, and EHV-1/EHV-4 demonstrate considerable variation (Fig. 1), their ability to inhibit TAP was found to be a common property of these varicellovirus gene products.
VZV infection has been shown to cause down-regulation of MHC class I expression at the cell surface [45]–[47]. This phenotype could not be reproduced by the VZV-encoded UL49.5 protein when expressed individually (Fig. 2A). To examine whether the absence of MHC class I down-regulation by VZV UL49.5 is due to a loss of the interaction of the viral protein with the TAP complex, TAP was immunoprecipitated from VZV UL49.5-expressing MJS cells that were solubilized in the presence of the mild detergent digitonin. The resulting protein complexes were separated by SDS PAGE and analyzed for the presence of UL49.5 by immunoblotting. Surprisingly, VZV UL49.5 was found to interact with the TAP complex (Fig. 3A).
Although VZV UL49.5 associates with TAP, this appears to be insufficient to inhibit peptide transport effectively (Fig. 2A and B). VZV UL49.5 could, however, interfere with peptide-loading and MHC class I-restricted antigen presentation in a different way. For instance, the US3 protein encoded by human cytomegalovirus binds both tapasin and TAP, without having an effect on TAP function. Instead, US3 impairs tapasin-dependent peptide loading and optimization of the MHC class I peptide cargo [48],[49]. To investigate whether VZV UL49.5 inhibits MHC class I-mediated antigen presentation via a mechanism similar to that of US3, functional T cell assays were performed using a panel of human leukocyte antigen (HLA)-A1 and HLA-A2-restricted CTL clones. It is known that especially HLA-A1-restricted peptide presentation strongly depends on the function of tapasin. Antigen-presenting phytohemagglutinin (PHA)-treated T cell blasts and the melanoma cell-line Mel518 were transduced to express the UL49.5 proteins of VZV and BHV-1. While the presence of BHV-1 UL49.5 greatly reduced specific lysis of the PHA-blasts by CTLs, the expression of VZV UL49.5 had no detectable effect (Fig. 3B). This was observed for HLA-A1 and HLA-A2-restricted CTL clones. VZV UL49.5-expressing and control target cells induced IFNγ production by the CTLs, while reduced IFNγ production was observed when BHV-1 UL49.5 was expressed by the target cells (Fig. 3C). This reflects effective inhibition of CTL recognition by BHV-1 but not VZV UL49.5. Thus, despite the interaction of the VZV UL49.5 protein with the peptide-loading complex, no interference with MHC class I-restricted antigen presentation could be detected.
Having observed that the UL49.5 proteins of BHV-1, PRV, EHV-1, and EHV-4 interfere with MHC class I-restricted antigen presentation when expressed individually, we next investigated whether the various UL49.5 proteins are responsible for TAP inhibition during infection with BHV-1, PRV, or EHV-1. Peptide transport activity was examined in natural host cells infected with wild type viruses or with the corresponding recombinant viruses lacking a functional UL49.5 gene [41],[43]. Whereas the wild-type viruses effectively blocked peptide transport, this inhibition was not observed in cells infected with the mutant viruses lacking UL49.5 (Fig. 4). These findings indicate that during infection with BHV-1, PRV and EHV-1, the UL49.5 gene products of these viruses are responsible for the inhibition of peptide translocation by TAP observed in virus-infected cells previously [50]–[52].
Next, the mechanism of TAP inhibition by the various UL49.5 proteins was investigated. Expression of BHV-1 UL49.5 strongly reduced TAP1 and TAP2 protein levels in human MJS cells [11]. It was shown that the cytoplasmic domain of UL49.5 is required for mediating proteasome-dependent degradation of TAP. To investigate whether BHV-1 UL49.5 has a similar mode of action in natural host cells, bovine MDBK cells were infected with wild type BHV-1 or a recombinant virus expressing a UL49.5 protein that lacks its cytoplasmic domain (UL49.5Δtail). Steady state protein levels of bovine TAP were evaluated by immunoblotting. Whereas bovine TAP was readily detectable in uninfected MDBK cells, it was no longer observed in cell lysates from wild-type BHV-1 infected cells (Fig. 5; upper panel, compare lanes 1 and 2). Interestingly, in cells infected with the recombinant virus expressing the UL49.5Δtail mutant, TAP1 steady state levels were not affected (compare lanes 2 and 3). As a control, α-tubulin was consistently detected in all samples (Fig. 5, middle panel). Immunoprecipitation of UL49.5 from the infected cells confirmed the expression of the wild-type and recombinant proteins (Fig. 5; lower panel). These findings indicate that the degradation of TAP by UL49.5 previously observed in human cells also occurs in bovine cells. In addition, like in human cells, the cytoplasmic domain of UL49.5 is critical to TAP degradation in the natural host cells.
To further address the molecular basis of TAP inhibition mediated by PRV, EHV-1, EHV-4, and CHV UL49.5, these proteins were stably expressed in human melanoma (MJS) cells. Like BHV-1 UL49.5, the PRV and EHV-1 UL49.5 proteins were capable of blocking human TAP (Fig. 6A). CHV UL49.5 did not inhibit peptide transport in human cells, while in canine cells some reduction in TAP activity was observed without a reduction of MHC class I surface expression (Fig. 2A and B). Expression of BHV-1 UL49.5 in MJS cells resulted in reduced TAP1 and TAP2 protein levels, which is in accordance with previous observations (Fig. 6B; compare lanes 1 and 2) [11]. In contrast, expression of the UL49.5 homologs of PRV and EHV-1 did not affect TAP1 and TAP2 steady state levels in MJS cells (Fig. 6B; lanes 3 and 4). These findings indicate that the UL49.5 homologs of PRV and EHV-1 inhibit peptide transport by TAP through a different mechanism than by mediating degradation of TAP.
The translocation of peptides into the ER lumen is initiated by the association of peptides with the peptide-binding site of TAP [53]. To investigate whether the inhibition of peptide transport by PRV and EHV-1 UL49.5 involves blocking of peptide binding to TAP, microsomes were isolated from MJS cells expressing PRV and EHV-1 UL49.5. Microsomes were incubated with a 125I-labeled reporter peptide (Fig. 7). At all concentrations tested, the peptide-binding capacity (Bmax) was similar for microsomes prepared from control cells and from cells expressing the PRV or EHV-1 UL49.5 proteins. Most importantly, the binding affinity (Kd) for the peptides was not changed by the viral inhibitors, demonstrating preservation of the peptide binding site of the TAP complex. This has also been observed for BHV-1 UL49.5 [11] and indicates that inhibition of TAP-mediated peptide transport by the BHV-1, PRV, and EHV-1 UL49.5 proteins does not rely on interference with peptide binding.
Since ATP-binding and hydrolysis are required to energize peptide translocation by TAP [54]–[56], it was investigated whether the expression of the PRV, EHV-1, and EHV-4 UL49.5 proteins affected binding of ATP to TAP. Previous experiments indicated that the BHV-1 UL49.5 protein did not influence the interaction of ATP with TAP [11]. The ATP-binding capacity of TAP in lysates from MJS cells (control) was compared to the binding in lysates from MJS cells stably expressing UL49.5 of PRV, EHV-1 or EHV-4, or the HCMV-encoded US6 protein. US6 is known to strongly inhibit ATP-binding to TAP [22],[24]. Cell lysates prepared in the presence of the mild detergent digitonin were incubated with ATP-agarose beads. Proteins bound to the ATP-agarose (Fig. 8; pellet “P”) were eluted from the beads with EDTA and displayed next to the unbound supernatant fractions (Fig. 8; “S”). TAP1 and TAP2 were detected by immunoblotting.
PRV UL49.5 did not alter the binding of ATP to TAP1 or TAP2 (Fig. 8A; compare lanes 2 and 4). As expected, the expression of US6 completely abolished the interaction of ATP with TAP (Fig. 8A; lane 6). These data show that TAP retains the capacity to bind ATP in the presence of PRV UL49.5.
In EHV-1 and EHV-4 UL49.5-expressing cells, neither TAP1 nor TAP2 could be detected in the ATP-agarose fraction (Fig. 8B; compare lane 2 with lanes 4 and 8). Since the C-terminus of UL49.5 is exposed in the cytosol, this domain might be responsible for the inhibition of ATP-binding to the nucleotide-binding domains of TAP. To evaluate whether the C-terminus of UL49.5 blocks ATP-binding to TAP, a truncated form of EHV-1 UL49.5 lacking the cytoplasmic domain was constructed and expressed in MJS cells. The EHV-1 UL49.5Δtail recombinant still interfered with ATP-binding to TAP (Fig. 8B; lane 6). When the association of wild type or mutant EHV-1 UL49.5 with TAP was disrupted by lysis of the cells in NP-40, the ability of TAP1 and TAP2 to bind to the ATP-agarose was restored (Fig. 8C, lanes 4 and 6; also compare Fig. 8B lanes 4 and 6 with Fig. 8C lanes 4 and 6, respectively). These results indicate that the EHV-1 UL49.5 protein is capable of interfering with the recruitment of ATP by human TAP independent of the cytoplasmic domain of UL49.5.
The ability of EHV-1 and EHV-4 UL49.5 to interfere with the binding of ATP to equine TAP was assessed in E. derm cells (data shown for EHV-4). Like human TAP (Fig. 8B lane 8), equine TAP2 was not able to bind ATP-agarose in the presence of EHV-4 UL49.5 (Fig. 8D; compare lanes 2 and 4). When the experiment was performed in the presence of NP-40, the ability of equine TAP2 to bind ATP was restored (Fig. 8D; compare lanes 4 and 6). These results indicate that the UL49.5 proteins of EHV-1 and EHV-4 inhibit human and equine TAP through similar mechanisms, rendering both human and equine TAP molecules incapable of recruiting ATP.
To obtain further insight into the strategies used by EHV-1 and PRV UL49.5 to block TAP transport, Fluorescence Recovery After Photobleaching (FRAP) assays were performed. With this technique, conformational changes of TAP that occur during peptide translocation can be indirectly visualized by measuring the lateral mobility of green fluorescence protein (GFP)-tagged TAP within the ER membrane. It has been shown that the lateral mobility of TAP is inversely proportional to its activity, as peptide-transporting TAP molecules diffuse at a slower rate than inactive, closed TAP complexes [57]. In the absence of ATP, the translocation cycle cannot be initiated and consequently TAP will have a closed, more compact conformation. In agreement with this, depletion of ATP results in increased mobility of TAP in the ER membrane (Fig. 9; control samples, compare black and grey bars). The complex can be trapped in the active conformation by adding long side chain peptides (l.s.c.p.). These peptides bind to TAP, but cannot be translocated over the ER membrane, which results in a retained open conformation and therefore a slow diffusion rate of TAP in the ER membrane [57] (Fig. 9; control samples, white bar).
Expression of EHV-1 and PRV UL49.5 results in a decreased mobility of TAP (Fig. 9; compare untreated samples/black bars). Whereas the diffusion rate of TAP increased considerably in ATP-depleted control cells, only a slight increase in TAP mobility was detected upon ATP depletion in the UL49.5-expressing cells (grey bars). The failure of TAP to respond to ATP depletion in the EHV-1 UL49.5 cells is in agreement with the observation that this protein interferes with ATP binding to TAP (Fig. 8B). Although ATP can still bind to TAP in the presence of PRV UL49.5 (Fig. 8A), ATP depletion induces only a minor change in TAP mobility in the PRV UL49.5 cells (Fig. 9). Apparently, the presence of PRV UL49.5 prohibits conformational transitions that normally follow ATP-binding.
In the presence of the UL49.5 proteins, l.s.c.p. were also unable to induce conformational changes within the TAP complex (Fig. 9). Since peptides can still bind to TAP in the presence of EHV-1 and PRV UL49.5 (Fig. 7), the failure of l.s.c.p. to induce conformational changes again suggests that the UL49.5 proteins arrest TAP in a translocation-incompetent state.
This study identifies the UL49.5 proteins of BHV-1, PRV, EHV-1, and EHV-4 as members of a novel class of viral immune evasion proteins. The UL49.5 gene products interfere with MHC class I antigen presentation by blocking the supply of antigenic peptides in the ER lumen through inhibition of TAP. Within the UL49.5 family of TAP inhibitors, heterogeneity is observed with respect to the mechanisms that underlie TAP inhibition. Whereas BHV-1 UL49.5 targets TAP for proteasomal degradation [11], PRV and EHV-1 UL49.5 do not diminish the steady state levels of TAP1 or TAP2. Interestingly, EHV-1 and EHV-4 UL49.5 interfere with the binding of ATP to TAP, a function that is not influenced by BHV-1 or PRV UL49.5. All TAP-inhibiting UL49.5 proteins arrest the transporter complex in a translocation-incompetent state.
UL49.5 homologs are encoded by all Herpesviridae analyzed to date [38]. However, the TAP-inhibiting capacities of these proteins appear to be restricted to certain members of the genus Varicellovirus. Members of this virus genus have co-evolved with their respective host species [39]. Viruses of even-toed ungulates or Artiodactyla like BHV-1 and PRV co-evolved with cattle and pigs; viruses of odd-toed ungulates or Perissodactyla (EHV-1 and EHV-4) with horses; the carnivore viruses FHV-1 and CHV with cats and dogs, and the Old World primate virus VZV with humans [39] (Fig. 10). The identification of the UL49.5 proteins encoded by BHV-1, PRV, EHV-1, and EHV-4 as members of the UL49.5 family of TAP inhibitors suggests that more UL49.5 proteins with this property may be found in varicelloviruses of even- and odd-toed ungulate hosts. Considering the shared evolution of (herpesviruses from) carnivores and (herpesviruses from) odd-toed ungulates [39], CHV UL49.5 was expected to inhibit TAP as effectively as EHV-1 and EHV-4 UL49.5. However, the reduction of TAP-dependent peptide transport caused by CHV UL49.5 was very moderate compared to the inhibition by the other TAP-inhibiting UL49.5 proteins. The identification of the UL49.5 domains contributing to TAP inhibition will provide more insights into these differences.
VZV infection of human cells results in reduced expression of MHC class I at the cells surface [45]–[47]. The VZV ORF66-encoded serine-threonine protein kinase has been shown to be one of the VZV proteins contributing to MHC class I down-regulation in VZV-infected cells [47]. However, a VZV recombinant lacking a functional ORF66 product still causes down-regulation of MHC class I surface expression, indicating that additional modulators of MHC class I-restricted antigen presentation are encoded by VZV. The observed down-regulation of MHC class I surface expression on VZV-infected cells [45]–[47] is not induced by UL49.5 when expressed individually. Despite the observed interaction between VZV UL49.5 and the peptide-loading complex, this protein alone did not block peptide transport by TAP and it had no effect on antigen recognition by HLA-A1 and HLA-A2-restricted CTL clones. As VZV occupies a somewhat isolated position in the phylogenetic tree of varicelloviruses (Fig. 10), it seems likely that evolutionary divergence has influenced VZV to acquire a separate mechanism to interfere with MHC class I-restricted antigen presentation. Alternatively, UL49.5 might co-operate with another unidentified VZV-encoded protein in order to reduce antigen presentation by MHC class I molecules. During virus infection, UL49.5 can be found in a complex with glycoprotein M (gM). However, the co-expression of VZV UL49.5 and glycoprotein M has no effect on the expression of MHC class I molecules at the cell surface [47], indicating that gM does not act as a modulator of UL49.5 with respect to TAP inhibition.
Interaction with the conserved viral membrane glycoprotein M appears to be a common property of all UL49.5 homologs, as is the presence of a single cysteine residue in their ER-luminal/extracellular domain [41]–[44]. This cysteine residue is involved in the interaction of UL49.5 with gM, with which it forms a disulfide-linked heterodimers [37],[40]. The complex of UL49.5 and gM is implicated in virion maturation and membrane fusion processes [58],[59]. Interestingly, the interaction of BHV-1 UL49.5 with gM interferes with its capacity to block TAP [42]. Nevertheless, UL49.5 blocks peptide transport by TAP in BHV-1-infected cells. This may be explained by the fact that UL49.5 is expressed prior to and in excess of the early-late gM [42].
Interference with TAP-mediated peptide transport is an effective way of reducing CTL recognition and is used by several other herpesviruses, including HSV-1 and -2, HCMV, MHV-68, and EBV [12]–[36]. Compared to the other herpesvirus-encoded TAP inhibitors, the cross-species activity of UL49.5 proteins is remarkable. Except for CHV UL49.5, the UL49.5 proteins of BHV-1, PRV, EHV-1, and EHV-4 all exhibit the ability to target human TAP. In addition, BHV-1 UL49.5 inhibits peptide transport by murine [60], rat, equine, and porcine TAP (D.K.L. and M.V., unpublished observations). Human, porcine, bovine, and rodent TAP1 and TAP2 demonstrate a substantial degree of amino acid identity (70–80%) [61]. Thus, the ability of UL49.5 proteins to act across species barriers most likely relies on structural homology within the TAP domains critically involved in UL49.5-TAP interaction. Apparently, this is less so for the domains within TAP that are targeted by US6, mK3 and BNLF2a, whose actions seem to be restricted largely to the natural host species. BHV-1 UL49.5 reduces TAP protein levels in bovine, human, and murine cells, and also mediates degradation of human TAP in insect cells when co-expressed with UL49.5 [11],[60],[62], indicating conservation of the pathway involved in this degradation process.
The UL49.5 proteins exhibit unexpected differences in their mechanisms of TAP inhibition, despite their close evolutionary relatedness. The cytoplasmic domain of BHV-1 UL49.5 is essential for mediating degradation of both human and bovine TAP. EHV-1 and PRV UL49.5 have no influence on the stability of TAP. Apparently, degradation of TAP is facilitated by a yet unknown signal within the C-terminal domain of BHV-1 UL49.5, which is not present in the other homologs. Studies to identify the nature of this sequence motif are in progress.
The interaction of EHV-1 and EHV-4 UL49.5 with TAP blocks ATP binding to TAP. This feature distinguishes EHV UL49.5 from the other homologs studied. Interestingly, removal of the cytoplasmic domain of the EHV-1 UL49.5 protein did not restore the ability of TAP to bind ATP. Therefore, a direct interaction of EHV-1 UL49.5 with the cytosolic nucleotide binding domains of TAP is unlikely. Instead, the viral protein appears to arrest TAP in a translocation-incompetent state, incompatible with ATP-binding. This may resemble the type of structural change caused by HCMV US6 [23],[63]. US6, a type I transmembrane protein, interacts with the luminal side of the TAP transporter and blocks ATP-binding by prohibiting essential conformational rearrangements within TAP. The inability of the BHV-1 and PRV UL49.5 homologs to interfere with ATP-binding could be due to a slightly different conformational change induced by these proteins.
Based on the results presented in this study, the UL49.5 proteins encoded by BHV-1, PRV, EHV-1, and EHV-4 can be classified as a new family of TAP-inhibiting proteins. These proteins share the ability of inducing a conformational arrest of TAP, which results in impaired peptide transport and inhibition of MHC class I-restricted antigen presentation. In view of these joint features it is likely that the TAP inhibiting UL49.5 proteins originate from a common ancestral protein, which acquired this capacity earlier during evolution. The VZV UL49.5 protein may be a rudimentary form with respect to TAP inhibition, or it may have lost its TAP inhibitory capacity later on. Alternatively, it may require additional VZV proteins for the inhibition of TAP.
This study has revealed unexpected variation among UL49.5 proteins of varicelloviruses with respect to their mechanisms of TAP inhibition. These differences can be related to distinct evolutionary pathways of these varicelloviruses. The UL49.5 family of TAP-inhibiting proteins does not demonstrate any structural or functional similarity to TAP-inhibiting proteins encoded by other herpesviruses, for instance ICP47, US6, mK3, or BNLF2a. This diversity of TAP-inhibiting proteins acquired by distantly related members of the subgroups of alpha-, beta-, and gammaherpesviruses is remarkable and presents a striking example of functional convergent evolution. At the same time, this identifies TAP as an Achilles' heel of the MHC class I antigen presentation pathway. Inhibition of TAP has apparently provided a strong advantage to these herpesviruses during co-evolution with their hosts.
Purified viral DNA from BHV-1 strain Lam and CHV strain Eva (Animal Sciences Group, Lelystad, The Netherlands), PRV strain Kaplan [41], EHV-1 strain Ab-4 (kindly provided by J. Rola; National Veterinary Research Institute, Pulawy, Poland) and EHV-4 (kindly provided by R. de Groot; Dept. of Infectious Diseases and Immunology, Utrecht University, The Netherlands), and VZV (viral DNA extracted from patient material; kindly provided by E. Klaas, Leiden University Medical Center, Leiden, The Netherlands) were used as a template for polymerase chain reaction (PCR) amplification. PCR-reactions were performed under standard conditions using Pfu DNA polymerase (Invitrogen) and specific primers (Table 1) for amplification of the full length coding sequence of the UL49.5 genes of BHV-1 [11], PRV, EHV-1, EHV-4, CHV and VZV UL49.5. The sequences of the primers are based on published sequences found in the NCBI database, except for the sequence of the CHV primers (Haanes, E. and Rexann, F. ‘Recombinant canine herpesviruses’, patent number EPO910406, publication date 1997-08-21). To generate the EHV-1 UL49.5Δtail construct, primers (Table 1) were used to obtain a PCR product lacking 3′-terminal 45 nucleotides, thereby deleting the 15 carboxy-terminal amino acids. PCR-generated products were sequenced and inserted into the retroviral expression vectors pLZRS-IRES-GFP or pLZRS-IRES-ΔNGFR, upstream of the internal ribosome entry site (IRES) element. pLZRS vector information can be obtained at www.stanford.edu/group/nolan/retroviral_systems/retsys.html).
The human melanoma cell line Mel JuSo (MJS), MJS TAP1-GFP [57] and Madin-Darby bovine kidney (MDBK) cells (American Type Culture Collection, ATCC) were maintained in RPMI-1640 medium; GP2-293 cells, porcine kidney (PK15) cells, the embryonic bovine trachea (EBTr) cell line, Madin-Darby canine kidney I (MDCK I) cells, and the equine epithelial cell line E.derm were maintained in DMEM medium. Media were supplemented with 10% heat-inactivated fetal bovine serum (FBS) (with the exception of E.derm cells that required 20%), 2 mM L-glutamine (Invitrogen), 140 IU/ml penicillin and 140 µg/ml streptomycin. PHA-treated T-cell blasts positive for HLA-A1 and HLA-A2 were generated from PBMCs by stimulation with 0.8 µg/ml PHA and were subsequently cultured in IMDM supplemented with 100 IU/ml IL-2 and 10% FBS. The HLA-A2-expressing melanoma cell line 518, Mel518 (a kind gift from E. Verdegaal, department of Clinical Oncology, Leiden University Medical Center, Leiden, The Netherlands) was maintained in DMEM containing 4.5 mM glucose, supplemented with 8% FBS, 2 mM L-glutamine (Invitrogen), 140 IU/ml penicillin and 140 µg/ml streptomycin.
Recombinant retroviruses were prepared using the Phoenix amphotropic packaging system as described previously (www.stanford.edu/group/nolan/retroviral_systems/retsys.html ). MJS, MDCK I, PK15, and E.derm cells were transduced with recombinant retroviruses to generate the following stable cell lines: MJS, MDCK I, PK15, and E.derm controls (containing BHV-1 UL49.5 in the anti-sense orientation, GFP+); MJS UL49.5BHV-1, PHA T-cell blast UL49.5BHV-1 and Mel518 UL49.5BHV-1 (containing BHV-1 UL49.5 in the sense orientation (SO), GFP+); MJS UL49.5VZV, PHA T-cell blast UL49.5VZV and Mel518 UL49.5VZV (containing VZV UL49.5 SO, GFP+); MJS UL49.5CHV and MDCK I UL49.5CHV (containing CHV UL49.5 SO, GFP+); MJS UL49.5PRV and PK15 UL49.5PRV (containing PRV UL49.5 SO, GFP+); MJS UL49.5EHV-1 and E.derm UL49.5EHV-1 (containing EHV-1 UL49.5 SO, GFP+); MJS UL49.5EHV-1Δtail (containing tail-less EHV-1 UL49.5 SO, GFP+); MJS UL49.5EHV-4 and E.derm UL49.5EHV-4 (containing EHV-4 UL49.5 SO, GFP+). In addition, MJS TAP1-GFP cells were transduced with recombinant retrovirus to generate MJS TAP1-GFP control (containing the empty pLZRS construct, ΔNGFR+); MJS TAP1-GFP UL49.5BHV-1 (containing BHV-1 UL49.5 SO, ΔNGFR+); MJS TAP1-GFP UL49.5PRV (containing PRV UL49.5 SO, ΔNGFR+) and MJS TAP1-GFP UL49.5EHV-1 (containing EHV-1 UL49.5 SO, ΔNGFR+). To generate recombinant retroviruses for MDBK cell line transductions, the GP2-293 pantropic packaging cell line was used according to the protocol obtained from BD Bioscience Clontech (www.bdbiosciences.com). In brief, 1×106 of GP2-293 cells were co-transfected with retroviral expression vector (pZLRS-IRES-GFP containing the BHV-1 UL49.5 gene in anti-sense or in the sense orientation) and pVSV-G construct (envelope vector) for retrovirus production. Retrovirus-containing medium was collected 48 hours post-transfection. MDBK cells were transduced four times with VSV-G containing recombinant retroviruses to generate the following stable cell lines: MDBK control (containing BHV-1 UL49.5 in anti-sense orientation, GFP+) and MDBK UL49.5BHV-1 (containing BHV-1 UL49.5 SO, GFP+). All cell lines generated in this study were selected for GFP or ΔNGFR expression using a FACSVantage cell sorter (Becton Dickinson). To obtain MJS cells stably expressing the HCMV-encoded US6 (MJS US6), MJS cells were transfected with pcDNA3-US6-IRES-NLS-GFP and selected for neomycin resistance [64].
The following antibodies were used in this study: anti-transferrin receptor (TfR) monoclonal antibody (mAb) 66Ig10, anti-TfR mAb H68.4 (Roche), anti-human MHC class I complexes mAb W6/32, anti-human MHC class I heavy chain mAb HC-10 (kindly provided by H. Ploegh, Whitehead Institute, Cambridge, Massachusetts, USA), anti-human class II HLA-DR mAb Tü36 (kindly provided by A. Ziegler, Institute for Immunogenetics, Universitätsklinikum Charité, Berlin, Germany), anti-TAP1 mAb 148.3 [54] and anti-TAP2 mAb 435.3 (kind gift from P. van Endert, Institut National de la Santé et de la Recherche Médicale, Paris, France). For the detection of equine TAP2, the polyclonal antibody anti-rat TAP2 Mac394 was used (kindly provided by M. Knittler, Institute of Immunology, Friedrich-Loeffler-Institute, Tübingen, Germany). For preparation of bovine TAP1 specific antibody, the bovine TAP1 ORF sequence encoding amino acid residues 117 to 167 were amplified from bovine genomic DNA and cloned into pGEX-4T-2 (GE Healthcare). The TAP1 polypeptide encompassing residues aa117-167 was purified as described previously [65]. The monoclonal antibody IL-A19 directed against bovine MHC class I molecules (a kind gift from Dr. J. Naessens, ILRAD, Nairobi, Kenya). The anti-equine and anti-canine MHC class I complexes mAb H58A and anti-porcine MHC class I mAb PT85A were purchased from VMRD Inc., Pullman, WA, U.S.A.
Mouse anti-BHV-1 UL49.5 serum was kindly provided by G.J. Letchworth (University of Wisconsin, Madison, Wisconsin, USA). Polyclonal rabbit anti-BHV-1 UL49.5 serum H11 was raised against a synthetic peptide representing the N-terminal sequence of BHV-1 UL49.5 and has been described [42]. In fig. 5, a different polyclonal rabbit anti-BHV-1 UL49.5 was used, obtained using a synthetic peptide corresponding to amino acid residues 27–41 of UL49.5 ([H] DAMRREGAMDFWSAGC*-[OH]). To facilitate conjugation to keyhole limpet hemocyanin, an additional irrelevant cysteine was added at the C terminus of the peptide (indicated by *). Rabbits were immunized by as described earlier [66]. The rabbit antiserum raised against PRV UL49.5 (gN) has been described [67], as was the anti-EHV-1 UL49.5 rabbit serum [43]. The VZV UL49.5-specific antibody was raised against two synthetic peptides: the N-terminal peptide EPNFAERNFWHASCSARGVYIDGSMITTLFKK and the C-terminal peptide RLFTRSVLRSTW. Both peptides were conjugated to glutathione S-transferase (GST) according to the methods described in [42]. The peptide-GST conjugates were mixed at a 1:1 ratio and emulsified in Freund's complete adjuvant for the first immunization and Freund's incomplete for the following immunizations. At 3-weeks intervals, the rabbit received four additional subcutaneous immunizations with the conjugates.
Cells were trypsinized and resuspended in phosphate-buffered saline (PBS) containing 1% bovine serum albumin (BSA) and 0.05% sodium azide. Cells were incubated with specific antibodies on ice for one hour. After washing, the cells were incubated with phycoerytrin (PE)-conjugated anti-mouse antibody for 45 min. Stained cells were analyzed by flow cytometry on a FACSCalibur flow cytometer (Becton Dickinson). To exclude dead cells, 7-aminoactinomycin D (7-AAD, Sigma-Aldrich) was added at a concentration of 0.5 µg/ml to all samples before analysis. Cells were analyzed using CellQuest software (Becton Dickinson).
The fluorescence-based peptide transport assay was performed as previously described [11],[68]. In brief, MJS cells were permeabilized with Streptolysin O (Murex Diagnostics Ltd.) at 37°C, followed by incubation with the fluorescein-conjugated synthetic peptide CVNKTERAY (N-core glycosylation site underlined) in the presence or absence of ATP. Peptide translocation was terminated by adding ice-cold lysis buffer containing 1% Triton X-100. After lysis, cell debris was removed by centrifugation, and supernatants were collected and incubated with Concanavalin A (ConA)-Sepharose (Amersham). After extensive washing of the beads, the peptides were eluted with elution buffer (500 mM mannopyranoside, 10 mM EDTA, 50 mM Tris-HCl pH 8.0) by vigorous shaking and further separated from ConA by centrifugation at 12,000×g for 2 minutes. The fluorescence intensity was measured using a fluorescence plate reader (CytoFluor, PerSeptive Biosystems; excitation 485 nm/emission 530 nm). The data were analyzed using the unpaired t-test. Statistical significance was set at p <0.05.
Cells were lysed in a buffer containing 1% (wt/vol) digitonin, 50 mM Tris·HCl (pH 7.5), 5 mM MgCl2, 150 mM NaCl, 1mM leupeptin, and 1 mM AEBSF (4-(2-Aminoethyl)-benzenesulfonyl fluoride), and subjected to immune precipitations using anti-TAP1 mAb 148.3 o/n. To determine steady state protein levels, cells were lysed in Nonidet P-40 (NP-40) lysis mix containing 50 mM Tris-HCl, pH 7.4, 5 mM MgCl2 and 0.5% NP-40, supplemented with 1 mM AEBSF (4-(2-Aminoethyl)-benzenesulfonyl fluoride), 1 mM leupeptin and 20 µM Cbz-L3 (Carbobenzoxy-1-Leucyl-1-Leucyl-1-Leucinal-H; Peptides International, Inc). The samples were kept on ice throughout the experiment. Protein complexes were denatured in reducing sample buffer (2% SDS, 50 mM Tris pH 8.0, 10% glycerol, 5% β-ME, 0.05% bromophenol blue) for 5 min at 96°C. Immunoblotting (IB) analysis was performed on denatured cell lysates separated by SDS-PAGE and blotted onto polyvinylidene fluoride (PVDF) membranes. Blots were incubated with the antibodies as indicated, followed by horseradish peroxidase-conjugated goat-anti-mouse or swine-anti-rabbit Igs (DAKO and Jackson Laboratories), and visualized by ECLplus (Amersham).
Steady-state labeling of MDBK cells with [35S]-methionine/cysteine and subsequent immunoprecipitations with rabbit BHV-1 UL49.5-specific antibody were performed as described [69]. Immunoblotting procedures with rabbit anti-bovine TAP1 and rabbit anti-α-tubulin have been described [65].
A total of 1,000 51Cr-labeled target cells were incubated with different CD8+ CTL clones at various effector to target ratios. The HY-A1 clone recognizes an HY epitope in the context of HLA-A1, and the HA2.27 clone recognizes the histocompatibility antigen HA-2 in the context of HLA-A2. After 4 hours of incubation at 37°C, 51Cr release into the supernatant was measured using standard methods. The mean percentage of triplicate wells was calculated as follows: % specific lysis = (experimental release–spontaneous release)/(maximal release–spontaneous release)×100. For analysis of IFN-γ production, 20,000 T-cells were co-cultured with 10,000 target cells. After 24 hours, the supernatant was harvested and the concentration of IFN-γ was measured by standard ELISA (Sanquin, Amsterdam, The Netherlands).
The wild type viruses used in this study were: BHV-1 strain Lam, BHV-1 strain Cooper (Fig. 5), PRV strain Kaplan and EHV-1 strain RacL1.The UL49.5 deletion mutant of PRV used in this study has been described before [41],[43]. The UL49.5 deletion mutant of EHV-1 (strain RacL11) was a gift from J. von Einem (College of Veterinary Medicine, Cornell University, Ithaca, NY, USA). Infections with wild type and mutant herpesviruses were carried out on the following cell lines: MDBK cells for BHV-1; PK15 cells for PRV and E.derm cells for EHV-1. The cells were washed once with PBS and infected with BHV-1 or PRV at an m.o.i. of 10, and with EHV-1 at an m.o.i. of 5 at 37°C in serum-free medium. After 2 hours, medium containing 10% FBS was added. Mock-infected cells were treated under the same conditions as infected cells. After 5 hours of infection, cells were collected and prepared for the peptide translocation assay. For immunoblotting and metabolic labeling experiments, MDBK cells were infected with BHV-1 wt and UL49.5Δtail viruses for 12 hours.
The BHV-1 UL49.5 mutant was generated by homologous recombination, using BHV-1 strain Lam as parent strain. The recombination region upstream of the UL49.5 gene was a 1.4 kb fragment running from nucleotide residue 7670 to 9061 (residue numbers based on the complete BHV-1 genome with NCBI accession number NC_001847, updated 30 March 2006). This fragment starts at a BstXI site 1.3 kb upstream of the start codon of the UL49.5 open reading frame and ends at its amino acid residue 31. The recombination region downstream of the UL49.5 gene was provided by a 1.9 kb fragment from nucleotide residue 9075 to 10972. This fragment starts at amino acid residue 36 of UL49.5 and ends at an FspI site 1.7 kb downstream of the UL49.5 open reading frame. A 2.2 kb NruI–PvuII fragment was cloned between the two UL49.5 recombination fragments that carries the hGFP gene in the expression cassette of pcDNA3 (Invitrogen). The complete recombination fragment (5.5 kb) was co-transfected with purified BHV-1 Lam DNA into EBTr cells using a calcium phosphate-based transfection method. After plating the supernatant of freeze/thawed transfected cells, a green plaque was found that, following three rounds of plaque purification, failed to react with anti-BHV-1 UL49.5 serum. The BHV-1-UL49.5 mutant could be grown to a titer of 107.0 TCID50/ml and was capable of penetrating bovine cells with the same kinetics as the wild type Lam strain.
The BHV-1 recombinant virus gN Am80, expressing a form of UL49.5 lacking its cytoplasmic domain, was constructed by introducing an amber mutation at gN residue 80R (AGG to TAG) by using a BHV-1 BAC clone (Liu and Chowdhury, manuscript in preparation).
Cellular microsomes were prepared as described [70]. Microsomes isolated from 7×106 homogenized cells were pre-incubated in 50 µl of AP buffer (5 mM MgCl2 in phosphate-buffered saline, pH 7.0) on ice for 45 min in the absence or presence of a 200-fold molar excess of the non-labeled TAP-specific viral inhibitor ICP47 [18]. Different concentrations of radiolabeled peptide (RR[125I]YQKSTEL) were added equally to the samples with or without ICP47 and incubated on ice [71]. Non-bound peptides were removed by washing the membranes with 400 µl of AP buffer and subsequent centrifugation at 20,000 g for 8 min. The amount of radioactivity bound to the membranes was quantified by γ-counting and corrected for the signal obtained in the presence of ICP47. All experiments were performed in triplicate.
TAP binding to ATP-agarose was assayed as described [11]. In brief, cells were solubilized in 1% (w/v) digitonin, 50 mM Tris-HCl (pH 7.5), 5 mM MgCl2, 150 mM NaCl, 5 mM iodoacatamide, and 1 mM AEBSF. Hydrated C-8 ATP-agarose (Fluka/Sigma) was added to the post-nuclear supernatant and incubated by rotation at 4°C. After 2 hours, the supernatant was separated from the ATP-agarose pellet by 5 minutes centrifugation. The resulting pellet was washed three times with 0.1% (w/v) digitonin, 50 mM Tris-HCl (pH 7.5), 5 mM MgCl2 and 150 mM NaCl. Proteins bound to the ATP-agarose were eluted with 500 mM EDTA and SDS sample buffer was added to both the supernatant and the pellet. The samples were separated using SDS-PAGE and analyzed by immunoblotting.
Confocal microscopy and Fluorescence Recovery After Photobleaching (FRAP) assays were performed as described [11],[57]. In short, a circular spot in the ER was bleached at full intensity, and an attenuated laser beam was used to monitor recovery of fluorescence. The half-time for recovery was calculated from each recovery curve after correction for loss of fluorescence caused by imaging (usually <4%). The diffusion coefficient D was determined from at least seven cells measured in different experiments.
UL49.5 sequence data used to generate the alignment shown in Fig. 1 and the phylogenetic tree shown in Fig. 7, have been obtained from the NCBI (www.ncbi.nlm.nih.gov) database with the accession numbers: bovine herpesvirus 1 (BoHV-1) [NP_045309], bovine herpesvirus 5 (BoHV-5) [NP_954898], cercopithecine herpesvirus 1 (CeHV-1) [AAP41468], cercopithecine herpesvirus 9 (CeHV-9) [NP_077423], cercopithecine herpesvirus 16 (CeHV-16) [YP_443897], equid herpesvirus 1 (EHV-1) [AAT67267], equid herpesvirus 4 (EHV-4) [CAA35670], gallid herpesvirus 1 (GaHV-1) [YP_182341], gallid herpesvirus 2 (GaHV-2) [NP_057812], gallid herpesvirus 3 (GaHV-3) [NP_066882], human herpesvirus 1 (HSV-1) [NP_044652], human herpesvirus 2 (HSV-2) [NP_044520], human herpesvirus 3 (VZV) [YP_068406], meleagrid herpesvirus 1 (MeHV-1) [AAG30090], psittacid herpesvirus 1 (PsHV-1) [AAQ73691], suid herpesvirus 1 (PRV) [YP_068325], transporter 1 ATP-binding cassette sub-family B [Bos taurus] [AAY34698]. Not obtained from the NCBI database are: bubaline herpesvirus 1 (BuHV-1)[MSRSLLVALATAALLAMVRGLDPLLDAMRREE AMDFWSAGCYARGVPLSEPPQAMVVFYAALTVVMLAVALYAYGLCFRLMSAGGPNKKEVRGRG; FAMR, unpublished], canid herpesvirus (CHV) [patent EPO910406 http://ep.espacenet.com ], cervid herpesvirus 1 (CvHV-1) [MARMPRSLLSALAVAALLAIAGARDPLLDAMRHEGAMDFWSASCYARGVPL SEPPQALVVFYVALAVVMFSVAVYAYGLCLRLVGADSPNKKDSRGRG; FAMR, unpublished]. |
10.1371/journal.pcbi.1006530 | Host contact dynamics shapes richness and dominance of pathogen strains | The interaction among multiple microbial strains affects the spread of infectious diseases and the efficacy of interventions. Genomic tools have made it increasingly easy to observe pathogenic strains diversity, but the best interpretation of such diversity has remained difficult because of relationships with host and environmental factors. Here, we focus on host-to-host contact behavior and study how it changes populations of pathogens in a minimal model of multi-strain interaction. We simulated a population of identical strains competing by mutual exclusion and spreading on a dynamical network of hosts according to a stochastic susceptible-infectious-susceptible model. We computed ecological indicators of diversity and dominance in strain populations for a collection of networks illustrating various properties found in real-world examples. Heterogeneities in the number of contacts among hosts were found to reduce diversity and increase dominance by making the repartition of strains among infected hosts more uneven, while strong community structure among hosts increased strain diversity. We found that the introduction of strains associated with hosts entering and leaving the system led to the highest pathogenic richness at intermediate turnover levels. These results were finally illustrated using the spread of Staphylococcus aureus in a long-term health-care facility where close proximity interactions and strain carriage were collected simultaneously. We found that network structural and temporal properties could account for a large part of the variability observed in strain diversity. These results show how stochasticity and network structure affect the population ecology of pathogens and warn against interpreting observations as unambiguous evidence of epidemiological differences between strains.
| Pathogens are structured in multiple strains that interact and co-circulate on the same host population. This ecological diversity affects, in many cases, the spread dynamics and the efficacy of vaccination and antibiotic treatment. Thus understanding its biological and host-behavioral drivers is crucial for outbreak assessment and for explaining trends of new-strain emergence. We used stochastic modeling and network theory to quantify the role of host contact behavior on strain richness and dominance. We systematically compared multi-strain spread on different network models displaying properties observed in real-world contact patterns. We then analyzed the real-case example of Staphylococcus aureus spread in a hospital, leveraging on a combined dataset of carriage and close proximity interactions. We found that contact dynamics has a profound impact on a strain population. Contact heterogeneity, for instance, reduces strain diversity by reducing the number of circulating strains and leading few strains to dominate over the others. These results have important implications in disease ecology and in the epidemiological interpretation of biological data.
| Interactions between strains of the same pathogen play a central role in how they spread in host populations. [1–7]. In Streptococcus pneumoniae and Staphylococcus aureus, for instance, several dozen strains can be characterized for which differences in transmissibility, virulence and duration of colonization have been reported in some cases [8, 9]. Strain diversity may also affect the efficacy of prophylactic control measures such as vaccination or treatment. Indeed, strains may be associated with different antibiotic resistance profiles [3, 5, 10, 11], and developed vaccines may only target a subset of strains [2, 3, 12]. With the increasing availability of genotypic information, it has become easy to describe the ecology of population of pathogens and to monitor patterns of extinction and dominance of pathogen variants [13–17]. However, the reasons for multi-strain coexistence patterns (e.g. coexistence between resistant and sensitive strains) or dominance of certain strains (e.g. in response to the selection pressure induced by treatment and preventive measures) remain elusive. One may invoke selection due to different pathogen characteristics, but also environmental and host population characteristics, leading to differences in host behavior, settings and spatial structure may affect the ecology of strains [14–19]. In particular, human-to-human contacts play a central role in infectious disease transmission [20]. This is increasingly well described thanks to extensive high-resolution data—including mobility patterns [21–23], sexual encounters [24], close proximity interactions in schools [25, 26], workplaces [27], hospitals [16, 28–31], etc.—that enable basing epidemiological assessment on contact data with real-life complexity [32, 33]. For instance, the frequency of contacts can be highly heterogeneous leading more active individuals to be at once more vulnerable to infections and acting as super-spreaders after infection [24, 33–35]. Organizational structure of certain settings (school classes, hospital wards, etc.) and other spatial proximity constraints lead to the formation of communities that can delay epidemic spread [36, 37]. Individual turnover in the host population is also described as a key factor in controlling an epidemic [20, 38]. It is likely that, since they impact the spread of single pathogens, the same characteristics could affect the dynamics in multi-strain populations. It was shown, indeed, that network structure impacts transmission with two interacting strains [39–46], the evolution of epidemiological traits [47–49] and the effect of cross-immunity [50, 51]. Yet in these cases, complex biological mechanisms—such as mutation, variations in transmissibility and infectious period, cross immunity—were used to differentiate between pathogens, thereby making the role of network characteristics difficult to assess in its own right.
For this reason, we focused on the dynamical pattern of human contacts and examined whether it contributes to shaping the population ecology of interacting strains under minimal epidemiological assumptions regarding transmission. We described a neutral situation where all strains have the same epidemiological traits and compete via mutual exclusion (concurrent infection with multiple strains is assumed to be impossible) in a Susceptible-Infected-Susceptible (SIS) framework. We studied the spread of pathogens in a host population during a limited time window, disregarding long-term evolution dynamics of pathogens. More precisely, new strains were introduced through host turnover rather than de novo mutation or recombination in pathogens. We quantified the effect of network properties on the ecological diversity in strain populations with richness and dominance indicators. We assessed in turn heterogeneities in contact frequency, community structure and host turnover by comparing simulation results obtained with network models exhibiting a specific feature. We then interpreted S. aureus carriage in patients of a long-term care facility in the light of these results.
We simulated the stochastic spread of multiple strains on a dynamical contact network of individuals (nodes of the network). Individuals can be either susceptible or infected with a single strain at a given time, and, for each strain, β and μ indicate the transmission and the recovery rate respectively. We assumed turnover of individuals, who enter the system with rate λin, and associated injection of previously unseen strains, carried by incoming individuals with probability ps. We considered synthetic network models, each displaying a specific structural feature, as well as a real network reconstructed from close-proximity-interaction data collected in a hospital facility. We calibrated all network models to the same average quantities—average population size V ¯, fraction of active nodes a ¯, average degree k ¯ and strength of the community repartition pIN, when applicable—that were chosen to correspond with the hospital network used in the application. Epidemiological parameters were motivated by the duration of S. aureus carriage in patients. A larger range of values was explored in some cases to address their impact on the dynamics. We analyze the structure of strain population at the dynamic equilibrium by computing, for each network model, ecological diversity measures, including species richness and evenness/dominance indices [52, 53]. All details about network models, numerical simulations and ecological indicators are described in the Materials and methods section.
In order to probe the effect of contact heterogeneity on strain ecology we compared a homogeneous model (HOM) in which all nodes have the same activity potential, i.e. they have equal rate of activation to establish contacts, with a heterogeneous model (HET), akin to the activity-driven model described in [34], where the activity potential is different across nodes and is drawn from a power-law distribution.
Fig 1 shows the results of numerical simulations comparing HOM and HET models. Sample epidemic trajectories are reported in Fig 1A. Here every strain is indicated with its own color to display its dynamics resulting from the interaction with the other strains. Fig 1B–1D shows summary statistics in varying strain transmissibility β. The prevalence presents a well-known behavior for both static and dynamic networks (Fig 1B): contact heterogeneities lower the transmissibility threshold above which total prevalence is significantly above zero, thus allowing the spread of pathogens with low transmissibility. At the same time, however, heterogeneities hamper the epidemic spread when β is large, reducing the equilibrium prevalence [35]. Fig 1C shows the average richness, i.e. the number of distinct strains co-circulating. For low values of β HET displays larger richness values compared to HOM. This trend reverses as β increases, and the richness is lower in HET consistently with the lower level of prevalence. The relation between richness and prevalence, however, is not straightforward. For instance, the reduction in richness for high β values is important even for the case with limited contact heterogeneity, when prevalence is barely affected. The scaling between prevalence and richness is not linear as β varies (Fig 1D), and the relation between the two quantities varies appreciably among contact networks. In correspondence of a fixed value of prevalence, heterogeneous networks have lower richness—e.g. a prevalence value of ∼0.8 corresponds to ∼20% lower richness in HET with respect to HOM, as highlighted in Fig 1D.
This fact can be explained by the dynamical properties of epidemics on heterogeneous networks. Active nodes, involved in a larger number of contacts, get infected more frequently [35]. Also, a randomly chosen node is likely surrounded by active nodes [33]. As a consequence, injected strains often find their propagation blocked by active infected nodes. In this way, contact heterogeneities enhance the competition induced by mutual exclusion and hamper the wide-spread of emerging strains, similarly to what was found in [46]. This mechanism is further confirmed by looking at the persistence time of strains (S2 Fig in the supporting information). Above the epidemic threshold, it is on average shorter in heterogeneous networks than in homogeneous ones. The distributions are however more skewed in heterogeneous networks, indicating that more strains are going extinct rapidly, while a few others can survive for a long time in the population.
If on the one hand hubs accelerate the extinction of certain strains, on the other they act as super-spreaders, amplifying the propagation of other strains. We find that this impacts profoundly the distribution of strains’ abundances, i.e. the strain-specific prevalence. Fig 2A shows that the latter is broader for the HET network, with the most abundant strain reaching a larger proportion of cases. This situation is synthesized by the Berger-Parker index, that quantifies the level of unevenness or dominance of a given ecological system [52, 53]. This is defined as the relative abundance of the most abundant strain (see Materials and methods section). Fig 2B shows that Berger-Parker index increases with β for all networks. This is expected since at low β strains′ transmission chains are short and barely interact, while they interfere more at higher values of transmission potential. The Berger-Parker index is always higher in a heterogeneous network, even when the comparison is made at fixed values of richness (Fig 2C). An alternative indicator, the Shannon evenness, shows a similar behavior as displayed in S3 Fig.
The fraction of strains going extinct also depends on stochastic effects in a finite size population. We indeed found that increasing network size, when temporal and topological properties were the same, led to an increase in both persistence time and richness (S4 Fig). This shows that interference among transmission chains is reduced in larger populations. However, the relative abundance distribution remained similar, showing that it is primarily affected by the nodes’ activity distribution (S5 Fig).
Eventually, we tested whether additional mechanisms of strain injection were leading to different results. In S6 Fig we assumed new strains to infect susceptible nodes already present in the system with rate qs, mimicking in this way transmissions originating from an external source, as it can happen in real cases. The plot of S6 Fig shows the same qualitative behavior described here.
We considered a community model (COM) with nC communities in which all nodes are as active as in HOM, but direct a fraction pIN of their links within their community and the rest to nodes in the remaining nC − 1 communities. The closer pIN is to 1, the stronger the repartition in communities is.
Fig 3A and 3B shows that a network with communities displays a higher richness for large β; even when community structure barely affects prevalence (Fig 3B). However, the effect is important only when communities are fairly isolated (pIN = 0.99) and the injection from the outside is not so frequent—otherwise the effect is masked by strain injection which occurs uniformly across communities. In particular, for the values of pIN = 0.78 and ps = 0.079, chosen to match the hospital application, the difference with the homogeneous case is very small. The limited role of community structure is also confirmed by the fact that once this feature is combined with heterogeneous activation—in a model with the activation scheme of HET and the stub-matching of COM—the latter property has the dominant effect and the richness decreases (S1 Fig).
The relation between richness and prevalence remains the same when adding the injection of new strains due to the transmission from an external source. This mechanism further increases the richness. When β is high and the fraction of infected nodes is close to one, however, such a mechanism is hindered by the fact that susceptible nodes, that can get infected from the external source, are rare (see S6 Fig). This is why richness starts to decrease for high values of β.
We tested the consequences of communities in strain dominance by plotting the Berger-Parker index in Fig 3C. For low β, the behavior of the Berger-Parker index follows the trend in richness. The initial decrease in this indicator is due to the increase in richness, that occurs at constant prevalence and is thus associated to a decrease in the average abundance [54]—green curve corresponding to pIN = 0.99 and ps = 0.01. At larger values of β, instead, increased competition levels induced higher dominance levels.
The increase in strain diversity is due to the reduced competition among strains introduced in different communities. When coupling among communities is low, indeed, strains may spend the majority of time within the community they were injected in, thus avoiding strains injected in other communities. Fig 3D confirms this hypothesis by showing the Inverse Participation Ratio (IPR) [55] that quantifies uniformity in the repartition of abundance across communities. Values close to zero indicate uniform repartition, while, conversely, values close to 1 indicate that, on average, a strain is confined within a single community for most of the time (more details are reported in the Materials and methods section). The strength of the community structure does not affect the repartition of the total prevalence (squares in the plot), however it alters the average IPR value computed from the abundance of single strains, thus strains become more localized as pIN increases. Notice that a certain degree of localization is present also in the homogeneous network, due to those strains causing very few generations before going extinct.
As a sensitivity analysis we tested whether the main results obtained so far are the same in a more realistic situation where additional heterogeneous properties of nodes are accounted for. We consider the case in which infectious duration varies across individuals, as happens for S. aureus colonization. S7 Fig shows that the inclusion of three classes differing in recovery rate reduces richness and increases the Berger-Parker index with respect to the homogeneous recovery. However, the effects discussed so far—e.g. reduction and amplification of richness in HET and COM, respectively—are still present.
Node turnover represents another important property of a network that may impact the ecological dynamics of strains for two reasons: incoming individuals contribute to richness by injecting new strains; on the other hand, the removal from the population of infected nodes breaks transmission chains and hampers the persistence of strains. The result of the interplay between these two mechanisms is summarized by the plot of richness as a function of β and node length of stay, τ,—Fig 4A. The figure, obtained with the HOM model, shows two distinct regimes. In the former case, richness decreases as τ increases, because replacement of individuals becomes slower and injections less frequent. In the high β regime, instead, the average richness at fixed β does not depend monotonically on the node turnover but it is instead maximized at intermediate τ. Interestingly, the optimal value of τ decreases as β increases. This behavior can be explained by looking at the balance between injection and extinction that determines the equilibrium value of richness, N ¯ S. This reads [56]:
N ¯ S = λ inp s T pers ( β , τ ) = V ¯ p s T pers ( β , τ ) τ , (1)
where λinps is the rate at which new strains are introduced and Tpers is the average persistence time of a strain. The trade-off between injection and extinction appears as the ratio between the two time scales, Tpers and τ. In the limit τ → 0 the spread plays no role, even for high β. As τ increases, newly introduced infectious seeds have a higher probability to spread, thus the average extinction time initially increases super-linearly with τ (see S8 Fig in the supporting information) resulting in an increase of richness. However, past a certain value of τ, Tpers does not grow super-linearly anymore, thus a further increase in τ is detrimental for pathogen diversity because it is associated to fewer introductions. This general behavior was not altered by the accounting for introductions by transmissions from an external source as shown in S6 Fig.
We derive an approximate formula for Tpers considering an emerging strain competing with a single effective strain formed by all other strains grouped together. This formulation, enabled by the neutral hypothesis, makes it possible to write the master equation describing the dynamics and to use the Fokker-Planck approximation to derive persistence times (see Materials and methods section). Analytical results well reproduce the behavior observed in the simulations, and, in particular, the value of the length of stay maximizing richness for different β as shown by the comparison between white stars and continuous line in Fig 4B. The quantitative match for other values of ps is reported in S9 Fig.
Unlike richness, Berger-Parker index always increases monotonically with the length of stay—Fig 4B. This behavior is due to the correlation of this indicator with average abundance, similarly to what we discussed in the previous section.
We conclude by analyzing the real-case example of the S. aureus spread in a hospital setting [10, 57]. We used close-proximity-interaction (CPI) data recorded in a long-term health-care facility during 4 months by the i-Bird study [16, 28, 31]. These describe a high-resolution dynamical network whose complex structure reflects the hospital organization, the subdivision in wards and the admission and discharge of patients [58]. Together with the measurements of contacts, weekly nasal swabs were routinely performed to monitor the S. aureus carriage status of the participants and to identify the spa-type and the antibiotic resistance profile of the colonizing strains.
The modeling framework considered here well applies to this case. The SIS model is widely adopted for modeling the S. aureus colonization [59, 60], and the assumption of mutual exclusion is made by the majority of works to model the high level of cross-protection recognized by both epidemiological and microbiological studies [61, 62]. The dynamic CPI network was previously shown to be associated with paths of strain propagation [16]. Consistently, we assumed that transmission is mediated by network links with transmissibility β. In addition, new strains are introduced in the population carried by incoming patients, or through contacts with persons not taking part in the study.
Fig 5A shows weekly carriage and its breakdown in different strains. Prevalence and richness fluctuate around the average values 87,3 ± 6,3 cases and 39,8 ± 2 strains, respectively. Simulation results are reported in Fig 5B, that displays the impact of transmission and introduction rate on richness and prevalence. When qs is low we find a positive trend between richness and prevalence, consistently with the synthetic case. For larger values of qs the trend appears instead different. As transmissibility increases, richness initially grows with prevalence and then decreases after a certain point. This behavior is the same as observed in S6 Fig and stems from the reduction of susceptible nodes, that causes a decline in the expected injection rate—see Materials and methods section.
To quantify the effect of contact patterns on S. aureus population ecology we compared simulation results with the ones on a network null model. Specifically, we built the RAND null model that randomizes contacts while preserving just the first and the last contact of every individual. The randomization preserves node turnover, the number of active nodes and links and destroys contact heterogeneities and community structure along with other correlations. Fig 5C shows the comparison for different transmissibility values. The effect of the network is consistent with the theoretical results described for a heterogeneous network, i.e. smaller richness values correspond to the same prevalence in the real network compared to the homogeneous one. We then quantified the level of dominance of the multi-strain distribution by means of the Berger-Parker index. We chose for each network the values of qs and β that better reproduce empirical richness and prevalence and, interestingly, we found that, for the two cases, same average richness and prevalence correspond to different levels of Berger-Parker index. The Berger-Parker index obtained with the real network is the highest and the one that better matches the empirical values—i.e. the empirical values are within one standard deviation of the mean for almost all weeks. Based on this result we argue that contact heterogeneities, along with the other properties of the contact network, contribute to the increased dominance of certain strains.
Multiple biological and environmental factors concur in shaping pathogen diversity. We focused here on the host contact network and we used a minimal transmission model to assess the impact of this ingredient on strain population ecology, quantifying the effects of three main network properties, i.e. heterogeneous activity potential, presence of communities and turnover of individuals. Results show that the structure and dynamics of contacts can alter profoundly strains’ co-circulation. Contact heterogeneities were found to shape the distribution of strains’ abundances. Highly active nodes are known to play an important role in outbreak dynamics by acting as super-spreaders [33]. At the same time, however, they were found to enhance the interference between the transmission chains of different strains, thus hindering the spread of an emerging variant [46]. Here we showed that the combination of these two dynamical mechanisms reduces the richness and increases the level of heterogeneity in strains’ abundances. In particular, hubs could allow strains with no biological advantage to generate a large number of cases and outcompete other equally fit strains. This mechanism may potentially bias the interpretation of biological data. Dynamical models that do not properly account for contact structure could overestimate the difference in strains’ epidemiological traits in the attempt to explain observed fluctuations in strain abundance induced in reality by super-spreading events. Moreover, these models could provide biased assessment of transmission vs. introduction rates.
The presence of communities causes the separation of strains and mitigates the effect of competition thus enhancing co-existence. A similar behavior was already pointed out before [46, 51, 59, 64], e.g. for the spread of S. pneumoniae, as induced by age assortativity [64], for the case of S. aureus where distinct settings were considered [59], and for a population of antigenic distinct strains in presence of cross-immunity [51]. We found that the impact of community structure is not so strong, and it is likely minor when individuals of different communities have frequent contacts. No appreciable variation was observed, indeed, for pIN = 0.78, chosen to match the inter-ward coupling of the hospital network. Similar results can be expected for school classes or workplace departments presenting a similar level of community mixing. The effect on richness becomes appreciable for low community coupling (e.g. pIN = 0.99 in Fig 3). This is consistent with a certain degree of diversity observed among strains belonging to separated communities, as it is the case of different hospitals [15].
Eventually, the analysis of turnover of individuals revealed major effects on strain diversity, when this mechanism is also the main driver of the introduction of strains in the population. When transmissibility is low richness decreases with host length of stay. When transmissibility is above the epidemic threshold we showed the existence of an optimal value of the length of stay that maximizes strain richness as a result of the interplay between two competing time scales, namely the typical inter-introduction time and the average persistence time of a strain. This provides insights for the spread of bacterial infections in transmission settings, such as hospitals or farms, that are of particular relevance for the spread of antimicrobial resistance and that are characterized by a rapid host turnover [15, 31, 65]. For the case of hospitals, for instance, they suggest that variations in patients’ length of stay, as induced by a change of policy, could have appreciable effects on the population structure of nosocomial pathogens.
We adopted a neutral model to better disentangle the relative role of the different network properties. A wide disease-ecology literature addressed the consequences of neutral hypotheses on multi-strain balance in order to provide a benchmark for interpreting the observed co-existence patterns and gauging the effect of selective forces potentially at play [11, 18, 66, 67]. Many of these works addressed, for instance, the co-existence between susceptible and resistant strains of S. pneumoniae [11, 66]. However, this assumption was rarely adopted in network models, that consider for the majority strains with different epidemiological traits with the aim of describing pathogen selection and evolution [47–49, 68]. Strains were assumed to have the same infection parameters in [50, 51], where the role of community structure and clustering was analyzed in conjunction with cross-immunity. With respect to these works, the minimal transmission model used here enabled a transparent comprehension of the role of the network. Multiple identical SIS processes can be mapped, in fact, on a single SIS process, in such a way that the wide literature of single SIS processes allows for a better understanding of the behavior recovered in the simulations [32, 33]. Strains can be also grouped in two macro-strains. This strategy allowed us to adopt the viewpoint of an emerging strain and study its competition with the others seen as a unique macro-strain. The associated master equation and Fokker-Planck approximation allowed computing the average extinction time, capturing the key aspects of the dynamics. In a future work this theoretical framework could be extended to consider other network topologies. It could, for instance, be coupled with the activity-block approximation to describe heterogeneous networks. Additional numerical analyses, based on a similar transmission model, could also address other properties known to alter spreading dynamics, such as heterogeneous inter-contact time distribution or topological and temporal correlations.
As a case study, we analyzed the spread of S. aureus in a hospital taking advantage of the simultaneous availability of contact and carriage information [16]. The temporal and topological features of the network lead to a lower prevalence and richness with respect to the homogeneous mixing (although the effect was quite small). In addition, similar prevalence and richness values are associated to different dominance levels in different networks—i.e. different values of the Berger-Parker index—with the real network leading to a higher dominance as observed in reality. This behavior can be explained by the theoretical results and can be attributed essentially to the effect of contact heterogeneities, considering that the community structure does not have appreciable effects for this network, as discussed above. The importance of accounting for host contacts and hospital organization in the assessment of bacterial spread and designing interventions has been recognized by several studies [16, 28–31, 63]. Here we show that this element may be critical also for understanding the population ecology of the bacterium. It is important to note however that, while the realistic network provides results that are closer to the data, this ingredient explains only part of the heterogeneity observed in the abundance. This shows that the contact network is a relevant factor, but other factors should be considered as well. The approach used here is intentionally simplified, as we focused on the main dynamical consequences of the contact network. Clearly, more detailed models can be designed to reproduce more closely the data. A certain degree of variation in the epidemiological traits could be at play, as for example the fitness cost of resistance [8]. Role of hosts in the network (e.g. patients vs. health-care workers), and heterogeneities in health conditions, antibiotic treatment and hygiene practices are also known to affect duration of carriage and chance of transmission [16, 28, 31, 63]. Eventually, we must consider that the comparison of model output with carriage data is also affected by the limitation of the dataset itself, already described in [16]. In particular, the weekly swabs may leave transient colonization undetected. Moreover, while the relevance of CPIs as proxies for epidemiological links has been demonstrated [16], the transmission through the environment (e.g. in the form of fomites) is also possible.
The understanding provided here can be relevant for other population settings, temporal scales and geographical levels. In addition, the modeling framework could be applied to pathogens other than S. aureus, such as human papillomavirus, S. pneumoniae and Neisseria meningitidis, for which the strong interest in the study of the strain ecology is justified by the public health need for understanding and anticipating trends in antibiotic resistance, or the long-term effect of vaccination [1, 2, 4, 5]. With this respect, if the simple framework introduced here increases our theoretical comprehension of the multi-strain dynamics, more tailored models may become necessary according to the specific case. In particular, we have considered complete mutual exclusion as the only mechanism for competition. In reality, a secondary inoculation in a host that is already a carrier may give rise to alternative outcomes, such as co-infection or replacement [69]. In addition, infection or carriage may confer a certain level of long-lasting strain-specific protection and/or a short-duration transcendent immunity [11, 50]. Eventually mechanisms of mutation and/or recombination are at play and their inclusion into the model can be important according to the time scale of interest.
We provide here details of the generative algorithms used for the contact network models. Network dynamics is implemented in discrete time according to the following rules common to all models:
Turnover dynamics: new nodes arrive according to a Poisson process with rate λin and leave after a random time drawn from an exponential probability distribution with average τ. After a short initial transient, population size is Poisson distributed with average V ¯ = λ in τ. Upon admission, a node i is assigned with an activity potential ai, i.e. an activation rate, drawn at random from a given probability distribution P(a). Any node retains this property throughout its whole lifespan.
Activation Pattern: each node i becomes active with rate ai. It then receives a number of stubs drawn from a zero-truncated Poisson distribution with parameter κ—we require active nodes to engage in at least one contact. The average number of stubs, computed among active nodes, is thus given by κ/(1 − e−κ), and the average degree can be computed by the latter quantity multiplied by the average activity potential. The active status lasts for a single time step.
Stub-matching: stubs are then matched according to the actual model considered.
We now describe in detail each network model:
HOM: in this model each node has the same probability aH to be active during each time step; the activity distribution is thus P(a) = δ(a − aH), where δ(x) is the Dirac’s delta function. Stubs are matched completely at random in order to form links, according to a configuration model [33]. We discard eventual self-links and multiple links that may occur during the matching procedure.
HET: here each node i has its own activity rate ai, drawn from a power-law distribution P(a) ∝ a−γ, with a ∈ (ϵ, 1]. We tune the variance by varying γ—lower γ higher variance. We then set ϵ to have the average activity a ¯ equal to aH in HOM. Stub-matching procedure is the same as in HOM. HET model is thus a variant of the activity driven model introduced in [34] with the difference that here contacts are created only among active individuals.
COM: incoming nodes are assigned to one among nC communities with equal probability—so that communities have the same size on average—and belong to the same community throughout their whole lifespan. Stubs are matched according to the community each node belongs to. Precisely, any stub is matched either with another stub of the same community, with probability pIN, or with a stub of a different community, with complementary probability. Here the stub-matching procedure results in a larger number of lost links—to eliminate multiple links and self-loops—compared to HOM and HET, due to the difficulty in matching stubs within small groups. Thus, the parameter κ has to be adjusted manually to recover the same average degree as in HOM and HET. Each node has the same activity potential aH as in HOM.
We use a dynamical contact network obtained from CPI data collected during the i-Bird study in a French hospital. Details of the network are already reported in [16]. Briefly, the dataset describes contacts occurring between 592 individuals from July to November 2009. The study involved both patients and health-care workers, distributed in 5 wards, as well as hospital service staff. Every participant wore a wireless device designed to broadcast a signal every 30 s containing information about its ID. Signal strength was tuned so that only devices within a small distance (around 1.5 m) were able to register a contact. CPIs were finally aggregated daily, keeping the information about their cumulative duration within each day.
We discard CPIs relative to the first 2 weeks and the last 4 weeks of dataset, corresponding to a period of adjustments in the measurements and progressive dismissal of the experiment, respectively. Simulations conducted with the CPIs network were compared with results obtained with a null model which we refer to as RAND. According to this randomization scheme the activity of a node is randomized while respecting the constraint that removal and addition of contacts must not alter the time of the first and the last contact of each node (tS and tL respectively). Notice that RAND preserves the number of nodes that are present at any time in the network by preserving their first contact tS and their length of stay tL − tS. Null models randomizing the latter properties lead to misleading results when node length of stay is heterogeneous and node turnover occurs [70]. RAND also sets all contact weights equal to the average weight value.
Spreading dynamics is stochastic and is performed in discrete time. At each time step of duration Δt, we update the state of each node: each infected node transmits the strain it is carrying to a susceptible neighbor with probability βΔt and it turns susceptible with probability μΔt. Notice that due to mutual exclusion, an individual can be infected by a single strain at a time [71]. Strain injection is given by the combination of two processes: incoming individuals bring a new strain with probability ps, and susceptible individuals turn infectious with a new strain with probability qsΔt. The two mechanisms mimic respectively incoming infectious individuals (e.g. admission of colonized patients) and transmission from an external source (in the hospital example this corresponds to contacts with individuals that were not participating in the study). The expected injection rate, which accounts for both introduction mechanisms, is thus given by ι = λ in p s + S ¯ q s, where S ¯ is the average number of susceptible individuals at the equilibrium. In the theoretical analysis in the main paper we assumed qs = 0 for simplicity, thus variations in ι were induced by variations in λin and ps. The case qs > 0 was considered in the supporting information.
Simulations on synthetic networks differ from those on the hospital network in the combination of the spreading and network dynamics. In the synthetic network case, at each time step of duration Δt = 1h, both network and spreading dynamics are simulated one after the other. On average, λinΔt new nodes enter in the population per time step, while existing nodes can leave with probability Δt/τ. Nodes then form contacts according to the specific generative network algorithm. Eventually, transmission and recovery are simulated as explained above. In order to reconstruct the equilibrium dynamics we run simulations for a sufficiently long time span, discarding a transient time of 4 ⋅ 104 time steps. We verified that the dynamical properties at the equilibrium are unaffected by initial conditions.
For the hospital example, the network is an external parameter fed into the simulations. Contacts were aggregated daily keeping the information of their total duration. We used this information by considering a weighted network with the link weight, wij, representing the number of contacts of duration 30 s registered during the day between i and j. We then assumed Δt = 1 day and computed the probability of infection depending on the weight as 1 − ( 1 − β δ) w i j, with δ = 30 s. We initialized the system with the same configuration observed in the data, i.e. the initial status for each node is set according to S. aureus carriage during the starting week. Simulation length is bound to the hospital contact network duration.
In order to facilitate the comparison between the synthetic and the real scenarios, parameters of the network models were set based on the properties of the hospital network. The average size, the average activity potential and the average degree were set equal to the values estimated from the hospital network, i.e. V ¯ = 306, a ¯ = 0 . 28, k ¯ = 0 . 89 respectively. For the COM model the number of communities (nC = 6) and one of the two explored values of pIN (pIN = 0.78) were also informed by the data. Additional values of V ¯ and pIN were also tested. Epidemiological parameters were informed by the data in some cases—ps = 0.079 as computed from carriage data -, or chosen among plausible values for the S. aureus colonization—i.e. μ−1, that was set equal to either 21 or 35 days with other values from 14 to 49 days explored in the supporting information. Values of β were explored systematically. For consistency, values of rates throughout the manuscript were always expressed per hour.
Carriage data was obtained from weekly swabs in multiple body areas, including the nares. Swabs that resulted positive to S. aureus were further examined. Spa-type and antibiotic resistance profiles (MSSA or MRSA) were then determined. In this work we regard two strains as different if they differ in spa-type and/or antibiotic resistance profile. We considered carriage data obtained from nasal swabs dismissing other body areas since the anterior nares represent the most important niche for S. aureus [72].
We described strain population diversity through standard ecological indicators. The abundance of a strain i, Ni, is the strain-associated prevalence. From this quantity we computed the relative abundance, f i = N i ∑ i N i, and the relative abundance distribution, being the frequency of strains with relative abundance f. The Berger-Parker index is the relative abundance of the dominant strain, i.e. maxi fi.
To analyze repartition of strains across communities we use the Inverse Participation Ratio (IPR) [55]. The general definition of this quantity is the following. Given a vector v → with l components {vi}i=1,…,l, all within the range [0, 1], the IPR is given by:
I P R = ∑ i = 1 l v i 4 . (2)
If all the components are of the order (l−1) then the IPR is small. Instead if one component vi ∼ 1 then IPR ∼ 1 too, reflecting localization of v →. The IPR for total prevalence is computed by setting vi equal to the fraction of infected individuals belonging to community i = 1, …, l = nC, while the IPR for a single strain is computed by setting vi equal to the fraction of individuals infected by that particular strain and belonging to community i. We can extend the IPR computation to HOM case by assigning nodes to different groups as in COM but without affecting the stub-matching scheme.
In order to estimate the value of the length of stay maximizing the average richness for a given value of β when the contact structure is given by the HOM network we consider a homogeneous mixing version of our system.
Due to Eq (1) the calculation of the average richness reduces to the calculation of the average persistence time. In order to estimate such quantity we focus on a particular strain, labelled as “strain A”, which is injected at t = 0 and we group all other strains under the label “strain B”. We are allowed to do so because all strains have identical parameters. We therefore reduce our initial, multi-strain problem, to a two-strain problem. Since all new strains that will be injected after t = 0 will be labeled as strain B, it is clear that A is doomed to extinction since there exists an infinite reservoir of B. The average time to extinction is therefore the average time to extinction of strain A.
Since HOM network realizes quite well homogeneous mixing conditions we regard our system as homogeneously mixed. Within this framework it is sufficient to specify the numbers of hosts infected by strain A (nA), hosts infected by strain B (nB) and susceptible hosts (ns). The master equation for the joint probability distribution P(nA, nB, ns) is given by [73]:
P ˙ ( n A , n B , n s ) = β ′ V ¯ − 1 ( n A − 1 ) ( n s + 1 ) P ( n A − 1 , n B , n s + 1 ) + β ′ V ¯ − 1 ( n B − 1 ) ( n s + 1 ) P ( n A , n B − 1 , n s + 1 ) + μ ( n A + 1 ) P ( n A + 1 , n B , n s − 1 ) + μ ( n B + 1 ) P ( n A , n B + 1 , n s − 1 ) + λ o u t ( n A + 1 ) P ( n A + 1 , n B , n s ) + λ o u t ( n B + 1 ) P ( n A , n B + 1 , n s ) + λ o u t ( n s + 1 ) P ( n A , n B , n s + 1 ) + λ o u t V ¯ p s P ( n A , n B − 1 , n s ) + λ o u t V ¯ ( 1 − p s ) P ( n A , n B , n s − 1 ) − [ ( n A + n B ) ( β ′ V ¯ − 1 n s + μ ) + λ o u t ( n A + n B + n s ) + λ o u t V ¯ ] P ( n A , n B , n s ) , (3)
Where β ′ = β k ¯. Terms appearing on the right-hand side of the equation represent the probability flow associated to each transition event. The first four terms describe, in order, the infection due to strain A, the infection due to strain B, the recovery from A and the recovery from B. The remaining terms are then associated to the discharge of either one of the three types of individuals—infected with A, infected with B and susceptibles—and to the admission of infected of type B and susceptibles respectively. In order to obtain some approximate solution to this equation we assume that the average number of individuals nA + nB + ns and the total prevalence nA + nB do not fluctuate in time and are therefore equal to V ¯ and i ( ∞ ) V ¯ respectively, where i(∞) is given by:
i ( ∞ ) = β ′ − μ − λ o u t + ( β ′ − μ − λ o u t ) 2 + 4 β ′ λ o u t p s 2 β ′ . (4)
After performing the Van-Kampen size expansion we are left with a Fokker-Planck equation for the density of A f ( x = n A V ¯ ) = P ( n A ):
∂ t f = − ∂ x ( D 1 ( x ) f ) + 1 2 V ¯ ∂ x 2 ( D 2 ( x ) f ) , (5)
where D1 = β′ (1 − i(∞)) x − μ − λout and D2 = β′ (1 − i(∞)) x + μ + λout are the so-called drift and diffusion coefficients respectively.
According to the theory of stochastic processes [73] the average extinction time Tpers(x0) (where x0 represents the initial density of strain A) satisfies:
D 1 ( x 0 ) d d x 0 T pers + 1 2 V ¯ D 2 ( x 0 ) d 2 d x 0 2 T pers = − 1 , (6)
with boundary conditions Tpers(0) = 0 and d d x 0 T pers ( i ( ∞ ) ) = 0. The solution is finally given by:
T pers ( x 0 ) = i ( ∞ ) λ o u t p s [ E i ( − α i ( ∞ ) ) ( e α x 0 − 1 ) − e α x 0 E i ( − α x 0 ) + l n ( α x 0 ) + γ E ] , (7)
where Ei(x) is the exponential integral function and γE is Euler-Mascheroni constant. When a new strain is introduced its prevalence is just 1, therefore we estimate the average extinction time using T pers ( x 0 = V ¯ − 1 ).
|
10.1371/journal.ppat.1003893 | Glutamate Utilization Couples Oxidative Stress Defense and the Tricarboxylic Acid Cycle in Francisella Phagosomal Escape | Intracellular bacterial pathogens have developed a variety of strategies to avoid degradation by the host innate immune defense mechanisms triggered upon phagocytocis. Upon infection of mammalian host cells, the intracellular pathogen Francisella replicates exclusively in the cytosolic compartment. Hence, its ability to escape rapidly from the phagosomal compartment is critical for its pathogenicity. Here, we show for the first time that a glutamate transporter of Francisella (here designated GadC) is critical for oxidative stress defense in the phagosome, thus impairing intra-macrophage multiplication and virulence in the mouse model. The gadC mutant failed to efficiently neutralize the production of reactive oxygen species. Remarkably, virulence of the gadC mutant was partially restored in mice defective in NADPH oxidase activity. The data presented highlight links between glutamate uptake, oxidative stress defense, the tricarboxylic acid cycle and phagosomal escape. This is the first report establishing the role of an amino acid transporter in the early stage of the Francisella intracellular lifecycle.
| Intracellular bacterial pathogens have developed a variety of strategies to avoid degradation by the host innate immune defense mechanisms triggered upon phagocytocis. We show here for the first time that glutamate acquisition is essential for phagosomal escape and virulence of an intracellular pathogen. Remarkably, inactivation of the glutamate transporter GadC of Francisella impaired the capacity of the bacterium to neutralize reactive oxygen species (ROS) production in the phagosome. Virulence of the gadC mutant was partially restored in mice with a defective NADPH oxidase. Importantly, we found that impaired glutamate uptake affected the production of tricarboxylic acid (TCA) cycle intermediates, highlighting novel links between the TCA cycle and bacterial phagosomal escape. Amino acid transporters are, thus, likely to constitute underscored players in microbial intracellular parasitism.
| Francisella tularensis is a Gram-negative bacterium causing the disease tularemia in a large number of animal species. This highly infectious bacterial pathogen can be transmitted to humans in numerous ways [1], including direct contact with sick animals, inhalation, ingestion of contaminated water or food, or by bites from ticks, mosquitoes or flies. Four different subspecies (subsp.) of F. tularensis that differ in virulence and geographic distribution exist, designated subsps. tularensis, holarctica, mediasiatica and novicida, respectively. The tularensis subspecies is the most virulent causing a severe disease in humans [2], [3]. F. tularensis subsp. novicida (F. novicida) is rarely pathogenic to non-immuno-compromized humans but is fully virulent for mice and is therefore widely used as a model to study Francisella intracellular parasitism.
F. novicida has the capacity to evade host defenses and to replicate to high numbers within the cytosol of eukaryotic cells [4]. The bacterium is able to enter and to replicate inside a variety of cells, and in particular in macrophages. After a transient passage through a phagosomal compartment, bacteria are released within 30–60 minutes in the host cell cytosol where they undergo several rounds of active replication [1]. Upon Francisella entry into macrophages, the phagosomal compartment transiently acidifies and the activation of NADPH oxidase leads to the production of noxious oxygen reactive species [5]. Although several genes required for phagosomal escape have been identified ([6], [7] and references therein), the molecular mechanisms underlying this complex process are still very poorly understood.
Protection against oxidative stress includes the production of anti-oxidant molecules (such as glutathione and NADPH) and of enzymes (such as catalases, superoxide dismutases glutaredoxin-related protein and alkylhydroperoxide reductases). Francisella subspecies encode a whole set of such oxidative stress-related enzymes [8]. Inactivation of the corresponding genes generally leads to increased sensitivity to oxidative stress, defective intracellular multiplication, and attenuated virulence [9], [10], [11]. Protection against oxidative and other stress also involves a number of dedicated protein chaperones and chaperone complexes [12].
In contrast, the importance of acid-resistance mechanisms in Francisella intracellular survival remains controversial [13], [14], [15] and their possible contribution to pathogenesis still largely unknown. One of the best characterized acid-resistance systems in bacteria couples the glutamate:γ-aminobutyrate exchanger GadC with the glutamate decarboxylase(s) GadA and/or GadB [16]. The decarboxylase replaces the α-carboxyl group of its amino acid substrate with a proton that is consumed from the cytoplasmic pool [17]. The capacity to produce γ-aminobutyric acid (GABA) through glutamate decarboxylation has been observed in both Gram-negative and Gram-positive bacteria. The GadC/GadB glutamate decarboxylase (GAD) system has been shown to play an essential role in acid tolerance in food-borne bacterial pathogens that must survive the potentially lethal acidic environments of the stomach before reaching the intestine. Some bacteria possess a unique permease-decarboxylase pair whereas others, like Listeria monocytogenes [18], encode several paralogues of each component.
Recent genome sequence analyses and genome-scale genetic studies suggest that an important proportion of genes related to metabolic and nutritional functions participate to Francisella virulence [19]. However, the relationship between nutrition and the in vivo life cycle of F. tularensis remain poorly understood. Francisella is predicted to possess numerous nutrient uptake systems to capture its necessary host-derived nutrients, some of which are probably available in limiting concentrations. Notably, we showed very recently that an asparagine transporter of the major facilitator superfamily of transporters was specifically required for cytoslic multiplication of Francisella and its systemic dissemination [20].
The amino acid-polyamine-organocation family of transporters (APC) is specifically involved in amino acid transport [19]. Remarkably, eight of the 11 APC members have been identified at least once in earlier genetic studies, and are likely to be involved in bacterial virulence. In particular, the gene encoding the GadC permease has been identified in several different genome-wide screens, performed in either F. tularensis subsp. holarctica [21] or F. novicida [22], [23], [24].
In the present work, we elucidate the functional role of the GadC protein in Francisella pathogenesis. We show that glutamate uptake plays a critical role in Francisella oxidative stress defense in the phagosomal compartment. Strikingly, the activity of GadC influences the expression of metabolic genes and the production of tricarboxylic acid (TCA) cycle intermediates, unraveling a relationship between oxidative stress defense, metabolism and Francisella virulence.
F. tularensis subspecies possess a unique putative GAD system, composed of the antiporter GadC and a decarboxylase GadB (encoded by genes FTN_0571 and FTN_1701 in F. novicida and hereafter designated gadC and gadB, respectively for simplification) (Figure S1A). The transcription of gadC is initiated 27 nucleotides upstream of the translational start from a predicted σ70 promoter (Figure S1B). This genetic organization is highly conserved in all the available F. tularensis genomes (not shown). The gene gadC encodes a protein of 469 amino acids sharing 98.7%, 99.1% and 99.6% identity with its orthologues in the subspecies mediasiatica (FTM_1423), holartica (FTL_1583) and tularensis (FTT_0480c), respectively.
The Francisella GadC protein is predicted as a putative glutamate:γ-aminobutyric acid (GAD) antiporter (KEGG database). Although it shows only modest homology (approximately 25% amino acid identity) with GadC of E. coli [25], secondary structure prediction (using the method for prediction of transmembrane helices HMM available at the internet site www.cbs.dtu.dk) indicates that the GadC transporter of Francisella also comprises 12 transmembrane helixes and has its N and C-terminal ends facing the cytoplasm (not shown).
The gadB gene encodes a putative glutamate decarboxylase protein of 448 amino acid residues that is highly conserved in F. tularensis subsp. tularensis (98.7% amino acid identity with FTT_1722c). However, the corresponding protein is truncated at its C-terminal end in the subspecies holarctica (FTL_1863) and mediasiatica (FTM_1673, and noted as a pseudogene in the KEGG database).
We constructed a strain with chromosomal deletion of the entire gadC gene in F. novicida by allelic replacement [26]. We confirmed that the ΔgadC mutation did not have any polar effect on the downstream gene FTN_0570 by quantitative qRT-PCR (Figure S1C). The growth kinetics of the parental F. novicida strain and the ΔgadC mutant were indistinguishable in tryptic soya broth (TSB) and chemically defined medium (CDM) [27] liquid media at 37°C (Figure S2), indicating that inactivation of gadC had no impact on bacterial growth in broth.
We examined the ability of wild-type F. novicida, the ΔgadC mutant and a ΔgadC mutant strain complemented with a plasmid-encoded copy of wild-type gadC, to survive in murine and human macrophage cell lines and primary bone marrow-derived mouse macrophages, over a 24 h-period. The ΔgadC mutant showed a severe growth defect in J774.1 cells, comparable to that of a mutant deleted of the entire Francisella pathogenicity island (ΔFPI mutant), with more than a 30-fold reduction of intracellular bacteria after 10 h and a 1,000-fold reduction after 24 h (Figure 1A). Impaired multiplication of the ΔgadC mutant was also observed in THP-1 macrophages (Figure 1B) as well as in bone marrow-derived macrophages (Figure 1C). In all cell types tested, introduction of the complementing plasmid (pKK-gadC) restored bacterial viability to same level as in the wild-type parent, confirming the specific involvement of the gadC gene in intracellular survival.
Next, in vivo competition assays in BALB/c mice were performed to determine if the GadC protein played a role in the ability of Francisella to cause disease. Five mice (6- to 8-week old) were inoculated by the intraperitoneal (i.p.) route with a 1∶1 mixture of wild-type F. novicida and ΔgadC mutant strains. Bacterial multiplication in the liver and spleen was monitored at day 2 post-infection (Figure 1D). The Competition Index (CI), calculated for both organs, was extremely low (10−6) demonstrating that the gene gadC played an essential role in Francisella virulence in the mouse model.
Upon Francisella entry into cells, Francisella initially resides in a phagosomal compartment that transiently acidifies and that acquires reactive oxygen species. We therefore examined the ability of wild-type and ΔgadC mutant strains to survive under acid or oxidative stress conditions. For this, bacteria were exposed either to pH 5.5 or to 500 µM H2O2 (Figure 2). Under the pH condition tested, the viability of two strains was unaffected (Figure 2A). It should be noted that at the lower pH of 2.5, the viability of both wild-type and ΔgadC mutant was equally reduced (approximately 2 logs, not shown). In contrast, the ΔgadC mutant strain appeared to be significantly more sensitive to oxidative stress than the wild-type strain in TSB (Figure 2B). After 40 min of exposure, it showed a 10-fold decrease in the number of viable bacteria and an approximately 50-fold decrease after 60 min of exposure to H2O2. Remarkably, in CDM, the wild-type and ΔgadC mutant strains were equally sensitive to H2O2 in the absence of glutamate supplementation (Figure 2C). However, upon glutamate supplementation, the wild-type strain showed increased resistance to H2O2 whereas the ΔgadC strain was unaffected (Figure 2D).
Confocal and electron microscopy analyses demonstrated that the ΔgadC mutant had lost the capacity to escape from the phagosomal compartment of infected macrophages.
Earlier phylogenetic studies have distinguished ten distinct subfamilies within the APC family of transporters, inferring possible substrate specificities. Consensus signature motifs were defined for each of them [34]. Inspection of the Francisella GadC protein reveals a signature sequence of the Glutamate-GABA antiporter subfamily in its N-proximal portion (Figure 4A), prompting us to test functional complementation of an E. coli gadC mutant by the Francisella gadC orthologue.
Functional complementation (Figure 4B) was determined by comparing the acid resistance (at pH 2.5) of a gadC-inactivated strain of E. coli (EF491) to the same strain carrying a plasmid-borne F. novicida gadC gene (pCRT-gadC). As a positive control, we used the E. coli gadC mutant complemented with the wild-type E. coli gadC gene (EF547). IPTG-induced expression of the Francisella gadC allele in the E. coli gadC mutant strain restored acid resistance to wild-type level, indicating that the Francisella GadC protein displays the acid-resistance function of the E. coli GadC protein.
To further support the role of GadC in glutamate entry, we quantified the amounts of intracellular glutamate by HPLC analysis, in the wild-type and ΔgadC strains grown in CDM supplemented with 1.5 mM of glutamate (in the presence or in the absence of hydrogen peroxide). As shown in Fig. 4C, the concentration of intracellular glutamate was significantly lower in the ΔgadC mutant than in the wild-type strain, both in the absence (84% reduction in concentration) or in the presence (31% reduction) of H2O2. We also quantified the amount of glutamate in culture supernatants of the two strains in the presence of H2O2 (not shown). External glutamate present in the culture medium of the wild-type strain was 39% lower than that of the ΔgadC mutant.
Altogether these data are compatible with a reduced capacity of the ΔgadC mutant to take up external glutamate.
We then directly evaluated the impact of gadC inactivation on glutamate uptake by live F. novicida. For this, we compared the uptake of radiolabeled glutamate (14C-Glu) by wild-type F. novicida to that of the ΔgadC mutant, over a broad range of glutamate concentrations (Fig. 4D). Incorporation of 14C-Glu was significantly affected in the ΔgadC mutant (representing only approximately 50% of the wild-type values at each concentration tested), confirming that GadC is a genuine glutamate transporter. The fact that glutamate uptake was not totally abolished in the ΔgadC mutant suggests that other transporter(s) allow the entry of glutamate in this strain.
We compared the amount of reactive oxygen species (ROS) in J774.1 cells infected either with wild-type F. novicida, ΔgadC or the ΔFPI strain, over a 60 min period. For this, we used the H2DCF-DA assay (Sigma-Aldrich Co). H2DCF-DA is a non-fluorescent cell-permeable compound that has been widely used for the detection of ROS [35]. Once inside the cell, this compound is first cleaved by endogenous esterases to H2DCF. The de-esterified product becomes the highly fluorescent compound 2′,7′-dichlorofluorescein (DCF) upon oxidation by ROS. The ROS content increased by 25% after 60 min in cells infected with wild-type F. novicida (Figure 5). A comparable increase was recorded with the ΔFPI mutant. However, in cells infected with the ΔgadC mutant, the ROS content was significantly higher than that recorded with the two other strains at each time point (25% higher at 15 min, and 55% higher after 60 min). These results suggest that the ΔgadC mutant is affected in its ability to neutralize the production of ROS in the phagosomal compartment. Alternatively, the ΔgadC mutant may trigger an increased production of ROS.
This result prompted us to evaluate the pathogenicity of the ΔgadC mutant in mice lacking a functional NADPH oxidase complex, both in vitro and in vivo.
Intracellular glutamate plays a central role in a wide range of metabolic processes in bacteria. In order to evaluate the potential impact of the gadC inactivation on bacterial glutamate metabolism, we first quantitatively monitored the transcription of selected genes connecting glutamate utilization to either the TCA cycle or to glutathione biogenesis. This analysis was done for wild-type F. novicida and for the ΔgadC mutant strain, grown in broth with or without H2O2 (Figure 7A).
Expression of FTN_0593 (sucD), FTN_0127 (gabD) and FTN_1532 (gdhA), was significantly decreased in the ΔgadC mutant under oxidative stress, whereas their expression was moderately increased in the wild-type strain. Expression of FTN_0277(gshA) and FTN_0804 (gshB) was reduced in both strains, under oxidative stress. However, the decrease was significantly less important (app. 4-fold) in the ΔgadC mutant than in the wild-type strain. Expression of FTN_1635 (sucA) was significantly decreased in both strains under oxidative stress. These data indicate that the absence of gadC affects the expression of several genes linked to glutamate metabolism under oxidative stress. The fact that expression of the gadC gene itself was significantly upregulated (approximately 10-fold) in the wild-type strain exposed to H2O2 stress (not shown) supports the importance of the GadC transporter in oxidative stress defense.
Direct quantification of TCA cycle intermediates present in the cytoplasm of the wild-type and ΔgadC strains, by gas chromatography coupled with mass spectrometry (see Materials and methods for details), revealed that gadC inactivation significantly affected succinate, fumarate, and oxoglutarate contents (Figure 7C). Indeed, in the ΔgadC mutant, the concentrations of succinate and fumarate were reduced ca. 60% as compared to the wild-type strain, whereas oxoglutarate was below the detection threshold of the assay. The concentrations of the three molecules increased up to 40% in the wild-type strain exposed to oxidative stress, suggesting an activation of the TCA under this condition. The concentrations of succinate and fumarate were not significantly modified in the ΔgadC mutant upon oxidative stress and oxoglutarate production was still below detection. The concentration of citrate was similar in the wild-type and the ΔgadC mutant and did not vary upon oxidative stress, in any of the two strains. The intracellular concentrations of glutathione were also almost similar in the wild-type and ΔgadC mutant (Figure 7B). Remarkably, under oxidative stress, the intracellular concentration of glutathione increased in both strains but only reached 65% of the level of the ΔgadC mutant in the wild-type strain.
These observations prompted us to evaluate the impact of supplementation with different TCA cycle intermediates on survival of the ΔgadC mutant in response to H2O2 challenge. For this, exponential phase wild-type and ΔgadC mutant strains, diluted in CDM supplemented with glutamate, were subjected to oxidative stress, in the presence or absence of either fumarate, succinate or oxoglutarate (Figure S5). The sensitivity to H2O2 of the ΔgadC mutant was not modified neither by fumarate nor by oxoglutarate. In contrast, supplementation with succinate increased significantly the survival of the ΔgadC mutant, to nearly wild-type level.
Intracellular pathogenic bacteria have adapted a variety of strategies and specific intracellular niches for survival and multiplication within their host [36]. Some reside in a vacuolar compartment whereas others have evolved to gain access to the host cytosol for multiplication. In mammalian host cells, Francisella intracellular replication occurs exclusively in the cytosolic compartment. We show here that inactivation of the GadC permease of Francisella prevents phagosomal escape, thus severely altering bacterial intracellular multiplication and virulence.
The data presented suggest that the GadC protein of Francisella is required to resist to the oxidative burst triggered by the NADPH oxidase in the phagosomal compartment of infected macrophages. We propose that GadC-mediated entry of glutamate contributes to fuel the tricarboxylic acid cycle and modulates the redox status of the bacterium. This work thus provides insights into the possible links between oxidative stress resistance, metabolism, and bacterial intracellular parasitism.
Inactivation of gadC in F. novicida led to a severe growth defect in all cell types tested and in vivo assays further demonstrated the importance of GadC in Francisella virulence. Confocal and electron microscopy analyses revealed that the severe intracellular growth defect of the mutant was due to its inability to escape from the phagosomal compartment of infected macrophages.
Interestingly, most of the mutant bacteria that remained trapped within the phagosome were still alive for at least 10 h post-infection, indicating that impaired escape was not due to bacterial death. Since the ΔgadC mutant showed increased susceptibility to oxidative stress in broth and failed to efficiently neutralize reactive oxygen species production in cells, it is likely that ROS may predominantly affect bacterial escape rather than survival.
F. tularensis produces enzymes that can metabolize and neutralize ROS, such as a superoxide dismutases (SodB, SodC), a catalase (KatG), a glutathione peroxidase and a peroxireductase [9], [10]. Acid phosphatases have also been implicated in the resistance of intracellular Francisella to H2O2 generated in the phagosomal compartment by the NADPH oxidase ([37], [38] and references therein). However, inactivation of the major phosphatase acpA in F. tularensis subsp. tularensis, had no impact on the activity of the NADPH oxidase in human neutrophils [5], thus confirming that other Francisella factors were involved in NADPH oxidase inhibition.
We show here that Francisella GadC is an important player specifically involved in oxidative stress defense. The existence of several paralogues of both the transporter GadC and the decarboxylase GadB in some bacterial species (for example in L. monocytogenes) might account for the fact that these have not yet been found to contribute to oxidative stress resistance and intracellular survival in standard genetic screens. Indeed, if functional paralogues exist, they must be simultaneously inactivated to observe a possible phenotypic defect. In addition, isofunctional antiporters with no significant amino acid sequence similarity to the GadC protein might exist in these bacteria.
The contribution of the GAD system to intracellular survival critically depends on the cellular compartment where bacterial survive and multiply. Indeed, bacteria residing in vacuolar compartments (such as Salmonella, Mycobacteria, Legionella, Brucella and Chlamydia) encounter different types of stresses (pH, oxidative, nutritional,…) than bacteria able to multiply in the host cell cytosol (such as Francisella, Listeria, Shigella and Rickettsia).
L-glutamate is very abundant in the intracellular compartment (reported concentrations vary between 2 and 20 mM) when compared to the extracellular compartment (app. 20 µM) [39]. Human macrophages have both the cystine/glutamate transporter and the Na-dependent high-affinity glutamate transporters (excitatory amino acid transporters, EAATs) that transport glutamate and aspartate. To maintain their intracellular pool of glutamate, macrophages may use either these transporters to import glutamate from the extracellular milieu or enzymatically convert cytosolic glutamine (via glutaminase) and aspartate (via aspartate transpeptidase) to glutamate. Glutamate might also be produced spontaneously intracellularly from pyroglutamate. Currently, nothing is known with respect to the content of glutamate in the phagosomal compartment. This might prove extremely difficult to establish, especially for pathogens such as Francisella or Listeria that reside only very transiently in this compartment.
A limited number of bacterial species have been shown to possess a GAD system [40]. These include E. coli, Lactobacillus, L. monocytogenes and Shigella species, in which the GAD system plays a major role in acid tolerance. It has been suggested that the GAD system is important for pathogenic microorganisms that, upon oral infection of mammalian hosts, need to survive the low pH of the stomach. However, some enteric pathogens like Salmonella do not possess a function GAD system and must thus rely on other anti-acidic pH strategies. Interestingly, the GAD system has been also found to contribute to oxidative stress defense in yeast and plant [41]. In bacteria, molecules such as the NADPH and NADH pools and glutathione (GSH), contribute to oxidative stress defense. Reduced GSH, present at mM concentrations, maintain a strong reducing environment in the cell. Specific enzymes are also dedicated to control the levels of reactive oxygen species (ROS).
Remarkably, the ΔgadC mutant was still outcompeted by wild-type bacteria in phox-KO mice. The different environments and the immune pressure, encountered by the bacterium during its systemic dissemination, are probably far more complex than in culture systems. In vivo, Francisella GadC is thus likely to contribute to other functions than combat ROS in the phagosomal compartment. It may, for instance, fulfill classical nutritional functions during bacterial cytosolic multiplication (in macrophages and/or in other infected non-phagocytic cells). Alternatively, GadC may be required during the bacterial blood stage multiplication and dissemination of the bacterium.
In E. coli, GABA produced by the glutamate decarboxylase is metabolized via the GABA shunt pathway. This leads to the production of succinate via the consecutive action of two enzymes: a GABA/oxoglutarate amino-transferase (GabT) that removes the amino group from GABA to form succinic semialdehyde (SSA) and Glu; and a succinic semialdehyde dehydrogenase (GabD) that oxidises SSA to form succinate. Very recently, Karatzas and co-workers have shown [42] that L. monocytogenes also possessed functional GabT and GabD homologues that could provide a possible route for succinate biosynthesis in L. monocytogenes. The GABA shunt pathway, allowing the bypass of two enzymatic steps of the TCA (from oxoglutarate to succinate; Figure 7 and S6), is thought to play a role in glutamate metabolism, anaplerosis and antioxidant defense. However, its physiological role in pathogenesis is yet poorly understood. Francisella genomes possess a gabD orthologue but lack gabT. The GABA shunt pathway may therefore be non-functional in Francisella. Interestingly, the isogenic glutamate decarboxylase ΔgadB mutant (Figure S4) that we constructed, showed a much less severe intracellular multiplication defect than the ΔgadC mutant, and as well as no (or only a very mild) attenuation of virulence. If the glutamate imported via GadC would serve to produce GadB-mediated GABA, one would expect gadB inactivation to cause the same defect as gadC inactivation. As already mentioned in the Introduction, the gadB orthologue encodes a truncated protein in the subspecies holarctica. Altogether, these data support the notion that GadC and GadB of Francisella do not function in concert, unlike in several other bacterial species, and that GABA production plays a marginal role in Francisella pathogenesis. Further work will be required to understand the exact contribution of GadB in Francisella metabolism.
Our data indicate that GadC of Francisella encodes a genuine glutamate transporter involved in oxidative stress, unlike most other GadC orthologues described thus far. Glutamate can be converted in the bacterial cytoplasm into a number of compounds (Figure 7), such as glutamine, glutathione, GABA or the TCA cycle intermediate oxoglutarate. Oxoglutarate is known to be a potent anti-oxidant molecule that can be converted, in absence of any enzymatic reaction, into succinate in the presence of H2O2. In addition, conversion of glutamate to oxoglutarate by the glutamate dehydrogenase GdhA increases the production of NADPH, which might also contribute to the anti-oxidant effect of glutamate acquisition.
Quantitative analyses of the intra-bacterial content of TCA cycle intermediates (Figure 7B) revealed a significant reduction of succinate and fumarate in the gadC mutant, as compared to wild-type F. novicida, and a striking decrease of oxoglutarate. These data support the notion that reduced entry of glutamate directly affects the production of these TCA cycle intermediates. In contrast, the amount of citrate remained unchanged in the mutant, suggesting refueling of the TCA cycle via other entry points (such as glycolysis or amino acid conversion).
Of note, mutants in genes gdhA (FTN_1533) and gabD (FTN_0127) were identified as required for replication in D. melanogaster S2 cells in a recent screen, supporting a role for these genes in intracellular bacterial survival [43]. The production and utilization of oxoglutarate by Francisella may thus constitute an efficient mean to modulate its cytoplasmic concentration of ROS.
In the absence of external glutamate, the pool of intracellular glutamate may be synthesized either from oxoglutarate, glutamine, GSH or even proline (according to KEGG metabolic pathways). Therefore, we evaluated the impact of gadC inactivation on the expression of genes involved in glutamate metabolism, under oxidative stress conditions. qRT-PCR analyses were performed in wild-type F. novicida and in the ΔgadC mutant, grown in chemically defined medium containing glutamate, in the absence or in the presence of H2O2 (Figure 7A). These assays revealed that gadC inactivation led to an important down-regulation of the genes involved the conversion of glutamate to oxoglutarate and succinate, upon oxidative stress (FTN_1532 and FTN_0127, respectively). Conversely, gadC inactivation only moderately decreased the expression of gshA and gshB, the two genes involved in glutathione biosynthesis (FTN_0277 and FTN_0804), upon oxidative stress whereas the expression of these genes was severely decreased in the wild-type strain.
These data are compatible with the notion that, under oxidative stress, the wild-type strain may favor the conversion of a fraction of its cytoplasmic pool of glutamate (neosynthesized and imported) to produce oxoglutarate and succinate rather than GSH. In contrast, when the cytosolic pool of glutamate is restricted to neosynthesized glutamate (i.e. in a gadC mutant or in a glutamate-depleted medium), the production of oxaloglutarate and succinate may be decreased to favor that of other molecules (including GSH).
In conclusion, we identified a glutamate transporter as a novel Francisella virulence attribute that suggests links between the oxidative stress response and the TCA cycle during the early stage of the bacterial intracellular life cycle. The importance of the TCA cycle in the homeostasis of reactive oxygen species has just started to be considered in pathogenic bacterial species [12], [44], [45], [46]. The development of specific inhibitors of transport systems involved in intracellular adaptation might constitute interesting anti-bacterial therapeutic targets.
All experimental procedures involving animals were conducted in accordance with guidelines established by the French and European regulations for the care and use of laboratory animals (Decree 87–848, 2001–464, 2001–486 and 2001–131 and European Directive 2010/63/UE) and approved by the INSERM Ethics Committee (Authorization Number: 75-906).
F. tularensis subsp. novicida (F. novicida) strain U112, its ΔFPI derivative, and all the mutant strains constructed in this work, were grown as described in Supplementary Material. E. coli strains (kindly provided by John Foster, University of South Alabama, USA) were grown as described in Supplementary Material. All bacterial strains, plasmids, and primers used in this study are listed in Supplemental Table 1.
Details of the construction and characterization of mutant and complemented strains; macrophage preparation and infections, are described in Supplementary Material. Quantitative (q)RT-PCR (real-time PCR) was performed with gene-specific primers (Supplemental Table 1), using an ABI PRISM 7700 and SYBR green PCR master mix (Applied Biosystems, Foster city, CA, USA).
Electron and confocal microscopy complete descriptions; real time cell death and phagosome permeabilization assays, are described in Supplementary Material.
J774.1 macrophage-like cells (ATCC Number: TIB-67) were propagated in Dulbecco's Modified Eagle's Medium (DMEM) containing 10% fetal calf serum, whereas human monocyte-like cell line THP-1 (ATCC Number: TIB-202) and bone marrow-derived macrophages (BMM) from BALB/c were propagated in RPMI Medium 1640 containing 10% fetal calf serum, respectively. J774.1 and BMM were seeded at a concentration of ∼2×105 cells per well in 12-well cell tissue plates and monolayers were used 24 h after seeding. THP-1 were seeded at a concentration of ∼2×105 cells per well in 12-well cell tissue plates 48 h before infection, and supplemented with phorbol myristate acetate (PMA) to induce cell differentiation (200 ng/ml). J774.1, BMM and THP-1 were incubated for 60 min at 37°C with the bacterial suspensions (approximately multiplicities of infection 100) to allow the bacteria to enter. After washing (time zero of the kinetic analysis), the cells were incubated in fresh culture medium containing gentamicin (10 µg mL−1) to kill extracellular bacteria. At several time-points, cells were washed three times in DMEM or RPMI, macrophages were lysed by addition of water and the titer of viable bacteria released from the cells was determined by spreading preparations on Chocolate agar plates. For each strain and time in an experiment, the assay was performed in triplicate. Each experiment was independently repeated at least three times and the data presented originate from one typical experiment.
Bacteria were centrifuged for 2 min in a microcentrifuge at room temperature and the pellet was quickly re-suspended in Trizol solution (Invitrogen, Carlsbad, CA, USA). Samples were either processed immediately or frozen and stored at −80°C. Samples were treated with chloroform and the aqueous phase was used in the RNeasy Clean-up protocol (Qiagen, Valencia, CA, USA) with an on-column DNase digestion of 30 min [47].
RNA Reverse transcription (RT)-PCR experiments were carried out with 500 ng of RNA and 2 pmol of specific reverse primers. After denaturation at 65°C for 5 min, 6 µL of the mixture containing 4 µL of 5× first strand buffer and 2 µL of 0,1 M DTT were added. Samples were incubated 2 min at 42°C and, then, 1 µL of Superscript II RT (Thermo Scientific) was added. Samples were incubated for 50 min at 42°C, heated at 70°C for 15 min and chilled on ice. Samples were diluted with 180 µL of H2O and stored at −20°C.
The following pair of primers was used to amplify the mRNA corresponding to the transcript of FTN_0570 (p13/p14), FTN_0571 (p15/p16), FTN_1700 (p27/9p28), FTN_1701 (p29/p30), FTN_1702 (p31/p32), FTN_1532 (p33/p34), FTN_0127 (p35/p36), FTN_0277 (p37/p38), FTN_0804 (p39/p40), FTN_0593 (p41/p42), FTN_1434 (p43/p44) and FTN_1635 (p45/p46) (Supplemental Table 1).
Wild-type F. novicida and mutant strains were grown at 37°C from OD600 ∼0.1. After 4 h of incubation, samples were harvested and RNA was isolated. For oxidative stress tests, samples were cultivated 30 min more with or without H2O2 (500 µM). The 25 µL reaction consisted of 5 µL of cDNA template, 12.5 µL of Fastart SYBR Green Master (Roche Diagnostics), 2 µL of 10 µM of each primer and 3.5 µL of water. qRT-PCR was performed according manufacturer's protocol on Applied Biosystems - ABI PRISM 7700 instrument (Applied Biosystems, Foster City, CA). To calculate the amount of gene-specific transcript, a standard curve was plotted for each primer set using a series of diluted genomic DNA from wild-type F. novicida. The amounts of each transcript were normalized to helicase rates (FTN_1594).
Stationary-phase bacterial cultures were diluted at a final OD600 of 0.1 in TSB broth or CDM with or without glutamate (1.5 mM final). Exponential-phase bacterial cultures were diluted to a final concentration of 108 bacteria mL−1 and subjected to either 500 µM H2O2 or pH 5.5.
Oxidative stress response was also tested in CDM supplemented with glutamate, in the presence or absence of the TCA cycle intermediates: oxoglutarate, succinate or fumarate (1.5 mM final). The number of viable bacteria was determined by plating appropriate dilutions of bacterial cultures on Chocolate Polyvitex plates at the start of the experiment and after the indicated durations. Cultures (5 mL) were incubated at 37°C with rotation (100 rpm) and aliquots were removed at indicated times, serially diluted and plated immediately. Bacteria were enumerated after 48 h incubation at 37°C. Experiments were repeated independently at least twice and data represent the average of all experiments.
J774.1 cells were infected with wild-type F. novicida, ΔgadC or ΔFPI strains for 1 h, 4 h and 10 h at 37°C, and were washed in KHM (110 mM potassium acetate, 20 mM Hepes, 2 mM MgCl2). Cells were incubated for 1 min with digitonin (50 µg/mL) to permeabilize membranes. Then cells were incubated for 15 min at 37°C with primary anti F. novicida mouse monoclonal antibody (1/500 final dilution, Immunoprecise). After washing, cells were incubated for 15 min at 37°C with secondary antibody (Ab) (Alexa Fluor 488-labeled GAM, 1/400 final dilution, Abcam) in the dark. After washing, cells were fixed with PFA 4% for 15 min at room temperature (RT) and incubated for 10 min at RT with 50 mM NH4Cl to quench residual aldehydes. After washing with PBS, cells were incubated for 30 min at RT with primary anti-LAMP1 Ab (1/100 final dilution, Abcam) in a mix with PBS, 0.1% saponine and 5% goat serum. After washing with PBS, cells were incubated for 30 min at RT with secondary anti-rabbit Ab (alexa 546-labeled, 1/400 dilution, Abcam). DAPI was added (1/5,000 final dilution) for 1 min. After washing, the glass coverslips were mounted in Mowiol. Cells were examined using an X63 oil-immersion objective on a LeicaTSP SP5 confocal microscope. Co-localization tests were quantified by using Image J software; and mean numbers were calculated on more then 500 cells for each condition. Confocal microscopy analyses were performed at the Cell Imaging Facility (Faculté de Médecine Necker Enfants-Malades).
Infection of J774.1 cells was followed by thin section electron microscopy as previously described [48].
To evaluate the viability of F. tularensis, labelings were adapted to use the cell-impermeant nucleic acid dye propidium iodide (PI). J774.1 macrophage-like cells were seeded at 5.105cells/ml on glass coverslips in 12-well bottom flat plates. Next day, cells were infected for 10 h with wild-type F. novicida or ΔgadC strain. After infection, cells were first permeabilized with digitonin for 1 min, washed three times with KHM and incubated for 12 min at 37°C with 2.6 µM PI (Life technologies, L7007) in KHM buffer to label compromised bacteria in permeabilized cells. Cells were washed three times with KHM and incubated for 15 min at 37°C with primary anti F. novicida mouse monoclonal antibody (1/500 final dilution). After washing, cells were incubated for 15 min at 37°C with secondary antibody (Ab) (Alexa Fluor 488-labeled GAM, 1/400 final dilution) in the dark. After washing, cells were fixed with PFA 4% for 15 min at room temperature (RT) and incubated for 10 min at RT with 50 mM NH4Cl to quench residual aldehydes. After washing with PBS, cells were incubated for 30 min at RT with primary anti-LAMP1 Ab (1/100 final dilution) in a mix with PBS, 0.1% saponine and 5% goat serum. After washing with PBS, cells were incubated for 30 min at RT with secondary anti-rabbit Ab (alexa 405-labeled, 1/400 dilution). After washing, the glass coverslips were mounted in Mowiol. Cells were examined using an X63 oil-immersion objective on a LeicaTSP SP5 confocal microscope. Analysis of cell fluorescence was performed with Image J software (http://rsb.info.nih.gov/ij).
Intracellular reactive oxygen species (ROS) were detected by using the oxidation-sensitive fluorescent probe dye, 2′,7′-dichlorodihydrofluorescein diacetate (H2DCF-DA) as recommended by the manufacturer (CM-H2DCF-DA, Molecular Probes, Eugene, OR). J774.1 cells were seeded at 5.104 cells/well. Cells were infected with bacteria for 15 min (MOI of 100∶1), washed three times with PBS and incubated with H2DCF-DA diluted in PBS (concentration). DCF fluorescence was measured with the Victor2 D fluorometer (Perkin-Elmer, Norwal, CT) with the use of excitation and emission wavelengths of 480 nm and 525 nm, respectively. Values were normalized by protein concentration in each well (Bradford). Samples were tested in triplicates in three experiments.
J774.1 cells were seeded at 5.104 cells/well. Cells were infected with bacteria for 15 min (MOI of 1,000∶1), washed three times with PBS and incubated with H2DCF-DA diluted in PBS for 1 h (5 µM). Images of the cells were captured with an Olympus CKX41 microscope and treated with Image J software. Cell counts were performed over 10 images of approximately 50 cells.
Acid resistance tests in E. coli were performed at pH 2.5 as described previously [49], by comparing the number of survival treated cells after 1 h of treatment over the number of cells at T0. We compared survival of wild-type E. coli strain (WT) with E. coliΔgadC (Δ) and E. coliΔgadC complemented with the F. novicida gadC gene (PCR-amplified gene FTN-0571 introduced into plasmid pCR2.1-Topo, in the correct orientation downstream of the plac promoter) (Comp). Acid challenge was performed by diluting 1∶100 the overnight (22 h) culture in LB, supplemented (Comp +) or not (Comp −) with 10−4 M final IPTG.
Glutamate detection and quantification was done by using HPLC analysis. Wild-type F. novicida and ΔgadC strains were tested in CDM supplemented with 1.5 mM of glutamate, with or without H2O2 (500 µM). For each condition, three independent cultures were prepared by overnight growth in CDM. The overnight cultures were diluted with 50 mL of fresh medium to OD600 of 0.05 and cultivated to an OD600 of app. 0.35. Bacteria were harvested by centrifugation at 4,000× g for 20 min, resuspended in 25 mL of pre-warmed appropriated medium and cultivated for 30 min. For extracellular glutamate dosage, 100 µL of each supernatant were resuspended with 400 µL of cold methanol and centrifuged at 12,000× g for 5 min. 20 µL of each preparation were derivatized with 80 µL of OPA. For intracellular dosage, each sample were resuspended in 600 µL of cold methanol. The bacterial suspensions were sonicated thrice for 30 sec at 4.0 output, 70% pulsed (Branson Sonifier 250). Lysates were then centrifuged at 3,300× g for 8 min, to remove debris.
Following steps were done with the standard procedure of Agilent using ZORBAX Eclipse AAA high as HPLC column. An amount equivalent to 2 µL of each sample was injected on a Zorbax Eclipse-AAA column, 5 µm, 150×4.6 mm (Agilent), at 40°C, with fluorescence detection. Aqueous mobile phase was 40 mM NaH2PO4, adjusted to pH 7.8 with NaOH, while organic mobile phase was acetonitrile/methanol/water (45/45/10 v/v/v). The separation was obtained at a flow rate of 2 mL min−1 with a gradient program that allowed for 2 min at 0% B followed by a 16-min step that raised eluent B to 60%. Then washing at 100% B and equilibration at 0% B was performed in a total analysis time of 38 min. To evaluate glutamate concentration, glutamate standard curve was made in parallel.
The procedure for the measurement of GSH was previously described [50]. Briefly, GSH were separated by HPLC, equipped with a Shimadzu Prominence solvent delivery system (Shimadzu Corp., Kyoto, Japan), using a reverse-phase C18 Kromasil (5 µm; 4.6×250 mm), obtained from AIT (Paris, Fr). The mobile phase for isocratic elution consisted of 25 mmol L−1 monobasic sodium phosphate, 0.3 mmol L−1 of the ion-pairing agent 1-octane sulfonic acid, 4% (v/v) acetonitrile, pH 2.7, adjusted with 85% phosphoric acid. The flow rate was 1 mL min−1. Under these conditions, the separation of aminothiol was completed in 20 min. Deproteinated samples were injected directly onto the column using a Shimadzu autosampler (Shimadzu Corp.). Following HPLC separation, GSH was detected with a model 2465 electrochemical detector (Waters, MA, USA) equipped with a 2 mm Glassy carbon (GC) analytical cell and potential of +700 mV were applied.
Cells were grown in Chamberlain medium to mid-exponential phase and then harvested by centrifugation and washed twice with Chamberlain without amino acid. The cells were suspended at a final OD600 of 0.5 in the same medium containing 50 mg/ml of chloramphenicol. After 15 min of pre-incubation at 25°C, uptake was started by the addition of L-[U-14C] glutamic acid (Perkin Elmer), at various concentrations (14C-Glu ranging from 1 to 50 µM). The radiolabeled 14C-Glu was at a specific activity of 9.25 GBq.mmol −1. Samples (100 µL of bacterial suspension) were removed after 5 min and collected by vacuum filtration on membrane filters (Millipore type HA, 25 mm, 0.22 mm) and rapidly washed with Chamberlain without amino acid (2×5 mL). The filters were transferred to scintillation vials and counted in a Hidex 300 scintillation counter. The counts per minute (c.p.m.) were converted to picomoles of amino acid taken up per sample, using a standard derived by counting a known quantity of the same isotope under similar conditions.
Succinate, fumarate, citrate and oxoglutarate were quantified by gas chromatography coupled with mass spectrometry (GC/MS). Wild-type F. novicida and ΔgadC strains were grown as for glutamate quantification. Briefly, after an overnight preculture in CDM, three independent cultures of wild-type and ΔgadC mutant were cultivated in 50 mL CDM to an OD600 of app. 0.35. Bacteria were harvested by centrifugation at 4,000× g for 15 min, resuspended to the same OD600 in pre-warmed CDM supplemented with glutamate (1.5 mM) and cultivated for 30 min±500 µM H2O2.
Metabolite measurements were normalized by checking that each sample contained equal amounts of total proteins.
Wild-type F. novicida and ΔgadC mutant strains were grown in TSB to exponential growth phase and diluted to the appropriate concentrations. 6 to 8-week-old female BALB/c mice (Janvier, Le Genest St Isle, France) were intra-peritoneally (i.p.) inoculated with 200 µl of bacterial suspension. The actual number of viable bacteria in the inoculum was determined by plating appropriate dilutions of bacterial cultures on Chocolate Polyvitex plates. For competitive infections, wild-type F. novicida and mutant bacteria were mixed in 1∶1 ratio and a total of 100 bacteria were used for infection of each of five mice. After two days, mice were sacrificed. Homogenized spleen and liver tissue from the five mice in one experiment were mixed, diluted and spread on to chocolate agar plates. Kanamycin selection to distinguish wild-type and mutant bacteria were performed. Competitive index (CI) [(mutant output/WT output)/(mutant input/WT input)]. Statistical analysis for CI experiments was as described in [51]. Macrophage experiments were analyzed by using the Student's t-test.
|
10.1371/journal.pgen.1005828 | Rab6 Is Required for Multiple Apical Transport Pathways but Not the Basolateral Transport Pathway in Drosophila Photoreceptors | Polarized membrane trafficking is essential for the construction and maintenance of multiple plasma membrane domains of cells. Highly polarized Drosophila photoreceptors are an excellent model for studying polarized transport. A single cross-section of Drosophila retina contains many photoreceptors with 3 clearly differentiated plasma membrane domains: a rhabdomere, stalk, and basolateral membrane. Genome-wide high-throughput ethyl methanesulfonate screening followed by precise immunohistochemical analysis identified a mutant with a rare phenotype characterized by a loss of 2 apical transport pathways with normal basolateral transport. Rapid gene identification using whole-genome resequencing and single nucleotide polymorphism mapping identified a nonsense mutation of Rab6 responsible for the apical-specific transport deficiency. Detailed analysis of the trafficking of a major rhabdomere protein Rh1 using blue light-induced chromophore supply identified Rab6 as essential for Rh1 to exit the Golgi units. Rab6 is mostly distributed from the trans-Golgi network to a Golgi-associated Rab11-positive compartment that likely recycles endosomes or transport vesicles going to recycling endosomes. Furthermore, the Rab6 effector, Rich, is required for Rab6 recruitment in the trans-Golgi network. Moreover, a Rich null mutation phenocopies the Rab6 null mutant, indicating that Rich functions as a guanine nucleotide exchange factor for Rab6. The results collectively indicate that Rab6 and Rich are essential for the trans-Golgi network–recycling endosome transport of cargoes destined for 2 apical domains. However, basolateral cargos are sorted and exported from the trans-Golgi network in a Rab6-independent manner.
| Cells in animal bodies have multiple plasma membrane domains; this polarized characteristic of cells is essential for their specific functions. Selective membrane transport pathways play key roles in the construction and maintenance of polarized structures. Drosophila photoreceptors with multiple plasma membrane domains are an excellent model of polarized transport. We performed genetic screening and identified a Rab6 null mutant with a rare phenotype characterized by a loss of 2 apical transport pathways with normal basolateral transport. Although Rab6 functions in the Golgi are well known, its function in polarized transport was unexpected. Here, we found that Rab6 and its effector, Rich, are required for multiple apical transport pathways but not the basolateral transport pathway. Our findings strongly indicate that the membrane proteins delivered to multiple polarized domains are not sorted simultaneously: basolateral cargos are segregated before the Rab6-dependent process, and cargos going to multiple apical domains are sorted after Rab6-dependent transport from the trans-Golgi network to the Golgi-associated Rab11-positive compartment, which presumably recycles endosomes. Our finding of the function of Rab6 in polarized transport will elucidate the molecular mechanisms of polarized transport.
| Many cells that make up animal bodies have multiple plasma membrane domains, which are the basis of their specific functions. Polarized vesicle transport is essential for establishing and maintaining these domains. However, the underlying mechanisms are not well elucidated. The Drosophila retina is a good genetics-based model system for studying polarized transport. A single retina contains approximately 800 ommatidia, which comprise 6 outer and 2 central photoreceptors with 4 distinct plasma membrane domains (Fig 1A and 1B, see also [1]). The first domain is the photoreceptive membrane domain, i.e., the rhabdomere, which is formed at the center of the apical plasma membrane as a column of closely packed rhodopsin-rich photosensitive microvilli. The second is the peripheral apical domain surrounding the rhabdomere, i.e., the stalk membrane, which is enriched in Crb and beta H-spectrin [2, 3]. “Eye shut” (Eys) is an extracellular matrix protein secreted from the stalk domain of photoreceptors from the early to mid-pupal stages [4, 5]; its secretion is necessary and sufficient to form the inter-rhabdomeric space (IRS). The third domain is the basolateral membrane, which is separated from the apical membrane by adherens junctions, where Na+K+ATPase localizes, similar to typical polarized epithelial cells [6]. Finally, the fourth domain comprises the axon and synapses, which extend below the retina to the brain.
Rh1, the rhodopsin expressed in R1–6 outer retinal photoreceptor cells, is the most accessible protein for investigating apical polarized transport in Drosophila photoreceptors. Blue light-induced chromophore supply (BLICS) releases Rh1 from the endoplasmic reticulum (ER) [7], allowing us to trace synchronous Rh1 trafficking. We previously used this method to show that Rab1 is involved in ER–Golgi transport [8], that GPI synthesis is essential for Rh1 sorting in the trans-Golgi network (TGN) [1], and that the Rab11/dRip11/MyoV complex is essential for post-Golgi Rh1 transport [9, 10]. In addition, other studies demonstrate that the exocyst complex tethers post-Golgi vesicles to the base of the rhabdomeres [11] and that Rab6Q71L overexpression inhibits Rh1 transport in the early secretory pathway [12].
Rab6 is a member of the Rab family of small GTPases, which regulate the specificity between donor and acceptor membranes for vesicle budding, docking, tethering, and fusion steps during transport [13–15]. Human Rab6 comprises 4 different isoforms—Rab6a, Rab6a′, Rab6b, and Rab6c—and is the most abundant Golgi-associated Rab protein [16]. Rab6 regulates retrograde transport from the Golgi complex to the ER [17, 18]; the retrograde transport of the B subunit of Shiga toxin from early endosomes or recycling endosomes (REs) to the Golgi complex [19, 20]; and the post-Golgi trafficking of herpes simplex virus 1 (HSV1) envelop proteins, tumor necrosis factor (TNF), and vesicular stomatitis virus G-protein (VSV-G) to the plasma membrane [21–23].
In the present study, we screened ethyl methanesulfonate (EMS)-mutagenized flies to find mutants with defective polarized transport. As one of the mutants identified was a null allele of Rab6, we reevaluated the function of Rab6 in Drosophila photoreceptors and found Rab6 is essential for transport towards both distinct apical domains (i.e., the stalk and rhabdomeres) but not towards the basolateral membrane. These results indicate basolateral cargos are sorted and exported from the TGN in a Rab6-independent manner before they are sorted into pathways destined for the stalk and rhabdomeres.
To identify the genes essential for rhabdomere morphogenesis and membrane trafficking, we previously performed retinal mosaic screening of P-element–inserted lines maintained in stock centers [24] using the FLP/FRT method [25] with in vivo fluorescent imaging [26] of Arrestin2::GFP [1]. To extend the screening to comprehensive coverage, we are currently engaged in genome-wide, large-scale screening of EMS-mutagenized chromosomes; thus far, we have isolated 233 Rh1-accumulation deficient mutant lines, which stochastically covers more than 80% of genes in more than 60% of Drosophila genome [27].
Flies including Drosophila melanogaster have an open rhabdom system, in which the rhabdomeres of each ommatidium are separated from each other and function as independent light guides. In contrast, bees and mosquitoes have a closed system, in which rhabdomeres within each ommatidium are fused to each other, thus sharing the same visual axis. Without expression or secretion to the IRS of the extracellular matrix glycoprotein Eys, fly ommatidia exhibit a closed rhabdom [4, 5]. By screening EMS-mutagenized left arms of the second chromosome, we isolated 11 mutant lines with a closed rhabdom phenotype similar to the eys mutant (Fig 1C and 1D); these lines were further classified into 4 complementation groups. The mutations in one group failed to complement a loss-of-function allele of the eys gene, eysBG02208 [4], indicating that mutations in this group represent new loss-of-function alleles of eys. On the other hand, the mutations in the other 3 groups complemented the eysBG02208 mutant. Fig 1C and 1D show Arrestin2-GFP localizations in 510U mutant in the eys complementation group as well as 546P in another complementation group; 546P was the sole member of the complementation group, and the homozygotes were lethal in the early larval stage. The closed rhabdom phenotypes of 510U and 546P had 100% penetrance: all adult flies observed for 510U or 546P using Arr2GFP had mutant ommatidia with a fused rhabdomere. We subsequently observed similar high penetrance of the phenotypes in immunostaining at the fly and cell levels unless otherwise noted.
Immunostaining of the mosaic retina of these mutants with anti-Eys antibody revealed complete loss of Eys protein in the whole mutant ommatidia of 510U (Fig 1E), which again strongly suggests the complementation group (which contains 510U) represents eys. On the other hand, in 546P mutant ommatidia, Eys was readily detected; however, it did not accumulate in the extracellular domain (i.e., the IRS) but rather within undefined cytoplasmic structures (Fig 1F). These results suggest the transport of Eys protein is inhibited in the 546P mutant.
We subsequently investigated the localization of several plasma membrane proteins in 546P mutant photoreceptors (Fig 2). In wild type photoreceptor cells, Rh1 localized at the photoreceptive rhabdomeres. However, Rh1 was mainly detected in the undefined cytoplasmic structures in 546P mutant photoreceptors (Fig 2A, blue). In contrast to Eys and Rh1, a basolateral membrane protein, Na+K+ATPase localized normally in the basolateral membrane and did not accumulate in the cytoplasm of 546P mutant photoreceptors (Fig 2A, green). Eys and Rh1 were co-localized in the same undefined cytoplasmic structures (Fig 2B). Two other rhabdomeric proteins—TRP and Chaoptin (Chp)—and a stalk protein—crumbs (Crb)—were also detected in the undefined cytoplasmic structures (Fig 2C–2E). On the other hand, DE-Cadherin (DE-Cad) was localized normally at the adherens junctions (Fig 2E). These results indicate inhibition of transport toward the rhabdomeres and stalks but not toward the basolateral membrane. Coinciding with the localizations of plasma membrane proteins, the basolateral membrane in 546P mutant photoreceptors was extended as in the wild type despite the shrunken stalk membrane and small rhabdomeres (Figs 2A, 2D and 2E and S5). None of the 222 other Rh1 accumulation-deficient lines clearly showed 546P-like phenotypes, closed rhabdomeres, or inhibition of transport toward the rhabdomere and stalk (but not toward the basolateral membrane), indicating these 546P phenotypes are characterized by an apical-specific transport deficiency rather than the result of a delay or loss of general membrane transport.
To clarify the details of the transport defects in 546P mutant photoreceptors, we investigated the dynamics of Rh1 transport using BLICS [1, 8]. Rh1 comprises an apoprotein, opsin, and the chromophore 11-cis retinal. Without the chromophore, opsin accumulates in the ER. Blue light illumination photoisomerizes all-trans retinal to 11-cis, inducing the synchronous release of Rh1 from the ER into the secretory pathway. Prior to BLICS, Rh1-apoprotein accumulated normally in the ER and Rh1 was transported to the Golgi units by 40 min after BLICS in 546P mutant photoreceptors (Fig 3A and 3B). However, in contrast to wild type photoreceptors, Rh1 did not reach the rhabdomeres in 546P mutant photoreceptors but instead localized in large globular structures in the cytoplasm with the late endosome marker Rab7 180 min after BLICS (Fig 3C and 3H). These results indicate Rh1 is normally synthesized in the ER and transported to the Golgi units in 546P mutant photoreceptors but is subsequently transported to the late endosomes rather than the rhabdomeres.
To determine how Rh1 is exported from the Golgi units to the late endosomes, we compared Rh1 staining in the Golgi units 30 and 60 min after BLICS in both wild type and 546P mutant photoreceptors (Fig 3D and 3E). There were no apparent differences between wild type and 546P mutant photoreceptors 30 min after BLICS. However, 60 min after BLICS, Rh1 staining became stronger in the Golgi units of mutant photoreceptors than wild type photoreceptors. Moreover, the Golgi units extended Rh1-positive tubular or globular structures (Fig 3E). Next, we triple stained 546P mutant photoreceptors with antibodies against the medial-Golgi marker p120 [28], late endosomal marker Rab7 [29], and Rh1 40 and 90 min after BLICS and investigated the Rh1-positive tubular or globular structures extending from the Golgi units (Figs 3F and S1). Golgi markers p120, Rab7, and Rh1 were frequently partially co-localized. Rab7 staining was often adjacent to p120 staining, and Rh1 staining overlapped with both of them (Fig 3F, arrows; S1 Fig). A Z-projection (Fig 3G) of 20 optical sections at 0.5-μm intervals (S1 Fig) showed most Golgi units faced the globular Rab7-positive structure, both of which were positive for Rh1; the Golgi units with Rh1 (turquoise) were often adjacent/connected to late endosomes containing Rh1 (yellow). These observations suggest Rh1 is first transported to the Golgi units in a normal manner but is not exported normally and subsequently accumulates there; finally, Rh1 directly exits the Golgi units to Rab7-positive tubular or globular structures.
We previously showed that the Rab11/dRip11/MyoV complex is essential for post-Golgi vesicle transport and that deficiency of any component of the complex induces cytoplasmic accumulation of Rh1-loaded post-Golgi vesicles [9, 10]. The pattern of Rh1-accumulation observed in the 546P mutant clearly differed from those of Rab11, dRip11, or MyoV mutants.
To investigate the epistatic interaction between mutations of the Rab11/dRip11/MyoV complex and the 546P mutation for Rh1 transport, we observed Rh1 localization in 546P mutant mosaic retina expressing the dominant-negative MyoV C-terminal domain. Rh1 did not accumulate in the cytoplasm of 546P/MyoV double-mutant photoreceptors like the MyoV single mutant but instead accumulated in the globular structures (Fig 3H). This phenotype was indistinguishable from that of cells with only the 546P mutation (Fig 2A–2C). This result indicates the 546P mutation is epistatic to mutations of the Rab11/dRip11/MyoV complex. Thus, in the 546P mutant, the Rh1 in Rab7-positive late endosomes directly comes from the Golgi units before being sorted into post-Golgi vesicles.
Kinetic and epistatic analyses of Rh1 transport in 546P mutant cells revealed that the processes between Golgi entry and before/upon post-Golgi vesicle formation are inhibited in the 546P mutant during Rh1 biosynthetic trafficking.
As the phenotype of 546P is drastic and important for understanding the mechanism of polarized vesicle transport, we identified the mutation responsible for the phenotype. Rough mapping using meiotic recombination and restriction fragment length polymorphisms (RFLP) [30] placed the 546P mutation between RFLP893 and RFLP977. We subsequently sequenced the whole genome of 546P homozygous animals using whole-genome amplification and a next-generation sequencer [27] and found 4 unique mutations in the mapped area; complementation tests over defined chromosomal deletions showed that only one of these, a nonsense mutation on the Rab6 gene (Rab6 R73 to a stop codon), is lethal like 546P homozygous larvae.
As a Rab6 null mutation was found in the mapped area, we subsequently determined if YFP::Rab6wt expression rescues the 546P phenotype. Using the UAS/Gal4 system, we expressed YFP::Rab6wt from early/late pupa in whole bodies using heat shock-Gal4 or in late pupal outer photoreceptors using Rh1-Gal4 (Fig 4). YFP::Rab6wt expression from 28% pupal development by heat shock-Gal4 completely rescued the 546P phenotype (Fig 4F and 4G), confirming that the Rab6 gene is responsible for the 546P phenotype. Therefore, we designated 546P as a new allele of the Rab6 gene, Rab6546P. Interestingly, YFP::Rab6wt expression starting from 68% pupal development by heat shock-Gal4 or by Rh1-Gal4 showed partial rescue; Rh1 accumulated in fused rhabdomeres lacking normal level of Eys in the IRS (Fig 4D, 4E, 4I and 4J). These results are concordant with previous reports on Eys/spacemaker [5] showing that Eys expression in early pupae is required to form open rhabdomeres.
We subsequently investigated whether photoreceptors homozygous for the other Rab6 null allele Rab6D23D have defects in Rh1 and Eys trafficking. Like Rab6546P mutant photoreceptors, Rab6D23D ommatidia exhibited a closed rhabdomere phenotype visualized by phalloidin staining (S2A Fig). Eys accumulated in the globular cytoplasmic structures, although a limited amount localized between photoreceptor apical membranes (S2A Fig). In Rab6D23D photoreceptors, most Rh1 was localized in the cytoplasm, whereas the basolateral membrane protein Na+K+ATPase localized normally in the basolateral membrane (S2B Fig). These results corroborate the notion that the 546P phenotype is due to Rab6 protein deficiency.
In yeast and mammalian cells, the Ric1/Rgp1 complex functions as a guanine nucleotide exchange factor (GEF) for Rab6 [31, 32]. The Drosophila Ric1p homolog Rich is a Rab6 effector in Drosophila [33]. Rich contains RIC1 and WD40 domains, both of which bind to fly Rab6. Although neither the GEF activity for Rab6 nor binding to the Drosophila rgp1 homolog CG1116 has been confirmed, Rich might function as a GEF together with other interacting proteins [33]. Rab6 and Rich interact genetically and are required for synaptic specificity in fly eyes and olfactory receptor neurons [33]. Therefore, we investigated the localizations of Rh1 and Eys in photoreceptors homozygous for a null allele in Rich, Rich1. Similar to Rab6546P and Rab6D23D photoreceptors, Rh1 and Eys accumulation in the rhabdomeres and IRS, respectively, were limited, and both proteins were detected in cytoplasmic structures (S2C and S2D Fig). On the other hand, the Na+K+ATPase localized normally in the basolateral membrane in Rich1 mutant photoreceptors (S2D Fig). The phenotypes were highly penetrant in Rab6D23D photoreceptors. However, in the case of Rich1 homozygous photoreceptors, some Rich1 mutant cells only exhibited a weak phenotype.
Detailed observation of Rh1 transport by BLICS in Rab6D23D and Rich1 mutants showed that Rh1 is synthesized and transported to the Golgi units in a normal manner; however, after leaving the Golgi units, most Rh1 enters the late endosomes instead of the rhabdomeres in the same manner as in Rab6546P photoreceptors (S3 Fig). The localizations of other membrane proteins were investigated in Rab6D23D and Rich1 mutants (S4 Fig). The rhabdomeric proteins Chp and TRP accumulated in the cytoplasm in both Rab6D23D and Rich1 mutants. The localization of the stalk protein Crb in the stalk membrane was limited in both Rab6D23D and Rich1 mutants, but its cytoplasmic accumulation was observed only in Rab6D23D mutants. More severe membrane trafficking defects were observed in Rab6D23D mutants than Rich1 mutants. On the other hand, DE-Cad localized normally in the basolateral membrane in both Rab6D23D and Rich1 mutants. These observations coincide with the Rab6546P phenotype. Quantification of the lengths of the stalk and basolateral membrane in tangential sections revealed that Rab6546P and Rab6D23D mutant photoreceptors had shorter stalk membranes than those of wild type but basolateral membranes of normal length (S5A–S5F Fig).
Furthermore, we determined if Rab6 deficiency affects the transport of another basolateral protein, FasIII, using ovarian follicle cells (S6 Fig). FasIII localized normally on the basolateral membrane, whereas the localization of the apical membrane protein, Notch, was diminished in follicle cells. Thus, the involvement of Rab6 in apical transport but not basolateral transport could be common in polarized cells.
To investigate the details of organelle structures and plasma membrane domains, we observed thin sections of late pupal wild type, Rab6546P, Rab6D23D, and Rich1 whole mutant ommatidia by electron microscopy (Fig 5A–5D and S7 Fig). A tangential section of a wild type ommatidium contains 7 round rhabdomeres separated by the IRS. However, Rab6546P, Rich1, and Rab6D23D ommatidia contained miniature rhabdomeres attached to each other, and a narrow IRS (Fig 5, purple) did not separate adjacent rhabdomere, similar to Eys395 and Eys1 mutants [4, 5]. The stalk membranes in Rab6546P, Rab6D23D and Rich1 photoreceptors were shorter than those of the wild type, whereas the basolateral membranes were normal (S5G–S5I Fig). Similar to the confocal microscopic observations, these phenotypes were less prominent in Rich1 ommatidia than Rab6546P and Rab6D23D ommatidia. Overall, these plasma membrane morphological phenotypes are consistent with the immunohistochemically detected transport phenotypes of Rab6 and Rich mutants.
The appearance of the Golgi units in Rab6546P, Rab6D23D, and Rich1 photoreceptors also differed from those in the wild type: both Rab6- and Rich-deficient Golgi units were larger and well developed, and their cisternae were dilated (Fig 5E–5H), similar to previous reports [21–23]. However, in contrast to the report of Storrie et al., (21) the vesicle or budding profiles of Rab6 and Rich-deficient Golgi units were not obviously increased. In addition to the differences in these plasma membrane domains, the IRS, and Golgi units, the number and size of multi-vesicular bodies (MVBs) were greater in Rab6546P, Rab6D23D, and Rich1 photoreceptors than the wild type (Fig 5A–5D, blue). We quantified the MVBs in dark-adapted homozygous wild type, Rab6546P, and Rich1 photoreceptors, because MVB formation is strongly triggered by light-dependent Rh1 endocytosis even in wild type photoreceptors (Fig 5I). MVBs 300–600 and ≥600 nm in diameter were categorized as small and large, respectively. Adult wild type photoreceptors contained (mean ± SD) 0.15 ± 0.06 small MVBs and zero large MVBs. On the other hand, single Rab6546P photoreceptors contained 0.33 ± 0.11 and 0.27 ± 0.09 small and large MVBs, respectively. Similarly, Rich1 photoreceptors contained 0.35 ± 0.09 and 0.27 ± 0.06 small and large MVBs, respectively. In addition to the increased number and size of MVBs, MVBs in Rab6- and Rich-deficient photoreceptors appeared different from those in the wild type photoreceptors; Rab6546P, Rab6D23D and Rich1 photoreceptors contained many MVBs densely packed with inner vesicles that appeared partially degraded and were often accompanied by cisternae, some of which were Golgi units (Fig 5E–5H). The consistent sizes and populations of MVBs suggest that MVBs are likely identical to the Rab7-positive undefined cytoplasmic structures that accumulate Rh1, Eys, TRP and Chp in the 546P mutant (Figs 2 and 3).
Further understanding the function of Rab6 requires knowing the localization of Rab6 protein in fly photoreceptors. In our first attempt, we used UAS-YFP::Rab6wt transgenic flies [34]. YFP::Rab6wt expressed by Rh1-Gal4 forms cytoplasmic foci in R1–6 photoreceptors. Most of the YFP::Rab6wt foci were associated with the cis-Golgi marker GM130 (Fig 6A, arrows), indicating these foci are located in Golgi units. GM130-negative YFP::Rab6wt foci at the base of the rhabdomeres were presumably post-Golgi vesicles bearing Rh1 (Fig 6A, arrowheads). Dispersed YFP signals were also observed in the cytoplasm and the rhabdomeres. The cytoplasmic YFP signal likely reflects GDP-bound form of YFP::Rab6wt associating with GDI. However, the reason for the YFP signal in the rhabdomeres is unclear; Rab6 might associate with some rhabdomeric proteins.
To confirm the localization of endogenous Rab6, we generated antisera against histidine-tagged full-length Rab6 protein in 1 rabbit and 2 guinea pigs (Rb anti-Rab6, GP1 anti-Rab6 and GP2 anti-Rab6). Western blotting showed that all antisera recognized a band around 22 kD in wild type head extract as well as a band around 60 kD in genomic Rab6EYFP heterozygous head extracts (S8A Fig). Immunohistochemistry showed cytoplasmic foci only in wild type photoreceptors in Rab6D23D mosaic retinas (Fig 6B, arrows and S8B Fig). Hence, these antisera specifically recognize Rab6. Rab6-positive cytoplasmic foci are also GM130-positive, suggesting these foci are Golgi units. Similar to YFP::Rab6wt, there were Rab6 foci at the bases of the rhabdomeres, which were GM130-negative, suggesting Rab6 also localizes on post-Golgi vesicles (Fig 6B, arrowheads). In Rich1 homozygous photoreceptors (Fig 6C, arrows), cytoplasmic Rab6-positive foci were not observed, suggesting that Rich is required for Rab6 recruitment to the Golgi membrane.
Photoreceptors at earlier stages (i.e., ~20–30% pupal development) contain well-developed Golgi units, often reaching >1 μm in width [10]. To investigate the Golgi–cisternal association of Rab6, we used these large Golgi units in young pupal photoreceptors expressing a CFP-GalT marker, i.e., a cyan fluorescent protein fused to the 81 N-terminal amino acids of human β-1,4-galactosyltransferase [10], which localizes at the trans-cisternae of the Golgi units and the TGN [35, 36]. Double staining of CFP-GalT–expressing retina with anti-Rab6 and anti-GM130 showed that Rab6 localized to the opposite side of CFP-GalT–positive compartment from the cis-Golgi marker GM130 (Fig 7A). Detailed observations indicate the Rab6-positive region includes the TGN and extends more distally (Fig 7C).
Most of the Golgi units were accompanied by Rab11-positive puncta adjacent to the trans side (S9 Fig). Unlike the late pupal stages, when Rh1 transport is active and many Rab11-positive puncta localize at the base of the rhabdomeres, there are only a few Rab11-positive puncta apart from Golgi units in the young pupal stages. As Rab11 is typically considered as RE marker and is close to the TGN, these results suggest that these Rab11-positive puncta are likely fly photoreceptor REs emerging from the TGN or transport vesicles going to REs. Double staining of a CFP-GalT–expressing retina with anti-Rab6 and anti-Rab11 showed that Rab6 spreads from the CFP-GalT–positive TGN to the Golgi-associated Rab11-positive recycling endosomes (Figs 7B and S9 [wide view]). These results suggest that Rab6 regulates transport between the TGN and RE.
The Clathrin heavy chain (Chc) is an essential coat protein of some post-Golgi vesicles, especially for the basolateral pathway in epithelial cells [37]. Therefore, we compared Rab6 localization with Chc by triple staining wild type retinas with anti-Rab6, anti-Rab11, and anti-Chc (Fig 7C) [38]. Interestingly, Rab6-signals overlapped with both Rab11 and Chc, whereas Chc and Rab11 are largely separate (Fig 7C, upper panel). These results suggest that Rab6 functions in a step close to but before Rab11-dependent post-Golgi trafficking.
In the present study, screening for the failure of Rh1 accumulation in Drosophila photoreceptor rhabdomeres led to the re-identification of Rab6 as an essential molecule for Rh1 transport. Our detailed observations of Rh1 trafficking by BLICS indicate the inhibition of transport in Rab6-null mutants occurs within the Golgi units: Rh1 is transported to the cis-Golgi with normal kinetics but does not exit the Golgi units into the plasma membrane. Consequently, Rh1 greatly accumulates on the Golgi membrane and finally exits directly into the endosomal compartment. Furthermore, in Rab6 mutant photoreceptors, other rhabdomere membrane proteins (TRP and Chp), a stalk membrane protein (Crb), and the apically secreted protein (Eys) are reduced and degraded in MVBs together with Rh1. In contrast, the basolateral membrane proteins Na+K+ATPase and DE-Cad are unaffected. Moreover, Rab6 proteins are distributed from the TGN to Golgi-associated Rab11-positive structures.
Many aspects of Rab6 functions have been reported, such as retrograde transport from endosomes to the TGN, intra-Golgi or Golgi to ER trafficking, anterograde transport from the TGN, Golgi homeostasis, and Golgi-ribbon organization [39]. Nevertheless, there is no concrete understanding of all Rab6 functions. Among them, the defective morphology of the Golgi units and the failure of Rh1 transport observed in the present study are consistent with the results of Rab6 knockdown in HeLa cells by Storrie et al. [21], who report amplification and dilation of Golgi stacks, accumulation of unreleased vesicles on the trans-Golgi, inhibition of Golgi–plasma membrane transport of VSV-G without delayed ER–Golgi transport, and increased MVBs around the Golgi. The present results are also concordant with a study by Januschke et al., who report that mutants of Rab6b and a Rab6 GEF BicD in Drosophila oocytes induce accumulation of the TGF-α homolog Gurken in “ring-like particles” that have the properties of endolysosomes [40]. Two recent reports on HSV1 production and TNF secretion indicate Rab6 depletion inhibits the post-Golgi transport of HSV1 envelope proteins and TNF, respectively [22, 23]; these proteins accumulate at a juxtanuclear Golgi-like location and Golgi membranes in Rab6-depleted cells, respectively. These results are concordant with Rh1 localization after BLICS in our studies. Thus, together with the Rab6 localization at the trans-Golgi, the present results indicate that Rab6 is crucial for apical cargo export from the TGN.
The constitutively active GTP-locked form of Rab6 overexpressed by Rh1-promoter drastically reduces Rh1 content and retains Rh1 in the ER form in Drosophila photoreceptors [12]. Although the observed phenotype appears specific to rhodopsin and deficiency of ER–Golgi transport rather than post-Golgi transport, the phenotype is different from that in the present study likely because of differences in the methods used: Eys, TRP, and Crb expressions peak during the mid-pupal stage, whereas Rh1-promoter begins to be expressed at the late pupal stage. Gain or loss of Rab6 function causes defects in protein transport via the Golgi; however, they have opposite effects on the structure of Golgi units. Constitutively active Rab6 forces Golgi proteins to relocate to the ER [41] in contrast to the accumulation of the Golgi cisternae in Rab6-knockdown conditions [21]. Therefore, Rh1 may remain in the ER in the Rab6 hypermorph but stall at the TGN in the Rab6 amorph.
The results of the present study show that the Rab6-binding protein Rich is involved in rhabdomere and stalk membrane transport. Although its biochemical activity as a Rab6 GEF is unconfirmed [33], the loss of Rab6 localization to the Golgi in Rich1 mutant indicates Rich is required to recruit Rab6 to the Golgi membrane and likely works as a Rab6GEF like yeast and mammalian Ric1 [31, 32]. Rich was originally identified as a regulator of DN-Cad trafficking to synapses; Rab6 is also involved in these processes [33]. The present and previous findings collectively suggest Rab6/Rich-dependent export from the TGN is required for the trafficking of some axonal proteins as well as rhabdomeres and stalk proteins.
The most important finding of the present study is that in Rab6 mutants, the transport of protein cargos destined for the 2 apical membranes (i.e., the rhabdomere and stalk) is arrested and they accumulate together in the MVBs, while the basolateral cargos are transported normally. Hence, Golgi-localized Rab6 is not required for all plasma membrane-directed transport pathways. These results have implications for previous work on the Drosophila germline cyst [42] suggesting the existence of Rab6-dependent and Rab6-independent exocytic pathways. Although the roles of these multiple exocytic pathways in the polarized membrane domains are not mentioned, the Rab6-independent pathway to the plasma membrane in the germline cyst might correspond to the basolateral transport in photoreceptors. In addition, in embryonic salivary glands, Rab6 and Rab11 are localized on cytoplasmic vesicles containing an overexpressed apical membrane protein Cadherin 99C (Cad99C) [43]. Moreover, there is phenotypic similarity between Rab6 null and sec5 null. These findings indicate Rab6 and Rab11 sequentially regulate the apical transport of Cad99C, similar to Rh1 transport. However, in this case, as basolaterally transported truncated Cad99D also associates with Rab6, Rab6 and Cad99C colocalization merely reflects the association of Cad99C with Golgi units rather than apical transport vesicles.
Clathrin and AP1B are involved in basolateral transport [44, 45] and are required to sort various apical proteins in both mice and C. elegans [46–49]. In fly photoreceptors, the sole fly AP1 is essential for Na+K+ATPase transport towards the basolateral membrane, and the loss of the AP1 subunit, gamma, or mu causes mistransport of Na+K+ATPase to the stalk membrane [1], similar to zebrafish hair cells lacking AP1β subunits [50]. Clathrin likely plays a role in Na+K+ATPase transport towards the basolateral membrane in fly photoreceptors. The partial colocalization of Chc and Rab6 as well as the separation of Chc and Rab11 at the distal TGN (Fig 7B and 7D) indicate that the basolateral transport pathway likely branches off the Rab6-dependent apical transport pathway at the TGN. We previously showed that Rh1 transport from the TGN to the rhabdomere is dependent on Rab11 [10]. In addition, Rab6 localization is not restricted to the TGN but extends further to the Rab11-positive puncta closely associated with the TGN. These results imply that Rab6 is involved in Rh1 transport from the TGN to the closely associated Rab11-positive puncta. As Rab11 is commonly considered as RE marker, these TGN-associated Rab11-positive puncta are likely REs emerging from the TGN or transport vesicles going to REs. These results collectively suggest Rab6 regulates the trafficking of apical cargos from the TGN to REs in fly photoreceptors.
The proposed model of polarized transport in Drosophila photoreceptors is illustrated in Fig 8. Membrane proteins are synthesized on the ER and transported together to the Golgi units. Within the Golgi units, the basolateral cargos are sorted into Chc/AP1 vesicles at the TGN, whereas apical cargos bound for the rhabdomere or stalk membrane are not segregated within the TGN but are rather transported together to REs. Rab6 is essential for the TGN–RE transport of apical cargos. The second sorting event at the RE segregates the cargos to the rhabdomere and stalk membrane, and carrier vesicles bearing the rhabdomere proteins are exported in a Rab11/dRip11/MyoV-dependent manner.
The proposed model postulates the second sorting occurs at REs; however, this might occur earlier, for example at the TGN or between the TGN and REs. In addition, the direction of Rab6-mediated transport (i.e., anterograde or retrograde) remains unresolved, because anterograde cargo transport can be due to the retrograde recycling transport of the TGN-resident proteins. The mechanisms of intra-Golgi transport and molecular role of Rab6 remain controversial despite substantial research. Three models of transport between the TGN and REs are based on models of intra-Golgi transport. First, the vesicle transport model posits that Rab6 is involved in the anterograde transport of tubules/vesicles bearing apical cargos from the TGN to REs. Second, in the cisternal maturation model, the TGN gradually maturates into REs as TGN-resident proteins are retrieved by Rab6-dependent retrograde tubule/vesicle-mediated transport [51, 52]. Third, in the cisternal progenitor model, in which Rab GTPases including Rab6 define the identity of membrane subcompartments, continual fusion fission and the “Rab cascade” achieve cargo transport from the TGN to REs [53, 54]. Future studies should elucidate the molecular mechanisms by which Rab6 functions in the transport of apical cargos from the TGN to REs.
Flies were grown at 20–25°C on standard cornmeal–glucose–agar–yeast food unless indicated otherwise. Carotenoid-deprived food was prepared from 1% agarose, 10% dry-yeast, 10% sucrose, 0.02% cholesterol, 0.5% propionate, and 0.05% methyl 4-hydroxybenzoate.
EMS mutagenesis and F1 or F2 live-imaging screening were performed as described previously (27). Starter strain with the second chromosome carrying proximal neoFRT at 40A was isogenized from the Bloomington Stock 5615 (Bloomington, IN, USA), which is used in single nucleotide polymorphism (SNP) mapping [30].
The tester line w; Rh1Arr2GFP ey-FLP/TM6B were used for live imaging of mutant lines. Meanwhile, y w ey-FLP; P3RFP FRT40A/SM1 was used for immunostaining. Furthermore, y w 70FLP; Ubi-mRFP.nls FRT40A was used for the mosaic analysis of ovarian follicle cells.
To quantify MVBs by electron microscopy, mutant whole-eye clones were made for Rich1 using the EGUF/hid method [55]. For Rab6546P and Rab6D23D, mutant whole-eye clones were not obtained probably because of the cell lethality of the mutants. Therefore, to partially rescue the lethality at early retinal development, YFP::Rab6wt was expressed by ey-Gal4, and Rab6546P mutant whole-eye clones were obtained as described in the mutant analysis of Sec6 [11].
The following fly stocks were used: Rh1-Gal4, heat shock-Gal4, EGUF40A (Bloomington stock number 5250, Bloomington, IN, USA), UAS-YFP::Rab6wt (Bloomington stock number 23251, Bloomington, IN, USA), UAS-GFP::MyoVCT [9], Rab6D23D FRT40A/CyO, Rich1 FRT80/TM3Kr>GFP, Rich2 FRT80/TM3, Kr>GFP (from Dr. Bellen, Baylor College of Medicine), and YRab6 (from Dr. Brankatschk, Max Planck Institute).
Fluorescent proteins expressed in photoreceptors were imaged by the water-immersion technique as described previously [8]. Briefly, late pupae with GFP-positive RFP mosaic retina were attached to the slide glass using double-sided sticky tape, and the pupal cases around the heads were removed. The pupae were chilled on ice, embedded in 0.5% agarose, and observed using an FV1000 confocal microscope equipped with a LUMPlanFI water-immersion 40× objective (Olympus, Tokyo, Japan).
Fixation and staining methods were performed as described previously by Satoh and Ready [56]. Primary antisera were as follows: rabbit anti-Rh1 (1:1000) [10], chicken anti-Rh1 (1:1000) [1], rabbit anti-GM130 (1:300) (Abcam, Cambridge, UK), rabbit anti-NinaA (1:300) (a gift from Dr. Zuker, Columbia University), mouse monoclonal anti Na+K+-ATPase alpha subunit (1:500 ascites) (DSHB, IA, USA), rat monoclonal anti DE-Cad (1:20 supernatant) (DSHB, IA, USA), mouse monoclonal anti-Notch (17.9C6) (1:10 supernatant) (DSHB, IA, USA), mouse monoclonal anti-FasIII (7G10) (1:10 supernatant) (DSHB, IA, USA), mouse monoclonal anti-Chp (24B10) (1:20 supernatant) (DSHB, IA, USA), rat anti-Crb (a gift from Dr. Tepass, University of Toronto), rabbit anti-TRP (a gift from Dr. Montell, Johns Hopkins University), rabbit anti-Rab7 (1:1000) (a gift from Dr. Nakamura, Kumamoto University, Kumamoto, Japan), mouse monoclonal anti-Eys (1:20 supernatant) (DSHB, IA, USA), chicken anti-GFP (1:1000) (Chemicon International Inc., Billerica, MA, USA), rabbit anti-Chc (1:500) (a gift from Dr. Kametaka, Nagoya University, Nagoya, Japan), mouse anti-Rab11 (1:250) [10], and rabbit anti-Rab11 (1:300) [10]. Secondary antibodies were anti-mouse, anti-rabbit, anti-rat and/or anti-chicken antibodies labelled with Alexa Fluor 488, 568, and 647 (1:300) (Life Technologies, Carlsbad, CA, USA) or Cy2 (1: 300) (GE Healthcare Life Sciences, Pittsburgh, PA, USA). Images of samples were recorded using an FV1000 confocal microscope (60× 1.42 NA objective lens; Olympus, Tokyo, Japan). To minimize bleed-through, each signal in double- or triple-stained samples was imaged sequentially. Images were processed in accordance with the Guidelines for Proper Digital Image Handling using ImageJ and/or Adobe Photoshop CS3 (Adobe, San Jose, CA, USA).
CFP was detected using the virtual channel function of the FV1000, which stacks CFP images taken by the other dichroic mirror on other channel images (i.e., Alexa Fluor 488, 568, and 647). However, the CFP laser beam axis was not perfectly aligned with the others probably because of the change in the dichroic mirror. Therefore, we scored this misalignment according to anti-GFP antibody staining of CFP with secondary antibodies conjugated with Alexa Fluor 488 and 647, and calibrated each of the 3 colored pictures obtained.
Newly eclosed flies fed carotenoid-deprived food were switched to carotenoid-deprived food with crystalline all-trans retinal (Sigma, St. Louis, MO, USA) in the dark. After 1 or 2 nights in the dark, the flies were irradiated with a 405-nm diode laser module at 30 mW for 20 min (Pepaless, Hyogo, Japan) to isomerize the all-trans retinal form to the 11-cis form and initiate Rh1 maturation.
Electron microscopy was performed as described previously [8]. Samples were observed on a JEM1400 electron microscope (JEOL, Tokyo, Japan), and montage images were taken by a CCD camera system (JEOL, Tokyo, Japan).
MVBs were counted in the tangential sections of 12–36 photoreceptor cells of individual flies, and the means and standard deviations were calculated from 4 flies for each allele.
Meiotic recombination mapping was performed according to the standard method [57]. Briefly, to allow meiotic recombination between the proximal FRT, the phenotype-responsible mutation, and a distal miniature w+ marker, flies carrying isogenized chromosome 546P were crossed with flies with isogenized P{EP0511}, which carry the miniature w+ marker near the distal end of arm 2L. Female offspring carrying the mutated chromosome and the miniature w+-marked chromosome were crossed with males carrying FRT40A and P3RFP on the isogenized second chromosome and Rh1Arr2GFP on the third chromosome. Live imaging was to judge whether the resultant adult offspring with w+ mosaic (i.e., maternally inherited FRT and w+) inherited the mutation responsible for the Rh1 transport defect. The recovered flies were individually digested in 50 μL 200 ng/μL proteinase K in 10mM Tris-Cl (pH 8.2), 1 mM EDTA, and 25 mM NaCl at 55°C for 1 h and subsequently heat inactivated at 85°C for 30 min and 95°C for 5 min. Then, 0.5 μL digested solution was used as the template for PCR amplification for RFLP analysis according to the method described in the FlySNP database ([58], http://flysnp.imp.ac.at/index.php). The mutation responsible for the Rh1 transport defect was mapped between SNP markers 893 and 977 defined in the FlySNP database.
For the whole-genome resequencing of the 546P mutant, the second chromosome was balanced over a balancer, CyO, P{Dfd-GMR-nvYFP}(Bloomington stock number 23230, Bloomington, IN, USA), to facilitate the isolation of homozygous embryo. Using the REPLI-G single-cell kit (QIAGEN, Hilden, Germany), the genomic DNAs were independently amplified from four 546P homozygous embryos. The pair-end library was prepared using the Nextera DNA sample prep kit (Illumina, San Diego, CA, USA) for each embryo, 2 × 250 bp reads were obtained using the MiSeq v2 kit (Illumina, San Diego, CA, USA). Reads were mapped to release 5 of the D. melanogaster genome using BWA 0.7.5a. The RFLP-mapped region of 546P was covered by reads with an average depth of 29.5x and width of 97.8%. Mapped reads were processed using Picard-tools 1.99 and the Genome Analysis Tool Kit 2.7–2 (GATK, Broad Institute, Cambridge, MA, USA). Single nucleotide variants and indels were called using Haplotypecaller in GATK, and those of the isogenized starter stock were subtracted to extract the unique variants in 546P and annotated using SnpSift [59] In the mapped region, there were 4 unique variants that could cause some defects in gene function (i.e., CG31705 P453L, Rab6 R73STOP, ACXE A193T, and nimB1 G255E). Complementation tests over defined chromosomal deletions covering the RFLP-mapped region (i.e., BSC213, BSC241, BSC242, BSC244, ED775, ED761, ED780, and ED791) revealed that only ED775 and ED761 were lethal to 546P homozygous larvae. The region deleted in both ED775 and ED761 contained the Rab6 R73STOP mutation in 546P, whereas this was not present in the other 3 variants described above.
The point mutation of Rab6 R73STOP on 2L:12108583 (Release 5) was verified by capillary sequencing of PCR-amplified fragment using the following primers: 5′-AGCGGAGCGAGAAGAGAGTT-3′ and 5′-GCTCCTGCTGTTGAAAAAGG-3′.
Full-length cDNA encoding Rab6 was amplified from a cDNA clone isolated from the fly retina cDNA library [60] and cloned into pQE60. The 6xHis-tagged Rab6 protein was expressed in E. coli pG-KJE8/BL21 (TAKARA) at 23°C and purified in native conditions using Ni-NTA Agarose (QIAGEN, Hilden, Germany). To obtain antisera, 1 rabbit and 2 guinea pigs were immunized 6 times with 200 μg 6xHis-Rab6 protein (Hashimoto, Wakayama, Japan). We designated the resultant antisera Rb anti-Rab6, GP1 anti-Rab6, and GP2 anti-Rab6.
Immunoblotting was performed as described previously [8]. The following antibodies were used: rabbit anti-Rab6 (1:2000 concentrated supernatant) (made in house), 2 guinea pig anti-Rab6 (1:2000 concentrated supernatant) (made in house) as primary antibodies; HRP-conjugated anti-rabbit or anti-guinea pig IgG antibodies (1:20,000, Life Technologies, Carlsbad, CA, USA) as a secondary antibody. Signals were visualized using enhanced chemiluminescence (Clarity Western ECL Substrate; Bio-Rad, Hercules, CA, USA) and imaged using ChemiDoc XRS+ (Bio-Rad, Hercules, CA, USA).
|
10.1371/journal.pcbi.1001011 | Reconstructing the Three-Dimensional GABAergic Microcircuit of the Striatum | A system's wiring constrains its dynamics, yet modelling of neural structures often overlooks the specific networks formed by their neurons. We developed an approach for constructing anatomically realistic networks and reconstructed the GABAergic microcircuit formed by the medium spiny neurons (MSNs) and fast-spiking interneurons (FSIs) of the adult rat striatum. We grew dendrite and axon models for these neurons and extracted probabilities for the presence of these neurites as a function of distance from the soma. From these, we found the probabilities of intersection between the neurites of two neurons given their inter-somatic distance, and used these to construct three-dimensional striatal networks. The MSN dendrite models predicted that half of all dendritic spines are within 100µm of the soma. The constructed networks predict distributions of gap junctions between FSI dendrites, synaptic contacts between MSNs, and synaptic inputs from FSIs to MSNs that are consistent with current estimates. The models predict that to achieve this, FSIs should be at most 1% of the striatal population. They also show that the striatum is sparsely connected: FSI-MSN and MSN-MSN contacts respectively form 7% and 1.7% of all possible connections. The models predict two striking network properties: the dominant GABAergic input to a MSN arises from neurons with somas at the edge of its dendritic field; and FSIs are inter-connected on two different spatial scales: locally by gap junctions and distally by synapses. We show that both properties influence striatal dynamics: the most potent inhibition of a MSN arises from a region of striatum at the edge of its dendritic field; and the combination of local gap junction and distal synaptic networks between FSIs sets a robust input-output regime for the MSN population. Our models thus intimately link striatal micro-anatomy to its dynamics, providing a biologically grounded platform for further study.
| The brain has an immensely complex wiring diagram, but few computational models of brain regions attempt accurate renditions of the wiring between neurons. Consequently, these models' dynamics may not accurately reflect those of the region. Key barriers here are the difficulty of reconstructing such networks and the paucity of critical data on neuron morphology. We demonstrate an approach that gets around these problems by using the available data to construct prototype neuron morphologies, and uses these to estimate how the probability of a connection between two neurons changes as we change the distance between them. With these in hand, we constructed artificial three-dimensional networks of the rat striatum and find that the connection distributions agree well with current estimates from anatomical studies. Our networks show features and dynamical implications of striatal wiring that would be difficult to intuit: the dominant input to the striatal projection neuron arises from other neurons just at the edge of its dendrites, and the main inhibitory interneurons are coupled locally by electrical connections and more distally by chemical synapses. Together, these properties set a unique state for the input-output computations of the striatum.
| The mammalian brain is a vastly complex structure at every level of description. Faced with the sheer breadth of neuron and receptor types, many researchers are abandoning attempts to intuit the ‘essential elements’ of a neural circuit, instead building large-scale models of neural circuits, modelling neuron-for-neuron [1]–[4]. This approach brings into sharp focus a further problem: how should we wire up the models? After all, the more accurate the underlying circuitry, the more confident we will be in linking dynamics of neural models to experimentally-recordable neural activity and, ultimately, to potential functions of the modelled structure. Typical modelling fall-backs of fully, regularly, or randomly connected networks are understandable choices when faced with this problem. Yet no neural circuit has these network topologies [5]–[8].
Establishing the detailed network of the striatum is a particular priority, given the large number of experimental and theoretical studies seeking to understand its computations [4], [9]–[16]. This large subcortical nucleus is the principal input structure of the basal ganglia, and is thought crucial for both motor control and learning [17], [18]. Profound deficits in both arise from diseases – such as Huntington's or Parkinson's – that directly affect the striatum or its primary afferents. Within the striatum lies a complex, predominantly GABAergic, microcircuit [19]. Medium spiny projection neurons (MSNs) are the only output neurons and comprise up to 97% of the cell population in rat, with GABAergic and cholinergic interneurons forming most of the remaining cell population. Despite their comparatively small number, the GABAergic fast-spiking interneurons (FSIs) in particular exert a very strong influence on the MSNs [20]–[22], receive input from similar sources, and are interconnected by both chemical synapses and gap junctions. However, the striatum's lack of layers and intermingling of neuron types has made it difficult to establish a detailed picture of its intrinsic network, hindering progress towards understanding the computations performed on its widespread cortical inputs [23].
One compelling reason for choosing to model at one-to-one scale is to explore a key question that can not be approached any other way: are there natural scales for the size of striatal regions involved in computing input-output functions? Much thought has been devoted to this question. The “domain” theory of striatum [9], [24], [25] began with the basic assumption that the natural computational element of the striatum was the network of MSNs within the radius of one MSN's dendritic tree – a sphere of approximately radius. Alexander and Crutcher [26] showed that microstimulation of primate sensorimotor striatum could elicit discrete movements of single joints, with each movement elicited from a small zone at most 1.2 mm in length. Graybiel and colleagues [27], [28] have argued that the pallidal-projecting regions of primate striatum are sub-divided into discernible cell clusters, each having a cross-sectional diameter of between and . More recently, Carillo-Reid et al [29] have shown that global excitation of an in vitro slice of striatum can induce the appearance of three or four cell assemblies – co-active groups of cells – within an region. All these lines of evidence point to different sizes and different reasons for defining a ‘computational element’ within striatum. Hence, by building at such scales we can look for the natural size of the computational element.
First though we had to build a model of the striatal network. Complete reconstructions of neural circuits are technically challenging, so quantitative data on the inputs and outputs of a single neuron are often incomplete or absent, while many published values are rough estimates. One way around this problem is to use reconstructions of stained dendrites and axons as guides [30]. Recent approaches test for appositions between cells by passing three-dimensional reconstructions of the morphology of several axonal and dendritic fields through each other [31]–[33], yielding statistics on the probability and location of synapses between two neurons. However, sets of complete, three-dimensional reconstructions of both axonal and dendritic morphologies are not available for most neural structures. Furthermore, building a network based on intersections of a sample of reconstructions may unknowingly limit the possible topologies.
To overcome these problems, we developed a stochastic approach based on the density of overlapping neurites, determining the densities from prototype dendrite and axon models. We applied this approach to reconstructing the three-dimensional GABAergic microcircuit of the adult rat striatum. Building prototype dendrite and axon models for MSNs and FSIs allowed us to determine any omissions or inconsistencies in existing quantitative data, and to establish constraints on the dendritic locations of afferent input. Using these models to reconstruct the three-dimensional network, we could address key questions about striatal micro-anatomy: how sparsely is the striatum connected? What are the comparative numbers of contacts for each type of connection? Are there natural spatial scales of the sub-networks within it? And do these scales correspond to previous electrophysiological [26] and theoretical [25] indications of functionally separate sub-regions of striatum? Finally, we could use our anatomical model as the basis for a dynamic model that showed the functional consequences of the network's structure. Our network models provided unique insights into striatal circuitry, overcoming the unintuitive nature of connectivity in three dimensions.
The striatal GABAergic microcircuit, shown in Figure 1, is formed by the connections between the GABAergic MSNs and FSIs. The MSNs are the only output neurons and comprise 90–97% of the neuron population in rat [19], [34], at a density of 84900 per [35]. The FSIs form 1–5% of the striatal neuron population [19], [36]. Stereological counting suggests that parvalbumin-immunoreactive neurons, the likely histochemical marker for FSIs [37], make up 0.7% of the striatum [38], [39]. As we will see, our model supports this lower estimate: only an FSI density of at most 1% resulted in numbers of FSI connections that are consistent with current data.
Four connection types make up the microcircuit. First, MSNs extend local axon collaterals that synapse on other MSN dendrites. Long-established anatomically [40], considerable electrophysiological evidence for them now exists [11], [21], [22], [41], [42]. Second, axon collaterals from FSIs synapse onto MSN dendrites and somas [43], and have a strong inhibitory influence [20]–[22], [44]–[46]. Third, FSI dendro-dendritic gap junctions [39], [47] electrically couple the paired cells [14], [20]. (Gap junctions between MSN dendrites probably occur only in immediate post-natal tissue [11], [48] so we do not consider them here). Finally, the FSI axon collaterals synapse onto other FSI dendrites: previously, evidence for these connections was indirect [47], with others finding no electrophysiological evidence of synaptic connection [20]; however, a recent study using transgenic mice found synaptic connections between pairs of striatal FSIs in the majority of cases [21]. We study the implications of this newly-described connection here.
The connection statistics between MSN pairs are partially known. Conservative anatomical estimates place 600 synapses on one MSN from other MSNs [19], [49]. Stimulating an afferent MSN elicits a post-synaptic response consistent with it making an average of 3 synapses on the target MSN [44]. This gives a lower bound of 200 MSNs afferent to each MSN. Other estimates suggest a single MSN contacts around 12–18% of the other MSNs within a radius axonal field, based on the observed frequency of synaptically-coupled pairs in stimulation studies [44], [50]. This gives an upper bound of about 470 MSNs afferent to 1 MSN within a radius, in good agreement with previous estimates [44], [51]. Finally, Planert et al [22] recently reported synaptic-coupling between 20% of all tested MSN pairs with somas within of each other.
There is less data on the statistics of FSI connectivity. Previous estimates of the number of FSIs afferent to a single MSN place bounds of 4–27 FSIs per MSN [19], [44]. Planert et al [22] recently reported synaptic-coupling between 74% of all tested FSI-MSN pairs with somas within of each other. Fukuda [39] reported densities between 500 and 4000 gap junctions per of striatal tissue, and observed typically 1–3 junctions per connected FSI pair. In Table 1, we use these data to calculate estimates for the number of FSIs connected to one FSI by gap junctions. These estimates show that we expect each FSI to be coupled to at most only a few others, and in many cases to have no gap junctions at all. As a consequence, and contrary to Fukuda's description of this network as “dense”, the FSI gap junction network seems to be very sparsely coupled.
Our aim was to construct a stochastic model of the three-dimensional network of the adult rat striatum, and study the statistics of contacts between the striatal GABAergic neurons. By “contact” we mean whether or not one neuron connects to another: a contact is one or more synapses or gap junctions. Our starting hypothesis was that, in a three-dimensional, non-laminar structure like the striatum, the minimum probability of contact between a pair of neurons is proportional to the density of their overlapping neurites. This is a passive process: numbers of contacts exceeding this minimum probability thus imply active processes, especially axon guidance towards specific types of target neurons. We encapsulate the role of active processes as an increase in the effective density of the axon. As we will see, this relatively simple model is able to capture the known statistics of the microcircuit's connectivity.
Figure 2 illustrates the steps in our approach to reconstructing contact probability functions, starting from models of dendrites and axons. We began by generating the dendrites and axons of both MSNs and FSIs using stochastic models (Figure 2A). For the dendritic trees we used an existing algorithm [52] that has been successfully applied elsewhere. However, some key parameters for this algorithm require data that are typically unavailable for most neuron types. We overcame this problem by finding these parameters using an evolutionary algorithm search of a fitness space defined by known properties (e.g. number of branch points) of the neuron type's dendritic tree. For the axon we created our own model based on known properties of MSN and FSI axons. By creating models for the dendrite and axon structure, we had a full set of data on the dendritic branches and axons at each distance from the soma, including their approximate volume (Figure 2B). Hence we produced a large number of dendritic trees and axons to estimate expected neurite volume.
We could then compute the expected spherical volume that was occupied by dendrite (or axon) at a given distance from the neuron body (Figure 2C,D). Then, in turn, we computed the expected volume of overlap between the spherical fields given the distance between neuron bodies for each connection type (Figure 2E). For every voxel in this overlapping volume, we computed the probability of its occupancy by both neurites (axon and dendrite or dendrite and dendrite, depending on the connection type) and thus the probability of intersection. Summing over all voxels in the overlapping volume thus gave us the expected number of intersections for each distance between neuron bodies (Figure 2F). We treat this as a probability of contact when constructing our three-dimensional networks. We elaborate on these steps below.
We chose the Burke algorithm [52] for reconstructing model dendrites. The Burke algorithm constructs dendrites in short, cylindrical segments long, each successive segment tapering in diameter as the tree extends away from the soma. Details of the Burke algorithm are given in Text S1. At each step of the algorithm, the current segment can either extend, branch, or terminate. If the segment branches, it bifurcates into two daughter segments, both narrower than the parent, and one larger than the other. If the segment terminates, the branch is complete and the algorithm moves to the segment on the next unfinished branch. This algorithm is repeated from a single starting segment to obtain each of the dendritic trees necessary to form a complete dendrogram: we built 6 trees for a complete MSN dendrogram [40], [53], [54], and 5 trees for a FSI [37]. The dendrogram records the diameter, distance from soma, parent segment, and end type (branch, termination, or continuation) of each dendritic segment.
The probability of a dendritic segment branching or terminating is a function of its diameter. To determine these probability functions, Burke et al [52] pooled morphological analyses of six spinal -motor neurons to obtain a distribution of the number of branch and termination points at each dendritic diameter, and found the probability per-unit-length of either termination or branching; all their resulting probability functions had the exponential form(1)for the probability of event (termination or branching ), given dendrite diameter and the free parameters and . A single function of this form was sufficient to fit the termination probability data; two functions of this form and were required to fit the branch probability data. The single branching probability was obtained by evaluating both and using the minimum value:(2)Figure 3A shows the termination and branch probability functions obtained from the -motor neuron data by [52] (for a segment length of ).
As for most neuron types, detailed data on the diameters of dendrites at branch and termination points are not available for MSNs and FSIs, and so we could not define the probability functions and apply the Burke algorithm directly. Instead, we gathered morphological data on the known properties of their dendritic trees (Table S1 in Text S1): branch order, dendritic radius, number of terminals, and terminal diameter. We then searched to find the parameters for the probability functions that resulted in dendrograms fitting all the constraints of the gathered data.
We use a simpler model for the axons, partly due to the absence of equivalent data to constrain a growth algorithm, partly because their structure is simpler, and partly because we can later use the axon model to encapsulate the process of attraction between axon and dendrites. The only quantitative description of local MSN axon collateralisation we are aware of is due to Preston et al [53], who described axons maintaining a diameter of over their initial length, then branching into 4 collaterals within of the soma, each with a diameter of (which terminate in extensive branching). This suggests an approximately two-fold increase in total diameter after all the branching had occurred. Similar branching patterns have been reported in [55]. We are not aware of any equivalent data for striatal FSI axons, and so use the same axon model for both as their axonal fields are similar [20].
Based on these observations, we proposed a sigmoidal model of the changes in axon diameter, in which the total axon diameter at distance from the soma is given by(3)We used and throughout for both MSNs and FSIs. With these values, the model captures all axonal branching occurring between and from the soma [53], [55], as illustrated in Figure 4A. We used a maximum distance from the soma of for both the FSIs [20], [37], [56] and the MSNs [53], [55].
Both MSNs [40], [63] and FSIs [20], [37] have approximately spherical dendritic and axonal fields. Following a mean-field approach, we thus made the simplifying assumption that the probability of finding the neurite is the same in all directions for a given distance away from the soma. We could then compute the following from the estimates of dendrite and axon volumes: the probability of finding dendrite (or axon) at a given distance from the soma, in a given volume of space; and hence the probability of intersection between two neurons' neurites in the same volume of space. To compute probabilities it was necessary to define the minimum volume required for a single intersection. The total volume of space was thus discretised into cubes or voxels that were on the side. We set to be consistent with the rat striatum's synaptic density of approximately 1 per [64]; this scale of individual intersections is also common to studies of rat cortical connectivity [31], [33].
As we are assuming that the probability of finding a neurite is invariant for a given distance from the soma, we proceed by considering successive spherical shells of width , the first shell wrapped around a sphere describing the soma. The voxels in a given shell will have the same probability of containing a neurite. The total volume of a shell at distance from the soma is(12)where is the radius of the soma: we used for both MSNs [65] and FSIs [20].
If the number of -on-the-side voxels in a shell at distance from the soma is(13)and the number of voxels occupied by dendrite in that shell is(14)(where is the total dendritic volume at that distance from the soma) then the ratio gives the probability of finding a dendrite-occupied voxel in that shell(15)Similarly, for axons occupying voxels of the shell, the probability of finding an axon-occupied voxel in that shell is(16)(For arbitrary distances from the soma, we could compute dendrite equations (14) and (15) using the continuous function fits or to the corresponding for FSIs and MSNs, respectively. Probabilities for the axons could be computed at arbitrary distances directly from equation (16) because the axon volume function (equation 11) is continuous).
Having obtained an estimate of the probability that a voxel contains an axon or dendrite, we could calculate the probability that a voxel contains an intersection between the neurites of two neurons. Let us denote the distance between the somas of the two neurons as . A given distance defines a volume of intersection between the two neurite-occupied spheres (Figure 2E). The centre of a given voxel in this intersecting volume is at distance from neuron 1 and distance from neuron 2. The probability of this voxel containing a neurite of the required type from both neurons is then(17)given the probabilities of finding a neurite from neuron 1 () and neuron 2 () in that voxel, from equation (15) or equation (16).
The total expected number of neurite intersections between two neurons at a given distance apart is then(18)where is the distance of the soma of neuron 1 from the th voxel, is similarly defined for neuron 2, and is the total number of voxels in the intersecting volume of the two neurite spheres. We calculated equation (18) for a range of inter-somatic distances , and fitted the resulting range of values with a continuous function so that we could obtain the expected number of intersections between a pair of neurons for an arbitrary distance between their somas. We did this for each of the four types of connection in the microcircuit (Figure 1), and the fitted functions are given in the Results – we add the additional subscript to denote which of the four connection types is being described.
We first define a volume of striatum we went want to model. The striatum contains 84900 MSNs per [35]; we added either 1% [19], 3% or 5% [36] of those as FSIs. We randomly assigned three-dimensional positions to each neuron, with a minimum distance of between neurons enforced, to model the non-laminar structure and intermingling of neuron types. To wire up the network, we treated the continuous functions giving the expected number of intersections between a pair of neurons as the probability of a contact between the pair of neurons. Hence, was treated as a contact with a probability of unity. Thus, given a particular distribution of neurons in space, with each pair at some distance , for each connection type we used as the binomial probability of a contact.
We explored the dynamical implications of some of our anatomical findings, using a computational model of the striatum drawn from our previous work [4]. In the model used here, the model neurons were wired together using our found intersection functions and the resulting network models; otherwise, the model neurons, synapses, gap junctions and inputs were as specified in [4]. Briefly, the neurons were simulated using the canonical, two-dimensional spiking model of Izhikevich [66], adapted to match the input/output properties of striatal MSNs and FSIs. We used conductance-based, single exponential synaptic models for intra-striatal connections (GABAa) and cortical input (AMPA and NMDA). As in the real striatum, we made synapses between model MSNs relatively weak, and the FSI synapses on MSNs relatively strong: following existing data [22], [44], the FSI-MSN synaptic conductance was five times greater than the MSN-MSN synaptic conductance. Gap junctions were modelled as a passive compartment between the coupled neurons, with a time-constant and conductance previously obtained by tuning to data on electrically-coupled cortical FSIs [4]. Cortical input was specified as the mean number of events/s arriving at excitatory synapses.
We ran two sets of simulations: one set used networks constructed within -on-the-side cubes of model striatum; the other used a network within a 1 mm-on-the-side cube. For the scale networks, we looked at the spontaneous activity of the striatal network in response to 10 seconds of background input of 475 events/s to every neuron (corresponding to around 1.9 spikes/s for 250 active afferents). For the 1mm-scale network, we selected the MSN closest to the centre of the cube as our reference neuron. We then stimulated all neurons in a series of wide spherical shells extending away from this central MSN. For each simulation, the central MSN and all neurons (MSNs and FSIs) in a shell were driven for 4 seconds with a mean of 1250 events/s (corresponding to around 5 spikes/s for 250 active afferents).
The evolutionary algorithm searches successfully found usable Burke algorithm parameters for both MSN and FSI dendrograms. The resulting parameters are given in Table 2. For MSNs, the top parameter set had a fitness of 83.3%, and was found on generation 44. For FSIs, the top parameter set had a fitness of 100%, and was found on generation 114. Both top sets were thus found well before the termination of search, and are likely to be close to the best available given the initial population. (Note that a fitness of 100% does not mean that the parameter set guarantees an accurate dendrogram every time, due to the stochastic nature of the Burke algorithm). We used these parameters to generate MSN and FSI dendrograms.
The resulting probability functions for branching and termination of the MSN and FSI dendrites are shown in Figure 3. The search results predict that, because for the second branching probability function is very small (Table 2), only a single exponential is effectively needed to describe the branching probabilities of both neuron species, rather than the two exponentials fitted by Burke et al [52] to their motorneuron data. This suggests some fundamental difference in the morphology of MSNs and FSIs, compared to the morphology of the motorneurons studied in [52].
The resulting MSN dendrogram models made some interesting predictions. Figure 5A shows that the predicted dendritic taper of the MSN model closely approximated the dendritic taper data recorded from real MSNs ([67]; data from C. Wilson, personal communication). The data from the real MSNs suggests a sharp initial decrease in diameter as the dendrite leaves the soma that is not captured by the model, but otherwise the tapering is of a similar form.
Second, the model MSN dendrograms predicted that existing data on total dendrite length and estimates of spine counts are mutually inconsistent. The median total dendrite length, averaged over all instantiated dendrograms, was (range 2693–4925), exceeding the previously obtained median value () and range of reported by Meredith et al [54] across 22 MSN reconstructions. The predicted number of spines on the whole dendritic tree was (mean 2 s.d.), a mean value lower than the bottom end of the previously predicted range of 6250–15000 spines per neuron based on the same original spine data [68]. The dendrogram model has thus shown that, even if the total dendritic length extends beyond the reported data, we cannot recover these total spine estimates.
A third prediction is that the spines are in abundance in the proximal dendrites. We plot the histogram of the MSN dendrograms' mean spine counts in Figure 5B and see that it is skewed, with half of all spines occurring within of the soma. The MSN model also shows us that the long-tailed fall-off of the number of spines when moving further away from the soma is primarily due to a corresponding fall in the number of processes across the whole dendrite (Figure 5C).
We used the instantiated dendrograms to find the mean total volumes of the MSN and FSI dendrites per step (equations 4–9). Having found these mean total volumes over a range of distances from the soma, they were fitted with functions of the form(19)to obtain functions and giving us the volume of MSN and FSI dendrite, respectively, at arbitrary distance from the soma. Table 3 gives the best-fit parameter values (found using non-linear least squares, as implemented by MATLAB function lsqcurvefit). Both the functional form and the transform in equation (19) were necessary to accurately fit the tails of the total volume distribution (Figure 6A). The transform overcomes the problem that using summed-squared error favours close fits to higher magnitude data-points, as the majority of ‘error’ occurs for them.
The importance of close-fitting to the tails becomes clear when we consider the probabilities of finding a neurite-occupied voxel, and the subsequent intersection calculations. When we compute the probability of finding a dendrite-occupied voxel (Figure 6B), we see that it falls faster than the dendrite volume (compare Figure 6A): the volume of the embedding spherical shell increases cubically with each step. Yet when we turn to compute the number of intersections, the number of voxels also increases cubically with each step. Hence, at intermediate distances from the soma, the very small probabilities of finding neurites are counteracted by the very large number of voxels checked for intersections. Poor fits to the tail thus incur noticeable changes in the number of expected intersections.
Finally we turn to actually computing the expected number of intersections for each of the MSN and FSI connection types in the striatal GABAergic microcircuit (Figure 1): local axon collaterals connecting MSNs [40]; projections from FSIs onto MSNs [20]; axo-dendritic synapses between FSIs [21]; and dendro-dendritic gap junctions between FSIs [20], [39], [47].
We used the expected intersection functions (equation 20) to construct model striatal networks, which we could examine for their predictions of striatal connectivity. We built networks within a cube of striatal tissue, giving us 84900 MSNs, with 1%, 3% or 5% FSIs added (see Materials and Methods). For every neuron within of the centre, we found all of its targets, afferents, and the distances to and from them. Restricting ourselves to this radius ensured that we could identify neurons that were little affected by their proximity to the edges of the volume, having complete afferent and efferent intra-striatal connectivity; hence we considered them the best candidates for comparing to, and making predictions about, the real striatum. To get sufficient numbers for analysis, we constructed 10 networks for each FSI percentage and pooled the data.
The constructed networks predict that, for all but the FSI gap junctions, the numbers of and distances between connected pairs of neurons have Gaussian distributions (Table 5), despite the complexity of the individual expected intersection functions (Figure 7). Figure 8 shows these Gaussian distributions for the MSN inputs to each MSN: each has MSN afferents, at distances of (note the distribution is truncated at the minimum distance of ). As we show in Figure 8C, the exception, consistent for each FSI percentage we tested, is the log-normal distribution of distances between gap-junction coupled FSIs.
Table 5 shows how the distributions of numbers and distances of contacts change for all connections across the 1, 3, and 5% FSI networks. The number of connected MSNs remains constant at around 728 MSNs afferent to one MSN. We found that if we restricted counting inter-connected MSNs to just those within of each other, then each MSN receives MSN afferents and, hence, has a probability of being connected with another MSN in that radius, in excellent agreement with previous estimates (see The microcircuit and connection statistics). The number of MSNs contacted by one FSI (‘1 FSI-MSNs’ in Table 5) stayed constant, as expected, at around 3000 MSNs per FSI. The number of FSIs afferent to a single MSN increased with increasing FSI percentage. The 1% FSI network predicts around 30 FSIs per MSN, in good agreement with previous estimates of 4–27 FSIs per MSN [19], [44] – the other FSI percentage networks fall well outside these bounds. Similarly, the numbers of synaptic and gap junctions contacts between FSIs increased when increasing the percentage of FSIs in the network models. The mean numbers of gap junction contacts per FSI are only in good agreement with our estimated ranges from Fukuda's [39] data (see Table 1) for an FSI density of 1%.
A striking prediction of the network model is that the mean afferent distances for FSI and MSN inputs to a MSN and for FSI synaptic inputs to other FSIs are all (for 1% FSI networks; the 3% and 5% networks have slightly lower mean distances for FSI input to other FSIs). This strongly suggests a natural spatial scale for the dominant inhibitory synaptic input to a MSN or FSI. Further, the network model predicts that a FSI's gap junction network is focussed locally around the neuron. Both these properties have implications for the dynamics of the striatum, which we illustrate below.
The models of the striatal network revealed two striking features that could play a key role in striatal dynamics: the differing spatial scales for the inter-FSI gap junction and synaptic contact networks, and the common mean distances of GABAergic afferents to one MSN. We show here that both features indeed have the potential to set the input-output relationships of the striatum. To do so, we use a computational model of striatum that took the developed models of neurons (MSNs and FSIs), synapses (AMPA, NMDA, and GABAa) and gap junctions from our previous work [4], but used the striatal network model developed here as the basis for wiring the neurons together.
We have established a complete protocol for constructing a biologically-realistic network from first principles. The process described here is of general interest: in principle it could be used to model any region of the brain. It is particularly suited to the reconstruction of three-dimensional networks in non-layered structures, and we used it to reconstruct the GABAergic microcircuit of the adult rat striatum.
Attempting to specify construction algorithms for the dendrites and axons showed specifically where quantitative morphology data were missing (we provide a complete list in Text S1). Building the MSN dendrite models revealed an inconsistency between previously reported total dendrite lengths and the number of spines on the MSN dendrites: the dendrite model had more wire, yet fewer spines. This suggests that the previously predicted range of 6250–15000 spines per MSN [68] is an overestimate: the model dendrites suggest a mean of 5932 spines per MSN – implying that, as each spine maintains a cortico-striatal synapse [40], [59], there are fewer cortical inputs to a MSN than previously estimated [68]. Moreover, the dendrite model predicts half of all spines are within of the soma, half the radius of the MSN dendritic tree. As cortico-striatal synapses occur only on the spines [40], [59], this suggests half of all cortical input is to the proximal dendrites.
Using the axon and dendrite models, we found that achieving the target probabilities from [22] for MSN-MSN and FSI-MSN contacts within required large axon density constants . Matching the target MSN-MSN probability of required an increase of the effective MSN axonal volume by a factor of ; matching the target FSI-MSN probability of required an increase of the effective FSI axonal volume by a factor of . Both these results imply a dominant role for active processes guiding axon to dendrite in wiring up the striatum, beyond passive intersection of dendrite and axon alone. We used the same FSI axon scaling factor to construct the synaptic connections between FSIs: the intersection function for this connection is, hence, currently a prediction of the model. By contrast, we found that the density of FSI gap junctions was captured by the model using passive intersections of dendrites alone.
The expected number of intersections between neurons had a product-of-exponentials form (Figure 7), with five parameters whose precise values (in Table 4) would be difficult to recover from anatomical data. Nonetheless, the characteristic double-exponential function (equation 20) could, in principle, be recovered qualitatively. Furthermore, we have shown that the probability of contact between two neurons need not be a simple exponential function of distance [31].
When we applied the intersection functions to construct the striatal network models, we found though that almost all distributions of numbers of contacts and their distances were Gaussian. The network models predicted that each MSN receives an average of 728 inputs from other MSNs, when considering the complete network. Confidence in this result stems not just from the tuning to match the data on probability of contact within , but also from the model having a mean number of 296 MSN-MSN contacts within , which is in excellent agreement with previous estimates (see The microcircuit and connection statistics). The network model predicts each FSI contacts around 3000 MSNs, which may explain why the FSIs, despite being few in number, are able to potently suppress MSN activity across the striatum [45].
The numbers of contacts for the other connection types were dependent on the percentage of FSIs in the network. Mean numbers of contacts in the 1% FSI network are consistent with existing estimates for the number of FSIs contacting one MSN [44], and the density of FSI-FSI gap junctions [39] (albeit at the lower end of the ranges we calculated from Fukuda's [39] data in Table 1). By contrast, the 3% and 5% FSI networks predict too many FSI inputs per MSN, and too many FSI gap junctions. Hence, the network model is consistent with recent estimates that at most 1% of striatal neurons are FSIs [19], [38], [39]. Given the decreasing density of FSIs over the dorsolateral-ventromedial axis of the striatum [12], [38], [70], it is plausible that even lower densities of FSIs occur in some striatal regions. Irrespective of the exact FSI percentage, the network models showed that the full three-dimensional network of the striatum is extremely sparse, forming around 1.7% of all possible MSN-MSN contacts and 7% of all possible single-FSI-to-many-MSNs contacts that could be made given the radii of dendritic and axonal fields.
The network models made two striking predictions about the spatial organisation of contacts in the striatum. First, that the networks of gap junctions and synapses inter-connecting FSIs were on different spatial-scales: the log-normal, left-skewed distribution of gap junction distances implies each FSI makes most of its electrical connections with immediately neighbouring FSIs; the Gaussian distribution of synaptic distances implies each FSI makes most of its synaptic connections with FSIs more distally. Second, that inputs to a MSN from either FSIs or other MSNs are on the same length scale of . This result illustrates the unintuitive nature of three-dimensional connectivity: the fall-off in the probability of connection is counteracted by an increase in the number of neurons to make connections with, so that the dominant distance of connections is some function of both.
The anatomical model results point to some intriguing implications for the spatial scales of computation in the striatum. The “domain” theory [9], [24], [25] suggests that the natural computational element of the striatum is the network of MSNs within the radius of one MSN's dendritic field. On the one hand, the network models confirm that MSNs have an approximately probability of contacting another MSN within that radius, far greater than the probability of contacting farther MSNs, suggesting the formation of a closely-knit network – if we consider only probability of connection. On the other hand, our network models show us that the number of connections give us the inverse of the domain concept: a MSN receives its greatest number of inputs from MSNs and FSIs whose soma lie just on edge of the main extent of its dendrites. The comparatively weak nature of the individual MSN-MSN synapse – with a mean conductance approximately five times smaller than the FSI-MSN synapse [21], [22], [44] – suggests that the number of MSN inputs is the key factor in understanding the influence of the local MSN axon collateral network. We showed in a computational model that this is indeed the case: the most potent inhibition of an active MSN was achieved by stimulating inputs around away (Figure 10E, Figure S1). Therefore, if a ‘computational element’ of the striatum is defined by the spatial scales of feedback and feedforward inhibition – from other MSNs and FSIs, respectively – then our models show it to be spread over the network, not concentrated locally within the dendritic field.
The spontaneous organisation of activity in striatum is consistent with such a widespread network of effective MSN-MSN connections. Carillo-Reid et al [29] showed that global excitation in vitro induced the appearance of a few cell assemblies within an plane, with each assembly comprising neurons spread over the plane. Models of this phenomenon from both us [4] – using distance-dependent connections, as here – and Ponzi and Wickens [16] – using uniform probability of connection – show that such cell assemblies are not formed by discrete groups in physical space. The data and models also showed that such assemblies contain comparatively small numbers of neurons (at most a few hundred) on the scale of other definitions for a striatal ‘computational-element’. Future work with the model reported here will examine the reasons for this discrepancy. What the network model, and the computational model built upon it, do make clear is that further understanding the computations performed by the striatal microcircuit requires better knowledge of the distribution of individual cortical inputs [63], [68], to understand if they are organised along any of the characteristic spatial scales of the striatal network.
The network model also showed that the density of FSIs affects both the number and spatial-scales of connections. We showed that these anatomical effects are reflected in changing dynamical properties of the FSI network in the computational striatum model. Changing the FSI density altered the distribution of FSI firing rates, decreasing the proportion of active FSIs, but increasing the range of rates. However, irrespective of the FSI density, the FSI network remained in a globally-asynchronous state, with many FSIs completely or nearly silent. Though contrary to previous reports that networks of spiking neurons coupled by both gap-junctions and inhibitory synapses promote globally synchronised activity (e.g. [71]), our findings are consistent with both our previous work [4], and with Lau et al's [15] report that asynchronous, partially-silent states dominate in such networks if the gap junction network is not wired together as a classic random network – that is, one with a uniform probability of connection. In the case of Lau et al [15], this wiring was a small-world network on a ring-lattice; here our striatal anatomy model deviates from a classic random network because of the distance-dependent probability of the gap junction connections between the FSIs. Taken together, Lau et al's [15] results, and ours here and previously [4], all point to the importance of considering both the wiring topology and its spatial embedding when considering the dynamics – and, hence, likely function – of interneurons coupled by both gap junctions and inhibitory synapses.
The local and distal networks formed respectively by the inter-FSI gap junctions and synapses produced characteristic properties of the MSN population dynamics too. The MSN population activity was remarkably consistent across changes in FSI density (Figure 9A), despite the changes in FSI activity just described. However, when we swapped the inter-FSI networks (gap junctions distally, synapses locally), the MSN firing rates now changed with changing FSI density. Hence, the combination of local gap junction and distal synaptic networks predicted by the model constrains the whole MSN population to a particular input-output regime, robust to changes in FSI density.
We also saw that in the absence of active FSIs, the MSN population activity was globally reduced compared to all normal models with active FSIs. This result in a larger and anatomically more detailed model confirms our previous finding that removing all FSIs reduces model MSN activity [4]. The unintuitive effect that increasing the number of GABAergic interneurons increases the firing rate of their target neurons has a clear underlying cause: the GABA reversal potential is above the typical MSN ‘down’-state membrane potential [11], and hence sporadic FSI input to MSNs will tend to keep their membrane potential relatively depolarised, allowing them to fire (and fire more often) to excitatory input (too much FSI input, however, would clamp the MSN membrane potential at the GABA reversal potential).
The results of this work have further applications in the study of both single-neuron and network-level dynamics. By using our found parameters for the Burke algorithm, it is possible to generate many MSN and FSI dendritic morphologies, each consistent with current morphological data. Hence, instantiating the same multi-compartmental model (e.g. [13], [67], [72]) on multiple instances of these generated morphologies will open up a wide range of applications, such as placing limits on post-synaptic potential summation, back-propagating action potentials, maximal conductance searches, and so on. At the network level we have shown how we gain the benefits of reconstructing the underlying structure, as argued at the outset of this paper. Particularly interesting will be the results of using these reconstructed networks – requiring only equation (20) – as the basis for other group's approaches to modelling the striatum [13]–[16], [72], [73].
We have focussed on the principal GABAergic microcircuit of the striatum here, as this provides the basis for the most immediate, powerful control over the output of the striatum [4], [45], [50]. The current work has thus omitted other interneuron types. A full striatal network reconstruction would include the giant cholinergic interneurons with their dense and long-reaching axonal ramifications that synapse on MSNs [37], [74], and the low-threshold spiking interneurons [37], which may form an inhibitory network between the cholinergic interneurons [75] or control gap junction efficacy through the release of nitric oxide [21], [76]. In addition, the network construction does not currently address what happens at the histochemically defined borders between the ‘patch’ and ‘matrix’ of the striatum [34], which many MSN dendrites do not cross [77].
Our model is a stochastic realisation of the adult striatum; the modelling of developing striatal connectivity is a stern challenge given the current paucity of data [78]. Nonetheless, we think modelling the development of connectivity is essential to capture elements of striatal wiring we have not accounted for in the present model. For example, recent work on BAC transgenic mice suggests a preferential direction of connection between the two populations of MSNs defined by their dominant dopamine receptor type (D1 or D2), with significantly fewer projections from D1-expressing to D2-expressing MSNs than any other combination [21], [22], [79]. With no data yet on how such selective connectivity might form, we must chalk this up as a future target for our models. Clearly we are only at the beginning of constructing realistic models of the striatum; but equally we have a promising start.
|
10.1371/journal.pcbi.1000096 | Model-Based Hypothesis Testing of Key Mechanisms in Initial Phase of Insulin Signaling | Type 2 diabetes is characterized by insulin resistance of target organs, which is due to impaired insulin signal transduction. The skeleton of signaling mediators that provide for normal insulin action has been established. However, the detailed kinetics, and their mechanistic generation, remain incompletely understood. We measured time-courses in primary human adipocytes for the short-term phosphorylation dynamics of the insulin receptor (IR) and the IR substrate-1 in response to a step increase in insulin concentration. Both proteins exhibited a rapid transient overshoot in tyrosine phosphorylation, reaching maximum within 1 min, followed by an intermediate steady-state level after approximately 10 min. We used model-based hypothesis testing to evaluate three mechanistic explanations for this behavior: (A) phosphorylation and dephosphorylation of IR at the plasma membrane only; (B) the additional possibility for IR endocytosis; (C) the alternative additional possibility of feedback signals to IR from downstream intermediates. We concluded that (A) is not a satisfactory explanation; that (B) may serve as an explanation only if both internalization, dephosphorylation, and subsequent recycling are permitted; and that (C) is acceptable. These mechanistic insights cannot be obtained by mere inspection of the datasets, and they are rejections and thus stronger and more final conclusions than ordinary model predictions.
| Insulin is a central player in maintaining energy balance in our bodies and in type 2 diabetes, where the effect of insulin on its target tissues is diminished. Insulin acts on cells by binding to specific insulin receptors (IRs) at the cell surface. This triggers a series of events, including attachment of phosphate to IR, activation of downstream proteins that eventually mediate the signal to specific targets in the cell, and internalization of IR to the inner cytosolic part of the cell. The importance, time relations, and interactions between these events are not fully understood. We have collected experimental time-series and developed a novel analysis method based on mathematical modeling to gain insights into these initial aspects of how insulin controls cells. The main conclusion is that either IR internalization and the subsequent recycling back to the cell surface or feedbacks from downstream proteins (or both) must be significantly active during the first few minutes of insulin action. These conclusions could not have been reached from the experimental data through conventional biological reasoning, and this work thus illustrates the power of modeling to improve our understanding of biological systems.
| Insulin is the primary hormone in control of whole body energy metabolism in human beings. The hormone is secreted to the blood circulation by the β-cells, located in the islands of Langerhans in the pancreas. The adipose tissue and the adipocytes are important targets for insulin control of energy metabolism. Failure of the adipocyte and other target cells to properly respond to insulin, insulin resistance, is often associated with obesity and is a distinguishing feature of type 2 diabetes.
Insulin controls cellular metabolism by binding to the insulin receptor (IR) at the surface of the cell (reviewed in [1]). In response to insulin-binding the intracellular β-subunits of the transmembrane receptor, which carry protein kinase activity, autophosphorylate on specific tyrosine residues. Thus autophosphorylated, the IR is active against a set of intracellular signal mediator proteins, in particular the insulin receptor substrate-1 (IRS1), which becomes phosphorylated on tyrosine residues. Phosphotyrosines in IRS1 are recognized by proteins, containing a SH2-domain, which by binding to phospho-tyrosine become activated to transduce the insulin signal further downstream. The signaling eventually affects cellular metabolism, for example through an increase of glucose uptake or inhibition of lipolysis. Many of the downstream intermediary steps in the insulin signaling network of the target cells remain unidentified. However, also apparently well-characterized early aspects of insulin signal transduction remain incompletely understood and may thus also reveal novel features of importance for insulin action, both in normal and in disease states.
At the plasma membrane of adipocytes, the IR has been shown to be localized in plasma membrane microdomains, invaginations of the membrane, referred to as caveolae [2]. It is important that in human fat cells, but for instance not in rat adipocytes, the IRS1 is co-localized with the IR in caveolae [3]. In conjunction with insulin-binding the IR is internalized by endocytosis [4],[5], but the function of IR endocytosis has not been demonstrated. It may be to turn off signaling, e.g., by dephosphorylation of the receptor, by downregulating the number of IRs at the cell surface, or by clearing insulin from the circulation. Conversely, endocytosis may be a part of the signal transduction per se, e.g., by gaining access to downstream signaling intermediates or by providing for compartmentalization of the signaling. It has not yet been possible to determine experimentally which of these alternatives that are of highest importance, at the various time-scales involved in the signaling [6]–[15]. Conversely, the insulin-controlled internalization of IR has been shown to depend on IR autophosphorylation [13],[16],[17], but to be independent of downstream activation of IRS or phosphatidylinositol-3 kinase [17].
To gain further insight into which mechanisms that are most active during the early events of insulin signaling, we have measured the transient phosphorylation of IR and IRS1 during the first ten minutes after a step increase in extracellular insulin concentration. The mechanistic explanation to such transient data is typically not evident from a mere inspection of the time courses. Nevertheless, such data contains valuable information on the active mechanisms in a complex system, and measurements of rapid transient responses is one of the most widely used methods for characterization of technical systems [18]. In such studies, the information in the data is typically extracted from the data using a model based hypothesis testing approach. Such an approach is different from the kind of large-scale gray-box modeling approaches that typically are used in systems biology studies. Two such related models are [19],[20], and large-scale gray-box models are in general characterized by the fact that many more interactions are included than can be tested from the existing data. Conversely, in the hypothesis testing tradition followed here, we do not include all known mechanisms in the models. This typically corresponds to setting parameters to zero in a comprehensive model, and a key question is whether the included mechanisms are sufficient, necessary, or not sufficient, to explain the data. This gives information on which mechanisms that may, must, and cannot be significantly active during the specific time-scale. Apart from the overall methodology, the work also makes use of several non-trivial theoretical results and methods that can be re-used in other analyses of signaling systems.
We examined the extent of phosphorylation of IR and IRS1 on tyrosine residues in human adipocytes. In three separate experiments, data were collected at 10 time points during 15 min, following a step increase from 0 to 0.1 µM in insulin concentration (Figure 1). The experimental set-up is limited to measurements of relative changes, i.e., all signals come with an unknown scaling factor. We measured phosphorylated and total IR and IRS1 by SDS-PAGE and immunoblotting. To achieve a robust measurement signal, the extent of phosphorylation of both IR and IRS1 were divided by total amount of IR and IRS1, respectively. The resulting signals are therefore proportional to the relative degree of phosphorylation of IR and IRS1. The rapid initial transient response was higher than the quasi-steady state level attained after about 5 min for both IR and IRS1 (Figure 1). This transient behavior is referred to as the overshoot in the data. The overshoot is clearly present both in each individual time course, and in their mean values. We now use a model based hypothesis testing approach, to translate these experimental observations to mechanistic insights.
Three hypotheses are considered as possible mechanistic explanations to the observed overshoot. The first of these hypotheses, hypothesis A, assumes that the overshoot is generated by an interplay between the autophosphorylation and protein phosphatase activity at the plasma membrane only. It is interesting to consider the possibility whether such mechanisms might be the only ones significantly active in the IR signaling subsystem, since we are only considering the first few minutes of the response. The analysis shows that this possibility can be rejected based on the information in the collected data.
While hypothesis B is like A with the additional possibility of endocytosis, hypothesis C is also like A but with the additional possibility of feedbacks from downstream signaling intermediates. There exist, indeed, in the literature suggestions of feedback from downstream signaling intermediates to, e.g., the phosphotyrosine protein phosphatase activity [22]–[25]. We here suggest an archetypical version of such a feedback, to show that hypothesis C also provides an acceptable explanation of the data. The suggested model structure (ℳf; Figure 3) includes activation of IRS1 and its subsequent activation of X, which refers to some non-identified downstream signaling intermediate. The notation X is chosen in order to illustrate the fact that it is impossible to conclude without further experiments which specific feedback that is most likely to generate the observed behavior in the experimental data, and only that any feedback of the given character would be sufficient. The feedback to PTP illustrates the archetypical feedback, which also could be illustrated by a direct feedback to the IR, by for instance its serine phosphorylation [26]. The agreement between this model structure and the data is just as convincing as that for the minimal model ℳi,b. Since it is sufficient that a single model structure from a given class produces a satisfactory explanation, in order for the whole class to be acceptable, we have now shown that hypothesis C is an alternative explanation to hypothesis B.
This paper has two parts. The first part reports a rapid overshoot in IR and IRS1 phosphorylation upon insulin stimulation of human fat cells. These observations, although interesting in themselves, do not provide any mechanistic insights by themselves, and mere inspection and reasoning around the data is not sufficient to evaluate which mechanisms that may and may not explain the given data in a satisfactory manner. The second part of the paper analyzes three biologically realistic and plausible mechanistic explanations: (A) direct phosphorylation and dephosphorylation of IR at the plasma membrane only; (B) the additional possibility of IR endocytosis; (C) the alternative additional possibility of feedback to IR from downstream intermediates. Our analysis has shown that A is not a satisfactory explanation, that B provides such an explanation if both internalization and subsequent recycling are included, and that hypothesis C provides such an explanation.
The mechanistic insights obtained here are the result of model based hypothesis testing and there are some important properties of such studies that should be pointed out. In a hypothesis testing framework, the most interesting result is when a model may be rejected. A rejection is also the kind of conclusion that is hardest to achieve. Ideally, all parameter values in all model structures belonging to the class of model structures corresponding to the tested explanation should be evaluated before a rejection has been shown. Conversely, evidence of the sufficiency of a mechanistic explanation is shown already by the existence of a single model structure at a single parameter point which gives a satisfactory agreement. Further, a model rejection is a strong statement since it will not be altered when new data are collected (unless, of course, the new data would point to errors in the previous data). The conclusions drawn here are thus not typical model predictions to be tested in validation experiments, but evaluations of possible mechanistic explanations for a given data set.
The significance of this modeling approach becomes evident when comparing with a previous modeling work by Sedaghat et al. [20]. That model structure is an example of a large-scale mechanistically detailed model for insulin signaling, and it includes both internalization of the insulin receptor and feedback effects from downstream metabolic intermediates to IRS1. Interestingly, the feedback signals do generate an overshoot in IRS1 phosphorylation. However, the Sedaghat model does not predict an overshoot in the IR phosphorylation, and must generally be revised to serve as a (single) explanation to our experimental data [27]. More importantly, however, the single model structure in [20] was evaluated at a single parameter point, and [20] is therefore a qualitatively different type of study than ours. A main drawback of such purely forward-simulation based studies is that most parameter values are unknown, especially in vivo. Analysis at a single parameter point is of course problematic if the chosen parameter values are unrealistic (which is the case for instance for the internalization constant in [20]). However, also if all parameter values are realistic, one does not know which model predictions and parameter values (i.e. active mechanisms) are necessary consequences of the given data and model structure, and which model predictions are merely outcomes of more or less arbitrarily chosen parameter values. An example of a stronger model prediction is for instance that for ℳm,PTP herein, which says that all parameter values that give an acceptable agreement with our experimental data must also give a steady state concentration of (IR⋅ins)⋅PTP larger than 25%. Finally, it should be noted that not even an ordinary model rejection, reporting a lack of agreement with the data, may be done without global searches among all realistic parameter values.
There are also other related works. Interestingly, a transient overshoot in the phosphorylation of internalized IR has been reported [14]. However, that work did not provide any mechanistic explanations. A simulated model of signaling by the epidermal growth factor (EGF) receptor has been found to exhibit a transient phosphorylation overshoot when endocytosis of the receptor is included in the model [19]. However, the EGF receptor has a different mechanism of activation than IR, and there does not exist a thorough hypothesis testing approach that evaluates which mechanisms that may, and may not, produce such an overshoot. In a more recent time-course modeling of IR phosphorylation and endocytosis in Fao hepatoma cells [28], no transient phosphorylation overshoot was included, neither in the experimental data, nor in the model. Further, the authors used the Akaike Information Criterion (AIC) hypothesis testing approach to distinguish between all possible model structures. The AIC simply chooses a model as the best one, by weighting model agreement against number of parameters. This means that AIC does not provide any statistical measure on whether any of the evaluated models show an acceptable agreement with the data, or what the statistical significance of the conclusions are. That means that the AIC test alone would not have been sufficient to find the main conclusions and rejections provided in this article.
Our statistical testings are based on a number of assumptions. For instance, the noise in the system is approximated by white and Gaussian signals appearing exclusively in the measurements. This means that intrinsic system noise has been neglected, as have the indications that experimental noise from immunoblotting might be log-normal. To compensate for this limited complexity of the noise model, the variance of the noise has been exaggerated, and for many analyses only the most prominent features of the data (primarily the overshoot) have been used for the rejections.
Other limitations in our assumptions are due to our usage of ordinary differential equations (ODEs). This means that stochastic effects from individual particles, or individual cells, and subtle spatial phenomena (everything besides the internalisation itself) all are disregarded. These approximations have been judged acceptable since the available data do not allow for a more detailed inspection of the processes. It will be an important step forward in our understanding of these processes when we can measure data containing spatially resolved single cell data, and when we can more realistically describe processes in micro-environments such as caveolae, where IR and IRS1 are situated. So far, we can only speculate what the corresponding conclusions might be in such studies. For instance, the number of IR proteins per fat cell has been estimated to >2×105 [29], and this should, according to generic studies such as [30], mean that molecular stochastic effects are insignificant, at least if the assumption of fast diffusion within the cell is valid. However, when it comes to incorporating the caveolae micro-environment properties, the fundamental kinetics will probably change (see e.g. [31]), and we have to-date no good guidelines for how such generalisations change the properties of a system.
In any case, despite these limitations, statistical assessments of the degree of uncertainty underlying model rejections do provide more detailed and objective statements than those based on simulations and/or subjective judgments alone. Most importantly, we have been able to draw mechanistic insights from a given set of time-series data; these mechanistic insights could not have been drawn using only classical biochemical reasoning.
Samples of subcutaneous abdominal fat were obtained from female patients at the University Hospital of Linköping. Patients with diabetes were excluded. Pieces of adipose tissue were excised, during elective abdominal surgery and general anesthesia, at the beginning of the operation. The study was approved by the Local Ethics Committee and participants gave their informed approval.
Rabbit anti-insulin receptor β-chain polyclonal and mouse anti-phosphotyrosine (PY20) monoclonal antibodies were from Transduction Laboratories (Lexington, KY, USA). Rabbit anti-IRS1 polyclonal antibodies were from Santa Cruz Biotech. (Santa Cruz, CA, USA). Insulin and other chemicals were from Sigma-Aldrich (St. Louis, MO, USA) or as indicated in the text.
Adipocytes were isolated by collagenase (type 1, Worthington, NJ, USA) digestion as described [32]. At a final concentration of 100 µl packed cell volume per ml, cells were incubated in Krebs-Ringer solution (0.12 M NaCl, 4.7 mM KCl, 2.5 mM CaCl2, 1.2 mM MgSO4, 1.2 mM KH22PO4) containing 20 mM Hepes, pH 7.40, 1% (w/v) fatty acid-free bovine serum albumin, 100 nM phenylisopropyladenosine, 0.5 U/ml adenosine deaminase with 2 mM glucose, at 37C on a shaking water bath. For analysis after 20–24 h incubation, cells were incubated at 37C, 10% CO2 in the same solution mixed with an equal volume of DMEM containing 7% (w/v) albumin, 200 nM phenylisopropyl adenosine, 20 mM Hepes, 50 UI/ml penicillin, 50 µg/ml streptomycin, pH 7.40. Before analysis cells were washed and transferred to the Krebs-Ringer solution. Cells were then incubated at 37C with 100 nM insulin for the indicated time period.
Cell incubations were terminated by separating cells from medium by centrifugation through dinonylphtalate. The cells were immediately dissolved in SDS and β-mercaptoethanol with protease and protein phosphatase inhibitors, frozen within 10 sec, and thawed in boiling water to minimize postincubation signaling modifications in the cells and protein modifications during immunoprecipitation [32]. Equal amounts of cells as determined by lipocrit, that is total cell volume, were subjected to SDS-PAGE and immunoblotting. After SDS-PAGE and electrotransfer membranes were incubated with indicated antibodies that were detected using ECL+ (Amersham Biosciences) with horseradish peroxidase-conjugated anti-IgG as secondary antibody, and evaluated by chemiluminescence imaging (Las 1000, Image-Gauge, Fuji, Tokyo, Japan).
By two-dimensional electrofocusing (pH 3–10) - SDS-PAGE analysis and immunoblotting against phosphotyrosine and IRS1, >95% of the tyrosine phosphorylated 180-kD band was determined to represent IRS1 [33].
In this paper we evaluate three mechanistic hypotheses for the explanation of experimentally observed phosphorylation dynamics. Each of these three hypotheses are too general to correspond to a single mathematical model that can make specific predictions, which can be compared with data. For this reason we consider classes of model structures in our analysis. In practise, these classes are approximated by a large number of specific models. In this paper we restrict ourselves to the consideration of models described by ODEs. The general form of an ODE is given by(1A)(1B)where x∈ℝn is the n-dimensional column vector containing the state variables (concentrations denoted by a square bracket), f is a well-behaved (e.g., continuous and differentiable) function, p∈ℝr contains the parameters, and y contains the measurement signals whose relation to the state variables and the parameters is given by the function h. Spatial transport in the form of endocytosis and recycling is described by the introduction of compartment specific state variables, where the subscript i denotes state variables that have been internalized. All the models are uniquely given by the figures according to standard interpretation of such figures; examples and more details are included in the Text S1.
Models are sought to be rejected in several different ways. The first way is rejection through analysis of a corresponding transfer function form. Transfer functions are commonly used for linear models [34], while the models considered here are nonlinear. Nevertheless, for the specific input studied (a step function), we can find equivalent linear models giving exactly the same responses, i.e., without approximations. This holds for all models accept ℳf, and to see it on a more general level, consider the systemwhere A(⋅) is an ℝn×n-valued function, x(0)∈ℝn is the state vector, e.g., concentrations of relevant substances, and u(t) is the input to the system. If u(t) changes from 0 to u0 at t = 0, the state vector x(t) will follow the same trajectory as for the system(2)where δ(t) is the Dirac function. In other words, we can study the impulse response of a linear system instead. Taking the Laplace transform of Equation 2 yields(3)
For instance, for ℳm,a we have the following state-space description (derived in the Text S1)(4A)(4B)and the following transfer function description(5)with X(s) being the Laplace transform of . Note the pole in s = 0, which means marginal stability. Biologically, this is due to the mass conservation, saying that [IR]+[IRp] is constant. For the same reason, all our considered model structures will contain a pole in s = 0.
Now, X(s) in Equation 5 can mathematically be interpreted as the step response of the transfer function(6)
This allows us to transfer standard results from linear systems theory to our specific application.
We have derived two general results that allow for rejection by direct inspection of the transfer functions, i.e., without considering specific parameter values; these are presented in the following two subsections.
Lemma 1 Consider a stable, linear time-invariant system with transfer function G(s) having real poles and no zeros. Then, the impulse response of G(s) is positive for all t>0, i.e., the system cannot display an overshoot.
Proof. Since G(s) has real poles, we can write G(s) as a cascade of first-order transfer functions Gi(s), i.e.,each with an impulse responsewhere H(t) is the Heaviside function, i.e., H(t) = 1 if t≥0, 0 otherwise. The lemma can now be proved by induction. Assume that has a positive impulse response yk(t). Then the impulse response yk+1(t) for satisfiesfor all t>0.
Now, since the step response of G(s) can be obtained by integration of the impulse response, it follows that if G(s) has only real poles and no zeros, its step response is monotonously increasing, which means that no overshoot may occur.
For more details on conditions for positive impulse responses, see [35].
Consider the system
Assume that it is stable and has real poles. As described above the impulse response of this system is equivalent to the step response of the model ℳm,c. The transfer function of the above system can be computed as
Note that the system has a pole at s = 0. Its impulse response equals the step response of
Since this transfer function has no zeros and real poles, its step response does not display any overshoot, according to Lemma 1. Therefore the final possibility would be that the overshoot in the models like ℳm,c would be generated by non-real poles. However, this generates damped oscillations, and this is not seen in the data. Nevertheless, to be sure that no erroneous conclusions are drawn because of this interpretation of the data, also the first models in Figure 2 have been rejected by a χ2 test.
For those models where a transfer function analysis is not sufficient for rejection, specific parameter values are needed: these are determined by parameter optimization. The resulting model is thereafter subjected to statistical tests, primarily χ2 tests. Models can also be rejected if they are biochemically unrealistic in some other way, even though they show an acceptable agreement with the data. All models that can not be rejected in any of these ways are considered as acceptable explanations of the given data set. Further details on the parameter optimization and on the statistical testing are available in the Text S1. Finally, note that even though all models except for ℳf may be analyzed using a transfer function study, this analysis gives a non-conclusive result for many more models than that, e.g., because the models may produce an overshoot, but it is unclear what its shape may be; for all those models we applied the more general optimization and statistical testing approach. |
10.1371/journal.pgen.1007273 | Ethylene induced plant stress tolerance by Enterobacter sp. SA187 is mediated by 2‐keto‐4‐methylthiobutyric acid production | Several plant species require microbial associations for survival under different biotic and abiotic stresses. In this study, we show that Enterobacter sp. SA187, a desert plant endophytic bacterium, enhances yield of the crop plant alfalfa under field conditions as well as growth of the model plant Arabidopsis thaliana in vitro, revealing a high potential of SA187 as a biological solution for improving crop production. Studying the SA187 interaction with Arabidopsis, we uncovered a number of mechanisms related to the beneficial association of SA187 with plants. SA187 colonizes both the surface and inner tissues of Arabidopsis roots and shoots. SA187 induces salt stress tolerance by production of bacterial 2-keto-4-methylthiobutyric acid (KMBA), known to be converted into ethylene. By transcriptomic, genetic and pharmacological analyses, we show that the ethylene signaling pathway, but not plant ethylene production, is required for KMBA-induced plant salt stress tolerance. These results reveal a novel molecular communication process during the beneficial microbe-induced plant stress tolerance.
| Plants as sessile organisms are facing multiple stresses during their lifetime. Among them, abiotic stresses, such as salt stress, can cause severe crop yield reduction, leading to food security issues in many regions of the world. In order to respond to growing food demands, especially in the context of the global climate change and increasing world population, it then becomes urgent to develop new strategies to yield crops more tolerant to abiotic stresses. One way to overcome these challenges is to take advantage of plant beneficial microbes, defined as plant growth promoting bacteria (PGPB). In this study, we report the beneficial effect of Enterobacter sp. SA187 on plant growth under salt stress conditions. SA187 increased the yield of the forage crop alfalfa when submitted to different saline irrigations in field trials. Moreover, using the model plant Arabidopsis thaliana, we demonstrate that SA187 mediates its beneficial activity by producing 2-keto-4-methylthiobutyric acid (KMBA), which modulates the plant ethylene signaling pathway. This study highlights a novel mechanism involved in plant-PGPB interaction, and proves that endophytic bacteria can be efficiently used to enhance yield of current crops under salt stress conditions.
| Abiotic stresses like salinity, drought or heat negatively affect plant growth and yield and belong to the most limiting factors of agriculture worldwide [1,2]. For example, salinity, known to affect almost one fourth of arable land globally, is a two-phase stress composed of a rapid osmotic and a slower toxic stress, resulting from Na+ ion accumulation and loss of K+ in photosynthetic tissues [3]. Salt stress reduces the rate of photosynthesis, leading to a decrease of plant growth and crop yield [4]. However, in the context of global climate change and an increasing world population, abiotic stress tolerant crops and sustainable solutions in agriculture are urgently needed to respond to growing food demands [5].
One way to overcome these challenges is to take advantage of plant-interacting microbes [6–8]. Indeed, plants and their rhizosphere host diverse microbial communities, selected from bulk soil [9–11], and beneficial bacteria, defined as plant growth-promoting bacteria (PGPB), can establish symbiotic associations with plants and promote their growth under optimal growth conditions or in response to biotic and abiotic stresses [12–18]. Direct plant growth-promotion mechanisms include the acquisition of nutrients by nitrogen fixation, phosphate and zinc solubilization, or siderophore production for sequestering iron. The modulation of phytohormone levels, such as auxin, ethylene, cytokinin or gibberellin, also largely contributes to the beneficial properties of PGPB [19–21]. Indirect mechanisms comprise the production of antimicrobial agents against plant pathogenic bacteria or fungi, or inducing systemic resistance against soil-borne pathogens [18,22].
Arid regions cover about one quarter of the Earth’s land surface and encompass many of the challenges for increasing agricultural productivity [23]. In contrast to better known dryland farming, desert agriculture can function only when crop plants are irrigated–usually with underground water with various levels of salinity [24]. Those areas face extreme environmental conditions, characterized by high levels of radiation, low rainfall, extreme temperatures, coarse soil which retains very little moisture, as well as low nutrients and typically high natural salinity, which all strongly limit the yield of crops [25]. Although deserts appear to be hardly inhabitable, a wide diversity of organisms has adapted to these extreme conditions. Plants along with their interacting microbial partners have evolved sophisticated mechanisms such as the production of osmoprotectants, reactive oxygen species scavengers or late embryogenesis abundant proteins to monitor the environment and reprogram their metabolism and development [26,27]. Therefore, this particular environment is an ideal reservoir to isolate and identify beneficial bacteria enhancing plant tolerance towards environmental stresses such as drought, heat or salinity [7].
To identify and characterize stress tolerance-promoting bacteria that can increase plant tolerance to abiotic stresses and therefore could be used for improving desert agriculture, we previously isolated and sequenced a number of rhizosphere and endophytic bacterial strains from nodules of desert pioneer plants [28–30]. Here, we report that Enterobacter sp. SA187, an endophytic bacterium isolated from root nodules of the indigenous desert plant Indigofera argentea [31], significantly increased yield of the agronomically important crop alfalfa (Medicago sativa) in field trials under both normal and salt stress conditions, demonstrating that SA187 has a high potential to improve agriculture under desert conditions. To better understand the molecular mechanisms for conveying enhanced stress tolerance of plants, we studied its interaction with Arabidopsis thaliana. SA187 could enhance Arabidopsis tolerance to salt stress, and GFP-labeled SA187 colonized surface and inner tissues of Arabidopsis roots and shoots. Moreover, transcriptome analyses uncovered that SA187-induced plant tolerance to salt stress is due to maintenance of photosynthesis and primary metabolism and a reduction of ABA-mediated stress responses. Using different plant hormone related mutants, ethylene sensing was found to play a primary role in SA187-induced salt stress tolerance. Indeed, Arabidopsis mutants impaired in ethylene perception were compromised in their beneficial response to SA187, while mutants deficient in ethylene synthesis remained unaffected. Gene expression analysis of SA187 indicated an upregulation of the methionine salvage pathway upon plant colonization, increasing the production of 2-keto-4-methylthiobutyric acid (KMBA), which is known to be converted into ethylene in planta [32]. KMBA alone could mimic the beneficial effects of SA187 on plant salt stress tolerance and 2,4-dinitrophenylhydrazine (DNPH), which specifically precipitates KMBA [33], could abrogate SA187-induced plant stress tolerance. These results unravel a novel communication process during beneficial plant-microbe interactions under stress conditions.
Since SA187 was an outstandingly performing bacterial isolate in a previous screen using Arabidopsis as a model plant [31], we evaluated the potential agronomic use of SA187 as a biological solution for agriculture. Therefore, we tested the beneficial activity of SA187 on different growth parameters of the crop plant alfalfa (Medicago sativa), which is largely used as animal feed in different regions of the world. Alfalfa seeds were coated with SA187 and tested in parallel with mock-coated seeds at the experimental field station Hada Al-Sham near Jeddah, Saudi Arabia. A randomized complete block design with a split-split plot arrangement with different replicates was used over two subsequent growing seasons (2015–2016 and 2016–2017). Using low saline water (EC = 3.12 dS·m-1) for irrigation, SA187-inoculated alfalfa plants showed an increase of 16 and 12% of fresh weight and 14 and 17% of dry biomass in the two growing seasons, respectively (Fig 1A). Using high saline water (EC = 7.81 dS·m-1) for irrigation, a similar beneficial impact on plant growth was observed over the two growing seasons (Fig 1B). However, the growth parameters in the second season were statistically not significant, most likely due to exceptional rainfall in that period (S1 Fig). We concluded that SA187 can efficiently improve crop productivity under extreme agricultural conditions.
To better understand the molecular mechanism by which SA187 confers stress tolerance to plants, we used the genetic model plant A. thaliana and first assessed the capacity of SA187 to affect the early stages of Arabidopsis development under normal conditions (½ MS agar medium, 22°C, 16 h of light). When compared to mock-inoculated plants, SA187 had no influence on the germination rate of Arabidopsis seeds (Fig 2A), and apart from considerably longer root hairs (Fig 2B and 2C), 5-day-old seedlings showed no morphological changes. Similarly, after transfer onto new ½ MS plates (S2 Fig), no differences between 17-day-old mock- and SA187-inoculated seedlings were recorded, when measuring root length, lateral root density, shoot morphology, or root and shoot fresh and dry weight of seedlings (Fig 2D–2F) indicating that SA187 has no significant effect on Arabidopsis development under normal growth conditions.
On the other hand, the stress tolerance and growth promoting capacity of SA187 on Arabidopsis was highlighted under salt stress. Five days after germination, SA187- and mock-inoculated seedlings were transferred onto ½ MS agar plates supplemented with 100 mM NaCl (S2 Fig), and the same growth parameters as abobe were evaluated up to 12 days after the transfer to salt plates. SA187-inoculated plants showed stress tolerance promoting activity on salt stress: the shoot and root systems of SA187-inoculated plants were significantly more developed than those of mock-inoculated plants (Fig 2E and 2F). While primary root length was similar between SA187- and mock-inoculated plants (Fig 2D), lateral root density was significantly increased (Fig 2F). Similarly to 5-day-old seedlings, SA187-inoculated plants at this stage had more than twice longer root hairs compared to the mock-inoculated ones under both normal and salt stress conditions (S3 Fig). Moreover, we proved that the beneficial activity of SA187 was largely linked to living bacterial cells as heat-inactivated SA187 cells did not induce any beneficial activity (S4A Fig). Overall, SA187 strongly enhanced Arabidopsis growth of both shoot and root under salt stress conditions, in contrast to normal conditions.
The concentration of sodium (Na+) and potassium (K+) ions in shoots is an important parameter for salt stress tolerance [34]. Therefore, the Na+ and K+ contents were determined in Arabidopsis organs in the absence and presence of SA187. Interestingly, both shoots and roots of SA187-inoculated plants accumulated similar levels of Na+ compared with mock-inoculated plants under normal and salt stress conditions (Fig 3A and 3D). However, increased K+ levels were found in SA187-inoculated plants (Fig 3B and 3E), resulting in significantly reduced shoot and root Na+/K+ ratios under saline conditions (Fig 3C and 3F), which may help the inoculated plants to keep high growth rate.
After recognition of the beneficial impact of SA187 on plant physiology, we wanted to characterize the interaction of SA187 with plants in more detail, and find whether SA187 is able to efficiently colonize Arabidopsis as its non-native host. SA187 cells were stably transformed to express GFP (SA187-GFP), which did not affect their beneficial effect on Arabidopsis seedlings (S4B Fig). Confocal microscopy revealed that SA187-GFP colonized both roots and shoots on ½ MS agar plates or in soil (Fig 4). On vertical ½ MS agar plates, the first colonies (formed by a small number of cells) were observed on the root epidermis in the elongation zone, preferentially in grooves between epidermal cell files (Fig 4A and 4B). In the differentiation zone and older root parts, colonies were larger and proportional with the age of the region (Fig 4C). A similar colonization pattern was observed in soil-grown seedlings, however, with a more random distribution of colonies (Fig 4D). SA187-GFP colonies were also often found in cavities around the base of lateral roots (Fig 4E). While it was rare to detect SA187-GFP cells inside root tissues in 5–7 days old seedlings, the apoplast of the root cortex and even of the central cylinder was regularly occupied by small scattered colonies in 3 weeks old seedlings (Fig 4F). Indeed, in our initial plant assays, SA187 could be re-isolated from surface sterilized Arabidopsis roots, indicating that SA187 was proliferating inside root tissues. Inspecting shoots, SA187-GFP colonies were found deep inside the apoplast of hypocotyls, cotyledons and the first true leaves, and in several cases, bacterial cells were directly observed to penetrate through stomata of these organs (Fig 4G–4I).
Furthermore, we evaluated colonization of root systems by SA187 (wild type strain) under normal and salt conditions. Plants were germinated on ½ MS agar plates containing SA187 wild type strains, transferred to new ½ MS plates with or without 100 mM NaCl after 5 days (S2 Fig), and parts of their root systems grown after the transfer were used for bacterial extraction after 5 more days. Interestingly, quantification based on counting of colony forming units (CFU) revealed that roots from salt conditions were twice more colonized than those from normal conditions (S5 Fig), suggesting that in our experimental system plants can probably facilitate their accessibility to colonization by beneficial bacteria under stress conditions.
To uncover how salt stress tolerance is achieved in SA187-inoculated Arabidopsis seedlings, we performed RNA-Seq analysis comparing the transcriptome of mock-inoculated to SA187-inoculated plants under non-saline (Mock, SA187), and salt stress conditions (Salt, SA187+Salt). Compared to “Mock” conditions, 545, 3113 and 1822 genes were found to be differentially expressed in the “SA187”, “Salt” and “SA187+Salt” samples, respectively (S1 Table). To obtain a global overview, the transcriptome data were organized by hierarchical clustering into 8 groups and analyzed for gene ontology enrichment (Fig 5, S2 Table).
Cluster 1 and 7 comprise the largest sets of differentially expressed genes with 1607 and 744 members, respectively, and consist of salt-stress regulated genes that were unaffected by the SA187 inoculation. Whereas Cluster 1 genes are strongly downregulated under salinity and are involved in water homeostasis, salicylic acid (SA) and defense response, those of Cluster 7 are highly upregulated and enriched in genes that are induced in response to water and salt stress or abscisic acid (ABA).
A specific effect of SA187 on the transcriptome of plants was found in Clusters 2, 3 and 4. Cluster 2 (354 genes) represents genes that are upregulated by SA187 independently of the growth conditions. This cluster is significantly enriched in plant defense genes such as chitin responsive genes but also in ethylene and jasmonic acid (JA) signaling (Fig 5). Importantly, Cluster 3 genes (246) are strongly downregulated in mock-inoculated plants under salt stress conditions but remain unaltered upon SA187-inoculation. These genes have a role in the primary metabolism, such as photosynthesis, carbon and energy metabolisms. On the contrary, Cluster 4 genes (464) are enriched in ABA and abiotic stress response and are upregulated in salt-treated plants, but not when the plants were inoculated with SA187.
In summary, these data indicate that SA187 colonization triggers in Arabidopsis the expression of genes involved in defense response as shown by the significant enrichment for chitin responsive genes and ethylene and JA signaling. Moreover, under saline conditions, SA187-inoculated plants release themselves from the impact of abiotic stress (ABA), maintain higher metabolic and photosynthetic activity, and can therefore grow better than mock-inoculated plants.
Since our transcriptome analysis indicated possible roles of several hormone pathways in the SA187-induced growth promotion under salt stress, we measured the levels of salicylic acid (SA), jasmonic acid (JA) and abscisic acid (ABA) in mock- and SA187-inoculated plants. SA187 did not significantly change plant SA levels in the absence or presence of salt (Fig 6A). Plant ABA and JA concentrations remained also unchanged upon SA187 colonization under normal conditions, but their salt-induced accumulation was significantly lower in SA187-inoculated plants (Fig 6B and 6C), indicating a partial attenuation of stress responses in these plants.
To assess the level of ethylene in Arabidopsis roots and possibly confirm the activation of the ethylene signaling pathway observed in Cluster 2, we used the ethylene-dependent pEBF2::GUS reporter [35]. In contrast to mock-inoculated seedlings, the reporter line showed strong GUS activity in root tips upon SA187-inoculation, similar to the treatment with the ethylene precursor aminocyclopropane-1-carboxylic acid (ACC) (Fig 6D), indicating the activation of the ethylene signaling pathway.
To substantiate the phytohormone quantifications, Arabidopsis hormone deficient or insensitive mutants were analyzed. The JA-receptor coi1-1 mutant [36], the JA-insensitive jar1-1 mutant [37], the ABA biosynthesis aba2-1 mutant [38] or the ABA receptor quadruple mutant pyr1-1 pyl1-1 pyl2-1 pyl4-1 (named here as pyr1/pyl) [39] maintained the SA187 beneficial activity upon salt stress, indicating that ABA or JA may not play a major role in this interaction (Fig 7A, S6 Fig).
However, the ethylene insensitive ein2-1 and ein3-1 mutants [40,41], impaired in ethylene perception, were strongly compromised in the beneficial effect of SA187, indicating that ethylene sensing could be of importance in SA187-induced tolerance of Arabidopsis to salt stress conditions. This result was confirmed by the up-regulation of the four ethylene-induced genes, ERF106, ERF018, RAV1 and SZF1, upon colonization by SA187 (Fig 7B). Moreover, application of 100 nM ACC during salt stress could largely mimic the beneficial activity of SA187 on plants (Fig 7C, S7 Fig).
In contrast, the heptuple ethylene-biosynthesis deficient mutant acs1-1 acs2-1 acs4-1 acs5-2 acs6-1 acs7-1 acs9-1 (called acs in this study) still showed full sensitivity to the beneficial activity of SA187 under salt stress (Fig 7A). Additionally, the SA187 beneficial effect was maintained when plants were treated with amino-ethoxy-vinyl glycine (AVG, 1 μM), an ethylene production inhibitor blocking ACC synthesis [42] (Fig 7D). However, when plants were treated with silver nitrate (AgNO3, 1 μM), which interferes with ethylene perception [42], SA187-inoculated plants did not exhibit any SA187-induced tolerance to salt stress (Fig 7D).
Altogether, these results indicate that the beneficial effect of SA187 may not be mediated by JA perception or the ABA pathway, but rather by the ethylene perception, as it was found to be necessary for SA187-induced salt stress tolerance on Arabidopsis plants.
The previous results suggested that ethylene most likely originates from SA187 cells rather than from the canonical plant ACC synthase (ACS) pathway. To support the hypothesis that SA187 provides ethylene to promote plant growth, we searched for bacterial genes encoding ACS or ethylene forming enzymes (EFE) [43] in the genome of SA187 [31]. No ACS- or EFE-related genes were found in SA187, which on the other hand, contains a conserved methionine salvage pathway (also known as 5’-methyl-thioadenosine cycle), and one of its components, KMBA, is known to be an ethylene precursor [44]. While SA187 alone did not produce ethylene when grown on synthetic media (S8 Fig), the expression level of most of the genes encoding proteins involved in the methionine salvage pathway were upregulated in SA187 upon plant colonization compared with bacteria incubated for 4h in liquid ½ MS with or without 100 mM NaCl in the absence of plants (Fig 8A).
To confirm that KMBA could function as an ethylene precursor during the beneficial plant-microbe interaction, we tested the effect of KMBA on Arabidopsis in comparison to SA187 inoculation. Under salt stress conditions, application of 100 nM KMBA induced a similar beneficial activity on Arabidopsis as SA187 resulting in a similar increase in both root and shoot fresh weight (Fig 8B, S7 Fig).
Finally, we took the advantage of 2,4-dinitrophenylhydrazine (DNPH), a known interactor of KMBA in vitro that was previously shown to precipitate KMBA produced by Botrytis cinerea and consequently impairs the production of ethylene by photo-oxidation [33]. Here, we could show that when plants were cultivated with 3 μM DNPH, the beneficial impact of SA187 on Arabidopsis growth under salt stress was greatly reduced from 68% to 14% (Fig 8C and 8D), showing the importance of KMBA in mediating SA187-induced plant tolerance to salt stress.
Enterobacter sp. SA187 was previously isolated from the desert pioneer plant Indigofera argentea Burm.f. (Fabaceae) [29,31]. In this work, we show that this bacterium promotes plant tolerance to salt stress, describing this strain as a stress tolerance-promoting bacterium. Indeed, under field conditions, using SA187 as an inoculum for alfalfa seeds and by monitoring growth parameters and yield over two different agriculture seasons, the inoculated plants showed a clear improvement in yield independently of the water regime applied (high or low salt stress). The data show similar effectiveness of the SA187 inoculations in both years. However, the differences for high and low-saline conditions were reduced during the second year (Fig 1), which could be explained by the increased rainfall (S1 Fig) during the 2nd growing season that may have diluted the salinity effects. We conclude that SA187 can efficiently improve crop productivity under extreme agricultural conditions and could be a simple biological solution to grow plants under extreme adverse conditions.
In order to understand the mechanisms underlying the beneficial plant interaction with SA187, Arabidopsis was used as a model system. SA187 colonizes both surface and inner tissues of Arabidopsis roots and shoots, supporting a functional plant-bacterial interaction (Fig 4). Colonization of both above- and under-ground organs is in agreement with the observation that leaf and root microbial communities share an important portion of their bacterial species [11]. While the mechanism of entry of SA187 into roots occurs most probably via cracks and/or by active penetration between epidermal cells [45], we observed that shoots were colonized through stomata, indicating that these apertures represent a major route of entry into plants not only by pathogenic but also by beneficial bacteria.
The capacity of SA187 to enhance salt stress tolerance of Arabidopsis was analyzed in detail. While SA187 induced only negligible morphological and physiological changes in plants under non-stress conditions (with the exception of longer root hairs), SA187 significantly enhanced root and shoot growth with increased fresh and dry weight under salt stress (Fig 2E and 2F). In addition, SA187 increased lateral root density, and thus the overall root surface area (Fig 2F) under salt stress. Changes in the root architecture have been considered to be beneficial for adaptation to various abiotic stress conditions including salinity [46], and very likely contribute to the SA187-induced salt tolerance in Arabidopsis.
The effect of salinity on plants includes two components: an osmotic component, being the consequence of an altered osmotic pressure due to an increased salt concentration, and a toxic ion effect as a result of the high Na+ concentration in shoots [47,48]. The toxic effects of the Na+ accumulation result in premature senescence, leading to a decrease in photosynthesis efficiency and impaired metabolic processes. Na+ also competes with K+ in membrane transport and enzymatic functions, reducing plant growth. Most plant cells possess mechanisms to counteract the harmful effects of Na+ accumulation by retaining K+ and actively excluding Na+ in roots and/or sequestering Na+ in vacuoles in shoots [47–50]. Several studies have shown that an inoculation of commercial crops, such as maize, strawberry and wheat by PGPBs under salt stress results in a decrease of Na+ and an increase of K+ in their shoots and leaves [51–53]. The inoculation of Arabidopsis thaliana and Trifolium repens (white clover) by Bacillus subtilis GB03 induced a decrease in the Na+ content in shoots in both species accompanied by an increase or no change in the K+ content [54,55]. In our study, we found no differences in Na+ contents in shoots or roots between SA187-inoculated and mock-inoculated plants in response to salt stress. However, K+ ion levels in both roots and, to a lesser extent, shoots increased upon the SA187 inoculation, resulting in reduced Na+/K+ ratios (Fig 3C and 3F), which might contribute to the higher salt tolerance of SA187-inoculated plants [56].
To analyze the interaction of SA187 with Arabidopsis at the molecular level, the transcriptome of Arabidopsis grown under salt and non-stress conditions in the absence or presence of SA187 was compared. The inoculation with SA187 dramatically reprogrammed the gene expression of plants grown either on ½ MS or on ½ MS with 100 mM NaCl. This was highlighted in Clusters 2, 3, and 4 of the RNA-Seq analysis (Fig 5). Cluster 3 genes, mostly related to photosynthesis and primary metabolism, were strongly downregulated under salt stress in mock-inoculated plants, confirming previously published reports which correlated such a downregulation with the inhibition of growth and development under salt stress conditions [57]. These results could therefore explain why SA187-inoculated plants grow better under stress conditions: SA187-inoculated plants only mildly reduce their photosynthetic capacity and maintain a functional metabolism allowing further growth in comparison to mock-inoculated plants. Cluster 4 genes are enriched in ABA-related stress genes and were induced upon salt stress in mock-inoculated plants, but not in SA187-inoculated plants. These results indicate that some salt stress-induced responses, including the enhancement of ABA levels, are dampened by SA187. However, they do not explain why plants are more salt stress tolerant. Indeed, the ABA biosynthesis aba2-1 mutant or the ABA receptor quadruple mutant pyr1-1 pyl1-1 pyl2-1 pyl4-1 still exhibited a similar growth improvement by SA187 as wild-type plants when exposed to salt stress, indicating that ABA production and signaling are dispensable in the presence of these beneficial bacteria (Fig 7A).
Induced salt stress tolerance by SA187 could be elucidated by Cluster 2, comprising genes specifically induced upon SA187-inoculation. This cluster is significantly enriched for genes involved in defense response to bacterium, and for chitin response. This latter GO term is not surprising in a plant-bacterial system, since pathogen associated molecular patterns (PAMPs) such as fungal chitin and bacterial flagellin are inducing a large set of common genes in plants, with more than 60% of overlap [58]. But the most interesting feature lies in the enrichment of the ethylene response pathway. Indeed, SA187 activates the ethylene perception pathway as shown by the qPCR analysis of ethylene-induced genes and by the ethylene reporter pEBF2::GUS (Figs 6D and 7B). Moreover, ACC and KMBA as ethylene precursors largely mimicked the beneficial effect of SA187 on plants under salt stress (Figs 7C and 8B). Finally, the involvement of ethylene was also supported by the observation of much longer root hairs (Fig 1B and 1C; S3 Fig), as ethylene plays an important role in root hair elongation [59,60]. Although the role of ethylene in plant abiotic stress tolerance is controversial [61], several pieces of evidence indicate that this phytohormone is important for plant adaptation to abiotic stresses. For example, the pre-treatment of Arabidopsis seedlings with ACC, or the use of the constitutive ethylene response (CTR1) or the EIN3 gain-of-function mutants were shown to enhance salt stress tolerance [62,63]. Furthermore, an ethylene overproduction in the eto1 mutant lead to salinity tolerance due to improved Na+/K+ homeostasis through an RBOHF-dependent regulation of Na+ accumulation [64].
Importantly, ethylene-related Arabidopsis mutants revealed that the beneficial activity of SA187 is to a major extent mediated via the perception of externally produced ethylene. Although the ein2-1 and ein3-1 mutants were compromised in their beneficial response to SA187, the disruption of the plant ethylene production in the heptuple acs mutant showed the same growth enhancement under salt stress when comparing SA187-inoculated plants to mock-inoculated plants (Fig 7A). This was supported by a parallel pharmacological approach, demonstrating that inhibition of the ACS activity using AVG did not block the stress tolerance promoting activity of SA187, while blocking the ethylene receptors by AgNO3 compromised the beneficial activity of SA187 on plants under salt stress (Fig 7D).
As plants were shown to perceive ethylene even without functional plant ethylene production, we suspect that SA187 could provide plants with ethylene or its precursor. Three main pathways for ethylene biosynthesis have been described in bacteria and other microbes. The mold Dictyostelium mucoroides and fungi Penicillium citrinum produce ethylene from methionine via S-adenosyl-methionine, through the sequential action of ACC synthases and ACC oxidases. S-adenosyl-methionine is first converted to ACC by ACC synthases, which is then oxidized by ACC oxidases to release ethylene and cyanide. The same pathway is well known to be responsible for ethylene biosynthesis in plants, where cyanide is converted to β-cyanoalanine to avoid toxicity [44,65]. Microbes can also produce ethylene from α-ketoglutarate and arginine by the action of the ethylene forming enzyme (EFE), which has been found in several microbial species such as Pseudomonas syringae and Penicillium digitatum [44,66]. A third pathway has been identified in a variety of bacteria such as Escherichia coli and Cryptococcus albidus, or in fungi like truffle or pathogenic Botrytis cinerea, where ethylene is produced via oxidation of KMBA, an intermediate of the methionine salvage pathway [32,33,44,67]. KMBA can be spontaneously converted to ethylene by photo-oxidation or through the action of peroxidases [33], which are abundantly present in the plant apoplast [68,69].
Based on P-BLAST homology searches, genome analysis of SA187 revealed that neither ACC synthase nor EFE genes are present in SA187. Instead, SA187 contains the entire methionine salvage pathway, suggesting that KMBA is most likely the precursor of ethylene in SA187. Interestingly, most of the methionine salvage pathway genes in SA187 are only actively expressed upon colonization of Arabidopsis (Fig 8A). Moreover, the application of KMBA could mimic the beneficial effect of SA187 on plants when subjected to salt stress (Fig 8B). Importantly, the SA187 beneficial activity towards plant was highly reduced when treated with DNPH, known to provoke KMBA precipitation and prevent thus its oxidation and ethylene release [33].
Taken together, the KMBA involvement in abiotic stress tolerance constitutes a novel mechanism in the field of plant-beneficial bacteria interaction. While the induction of the ethylene signaling pathway by PGPB has been reported in several studies to play an important role in the induced systemic resistance in plants [18,70,71], PGPB activity in the context of abiotic stress has been commonly attributed rather to a reduction of the plant ethylene level through the activity of bacterial ACC deaminases [53,72–74], or shown to be independent of the ethylene signaling pathway [75,76]. Several reports hypothesized the involvement of ethylene signaling in abiotic stress tolerance induced by rhizosphere bacteria, with evidences that were largely based on emissions of unidentified volatiles or by comparison with plant-fungal interactions [32,77,78]. Recently, it has been reported that the beneficial bacterium Burkholderia phytofirmans PsJN enhanced plant growth through an auxin/ethylene-dependent signaling pathway under optimal conditions, but in contrast to the present study, the authors hypothesized that the plant intrinsic ethylene production was fundamental in that interaction [79].
In conclusion, we provide evidence that the endophytic bacterium Enterobacter sp. SA187 induces salt stress tolerance in Arabidopsis via production of KMBA to activate the ethylene pathway. SA187 enhances plant salt stress tolerance under controlled conditions in the model plant Arabidopsis thaliana and under field conditions in the crop plant alfalfa. These results show the potential use of SA187 for bringing saline agriculture of current crops a step closer to reality.
To inoculate alfalfa (Medicago sativa var. CUF 101) seeds, a slurry was prepared consisting of sterilized peat, a broth culture of SA187, and sterilized sugar solution (10%) in the ratio 5:4:1 (w/v/v). Subsequently, alfalfa seeds were coated with the slurry at a rate of 50 mL·kg-1. As a control, seeds were coated with a similar mixture without bacteria. Field trial was conducted at the experimental station in Hada Al-Sham (N 21°47'47.1" E 39°43'48.8"), Saudi Arabia, in winter seasons 2015–2016 and 2016–2017. The experiment was a randomized complete block design with a split-split plot arrangement of four replicates in the for season 2015–2016 season and three replicates in the 2016–2017 season, plots (2 × 1.5 m) with seed spacing 20 cm row-to-row. The field was irrigated using groundwater with two different salinity levels: low salinity (EC = 3.12 dS·m-1), and high salinity (EC = 7.81 dS·m-1). The soil had an average pH 7.74 and salinity EC = 1.95 dS·m-1. Agronomical data (plant height, fresh biomass, and dry biomass) were recorded every 25–30 days from each harvest; three harvests were done in the first season, four harvests in the second season. Field trials data were analyzed as a randomized complete block design using a Factorial ANOVA Model, followed by least significant difference (LSD) test for pairwise comparisons. Results with a p-value < 0.05 were considered significant. All statistical analysis was carried out using SAS/STAT software (https://www.sas.com/).
Enterobacter sp. SA187 was previously isolated from root nodules of the leguminous pioneer plant Indigofera argentea in the Jizan region of Saudi Arabia [29,31]. Arabidopsis seeds were obtained from publicly available collections. The following mutant lines used in this study were published previously: the JA-receptor coi1-1 mutant [36], JA-insensitive jar1-1 [37], the ABA biosynthesis aba2-1 mutant [38], the ABA receptor quadruple pyr1-1pyl1-1pyl2-1pyl4-1 mutant [39], the ethylene insensitive ein2-1 [40] and ein3-1 mutants [41], the heptuple ethylene-biosynthesis deficient mutant acs1-1acs2-1acs4-1acs5-2acs6-1acs7-1acs9-1 [80], and the ethylene-dependent pEBF2::GUS reporter [35].
Prior to every experiment, A. thaliana seeds were surface sterilized 10 min in 70% ethanol + 0.05% sodium dodecyl sulfate on a shaker, washed 2 times in 96% ethanol and let to dry. To ensure SA187-inoculation, sterilized seeds were sown on ½ MS plates (Murashige and Skoog basal salts, Sigma) containing SA187 (2·105 cfu·ml-1), stratified for 2 days at 4°C in the dark and then placed vertically to growth conditions for 5 days as shown as in S2 Fig. The ½ MS plates with SA187 were prepared by addition of 107 bacteria to 50 ml pre-cooled agar medium during plate preparation.
Average length of root hairs was determined based on images of 5-day-old roots (1 image per root at constant distance from the root tip, 25 seedlings per condition) or 16-day-old roots (along the whole primary root length grown after transfer) captured by a Nikon AZ100M microscope equipped with an AZ Plan Apo 2x objective and a DS-Ri1 camera (Nikon). All root hairs in focus were measured using ImageJ (https://imagej.nih.gov/ij/). Average values and standard deviations were calculated from 10% longest root hairs to eliminate non-developed root hairs and describe the maximal elongation capacity of root hairs.
For salt stress tolerance assays, 5-day-old seedlings were transferred onto ½ MS plates with or without 100 mM NaCl (Sigma). Primary root length was measured every 2 days using ImageJ software after scanning the plates. Lateral root density was evaluated as detectable number of lateral roots under a stereo microscope divided by the primary root length. Fresh weight of shoots and roots was measured 12 days after transfer of seedlings. Dry weight was measured after drying shoot and shoots for 2 days at 70°C. Following Koch’s postulate, SA187 was re-isolated from Arabidopsis root system at the end of an initial experiment to confirm the genotype of the inoculated strain. To address the ethylene involvement in Arabidopsis adaptation to salt stress, ACC (1-aminocyclopropane-1-carboxylic acid, Sigma), KMBA (2-keto-4-methylthiobutyric acid, Sigma), AVG (aminoethoxyvinylglycine, Sigma), AgNO3 (silver nitrate, Sigma) were added into pre-cooled ½ MS agar medium together with 100 mM NaCl. For DNPH (2,4-dinitrophenylhydrazine, Sigma), 5 mM solution was prepared by solubilizing DNPH into 2M HCl (hydrochloric acid, Sigma) as described previously [81], then the solution was diluted until reaching 1 mM, and equilibrated to the same pH as MS medium (pH 5.8) using 2M KOH (potassium hydroxide, Sigma). DNPH was used at final concentration 3 μM.
All plants were grown in long day conditions in growth chambers (Percival; 16 h light / 8 h dark, 22°C). Each experiment was performed at least in three biological replicates.
Dry rosettes and root systems were weighted. All samples were measured individually except for salt-treated root systems, whereby pools of three root systems were measured to ensure proper weight measurements. Sodium and potassium concentrations were prepared for shoot and root dry samples by adding 1 mL of freshly prepared 1% HNO3 (nitric acid, Fisher Scientific) to the pre-weighed samples. The concentrations of sodium and potassium were determined, using Inductively Coupled Plasma Optical Emission Spectrometer (Varian 720-ES ICP OES, Australia).
SA187 was genetically labeled with the GFP expressing cassette by taking advantage of the mini-Tn7 transposon system [82]. In order to specifically select for a bacterium carrying the GFP integration in the genome, a spontaneous rifampicin resistant mutant of the strain was obtained first [83]: an overnight-grown culture of SA187 was plated on LB plates supplemented with 100 μg·mL-1 of rifampicin, and the plates were incubated for 24 h at 28°C. At least 10 colonies, representing spontaneous rifampicin resistant (RifR) mutants of the strain were streaked twice on LB plates containing 100 μg·mL-1 of rifampicin and thereafter twice on LB plates supplemented with 200 μg·mL-1 of rifampicin. The GFP expressing cassette was introduced in the SA187 RifR strain by conjugation as described in Lambertsen et al. (2004) [84]. Briefly, 1010 cells of SA187 RifR strain were mixed with 109 cells of E. coli SM10λpir harboring the helper plasmid pUX-BF13, the GFP donor (a mini-Tn7) plasmid and mobilizer pRK600 plasmid. The mixed culture was incubated on sterile nitrocellulose filter for 16hrs. The conjugation culture of bacterial cells was resuspended in saline buffer (9 g/L NaCl) and spread on selective media with a propitiate antibiotics to select transformed SA187. The selected colonies were screened by fluorescence microscopy for GFP fluorescence and positive colonies were further subjected to genotype confirmation by 16S rRNA gene sequencing.
GFP-labeled SA187 on Arabidopsis roots was imaged using an inverted Zeiss LSM 710 confocal microscope equipped with Plan-Apochromat 10x/0.45, Plan-Apochromat 20x/0.8, and Plan-Apochromat 40x/1.4 Oil objectives. Seedlings grown for 3–21 days on vertical ½ MS agar plates or in soil inoculated with SA187-GFP were washed gently in sterile distilled water and transferred on a sterile agar plate. A block of agar with several seedlings was immediately cut out and placed upside-down to a chambered cover glass (Lab-Tek II) with 30 μM propidium iodide (PI) in water as mounting medium. The GFP and PI fluorescence was excited using the 488nm laser line, and captured as a single track (emission of 493–537 nm for the GFP channel, 579–628 nm for the PI channel, 645–708 nm for chloroplast autofluorescence). For 3D reconstructions, 1 μm-step Z-stacks were taken, and images were generated in the integral 3D view of the Zen software (Zeiss).
Col-0 seedlings were germinated on SA187-inoculated ½ MS agar plates and transferred to new ½ MS plates with or without 100 mM NaCl 5 days after germination (10 seedlings per plate). Parts of their root systems grown after the transfer were cut, gently washed by dipping in distilled water to remove non-attached bacterial cells, and then ground in Eppendorf tubes using Teflon sticks. Each sample was resuspended in 1 ml of extraction buffer (10 mM MgCl2, 0.01% Silwet L-77), sonicated for 1 min and subsequently vortexed for 10 min. Samples were diluted 10-fold, and then spread on LB agar plates, and colony forming units (CFUs) were counted after overnight incubation at 28°C. Calculated number of CFUs was normalized per centimeter of root length (total root length was determined based on images of root systems before their harvest). The experiment was conducted in three biological replicates, each with three technical replicates per condition; each sample consisted of five roots.
Total RNA was extracted from 5-day-old plants either or not inoculated with SA187 and transferred for 10 more days on ½ MS plates with or without 100 mM NaCl using the Nucleospin RNA plant kit (Macherey-Nagel), including DNaseI treatment, and following manufacturer’s recommendations.
RNA samples were analyzed by Illumina HiSeq deep sequencing (Illumina HiSeq 2000, Illumina). Three biological replicates were processed for each sample. Paired-end sequencing of RNA-Seq samples was performed using Illumina GAIIx with a read length of 100 bp. Reads were quality-controlled using FASTQC (http://www.bioinformatics.babraham.ac.uk/projects/fastqc/). Trimmomatic was used for quality trimming [85]. Parameters for read quality filtering were set as follows: Minimum length of 36 bp; Mean Phred quality score greater than 30; Leading and trailing bases removal with base quality below 3; Sliding window of 4:15. TopHat v2.0.9 [86] was used for alignment of short reads to the A. thaliana genome TAIR10, Cufflinks v2.2.0 [87] for transcript assembly and differential expression. To identify differentially expressed genes, specific parameters (p-value: 0.05; statistical correction: Benjamini Hochberg; FDR: 0.05) in cuffdiff were used. Post-processing and visualization of differential expression were done using cummeRbund v2.0.0 [88]. Gene was considered as regulated if fold change > log2|0.6| and q-value < 0.05 compared to Mock condition. RNA-Seq data set can be retrieved under NCBI geo submission ID GSE102950.
For qPCR analysis, mock and SA187-inoculated plants were used for RNA extraction as described above. Samples were used for analysis of either plant or SA187 gene expression. For bacteria alone, SA187 incubated for 4h in liquid ½ MS or ½ MS with 100 mM NaCl at 28°C and dark were used for RNA extraction, using the RiboPure RNA Purification Kit (Ambion), following manual instructions for Gram-negative bacteria, with the exception that no beads were added during bacterial lysis. RNA extraction was followed by DNAseI treatment.
cDNAs were using SuperscriptIII (Invitrogen): 1 μg of total RNA, oligo-dT as a primer, following manufacturer’s recommendations. For Arabidopsis gene expression analyses, ACTIN2 (At3g18780) and UBIQUITIN10 (At4g05320) were used as reference genes. For SA187 gene expression analyses, infB, rpoB and gyrB were used as reference genes. All reactions were done in a CFX96 Touch Real-Time PCR Detection System (BIO-RAD) as follows: 50°C for 2 min, 95°C for 10 min; 40× [95°C for 10 sec and 60°C for 40 sec]; and a dissociation step to validate PCR products. All reactions were performed in three biological replicates, and each reaction as a technical triplicate. Gene expression levels were calculated using the Bio-Rad CFX manager software. Primer sequences used in this analysis are listed in S3 Table.
Arabidopsis regulated genes were used to generate HCL tree using Multi Experiment Viewer (MeV 4.9.0 version, TM4, https://sourceforge.net/projects/mev-tm4/files/mev-tm4/MeV%204.9.0/). Raw data were normalized for every gene. Hierarchical clustering was performed using Euclidian distances, average linkage and leaf order optimization.
Gene enrichment analyses were performed using AmiGO website (http://amigo1.geneontology.org/cgi-bin/amigo/term_enrichment). All clusters were analyzed using default parameter (S2 Table).
For each sample, 10 mg of freeze-dried powder were extracted with 0.8 mL of acetone/water/acetic acid (80/19/1 v:v:v). For each sample, 2 ng of each standard was added to the sample: abscisic acid, salicylic acid, jasmonic acid, and indole-3-acetic acid stable labeled isotopes used as internal standards were prepared as described previously [89]. The extract was vigorously shaken for 1 min, sonicated for 1 min at 25 Hz, shaken for 10 minutes at 4°C in a Thermomixer (Eppendorf), and then centrifuged (8000 g, 4°C, 10 min). The supernatants were collected, and the pellets were re-extracted twice with 0.4 mL of the same extraction solution, then vigorously shaken (1 min) and sonicated (1 min; 25 Hz). After the centrifugations, three supernatants were pooled and dried.
Each dry extract was dissolved in 140 μL of acetonitrile/water (50/50; v/v), filtered, and analyzed using a Waters Acquity ultra performance liquid chromatograph coupled to a Waters Xevo Triple quadrupole mass spectrometer TQS (UPLC-ESI-MS/MS). The compounds were separated on a reverse-phase column (Uptisphere C18 UP3HDO, 100 × 2.1 mm, 3 μm particle size; Interchim, France) using a flow rate of 0.4 mL·min-1 and a binary gradient: (A) acetic acid 0.1% in water (v/v) and (B) acetonitrile with 0.1% acetic acid. For ABA, salicylic acid, jasmonic acid, the following binary gradients were used (time, % A): (0 min, 98%), (3 min, 70%), (7.5 min, 50%), (8.5 min, 5%), (9.6 min, 0%), (13.2 min, 98%), (15.7 min, 98%), and the column temperature was 40°C. Mass spectrometry was conducted in electrospray and multiple reaction monitoring scanning mode (MRM mode), in the negative ion mode. Relevant instrumental parameters were set as follows: capillary 1.5 kV (negative mode), source block and desolvation gas temperatures 130°C and 500°C, respectively. Nitrogen was used to assist the cone and desolvation (150 L·h-1 and 800 L·h-1, respectively), argon was used as the collision gas at a flow of 0.18 mL·min-1. Samples were reconstituted in 140 μL of 50/50 acetonitrile/H2O (v/v) per mL of injected volume. The limit of detection (LOD) and limit of quantification (LOQ) were extrapolated for each hormone from calibration curves and samples using Quantify module of MassLynx software, version 4.1.
Seedlings were vacuum infiltrated with the pre-fixation buffer [0.3% formaldehyde, 0.28% mannitol, 50 mM sodium phosphate buffer (pH 7.2)], washed with phosphate buffer and incubated in staining solution [250 μM K3Fe(CN)6 (potassium ferricyanide), 250 μM K4Fe(CN)6 (potassium ferrocyanide), 2% Triton-X, 1 mM 5-bromo-4-chloro-3-indolyl-b-D-glucuronic acid (X-GlcA; Duchefa), 50 mM sodium phosphate buffer (pH 7.2)]. Tissue was cleared with Visokol (Phytosys) overnight and observed with Axio Imager 2 (Zeiss) equipped with Plan-Neofluar 10x/0.45 objective.
A fresh SA187 culture was prepared by inoculation of 50 mL of liquid LB medium with 1 mL of overnight-grown culture. Subsequently, 2 mL of fresh culture was transferred to 10 mL chromatography vials and sealed with a rubber plug and snap-cap (Chromacol) after 0, 1, 2 or 4 hours of growth on a shaker incubator (220 rpm, 28°C). The sealed vials were again transferred to the shaker incubator for another 2 hours to allow ethylene accumulation. Three biological replicates were prepared at each time point along with 3 controls to correct for background ethylene emanation. Ethylene emission was measured with a laser-based photo-acoustic detector (ETD-300 ethylene detector, Sensor Sense, The Netherlands) [90]. Immediately after the ethylene measurement, OD600 was determined with Implen NanoPhotometer NP80 (Sopachem Life Sciences, Belgium) to correct for the total amount of bacterial cells present in the samples.
RNA-Seq data are available under the ID GSE102950 (http://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE102950)
|
10.1371/journal.pbio.1000414 | The Role of Nucleosome Positioning in the Evolution of Gene Regulation | Chromatin organization plays a major role in gene regulation and can affect the function and evolution of new transcriptional programs. However, it can be difficult to decipher the basis of changes in chromatin organization and their functional effect on gene expression. Here, we present a large-scale comparative genomic analysis of the relationship between chromatin organization and gene expression, by measuring mRNA abundance and nucleosome positions genome-wide in 12 Hemiascomycota yeast species. We found substantial conservation of global and functional chromatin organization in all species, including prominent nucleosome-free regions (NFRs) at gene promoters, and distinct chromatin architecture in growth and stress genes. Chromatin organization has also substantially diverged in both global quantitative features, such as spacing between adjacent nucleosomes, and in functional groups of genes. Expression levels, intrinsic anti-nucleosomal sequences, and trans-acting chromatin modifiers all play important, complementary, and evolvable roles in determining NFRs. We identify five mechanisms that couple chromatin organization to evolution of gene regulation and have contributed to the evolution of respiro-fermentation and other key systems, including (1) compensatory evolution of alternative modifiers associated with conserved chromatin organization, (2) a gradual transition from constitutive to trans-regulated NFRs, (3) a loss of intrinsic anti-nucleosomal sequences accompanying changes in chromatin organization and gene expression, (4) re-positioning of motifs from NFRs to nucleosome-occluded regions, and (5) the expanded use of NFRs by paralogous activator-repressor pairs. Our study sheds light on the molecular basis of chromatin organization, and on the role of chromatin organization in the evolution of gene regulation.
| Divergence in gene regulation plays a major role in organismal evolution. Evidence suggests that changes in the packaging of eukaryotic genomes into chromatin can underlie the evolution of divergent gene expression patterns. Here, we explore the role of chromatin structure in regulatory evolution by whole-genome measurements of nucleosome positions and mRNA levels in 12 yeast species spanning ∼250 million years of evolution. We find several distinct ways in which changes in chromatin structure are associated with changes in gene expression. These include changes in promoter accessibility, changes in promoter chromatin architecture, and changes in the accessibility of specific transcription factor binding sites. In many cases, changes in chromatin architecture are coupled to physiological diversity, including the evolution of a respiration- or fermentation-based lifestyle, mating behavior, salt tolerance, and broad aspects of genomic structure. Together, our data will provide a rich resource for future investigations into the interplay between chromatin structure, gene regulation, and evolution.
| Regulatory differences affecting gene expression can play a major role in species evolution [1] and can help elucidate the functional mechanisms that control gene regulation [2],[3]. Although specific examples of regulatory divergence are known in bacteria [4], fungi [5],, flies [9], and mammals [10], a general understanding of the evolution of gene regulation is still lacking. The recent availability of many sequenced genomes and accessibility of genomic profiling approaches open the way for comparisons of gene regulation across multiple species.
Among eukaryotes, the Hemiascomycota yeasts (Figure 1A), which span over ∼250 million years of evolution, are particularly suitable for studying evolution of gene regulation. This is due to the genetic tractability of yeasts, the wealth of knowledge about the model organism Saccharomyces cerevisiae, the large number of sequenced genomes, and the diversity of yeast lifestyles [3]. Notably, Hemiascomycota yeasts diverged before and after a whole genome duplication event (WGD, Figure 1A) [11], which marked a shift from using respiration for energy production in pre-WGD species to primarily using fermentation in post-WGD species [12].
Nucleosomes modulate eukaryotic gene regulation by affecting the accessibility of other proteins to the DNA, which can impact gene activation and repression [13]. In particular, many genes have nucleosome-depleted “Nucleosome Free Regions” (NFRs) in their proximal promoters (Figure 1B, top), providing access to sequence specific transcription factors (TFs) and to the basal transcription machinery [14],[15],[16],[17]. Three major determinants have been proposed to impact nucleosome depletion at NFRs: (1) active transcription by RNA polymerase II results in eviction of the −1 nucleosome [18],[19], (2) intrinsic “anti-nucleosomal” DNA sequences such as Poly(dA:dT) bind histones with low affinity and can “program” NFRs constitutively [20],[21],[22],[23],[24], and (3) trans-acting proteins can move nucleosomes away from their thermodynamically preferred locations [25],[26].
Recent studies in yeast suggest a broad role for chromatin organization in regulatory evolution. Most regulatory divergence between closely related S. cerevisiae strains is associated with divergence in unlinked (trans) chromatin remodelers [27],[28]. Conversely, many transcriptional differences between S. cerevisiae and S. paradoxus (Last Common Ancestor (LCA) ∼2 million years ago (MYA)) are due to linked cis polymorphisms predicted to affect nucleosome occupancy [29],[30]. Furthermore, a recent study suggested that changes in the regulation of mitochondrial ribosomal protein (mRP) genes between the distant species C. albicans and S. cerevisiae (LCA ∼200 MYA) were associated with a change in nucleosome organization [31],[32]. In particular, the higher expression of mitochondrial genes in respiratory C. albicans is accompanied by enrichment for the PolyA-like “RGE” binding site in the mRP gene promoters [31], which appears to “program” the constitutive presence of wider, more open NFRs at these genes [32]. All of these are absent from the promoters of mRPs in the fermentative S. cerevisiae. Finally, a recent study [33] compared genome-wide nucleosome positioning in S. cerevisiae and S. pombe (LCA ∼300M–1 BYa), finding changes in global nucleosome spacing and in the apparent sequences that intrinsically contribute to nucleosome positioning in vivo.
While these examples are intriguing, they are limited in their phylogenetic coverage (a pair of species) or their functional scope (one regulon). Thus, we understand little about the evolutionary interplay between gene expression, regulatory sequence elements, and chromatin organization. How does chromatin organization change over evolutionary time scales? Are the mechanisms underlying chromatin packaging of functional gene modules conserved? If not, how do they evolve and what is the role of different factors in this divergence? Are changes in chromatin organization related to changes in gene regulation? Can phylogenetic comparisons shed light on the distinct mechanisms that help establish chromatin organization?
Here, we present the first large-scale experimental and computational study of chromatin organization across a eukaryotic phylogeny. We measured genome-wide nucleosome locations and mRNA abundance in 12 Hemiascomycota yeast species, spanning over 250 million years of evolution (Figure 1A). We developed an analysis framework that integrates the experimental data with genome sequences, functional gene sets, and TF binding sites across the 12 species.
Our analysis uncovers several major principles that govern the evolutionary and functional relationship between chromatin organization and gene regulation in this phylogeny. (1) While qualitative features of chromatin organization are conserved in all species, quantitative features such as nucleosome packing, NFR length, and NFR to ATG distance have substantially diverged; (2) promoter chromatin organization and gene expression levels of “growth” and “stress” genes follow distinct patterns, and this dichotomy is conserved in all species; (3) evolutionary divergence in gene expression is often accompanied by transition of chromatin organization from a “growth” to a “stress” pattern; (4) changes in transcription levels, gain/loss of anti-nucleosomal sequences, and gain/loss of binding sites for “general regulatory factors” (GRFs) all play substantial and complementary roles in divergence of chromatin organization; (5) the loss of anti-nucleosomal sequences and parallel gain of binding sites for GRFs drive shifts from intrinsic to trans-regulated chromatin organization; (6) regulatory divergence can also occur by re-positioning of binding sites relative to nucleosome positions or by expanding the use of accessible sites by paralogous TFs. These mechanisms played a role in the evolution of respiro-fermentation, as well as in the evolution of regulation of other key regulons at different phylogenetic points, including mating, meiosis, RNA polymerase subunits, proteasomal, and splicing genes. Together, they uncover novel insights into the general roles for chromatin in regulating genomic access and in the evolution of regulatory programs, and provide a rich resource for future investigation.
We mapped nucleosome positions genome-wide in 12 Hemiascomycota species (Figure 1A) [34] by Illumina sequencing of mononucleosomal DNA [19],[21],[35] isolated from mid-log cultures (Materials and Methods, Figures 1A and S1). To minimize condition- and stress-related differences, we grew all species in the same rich medium, where the growth rate of each species was at least ∼80% of its maximal measured rate in any of over 40 tested media formulations. In order to compare our data to transcriptional output, we also used species-specific microarrays to measure mRNA abundance in all species in the same mid-log cultures used for nucleosome mapping (Table S2, Materials and Methods).
Aligning nucleosome reads to each genome and averaging over all genes showed remarkably similar profiles in all species studied (Figures 1A, S2, S3). All gene-averaged profiles are dominated by a pronounced depression upstream of the ATG that corresponds to the NFR [14],[15],[16],[17],[36]. To quantitatively compare chromatin structure between various genes, we first called nucleosome positions, identified 5′ and 3′ NFRs, and measured a number of nonredundant features that describe the chromatin organization at each gene (Materials and Methods, Figures 1B and S4). Below, we will study each feature at three levels: (1) globally, averaged across all genes in a genome; (2) functionally, averaged across all genes in a functional category; and (3) locally, at a single gene.
Several qualitative chromatin features have previously been identified in all eukaryotes studied [14], and these are conserved across all 12 species (Figures 1A, S2, and S3). These include an abundant 5′NFR, a common 3′NFR, a well-positioned +1 nucleosome (Nuc+1), and increasing nucleosome fuzziness over the body of genes (Figures S2 and S3, Table S3), which is consistent with statistical positioning of nucleosomes [23],[37],[38].
In contrast, quantitative global features were often variable between species (Figures 1C–F and S5, Table S3). Our measurements recapitulated previous predictions or bulk assays in the few cases where these were available, thus validating our dataset and analytical methods. For example, nucleosome spacing in coding regions was variable between species (Figure 1C,D), consistent with observed nucleosome laddering on gels [39],[40]. This leads to variation in the specific coding sequences exposed in linker DNA and could affect patterns of sequence variation [41],[42],[43] and higher-order packaging into the 30 nm fiber [44]. The distance between the NFR and a gene's start codon (Figures 1E,F and S5) is also variable between species, consistent with prior computational predictions [45].
Other evolutionary variations in global features were not previously described, showing that additional major aspects of chromatin architecture can substantially diverge. Most notably, the median NFR width was highly variable between species (Table S3), ranging from 109 to 155 nucleotides. This likely reflects the variation in the length and abundance of anti-nucleosomal Poly(dA:dT) tracts between species (discussed below). Shorter NFRs may constrain regulatory information into more compact promoters.
We next explored possible functional implications of chromatin organization in specific sets of genes with related function. Prior studies in S. cerevisiae and C. albicans have shown that in both species, “growth” genes, defined by their co-expression with cytoplasmic ribosomal proteins (cRPs), have a more open chromatin organization on average [32]. Conversely, “stress” genes, whose expression is anti-correlated to that of growth genes, have a more closed chromatin organization in both species.
To assess the generality of this observation, and identify additional trends, we tested in each species thousands of functional gene sets for enrichment of each of 22 distinct chromatin parameters. We used gene orthology [34] to project functional gene sets defined in S. cerevisiae across species (Materials and Methods). For a given gene set in each species we calculated whether its constituent genes tended to have high or low values of each of the chromatin features (Figure 1B), relative to the background of that feature's overall distribution in that species (Kolmogorov-Smirnov (K-S) test, Figure 2A,B). This provides a comprehensive overview of chromatin organization at 5′ promoters and 3′ ends for each functional gene set across the 12 species (Figure 2C–J, middle panels, Figures S6 and S7, and Tables S4–S5). In order to compare chromatin changes to gene expression levels, we also calculated the enrichment for high or low mRNA expression in all gene sets for each species (K-S test, Figure 2C–J, left panels).
We confirm a strong dichotomy in the promoter chromatin architecture of most “stress” and “growth” genes in S. cerevisiae [19],[46],[47],[48],[49] and C. albicans [32] and find that it is conserved across all 12 species (Figures 2C,D and S6–S8). Promoters of “growth” genes (e.g., ribosomal, proteasomal, and nuclear pore proteins, Figure 2C,E,G) exhibit long and deep (low occupancy) 5′NFRs. Conversely, those of “stress” genes (e.g., toxin-response genes, integral membrane proteins, Figure 2D) exhibit a more variable chromatin architecture, with shallower (higher occupancy) and narrower 5′NFRs. A host of other chromatin features also distinguish between the two functional groups (Figure S6). Thus, the separation of the “growth” and “stress” axes is a hallmark of Hemiascomycota gene regulation [2],[3] and imposes strong constraints at all levels from evolution of gene content [34] to chromatin organization. There are, however, several exceptions to this rule. Most notably, several key “growth” genes, including glycolysis genes and endoplasmic reticulum genes, are highly expressed, yet do not exhibit deep NFRs in any species (Figure 2F).
We identify a range of additional conserved patterns of chromatin architecture associated with other specific functions, which were not previously reported. For example, a number of gene sets (e.g., reproduction, cell wall, inositol phosphate, benzoate, and nicotinamide metabolism genes) have conserved long NFR to ATG distances (Figure S6), but have few other hallmarks of stress genes, and are expressed at average levels. In S. cerevisiae, these genes have long 5′ untranslated regions (5′UTRs) [50], suggesting that relatively long 5′UTRs are conserved at their orthologs in all 12 species. This may indicate a conserved role for translational control in the regulation of these functions [51].
On this backdrop of conservation, we find that coordinated changes have occurred in chromatin organization of specific functional gene sets, consistent with major phenotypic changes. Most notably, respiration and mitochondrial genes have switched from a “growth”-like chromatin pattern in pre-WGD species (where they are highly expressed) to a more “stress”-like pattern post-WGD (Figures 2H and S6). We confirm the previously reported change between S. cerevisiae and C. albicans for genes involved in respiratory metabolism [32]. We further extend these results across the full phylogenetic scope and to several other gene sets of related function (Figures 2H and S6). This change corresponds to a major change in lifestyle from respiration to respiro-fermentation after the WGD [12],[31],[32],[52]. We also discover the converse evolutionary pattern (Figure 2I): a number of gene sets involved in cytoskeletal organization are packaged into deeper NFRs in post-WGD species than in pre-WGD species. Surprisingly, the expression level of these genes has not substantially changed with this transition.
Changes in chromatin organization have also occurred at other phylogenetic points of phenotypic evolution, suggesting a general evolutionary mechanism. For example, we discovered that in Yarrowia lipolytica spliceosome genes are associated with long and deep NFRs, but in all other species they are enriched for short and shallow NFRs (Figure 2J, middle panel). This switch between deep and shallow NFRs is accompanied by a decrease in expression of these genes (Figure 2J, left panel) and is consistent with the much larger number of introns in Yarrowia lipolytica genes [53] and with the loss of introns and reduction of splicing in the subsequently diverged species.
We next asked what mechanisms contribute to conservation and variation in chromatin organization across species. Three determinants have been previously implicated in establishing NFRs in S. cerevisiae [14]: (1) the expression level of the gene, as RNA polymerase recruitment affects NFR width; (2) the presence of intrinsic anti-nucleosomal sequences such as Poly(dA:dT) tracts in the gene's promoter; and (3) the binding of proteins such as chromatin remodelers that actively evict or move nucleosomes. We first consider these three determinants independently, and then assess their relative contributions.
In some cases, variation in chromatin organization in a gene set, both within and between species, correlates with gene expression level. Within each species, many highly expressed “growth” genes (e.g., RP genes) are packaged with wide and deep NFRs, while many poorly expressed stress genes have shorter, occupied NFRs (Figures 2C,D, S6). Between species, evolutionary shifts from high to low expression levels were sometimes accompanied by corresponding changes in chromatin organization (e.g., mitochondrial RP and splicing genes, Figure 2H,J).
However, transcription level is insufficient to solely explain the NFR occupancy measured across the 12 species. Globally, expression level alone explains only 1.7%–13.1% of the variation in NFR occupancy in each of the 12 species (Lowess fit, Figure S9A,C,E, Materials and Methods). Furthermore, when we use Lowess subtraction to correct for the relationship between mRNA level and each chromatin feature, the enrichments of most gene sets for high or low values of chromatin features were maintained (Figure S10, Materials and Methods). Within species, the discrepancy is prominent in some of the gene sets (e.g., glycolysis, gluconeogenesis) that are highly expressed in all species but do not exhibit the expected deep NFRs (Figure 2F). Between species, cytoskeleton and nuclease-related gene sets have shifted from shallow to deep NFRs at the WGD, often without a concomittant change in expression levels (Figure 2I). The failure of transcript levels to fully explain NFR width and depth is consistent with recent experimental results in S. cerevisiae, where the distinctive chromatin organization of growth and stress genes was largely maintained even after genetically inactivating RNA Pol II [19].
We next tested an alternative hypothesis that chromatin organization at the NFR is determined by intrinsic “anti-nucleosomal” sequences with low affinity for the histone octamer, such as Poly(dA:dT) tracts [20],[21],[22],[24],[54],[55]. We estimated the average extent of nucleosome depletion over a variety of Poly(dA:dT) elements (Materials and Methods) for each species (Figures S11, S12). We then tested if functional gene sets in each species were enriched or depleted for strongly anti-nucleosomal sequences in their NFRs. Finally, we compared this pattern to their chromatin organization (Figure 2C–J, right versus middle panels).
In some cases, the variation in chromatin organization within and between species is associated with variation in intrinsic “anti-nucleosomal” Poly(dA:dT) tracts. Within each species, Poly(dA:dT) sequences are enriched upstream of many highly expressed, nucleosome-depleted, “growth” gene sets, consistent with previous observations in S. cerevisiae [48],[49]. Between species, we found that gain and loss of polyA sequences is associated with changes in chromatin organization at several gene sets and phylogenetic points, suggesting that this is a common evolutionary mechanism used more than once in this phylogeny. We confirmed a prior observation [32] that the change in chromatin organization at mitochondrial ribosomal protein (mRP) genes in post-WGD respiro-fermentative species is accompanied by the loss of PolyA-like sequences from these promoters (Figure 2H). In addition, we found that the deeper and wider NFRs at splicing genes in Y. lipolytica are associated with a greater length and number of PolyA sequences at these genes (Figure 2J). Conversely, the relatively shallow NFRs of gluconeogenesis genes observed in S. castellii are associated with concomitant depletion of polyA sequences in this species (Figure 2F).
Nevertheless, intrinsic anti-nucleosomal sequences explain only 8.6%–25.7% of the global variation in NFR occupancy within a given species (Figure S9). Even when combining expression levels and sequence information together, these can only explain 13%–29% of the global variation in nucleosome organization in the 12 species (Figure S9E). Similar results are obtained when considering other measures of intrinsic anti-nucleosomal sequences, such as those based on computational models [21],[48] derived from in vitro data (unpublished analysis).
Thus, anti-nucleosomal sequences and expression patterns are insufficient to fully explain either conservation or divergence in chromatin organization across species. For example, proteasomal genes are highly expressed and have deep NFRs conserved in all species, but are not associated with intrinsic anti-nucleosomal sequences (Figure 2E). Furthermore, RNA Polymerase II subunits, RNA export, and nuclear pore genes are highly expressed with deep NFRs conserved in most species, but are enriched for intrinsically anti-nucleosomal sequences in only a subset of species (Figure 2G, see below). Conversely, peroxisome genes are highly expressed in D. hansenii, C. albicans, and Y. lipolytica, where they are packaged with long (but not deep) NFRs, despite no enrichment for Poly(dA:dT) tracts (see below). In these and other cases, even when we consider expression levels, much of the depletion in NFRs remained unexplained (Figures S9,S10).
We therefore wished to explore the role that the third mechanism—nucleosome eviction by chromatin remodelers—plays across the 12 species. We hypothesized that changes in chromatin remodeling would be accompanied by variation in the cis-regulatory elements bound by GRFs that likely recruit chromatin remodelers [25],[56],[57]. Unlike intrinsic anti-nucleosomal sequences that establish constitutively programmed NFRs, binding sites for GRFs likely establish regulated NFRs that can change based on trans inputs.
We first assessed the potential contribution of chromatin remodelers to chromatin organization based on the presence in NFRs of the known binding sites for the two best-studied S. cerevisiae GRFs: Abf1 and Reb1 (Figure S9E, Materials and Methods). Together, the two motifs explain 1.2%–15.1% of the observed variation in nucleosome organization in the 12 species. Furthermore, Abf1 and Reb1 can explain up to 12.6% of the residual variation after accounting for the contribution of expression levels and intrinsic sequences (Successive Lowess, ). Thus, GRFs can play an important role in explaining global chromatin organization.
Notably, the Abf1 and Reb1 sites explain little of the variation in D. hansenii, C. albicans, and Y. lipolytica—the species from the two clades most distant from S. cerevisiae. In particular, the Abf1 binding site explains less than 1% of the variation in each of these species, consistent with the absence of the Abf1 ortholog from their genome, and validating the specificity of our approach. Furthermore, although the Reb1 ortholog is present in each of these species, its contribution is substantially reduced (compared to, e.g., S. kluyveri). This loss of predictive power by Abf1 and Reb1 sites at increasing phylogenetic distance led us to hypothesize that other GRFs, with distinct binding specificity, are active in these species.
To identify novel GRF cis-elements, we therefore searched for short sequence elements that are depleted of nucleosomes in vivo but not in vitro [21]. We calculated the extent of nucleosome depletion over every 6- and 7-mer sequence in each of our species (Table S6, Materials and Methods) and identified those sequences whose depletion score in vivo in at least one species is significantly greater than expected from published in vitro data (Figure 3A, Figure S13–S14) [21]. This procedure automatically identified in vivo-specific depletion over 7-mers consistent with the binding sites for known S. cerevisiae GRFs such as Reb1 (Figure 3A, orange) [58],[59] and the Rsc3/30 components of the RSC ATP-dependent chromatin remodeling complex (Figure 3A, green) [25],[58],[59], validating our approach. Consistent with our hypothesis, it also revealed a number of sequence motifs that were specifically nucleosome-depleted in vivo in some species but not in S. cerevisiae, such as the CACGTG motif that serves as the binding site for Cbf1 in S. cerevisiae and C. albicans (Figure 3A, blue) [8],[58],[59],[60],[61]. We therefore propose that these sites are candidates for putative GRF function in these species.
When we compared the GRF sequences between species we discovered extensive divergence that largely conforms to phylogenetic distance (Figures S13, S14). The extent of nucleosome depletion over short sequence elements is well conserved between closely related species, such as S. cerevisiae and S. mikatae (∼2–5 MYA, Figure S13B). In contrast, there are much more dramatic differences in the in vivo depleted sites (e.g., Rsc3/30, Cbf1) between the more distant S. cerevisiae and K. lactis (∼150 MYA, Figures S13C, S14). Finally, there are also gradual changes in the specific Rsc3/30 CGCG-containing motifs that were nucleosome-depleted in each species (Table S6), consistent with co-evolution of a GRF and its binding site, as previously observed for TFs [3],[5].
The use of different GRF sites often follows strong phylogenetic patterns, allowing us to trace transitions from the dominant use of one repertoire of GRFs to that of another, and suggesting compensatory evolution of GRF use. Most notably, we find a major and gradual transition from the use of Cbf1 as a major GRF in pre-WGD species to the use of Reb1 as a GRF in post-WGD species (Figure 3B). The Cbf1 binding sequence CACGTG is nucleosome-depleted in vivo in most pre-WGD species (except Y. lipolytica and C. albicans) but not in post-WGD species (except C. glabrata) (Figure 3B). Conversely, Reb1 sites are nucleosome depleted in all post-WGD species but not in most pre-WGD species (except K. lactis) (Figure 3B). This complementary phylogenetic pattern suggests an evolutionary scenario where Cbf1 was a major ancestral GRF, Reb1 emerged as a GRF before the WGD, and gradually “took over” Cbf1's global functionality. Similar evolutionary patterns were previously observed for TFs [3],[7],[61],[62], and this is the first demonstration to our knowledge of such a “mediated replacement” for GRFs. Evolutionary transitions in GRF usage are sometimes limited to one or a few species. For example, we found a set of novel motifs that were nucleosome-depleted only in Y. lipolytica (Table S6), the earliest diverging species in our panel.
Finally, we observe changes in the relative balance between nucleosome depletion via GRFs and constitutively programmed depletion via Poly(dA:dT) sequences, suggesting a global mode of compensatory evolution. Most notably, A7/T7 is less nucleosome-depleted at D. hansenii promoters than at promoters of any other species, whereas Cbf1-like and Rsc3/30-like sites are strongly nucleosome-depleted in D. hansenii (Figure S14). This transition is likely due to the shorter lengths of Poly(dA:dT) stretches in D. hansenii (Figure S11C, Table S7), a sequence change that may be an adaptation to the high salt concentrations in this species' ecological niche (secondary to increased DNA flexibility in high salt). As noted above, D. hansenii has a very short average NFR width (Table S3, Figure S11D), consistent with diminished nucleosome repulsion at its shorter Poly(dA:dT) sequences. We hypothesize that the expansion in use of the Cbf1 and Rsc3/30 GRFs is a mode of compensatory evolution needed to adapt to a change in genome sequence in a unique niche; it also suggests that D. hansenii NFRs may be more responsive to environmental signals.
We next hypothesized that the identified GRFs are important for the observed chromatin organization in functional gene sets across species. To test this hypothesis, we assessed the enrichments of GRF motifs in the NFRs of each gene set across the 12 species (Table S8).
In some cases, GRF motifs (but not Poly(dA:dT) tracts) were enriched in a gene set across multiple species, strongly indicating a conserved regulatory mechanism. For example, the Abf1 site is enriched in RNA polymerase genes across the clade spanning S. cerevisiae and S. kluyverii (Figure S15D). However, since the spectrum of GRFs is species-specific (Figures 3B, S14), we found no gene set associated with the same GRF site across the entire phylogeny.
Instead, we found a number of cases where a single gene set has a conserved chromatin architecture but is associated with distinct GRF sites in different species, consistent with changes in the global GRF repertoire. This is most notable in proteasome genes, which are uniformly associated with wide/deep NFRs but are depleted of Poly(dA:dT) tracts (Figure 2E). The establishment of NFRs at these genes has likely transitioned from a mechanism dependent on the CACGAC sequence in the Candida clade to an Abf1-dependent mechanism in later lineages, with additional contribution from Reb1 and Rsc3/30 sites, as these GRFs gained dominance in specific species and clades (Figures 3C and S15E). Although the specific GRF mechanism underlying NFRs in proteasome genes has diverged, the establishment of wide/deep NFRs by a GRF-regulated mechanism (rather than polyA/constitutive mechanism) is conserved in all species. We hypothesize that GRF-regulated NFRs at proteasome genes may be related to the unusual transcriptional regulation of proteasome genes: these are among the few highly expressed “growth” genes (with open accessible promoters) that are further upregulated (rather than downregulated) during stress responses [63].
Could promoters evolve from having constitutively programmed NFRs to regulated ones? To test this, we searched for gene sets where chromatin organization is conserved, while the underlying anti-nucleosomal sequences have diverged in a phylogenetically coherent pattern. We found that genes encoding RNA polymerase subunits exhibit deep NFRs across most of the phylogeny (Figure S15D). These genes' promoters are associated with Poly(dA:dT) tracts in Y. lipolytica and the species of the Candida clade, with both Poly(dA:dT) and the site for the Abf1 GRF in species from S. kluyveryi to S. bayanus, and only with Abf1 in the clade spanning S. mikatae, S. paradoxus, and S. cerevisiae (Figures 3D and S15D). Similar behavior is seen at a number of other gene sets, such as those encoding nuclear pore components (unpublished analysis). This profile suggests an evolutionary scenario where the ancestral mechanism relied on Poly(dA:dT). With the emergence of Abf1 in the LCA of the pre- and post-WGD species [34], it gained additional control of the NFRs in this gene set, alongside Poly(dA:dT) tracts. Then, after the divergence of S. bayanus, Poly(dA:dT) tracts were lost from the genes' promoters, leading to a complete switch from a constitutively programmed to a regulated NFRs. This compensatory evolution is consistent with patterns observed for TF binding sites in functional regulons [3],[62] and with the global transitions in GRFs described above.
In some cases, the gain or loss of binding sites for GRFs can contribute to divergence in chromatin organization, coupled to phenotypic changes. Most notably, peroxisomal genes are associated with wider NFRs in Y. lipolytica, C. albicans, and D. hansenii, and shorter NFRs in subsequently divergent species (Figures 3E and S15F), but are not associated with intrinsic anti-nucleosomal poly(dA:dT) tracts in any of the 12 species. Instead, we find that these genes' promoters are enriched for PolyG and Rsc3/30-like sites in Y. lipolytica, C. albicans, and D. hansenii, but not in other species. This suggests an evolutionary scenario where either a Rsc-like motif or PolyG-based nucleosome depletion was the ancestral mechanism controlling peroxisomal genes, and was subsequently lost in the LCA of the clade spanning S. kluyverii and S. cerevisiae. This scenario is consistent with the higher expression of peroxisomal genes in Y. lipolytica (where peroxisomes are particularly central for carbon metabolism) and C. albicans (where peroxisomes play a key role in virulence).
Even when NFR positions and their underlying mechanisms are largely conserved, they can play an important role in regulatory divergence. Nucleosomes are generally inhibitory to TF binding [13], and in S. cerevisiae most functional TF binding motifs are found in NFRs [23]. Precise positioning of TF binding sites relative to nucleosomes has regulatory consequences such as changing signaling thresholds [64] or logic gating [65]. We therefore hypothesized that an evolutionary change in the location of TF-binding motifs relative to the nucleosomes in a gene's promoter can lead to regulatory divergence between species.
To test this hypothesis, we examined the location of known TF binding motifs (from S. cerevisiae; [58],[59],[60],[66]) relative to nucleosome positions in each of the 12 species (Materials and Methods). Consistent with our expectations, in S. cerevisiae (Figure 4A,B), up to 90% of the binding sites for growth-related TFs are localized to NFRs (e.g., REB1, ABF1, RAP1, and FHL1), whereas as few as 25% of sites for stress-related TFs are at NFRs (e.g., HSF1, YAP6, HAP2/3/5, GZF3, and CRZ1). Thus, sequences that are mostly occluded by nucleosomes tend to be the binding sites for inactive TFs, and we can use chromatin information to infer TF activity under our growth conditions in each species. We therefore calculated for each motif the fraction of its instances located in NFRs in each of the 12 species (Figures 4C and S16).
The NFR positioning of many key motifs is strongly conserved. For example, sites for growth-related factors such as SWI4/6 and GCN4 were similarly NFR-exposed in all species in this phylogeny. Notably, this conservation is observed despite the fact that many motifs, which were experimentally defined for S. cerevisiae proteins, were globally less NFR-localized in distantly related species (Figure 4C, Figure S16B). This can be attributed in some cases to divergence of binding site preferences of the cognate TFs, and in other cases to the absence of the TF's ortholog from the genome (Figure 4C, white). Nevertheless, many motifs showed robust conserved positioning in NFRs.
Conversely, the motifs for key TFs associated with regulation of respiration and carbohydrate metabolism have repositioned relative to NFRs at the WGD, consistent with regulatory divergence in these functions (Figure 4D). For example, the sites for the HAP2/3/4/5 complex (a regulator of respiration genes) and for YAP6 (a regulator of oxidative functions) have re-positioned from NFRs to nucleosome-occluded positions post-WGD, consistent with the reduction in expression of respirative genes. In contrast, the sites for the carbon catabolite repressor MIG2 and for the glucose-responsive TF RGT1 have repositioned from nucleosomes into NFRs in post-WGD species, consistent with these factors' role in establishing a fermentative strategy through gene repression.
Motif re-positioning has also occurred at other phylogenetic points and gene sets, suggesting that this is a general regulatory and evolutionary mechanism (Figure 4E,F). For example, the mating-related STE12 motif is significantly enriched upstream of reproduction and mating-related genes in species from S. cerevisiae to S. kluyverii, including C. glabrata. Although STE12 sites are found in NFRs at mating genes for most of these species, they are largely nucleosome-occluded in C. glabrata (Figure 4E), an organism which has never been observed to mate [67]. We speculate that occlusion of STE12 sites under nucleosomes may contribute to this species' reluctance to mate, but the continued enrichment of STE12 upstream of mating genes and the retention of many meiosis-related genes [34] in C. glabrata suggests that it may still be capable of mating under special conditions. We therefore predict that conditions (environmental or perhaps genetic) that either mobilize or destabilize the nucleosomes covering STE12 sites at pheromone-response genes might enable mating in this species. Similarly, motifs for UME6, a major regulator of meiosis genes in S. cerevisiae [68], are globally NFR-positioned in all species except C. glabrata (Figure 4F), despite the fact that UME6 sites are enriched upstream of orthologs of meiosis-related genes in C. glabrata. Thus, the relative re-positioning of NFRs and TF binding sites may help explain the molecular underpinnings of dramatic changes in regulatory and phenotypic evolution.
Finally, we asked whether chromatin information could be used to infer the regulatory effect of exposed TF binding sites from the expression level of their target genes. We expect exposed TF binding sites to have different regulatory consequences depending on whether or not the TF is active and whether it acts as an activator or a repressor. We reasoned that an NFR-positioned site for an active positive regulator will be associated with a higher expression of the target genes. Conversely, an NFR-positioned site for an active negative regulator will be associated with a lower expression of the target genes. We therefore compared the expression level of all genes where a given TF motif was located within nucleosomes versus those in which the motif was located within promoter linkers (largely the NFR, Figure 5A). Consistent with our expectation, in S. cerevisiae, transcriptional activators known to be active in mid-log phase, such as RPN4 or PBF1, were associated with higher expression levels at genes carrying an accessible, linker-positioned motif. In contrast, NFR-positioned motifs for transcriptional repressors known to be active in mid-log (e.g., MIG1, SUM1, NRG1, DIG1, STB1/2, or RIM101; Figure 5A) were associated with lower downstream gene expression. Thus, we devised a novel approach to predict whether a given motif is associated with an activator or repressor in vivo in the growth condition tested.
When we extended this analysis to all 12 species (Figure S17), we found substantial divergence in the regulatory logic of the same NFR-positioned motif, most notably at the WGD (Figure 5B). We found a host of motifs which, when present in NFRs, were associated with differences in RNA expression levels between pre- and post-WGD species. Many of those (∼100) appeared to shift from activator-like behavior in pre-WGD species (higher target expression when in NFR) to repressor-like behavior in post-WGD species (lower target expression when in NFR). These included sites for a surprisingly large number of TFs involved in repression of metabolic genes in S. cerevisiae, including MIG1, GIS1, RGT1, and GAL80. Interestingly, several of these genes are found in a single copy in pre-WGD species but were retained as duplicates [34] with similar DNA-binding specificity following the WGD (e.g., GIS1/RPH1, RGT1/EDS1; Figure 5B,C). This suggests that widespread usage of competing activator/repressor pairs in S. cerevisiae may have been facilitated by the generation of such TF pairs at the WGD. Such duplication of trans-factors can serve as an alternative evolutionary mode to expand and evolve regulatory capacity [69] even when NFRs and motif positioning may be conserved.
In this work we used a comparative functional genomics approach to study the evolutionary interplay between chromatin organization, gene expression, and regulatory sequence elements. We aimed to achieve two goals: (1) understand the determinants of chromatin organization and function using comparative genomics and (2) characterize the role of chromatin organization in the evolution of gene regulation.
What establishes the nucleosomal organization of a genome? While it has been argued that intrinsic DNA sequence can almost fully explain nucleosome organization [21], recent analysis of in vitro reconstitution data showed that the major intrinsic contributor to nucleosome positioning in budding yeast is the anti-nucleosomal behavior of Poly(dA:dT) and related sequences [21],[24],[70]. Conversely, recent reports indicate that in S. pombe Poly(dA:dT) plays only a minor role in nucleosome exclusion in vivo [33], indicating that even the best-understood sequence contributor to chromatin organization plays variable roles in chromatin structure in different species.
Our analysis provides several lines of evidence that expression levels, intrinsic anti-nucleosomal sequences, and binding sites for GRFs that may recruit chromatin modifiers all play a role in establishing promoter chromatin architecture, and that the balance between these three contributors changes in evolution and between functional groups of genes. (1) We show that a sequence-based model based on in vitro depletion alone [21] can only account for 8.6%–25.7% of variance in NFR depth within any of the 12 species, including S. cerevisiae (10.6%). Similarly, expression levels alone can only account for 1.7%–13.1% of the variation in each species. Even when combining both the expression and intrinsic models we can only explain 13%–29% of the variation within any single species. (2) Although changes in intrinsic sequences and expression levels can explain changes in chromatin across species for some gene sets (e.g., mRPs or splicing genes; Figure 6A and B), they are insufficient to explain conserved chromatin behavior across the phylogeny (e.g., RNA Polymerase subunit genes; Figure 6D), nor do they explain changes in chromatin organization across species in other groups of genes (e.g., peroxisome genes; Figure 3E). Thus, these two determinants (alone or in combination) are insufficient to explain both intra- and inter-species variation. (3) In contrast, by comparing our in vivo data in each species to two in vitro datasets [21],[32], we find in each species a host of sequences that exhibit significantly greater nucleosome depletion in vivo than in vitro. Many of these correspond to binding sites for known GRFs that play an active role in nucleosome eviction in S. cerevisiae [14],[25],[56],[58], whereas others represent novel candidate GRF sequences (Figures 3 and 6C). (4) The relative contribution to nucleosome organization from GRFs, intrinsic sequences, and expression levels varies between different gene sets (in all species). For example, we show that intrinsic anti-nucleosomal sequences are enriched at NFRs in cytoplasmic RPs (in all species; Figure 2C), whereas GRFs fulfill this role in proteasome genes (in all species; Figure S15). (5) We also show that the relative contribution of one mechanism versus another can change in evolution (across species), both globally (as in the halophile D. hansenii, that relies more on GRFs) and in specific gene sets (as in the RNA polymerase gene set that shifted from intrinsic to regulated NFRs; Figure 6D). (6) Globally, even when we consider only the binding sites for the two best-characterized GRFs from S. cerevisiae (Abf1 and Reb1), GRFs alone can explain 5.2%–15.1% of the variation in nucleosome organization (in species where their orthologs are present), and 3.7%–12.6% of the residual variation after considering the contribution from expression and Poly(dA:dT). Taken together, this analysis points to a complex interplay between the different factors that control nucleosome positions, allows us to assess their contributions, and recognizes the plastic and evolvable nature of all the determinants.
Our study also discovers an intricate and intimate relationship between conservation and divergence of chromatin organization and evolution of gene regulation. At one extreme, we found a broad functional dichotomy in chromatin organization between “growth” and “stress” genes, which is largely conserved. At the other extreme, we found that chromatin organization has diverged at a major evolutionary scale, as has happened during the evolution of respiro-fermentation, and at other points of phylogenetic and phenotypic divergence.
We found five major mechanisms by which chromatin organization can be associated with divergence of gene expression. Each of these was “used” more than once in the phylogeny, and is associated with more than one phenotypic or regulatory change, including the changes described in carbon metabolism, mating, meiosis, and splicing genes. These include (1) gain or loss of intrinsic (PolyA) sequences can open or close NFRs (Figure 6A,B) [32]; (2) conserved NFRs can be controlled by different GRF determinants, through compensatory evolution (Figure 3C); (3) NFRs can shift between constitutive and regulated determinants by compensatory (“balanced”) gain/loss of intrinsic anti-nucleosomal sequences and GRF binding sites (Figure 6D); (4) motifs can re-position relative to NFRs to change transcriptional output (Figure 6E–G); and (5) duplication and divergence of trans-factors can expand the regulatory behavior of conserved NFRs and binding sites (Figure 6H).
The evolution of the respiro-fermentative lifestyle following the WGD required a major reprogramming of the yeast transcriptional network and involved all of the mechanisms we describe. The shift thus included loss of intrinsic Poly(dA:dT) anti-nucleosomal sequences in some functional modules (e.g., mitochondrial RP genes), and the loss or switch of putative GRF sequences in others (e.g., oxidation-reduction genes). Furthermore, sites for certain respiratory TFs (e.g., HAP2/3/5, YAP1/3/6) have re-positioned out of NFRs, and those for glucose repression TFs have re-positioned into NFRs (e.g., RGT1, MIG1). In yet other cases, the WGD has resulted in the retention of paralogous activator-repressor pairs that control several modules in carbohydrate metabolism. Notably, each of these mechanisms has acted also at other phylogenetic points, suggesting that they point to general principles, and emphasizing the utility of the WGD as a model to study regulatory evolution.
Our work provides a general framework for the study of chromatin organization, function, and evolution. This includes a comprehensive genomics resource (http://www.broadinstitute.org/regev/evolfungi/) and a host of analytical approaches with broad applicability. Future studies can use our resource and methods to decipher more detailed models of the relationship between sequence elements, trans-factors, and gene expression, as well as on the evolution of regulatory systems. Finally, our comprehensive study in the emerging field of comparative functional genomics demonstrates how to combine the power of functional assays with extensive phylogenetic scope, to shed light both on mechanistic and evolutionary principles.
We used the following strains in the study: Saccharomyces cerevisiae, BY4741, Saccharomyces cerevisiae, Sigma1278b L5366, Saccharomyces paradoxus, NRRL Y-17217, Saccharomyces mikatae, IFO1815, Saccharomyces bayanus, NRRL Y-11845, Candida glabrata, CLIB 138, Saccharomyces castellii, NRRL Y-12630, Kluyveromyces lactis, CLIB 209, Kluyveromyces waltii, NCYC 2644, Saccharomyces kluyveryii, NRRL 12651, Debaryomyces hansenii, NCYC 2572, Candida albicans, SC 5314, Yarrowia lipolytica, CLIB 89.
All cultures were grown in the following medium: Yeast extract (1.5%), Peptone (1%), Dextrose (2%), SC Amino Acid mix (Sunrise Science) 2 grams per liter, Adenine 100 mg/L, Tryptophan 100 mg/L, and Uracil 100 mg/L. This in-house recipe was designed to mitigate differences in growth rates between species.
Overnight cultures for each species were grown in 450 ml of media at 220 RPM in a New Brunswick Scientific air-shaker at 30°C until reaching mid log-phase (OD600 = 0.5, WPA biowave CO 8000 Density Meter). Before formaldehyde fixation, 50 ml of the culture were transferred to a 50 ml conical and spun down immediately. The isolated cell pellets were then placed in liquid nitrogen, stored at −80°C, and were later archived in RNA later for future RNA extraction. Nucleosomal DNA isolation was carried out as previously described [23] with the following slight modifications. For different species, cells were spheroplasted with zymolase between 30 and 40 min, depending on how much time was necessary to fully remove each species' cell wall. MNase digestion levels for all samples were uniformly chosen across species to contain a slightly visible tri-nucleosome band (Figure S1). Mononucleosomes were size-selected on a gel and purified using BioRad Freeze-N-Squeeze tubes followed by phenol-chloroform extraction. Selected DNA was prepared for sequencing using the standard Illumina protocol that includes blunt ending, adaptor ligation, PCR amplification, and final size selection plus gel purification [35]. Libraries were sequenced on an Illumina 1G Analyzer, to generate 36 bp reads.
Total RNA was isolated using the RNeasy Midi or Mini Kits (Qiagen) according to the provided instructions for mechanical lysis. Samples were quality controlled with the RNA 6000 Nano ll kit for the Bioanalyzer 2100 (Agilent). Genomic DNA was isolated using Genomic-tip 500/G (Qiagen) using the provided protocol for yeast. DNA samples were sheared using Covaris sonicator to 500–1000 bp fragments, as verified using DNA 7500 and DNA 12000 kit for the Bioanalyzer 2100 (Agilent). Independently sheared samples labeled with Cy3 and Cy5 were highly correlated (R>.97 in each of 4 independent hybridizations), indicating that the shearing procedure is reproducible and unbiased. Total RNA samples were labeled with Cy3 (cyanine fluorescent dyes) and genomic DNA samples were labeled with Cy5 using a modification of the protocol developed by Joe Derisi (UCSF) and Rosetta Inpharmatics (Kirkland, WA) that can be obtained at www.microarrays.org.
Between three and four biological replicates of Cy3-labeled RNA samples were mixed with a reference Cy5 labeled genomic DNA sample and hybridized on two-color Agilent 55- or 60-mer oligo-arrays. We used the 4×44 K format for the S. cerevisiae strains (commercial array; 4–5 probes per target gene) or a custom 8×15 K format for all other species (2 probes per target gene, designed using eArray software, Agilent). After hybridization and washing per Agilent's instructions, arrays were scanned using an Agilent scanner and analyzed with Agilent's feature extraction software version 10.5.1.1. For each probe, the median signal intensities were background subtracted for both channels and combined by taking the log2 of the Cy3 to Cy5 ratio. To estimate the absolute expression values for each gene, we took the median of the log2 ratios across all probes. The experiments were highly reproducible; most biological replicates correlated at R = 0.99 and replicates with R<0.95 were removed. Different biological replicates were combined using quantile normalization to estimate the absolute expression level per gene per species.
We used BLAT [71] to map sequenced reads from each experiment to the corresponding reference genome, keeping only reads that mapped to a unique location and allowing for up to 4 mismatches. Each uniquely mapped read was then extended to a length of 100 bp. To generate a genomic nucleosome occupancy landscape, we summed all extended reads covering each base pair. We then masked all repetitive regions along each track, defining repetitive regions as locations in the genome that cannot be uniquely defined by the length of a read (36 bp). We also masked all regions of nucleosome occupancy greater than 10 times the median occupancy, to remove outlier effects that occur in places such as the rDNA locus. To normalize for sequencing depth for each genomic nucleosome track, we divided the occupancy at each location by the mean nucleosome occupancy per base pair. These normalized maps were used to generate the average nucleosome occupancy plots (Figures 1, 2A, and S2–S3).
To infer the location of nucleosomes from the data, we used a Parzen window approach similar to that previously described [35],[46]. Our modified approach uses three parameters—the average DNA fragment length, the standard deviation of the Parzen window, and the maximum allowable overlap between nucleosomes. To estimate the mean DNA fragment length in each experiment, we shifted reads from one strand and then correlated them with the reads of the opposite strand. For each species, we observed a peak in the cross-correlation at a shift between 127 and 153 bp, which we used to estimate the mean DNA fragment length per experiment. We chose a standard deviation of the Parzen window of 30 bp for all species, since it closely matched the observed standard deviation around the cross-correlation peak of each experiment. Finally, we set the maximum allowable overlap between nucleosomes to 20 bp. We then shifted all read start locations by half of the mean DNA fragment length in the direction towards the dyad of the nucleosome they represent. Our approach places a normal distribution with a standard deviation of 30 bp at each read's shifted location. Summing all individual curves for all loci leads to a smoothed probability landscape of nucleosome occupancy. We next identify all peaks along the landscape, which represent nucleosome centers. The algorithm then places nucleosomes along the genome in the order of decreasing peak heights (greedy approach) and iteratively masks out these regions to prevent more than 20 bp overlap between nucleosomes.
We define 5′ and 3′ NFRs as the linker DNA of “significant length” closest to the 5′ and 3′ end of each gene, respectively. To find NFRs, we first created a nucleosome call landscape for each genome, normalized for sequencing depth in the same manner as the nucleosome occupancy maps (above). NFR boundaries were often obscured by very low occupancy nucleosome calls. We therefore removed all nucleosome calls with occupancy less than 40% of the average nucleosome occupancy from the map. We searched for 5′ or 3′ NFRs within 1,000 bases upstream/downstream of the 5′ or 3′ end of each gene, truncated when neighboring ORFs overlapped this region. We then defined an NFR as the linker DNA longer than 60 bp closest to the 5′ or 3′ end of each gene. If no linker longer than 60 bp was found in this search, we defined the NFR as the first linker from the 5′ or 3′ end. Our method was highly predictive of transcription start sites (TSSs) in S. cerevisiae [50]—the NFR boundary closest to the 5′ end of the gene was able to predict 84% of TSSs within 50 bp. Linker lengths of 50 bp or 70 bp and occupancy thresholds of 30% or 50% produced highly similar results (unpublished data).
Since 5′NFR-ATG distances vary substantially between species, an analysis of nucleosome organization that relies on alignment by ATG can be misleading. For example, the average nucleosome organization of C. glabrata and S. castellii look similar when aligned by the +1 nucleosomes (Nuc+1) but very different when aligned by ATG (unpublished data). A previous study [32] defines a promoter nucleosome depleted region (PNDR) score as mean nucleosome occupancy of the most depleted 100 bp region within 200 bp upstream of the ATG. Since some species have longer 5′NFR-ATG distances we reasoned that the NFR of some genes may not be contained within a 200 bp window (e.g., only a third of C. glabrata NFRs are contained within 200 bp, while 90% are contained within 500 bp). To avoid such pitfalls and analyze nucleosome organization consistently in all species we aligned the data by Nuc+1, which is consistent with alignment by TSS.
For S. cerevisiae we used functional gene sets from several sources: KEGG [72], GO categories [73], MIPS [74], and BioCyc [75], as previously described [34]. For all other species, we project these gene sets based on gene orthologies [34] using the ortholog mapping at www.broad.mit.edu/regev/orthogroups.
The chromatin features used in our analysis are listed and defined in Table S1. To quantify the enrichment for a given feature within a functional category we used the two-sample K-S test. For each K-S test, we defined our two sample sets as genes within a given functional group and all other genes in the genome. The K-S test quantifies the distance between the distributions of a given chromatin feature for the two sets. The K-S statistic KK-S is defined as the maximum absolute difference between the cumulative distribution functions (CDFs) of the two samples. We estimated the P value, PK-S, for the statistical significance of this difference as follows:
For further analysis, we converted P values to K-S scores, SK-S, where SK-S = ±log10(PK-S) if the difference realizing the statistic KK-S is positive/negative, respectively. To account for multiple hypotheses testing, we only considered PK-S as significant if it was below the P value threshold for a False Discovery Rate of 5% [76]. This analysis was also applied to absolute expression levels, Poly(dA:dT) strength in NFRs, trans factor motif affinity scores in NFRs, and comparison in expression of sites located in NFRs versus sites located in nucleosomes.
To subtract the effect of expression on observed chromatin features, we used robust Lowess smoothing. We smoothed the scatter data of each chromatin feature versus expression level using a Lowess linear fit and a smoothing window set to 10% of the span of expression level values. We assigned zero weight to outliers, defined as data more than six standard deviations from the mean. To remove the effect of expression, the Lowess fit was subtracted from its corresponding chromatin feature value. K-S functional enrichments for the Lowess subtracted chromatin features were calculated as described above.
We assessed transcriptional activity by absolute RNA expression. We assessed intrinsic anti-nucleosomal sequence by Poly(dA:dT) strength in NFRs, since it explains the vast majority of the intrinsic sequence information and generalizes to all species in an unbiased manner. Other models of intrinsic sequence contribution [21],[48] yielded similar results (unpublished data). We assess the contribution of chromatin modifiers based on the Abf1 and Reb1 motif affinity scores in NFRs. This is a conservative estimate, since we only considered the two most established GRFs. To quantify the contribution of each of these factors on NFR occupancy, we used robust Lowess smoothing as described above.
To compute the percent of variance explained by the robust Lowess fit, each NFR occupancy was assigned a “fitted” value Fi from the Lowess fitting line based on each of the three determinants. Then the variance of the residuals σ2R = var({Fi−Zi}) was compared to the variance of the original data σ2D = var({Zi}). The percent of variance explained is defined as (1−(σ2R/σ2D))*100. To find the percent variance explained by all determinants we first fit NFR occupancy versus one determinant, then iteratively took the residual, and fit it against the next determinant. For the figures, we first fit expression, then fit the successive residual versus Poly(dA:dT) tracts, and then fit the residual versus Abf1 and Reb1 motif affinity scores. Changing the order of the successive fits did not significantly reduce the total percent variance explained.
Promoter sequences for each gene were defined as 1,000 bases upstream, truncated when neighboring ORFs overlapped with this region. We collected a library of Position Weight Matrices (PWMs) for several hundred S. cerevisiae DNA-binding proteins as previously defined [58],[59],[60],[66]. Motif targets were identified via the TestMOTIF software program [77] using a 3-order Markov background model estimated from the entire set of promoters per genome. We considered all motif instances with P value <0.05 as significant. Since a few motifs had thousands of instances for this cutoff, we also limited the number of promoters with significant sites to the top 1,000. The upper bound was chosen to exceed the maximal number of promoters bound (866, P value <0.05) by any TF in S. cerevisiae, as measured by ChIP-chip [60]. For all subsequent motif analyses, we used the above criterion to define two sets of sites: (1) all significant sites within allowed promoters and (2) the best sites per allowed promoters.
All motif instances were binned into five regions (Nuc+1, 5′NFR, Nuc−1, Nuc−2, and NFR2—the linker between Nuc−1 and Nuc−2) if their centers overlapped with the defined regions. In addition, sites were also split into two categories: Linkers (5′NFR and NFR2) and Nucs (Nuc+1, Nuc−1, and Nuc−2). We assigned the expression level of each gene to each site in the upstream promoter of that gene. We used a two-sample K-S test (as described above) to quantify the difference in expression levels between sites in Linkers versus Nucs.
To quantify the preference of a motif for nucleosome depleted regions, we compared the mean log2 normalized nucleosome occupancy at all sites (x) against the mean log2 normalized nucleosome occupancy over the corresponding promoters (y). To estimate the significance of the difference of the two vectors (x-y), we used the paired Wilcoxon signed rank test that assigns a P value for rejecting the null hypothesis that x-y comes from a continuous, symmetric distribution with a zero median.
To estimate the probability that k or more elements intersect subsets of n and m members at random in a superset of size N (or the P value for overlap of k, PHG) we summed over the right tail of a hypergeometric distribution:
Using the hypergeometric P values, we estimated the significance of k overlaps between n genes with sites in their upstream promoter and m genes within a GO category, for a species with N genes.
We represent each motif of length L by a position specific scoring matrix (PSSM) P, or the probability distribution P(S1, …, SL) of that motif occurring over any sequence S1…SL. This is a standard approximation to a factor binding energy for sequence S1…SL. We also learned the 0th-order Markov background probability distribution B(S1, …, SL) for each sequence S1…SL, set to the frequency of the four nucleotides in the promoter regions of a given species. We calculate A(P,S), a motif's affinity score for an NFR sequence S, by summing the contributions of P(S1, …, SL)/B(S1, …, SL) over all allowable positions k in S as follows:
Here, b(Sk+j-1) is the background probability of the nucleotide Sk+j-1 of sequence S, and p(Sk+j-1,j) is the probability for nucleotide Sk+j-1 in position j of the motif's PSSM. For the results in this study, we combined the contributions of both forward and reverse strands of each NFR. Also, normalizing the affinity by the length of each NFR sequence did not affect our results significantly.
Prior to analysis, we log2-transformed the normalized nucleosome occupancy data (Data post-processing, above), subtracted the mean, and divided by the standard deviation. Hence, the global nucleosome occupancy data for each species is approximately normal with zero mean and unit variance. We also used the same procedure for processing published in vitro data [21].
For each N-mer, we define the in vivo depletion score as the mean −log2 normalized nucleosome occupancy across all instances and all instances of the reverse complement. We also defined the depletion score relative to in vitro as power 2 of the difference between the in vivo depletion scores in each species and the in vitro depletion scores in S. cerevisiae (also repeated for in vitro data from C. albicans [32]). The analysis was done for N = 5, 6, 7, 8 and also repeated for N-mers found only in coding regions and only in upstream promoter regions.
To annotate all Poly(dA:dT) tracts in each species and determine their nucleosome repelling strength we used an approach similar to a previously described one [48]. In summary, for each species' genome we found all PolyA or PolyT tracts of length L of 5 bp or more. We define the depletion score for a tract of length L as the mean of the −log2 normalized nucleosome occupancy across all instances of that length. This was calculated both using in vitro data from S. cerevisiae [21] and the in vivo data from each species. For long Poly(dA:dT) tracts with very few occurrences in a given genome we noticed a larger variation in the depletion score, likely due to small sample size. To mitigate this problem, we fit a line for depletion scores versus L using a weighted linear least squares fit with weights proportional to the number of occurrences for tracts of length L. We then used the line as an estimate for long tracts with fewer than 100 occurrences in a given genome. We iterated this procedure for all maximal Poly(dA:dT) tract with k allowed mismatches, k = 1, …,20. The depletion score increases linearly with L for tracts with different k, confirming that a linear fit is appropriate (Figure S11).
To aggregate all non-overlapping Poly(dA:dT) tracts within a given genome, we first quantized the strengths for each L. We define the fold depletion score of all tracts of length L as power 2 of the depletion score. We then quantized all Poly(dA:dT) tract fold depletion scores to the highest fold depletion level exceeding 2, 4, 8, 16, and 32. For example, a tract with a depletion score of 3.5 is 23.5 = 11.3-fold depleted in nucleosomes relative to average, and would be assigned a fold depletion score of 8. We next iterated over all Poly(dA:dT) tracts with mismatches k = 0, …,20, replacing overlapping tracts only if the tract with more mismatches had a higher quantized fold depletion score.
Data have been submitted to GEO, accession #GSE21960.
Supplementary website http://www.broadinstitute.org/regev/evolfungi/
|
10.1371/journal.ppat.1000169 | Molecular Mechanisms Involved in Vascular Interactions of the Lyme Disease Pathogen in a Living Host | Hematogenous dissemination is important for infection by many bacterial pathogens, but is poorly understood because of the inability to directly observe this process in living hosts at the single cell level. All disseminating pathogens must tether to the host endothelium despite significant shear forces caused by blood flow. However, the molecules that mediate tethering interactions have not been identified for any bacterial pathogen except E. coli, which tethers to host cells via a specialized pillus structure that is not found in many pathogens. Furthermore, the mechanisms underlying tethering have never been examined in living hosts. We recently engineered a fluorescent strain of Borrelia burgdorferi, the Lyme disease pathogen, and visualized its dissemination from the microvasculature of living mice using intravital microscopy. We found that dissemination was a multistage process that included tethering, dragging, stationary adhesion and extravasation. In the study described here, we used quantitative real-time intravital microscopy to investigate the mechanistic features of the vascular interaction stage of B. burgdorferi dissemination. We found that tethering and dragging interactions were mechanistically distinct from stationary adhesion, and constituted the rate-limiting initiation step of microvascular interactions. Surprisingly, initiation was mediated by host Fn and GAGs, and the Fn- and GAG-interacting B. burgdorferi protein BBK32. Initiation was also strongly inhibited by the low molecular weight clinical heparin dalteparin. These findings indicate that the initiation of spirochete microvascular interactions is dependent on host ligands known to interact in vitro with numerous other bacterial pathogens. This conclusion raises the intriguing possibility that fibronectin and GAG interactions might be a general feature of hematogenous dissemination by other pathogens.
| Many bacterial pathogens can cause systemic illness by disseminating through the blood to distant target sites. However, hematogenous dissemination is still poorly understood, in part because of an inability to directly observe this process in living hosts in real time and at the level of individual pathogens. We recently engineered a fluorescent strain of Borrelia burgdorferi, the Lyme disease pathogen, and visualized its dissemination from the microvasculature of living mice using intravital microscopy. We found that dissemination was a multistage process that included tethering, dragging, stationary adhesion and extravasation. In the study described here, we used quantitative real-time intravital microscopy to investigate the mechanistic features of the vascular interaction stage of B. burgdorferi dissemination in living hosts. We found that tethering and dragging interactions (collectively referred to as initiation interactions) were mechanistically distinct from stationary adhesion. Initiation of microvascular interactions required the B. burgdorferi protein BBK32, and host ligands fibronectin and glycosaminoglycans. Initiation interactions were also strongly inhibited by the low molecular weight clinical heparin dalteparin. Since numerous bacterial pathogens can interact with fibronectin and glycosaminoglycans in vitro, these observations raise the intriguing possibility that fibronectin and glycosaminoglycan recruitment might be a feature of hematogenous dissemination by other pathogens.
| Hematogenous dissemination of pathogenic organisms is an important feature of disease progression. However, dissemination is poorly understood, in large part because of the difficulty in studying this process directly in living organisms under the shear stress conditions that characterize the host vasculature. One such disseminating pathogen is the spirochete Borrelia burgdorferi, a primarily extracellular bacterium causing Lyme disease, also referred to as Lyme borreliosis [1].
Pathogenic spirochetes cause a number of emerging and re-emerging diseases, including syphilis, leptospirosis, relapsing fever and Lyme disease [2]–[5]. B. burgdorferi is transmitted to the dermis of vertebrate hosts during the blood meal of Ixodes ticks, and subsequently disseminates to other tissues and organs during the hematogenous phase of infection [1]. B. burgdorferi and other spirochetes interact with endothelial cells under static conditions in vitro [6]–[8]. However, until recently, spirochete-vascular interactions have never been directly examined in the host itself, or under the fluid shear forces that are present at dissemination sites [9].
To facilitate direct study of hematogenous dissemination we recently generated a fluorescent infectious strain of B. burgdorferi, and used intravital microscopy (IVM) to directly visualize its interaction with and extravasation from the microvasculature of living murine hosts (as summarized in Fig. 1A–C) [9]. IVM is a powerful tool for studying the dissemination and transmigration of tumor and immune cells in living hosts, but it is only recently that this technique has begun to be applied to the study of host-pathogen interactions [10],[11].
The results of our recent study indicated that B. burgdorferi dissemination from the host microvasculature in vivo is a progressive, multi-stage process consisting of several successive steps: transient and dragging interactions (collectively referred to as short-term interactions), followed by stationary adhesion and extravasation. Short-term interactions constitute the majority of spirochete-endothelial associations (89% and 10% for transient and dragging interactions, respectively), take less than one second (transient interactions) or 3–20s (dragging interactions) to travel 100 µm along the vessel wall, and occur primarily on the surface of endothelial cells and not at endothelial junctions [9]. Transient interactions are characterized by a tethering-type attachment-detachment cycle of association in which part of the spirochete adheres briefly to the endothelium before being displaced by blood flow, whereas dragging spirochetes adhere along much of the length of the bacterium, and creep more slowly along the vessel wall [9]. In contrast, stationary adhesions (1% of interactions) do not move along the vessel wall, occur chiefly, but not exclusively, at endothelial junctions, and entail a more intimate association with the endothelium than short-term interactions [9]. Finally, spirochete extravasation (<0.12% of interactions) also occurs primarily, but not exclusively, at endothelial junctions, and is a triphasic process consisting of a rapid, end-first initial penetration of the endothelium, followed by a prolonged period of reciprocating movement, and ending with a rapid exit phase in which the bacterium bursts out of the vessel and migrates rapidly into the surrounding tissue [9].
In vitro studies have shown that B. burgdorferi binds several host molecules that might mediate endothelial interactions in vivo, including fibronectin (Fn), integrins, heparan sulfate-type glycosaminoglycans (GAGs) and regulators of the complement cascade [12]-[20]. A broad array of pathogens have been shown to interact with these ubiquitous host molecules in direct binding assays and tissue culture models in vitro; most of these studies have been performed in the absence of shear forces, microvascular endothelium or a functioning immune system, and so the potential contribution of such interactions to hematogenous dissemination in the living host is unknown. To date, 19 candidate adhesin genes have been identified in B. burgdorferi, two of which are known to interact with integrins (P66 and BBB07), and two of which can associate with heparan sulfate GAGs (BBK32 and Bgp) [16], [17], [19], [21]–[25]. B. burgdorferi encodes one characterized Fn binding protein, BBK32, and appears to express a number of others [12],[16],[24]. Five B. burgdorferi CRASP proteins that interact with complement cascade regulating proteins factor H, FHL-1 and FHR-1 have also been identified [20],[26],[27], but their potential contributions to endothelial cell adhesion are unknown.
In the work described here we used IVM to explore the mechanistic basis for B. burgdorferi interactions with the microvasculature of living mice. We found that the initiating and stationary adhesion stages of microvascular interactions were mechanistically distinct but inter-dependent events, and that BBK32, Fn and GAGs played a substantial role in initiation events. These findings and the methodology described here provide a framework for investigating the role of Fn and GAGs in vascular interactions during hematogenous dissemination by B. burgdorferi and possibly other pathogens.
To quantitatively analyze B. burgdorferi interactions with the microvascular endothelium in vivo we employed conventional epifluorescence IVM to examine interactions in the flank skin of mice after intravenous inoculation with 4×108 spirochetes (Fig. 1 and Videos S1 and S2). Conventional epifluorescence IVM was used instead of spinning disk confocal IVM because it is more effective for imaging the rapid associations that constitute the bulk of B. burgdorferi microvascular interactions (Fig. 1). Analysis was performed in post-capillary venules, where interactions could be most accurately quantified. For all experiments reported in this manuscript, data describing the numbers of recorded vessels and mice for each experimental condition, as well as the average time after spirochete injection at which recordings were made are provided in the Figure Legends. As we have recently reported, during the experimental observation period no signs of endothelial or leukocyte activation were detected [9]. In addition, leukocyte adhesion in dermal postcapillary venules is an indicator of local activation, and can be measured by using the dye rhodamine 6G to fluorescently label all circulating leukocytes and then counting the number of adherent leukocytes in a 100 µm length of vessel [28]. The presence of infectious B. burgdorferi in the vasculature for as long as 70 minutes after injection of spirochetes did not significantly alter leukocyte adhesion from the baseline levels observed in the absence of B. burgdorferi (1.56−/+0.29 vs 1.75−/+1.15 adhered leukocytes/100 µm, respectively; P = 0.797; N = 50 vessels from 5 mice). The observed number of leukocyte adhesions was normal for the dermal microvasculature of mice [28].
As shown in Fig. 2, the ability to interact with the microvascular endothelium was specific to infectious spirochetes. When mice were injected with non-infectious B. burgdorferi exhibiting the same fluorescence intensity as infectious spirochetes (Fig. 2A and B), transient interactions were reduced by 94% (Fig. 2C). Furthermore, non-infectious B. burgdorferi did not form many dragging interactions and no detectable stationary adhesions (Fig. 2B and C). Non-infectious spirochetes were never observed escaping the microvasculature. These observations indicated that early-stage interaction events were essential for sustained association and vascular escape. These observations also demonstrated that microvascular interactions were dependent on B. burgdorferi proteins expressed only in the infectious strain.
Many bacterially-encoded proteins interact with host cells via GAGs [29], and a number of previous in vitro studies performed under static conditions have found that B. burgdorferi can bind to GAGs and that exogenously applied GAGs can competitively inhibit interaction of B. burgdorferi with cell monolayers [13],[15],[30],[31]. Therefore, we investigated the potential role of endothelial host cell GAGs in spirochete microvascular association in vivo. Interaction rates were first examined in the presence and absence of a therapeutic low molecular weight heparin compound, dalteparin (Fragmin, average molecular weight 5 kDa). Dalteparin was used at a concentration previously shown to block leukocyte rolling in vivo [32]. Dalteparin (200 µl of a 25 I.U./µl solution) was injected via the femoral vein 15 minutes before intravenous inoculation with infectious spirochetes (see Materials and Methods). As shown in Fig. 3A, dalteparin treatment did not cause any significant change in transient interactions between fluorescent, infectious B. burgdorferi and the vascular endothelium. However, dragging interactions were significantly reduced by 72% (Fig. 3B) while the number of stationary adhesions were also reduced by dalteparin treatment to a similar extent (Fig. 3C, 76% of controls).
Similar experiments were performed with dextran sulphate (Fig. 3D–F), a 500 kDa high molecular weight GAG analogue, which interacts in vitro with infectious but not non-infectious B. burgdorferi [30]. In these experiments, dextran sulfate was incubated with infectious B. burgdorferi for 30 minutes, followed by extensive washing, prior to B. burgdorferi administration to the animal, as this compound was toxic when injected directly into the mouse bloodstream. The concentration of dextran sulfate used in preincubations (20 µg/ml) was the dose that has been previously shown in vitro to maximally inhibit B. burgdorferi interaction with endothelial cells without altering spirochete morphology or motility [30],[31]. Preincubation of infectious B. burgdorferi with dextran sulfate caused a slight (30%) reduction in the number of transient interactions, but dragging interactions were reduced by 80% (Fig. 3D and E). A similar reduction in the number of stationary adhesions was also observed (Fig. 3F).
The results from the dalteparin and dextran sulphate experiments indicated that host GAGs play an important role in dragging interactions between B. burgdorferi and the microvascular endothelium in vivo, and that competition with a high molecular weight GAG analogue (dextran sulphate) also inhibited transient interactions. The similar levels of inhibition of dragging interactions and stationary adhesions caused by treatment with GAGs suggested that reductions in stationary adhesion were the result of inhibition of dragging. This in turn implied that additional host and spirochete molecules might contribute to stationary adhesion. However, the results of these experiments alone did not rule out the possibility that GAGs played a role in stationary adhesion.
Many bacterial adhesins can interact with GAGs, either directly, or indirectly via their association with host molecules such as fibronectin, regulators of the complement cascade and components of the coagulation system; furthermore, GAGs can act as bridging molecules that facilitate interactions between pathogen adhesins and host receptors [29]. In an effort to identify spirochete adhesins mediating GAG-dependent microvascular interactions in vivo, we PCR-amplified and sequenced all candidate B. burgdorferi adhesin genes identified to date [16], [17], [19], [21]–[25], using genomic DNA extracted from our fluorescent infectious and non-infectious strains (Table S1). This approach indicated that the genes encoding BBK32, VlsE, OspF, ErpL and ErpK were absent or mutated in the non-infectious strain (Table S1). It is possible that VlsE, OspF, ErpL and ErpK could mediate GAG-dependent host interactions directly or through recruitment of host molecules such as complement cascade regulators; however, interaction of these proteins with GAGs has not been directly demonstrated. In contrast, BBK32 has recently been shown to bind to host GAGs and to rescue the ability of non-infectious B. burgdorferi to interact with endothelial cells in vitro [25]. It was, therefore, of interest to investigate a possible role for BBK32 in B. burgdorferi interactions with the microvasculature in the living mouse. The bbk32 coding sequence, under the control of the ospC promoter [25], was cloned into the GFP expression construct and the resulting plasmid was used to transform the parental non-infectious B. burgdorferi strain. Both parental and complemented strains had the same endogenous plasmid content (data not shown). Expression of BBK32 in the complemented strain was lower than the expression observed in the infectious strain (8.0−/+1.8%), but even this reduced expression was sufficient to restore transient and dragging interactions to the level observed with the infectious strain (Fig. 4A and B). However, stationary adhesion rates in the bbk32 complementation strain did not reach the same levels as in the infectious strain (Fig. 4C), implying either a greater dependence upon BBK32 or a dependence upon additional spirochete factors that were missing in the complemented non-infectious strain. Attempts to genetically disrupt the bbk32 locus in the infectious strain were successful, but did not result in usable constructs due to loss of endogenous plasmids (lp28-1 and others) in all recovered strains.
PCR amplification and sequencing of all candidate B. burgdorferi adhesin genes identified to date also indicated that other candidate adhesin genes (bbf32, bbk2.10, bbO39 and bbm38, encoding VlsE, OspF, ErpL and ErpK, respectively) were absent or mutated in the non-infectious parental strain (Table S1). Hence, these genes were not essential for transient and dragging interactions with the microvasculature of murine skin in vivo, since expression of BBK32 alone in this strain was sufficient to restore transient and dragging interactions. However, the possibility still exists that some of these genes play a role in stationary adhesion.
Examination of the sequence, expression and localization of two other major adhesins, P66 and Bgp, which have been shown to associate with integrins and GAGs respectively under static conditions in vitro, indicated that these proteins were expressed and localized similarly in non-infectious and infectious strains, and were not mutated (Fig. 4D; Table S1). Therefore, neither P66 nor Bgp expression nor localization was sufficient for transient or dragging interactions in the absence of BBK32 expression. Although the genomic sequence of the P66- and Bgp-encoding genes was identical in both infectious and non-infectious strains, it remains possible that secondary mutations elsewhere in the genome of non-infectious B. burgdorferi could have negatively affected transient and dragging interactions.
Because BBK32 binds Fn in addition to GAGs, we also investigated a possible role for Fn in the adhesion of B. burgdorferi to the endothelium in vivo. Rabbit serum, which contains fibronectin, is an important component of the BSK-II medium used to propagate B. burgdorferi; therefore, we investigated whether antibodies to rabbit plasma Fn could disrupt B. burgdorferi microvascular interactions in vivo. Anti-Fn IgGs did not alter spirochete morphology or motility in vitro, implying that they were not toxic to B. burgdorferi. The tethering, dragging and stationary interactions/min for infectious B. burgdorferi treated with αFn IgGs were compared to the interaction rates of untreated spirochetes, and of spirochetes treated with nonspecific goat IgGs (Fig. 5). Preincubation of infectious spirochetes for 20 minutes with the IgG fraction of goat antiserum to rabbit plasma Fn, together with intravenous injection of this IgG fraction into the blood stream of mice, reduced transient and dragging microvascular interactions by 92% and 99%, respectively (Fig. 5A and B). When the same treatment regimen was performed using nonspecific goat IgGs, no effect on interaction rates was observed, indicating that the reduction in interactions following treatment with αFn IgGs was specific. Although stationary adhesions were essentially abolished by the αFn treatment (Fig. 5C), the reduction in transient and dragging interaction rates was so great that we could not determine if stationary adhesion rates were specifically affected by treatment with anti-Fn IgGs. Interestingly, interaction rates returned to normal levels 15–20 minutes after injection of the spirochetes and antibody (data not shown), suggesting that antibody-blocked rabbit Fn bound to spirochetes might have been replaced by mouse Fn in vivo, thus restoring microvascular interactions. The long population doubling time of B. burgdorferi (6–8h) precludes the possibility that restored interaction rates were caused by spirochete replication. The dramatic reduction in transient and dragging interactions resulting from Fn antibody treatment suggested that B. burgdorferi exploits host Fn for these interactions with the host microvasculature in vivo.
Although experiments performed with anti-Fn IgGs suggested that Fn played a major role in the initiation of microvascular interactions, it was possible that IgG-dependent inhibition was partly a result of factors such as steric hindrance of interactions by bulky IgGs. Therefore, we also investigated the Fn dependence of interactions using Fn peptides. Fn is a structurally and functionally complex molecule (reviewed in [33]). Briefly, the N-terminal Type I Fn repeats and gelatin-binding region interact with Fn-binding proteins from B. burgdorferi, Staphylococci and Streptococci in vitro [16],[34],[35]. The central cell-binding domain contains multiple integrin-binding sites, including the canonical RGD sequence, which binds to most integrins that have been implicated in B. burgdorferi-host cell interactions to date [17],[19],[33]. Finally, the Fn C-terminus contains a high affinity heparin-binding domain that also interacts with host cell GAGs [33].
To investigate endothelial cell molecules associating with spirochete-bound Fn, we used peptides derived from the C-terminal heparin-binding domain and the integrin-interacting cell-binding domain in an attempt to block B. burgdorferi-microvascular interactions in vivo (Fig. 6). The heparin domain peptide (FN-C/H II: KNNQKSEPLIGRKKT) inhibits Fn-mediated cell adhesion and heparan sulfate binding [36], and the GRGDS cell-binding domain peptide is a well-studied competitive antagonist of integrin binding [37] that also inhibits B. burgdorferi interactions with integrins αIIbβ3, αvβ3 and α5β1 in vitro [15]. Peptides were injected via the femoral vein immediately before inoculation with infectious spirochetes, at concentrations (∼50 µg/ml of circulating blood) that disrupt leukocyte adhesion and recruitment in vivo [38]. Microvascular interactions in dermal postcapillary venules were recorded for no longer than 20 minutes after injection of peptide as the effect of peptide treatment on interaction rates was diminished at later time points, presumably because linear peptides are rapidly cleared from the mouse circulation [39].
Intravenous injection of 100 µg of the heparin-binding domain peptide reduced transient interaction rates by 52%, and impaired both dragging interactions and stationary adhesion levels by 84%, confirming the role of GAGs in early stages of microvascular interaction (Fig. 6A–C). In contrast, competition with the RGD peptide did not significantly inhibit any class of interaction (Fig. 6D–F), even though the estimated final concentration of this peptide in the mouse circulation (100 µM) was twice as high as the dose known to reduce in vitro B. burgdorferi-integrin interactions by at least 75% in vitro [15]. Administration of twice as much RGD peptide (∼200 µM final concentration) did not inhibit interactions, nor did intravenous administration of anti-CD41 and CD49e antibodies that respectively target RGD-dependent B. burgdorferi-interacting integrins containing αIIb or α5 chains (platelet glycoprotein αIIbβ3 and the α5β1 Fn receptor; data not shown). The effect of treatment with antibodies to αvβ3 was not examined; however, since treatment with RGD peptide in vitro has been shown to strongly inhibit B. burgdorferi binding to this integrin as well as glycoprotein αIIbβ3 and integrin α5β1, it seems unlikely that integrin αvβ3 mediated the early stages of microvascular interactions. The conclusion that RGD-dependent integrins are not required for microvascular recruitment is consistent with the localization of known B. burgdorferi-associating RGD-dependent integrins, which are found at sites of endothelial attachment to extracellular matrix, and not in the lumen [15]. Collectively, these results implied that Fn-dependent transient and dragging interactions in vivo were mediated by host GAGs and not by RGD-dependent integrin interactions.
In this study we employed intravital microscopy, a live cell imaging technique commonly used to analyze leukocyte recruitment and tumor dissemination in situ [40],[41], to investigate the molecular basis of B. burgdorferi dissemination in vivo. Our results demonstrate that IVM can provide critical insight into the mechanisms of pathogen dissemination. Fig. 7 provides a summary of the features of B. burgdorferi dissemination which we have identified using IVM, based on data described in this study and in a recent companion report [9].
This study revealed pivotal roles for the B. burgdorferi adhesin BBK32 as well as host GAGs and fibronectin in the initiation of spirochete-microvascular interactions (see below) (Fig. 7). Although other host and B. burgdorferi molecules may also contribute to microvascular interactions in vivo, the involvement of GAGs and Fn in this process is especially interesting, since a broad array of pathogens are known to interact in vitro with these host molecules in direct binding assays and tissue culture models (reviewed in [29], [42]–[45]. However, the potential contribution of these interactions to processes such as hematogenous dissemination has not been directly examined in living hosts.
We found that BBK32 and its host ligands Fn and GAGs played major roles in transient and dragging interactions. Although we cannot rule out the possibility that these molecules also contribute to stationary adhesion, the results of the bbk32 complementation experiments indicate that additional spirochete molecules are likely required for stationary adhesion. This implies that stationary adhesion is a mechanistically distinct step in B. burgdorferi dissemination. This conclusion is supported by our previous observations that: 1) stationary adhesions form primarily at endothelial junctions, whereas short-term interactions occur chiefly on endothelial cells themselves; and 2) stationary adhesions associate more intimately with the endothelium than short-term interactions and appear to traverse the surface of the endothelium when these cells are labeled with PECAM-1 (Fig. 7) [9]. These observations imply that B. burgdorferi dissemination shares functional similarities with the sequence of events that constitute the leukocyte recruitment cascade [46],[47], as well as the events associated with dissemination of circulating tumor cells [48]. Leukocyte recruitment is initiated by selectin-mediated tethering and rolling interactions that permit firm adhesion, which is mediated by integrins. The initiation phase of leukocyte recruitment is a rate-limiting step, as it is essential for all subsequent events in the recruitment cascade. Similarly, we propose that transient and dragging interactions mediated by GAGs and Fn together constitute the corresponding initiation phase of B. burgdorferi dissemination, while other host and spirochete molecules become essential at the stationary adhesion phase.
Though our data indicated that transient and dragging associations were mediated by the same host and spirochete molecules, the observation that the low molecular weight heparin dalteparin inhibited only dragging interactions was surprising. The reason for this is currently unknown, but may result from differences in total charge, chain length and chemical composition of the carbohydrate moieties.
This study identified a central role for the B. burgdorferi protein BBK32, host GAGs and Fn in the initiation of microvascular interactions. This observation was unexpected, since previous studies have shown that genetic disruption of bbk32 attenuates but does not abolish infectivity [49],[50]. However, bbk32 disruption mutants still bind Fn [49],[50], implying that other functionally redundant Fn-binding proteins in B. burgdorferi might also mediate the initiation of dissemination. The simplest interpretation of our data is that initiation is mediated by BBK32 interactions with GAGs, either independently or via a fibronectin bridge. It is possible that initiation might also be mediated by RGD-independent integrins such as α3β1, which interacts with Fn, GAGs and the B. burgdorferi protein BBB07 [18],[19]; however, this integrin is expressed at endothelial junctions [51] implying that it is more likely to mediate stationary adhesion or extravasation than initiation interactions. Furthermore, the activation of adhesive properties by endothelial integrins generally requires endothelial activation [46], which is not detected in the short time frame of our experiments [9]. Taken together, these data make it unlikely that integrins play a role in the initiation of vascular adhesion.
All molecules to date implicated in tethering under shear force conditions (selectins, von Willebrand factor the E. coli FimH adhesin) interact with sugar-containing ligands [52],[53], suggesting that Fn-dependent or -independent interactions between BBK32 and GAGs might promote B. burgdorferi tethering by a similar mechanism. The affinity of BBK32 for GAGs is unknown, but in the absence of shear forces BBK32 associates with high specificity and probable high affinity to the Fn N-terminus via a tandem β-zipper mechanism shared with Fn-binding proteins of Staphylococcus aureus and Streptococcus pyogenes [34],[35],[54],[55]. In the absence of shear forces, the affinity of plasma Fn for heparin (Kd = 0.1–1.0 µM) is within the affinity range of P- and E-selectins for their ligands (Kd = 1.5 µM and 109 µm, respectively) [56],[57], suggesting that BBK32-, GAG- and Fn-dependent initiation interactions may be mechanistically feasible.
Although under shear stress conditions Fn does not bind to the leukocyte Fn receptor VLA-4, which mediates tethering to endothelial VCAM-1 under flow [58], previous reports indicate that both platelets and Mycobacterium tuberculosis can bind to immobilized Fn in vitro under shear stress conditions that mimic those found in postcapillary venules [59],[60]; interestingly, platelet-Fn interactions are almost completely blocked by treatment with unfractionated or high molecular weight heparin [60]. This suggests the possibility that Fn-dependent tethering interactions entail cooperative GAG binding, a conclusion that is consistent with our observation that expression of the Fn- and GAG-binding BBK32 protein was sufficient to restore initiation interactions to wild-type levels.
Another possibility is that BBK32-induced conformational changes in Fn might facilitate Fn- and GAG-dependent tethering interactions. This hypothesis stems from recent data from the Höök laboratory indicating that BBK32 binding to Fn induces the formation of superfibronectin (S. Prabhakaran and M. Höök, personal communication), a high molecular weight Fn complex that substantially enhances adhesion of cells to Fn by integrin-dependent and independent mechanisms [61]. Further analysis of the precise mechanisms underlying BBK32-, Fn- and GAG-dependent dissemination under shear force conditions will be required.
The results of this study emphasize the importance of directly investigating host-pathogen interactions in a native context where major regulators of interaction such as fluid shear stress are present. The methodology and observations presented here provide the first direct insight into the role of host GAGs, Fn and a B. burgdorferi protein that binds both of these host components, in host microvascular interactions in situ. These results may have broad-reaching implications for our understanding of processes underlying the dissemination of a variety of other bacterial pathogens that interact with Fn and GAGs.
Plasmid pTM170 was constructed by PCR amplification of the PospC-driven bbk32 cassette from pBBK32 [25] with flanking KpnI and FspI sites, using primers B1093 (5′-GGTACCTTAATTTTAGCATATTTGGCTTTG-3′) and B1094 (5′-GGCCTGCGCATTAGTACCAAACGCCATTCTTG-3′). The PCR product was cloned into the GeneJet plasmid, using the Gene Jet blunt cloning kit (MBI Fermentas) to generate pTM169. The KpnI/FspI-digested pTM169 insert was cloned into the KpnI/FspI sites of the GFP-encoding plasmid pTM61 [9] to yield pTM170.
B. burgdorferi strains used in this study were GCB705 (non-infectious strain B31-A transformed with pTM61) [9],[62], GCB726 (infectious B. burgdorferi strain B31 5A4 NP1 transformed with pTM61) [9],[63] and GCB769 (non-infectious B31-A transformed with pTM170). The plasmid content of these strains is noted in Table S1. All strains were grown in BSK-II medium prepared in-house [64]. Electrocompetent B. burgdorferi strains were prepared as described [9],[65]. Liquid plating transformations were performed with 50 µg pTM61 or pTM170 in the presence of 100 µg/ml gentamycin as described [66],[67]. Gentamycin-resistant B. burgdorferi clones were screened for: 1) the presence of aacC1 sequences by colony screening PCR performed with primers B348 and B349 as described [68]; 2) GFP expression by conventional epifluorescence microscopy, and 3) BBK32 expression, as detected by immunoblotting (for bbk32 complementation strains). The presence of plasmids in non-integrated form in fluorescent strains was confirmed by agarose gel electrophoresis of total genomic DNA prepared on a small scale as described [69]. PCR screening for native plasmid content was performed as described [68],[70].
Gene sequences were amplified from genomic DNA preparations of GCB705 and GCB726. PCR was performed with Phusion DNA polymerase (NEB, Pickering, Ontario, Canada), according to manufacturer's instructions. Sequencing was performed by the University of Calgary DNA Services. All primers used for PCR amplification and sequencing are provided in Table S1.
The expression and outer membrane localization of adhesins P66 and Bgp were analyzed as previously described [16],[25]. Briefly, for each strain, two pellets containing 5×107 spirochetes were washed twice with PBS+2% BSA. PBS containing 5 mM MgCl2 was added without dislodging pellets. Proteinase K was added to one pellet to a concentration of 4 mg/ml. After 30 min incubation at room temperature, reactions were stopped with 150 µg phenylmethylsulfonyl fluoride, spirochetes were pelleted and washed twice with PBS+0.2% BSA, and pellets were lysed using SDS-PAGE loading dye. Proteins were resolved by electrophoresis on 12% SDS-PAGE gels, transferred to nitrocellulose membranes, followed by immunoblotting with antibodies to P66, Bgp or BBK32, as previously described [17],[22].
These conditions have been described in detail previously [9]. Quantification of spirochete interactions was performed as recently described [9]. All animal studies were carried out in accordance with the guidelines of the University of Calgary Animal Research Centre.
Leukocyte recruitment studies were carried out as previously described [28]. Briefly, animals were injected with 50 µl of 0.05% (i.v.) rhodamine 6G (Sigma-Aldrich). Fluorescence was visualized by epi-illumination using 510 and 560 filters. Leukocytes were considered adherent to the venular endothelium if they remained stationary for 30 s or longer. Experiments were performed in mice that had been intravenously inoculated with 4×108 infectious B. burgdorferi grown for 48h in 1% mouse blood, as previously described, and also with mice that were not inoculated with spirochetes. Leukocyte adhesions were counted in the dermal postcapillary venules of infected and non-infected mice from 5 minutes after injection of spirochetes and/or rhodamine until at least 1 hour from injection, in order to monitor leukocyte recruitment during the time frame that is used for all experiments reported in this study.
GCB726 spirochetes prepared as described above were resuspended to 2×109/ml in PBS. The IgG fraction of polyclonal goat anti-rabbit plasma Fn serum or non-specific goat IgGs (Cappel/MP Biomedicals, Solon, OH) were added to 1 mg/ml final. After mixing for 30 min at room temperature, spirochetes were directly injected into the mouse bloodstream, along with 2 mg of Fn antiserum IgGs or non-specific IgGs.
GCB726 spirochetes were prepared and injected as described above, together with 100 µg of GRGDS or FN-C/H II peptide (Sigma Canada, Oakville, ON; catalogue numbers G4391 and F7049, respectively), injected via the femoral vein. Peptides injected at this amount are known to disrupt leukocyte adhesion and recruitment in vivo [38].
Two hundred µl of a 25 I.U./µl solution of dalteparin (Fragmin: Pfizer Canada, Kirkland, PQ) were injected via the femoral vein 15 minutes before intravenous inoculation with infectious spirochetes. This concentration has previously been shown to inhibit leukocyte rolling in vivo [32]. Dextran sulfate-treated spirochetes were incubated with 20 µg/ml dextran sulfate (500 kDa; Fisher Scientific Canada, Ottawa, ON) in a final volume of 100 ml PBS for 30 min at RT°C, followed by 2 100ml washes with PBS. Spirochetes were resuspended to 2×109/ml in PBS, and injected as previously described [9]. The concentration of dextran sulfate used in these preincubations (20 µg/ml) is the dose that maximally inhibits B. burgdorferi interaction with endothelial cells in vitro, and does not affect spirochete morphology or motility [30],[31].
One hundred µg anti-CD41 monoclonal Ab (Clone MwReg30; Becton Dickinson, San Diego, CA), or 20 µg CD49e monoclonal Ab (clone 5H10-27; Pharmingen, Oxford, UK) were intravenously administered immediately prior to injection of spirochetes. These quantities of anti-CD41 and CD49e antibodies are those that respectively protect against Plasmodium berghei infection in vivo [71], and which inhibit neutrophil migration in vivo [72].
For quantitative analysis, average and standard error values for different variables were calculated and plotted graphically for all vessels from all mice using GraphPad Prism 4.03 (GraphPad Software, Inc., San Diego, CA). Statistical significance was calculated in GraphPad Prism using a two-tailed non-parametric Mann Whitney t-test with a 95% confidence interval.
|
10.1371/journal.pbio.3000091 | Non-proteolytic ubiquitin modification of PPARγ by Smurf1 protects the liver from steatosis | Nonalcoholic fatty liver disease (NAFLD) is characterized by abnormal accumulation of triglycerides (TG) in the liver and other metabolic syndrome symptoms, but its molecular genetic causes are not completely understood. Here, we show that mice deficient for ubiquitin ligase (E3) Smad ubiquitin regulatory factor 1 (Smurf1) spontaneously develop hepatic steatosis as they age and exhibit the exacerbated phenotype under a high-fat diet (HFD). Our data indicate that loss of Smurf1 up-regulates the expression of peroxisome proliferator-activated receptor γ (PPARγ) and its target genes involved in lipid synthesis and fatty acid uptake. We further show that PPARγ is a direct substrate of Smurf1-mediated non-proteolytic lysine 63 (K63)-linked ubiquitin modification that suppresses its transcriptional activity, and treatment of Smurf1-deficient mice with a PPARγ antagonist, GW9662, completely reversed the lipid accumulation in the liver. Finally, we demonstrate an inverse correlation of low SMURF1 expression to high body mass index (BMI) values in human patients, thus revealing a new role of SMURF1 in NAFLD pathogenesis.
| Nonalcoholic fatty liver disease (NAFLD) is a disease associated with abnormal fat accumulation in the liver and other metabolic symptoms. Among its many social–behavioral and genetic causes, dysregulation of peroxisome proliferator-activated receptor γ (PPARγ) is an investigative focal point for therapeutic intervention. This lipid-sensing nuclear receptor plays a major role in promoting lipogenesis in adipose tissues, whereas its expression is low in the liver. We show here that in the absence of ubiquitin ligase (E3) Smurf1, PPARγ expression increases dramatically in the liver, causing fatty acid uptake and fat accumulation in hepatocytes. We also found that the low SMURF1 expression in human populations correlates with high body mass index (BMI) values. We demonstrate that Smurf1 catalyzes the lysine 63 (K63)-linked non-proteolytic modification of PPARγ that suppresses the transcriptional activity of PPARγ and breaks the positive feedback loop governing its own expression. Our data further indicate that treating this mouse model with a PPARγ antagonist, GW9662, completely reverses the fat accumulation in the liver.
| Nonalcoholic fatty liver disease (NAFLD) is a chronic liver condition associated with obesity, non–insulin-dependent diabetes, and hyperglyceridemia [1]. Although presenting few clinical symptoms at early stages, a subset of patients with NAFLD will progress to nonalcoholic steatohepatitis (NASH) consisting of hepatic steatosis and inflammation, which can ultimately lead to cirrhosis and even liver cancer [2]. Myriad social–behavioral and genetic causes of NAFLD are now known, but the roles of peroxisome proliferator-activated receptors (PPARs) have emerged as crucial molecular underpinnings of these metabolic imbalances and targets of several investigational drugs [3–5]. A thorough understanding of regulatory mechanisms governing PPAR activities will undoubtedly aid in the development of much-needed treatments.
PPARs are nuclear hormone receptors that heterodimerize with retinoid X receptors to modulate metabolic transcriptional programs in response to nutritional inputs [6]. Of three PPARs encoded by distinct mammalian genes, PPARα, which is highly expressed in the liver, kidney, and muscle, directs expression of a network of genes that promote utilization of fat as an energy source. PPARγ, on the other hand, is normally expressed in adipose tissues, where it activates target genes involved in fatty acid uptake, transport, and lipogenesis to promote lipid storage. In the liver, PPARγ expression is normally low but becomes drastically induced as hepatic steatosis develops [7]. Reports in the literature have shown that overexpression of PPARγ promotes the accumulation of lipid droplets in the liver, whereas hepatic disruption of PPARγ improves the fatty liver condition in leptin-deficient obese mice or mice that were fed on a high-fat diet (HFD) [8, 9]. In adipose tissues, ligand binding was reported to induce degradation of PPARγ via the ubiquitin-proteasome system, whereas small ubiquitin-like modifier (SUMO)ylation of PPARγ was shown to repress its transcriptional activity [10]. However, how steatogenic activities of PPARγ are regulated in the liver remains to be determined.
Smad ubiquitin regulatory factor 1 (Smurf1) and its close relative, Smurf2, are members of homologous to E6-AP carboxyl terminus (HECT) domain–containing ubiquitin ligases (E3s), which were initially identified as negative regulators of transforming growth factor-β (TGF-β) and bone morphogenetic protein (BMP) signaling pathways [11–14]. Subsequent studies broadened the repertoire of Smurf substrates and extended their function to cell differentiation, polarity, and DNA repair [15–18]. During our ongoing quest for physiological functions of Smurfs, we found abnormal accumulation of lipid droplets in the livers of 9–12-month-old Smurf1 knockout (KO) mice and other signs that phenocopy NAFLD in human patients. Here, we report that Smurf1 induces non-proteolytic ubiquitination of PPARγ and inhibits PPARγ transcriptional activity in hepatocytes, thereby acting as a critical safeguard against the development of hepatic steatosis.
We previously reported an increased bone density phenotype in aged Smurf1KO mice that were commonly observed under mixed black Swiss × 129/SvEv (BL) and C57BL/6N (B6) genetic background [18]. Further analysis revealed a conspicuous accumulation of lipid droplets in the livers of aged Smurf1KO mice that was unique to the BL background (S1 Table). The liver sections of these mice were characterized by large, clear, sharp-bordered cytoplasmic vacuoles upon hematoxylin–eosin (HE) staining (Fig 1A). The bright red staining of frozen sections by Oil Red-O confirmed the high fat and neutral lipid content therein (Fig 1A). This phenotype was observed in 12 out of 15 male and female mice examined beyond 9 months of age, implying a 75% penetrance. Microscopic quantification of HE-stained sections reaffirmed the statistically significant increase of steatosis in the livers of Smurf1KO mice compared with that of the wild-type (WT) controls (Fig 1B). Surprisingly, this steatosis phenotype was not observed in the livers of Smurf2KO mice (Fig 1A and 1B), suggesting that it is specifically associated with disruption of the Smurf1 function. To determine which lipid fractions were increased, we carried out colorimetric assays in liver lipid extracts prepared from Smurf1KO mice at 9–12 months of age, and the results showed that the level of triglycerides (TG) increased more than 3-fold compared with that of WT or Smurf2KO mice (Fig 1C). Moreover, the levels of total cholesterol (CHO) and free fatty acids (FFAs) were also increased significantly in Smurf1KO livers (Fig 1C). Compared with WT mice, Smurf1KO mice were approximately 30% heavier in body weight, bore more white adipose tissue (WAT), and had a higher liver to body weight ratio (Fig 1D). Nevertheless, despite exhibiting ostensible steatosis, the mutant livers appeared to function normally, as indicated by aspartate transaminase (AST) and alanine transaminase (ALT) activity measurements (Fig 1E). Because the manifestation of hepatic steatosis is usually accompanied by a constellation of adverse alterations in glucose metabolism, we conducted glucose tolerance and insulin resistance tests. At the fasting state, there was not much difference in plasma glucose levels between aged (9–12 months old) WT and Smurf1KO mice that had developed steatosis; however, following intraperitoneal injection of glucose, the blood glucose level of the mutant mice showed a more dramatic flash increase of the blood glucose level within 30 minutes of injection and more than 100% accumulative gain in the area under the curve (AUC) (Fig 1F). On the other hand, after an initial dip following the insulin injection, the blood glucose level in aged mutant mice recovered more rapidly and to a higher extent than that in WT controls (Fig 1G). The AUC of the insulin resistance test of aged Smurf1KO mice was 13.5% more compared with that of WT mice. Because young Smurf1KO mice (at 4–5 months of age) that had yet to develop steatosis scored no difference from their WT counterparts in both the tests (S1 Fig), the systemic change in glucose metabolism observed in aged mutant mice was most likely associated with the steatosis. Taken together, the phenotypes of hepatic steatosis, obesesity, glucose intolerance, and insulin resistance make these aged Smurf1KO mice a good mouse model of NAFLD.
In rodents, difference in genetic background has a well-known influence on the susceptibility to obesity and hepatic steatosis [19–21]. Although the spontaneous steatosis hereto described was only observed at old age, young Smurf1KO mice of the BL background were grossly normal except for a higher body fat content compared with their age- and background-matched WT counterparts and showed no sign of steatosis (Fig 2A and 2B) when fed on normal diet (ND). Mice of this strain background are notoriously known for their resistance to HFD-induced obesity, as evident by the lack of apparent gain in body weight and ratio of fat-to-lean mass in young WT mice that were put on a HFD feeding regimen beginning at 10–12 weeks of age and continuing for 8 consecutive weeks (Fig 2A, S2 Table). In contrast, HFD feeding significantly increased fat content in the Smurf1KO mice (Fig 2A, S2 Table). Despite a lack of significant weight gain, HFD feeding did cause mild steatosis (Fig 2A and 2B), as well as an increase in liver TG content in WT mice (Fig 2C); however, these changes were all dramatically exacerbated in BL-Smurf1KO mice (Fig 2A–2C).
As alluded earlier, Smurf1KO mice of the B6 background did not show accumulation of lipid droplets in the liver (S1 Table), and they were not overweight or overtly obese either (S2A Fig). To ascertain that the steatosis pheneotype was not a mere coincidence unique to the BL background, we carried out the HFD feeding study on WT and Smurf1KO mice of the B6 background with the same regimen as for the young BL mice. In contrast to BL mice, B6 mice gained body weight and fat content on HFD as expected, regardless of the presence of Smurf1 gene (S2A Fig, S2 Table). However, the B6-Smurf1KO mice on HFD became ostensibly more obese (S2A Fig) and showed more severe lipid droplet accumulation in the liver compared to WT mice of the same background (S2A and S2B Fig). In addition, the increase in the liver TG content was also more pronounced (S2C Fig). Thus, the steatosis associated with Smurf1 loss is likely the result of an overall gain in body fat content in both strain backgrounds, suggesting that Smurf1 may have a systemic role in regulating lipid metabolism.
To address if what we learned from the Smurf1KO mice is applicable to human populations, we took the advantage of the non-tumor liver tissue data sets compiled from a cohort of 247 Chinese liver cancer patients from the Liver Cancer Institute (LCI) [22]. According to the SMURF1 mRNA expression levels retrieved from the gene expression profile (GEO: GSE14520), we separated non-tumor liver tissue samples into the high SMURF1 expression (top 25%) group (n = 61) and the low SMURF1 expression (bottom 25%) group (n = 59) (Fig 2D, left panel). We then graphed the body mass index (BMI) of these 120 patients against these two groups of SMURF1 expression, and found that patients with the low SMURF1 expression have a statistically significant higher BMI (Fig 2D, right panel). It is worth noting that the average BMI of the Asian population is lower than that in the United States and European countries, and an Asian with BMI > 27.5 is considered obese [23, 24]. This inverse correlation was further corroborated with non-tumor liver tissue data sets from the cancer genome atlas-liver hepatocellular carcinoma (TCGA-LIHC) (Fig 2E). Because there are only 37 cases of non-tumor liver samples that have linked BMI values in the TCGA data set, the median SMURF1 expression level was used as the cutoff to plot BMI values (Fig 2E). Because the BMI is widely used in clinics as a surrogate prognostic indicator for fatty liver [25, 26], these results suggest that low Smurf1 expression appears to be associated with high fat accumulation in humans, as well.
To investigate the underlying causes of steatosis associated with Smurf1 loss, we compared hepatic gene expression profiles of 11-month-old Smurf1KO, Smurf2KO, and their respective matching WT mice from the BL background, and selected genes that showed either increased or decreased expression by a cutoff of 1.5-fold (false discover rate [FDR] <0.1). The results showed that 987 genes in the Smurf1KO livers were differentially expressed over their WT controls, whereas only 13 genes were differentially expressed in the Smurf2KO livers (Fig 3A, left panel, and S3 Table). This result is in line with the notion that Smurf1 plays a more prominent role in the liver than Smurf2. Many genes that are involved in the lipid metabolism were up-regulated in Smurf1KO livers, and the Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway enrichment analysis of differentially expressed genes between Smurf1KO and WT livers revealed a number of metabolically relevant pathways (Fig 3B). We were intrigued by the enrichment of the PPAR signaling pathway that has known strong effects on steatosis [4]. Of the three PPAR genes, Pparγ encodes two protein isoforms, PPARγ1 and PPARγ2, whose mRNAs are transcribed from two separate promoters [27, 28]. Quantitative real-time PCR (qRT-PCR) analyses showed severalfold increases of both Pparγ isoforms in the livers of aged Smurf1KO but not Smurf2KO mice (Fig 3C). Interestingly, the expression of Pparα was not altered in the liver of any mouse examined (Fig 3C). In young BL mice (10–12 weeks of age) that had yet to develop steatosis, loss of Smurf1 increased the expression of total Pparγ (about 1.57-fold) when the mice were fed on ND (S3A Fig), suggesting that Smurf1 has a direct causal effect on Pparγ expression. HFD feeding further exacerbated the difference of Pparγ expression to 3.42-fold between WT and Smurf1KO livers (S3A Fig). On the other hand, no difference was observed in TNFα and F4/80 expression (S3B Fig), two genes involved in inflammatory response, which is consistent with the absence of any inflammation in Smurf1KO mice (S1 Table and S1 Data). Western blot analyses confirmed the corresponding up-regulation of the PPARγ protein in the livers of aged Smurf1KO mice (Fig 3D). According to data from The Human Protein Atlas (https://www.proteinatlas.org/ENSG00000198742-SMURF1/tissue), Smurf1 protein is highly expressed in visceral organs, but its expression levels in muscle and adipose tissues are extremely low or moderate, respectively. This likely accounts for the dramatic increase of PPARγ in the Smurf1KO livers, where Smurf1 function is expected to be robust. Consistent with tissue distribution of Smurf1 expression, qRT-PCR revealed that total Pparγ expression increased dramatically in the liver and WAT but did not change in the muscle of Smurf1KO mice (Fig 3E). Finally, loss of Smurf1 cast a profound impact on the hepatic expression of PPARγ transcriptional target genes that are involved in fatty acid synthesis, uptake, and transport (Fig 3F), thus lending further support to the activation of PPARγ and its signaling pathway in aged Smurf1KO mice.
PPARγ is a strong lipogenic factor essential for steatosis [7]. Although our qRT-PCR analysis alluded that loss of Smurf1 has a direct causal effect on PPARγ1 up-regulation, further evidence is needed to confirm this finding. Toward this end, we silenced Smurf1 using short interfering RNA (siRNA)s in human hepatocarcinoma Hep3B cells and mouse normal hepatocyte AML12 cells. Relative to the effect by non-silencing control siRNA (siNS), knockdown by siSmurf1 significantly increased the level of PPARγ but not PPARα or PPARδ in both cell lines (Fig 4A). As expected, siSmurf2 had little effect in either of these two cell lines (Fig 4A). Because Pparγ is a direct transcriptional target of itself in a positive feedback loop [29], siRNA-mediated silencing of Smurf1 drastically increased the expression level of Pparγ, but not other paralogous Ppars or their transcriptional regulatory partners retinoid x receptor (Rxr)α and Rxrβ (Fig 4B). In adipose tissues, transcription of Pparγ genes is under the control of CCAAT enhancer binding protein (CEBP)α/β [30, 31]; however, we were unable to detect any increase of either Cebpα or Cebpβ mRNA by qRT-PCR (S4A Fig), suggesting that the regulation of PPARγ by Smurf1 is by way of a C/EBPα/β-independent mechanism. In line with the low expression of PPARγ in AML12 cells, introducing siPPARγ showed little effect on the expression of PPARγ transcriptional target genes, Fabp1, Cd36, Acacb, and Apoc3, but siSmurf1 significantly increased the expression of these genes (Fig 4C, and S4B Fig). Furthermore, introducing siPPARγ completely blocked the enhancing effect of siSmurf1 (Fig 4C), thus confirming the direct causal relationship between Smurf1 and PPARγ. The fact that up-regulation caused by siSmurf1was particularly pronounced in Fabp1 and Cd36, two genes that are essential for fatty acid uptake [32, 33], suggested a strong connection between Smurf1 and fatty acid uptake. Indeed, using 3H-labelled palmitic acid as a tracer, we observed a 20% increase in fatty acid uptake by AML12 cells upon Smurf1 depletion (Fig 4D). We also measured lipid synthesis in AML12 cells by measuring the incorporation of 3H-labelled acetate into lipids and found it was increased by siSmurf1 as well (Fig 4E). Once again, these two effects of Smurf1 loss were specifically mediated through PPARγ as they were reversed by siPPARγ (Fig 4D and 4E).
To further show if Smurf1 actually regulates lipid metablism in vivo, we injected fluorescent 4,4-Difluoro-5,7-Dimethyl-4-Bora-3a,4a-Diaza-s-Indacene-3-Hexadecanoic Acid (BODIPY-FL-C16) into the peritoneal cavities of WT and Smurf1KO mice and found that the fatty acid uptake was greatly enhanced in the liver and WAT tissues but not the muscles of Smurf1KO mice compared with that of WT mice (Fig 4F). We also repeated the 3H-labelled acetate incorporation experiment in primary hepatocytes isolated from WT and Smurf1KO mice and confirmed the enhancement effect of Smurf1 ablation on lipid synthesis (Fig 4G).
The increased body fat content in aged BL-Smurf1KO mice and HFD-fed young Smurf1KO mice from both background suggests that Smurf1 may also regulate adipogenesis. To determine if this was the case, we took advantage of an in vitro adipogenic differentiation system using 3T3-L1 pre-adipocytes. Following a 6-day differentiation protocol, both PPARγ1 and PPARγ2 as well as their target Cd36 were all induced, as shown by western blot analysis, and the induction was greatly enhanced by siSmurf1 but reversed by the double transfection of siSmurf1 and siPPARγ (S5A Fig). In keeping with the western blot analysis results, Oil Red-O staining of these differentiated 3T3-L1 cells was also enhanced by siSmurf1 and reversed by siSmurf1 and siPPARγ double transfection (S5B Fig). Finally, expression of a cohort of adipogenic target genes of PPARγ also followed the same pattern as influenced by siSmurf1 and siPPARγ (S5C Fig). Taken together, these data indicate that Smurf1 has an intrinsic role in controlling adipogenesis and lipid metabolism through PPARγ.
The WW domains of HECT E3 ligases recognize a PPxY (PY) motif that is present in the primary sequence of many of their targets [34]. There is one such sequence motif in both human and mouse PPARγ but not in PPARα, which might potentially account for the lack of an effect on this closely related protein by the loss of Smurf1 (Fig 3C). By co-immunoprecipitation experiments, we found that endogenous Smurf1 interacted specifically with PPARγ in the AML12 cells (Fig 5A), and the PY motif of PPARγ contributed to the interaction, because removing it considerably weakened the interaction between Myc-tagged Smurf1 and FLAG-tagged PPARγ, as assayed in transiently transfected AML12 cells (Fig 5B). Also in AML12 cells, Smurf1 but not Smurf2 showed the propensity to ubiquitinate both PPARγ1 and PPARγ2 isoforms (Fig 5C). The substrate and enzyme relationship was further demonstrated in Smurf1KO mouse embryonic fibroblasts (MEFs), in which exogenous Smurf1 but not the catalytically inactive Smurf1 C699A (CA) mutant ubiquitinated PPARγ (Fig 5D), as well as in a reconstituted in vitro reaction with recombinant Smurf1 and PPARγ (Fig 5E). Finally, the ubiquitin chain of the modified PPARγ is likely of the K63 linkage, as only the ubiquitin mutant with a single lysine residue at the amino acid residue position 63 supported the polyubiquitination of PPARγ in the reconstituted in vitro reaction, whereas other single-lysine ubiquitin mutants with lysine at other positions did not (Fig 5F). In light of this result and the fact that co-expressing Smurf1 with PPARγ did not alter the stability of the latter (Fig 5B and 5C), we concluded that Smurf1 mediates a non-proteolytic ubiquitin modification of PPARγ.
PPARs recognize a consensus sequence of PPAR response element (PPRE) that consists of two AGGTCA-like sequences arranged in tandem with a single nucleotide spacer and is present in all PPAR target gene promoters [35, 36]. In AML12 cells, where PPARγ expression is very low, overexpressing Smurf1 had little effect on a luciferase reporter driven by PPRE, whereas both PPARγ1 and PPARγ2 significantly activated it; however, co-expressing Smurf1 with either PPARγ1 or PPARγ2 severely curtailed their transcriptional activity (Fig 6A). Because Smurf1 has no effect on PPARγ protein levels per se (Fig 6A, right panel), these results suggested that Smurf1 inhibits the transcriptional activity of PPARγ. The regulation by Smurf1 depends on its E3 ligase activity because a ligase-deficient mutant, Smurf1CA, could not reverse the activation of PPRE-luc by PPARγ (Fig 6B). Chromatin immunoprecipitation (ChIP) experiments on Pparγ1, Pparγ2, and Fabp1 promoters indicated that the binding of PPARγ to these promoters was blocked when it was co-expressed with Smurf1 (Fig 6C). Once again, the E3 ligase activity of Smurf1 is required for its ability to block DNA binding of PPARγ (Fig 6C). ChIP experiments performed in liver extracts isolated from WT and Smurf1KO mice also revealed a much stronger binding of PPARγ to its own Pparγ1 and Pparγ2 promoters, as well as its target Fabp1 promoter in the absence of Smurf1 (Fig 6D), thus lending further support to Smurf1 regulating transcriptional activity of PPARγ.
To directly test if the increased PPARγ activity and expression are responsible for steatosis associated with Smurf1 loss, we treated a group of WT and Smurf1KO mice from the BL background with the PPARγ antagonist GW9662 [37]. The compound was administered by intraperitoneal injection starting at 7–9 months of age, and the treatment lasted for 2 months; in this time period, the steatosis was expected to fully develop in Smurf1KO mice. The GW9662 treatment decreased body weight of both WT and Smurf1KO mice (Fig 7A), but because the average beginning weight of Smurf1KO mice was higher, the reduction thereof was more dramatic than that of the WT controls (about 10% reduction versus about 5%). The body fat mass content in Smurf1KO mice was also significantly lowered, to an extent that was comparable to that of the untreated WT mice (Fig 7B). Commensurate to the systemic reduction in obesity, the lipid droplets were essentially cleared from Smurf1KO livers by GW9662 (Fig 7B and 7C). Although the GW9662 treatment caused no significant change in the serum TG and CHO levels (Fig 7D), hepatic contents of TG, CHO, and FFA were all reduced to normal levels (Fig 7E) and so was hepatic expression of Pparγ2, as well as several PPARγ target genes (Fig 7F). These results unequivocally demonstrated that the elevated PPARγ activity and expression account for the NAFLD phenotypes observed in Smurf1KO mice.
PPARγ is a nuclear hormone receptor with principle functions of increasing insulin sensitivity and promoting lipid storage in adipose tissues [6]. In the liver, the physiological function of PPARγ is less clear, although its expression is associated with injury-induced activation of hepatostellate cells and provides an anti-fibrogenic protection [3, 5]. PPARγ up-regulation is also a known property of steatotic livers, and liver-specific disruption of PPARγ was reported to protect leptin-deficient mice or HFD-fed mice from developing fatty liver [7–9]. Here, we show that mice deficient for HECT-domain E3 ligase Smurf1 in the mixed BL genetic background develop hepatosteatosis spontaneously as they age or are more susceptible to HFD-induced hepatosteatosis. These mutant mice are overweight and obese, as well as glucose intolerant and insulin resistant. These NAFLD phenotypes can be attributed to the heightened transcriptional activity of PPARγ, which in turn increases the expression of itself and genes involved in lipogenesis and fatty acid transport via a positive feedback loop. We further demonstrate that Smurf1 catalyzes the K63-linked non-proteolytic ubiquitination that normally attenuates PPARγ transcriptional activity, and show an inverse correlation between low SMURF1 expression and high BMI values in human patients. This investigation thus reveals a previously unknown mechanism that regulates the lipogenic activity of PPARγ and sheds light on a new role of Smurf1 in NAFLD pathogenesis.
Different HECT E3 ligases are known to catalyze ubiquitination with different ubiquitin chains that mark modified protein substrates for different fates [38]. Members of the neural precursor cell expressed developmentally down-regulated protein 4 (NEDD4) family E3 ligases preferentially support monoubiquitin modification or K63-linked chains associated with non-proteolytic functions, but can also assemble lysine 48 (K48)-linked chains that target proteins for proteasome-mediated degradation [39]. As members of this E3 ligase family, Smurf1 and Smurf2 have been shown to target many proteins for K48-linked ubiquitination and degradation [16]. Smurf2 was also shown to induce multi-monoubiquitin modification of Smad3, thereby inhibiting Smad3 activity [40], but the K63-linked ubiquitination by Smurfs has not been reported in mammalian species. Recently, NEDD4 itself was shown to induce both K48- and K63-linked ubiquitination of PPARγ in adipocytes, with different functional outcomes [41]. In our study, Smurf1 inhibits PPARγ activity, and deletion of Smurf1 enhances PPARγ activity and up-regulates PPARγ levels through a positive feedback mechanism. In contrast, NEDD4 was shown to stabilize PPARγ, and knockdown of NEDD4 reduced PPARγ expression [41]. Moreover, the PY motif in PPARγ played a role in mediating interaction with Smurf1, but it was not demonstrated for NEDD4 as such. Perhaps these apparent discrepancies reflect the differences in experimental conditions conducted in different cell types, or the mixed linkages in ubiquitin chains formed by NEDD4 could have compounded the functions of modified PPARγ. In any event, the steatosis observed in Smurf1KO mice is consistent with the heightened PPARγ activity in the liver. Despite the conspicuous steatosis, overweight, and obesity that were present in 75% aged BL-Smurf1KO mice, their liver functions were nevertheless normal. Because these animals were well shielded from inflammatory insults by their accommodative housing facility, it is likely that the elevated PPARγ activity unleashed by the loss of Smurf1 was only sufficient to manifest a restricted impact in bringing about the early-stage NAFLD phenotypes. Future studies are necessary to ascertain the tissue origin of the steatogenic effect of Smurf1 ablation using conditional knockout approaches and to determine if and how BL-Smurf1 mice could be enticed to progress through NASH or even liver cancer to model the entire NAFLD disease spectrum. Regulation of PPARγ by the Smurf1-mediated K63-linked ubiquitin modification centers on its transcriptional activity. Because PPARγ is also a transcriptional target of its own, a disturbance of Smurf1 would create an “all or none” effect: a rise or fall of Smurf1 across a threshold level would either maximize or minimize PPARγ activity. This scenario may normally operate to keep the lipogenic activity of PPARγ to a minimum in the liver but maximized in the adipose tissues.
Epidemiology studies indicate that an estimated 27%–34% of the general population within North America have NAFLD [42], for which there is no approved treatment available at present. Current NAFLD drug developmental effort centers on repurposing fibric acid derivatives, which are lipid-lowering PPARα agonists and insulin sensitivity–improving PPARγ agonists, thiazolidinediones, but the clinical trials yielded mixed results [3, 4]. Because of the opposite actions of PPARα and PPARγ on hepatic steatosis, the “spillover” effects of these PPAR agonists might prevent a net gain in their ability to reduce TG accumulation in the liver. As to PPARγ agonists, although clinical trials for rosiglitazone in patients with type 2 diabetes reported improvement of steatosis by a median of 20% during the first year, no further improvement was found after 2 additional years of treatment, and the trials exposed severe cardiovascular risks and weight gain [43]. Intuitively, it is possible that the benefit is derived from the systemic lipid clearance by increased fat storage in adipose tissue, because PPARγ is normally expressed in adipose tissues, and its activation in the liver was clearly linked to fatty liver formation. Given our current finding of Smurf1 in protecting the liver from steatosis, a viable strategy to treat NAFLD may be to curtail the transcriptional activity of PPARγ by turning on Smurf1-mediated non-proteolytic ubiquitin modification.
All mice were maintained and handled under protocols (LCMB-014, ASP 10–214, 13–214, 16–214) approved by the Animal Care and Use Committee of the National Cancer Institute, National Institutes of Health (NIH), according to NIH guidelines.
Generation of Smurf1KO and Smurf2KO mice in the mixed BL and pure C57BL/6N (B6) background was described previously [18, 40]. For spontaneous hepatosteosis development, animals were maintained on a ND, monitored weekly, and euthanized and necropsied at 9–12 months of age. For the HFD treatment, male mice were maintained on a ND until 10–12 weeks of age before they were given HFD (Research Diets, Cat# D12266B) containing 16.8% kcal protein, 31.8% kcal fat, and 51.4% kcal carbohydrate for 8 weeks. For the GW9662 treatment, a dose of 1 mg/kg of GW9662 dissolved in DMSO was injected intraperitoneally (i.p.) into 7–9-month-old BL-WT and BL-Smurf1KO mice for 5 days per week for 2 months. Age and sex of mice used in these studies are listed in S1 Table.
Body composition was determined using an EchoMRI mouse scanner (EchoMRI, Houston, TX). Mouse liver and epididymal fat pad were dissected, weighed, then either snap-frozen in liquid N2 or fixed in 10% neutral buffered formalin prior to paraffin embedding. Frozen liver tissues were used for Oil Red-O staining. Liver and fat tissue histology were read by board-certified veterinary pathologists in the Pathology and Histotechnology Laboratory of the Frederick National Laboratory for Cancer Research.
Serum TG, CHO, and albumin concentrations, as well as ALT and AST activities were measured by standard methods with a Vitro 250 dry slide analyzer (Ortho Clinical Diagnostics) in the Pathology and Histotechnology Laboratory of the Frederick National Laboratory for Cancer Research. Liver TG, CHO, and FFA concentrations were determined using the EnzyChrom TG, CHO, and FFA assay kits (Bioassay Systems) after extracting total lipids from 50-mg liver tissues as described [44].
To perform the gluclose tolerance test (GTT) or insulin tolerance test (ITT), mice were fasted overnight before receiving an i.p. injection of 20% glucose (2 g/kg body weight) or recombinant insulin (Humulin R, 0.75 U/kg; Lily), respectively. Blood samples were collected from the tail 0, 0.5, 1, 2, and 4 hours later, after injection for analysis using the Accu-Chek Compact Plus blood glucose meter (Roche Diagnostics).
AML12 cells (ATCC CRL-2254) were cultured in DMEM/F12 supplemented with 10% fetal bovine serum (FBS), 0.005 mg/mL insulin, 0.005 mg/mL transferrin, 5 ng/mL selenium, and 40 ng/mL dexamethasone. Hep3B cells were cultured in MEM supplemented with 1% Non-Essential Amino Acids (NEAA) and 10% FBS. Smurf1KO MEFs were cultured in DMEM supplemented with 10% FBS. Primary hepatocytes were isolated by a two-step collagenase perfusion of the liver and cultured as described [45]. Flag-tagged PPARγ1, PPARγ2 plasmids, and PPRE-Luc reporter plasmids were obtained from Addgene. Flag-tagged PPARγ2ΔPY plasmid was generated using Site Directed Mutagenesis Kit (Agilent Technologies). Myc-tagged Smurf1, Smurf2, and Smurf1CA mutant, HA-tagged Ubiquitin plasmids were described before [13, 18, 46]. Anti-Smurf1 (Novus, 1D7); anti-Smurf2 (Abcam, EP629Y3); anti-PPARγ (Santa Cruz, sc-7273); anti-PPARα (Rockland, 600-401-4215); anti-PPARδ (ThermoFisher, PA1-823A); anti-HSC70 (Santa Cruz, B-6); Anti-Flag-Peroxidase (A8592, Sigma); anti-HA (Covance, HA11); and anti-Myc (Santa Cruz, 9E10) were used for western blotting and immunoprecipitation. Knockdown experiments were performed using the following siRNAs: siPPARγ (J-040712-05 and J-040712-07, Dharmacon). Validated siSmurf1, siSmurf2 and siNS were previously described [47, 48].
Lipogenesis assay in AML12 cells and primary hepatocytes were performed using 3H-acetate as described [45]. Fatty acid uptake assay in AML12 cells was performed in 12-well plates. Briefly, AML12 cells were incubated with assay buffer (Hanks’ balanced buffer containing 1% BSA and 5 μCi/mL 3H-palmitic acid) for 60 minutes at 37°C. The cells were then washed twice with ice-cold PBS and lysed with 0.3 M NaOH. The radioactivity of the cell lysates was measured by liquid scintillation counting. In vivo fatty acid uptake assays were performed as described [49]. Briefly, mice were i.p. injected with BODIPY-FL-C16 (Life Technologies) after being fasted for 4 hours, then were euthanized at 5 hours after injection; liver, epididymal fat pad, and skeletal muscle were collected. Fluorescence was analyzed from cleared tissue homogenate using a plate reader and normalized to tissue weight.
Preadipocytes 3T3-L1 (ATCC, CL-173) were cultured in basal medium (DMEM supplemented with 10% FBS). Two days after transfection with siRNA, basal media were changed to differentiation media (day 0), which is DMEM supplemented with 10% FBS, 0.5 mM IBMX, 1μM dexamethasone, and 4 μg/mL insulin, for 2 days, then replaced with basal media with 2 μg/mL insulin for another 4 days. After 6 days of differentiation, cells were harvested for protein and mRNA analysis or subjected to Oil Red staining.
The purified recombinant PPARγ (0.25 μg) (Abcam, ab81807) and His6-Smurf1 (1.5 μg) were used in in vitro ubiquitination assay, which was carried out for 1 hour at 37°C in 30 μl reaction buffer supplemented with 2 mM Mg-ATP, 1 μg E1, 1 μg of recombinant UbcH5c, and 20 μg HA-ubiquitin or HA-ubiquitin variants (all from Boston Biochem).
Total RNA from AML12 cells or liver tissues was extracted by RNeasy Mini Kit (Qiagen) according to the manufacturer’s instructions. High Capacity Reverse Transcription Kit (ABI, Life Tech) was used to generate cDNA from RNA (500–2,000 ng). qRT-PCR was performed with Power SYBR Green PCR Master Mix (Life Technologies) using specific oligonucleotide primers as specified (S4 Table).
ChIP assays were carried out with an EZ-ChIP Chromatin Immunoprecipitation Kit (Millipore) according to the manufacturer’s instructions. Immunoprecipitations were carried out using anti-PPARγ antibody (Abcam, A3409A) and an isotype-matched IgG as the control. Reporter assays were performed in 12-well plates using PPRE-Luc (0.5 μg) and pRL-TK (0.2 μg) reporter plasmids, and the luciferase activities were determined using Dual Luciferase Reporter Assay System (Promega).
Microarray experiments for mouse liver tissues were performed on Affymetrix GeneChip Mouse Gene 1.0 ST arrays according to the standard Affymetrix GeneChip protocol at the Affymetrix service core in the Frederick National Laboratory for Cancer Research. The raw array data were then analyzed with packages oligo and lima under R platform, as described before [50, 51], to identify differentially expressed genes among groups (fold > 1.5, FDR cutoff = 0.1), and results were visualized using VennDiagram (https://cran.r-project.org/web/packages/VennDiagram) and gplots (https://cran.r-project.org/web/packages/gplots) under R platform. Data were submitted to GEO (accession number GSE113995). KEGG pathway analysis was performed by gage package, as described [52], to identify significantly enriched pathway (FDR q-value cutoff = 0.1) between Smurf1KO and WT liver samples.
The microarray analysis for human liver tissues from the LCI cohort of 247 Chinese patients was previously published [22] and data are accessible through GEO (accession number GSE14520). TCGA non-tumor liver tissue gene expression data were downloaded from TCGA-LIHC (https://portal.gdc.cancer.gov).
Unless indicated in the figure legends, two-tailed Student t test was used for statistical analysis.
|
10.1371/journal.pcbi.1003174 | Conditional Cooperativity of Toxin - Antitoxin Regulation Can Mediate Bistability between Growth and Dormancy | Many toxin-antitoxin operons are regulated by the toxin/antitoxin ratio by mechanisms collectively coined “conditional cooperativity”. Toxin and antitoxin form heteromers with different stoichiometric ratios, and the complex with the intermediate ratio works best as a transcription repressor. This allows transcription at low toxin level, strong repression at intermediate toxin level, and then again transcription at high toxin level. Such regulation has two interesting features; firstly, it provides a non-monotonous response to the concentration of one of the proteins, and secondly, it opens for ultra-sensitivity mediated by the sequestration of the functioning heteromers. We explore possible functions of conditional regulation in simple feedback motifs, and show that it can provide bistability for a wide range of parameters. We then demonstrate that the conditional cooperativity in toxin-antitoxin systems combined with the growth-inhibition activity of free toxin can mediate bistability between a growing state and a dormant state.
| The effectiveness of antibiotics on many pathogenic bacteria is compromised by multidrug tolerance. This is caused by a small sub-population of bacteria that happen to be in a dormant, non-dividing state when antibiotics are applied and thus are protected from being killed. These bacteria are called persisters. Unraveling the basic mechanism underlying this phenomenon is a necessary first step to overcome persistent and recurring infections. Experiments have shown a connection between persister formation and the battle between a toxin and its antitoxin inside an E. coli cell. Toxin inhibits the cell growth but is neutralized by the antitoxin by forming a complex. The proteins also regulate their own production through this complex, thereby forming a feedback system that controls the growth of the bacterium. In this work we provide mathematical modeling of the feedback module and explore its abilities. We find that the auto-regulation with reduced growth associated with free toxins allows the cell to be bistable between two states: an antitoxin-dominated, normal growing one, or a dormant one caused by the activity of the toxin. The latter can be the simplest description of persister state. The toxin-antitoxin system presents a powerful example of mixed feedback design, which can support epigenetics.
| Many bacteria and archaea have multiple Toxin-Antitoxin (TA) loci [1], where the toxin normally inhibits cell growth, while the antitoxin neutralizes the activity of the toxin by forming a tight TA complex. One of the known functions of TA loci is to respond to nutritional stress [2], namely, toxins are activated upon nutritional starvation and slow down the rate of translation. When cells are under normal fast growth conditions, on the other hand, the majority of the cells will be in the antitoxin-dominated state, such that toxin activity is fully inhibited.
It has been found that many bacterial TA loci are auto-regulated at the transcriptional level by a mechanism called “Conditional Cooperativity” (CC) [3], where the transcription factor can bind cooperatively to the operator only if the concentrations of two different proteins satisfy a certain stoichiometric ratio. CC was quantitatively studied in one of the Escherichia coli TA loci, relBE [3]–[6]. Here the two proteins, the toxin (mRNase) RelE and the antitoxin RelB, are encoded by the same operon, which is negatively auto-regulated. The tight dimer is a weak transcriptional auto-repressor, but this repression is strongly enhanced by the presence of RelE and becomes strongest at ratio . Over-expression of RelE above twice of , though, will result in an abrupt de-repression of the promoter. This unique behavior is a consequence of formation of alternative hetero-complexes of RelB and RelE; and . Two bind to the promoter site cooperatively to repress the promoter strongly, while does not bind to the promoter.
Interestingly, all plasmid and chromosome-encoded TA loci investigated are found to be regulated by CC so far, including relBE of E. coli [3], [4], vapBC of Salmonella enterica [7], phd/doc of plasmid P1 [8], [9] and ccdA/ccdB of plasmid F [10]. This suggest that CC is a common feature for TA loci.
In our previous work, we have explored the function of CC in the starvation response of the RelBE system, and showed that CC prevents random toxin activation and promotes fast translational recovery when starvation conditions terminate. However, to reproduce the full dynamics of the starvation response, we took into account details of the RelBE system, which made the model rather specific to it. The primary purpose of this paper is to construct a simple mathematical model that demonstrates the functions of CC in a more general perspective.
TA loci have been suggested to be involved in persister formation [11]–[16]. When an antibiotic is applied to a growing bacterial population, the majority of the bacteria are killed. However, a very small fraction of them survives and re-grows after the antibiotic is removed. If the progeny of the bacteria is again sensitive to the same antibiotic, they are called persisters, in contrast to the resistant bacteria that have acquired resistance to antibiotic by mutation. Persisters are genetically identical to the sensitive cells, but believed to be in a non- or slow-growing, dormant state. Since the majority of antibiotics interferes with the cell growth and division process, cells can survive if they grow slowly or not at all.
The exact molecular mechanism underlying persistence is not fully understood. However, it has been found that mutations in hipAB genes severely increase the level of of persisters formation. Interestingly hipAB is one of the TA loci in E. coli [11], [13], [14]. In addition, recent experiments [15] showed that removal of 10 mRNase-encoding TA loci reduced the persister fraction significantly. These observations strongly suggest that TA loci are important factors for persister formation.
One of the possible explanations is that stochastic activation of the toxin will slow down cell growth, resulting in a dormant state. This will be possible if the TA locus dynamics exhibits bistability, where a cell can be either in the antitoxin-dominated state that ensures the growth or in the toxin-dominated state that inhibits the growth. This viewpoint is also consistent with the observation that the persister state can be described as a metastable state with a constant stochastic switching rate to and from normal growing state [12].
This idea was theoretically pursued by Lou et al. [17] with a simple mathematical model that did not take CC into account. They concluded that, for bistability to be achieved, high cooperativity (Hill-coefficients ) is necessary, both in transcriptional auto-regulation of the TA operon and in the free toxin activity.
In this paper, we explore the basic features of CC as a regulation mechanism mediated by heteromer formation. We demonstrate that CC provides bistability in a simple feedback motif in a wide range of the parameters. We then construct a simplified model of TA system regulation and demonstrate that CC with growth rate-mediated feedback via toxin activity can provide the bistable alternatives between the antitoxin-dominated and the toxin-dominated states.
In this section, we construct a simple model of TA activity control with CR, a model aimed at capturing the central features of persister formation. We use the RelBE system as a reference because the molecular interactions and parameters are best known here. The reference parameters are listed in Materials and Methods.
In RelBE [6], the antitoxin RelB and the toxin RelE are encoded by the same operon, and transcriptionally auto-regulated by CC. RelE is metabolically stable, and its concentration decreases only by dilution due to cell division (generation time ∼30 min in log phase growth in rich medium). On the other hand, RelB is actively degraded by protease Lon, resulting in its very short half-life of min. In spite of this, the RelB concentration in a normally growing cell is about 10 times of that of RelE [4], suggesting that the RelB mRNA is translated about 100 times more often than RelE mRNA [6].
This situation is depicted in Fig. 3A1. Since both toxin T and antitoxin A are regulated by the same promoter, the corresponding equations apply:(5)where and are the maximal production rate for and for , respectively. The dilution rate of is given by cell division, and is taken as a unit rate, while is the active degradation rate of .
This motif, however, cannot exhibit bistability. Fig. 3A2 shows example null-clines, which have only one stable fixed point at the antitoxin dominated state. We performed parameter scan spanning from 1/8 to 8 fold relative to the values used for Fig. 3A2, but did not find any combination of parameters that gives bistability, even if we allow cooperative binding of AT to DNA with Hill coefficient 2 (data not shown). This absence of bistability is due to A being regulated identically to T. Accordingly, the de-repression of the promoter around increases not only the toxin production but also the antitoxin production, and the latter is so large that the system remains in the antitoxin-dominated state.
When we include the activity of free toxin on cell growth, however, the model system can show bistability. This is because the toxin-induced arrest of cell growth prolong lifetime of T, while leaving A being degraded by Lon at a high rate. The mathematical formulation of this extended model is(6)(7)expressing that reduces all protein production, and accordingly also decreases the dilution by cell growth. represents the reduction of protein expression per free toxin () molecule, and represents the growth inhibition per free toxin molecule. Notice that does not influence degradation of A, because it is anyway so unstable that cell division hardly affects its concentration.
These terms correspond to the growth-rate dependent feedback [17], [26], [27]. The reduction of the protein production ( term) can account for both direct activity of free toxin to TA locus and the global slowdown of the transcription rate due to change of physiological conditions [26]. Comparison of the present model with the steady state growth data in Ref.[26] is given in Text S1. We expect because the slowing down of the growth rate is due to the global slowing down of the protein production. At the same time, there can be some quantitative difference because may include the effect specific to the TA locus.
The growth-rate reduction mediated by T constitutes a positive feedback [17], [26], [27] on T accumulation, which is essential for bistability and persister formation. The term with reduces the production of both antitoxin and toxin, and thus overall weaken the ability to maintain the bistability. Note that primarily influences the transition state from A to T dominated state, because the reduction of production targets the short lived A protein first.
Fig. 3B1 examines eqs. (6)–(7) with parameters extracted from the RelBE system [6] (see the figure caption of Fig. 3). The null-clines in Fig. 3B2 are from the case, exhibiting two stable fixed point, one at the antitoxin-dominated state (the low- state, , ) and another at the toxin dominated state (the high- state, , ). Note that the antitoxin dominated state has almost the same concentrations as the stable fixed point in Fig. 3A2 with . The antitoxin dominated state scarcely depends on and , since there is almost no free toxin () in the antitoxin dominated state.
Figure 4A shows the ratio between the dilution rates at the low and high steady state, . The figure illustrates that our model predicts bistability for a wide range of parameters, and further that this bistability is indeed governed by the increase in cell generation parameterized by the term. For too large the bistability is counteracted because the toxin production is reduced too much by free toxin to accumulate enough for the stable high toxin state. Remarkably, for proportional reduction of protein production and increased cell generation, , the model predicts bistability for all .
We also studied the robustness of the bistability against parameter change. One of the most crucial parameters for the bistability is the ratio , because this determines the difference of the concentration of and . We therefore varied with keeping constant, and searched for the bistable regime in space. The rest of the parameters are kept same as those used in Fig. 4A. Only is considered, because lower ratios prevent antitoxin domination due to its 10 times higher degradation rate. For rather small (), too large makes the anti-toxin dominated state unstable, because very small amount of free toxin is enough to activate the positive feedback to toxin via the growth rate. With even larger , stronger feedback is needed to stabilize toxin-dominated state, reflected in larger values of and .
We further performed scanning of other parameters. We fixed one parameter at a time and sampled the rest of the parameters randomly to test 1000 samples in logarithmic scale within the range between 1/8 to 8 fold of the reference values. We then systematically changed the fixed parameters between 1/8 to 8 fold and repeated the procedure, to see the effect of the parameter. We found that 20% to 80% of the samples showed bistability. The detailed results are given in Text S2. We also explored the effect of the dissociation constant and more intensively, by changing from the reference value to 64 fold, since they describe the sharpness of the CR and this is expected to influence the bistability. We find that the number of bistability parameter sets decreases gradually with the fold change of and . Details are given in Fig. S4.
Using known parameters for the RelBE system in E. coli, we constructed a minimal model for TA activity, combining conditional regulation with a feedback from free toxin to the cell growth. It was demonstrated that this model shows bistability for a wide range of parameters, with a stable state corresponding to the antitoxin-dominated, normal growing state, and another metastable state corresponding the toxin dominated state, potentially corresponding to the persister state.
Noticeably, the model eqs. (6)–(7) did not rely on details of the molecular mechanisms of how the toxin works, and therefore the model is not limited to the RelBE system. The important assumptions are: (i) The TA system is conditionally regulated, (ii) toxins are stable and diluted mainly by cell division, while antitoxins are metabolically unstable, and (iii) free toxins reduce the productions of proteins and hence cell growth. All the conditions are satisfied in the TA loci that are confirmed to be regulated by conditional cooperativity [3], .
Our simple model pinpoints minimal ingredients for obtaining a persister state, but did not include stochastic production and/or degradation, and therefore cannot address the switching rates. In order to understand stochastic persister formation in E. coli, just performing stochastic simulation of the present motif is not enough, because the frequency of persisters depends on multiple parallel TA systems. In E. coli, 11 simultaneously interfering TA systems maintain a probability of persisters to be about 0.01%, while this probability is changed substantially first when about 50% of the TA systems is removed [15]. This clearly suggests that the interference of parallel systems has a strong influence to the switching behavior. Furthermore, comparing the stochastic simulations with the experimentally observed frequency of persisters requires a knowledge of the underlying distribution of the expression levels and corresponding growth rates in the cell population. It is not a simple task when the single cell growth rate depends on expression levels, because it feedbacks to the frequency of the cells as pointed out by Nevizhay et al. in [28]. In addition, it has been suggested that there is a strong link between the activation of the protease Lon and the TA-mediated persister formation, through the increase of the antitoxin degradation rate [15], [16]. The fluctuation of the Lon activity may be particularly important in determining switching rates, because it can provide coherent noise that favours simultaneous switching of many TAs to the persister state. It should also be noted that the Lon activity is activated by polyphosphate, which is regulated by the stringent response signalling molecule (p)ppGpp [16]. We plan to extend the present model to include these features and study the switching behavior in near future.
It is still interesting to think about possible implication of the observed switching rate to the present model. The fact that the persister formation is a rare event may indicate that the actual parameter value in the real system is located close to the boundary between the bistable region and the monostable region of the antitoxin-dominated state. Such parameter values can be chosen through selection process in a fluctuating environment, where slow growth of the persister pays off as a risk hedging strategy; the switching rate is expected to reflect the time scale of the temporal fluctuation of the environment [29].
Conditional regulation is an example of mixed feedback motifs [30], where protein-protein interactions and transcriptional repression are combined. In natural systems, protein-protein interaction mediated bistable switch was previously found for example in the epigenetic switch of the TP901 phage [23], [25] and in the sigma-factor/antisigma-factor system [24]. Conditional cooperativity in TA systems opens for a toolbox of regulatory units that can exhibit sufficient bistability to support also epigenetics. When removing the toxic ability of toxin, which has been done for RelE [3], and separating antitoxin from the operon to allow independent control, the strong binding between RelE and RelB should provide extreme ultrasensitivity, and thus very well separated metastable states. This conditional cooprativity-mediated bistability is the base for the bistability in full TA systems, and thus for the type II persister formation [12], [13], where a cell can spontaneously switch between the dormant state and the growing state (Fig. 5).
While simple protein-protein heteromers could produce ultrasensitivity, the non-monotonicity of the conditional cooperativity also secure that the antitoxin dominated state has a substantial amount of toxins present (Fig. 5). These toxins' activity is normally inhibited by short lived antitoxins, but the stored toxins can be used for faster switching to a dormant state if overall protein productions are externally inhibited, for example by starvation (Fig. 5). Therefore, the non-monotonicity may enhance the transition to type I persister formation [12], [13], where environmental stress triggers persister formation.
The importance of the protein-protein interaction mediated ultrasensitivty [22]–[25] and the growth rate-mediated feedback [17], [26]–[28] to bistable systems have been discussed as independent regulatory features in recent literature [31]. The uniqueness of the bistability in the TA system is that it combines both of these mechanisms. The need for combining these two mechanisms is closely associated with the fact that T and A are produced from the same operon, and thus are exposed to identical transcription regulation. Though it is difficult to get bistability with only one of the mechanisms [17], the TA system realizes a persister state by regulating the products of one operon through a combination of growth modulation and hetero-complex formation.
All the numerical analyses are done using C++ codes developed by the authors. When necessary, was calculated by solving algebraic equations (2) and (3) with conservation of mass for a given amount of by Newton's method [32]. The bistable solutions in Fig. 2 B (Fig. 4) were obtained by finding the fixed points for with eq. (4) ( and with eqs. 6 and 7) by Newton's method and then evaluating the stability based on the Jacobian. The trajectories that constitute the flux in Figs. 3A2 and 3B2 were calculated by the 4th-order Runge-Kutta method [32].
The values of the parameters used in the ODEs correspond to a conversion to dimensionless numbers of the parameters relative to the system we studied in [6].
In particular we used the lifetime of in exponential growth conditions () as time-unit () and the maximal amount of proteins produced in the unit time as concentration unit (). In the system nM thus fixing we obtain nM, while min. The value of in the starved condition [6] was evaluated to be around in this units. However, it is expected to be smaller in the normal condition, since RelE cleaves mRNA at the ribosomal A-cite, which is expected to be more accessible at the starvation. Therefore, we mostly explore values smaller than 1000.
The reference parameters are shown in table 1.
|
10.1371/journal.ppat.1005954 | SIRT1-PGC1α-NFκB Pathway of Oxidative and Inflammatory Stress during Trypanosoma cruzi Infection: Benefits of SIRT1-Targeted Therapy in Improving Heart Function in Chagas Disease | Chronic chagasic cardiomyopathy (CCM) is presented by increased oxidative/inflammatory stress and decreased mitochondrial bioenergetics. SIRT1 senses the redox changes and integrates mitochondrial metabolism and inflammation; and SIRT1 deficiency may be a major determinant in CCM. To test this, C57BL/6 mice were infected with Trypanosoma cruzi (Tc), treated with SIRT1 agonists (resveratrol or SRT1720), and monitored during chronic phase (~150 days post-infection). Resveratrol treatment was partially beneficial in controlling the pathologic processes in Chagas disease. The 3-weeks SRT1720 therapy provided significant benefits in restoring the left ventricular (LV) function (stroke volume, cardiac output, ejection fraction etc.) in chagasic mice, though cardiac hypertrophy presented by increased thickness of the interventricular septum and LV posterior wall, increased LV mass, and disproportionate synthesis of collagens was not controlled. SRT1720 treatment preserved the myocardial SIRT1 activity and PGC1α deacetylation (active-form) that were decreased by 53% and 9-fold respectively, in chagasic mice. Yet, SIRT1/PGC1α-dependent mitochondrial biogenesis (i.e., mitochondrial DNA content, and expression of subunits of the respiratory complexes and mtDNA replication machinery) was not improved in chronically-infected/SRT1720-treated mice. Instead, SRT1720 therapy resulted in 2-10-fold inhibition of Tc-induced oxidative (H2O2 and advanced oxidation protein products), nitrosative (inducible nitric oxide synthase, 4-hydroxynonenal, 3-nitrotyrosine), and inflammatory (IFNγ, IL1β, IL6 and TNFα) stress and inflammatory infiltrate in chagasic myocardium. These benefits were delivered through SIRT1-dependent inhibition of NFκB transcriptional activity. We conclude that Tc inhibition of SIRT1/PGC1α activity was not a key mechanism in mitochondrial biogenesis defects during Chagas disease. SRT1720-dependent SIRT1 activation led to suppression of NFκB transcriptional activity, and subsequently, oxidative/nitrosative and inflammatory pathology were subdued, and antioxidant status and LV function were enhanced in chronic chagasic cardiomyopathy.
| In this study, we determined whether enhancing the activity of sirtuin 1 (SIRT1) would be beneficial in maintaining heart health in Chagas disease. SIRT1 senses the redox shifts and integrates mitochondrial metabolism and inflammation. We found that treatment with SIRT1 agonists, given in a therapeutic window of time after Trypanosoma cruzi infection, had no beneficial effects in reducing the cardiac remodeling and mitochondrial biogenic defects in chagasic mice. SIRT1 agonist, however, controlled the NFκB signaling of oxidative and inflammatory responses and helped preserve the left ventricular function in chagasic mice. Co-delivery of SIRT1 agonists with other small molecules that inhibit mitochondrial dysfunction, cardiac fibrosis, and parasite persistence will potentially form a complete therapeutic regimen against Chagas disease.
| Trypanosoma cruzi (T. cruzi or Tc) is the etiological agent of Chagas disease that is endemic in Latin America [1]. After an exposure to parasite, infected individuals develop mild-to-no overt clinical symptoms. However, several decades later, ~30% of the infected individuals progress to heart failure associated with cardiac fibrosis, ventricular dilation, and thrombosis [2,3]. Vectors infected with T. cruzi are also present in the southern US [4], and CDC estimates that >300,000 infected individuals are living in the US [5,6]. Currently only two drugs are available for the treatment of T. cruzi infection: nifurtimox and benznidazole. These drugs are curative in early infection phase, but exhibit high toxicity and limited-to-no efficacy against chronic infection [7]. Thus, there is a need for new drugs for the treatment of chronic Chagas disease.
Mitochondria are the prime source of energy, providing ATP through oxidative phosphorylation (OXPHOS) pathway. A high copy number of mitochondrial DNA (mtDNA), reported to be ~6500 copies per diploid genome in myocardium [8], as well as the integrity of each mtDNA molecule is required to meet the high energy demand of the heart [9]. The mtDNA encodes 13 proteins that are essential for the normal assembly and function of the respiratory chain complexes. Peroxisome proliferator-activated receptor gamma coactivator-1α (PGC1α) is a member of the PGC family of transcription coactivators. PGC1α plays an important role in the expression of nuclear DNA and mtDNA encoded genes that drive mitochondrial biogenesis and increase the oxidative phosphorylation (OXPHOS) capacity [10]. Recently, we showed the mitochondrial respiratory chain activity and oxidative phosphorylation capacity were compromised in the myocardium of chronically infected rodents [11]. Further, mtDNA content and mtDNA encoded gene expression were decreased in Tc-infected cardiac myocytes and cardiac biopsies of chagasic patients [12]. Whether PGC1α activation of mitochondrial biogenesis and oxidative metabolism is compromised in CCM is not known.
Besides mitochondrial metabolic defects, chronic oxidative and inflammatory stress are hallmarks of Chagas disease. Acute infection by T. cruzi results in intense inflammatory activation of macrophages and CD8+T lymphocytes accompanied by increased expression of inflammatory mediators such as cytokines, chemokines, and nitric oxide synthase (NOS) in the heart (reviewed in [13,14]). Further, reactive oxygen species (ROS) are produced by neutrophils and macrophages activated by T. cruzi infection [14]. Besides infiltration of inflammatory infiltrate, cardiomyocytes are also reported to produce cytokines and mitochondrial ROS in response to T. cruzi infection [15,16]. The ROS induced adducts of DNA, protein and lipids were exacerbated in the myocardium of chronically infected rodents and human patients [12,17]. NFκB transcriptional factor signals oxidative and inflammatory responses [18], though mechanistic role of NFκB in chronic oxidative and inflammatory stress during CCM is yet to be elucidated.
Sirtuin 1 (SIRT1) is a highly conserved member of the family of NAD+-dependent Sir2 histone deacetylases, which deacetylates PGC1α at multiple lysine sites, consequently increasing PGC1α activity [19]. SIRT1 has also been reported to sense the redox shifts and integrate mitochondrial metabolism and inflammation through post-transcriptional regulation of the transcription factors and histones [20]. Several small molecule agonists of SIRT1 have been reported in literature. For example, resveratrol (3,5,4'-trihydroxy-trans-stilbene), a polyphenol found in red grape skins and red wine, is a natural agonist of SIRT1, and has been shown to increase mitochondrial number and the expression of genes for oxidative phosphorylation [21]. SRT1720 is a selective small molecule activator of SIRT1 and it is 1,000-fold more potent than resveratrol [22]. SRT1720 has been demonstrated to improve mitochondrial oxidative metabolism [23], and attenuate aging-related cardiac myocyte dysfunction [24].
In this study, we aimed to determine whether treatment with SIRT1 agonists will be beneficial in improving the heart function in Chagas disease. C57BL/6 mice were infected with T. cruzi, and treated with small molecule agonists of SIRT1. We demonstrate the therapeutic window of the efficacy of SIRT1 agonists in arresting the cardiac dysfunction in chagasic mice. Our results demonstrate a link between SIRT1, PGC1α, and NFκB in regulating ROS and inflammatory responses during T. cruzi infection and CCM.
We first determined if enhancing the SIRT1 activity would preserve the cardiac function in Chagas disease. Mice were infected with T. cruzi and then treated with resveratrol or SRT1720 as described in Materials and Methods. In vivo transthoracic echocardiography was performed at ~150 days pi to evaluate the changes in LV function (Table 1, Fig 1). Chagasic mice, as compared to normal controls, exhibited a substantial increase in LV end systolic volume (ESV, >2-fold), and a decline in stroke volume (SV, 35%), cardiac output (CO, 59%), and ejection fraction (EF, 33%) (Fig 1A–1D, all, p<0.001). The LV internal diameter at systole (LVIDs) was increased by >2-fold, and consequently, fractional shortening (FS) was decreased by 53% in chagasic (vs. normal) mice (Fig 1E & 1F, p<0.001). The Tc-infected/SRT1720-treated mice, in comparison to normal controls, exhibited only 20%, 27% and 16% decline in SV, CO, EF, respectively, that were not statistically significant. In comparison to Tc-infected/untreated mice, infected/SRT1720-treated mice exhibited 55% decline in ESV and 29%, 33%, 37%, and 36% increase in SV, CO, EF, and FS, respectively (Fig 1Aa–1Ae, Table 1, all, #p<0.05–0.001). Tc-infected/resveratrol-treated mice (vs. infected/untreated mice) exhibited a moderate (up to 20%) but significant improvement in Tc-induced loss in ESV, SV and CO; and a modest, but statistically insignificant, improvement in EF and FS (S1 Fig, panels a-f). These results suggested that SRT1720 treatment was effective in arresting the LV dysfunction in chagasic mice. Resveratrol provided a partial recovery of cardiac output in Tc-infected mice.
Echocardiography imaging in M mode was performed to gain an anatomo-pathological view of the heart in chronically infected mice. These data showed that systolic and diastolic thickness of inter-ventricular septum (IVS), LV area, and LV mass were increased by 44%, 40%, 28%, and 28% respectively, while LV posterior wall (LVPW) was thinned by 41% in chagasic (vs. normal) mice (Fig 2Aa–2Ae, all, p<0.001, Table 1). Histological evaluation of the tissue sections subjected to Masson's Trichrome staining showed an increase in diffused collagen deposition in chagasic myocardium (score: 4.0 ± 0.4 vs. 0.3 ± 0.04, Tc-infected vs. normal controls, p<0.05, Fig 2Ba & 2Bb). An increase in cardiac fibrosis in chagasic myocardium was also evidenced by 1.6-fold, 1.8-fold and 3.2-fold increase in mRNA levels for COLI, COLIII, and αSMA, respectively (Fig 2Ca–2Cc, all, p<0.01). In infected/SRT1720-treated mice, LV area (systole) was normalized (Fig 2Ae, #p<0.05), though SRT1720 treatment provided no benefits in normalizing the IVS and LVPW thickness and LV mass in chagasic mice (Fig 2A). The myocardial deposition of collagen (score: 2.5 ± 0.8) and collagen-related gene expressions were also not significantly changed in SRT1720-treated (vs. untreated) chagasic mice (Fig 2B and 2C). Chagasic mice treated with resveratrol also exhibited modest control of LV mass, but no improvement in the IVS and LVPW thickness and LV area (S2 Fig, panels a-e, Table 1). Together, the results presented in Fig 2 and S2 Fig suggested that a) an increase in passive stiffness (enhanced ESV, IVSs, IVSd, LVIDs) alongside thinning of the LVPW contributed to depressed LV function in chagasic mice, and b) SRT1720 benefits in arresting the LV dysfunction were not delivered through control of cardiac collagenosis and hypertrophy in chagasic mice.
T. cruzi infection results in respiratory chain inefficiency in mice and humans [25,26]. SIRT1 deacetylates PGC1α, and PGC1α coactivation of nuclear respiratory factor (NRF1) signals the expression of key metabolic genes required for respiration and mtDNA transcription and replication. Western blotting showed the total and nuclear levels of SIRT1, PGC1α and NRF1 proteins were either increased or not changed in the myocardium of infected/untreated and infected/SRT1720-treated mice as compared to that noted in normal controls (Fig 3Aa, 3Ab and 3Ba, 3Bb). However, total and nuclear concentration of acetylated PGC1α (inactive form) were increased by >9-fold (Fig 3Aa, 3Ab and 3Ba, 3Bb, p<0.01), and associated with a 53% decline in SIRT1 activity (Fig 3C, p<0.05) in the myocardium of chronically infected (vs. normal) mice. SRT1720 treatment of chagasic mice resulted in 84% and 58% decline in total and nuclear levels of the acetylated PGC1α level, respectively (Fig 3A and 3B, #p<0.001) and 60% increase in SIRT1 activity (Fig 3C, #p<0.05). The changes in mitochondrial biogenesis were examined by measuring the mitochondrial markers at the DNA, gene expression and protein levels. The mtDNA levels for CYTB and COII sequences, normalized to nuclear DNA sequence for β-globin, were decreased by 25% and 24%, respectively, in the myocardium of chronically infected mice (Fig 4Aa & 4Ab, *p<0.05). No difference in the citrate synthase activity, a marker of mitochondrial mass, was noted in chagasic vs. normal mice. The mRNA levels for mtDNA encoded ND1, COIII, and ATP6 subunits that are essential components of the CI, CIV, and CV respiratory complexes and required for maintaining the oxygen consumption and coupled OXPHOS, were decreased by 34%, 55% and 54%, respectively, in chagasic (vs. normal) murine myocardium (Fig 4Ba–4Bc, *p<0.05). The protein levels of mtDNA-encoded CYTB and COI were also decreased by 60% and 33%, respectively, in chagasic myocardium (Fig 4Ca & 4Cb, *p<0.01). The decline in OXPHOS-related transcripts could be a result of changes in mtDNA replication/transcription efficiency. Our data showed the myocardial mRNA levels for mtDNA replication machinery, POLG1, SSBP1, and TOP1, were decreased by 51%, 47%, and 37%, respectively, in chagasic (vs. normal) mice (Fig 4Bd–4Bf, *p<0.05). We also noted 35%-90% decline in POLG and TOP1 protein levels in chagasic mice (Fig 4Ca & 4Cb, *p<0.01). Treatment with resveratrol resulted in 30–40% increase in mtDNA level (S3 Fig, panels A.a&b, #p<0.05) and 40–55% increase in mRNA levels for mtDNA encoded ND1, COIII, and ATP6 genes (S3 Fig, panels B.a-c, #p<0.05), and no significant improvement in the mRNA levels for POLG1, SSBP1 and TOP1 (S3 Fig, panels B.d-f) in chagasic mice. Surprisingly, SRT1720-treated/chagasic mice exhibited no significant improvement in the PGC1α/NRF1-dependent mtDNA content (COII and CYTB levels, Fig 4Aa & 4Ab), mtDNA encoded gene expression (ND1, COIII, ATP6, Fig 4Ba–4Bc), and mtDNA replication machinery (Fig 4Bd–4Bf) that were compromised in the myocardium of chagasic mice. Likewise, protein levels of the mtDNA-encoded proteins (e.g. CYTB, COI) and the mtDNA replication/transcription machinery (POLG1, TOP1) were not improved in SRT1720-treated chagasic mice (Fig 4Ca & 4Cb). Together, the results presented in Fig 3 and Fig 4 and S3 Fig suggested that a) mtDNA content and mRNA and protein levels of the mtDNA-encoded genes were significantly decreased in the myocardium of chronically infected mice, and this outcome was associated with a decline in mtDNA replication machinery, and b) SRT1720 treatment was effective in activation of SIRT1/PGC1α in the chagasic myocardium. However, c) SRT1720-mediated increase in SIRT1 activity and deacetylated PGC1α did not improve the mitochondrial biogenesis in chagasic mice.
We next determined if SIRT1 agonists controlled the chronic oxidative and inflammatory stresses that are hallmarks of Chagas disease [11]. Fluorometric evaluation of ROS showed 2.3-fold increase in H2O2 levels in the myocardial homogenates of chronically infected (vs. normal) mice (Fig 5A, *p<0.001). Advanced oxidation protein products (AOPPs) are formed by HOCl-induced chlorination of amines and considered a biomarker of inflammatory and oxidative pathology. Our data showed a 57% increase in AOPP content in chagasic (vs. normal) myocardium (Fig 5B, *p<0.01). The expression of inducible NOS (iNOS, a major source of nitric oxide) and the levels of the oxidative/nitrosative stress markers 4-hydroxynonenal (4HNE) and 3-nitrotyrosine (3NT) were increased by 20-fold, 10-fold, and 8-fold, respectively, in chagasic (vs. normal) myocardium (Fig 5Ca & 5Cb, all, p<0.001). In contrast, protein level of Nrf2 (transcriptional regulator of antioxidant gene expression) and total antioxidant capacity were decreased by 60% and 41%, respectively, in chagasic (vs. normal) myocardium (Fig 5C and 5D, *p<0.001). Resveratrol treatment was not effective in controlling the chronic oxidative stress in chagasic mice (S4 Fig, panels a&b). However, SRT1720 treatment resulted in a 57% and 90% decline in Tc-induced H2O2 and AOPP levels, respectively (Fig 5A and 5B, #p<0.01); 76%, 64% and 62% decline in Tc-induced iNOS, 4HNE and 3NT levels, respectively (Fig 5Ca & 5Cb, all #p<0.01); and 57% and 50% improvement in Tc-induced loss in Nrf2 expression and antioxidant capacity, respectively (Fig 5C and 5D, #p<0.05). These results suggested that SRT1720 activation of the SIRT1 was beneficial in controlling the chronic oxidative stress in chagasic myocardium.
Histological studies showed the myocardial level of inflammatory infiltrate constituted of diffused inflammatory foci (histological score: 2–3) was increased in heart tissue of chronically-infected/untreated (vs. normal) mice (Fig 6Aa & 6Ab). Chagasic mice exhibited a high degree of myocardial degeneration with enlarged myocytes. The cytokine gene expression was predominantly of proinflammatory nature evidenced by 9-fold, 3-fold and 2-fold increase in IFNγ, IL1β, and TNFα mRNA (Fig 6Ba–6Bc, all, *p<0.01) and 29% and 43% increase in IL10 and arginase 1 (Arg1) mRNA (Fig 6Bd & 6Be, *p<0.05), respectively, in chagasic myocardium. The myocardial IL10 protein level was increased by 3-fold in chagasic mice (Fig 6Ca & 6Cb, *p<0.05). Resveratrol treatment resulted in a modest (but not statistically significant) control of pro-inflammatory cytokine expression in chagasic heart (S5 Fig, panels a-c). However, myocardial inflammation was significantly subsided in infected/SRT1720-treated mice, evidenced by the detection of minimal tissue inflammatory infiltrate (histological score: 0–1, Fig 6Ac). Further, SRT1720-treated chagasic mice exhibited 80%, 36% and 46% decline in the expression of IFNγ, IL1β and TNFα, respectively (Fig 6Ba–6Bc, all, #p<0.05), and no change in the myocardial expression of IL10 and Arg1 (Fig 6B & 6C). Chronic persistence of parasite was noted in all infected mice, and was not changed by SRT1720 treatment (Fig 6D). These results suggested that SRT1720 was beneficial in attenuating the myocardial inflammatory infiltrate and proinflammatory cytokine response in chagasic mice.
NFκB family of transcription factors is of central importance in inflammation and immunity. Rel A (p65) is an important subunit of activated NFκB dimers (p50/p65, p65/p65, and p65/c-Rel). Western blotting showed the nuclear level of NFκB-p65 was increased by 76% in the myocardium of infected/untreated (Fig 7Aa & 7Ab, *p<0.05) mice, and normalized to control levels in infected/SRT1720-treated mice (#p<0.05), thus suggesting that SIRT1 might regulate NFκB activation in CCM. To verify this, we utilized an in vitro system. Cardiac myocytes were infected with T. cruzi and incubated for 24 h in presence or absence of SRT1720 or emodin (blocks IκB degradation and p65/Rel A release for nuclear translocation). As in chagasic heart, no changes in total levels of p65 were noted in any of the treatment groups, while the cytosolic level of p65 was decreased in Tc-infected cells (Fig 7B). Further, the nuclear translocation of p65 and acetylated-p65 were increased in Tc-infected cardiac myocytes (Fig 7B) and associated with 7-fold and 5-fold increase in the mRNA levels for IL1β and IL6, respectively (Fig 7Ca & 7Cb, *p<0.01). The Tc-induced cytokine gene expression was abolished by 59%-72% by emodin treatment (Fig 7Ca & 7Cb, #p<0.05), thus, verifying the role of NFκB in signaling inflammatory responses in infected cardiomyocytes. SRT1720 treatment normalized the nuclear p65 content to control levels, substantially diminished the nuclear acetylated-p65 level (Fig 7B), and decreased the cytokine gene expression by 41%-43% (Fig 7C, #p<0.05) in infected cardiomyocytes. SRT1720 treatment also decreased the Tc-induced oxidative stress (iNOS, 4HNE, 3NT) in cardiomyocytes (Fig 7D). We performed a dual reporter assay to evaluate the NFκB activity. HEK293 cells were transiently transfected with NFκB-TATA-luciferase reporter plasmid and pRL-TK plasmid (expresses renilla luciferase, control for transfection efficiency), infected with T. cruzi for 24 h, and NFκB-dependent luciferase activity was monitored. These data showed the NFκB-dependent luciferase activity (normalized to renilla luciferase) was increased by 2-fold in Tc-infected cells (Fig 7E, *p<0.01) and controlled by 66% when infected cells were treated with SRT1720 (Fig 7E, #p<0.05). Together, these results suggested that SIRT1 deacetylation of NFκB-p65 regulated inflammatory responses that otherwise were pronounced in T. cruzi-infected cardiomyocytes.
In this study, we demonstrated that SIRT1 activity was decreased in chagasic heart, and treatment with SIRT1 agonist during a therapeutic window, i.e., after the immune control of acute parasitemia and before the onset of myocarditis, was beneficial in preserving cardiac function in CCM. The decline in SIRT1/PGC1α activity was not the key mechanism in mitochondrial biogenic defects in Chagas disease, and therefore SIRT1-targeted therapy did not normalize the PGC1α/NRF1-dependent mitochondrial biogenesis and cardiac remodeling in chagasic disease. Instead, SIRT1 deacetylation of NFκB-p65 repressed the Tc-induced inflammatory stress and preserved the antioxidant/oxidant balance in the myocardium of SRT1720-treated chagasic mice. Our results, to the best of our knowledge, provide the first evidence for potential utility of SRT1720 mediated protection of LV function in CCM.
Others and we have shown the mitochondrial respiratory complexes and OXPHOS capacity are compromised in the cardiac biopsies of Tc-infected experimental animals and chagasic human patients [12,17,27]. Our findings in this study suggested that a decline in mitochondrial biogenesis constituted at least one of the mechanisms involved in OXPHOS inefficiency in Chagas disease. This is because mtDNA content as well as the expression of the mtDNA encoded genes at mRNA and protein levels were significantly suppressed in chagasic myocardium (Fig 4); and mtDNA encoded 13 polypeptides are essential for normal assembly and function of the CI, CIII, CIV and CV complexes of the respiratory chain. The expression levels of SIRT1, PGC1α and NRF1 that are involved in regulating the mitochondrial biogenesis were not changed, yet SIRT1 activity and deacetylated-PGC1α (active form) were significantly decreased in chagasic heart (Fig 3). Though a decline in PGC1 isoforms (PGC1α and PGC1β) is noted in other metabolic diseases, such as obesity and diabetes [28–31]; it is generally accepted that deacetylation, and not the changes in the expression level, of PGC1α is required for maintaining the mitochondrial biogenesis.
We postulated that SIRT1 agonists, via enhancing the SIRT1/PGC1α activity, would offer a therapeutic option to improve the mitochondrial biogenesis, and subsequently, the heart function in CCM. We, first, used resveratrol as a therapeutic candidate for the treatment of chronic CCM. Resveratrol has been shown to induce mitochondrial biogenesis in many tissues [32,33], control pressure overload induced hypertrophy and contractile dysfunction in mice [34,35]; and reverse the ischemia/reperfusion induced loss in renal mitochondrial mass by an increase in the expression of PGC1α and its downstream mediators [36]. In the present study, though resveratrol partially improved the heart function (S1 Fig), an overall lackluster performance of resveratrol in arresting Tc-induced cardiac remodeling and mitochondrial biogenic defects was noted (S2–S5 Figs). This was despite the fact that we have used biologically relevant concentrations of resveratrol as was used in other studies. Our data suggest that delayed treatment in chronic phase when oxidative/inflammatory pathology have already caused tissue damage in the heart was at least partially responsible for resveratrol inefficacy in CCM. The data discussed below with SRT1720 treatment allow us to propose that SIRT1 agonists offered during the clinically asymptomatic phase when host has controlled the acute parasitemia but yet not entered the chronic phase of progressive cardiomyopathy, will be most beneficial in arresting the adverse clinical outcomes in Chagas disease.
We treated mice with SRT1720 (specific and potent SIRT1 agonist) for three weeks during 45–66 days pi. In contrast to untreated/infected mice that developed significant LV systolic dysfunction by ~150 days pi; short-term SRT1720 treatment in the clinically asymptomatic phase was effective in preserving the heart function in chronically infected mice (Fig 1). This is the first study demonstrating that SRT1720 treatment rescued the heart function following a chronic T. cruzi infection. The effects of SRT1720 in improving the LV function in chagasic mice were associated with a significant increase in SIRT1 activity and deacetylation of PGC1α (Fig 3), as has also been noted in models of metabolic disease [37,38]. Others have shown that long-term SRT1720 treatment produced benefits in increasing the organ function and lifespan in mice [24]. SRT1720 stimulated the mitochondrial biogenesis and effectively reversed the conditions associated with metabolic deficiencies [37,39,40]. Surprisingly, despite SIRT1 activation and PGC1α deacetylation suggested to be required for mitochondrial biogenesis, SRT1720 treatment did not reverse the mitochondrial biogenic defects in the myocardium of chagasic mice (Fig 4). A recent study showed the deacetylation by SIRT1 decreased PGC1α activity and mitochondria number in myotubes [41]. Others have shown that kidney-specific overexpression of SIRT1 was protective against metabolic kidney disease though mitochondrial number was not changed. Further studies will be required to delineate the SIRT1/PGC1α dynamics in maintaining the mitochondrial biogenesis in normal and disease conditions. Yet, our data allows us to surmise that activators of the sirtuin family of proteins may be important in the development of new therapeutic strategies for treating cardiac dysfunction in Chagas disease.
Multiple sources of ROS including mitochondrial electron transport chain leakage and NADPH oxidases, sometimes in response to cytokines and growth factors, are noted in Chagas disease (reviewed in [42]). In this study, we found that SIRT1 agonists enhanced the antioxidant capacity and reversed the oxidative/nitrosative injuries (Fig 5 & S4 Fig), inflammatory cytokine response (Fig 6 and Fig 7 and S5 Fig), and infiltration of inflammatory infiltrate in the myocardium of chronically infected mice (Fig 6). Consistent with these results, SRT1720 has been reported to decrease the levels of 3-nitrotyrosine and iNOS in ischemia perfusion induced renal injury in mice [36]. SRT1720 was also shown to ameliorate vascular endothelial dysfunction by enhancing COX2 signaling and reducing oxidative stress and inflammation with aging in mice [43]; and increase the levels of catalase, thus reducing ROS level and apoptosis and retaining kidney function in mice [44]. SIRT1 can deacetylate the FoxO factors and stimulate the expression of antioxidants [45], and inhibit NFκB signaling that is a major inducer of inflammatory responses [46]. The role of FoxO in preserving antioxidant/oxidant balance in CCM remains to be investigated. However, others and we have shown the activation of NFκB by T. cruzi in a variety of immune and non-immune cells [16,47]. A variable degree of loss in SIRT1 activity associated with steady hyper-activation of NFκB-p65 is observed in many chronic inflammatory diseases [48], including CCM in this study (Fig 7). SRT1720 treatment inhibited the nuclear translocation of p65/Rel A, NFκB transcriptional activity, and NFκB-dependent inflammatory cytokines’ gene expression in cells infected by T. cruzi (Fig 7). SIRT1 influenced the chronic inflammation in chagasic disease by directly deacetylating the p65/Rel A (Fig 7). SIRT1 can also inhibit the NκB target genes by co-localizing with p65 and p300, the latter a histone acetyl transferase with a broad range of substrates. While SIRT1’s ability to regulate NFκB activity is shown in macrophages [49,50], ours is the first observation demonstrating SIRT1 regulation of NFκB in stressed cardiomyocytes. The observation of no increase in T. cruzi burden in mice treated with SRT1720 (Fig 6) implies that NF-κB-induced inflammatory responses were more detrimental to the host than to the parasite. Further, our finding that SIRT1 agonist (SRT1720) restricted the ROS and oxidative stress markers (3NT and 4HNE) that otherwise were significantly induced by T. cruzi infection (Fig 6) suggest that SIRT1/NFκB axis coordinates the oxidative stress as well as inflammation in chronic CCM and heart failure.
In summary, we have shown that mitochondrial biogenesis is compromised in chronic chagasic mice. A loss of SIRT1 activity contributed to NFκB-p65 activation and chronic cardiac pathology and heart failure in CCM. SRT1720 treatment enhanced the SIRT1/PGC1α activity but failed to improve the mitochondrial biogenesis in CCM. Instead, SRT1720 influenced the SIRT1/NFκB regulation of oxidative, nitrosative, and inflammatory responses, and, consequently, preserved the cardiac function in chagasic mice. We conclude that activators of the sirtuin family of proteins will provide promising new therapeutic strategies for treating cardiac dysfunction in chronic Chagas disease.
All animal experiments were performed according to the National Institutes of Health Guide for Care and Use of Experimental Animals, and approved by the Institutional Animal Care and Use Committee (IACUC) at the University of Texas Medical Branch, Galveston (protocol number: 0805029).
All chemicals were of molecular grade, and purchased from Sigma-Aldrich (St. Louis, MO) unless otherwise stated. T. cruzi trypomastigotes (SylvioX10 strain, ATCC 50823) were propagated by in vitro passage in C2C12 cells. C57BL/6 mice were purchased from Harlan Laboratories (Indianapolis, IN). Mice (6-weeks-old) were infected with T. cruzi (10,000 trypomastigotes/mouse, intraperitoneal), and harvested at days 150 post-infection (pi) corresponding to chronic disease phase. To enhance the SIRT1 activity, two approaches were applied. One, mice were treated with resveratrol (20 mg/ml in drinking water) for three weeks, during days 90–111 pi. Two, mice were given SRT1720 (1 mg/100 μl/mouse, intraperitoneally, Selleck Chemicals, Houston, TX) three times a week during days 45–66 pi. Tissue samples were stored at -80°C. Protein levels were determined by using the Bradford Protein Assay (Bio-Rad, Hercules CA).
Human cardiomyocytes were cultured and maintained in Dulbecco's modified Eagle's medium/F-12 medium with 12.5% fetal bovine serum. Cardiomyocytes were seeded in T75 flasks (3×106 cells per flask, 70% confluence), and infected with T. cruzi trypomastigotes (cell: parasite ratio, 1:3). Cells were incubated in presence or absence of SIRT1 agonist (1 μM SRT1720) and NF-κB inhibitor (50 μM emodin) for 24 h.
Mice were continuously anesthetized by inhalant 1.5% isoflurane/100% O2 to maintain a light sedation level. Mice were placed supine on an electrical heating pad at 37°C and heart rate and respiratory physiology were continuously monitored by electrocardiography. Mice chests were shaved, and warm ultrasound gel was applied to the area of interest. Transthoracic echocardiography was performed using the Vevo 2100 ultrasound system (Visual Sonics, Toronto, Canada) equipped with a high-frequency linear array transducer (MS400, 18–38 MHz) [51]. Heart was imaged in B-mode and M-mode to examine the parameters of left ventricle (LV) in diastole (-d) and systole (-s). All measurements were obtained in triplicate and acquired in long-axis and short-axis views. Data were analyzed by using Vevo 2100 standard measurement software.
Histological preparation and staining of the tissues was performed at the Research Histopathology Core at the UTMB. Briefly, tissue sections were fixed in 10% buffered formalin, dehydrated in absolute ethanol, cleared in xylene, and embedded in paraffin. Five-micron tissue sections were subjected to Masson’s Trichrome or Hematoxylin and Eosin (H&E) staining, and evaluated by light microscopy using an Olympus BX-15 microscope equipped with a digital camera and Simple PCI software (v.6.0, Compix, Sewickley, PA). In general, we analyzed each tissue section for >10 microscopic fields (20X magnification), and examined three different tissue sections/mouse (4 mice/group). The collagen area as a percentage of the total myocardial area was assessed as a measure of fibrosis. All pixels with blue stain in Masson’s trichrome-stained sections were selected to build a binary image, subsequently calculating the total area occupied by connective tissue. Sections were categorized based on percent fibrotic area as: (0) <1%, (1) 1–5%, (2) 5–10%, (3) 10–15%, and (4) >15% [52].
Myocarditis (presence of inflammatory cells) was scored as 0 (absent), 1 (focal/mild, ≤1 foci), 2 (moderate, ≥2 inflammatory foci), 3 (extensive coalescing of inflammatory foci or disseminated inflammation), and 4 (diffused inflammation, tissue necrosis, interstitial edema, and loss of integrity). Inflammatory infiltrate was characterized as diffused or focal depending upon how closely the inflammatory cells were associated [52].
Heart tissue sections (10 mg) were homogenized in 500 μl of TRIzol reagent (Invitrogen, Carlsbad, CA), and RNA was extracted by chloroform/isopropanol/ethanol method. Total RNA (2 μg) was reverse transcribed by using poly(dT)18 with an iScript kit (Bio-Rad). The cDNA was utilized as template with SYBR-Green super-mix (Bio-Rad), and real time quantitative PCR (qPCR) was performed on an iCycler Thermal Cycler. The gene-specific oligonucleotide pairs used for amplifying the mRNAs are listed in supplemental S1 Table. The PCR Base Line Subtracted Curve Fit mode was applied for threshold cycle (Ct), and mRNA level was calculated by iCycler iQ Real-Time Detection Software (Bio-Rad). The Ct values for target mRNAs were normalized to geometric mean of GAPDH mRNA, and the relative expression level of each target gene was calculated as 2−ΔCt, where ΔCt represents the Ct (sample)—Ct (control) [53,54].
Freshly harvested heart tissue sections (30 mg) were washed with ice-cold Tris-buffered saline and homogenized in RIPA buffer (tissue: buffer ratio, 1: 10, w/v) [17]. Homogenates were centrifuged for 10 min at 10,000 g, and supernatants were stored at −80°C. For the preparation of nuclear and cytosolic fractions, heart tissue sections (50 mg) were homogenized in ice-cold HMK buffer (10 mM HEPES pH 7.9, 1.5 mM MgCl2, 10 mM KCl) containing 1 mM DTT and 1% Protease Inhibitor Cocktail (Sigma-Aldrich). Tissue lysates were centrifuged at 4°C, 10000 g for 20 min and supernatants stored as a cytosolic fraction. The pellets were re-suspended in HMK buffer containing 0.42 M NaCl, 0.2 mM EDTA, and 25% (v/v) glycerol, centrifuged at 4°C, 20, 000 g for 5 minutes, and nuclear fractions were stored at −80°C.
Cardiomyocytes were infected with T. cruzi and incubated in presence or absence of SRT1720 (with 1 μM) for 24 h. Cells were lysed for 30 min on ice in lysis buffer containing 50 mM Tris pH 7.5, 150 mM NaCl, 1 mM EDTA, 1 mM EGTA, 1% NP-40, 2.5 mM KH2PO4, and 1 mM Na3VO4. Cell lysates were centrifuged at 3000 g at 4°C for 15 min and the resultant supernatants were stored at −80°C. For the preparation of nuclear and cytosolic fractions, cells (7×106/ml) were incubated on ice for 30 minutes in buffer A (10 mM HEPES, pH 7.9, 10 mM NaCl, 0.1 mM EDTA, 0.1 mM EGTA, 1 mM DTT, 1 mM PMSF) containing 0.625% NP-40 and 1% protease inhibitor cocktail. Cell lysates were centrifuged at 4°C at 10,000 g for 1 min and supernatants stored as a cytosolic fraction. Pellets were washed with buffer A containing 1.7 M sucrose, re-suspended in buffer B (20 mM HEPES pH 7.9, 0.4 M NaCl, 1 mM EDTA, 1 mM EGTA, 1 mM DTT, and 1 mM PMSF), and centrifuged at 4°C at 13 000 g for 5 minutes. The resultant supernatants were stored at −80°C as nuclear extracts.
Heart or cell homogenates and nuclear fractions (30 μg protein) were electrophoresed on a 4–15% Mini-Protein TGXTM gel using a Mini-PROTEAN electrophoresis chamber (Bio-Rad), and proteins were transferred to a PVDF membrane using a Criterion Trans-blot System (Bio-Rad). Membranes were blocked with 50 mM Tris, 150 mM NaCl (TBS) containing 5% non-fat dry milk (NFDM), washed three times for 10 min each with TBS—0.1% Tween 20 (TBST), and incubated overnight at 4°C with antibodies against CYTB (Santa Cruz Biotech, Dallas, TX, sc11436), COI (Abcam, Cambridge, UK, ab147053), GAPDH (Cell signaling, Danvers, MA, 3683), 4-hydroxynonenal (4HNE, Alpha Diagnostic Inc, San Antonio, TX, HNE11-S), iNOS (Abcam, ab49999), Lamin A/C (Santa Cruz, sc20681), NF-κB-p65 subunit (Santa Cruz, F-6 clone, sc-8008); NF-κB-acetyl- p65 (Abcam, acetyl K310 clone, ab198703), nitrotyrosine (3NT, Merck Millipore, Billerica, MA, 06–284), NRF1 (Santa Cruz, sc33771), Nrf2 (Santa Cruz, sc722), PGC1α (Santa Cruz, sc13067), POLG (Santa Cruz, 390634), SIRT1 (Abcam, ab32441), TOP1 (Abcam, ab3825), IL-10 (A2, Santa Cruz, sc-365858) and anti-acetylated-lysine antibody (Cell Signaling, 9441). All antibodies from Santa Cruz were used at 1:200 dilution in TBST-5% NFDM. All other antibodies were used at 1:1000 dilution in TBST-5% NFDM. Membranes were washed as above, incubated with HRP-conjugated secondary antibody (1:10,000 dilution, Southern Biotech, Birmingham, AL), and images were acquired by using an ImageQuant LAS4000 system (GE Healthcare, Pittsburgh, MA). Immunoblots were subjected to Ponceau S staining to confirm equal loading of samples. Densitometry analysis of protein bands of interest was performed using a Fluorchem HD2 Imaging System (Alpha-Innotech, San Leandro, CA), and normalized against GAPDH (tissue homogenates) or Lamin A/C (nuclear fractions).
SIRT1 deacetylase activity was measured by using a SIRT1 Fluorometric Assay Kit (Abcam, ab156065). Briefly, nuclear fractions (100 μg) isolated from heart tissues were added to acetylated Lys382 p53 peptide (50 μl) that is coupled with fluorophore and quencher at the amino terminal and carboxyl terminal, respectively. The reaction was initiated with addition of 100 μl NAD+, and SIRT1 dependent deacetylation of the substrate peptide coupled with its digestion by the action of proteases, and fluorescence release recorded. The fluorescence intensity (Ex350nm/Em440nm) was measured at two min intervals for 60 minutes using a SpectraMax M5 microplate reader (Molecular Devices, Sunnyvale CA). Standard curve was prepared with recombinant SIRT1 (0–120 ng), and results were presented as relative fluorescence units per μg protein.
Tissue sections (5 mg) were subjected to Proteinase-K lysis, and total DNA was purified using a DNeasy Blood & Tissue Kit (Qiagen, Hilden, Germany). Total DNA (20 ng) was utilized as template with SYBR Green Super-mix (Bio-Rad) and primer pairs specific for mtDNA-encoded cytochrome b (CYTB) and cytochrome oxidase 2 (COII) regions, and real-time qPCR was performed on an iCycler thermal cycler. The mtDNA content was normalized to β-globin nuDNA. For the evaluation of parasite burden, Tc18SrDNA-specific primers were utilized for qPCR, and data were normalized to GAPDH.
Citrate synthase (CS) activity is a sensitive measure of mitochondrial mass, and was measured by using the MitoCheck Citrate Synthase Activity Assay Kit (Cayman, Ann Arbor, MI) following the protocol provided by the manufacturer.
To measure the total H2O2 levels, tissue homogenates (100 μg) were added in triplicate to flat-bottom (dark-walled) 96-well plates. The reaction was started with addition of 33-μM amplex red (10-acetyl-3, 7-dihydroxyphenoxazine) and 0.1U/ml horseradish peroxidase (final reaction volume: 150 μl). The oxidation of amplex red to fluorescent resorufin by H2O2 (Ex563nm/Em587nm) was recorded on a SpectraMax M2 microplate reader (Molecular Devices). Standard curve was prepared with 0–5 μM H2O2.
Advanced oxidation Protein Products (AOPP) are produced through the reaction of proteins with chlorinated oxidants (e.g. chloramines, hypochlorous acid), and provide a sensitive measure of total oxidative stress. AOPP contents were assayed by using the OxiSelect AOPP Assay Kit (Cell Biolabs, San Diego, CA). Briefly, tissue homogenates (10 μg) were mixed with 10 μl of 1.16 M KI and 20 μl of 100% acetic acid (final volume: 200 μl). Reaction was stopped after 5 min and absorbance was recorded at 340 nm. AOPP concentration was expressed as chloramine-T equivalents (standard curve: 0–100 μmol chloramine-T/ml) [55].
A commercially available kit (Abcam ab65329) was utilized to measure the total antioxidant capacity (TAC). The assay uses lag time by antioxidants against the myoglobin-induced oxidation of 2,2'-azino-di(3-ethylbenzthiazoline-6-sulfonic acid (ABTS) with H2O2. Briefly, 20 μl of tissue homogenates (diluted 1:20, v/v) were added in triplicate to 96-well plates, and mixed with 90 μl of 10 mM PBS (pH 7.2), 50 μl of myoglobin solution, and 20 μl of 3 mM ABTS. Reaction was initiated with H2O2 (20 μl) and change in color monitored at 570 nm (standard curve: 2–25 μM trolox).
The NFκB-TATA-luciferase reporter plasmid was graciously provided by Dr. Shao-Cong Sun (University of Texas MD Anderson Cancer Center). The pRL-TK vector containing Renilla luciferase (positive control) was purchased from Promega (Madison, WI). HEK-293 cells (CRL-1573, ATCC Manassas, VA) were cultured in Dulbecco's Modified Eagle Medium (DMEM) media in T75 flasks. Cells were seeded in 96-well tissue culture plates (2.5X104 cells/well), allowed to adhere for 1 h, and acclimatized overnight to antibiotic-free Opti-MEM medium. Cells were transfected with NFκB-TATA-luciferase (100 ng) and pRL-TK (10 ng) using Lipofectamine 2000 (Invitrogen) according to the instructions provided by the manufacturer. After 6 h of incubation, cells were washed, replenished with complete medium for 3–5 h, and then infected with T. cruzi (cell: parasite ratio, 1:3). Cells were incubated in presence or absence of 1 μM each of SRT1720 for 24 h. The relative NFκB transcriptional activity was measured by using a Dual Luciferase Reporter Assay System (Promega, Madison, WI) and data were normalized to Renilla luciferase activity.
All experiments were conducted with triplicate observations per sample (n = 6 mice per group per experiment, at least two experiments per group), and data are expressed as mean ± standard deviation (SD). All data were analyzed using GraphPad Prism 5 (GraphPad Software, La Jolla, CA). Data were analyzed by Student’s t test (comparison of 2 groups) and one-way ANOVA with Tukey’s test (comparison of multiple groups). Significance is presented by * (infected vs. normal) or # (infected/treated vs. infected/untreated) (*, #p<0.05, **,##p<0.01, ***,###p<0.001).
|
10.1371/journal.pcbi.1000944 | Experimental and Computational Analysis of Polyglutamine-Mediated Cytotoxicity | Expanded polyglutamine (polyQ) proteins are known to be the causative agents of a number of human neurodegenerative diseases but the molecular basis of their cytoxicity is still poorly understood. PolyQ tracts may impede the activity of the proteasome, and evidence from single cell imaging suggests that the sequestration of polyQ into inclusion bodies can reduce the proteasomal burden and promote cell survival, at least in the short term. The presence of misfolded protein also leads to activation of stress kinases such as p38MAPK, which can be cytotoxic. The relationships of these systems are not well understood. We have used fluorescent reporter systems imaged in living cells, and stochastic computer modeling to explore the relationships of polyQ, p38MAPK activation, generation of reactive oxygen species (ROS), proteasome inhibition, and inclusion body formation. In cells expressing a polyQ protein inclusion, body formation was preceded by proteasome inhibition but cytotoxicity was greatly reduced by administration of a p38MAPK inhibitor. Computer simulations suggested that without the generation of ROS, the proteasome inhibition and activation of p38MAPK would have significantly reduced toxicity. Our data suggest a vicious cycle of stress kinase activation and proteasome inhibition that is ultimately lethal to cells. There was close agreement between experimental data and the predictions of a stochastic computer model, supporting a central role for proteasome inhibition and p38MAPK activation in inclusion body formation and ROS-mediated cell death.
| Neurodegenerative diseases feature concentration of misfolded or damaged proteins into inclusion bodies. There is controversy over whether these entities are protective, detrimental, or relatively benign. The formation of inclusion bodies may be accelerated by inefficient protein degradation and may promote activation of stress signalling pathways. Each of these events may promote the generation of reactive oxygen species which may exacerbate the problem by damaging more proteins, possibly damaging components of the UPS itself, but in either case further impeding the function of cellular proteolytic systems. To determine how these events are related and which are critical, we generated a live cell imaging system in which inclusion formation and proteolytic efficiency can be evaluated, and created a stochastic computer model incorporating the same components. Laboratory data and computer simulations were found to be in close agreement, supporting a mechanism wherein misfolded protein induced a vicious cycle of stress kinase activation, ROS generation, and proteasome inhibition which was ultimately cytotoxic. Inclusion body formation partially alleviated the burden on the proteolytic system, but may not provide long term benefit. Pharmacological blockade of a stress-activated kinase was effective in breaking the vicious cycle, as predicted by the computer model and confirmed experimentally.
| A hallmark feature of human neurodegenerative diseases is the accumulation of misfolded or otherwise abnormal proteins which become concentrated into large aggregates. Inclusion bodies are large nuclear or cytoplasmic protein aggregates whose predominant constituents may be characteristic of particular diseases. In many cases inclusion bodies (IB) are immunoreactive for ubiquitin and proteasome components [1], indicative of abortive or incomplete proteolysis. The sustained expression of mutant protein with the propensity to misfold may ultimately overwhelm the ubiquitin/proteasome system (UPS) and promote the formation of inclusions. This process may be accelerated by an age-related decline in UPS efficiency (discussed in [2]), which may explain why genetically transmitted neurodegenerative disorders typically affect older individuals. Consistent with the proteasome impairment hypothesis, IB form in the neurons of mice in which proteasome function has been genetically compromised [3]. Because misfolded, damaged, or genetically abnormal proteins are aggregation-prone their sequestration into inclusion bodies may actually alleviate the load on the UPS and promote neuronal survival, at least in the short term. Time lapse microscopy of a fluorescent proteasome reporter in cultured neurons has indicated that the UPS load is partially alleviated upon IB formation [4], and there is evidence that cultured cells forming such inclusions have a survival benefit [5] over the course of the experiment. In the longer term, however, it is possible that deleterious effects from IB formation would become pronounced. Apart from potential physical perturbations imposed by large proteinaceous inclusions (in axons, for example) these entities may wreak havoc by depleting essential cellular components (reviewed in [6]) or by biochemical means. In Huntington's disease, IB form when a polyglutamine tract in the N-terminal region of the huntingtin protein exceeds the threshold length of approximately forty glutamine residues; early onset and severe disease are correlated with very long tracts, whereas huntingtin proteins with polyglutamine tracts shorter than the threshold do not form IB and are not pathogenic [7]. The nuclear IB formed by the mutant huntingtin protein are generators of reactive oxygen species [8], and expression of such an expanded polyglutamine protein results in sustained and ultimately cytotoxic activation of p38MAPK [9]. It is likely that proteasome inhibition, ROS generation, and p38MAPK activation all feature in the death of cells containing IB, but their relative importance and potentially complex interdependencies are poorly understood. We have combined live cell imaging with mathematical modeling to explore such relationships. Our data point to a positive feedback loop between IB formation and p38MAPK activation that likely involves ROS. The existence of this loop is supported by the close agreement of laboratory data and simulations generated by a stochastic computer model.
We have previously demonstrated sustained activation of p38MAPK in cultured mammalian cells expressing expanded polyglutamine proteins and in a transgenic mouse model of expanded-polyglutamine disease [9]. This activation could be abrogated by treatment with the specific p38MAPK inhibitor SKF86002. To confirm the activation of p38MAPK and its inhibition by SKF86002 we performed western blot analysis of U87MG cells expressing HttQ103 (an amino terminal fragment of the human huntingtin protein containing a 103 glutamine tract). Expression of the expanded polyglutamine protein resulted in extensive phosphorylation of HSP-27, a downstream target of p38MAPK (Figure 1A). Reduced phosphorylation of HSP-27 was detected in cells expressing HttQ25 and HSP-27 phosphorylation was undetectable in untransfected control cells. Although the HttQ25 protein contains a polyglutamine tract below the threshold length for pathogenicity, its overexpression upon transfection may be sufficient to activate p38MAPK at a low and nontoxic level as we have documented previously [9]. The phosphorylation of HSP-27 was precluded by pre-treatment of transfected cells with a pharmacological inhibitor of p38MAPK (Figure 1A) or by co-expression with a dominant-negative (kinase dead) variant of p38MAPK (Figure S1) confirming that the phosphorylation of HSP-27 was due to the activation of the p38MAPK pathway by Htt103.
To determine if blockade of p38MAPK activity affects cell survival, we transfected U87MG cells with either HttQ25, HttQ103 or a GFP control plasmid. At 30 hours post-transfection, flow cytometry was performed using propidium iodide exclusion, and revealed that HttQ103 expressing cells exhibited the highest levels of death (25%, Figure 1B) whereas the death associated with Htt25 expression was similar to that of GFP (10–12%). Pre-treatment of HttQ25 and HttQ103cells with SKF86002 resulted in a decrease in cell death that was most pronounced in cells expressing HttQ103 (Figure 1B). The HttQ103 protein is known to inhibit proteasome activity in a cell-based assay [10]; to determine if the cytotoxicity of the expanded-polyglutamine proteins could be enhanced by further proteasome inhibition a pharmacological proteasome inhibitor (PI) was added to HttQ25 and HttQ103 transfected cells. Cells were pre-treated with PI for 6 hours prior to assessing cellular death by flow cytometry. PI-treated U87MG cells expressing HttQ103 exhibited a significant increase in cell death when compared to their untreated counterparts (Figure 1B), and was 15% greater than PI treated HttQ25 expressing cells. Under the same experimental conditions the amount of cell death induced by PI in untransfected cells was approximately 5% (not shown), roughly equivalent to the increase in cell death mediated by PI in cells expressing HttQ25 and HttQ103. Under our conditions proteasomes cannot therefore be fully inhibited by expression of the polyQ proteins alone. The pharmacological data support the argument that the cytotoxicity of misfolded proteins is mediated by proteasome inhibition and p38MAPK activation, but do not reveal whether these activities are independent.
To simultaneously assess functioning of the UPS and inclusion body (IB) formation at a single cell level we created a bicistronic construct that encodes an expanded polyglutamine protein and fluorescent proteasome substrate on the same transcript (see schematic diagram in Figure 2A). This construct was designed to investigate the temporal order of events leading to HttQ103 induced cellular and proteasome toxicities and to dissect the role of p38MAPK in these events.
U87MG cells were transfected with HttQ103YFP-pIRES-mRFPu and were imaged at 10 minute intervals from 24 to 48 hours post-transfection. The time lapse images revealed formation of IB at 36 hours post-transfection in cells expressing HttQ103 (Figure 2B, first panel; time lapse videos are provided as Videos S1 to S8). Values of mRFPu intensity were graphed as a function of time revealing an increase in mRFPu fluorescence prior to the formation of an IB, followed by a period of constant mRFPu intensity. The single cell analysis revealed that HttQ103-induced cellular death is preceded by gradual UPS impairment. Once this impairment reaches a threshold level, IBs begin to form and their formation correlates with a momentary recovery of UPS efficiency as measured by mRFPu intensity. These results are consistent with previously published findings in primary neuron cultures [4].
To examine the extent to which IB formation was dependent on UPS dysfunction U87MG cells expressing HttQ103YFP-pIRES-mRFPu were treated with proteasome inhibitor. As expected PI treatment resulted in a rapid and persistent increase in mRFPu intensity. Under these conditions IB formation was accelerated relative to untreated controls (Figure 2B, second panel). These data suggest an immediate relationship between proteasomal inhibition and IB formation.
Damaged proteins are normally eliminated by the UPS, and we speculated that an increase in cellular ROS levels would lead to oxidative damage to proteins and inflict an additional burden on proteasomes that may affect the kinetics of IB formation. To test this hypothesis, we depleted reduced glutathione levels in HttQ103YFP-pIRES-mRFPu tranfected cells by treating cells with buthionine sulphoximine (BSO) at 24 hours post-transfection. In BSO treated cells we observed a constant increase in mRFPu intensity (Figure 2B, third panel) consistent with a cumulative UPS burden.
Having previously established that the activation of p38MAPK in HttQ103 expressing cells contributes to cytotoxicity, we sought to determine what effects inhibition of p38MAPK signalling pathway would have on IB formation and UPS dysfunction. U87MG cells expressing HttQ103YFP-pIRES-mRFPu were therefore treated with SKF86002. These cells exhibited a low level of mRFPu fluorescence along with a delay in IB formation (Figure 2B, fourth panel). The mRFPu fluorescence remained low and did not feature a rapid increase as seen in the untreated counterparts. These data suggest that inhibition of the p38MAPK pathway decouples the proteasome inhibition from HttQ103 protein expression, resulting in delayed formation of IB.
We quantified IB formation by recording the number of cells with IB at 6 hour intervals starting at 24 hours post-transfection. The percentage of cells with IB was graphed as a function of time (Figure 3A). PI treatment generated the greatest number of IB compared to untreated cells while SKF86002-treated cells were found to have the fewest IB. Comparative single-cell analysis of IB formation corresponded to time-points in which 40% of transfected cells had formed IB. Similarly, we quantified average mRFPu fluorescence from many cells and graphed the fluorescence intensities as a function of time. Treatment with PI and BSO generated the highest amounts of mRFPu fluorescence, while SKF86002 treatment resulted in the lowest levels of mRFPu in comparison to untreated cells (Figure 3B). Findings from multiple cells were consistent with the mRFPu fluorescence levels observed in the single cell analysis.
We have previously used stochastic computer modeling to study the age-related decline of proteolysis [11], and have adapted this model to incorporate p38MAPK in an effort to better understand polyglutamine-mediated cell death. Our objective was to determine if a relatively simple mathematical model incorporating the components thought to be critical for polyQ-mediated cytotoxicity could recapitulate our laboratory findings; if not some critical component must have been overlooked or one or more of the starting assumptions must be invalid. Conversely, a good fit would suggest that the assumptions are valid and no critical elements have been overlooked. The stochastic computer model is represented schematically in Figure 4. The model was constructed using the Systems Biology Markup Language as described in the Materials and Methods section; details of molecular species and reactions are given in Text S1. The model predicted that treatment with PI would lead to reduced cell death at 30 hours (Figure 5A), a short term benefit from reduced levels of small aggregates binding to proteasomes (the consequence of which would be reduced by concentration of aggregates into IB). The experimental data, however, indicated a slight increase in the proportion of cell deaths under PI suggesting that the situation is more complex than is currently accounted for in the model. On the other hand, the model predicted that inhibition of p38MAPK activity should lead to much lower cell death, in close agreement with the experimental data. Also in agreement with the experimental findings described above, the model predicted that proteasome inhibition should lead to an increase in the rate of IB formation compared to untreated cells (Figure 5B). When p38MAPK activity is inhibited the model predicts a lower rate of IB formation at early time-points (Figure 5B) although from 36–48h, the rate of increase is similar to untreated cells (note lines are parallel for polyglutamine and p38MAPK inhibition during this time interval). The later increase in inclusion formation when p38MAPK activity is inhibited is probably due to ROS generation via the aggregated protein. The computer model predicts that much higher levels of mRFPu will be observed in PI treated cells than in untreated cells, as expected for a proteasome substrate (Figure 5C). Critically, inhibition of p38MAPK activity reduced the accumulation of mRFPu in the computer model. Overall the simulations and experimental data are in good agreement, indicating that the model describes the important molecular relationships of the system as portrayed in Figure 4.
Based on our single-cell analysis data, we speculated that the activation of p38MAPK and UPS impairment were contributing to the production of reactive oxygen species (ROS) and that SKF86002 may be counteracting this cellular response. A genetic approach was adopted to test this hypothesis, utilizing transfection of cells with expression vectors encoding wild type or kinase dead versions of p38MAPK. Whereas pharmacological inhibition may affect multiple p38MAPK isoforms any modulation of cellular response observed with the genetic approach would be attributable to the alpha isoform of p38MAPK exclusively. We first confirmed that in our U87MG cell system the expression vectors were capable of modulating p38MAPK activity using phosphorylation of HSP27 as a proxy marker (Figure S1). Cells were then transfected with the p38MAPK expression vectors or with an empty vector control construct and lysates were assayed for reduced glutathione (GSH) content, a marker of oxidative status within cells. We found that overexpression of wild type p38MAPK resulted in a significant decrease in GSH levels whereas the GSH content in cells expressing kinase dead p38MAPK was less affected in comparison to controls (Figure 6A). To determine if overexpression of wild type p38MAPK resulted in dysfunction of the UPS, we transfected the p38MAPK expression constructs into a cell line stably expressing GFPu, a well characterized proteasome sensor [10]. For these experiments we made use of a stable NIH 3T3 cell line we had previously generated; the GFPu reporter is expressed at a lower level in the stable line and does not accumulate as an artefact of transfection-mediated overexpression. By flow cytometric analysis, we found that cells overexpressing wild type p38MAPK exhibited the highest levels of GFPu intensity whereas the levels of GFPu in cells expressing the kinase dead p38MAPK or the empty vector were similar (Figure 6B). These data suggested that the activation of p38MAPK was negatively affecting UPS function. To test whether this inhibition of the proteasome was a direct affect of p38MAPK activation, we transfected cells with the same set of plasmids and assayed their proteasome activity using fluorogenic substrates. No significant differences were noted (Figure S1). Taken together, the data indicate that the expression of p38MAPK does affect the oxidative status of cells but does not directly inhibit the proteasome.
Since aggregated protein leads to increased levels of ROS, inhibition of p38MAPK may simply delay cell death and thereby allow more aggregation to take place. To determine the predicted outcome should p38MAPK not be involved in generating more ROS we removed the reaction for ROS generation via p38MAPK from the model and repeated the computer simulations. Without p38MAPK-generated ROS less cell death was predicted under all conditions (Figure 7A). With the feedback loop broken in this way the model also predicted that there would be no significant difference in the numbers of inclusions at each time point between untreated cells and cells treated with a p38MAPK inhibitor (Figure 7B). We also refitted the model for HttQ103 without treatments and no feedback loop using the experimental data for cell death and inclusion formation (Figure 7C–D). Since the model predicted less cell deaths and lower levels of inclusions than the data, we increased the parameters for inclusion formation (kaggPolyQ) and cell death (kp38death and kPIdeath). We also increased the parameter for ROS generation via p38MAPK (kgenROSp38) since this had the effect of increasing both the levels of inclusions and the number of cell deaths. We then ran the model with the treatment for proteasome inhibition and p38MAPK inhibition. The model was not able to reproduce the decline in cell death or the lower levels of inclusions when p38MAPK is inhibited (evident from the comparison of Figure 7C–D with Figure 5A–B). Therefore the model indicates that a feedback loop from p38MAPK to ROS is required to explain the experimental data.
Based on the experimental data and the computer simulations reported herein we propose a vicious cycle mechanism of polyglutamine-mediated cytotoxicity (Figure 8). In the proposed mechanism the initiating event is the inhibition of the proteasome by small aggregates of misfolded protein. It has been previously shown that proteasome inhibition leads to the generation of ROS (reviewed in [12]) and the activation of MKK3 and MKK6 kinases upstream of the stress kinase p38MAPK [13], [14]. Activation of p38MAPK by its upstream regulators may exacerbate the problem by promoting downstream ROS production (through a mechanism discussed below). By damaging other cellular proteins the reactive oxygen would provide an additional burden to the UPS, which is normally charged with the proteolytic degradation of damaged or abnormal proteins. Increasing proteasome inhibition would lead to further accumulation of misfolded proteins, ultimately coalescing into inclusion bodies. By reducing local concentrations of the small aggregates that are thought to be most inhibitory to the proteasome the IBs may temporarily alleviate proteasome inhibition, but by concentrating iron they may promote the further generation of ROS, ensuring yet more damage and the conditions that will sustain p38MAPK activation. Whether or not there is self-amplification of the vicious cycle as a consequence of increasing ROS production we postulate that it is sustained p38MAPK activity and proteasome inhibition that will ultimately lead to cell death.
The mathematical model was based on our previous model of the ubiquitin/proteasome system which we modified and extended to include turnover and aggregation of polyglutamine proteins. The model predictions were in close agreement with the experimental data indicating that the proposed network in Figure 4 captures the important components in the system. However, we found that there were some discrepancies between the model predictions and experimental data regarding cell death at early time-points under conditions of proteasome inhibition. This suggests that there is something missing from the model. Since proteasome inhibition affects all protein turnover, the missing part could be a pro-apoptotic protein. One such candidate is the transcription factor p53 which is normally rapidly turned over by the proteasome. A future extension of our current model to include a pro-apoptotic protein would be fairly straightforward as we have previously modelled the p53 system [15]. An advantage of modelling over laboratory experiments is that it is easy to manipulate the system on a computer, whereas the same experiments in the laboratory may be impossible or very costly to do. For example, although it would be difficult to prove experimentally that ROS generation by p38MAPK is required to explain the experimental data, the hypothesis could easily be tested in the mathematical model by simply removing the reaction of p38MAPK-dependent ROS generation and repeating the simulations. The model confirmed that p38MAPK is involved in generating more ROS since with the feedback loop broken the model output no longer fitted the experimental data.
We have previously shown that pharmacological blockade of p38MAPK can protect cells from polyglutamine-mediated cell death [9], and the current experimental data and mathematical modeling provide an explanation for the efficacy of this intervention: by breaking the vicious cycle the p38MAPK inhibitor precludes further damage and proteasome inhibition. The data we present herein support a central role for ROS in the proposed cycle of p38MAPK activation, proteasome inhibition, and protein aggregation that ultimately leads to cell death. The importance of ROS in protein misfolding disorders is not a new concept; indeed the cytotoxicity of elevated ROS generated by mutant huntingtin has been convincingly demonstrated by the laboratory of Rubinsztein [16]. Intriguingly, the proteasome is itself an important regulator of oxidative damage in neurons and proteasome inhibition is known to induce mitochondrial dysfunction and promote oxidative damage to DNA and protein (reviewed in [12]). Once some threshold of polyQ-mediated proteasome inhibition is reached it seems entirely plausible that a self-perpetuating loop of ROS generation and p38MAPK activation could lock the cell into a dysfunctional state. Indeed, a directly analogous positive feedback loop involving ROS and p38MAPK activation was recently proposed for the induction of senescence in mammalian cells. By combining bioinformatics, stochastic computer modeling, and direct experimental interventions Passos et al. have provided convincing evidence that sustained activation of p38MAPK is required for generation of mitochondrial ROS, which by generating DNA damage ensures the continued activation of p38MAPK in senescent cells [17]. The ROS generated as a consequence of proteasome inhibition also appears to be of mitochondrial origin [18], so the cascade of events in cells expressing misfolded protein may be very similar to that in cells in which the initiating event is exposure to ionizing radiation (as was the case in the Passos paper). If a similar loop were operating in cells expressing expanded polyglutamine proteins one might expect to find ROS-mediated DNA damage leading to activation of ATM and phosphorylation of histone H2AX, as has indeed been documented in cells from Huntington's disease and SCA-2 patients [19]. The activation of ATM and formation of H2AX repair foci appears to precede the formation of IB [20], but may be coincident with p38MAPK activation and the generation of ROS. If our model is correct pretreatment of cells with the p38MAPK inhibitor should reduce the number of H2AX foci in cells expressing HttQ103. We are currently testing this hypothesis.
Enhanced levels of autophagy may also contribute to the protective effects of the p38MAPK inhibitor we have observed. Although the effects may be dependent on cell type, there is evidence that the alpha isoform of p38MAPK inhibits autophagy [21] [note that this is the same isoform utilized in our genetic experiments, and may be the target of primary importance to all of the interventions described herein]. Low level inhibition of the proteasome is known to promote autophagy [22], but as proteasome inhibition increases the activation of p38MAPK may limit autophagic clearance of protein aggregates. Enhancement of autophagy has been proposed as a therapeutic strategy for the treatment of polyglutamine disorders such as Huntington's disease [23], and pharmacological blockade of p38MAPK may provide benefit by multiple mechanisms. It may break the vicious cycle of ROS generation while promoting autophagy-mediated clearance of aggregates in cells already compromised for proteasome function.
We have performed the current set of experiments in U87MG cells, a human glioblastoma cell line. These cells were chosen for their ease of handling, including their high transfection efficiency. Because the critical components of the proposed vicious cycle are universally present in mammalian cells (including p38MAPK, proteasomes, and ROS from mitochondria) it is likely that a self-perpetuating loop could be triggered by expanded polyglutamine proteins in any cell type, and we have previously demonstrated the protective effect of a p38MAPK inhibitor in normal and transformed cells of both human and mouse origin [9]. We nevertheless recognize that the kinetics of the events we have described may be markedly different for neurons in situ.
Although the short term benefit of IB formation (through improved proteasome function) has been demonstrated in cultured neurons [4], [5] the longer term implications of IB in neurodegenerative diseases must be considered. A transient alleviation by IB of the polyglutamine-mediated proteasome burden was recently demonstrated in vivo in an inducible model of Huntington's disease [24], but no long term impairment of proteasome function was observed in the brains of these mice. If the IB concentrates iron and promotes ROS generation through Fenton chemistry [8] the IB, once formed, might ensure the perpetuation of the vicious cycle proposed in Figure 3C. The failure to detect this effect in older mice expressing a polyQ protein may relate to the inability of the reporter system to detect less than a 40% decrease in proteasome inhibition [24] or may relate to differences between the mouse model and the human disease. In human polyglutamine disorders the disease may progress over years if not decades, and the later consequences of IB formation might therefore be more severe. Clinical evidence supports a deleterious role for IB in human disease. In polyglutamine repeat diseases such as HD, frontotemporal lobar degeneration, and RNA-mediated diseases such as myotonic dystrophy, inclusion-mediated titration of transcription factors (like CBP), tar DNA-binding protein 43 (TDP-43), and RNA splicing factors (for example muscleblind), respectively, may represent an important molecular mechanism of disease [6]. Added to these effects would be the deleterious effects we ascribe to the ROS-mediated vicious cycle.
The pEGFP-N1 expression construct was purchased from Clontech (Palo Alto, California, USA). Htt-Q25 and Htt-Q103 expression constructs containing a synthetic insert encoding exon 1 or human Huntingtin containing a polyglutamine tract of either 25Q or 103Q fused to the yellow fluorescent reporter protein (YFP) were generous gifts from Dr. Ron Kopito (Stanford University). These are designated HttQ25YFP or HttQ103YFP. A red fluorescent proteasome reporter was generated by PCR-mediated transfer of the degron sequence from the GFPU reporter (Bence et al.; the gift of Dr. Ron Kopito) to the C terminus of the monomeric red fluorescent protein (the gift of Dr. Robert Campbell, University of Alberta). Under normal conditions, mRFPu is quickly degraded by the 26S proteasome, but during conditions of proteasomal impairment, turnover of mRFPu is reduced, leading to an accumulation of mRFPu that is visible by fluorescent microscopy. To simultaneously express the expanded YFP-tagged polyglutamine proteins and the red fluorescent proteasome reporter the former was inserted into NheI site upstream of the internal ribosome entry site (IRES) in the vector pIRES (Clontech, Palo Alto, California, USA) and the latter was inserted between the Xba I and Sal I sites downstream of the IRES element. The wild type and kinase dead p38MAPK variants were generous gifts from Dr. J. Han (The Scripps Research Institute, La Jolla, CA). The hyper-active p38MAPK construct was a gift from Dr. Oded Livnah (The Hebrew University of Jerusalem).
The human U87MG glioblastoma cells (a gift from Dr. I. Lorimer at the Ottawa Hospital Research Institute) were maintained in Dulbecco's modified Eagle's medium (DMEM) and supplemented with 10% FBS and maintained in a 37°C incubator with 5% CO2. For transient transfections, cells were plated in either 96- or 6 well dishes 24hours prior to transfections. Subsequently, they were transfected using GeneJuice Transfection Reagent (Novagen, Madison, WI, USA) as per the supplier's protocol. 0.5µg or 3.0µg of plasmid DNA was used in each well of a 96 or 6 well dish. For p38MAPK inhibition experiments, cells were pre-treated for 2h with 20µM SKF86002 (Calbiochem) prior to transfection with various expression constructs.
U87MG cells were harvested in protein lysis buffer consisting of 100mM Tris pH 6.8, 20mM DTT, 4% SDS, 5% glycerol. Protein concentrations were determined using the Bradford assay reagents (Bio-Rad, Hercules, CA, USA). Reduced proteins were resolved on a 10% SDS-polyacrylamide gel and electro-blotted onto a Hybond C nitrocellulose membrane (Amersham Bioscience Corp, Baie d'Urfé, QC). The membranes were stained with Ponceau S prior to immunoblotting with phospho-HSP-27 (polyclonal rabbit), total HSP-27 (rabbit polyclonal) (Cell Signaling, Danvers, MA), or Actin (Sigma-Aldrich). Proteins were detected using the HRP method and SuperSignal West Pico Chemiluminescent Substrate reagents (Pierce, Rockford, IL, USA). Proteins were visualized using the GeneGnome (Syngene, Frederick, MD, USA).
Cell viability was assessed by flow cytometry using propidium iodide exclusion. Adherent and non-adherent cells were transfected with various constructs for 30 hours, harvested and stained with Propidium Iodide. For p38MAPK inhibition experiments, cells were pre-treated for 2h with 20µM SKF86002 (Calbiochem) prior to transfection with various expression constructs. For proteasome inhibition experiments, cells were treated with Proteasome Inhibitor I (Calbiochem) 24h post-transfection at a final concentration of 25µM. Fluorescent detection was analyzed by flow cytometry using a Beckman Coulter Quanta SC MPL. Data and analysis were done using Quanta Analysis software (Beckman Coulter, Inc., Brea, CA, USA).
75 000 U87MG glioblastoma cells were seeded onto a Delta T4 culture dish system (Bioptechs, Butler, PA) and maintained in a 37°C incubator with 5% CO2 for 24h hours. Cells were transfected with 2ug of plasmid DNA encoding HttQ103YFP-pIRES-mRFPu for 24 hours before being transferred onto a heated stage maintained at 37°C and at 5% CO2 using a Delta T4 culture dish temperature controller and cell perfusion system (Bioptechs, Butler, PA). For p38MAPK inhibition experiments, cells were pre-treated for 2h with SKF86002 for a final concentration of 20µM to preclude kinase activation upon transfection. For proteasome inhibition experiments, cells were treated with Proteasome Inhibitor I (Calbiochem) 24h post-transfection at a final concentration of 25µM (treatment with PI prior to transfection is not possible due to its immediate toxicity). For buthionine sulphoximide (BSO)-induced depletion of glutathione experiments, cells were treated 24h post-transfection with BSO (Sigma) 24h post-transfection at a final concentration of 5mM. Microscopy was performed 24 hours post-transfection on a Zeiss Axiovert 200M inverted fluorescent microscope for a total of 24 hours. Fully automated multidimensional acquisition was controlled using Axiovision 4.8 software. Images were acquired using a 10× objective (EC Plan-Neofluar) with a side-mounted AxiocamHRm camera. Yellow fluorescent protein or red fluorescent protein was excited using the Zeiss Colibri LED illumination system (LEDmodule 505nm or LED module 590nm) and detected using the appropriate filters (46HEYFP or 61HEGFP/HcRED, respectively). Fixed exposure times were as follows: Brightfield phase contrast 1ms; YFP 100ms; RFP 188ms. Images were taken at 10 minute intervals for 48 hours and compiled into video files using Axiovision 4.8 software (Carl Zeiss, Thornwood, NY).
Cell lysates from U87MG cells over-expressing wild type p38MAPK or kinase dead p38MAPK were analyzed for reduced GSH content using a luciferase kit (GSH-Glo) from Promega (Madison, WI). 10,000 cells were seeded in 96-well plates and transfected with 0.5µg of DNA for 48 hours. Cells were collected and analyzed for GSH following the manufacture's protocol.
Mouse NIH 3T3 cells were co-transfected with GFPu and a PGK-driven puromycin resistance gene (gift of Dr. M. McBurney, Ottawa Hospital Research Institute). Cells stably expressing the proteasome reporter GFPu were selected over 2 weeks in a final concentration of 2.0µg/ml. For proteasome assays, wild type p38MAPK, kinase dead p38MAPK, or pcDNA were transfected into 3T3-GFPu cells for 48 hours. Cells were collected and analyzed for GFPu expression by flow cytometry (Beckman Coulter Quanta SC MPL). Mean GFP intensity was analyzed using the Quanta Analysis software and subsequently graphed using Excel (Microsoft).
Statistical significance was determined by a two tailed Student's t-test. Unless otherwise indicated, values were considered significant when p<0.05.
The model was developed to mimic the experimental system so that simulations could be performed to see which parameters affected the different cellular outcomes. The model was initially fitted to experimental data where HttQ103 had been added to cells but without any inhibitors. The model was then used to mimic the experimental treatments and the model predictions were compared to the experimental results. If there were discrepancies between the model predictions and experimental results, the model was modified (either by changing parameter values or by the addition of further reactions). The model was then re-run for the experiment without any treatments to check that it still fitted the experimental data. If the model did not fit, further adjustments were made and the procedure repeated. We originally started with a model that did not contain a feedback loop from p38MAPK to ROS but found that it was necessary to include this loop in order to get the model to fit both the data for the treatment with p38MAPK inhibitor and the data without the treatment. Further details are given in Text S1.
The model was encoded in the Systems Biology Markup Language (SBML) as this standard allows models to be easily shared, modified and extended [25]. SBML is a way of representing a network of interactions so that it can be simulated on a computer and the evolution of the system over time can be followed. SBML shorthand was used to create the SBML code which was then converted into full SBML [26]. The network diagram is given in Figure 4, and Tables S1 and S2 in Text S1 give details of the species and reactions respectively. Text S1 also contains more detail of events and parameter values (Tables S3 and S4 in Text S1). Since there is large variability in cellular outcomes in terms of both inclusion formation and cell death, we used stochastic simulation. This was based on the Gillespie algorithm [27] which assumes that collisions of molecules occur within a reaction vessel and that at most only two molecules can collide. We chose this method of simulation as there are low copy numbers of many of the species and random effects play a major role in this model, as can be seen by the cell to cell variability of the model output. It should be noted that we have a few reactions which have more than two molecules in the list of reactants. These are the reactions for the aggregation of polyQ where we assume that ROS affects the reaction kinetics although ROS itself is not consumed by the reaction (and so also appears in the list of products). We use a function of ROS in the kinetic law so that we have a pseudo second order reaction rather than a third order reaction. Although this may seem to be a violation of the assumptions of the Gillespie algorithm, this provides a simple way to allow for the effects of ROS on the aggregation process rather than adding many more reactions and parameters. The model is available from the Biology of Ageing e-Science Simulation and Integration (BASIS) system ([28], [29]) and the Biomodels database (ID:MODEL1002250000) [30].
We assume that the addition of the polyQ gene to the cell resulted in continuous synthesis of the polyQ protein. It is also degraded by the proteasome so that total levels remain fairly constant with a half-life of about 20 hours [31]. We set the levels of polyQ sythesis and degradation so that the half-life would be 20 hours if the proteasome did not become inhibited by aggregates. Two molecules of polyQ interact to form a small aggregate (AggPolyQ1). We assume that ROS affects the aggregation kinetics if it rises above basal levels. The aggregate can grow in size by the addition of further polyQ proteins in a reversible manner. However, when the aggregate reaches a certain threshold size, we assume that disaggregation can no longer take place and that instead an inclusion forms (SeqAggP). This threshold represents the seed and is assumed to be of size six based on data for amyloid fibril polymerization [32]. Since mutant huntingtin forms amyloid-like filaments, it is reasonable to assume that it has similar aggregation kinetics [33], [34]. A very recent study shows that mutant huntingtin forms three major pools: monomers, oligomers and inclusion bodies [35]. Interestingly, the study showed that the pool of oligomers as a proportion of total huntingtin did not change over a time period of 3 days despite continued conversion of monomer to inclusion bodies. We also compared the levels of oligomers (represented by the species AggPolyQ[i], where i = 1–5) in our simulation output in cells which formed inclusion bodies (Figure S2) and discuss the results in the Text S1 section. It has been shown that small aggregates bind to proteasomes and inhibit proteasomal function [36]. Therefore, we assume that AggPolyQ can bind to the proteasome and so reduce the pool of available proteasomes. However we assume that inclusions do not interfere with the degradation machinery. We also include a species to represent mRFPu and assume that this is turned over with a half-life of about 30 minutes [36]. We assume that ROS is continuously generated and removed with a half-life of 1 hour [37] but that basal levels are low (about 10 molecules). We assume that small aggregates may generate ROS, so that the level of ROS is dependent on the amount of small aggregates (either bound to the proteasome or free pools). We also assume that the presence of inclusions will increase levels of ROS but with a much smaller effect than small aggregates. We represent p38MAPK in two forms: unphosphorylated (p38) and phosphorylated (p38-P) with p38-P being the active state. We assume that high levels of ROS activate p38MAPK and that high levels of p38-P initiate a signalling cascade that results in cell death. We set the rate of this reaction so that it is unlikely to occur when p38-P levels are low and the probability of the reaction occurring increases with increasing levels of p38-P. However, since the model is stochastic, it is possible that even low levels of p38-P will occasionally signal for cell death. We also assume that if the level of proteasomes bound by aggregates increases above a threshold of about 50%, then another signalling pathway leads to cell death due to the accumulation of the pro-apoptotic protein p53. As in the p38 death pathway, cell death due to aggregates inhibiting the proteasome may occur even when levels of AggP Proteasome are fairly low. The reactions for the cell death pathways are shown in Table S2 in Text S1. After cell death occurs, a dummy parameter kalive is set to zero to prevent further reactions occurring, and a dummy species to record the cause of cell death is set to 1. This makes it possible to plot the time of cell death, the cause of death and to count the number of cell deaths of each type in multiple simulations.
We also assume that proteasomes bound by AggP, polyQ or mRFPu may be sequestered into inclusions if degradation does not take place. If AggP is sequestered into inclusions, then this will help alleviate the increase in ROS due to protein aggregation, since we assume that small aggregates lead to greater ROS generation than inclusions.
We also include a generic pool of protein (NatP) which can misfold to become (MisP). We assume that misfolded protein can be either refolded, ubiquitinated and degraded or at high concentrations it may start to aggregate. Once an inclusion forms, misfolded protein may be sequestered into the inclusion body, including MisP bound to proteasomes.
|
10.1371/journal.pntd.0003447 | Review on Dog Rabies Vaccination Coverage in Africa: A Question of Dog Accessibility or Cost Recovery? | Rabies still poses a significant human health problem throughout most of Africa, where the majority of the human cases results from dog bites. Mass dog vaccination is considered to be the most effective method to prevent rabies in humans. Our objective was to systematically review research articles on dog rabies parenteral vaccination coverage in Africa in relation to dog accessibility and vaccination cost recovery arrangement (i.e.free of charge or owner charged).
A systematic literature search was made in the databases of CAB abstracts (EBSCOhost and OvidSP), Scopus, Web of Science, PubMed, Medline (EBSCOhost and OvidSP) and AJOL (African Journal Online) for peer reviewed articles on 1) rabies control, 2) dog rabies vaccination coverage and 3) dog demography in Africa. Identified articles were subsequently screened and selected using predefined selection criteria like year of publication (viz. ≥ 1990), type of study (cross sectional), objective(s) of the study (i.e. vaccination coverage rates, dog demographics and financial arrangements of vaccination costs), language of publication (English) and geographical focus (Africa). The selection process resulted in sixteen peer reviewed articles which were used to review dog demography and dog ownership status, and dog rabies vaccination coverage throughout Africa. The main review findings indicate that 1) the majority (up to 98.1%) of dogs in African countries are owned (and as such accessible), 2) puppies younger than 3 months of age constitute a considerable proportion (up to 30%) of the dog population and 3) male dogs are dominating in numbers (up to 3.6 times the female dog population). Dog rabies parenteral vaccination coverage was compared between “free of charge” and “owner charged” vaccination schemes by the technique of Meta-analysis. Results indicate that the rabies vaccination coverage following a free of charge vaccination scheme (68%) is closer to the World Health Organization recommended coverage rate (70%) than the achieved coverage rate in owner-charged dog rabies vaccination schemes (18%).
Most dogs in Africa are owned and accessible for parenteral vaccination against rabies if the campaign is performed “free of charge”.
| Rabies is one of the most fatal diseases in both humans and animals. A bite by a rabid dog is the main cause of human rabies in Africa. Parenteral mass dog vaccination is the most cost-effective tool to prevent rabies in humans. Our main objective was to review research articles on the parenteral dog rabies vaccination coverage in Africa. We aimed to review published research articles on percentage of dogs owned and percentage of dogs vaccinated against rabies, and on the relation between vaccination coverage and cost recovery. We followed the standard procedures of a systematic literature review resulting in a final review of 16 scientific articles. Our review results indicate that only a small percentage of African dogs is ownerless. Puppies younger than 3 months of age constitute a considerable proportion of the African dog population. There are considerably more male dogs than female dogs present within the dog population. The dog rabies parenteral vaccination coverage following a “free of charge” vaccination scheme (68%) is closer to World Health Organization recommended threshold coverage rate (70%) compared to the coverage rate achieved in “owner-charged” dog rabies vaccination schemes (18%). In conclusion, most dogs in Africa are owned and accessible for vaccination once the necessary financial arrangements have been made.
| Rabies is one of the infectious diseases with the highest human case fatality rate (almost 100%)[1]. Globally, rabies is responsible for more than 60,000 human deaths, while approximately 15 million people receive rabies post exposure prophylaxis (PEP) annually. More than 95% of the global deaths occur in Asia and Africa, where canine rabies is enzootic [2]. Africa contributes to 43% of the human deaths due to rabies [3]. In addition to human life losses, rabies is also a cause of substantial livestock losses [4] and a threat to rare carnivores like the Ethiopian wolf (Canis simensis) [5] and the African wild dog (Lycaon pictus) [6]. Despite these consequences, rabies has been seriously neglected in Africa [7].
The main cause of transmission of rabies to human in Africa is by a bite of a rabid dog [8]. Once bitten by a rabid dog, development of the disease in human can be prevented by an appropriate post-exposure prophylaxis (PEP). However, PEP is relatively expensive and not always available. Moreover, PEP lacks long-term benefits as it will not stop the virus transmission from rabid dogs to other humans or dogs [9]. Dog rabies parenteral vaccination is therefore more cost-effective measure in preventing human rabies [10].
To eliminate rabies from the dog population in an endemic area at least 70% of the dog population needs to be vaccinated during an annual rabies mass vaccination campaign [11]. In many African countries, the proportion of dogs vaccinated against rabies is far below 70% [12]. Accessibility of free roaming dogs for vaccination is often mentioned as an operational constraint [13] with the assumption that parenteral dog vaccination requires catching and restraining dogs physically. Catching free roaming dogs is easier if the dogs are owned. Therefore, dog ownership is an important factor in determining the percentage of dogs vaccinated during a campaign. Dog ownership status and management factors in developing countries in relation to dog rabies vaccination have extensively been addressed in literature (see for example [14]). But, as developing countries at different continents have a wide variation in social and cultural context, studies on African specific socio-economic situations related to dogs and rabies are a necessity for a valid interpretation and practical application of effective vaccination campaigns in Africa. Besides, there is no valid evidence to what extent charging owners for the costs of dog vaccination against rabies contribute to a low vaccination coverage.
Therefore, the objective of our study is to systematically review articles on parenteral vaccination coverage on dog rabies achieved in Africa, in relation to dog demographics and financial arrangements on vaccination costs.
To obtain insight in the trend of peer reviewed articles focussing on “dog rabies control in Africa” during the last 20 years (1994–2013), a systematic search was made in the databases of CAB abstracts (EBSCOhost and OvidSP), Scopus, Web of Science, PubMed, Medline (EBSCOhost and OvidSP) and AJOL (African Journal Online).
Subsequently, a search was made in the above mentioned databases for peer reviewed articles in the themes: 1) “rabies control”, 2) “dog vaccination coverage” and 3) “dog demography”. All theme searches were limited to papers regarding the continent of Africa. The search for each theme was conducted in the search items “title/abstract/key words” using the following search protocol: 1) ((dog? OR canine OR livestock OR human? OR wild? life) AND rabies AND control AND Africa?), 2) ((dog? OR canine) AND rabies AND vaccine* AND coverage AND Africa?) and 3) ((dog OR canine) AND (demography OR population) AND Africa?). The search protocol was designed, based on standard procedures of a systematic literature search [15]. However, as the AJOL database webpage has no feature to select the search protocol in title/abstract/keywords, the search in AJOL was done within the entire article. The systematic literature search included articles published between 1990 and January 2014.
Publications were screened systematically according to the schematic framework as shown in Fig. 1 using EndNote X5 (Endnote @ 2013) reference manager. First, an evaluation of titles and abstracts was performed followed by a removal of duplicates (i.e., publications indexed in more than one databases and published in more than one format, including conference proceedings and book chapters). Several inclusion and exclusion criteria were considered including year of publication (viz. ≥ 1990), type of study (cross sectional), objective(s) of the study (parenteral vaccination coverage rates on dog rabies in Africa, dog demographics and financial arrangements with respect to vaccination costs), language of publication (English) and geographical focus (Africa). As a result, a publication could be excluded for more than one reason, making it impractical to reflect the number of publications excluded per criteria. Articles, of which the full text was not electronically available, were requested from the Royal College of Veterinary Surgeons Trust Library (in United Kingdom).
For each selected article, a record was made in Microsoft Excel describing the studied dog population by the main purpose for keeping dogs, dog demography (mean age, age distribution, sex ratio age and sex distribution), ownership status (percentage of owned, free roaming and ownerless dogs), and parenteral vaccination coverage based on either”free of charge” or”owner charged” financial arrangements. The selected articles encompassed research from Southern (South Africa, Madagascar, Zimbabwe, Zambia), Central (Chad), Northern (Tunisia), Eastern (Tanzania, Kenya, Ethiopia) and Western (Nigeria) Africa. The methodological quality of candidate peer review articles was critically assessed by the Assessment of Multiple Systematic Reviews (AMSTAR) measurement tool [16, 17] and (supporting information: S1 Checklist).
Results on dog demography (i.e. sex ratio and mean age of dogs) and ownership status were extracted from the selected articles and presented in tabular form without further analysis. In a meta-analysis, we evaluated the difference in the percentage of dogs vaccinated against rabies (i.e. vaccination coverage) by the applied financial arrangement on vaccination costs; i.e. whether the vaccination was provided for free to the dog owners or not (i.e. free versus charged). From the selected articles the presented parenteral vaccination coverage was entered as an event rate in the meta-analysis software (Comprehensive meta-analysis V2, 2013). A Forest plot was created to serve as a visual representation of the data in a combined point estimate for the free and charged vaccination study groups, bounded by its confidence interval. Statistical differences, called heterogeneity tests, between the two groups of studies were tested as indicated by I2 and tau square. I2 represents the percentage of the total variation across studies due to heterogeneity across studies within a group and across a group. It takes values from 0% to 100%, with the value of 0% indicating no observed heterogeneity. Tau square is an estimate of the between study and between group variance. If greater than 1, it suggests the presence of substantial statistical heterogeneity in each group, which is a statistical variation due to heterogeneity rather than chance between the free-of-charge and charged study groups [18, 19].
Based on the systematic literature search using the phrase “Rabies control in Africa”, the highest number of research publications was found indexed by the Web of Science/Knowledge database. In Fig. 2 the trend in the number of scientific publications on the topics “Rabies in Africa”, “Dog/Canine rabies in Africa” and “Control of “Dog/Canine rabies in Africa” during the last 20 years indicates an increase in scientific interest on rabies in Africa (Fig. 2). While a worldwide search on “Rabies” during the same period resulted in 9,836 selected entries, only 328 of them were specifically referring to the African situation. Of these 328 papers, approximately half focussed specifically on dog rabies (“Dog/Canine rabies in Africa”, n = 172) and one fifth on dog rabies control (“Control of “Dog/Canine rabies in Africa”, n = 76).
The systematic search from the databases by the defined framework resulted in 1,239 articles (Fig. 1). After removal of all duplicates and exclusion of publications not fulfilling the selection criteria, 16 peer reviewed articles remained to be included in the study on dog demography (sex ratio and mean age of dogs) and ownership status (owned confined, owned but free roaming and proven to be ownerless). Seven of these papers have been published during the last five years. For the comparison of dog parenteral vaccination coverage against rabies related to the dog owners’ costs of vaccination, 11 peer reviewed articles with 15 entries (including four studies comparing free and owner charged vaccination arrangements) remained. The majority of these papers (7 out of 11) has been recently published (e.g., between 2009–2013).
The majority (up to 98%) of dogs in African countries is kept for socio-economic purposes including guarding livestock from predators, homestead from intruders, crops from wildlife and hunting. Dogs are also used as pets, income generation means and as a protein source (Table 1) [20, 21, 22, 23]. Furthermore, puppies younger than 3 months of age constitute up to 30% of the dog population [24]. Male dogs dominate the female dogs up to 3.6 times in number within the population [22]. The mean age of the dogs varies between 1.8 and 3.4 years. Studies accounting for ownership of dogs (Table 1) showed that the percentage of ownerless dogs ranges between 0.7% and 20% of a dog population within the 11 represented African countries. Except for a study in Tanzania [24], all studies reported that more than two third of the free roaming dogs has a responsible owner[20, 22, 23, 25, 26, 27]. Owned dogs with confined housing constitute 18.5% to 60.9% of the dog population.
The published studies selected for vaccination coverage comparison by vaccination costs arrangement schemes consisted of eleven studies in eight different countries representing all regions of Africa (Table 2). Four studies compared vaccination coverage under “free of charge” and “charged” arrangements schemes, four studies evaluated vaccination coverage resulting from “charged” vaccination arrangement schemes only and three studies estimated parenteral vaccination coverage by a “free-of-charge” scheme. The Forest plot (Fig. 3) shows a coverage of less than 50% in the charged groups except for one study, while all studies under free of charge arrangements resulted in a coverage above 50%. The vaccination coverage in studies based on free of charge vaccination (68%) is significant higher ((P<0.001) than the studies based on a charged vaccination campaign (18.1%).
Table 3 provides the heterogeneity test results from the two groups of vaccination cost arrangement schemes. I² describes the percentage of variation across groups due to heterogeneity rather than chance. This study shows 99.9% heterogeneity between free and owner charged groups, indicating significant difference between the vaccination coverage in the two groups. Tau square is an estimate of the between-study variance in the meta-analysis. As Tau square between studied groups is larger than one (i.e. 1.45), it shows a substantial heterogeneity between the studied groups, while Tau squares within the free and charged groups were smaller than one (0.16 and 0.54, respectively).
According to the World Health Organization (WHO), the adequate vaccination coverage of a dog population in a community vaccinated annually against rabies should be at least 70% in order to block the occurrence of an outbreak [1]. In this study dog accessibility for parenteral vaccination reflected by the ownership status and vaccination costs arrangement schemes were assessed to explore their influence on the realised vaccination coverage in Africa
When resources have to be allocated to the control of a disease, this should be done on scientific evidence. For instance, the organized efforts of the Pan American Health Organization (PAHO) in Latin America [28] and the Bohol Rabies Prevention and Elimination Project of the Philippines [29] have witnessed the possibility of reducing the incidence and burden of rabies with concerted efforts of experts. In Africa also, as growing scientific interest was shown through publications produced in the last few decades, it is possible to control rabies with organization of resources from different stakeholders together with a high local community involvement.
Accessibility of dogs is perceived to be the major operational constraint to achieve adequate coverage for dog vaccination against rabies through mass dog vaccination schemes[13]. Our study shows that the majority of dogs in Africa is free roaming but owned. Dogs having responsible owners are accessible for parenteral mass vaccination indicating the possibility of achieving the minimum proportion of dogs that ought to be vaccinated to reduce the incidence of rabies. However, it doesn’t mean that all owned dogs are presented for vaccination[30]. Oral rabies vaccination could be an option for those dogs that are difficult to capture, whether these dogs are owned or ownerless [31, 32]. As long as the proportion of ownerless dogs is less than 20% it is still possible to obtain sufficient immunity coverage by focussing on the mass vaccination of owned dogs. The relative impact of ownerless dogs could be studied by looking at the proportion of ownerless dogs compared to owned dogs in reported cases of human bites. For instance, in Nigeria only 9.7% of the dog bites could not be traced back to a dog with a responsible owner [33]. A study in South Africa showed that only a small proportion of dog bite reports resulted from unknown dogs [34]. In Chad, only 3% of the biting dogs were ownerless or from an unknown owner [35].
Studies referred in this review showed that the mean age of the African dog is between 1.9 and 3.4 years indicating an average turnover rate between 53% and 29%. These numbers are higher when compared to the turnover rates in industrialized countries as for instance, in North America and Europe where the dogs have an average life expectancy of respectively 4.5 years [36] and 5.7 years [37] resulting in an average turnover rate of 20%. Insight in the dog population demography, population size and turnover rates supports the selection of a vaccine in a vaccination scheme in terms of protection time period and frequency of required boosting to keep the required level of immunity.
Male dogs represent a considerable higher proportion of the African owned dog population than female dogs. Male dogs are more aggressive than female dogs and are, therefore, preferred for guarding and hunting. This might have implications in the transmission of rabies to humans and other dogs. For instance, a study in Chad showed that 80% of the human dog bites originated from male dogs [38]. Male dogs are also more likely to be diagnosed positive for rabies than females [39].The risk of acquiring an infection may possibly be influenced by males’ fighting over females during the breeding season [40].
According to the performed meta analysis, the vaccination costs recovery arrangement is one of the factors determining the proportion of dogs vaccinated against rabies in a community. In most African countries, where dog vaccination is not free of charge, the coverage is as low as 9% (Tanzania) [12]. The willingness of dog owners to pay for dog rabies vaccination is generally low as they perceive no direct economic and/or health benefit for themselves. This is supported by the knowledge that the majority of humans bitten were not bitten by their own dogs [29, 34].
The majority of the reviewed studies on charged vaccination schemes did not stated explicitly which costs were charged from the dog owners. In the study of Dürr et al. [41]in Chad, it was indicated that only 24% of the dogs was vaccinated during a parenteral mass dog vaccination campaign in which owners had to pay 21% of the vaccination costs themselves. Dog owners in Chad [42] were willing to pay ≈ 400–700 CFA francs per animal, while the average vaccination costs corresponded to 4000 CFA francs per animal (e.g. 10–17.5% of total costs). These findings also indicate the need for substantial subsidise to vaccinate the required >70% of dogs to interrupt rabies transmission.
Within the African context, the percentage of dogs vaccinated under free of charge vaccination schemes is much higher compared to the owner charged vaccination coverage. However, in developing countries in Asia, like Indonesia, a different situation was observed. Despite the application of a free of charge vaccination campaign [30] vaccination coverage remained as low as 33%. This difference might be explained by the diverse levels of awareness, beliefs and socio-economic factors among the different continents. Similarly, studies in Asia showed that educational level, dog ownership and veterinary service access are important factors affecting the vaccination coverage[29, 43].
The estimation of percentage of dogs vaccinated in owner charged schemes might have been slightly overestimated compared to the vaccination coverage in costs free schemes, because only owned dogs are considered in the case of owner charged vaccination schemes.
Despite the fact that puppies younger than 3 months of age are generally excluded from vaccination campaigns [41], they contribute to a significant proportion of the dog population. This will influence the vaccination coverage and also the risk of human rabies especially the risk for children who have more frequent contacts with puppies than adults. The general exclusion of this age group is often a result of acting upon vaccine manufactures’ guidelines and recommendations. Many rabies vaccines are licensed and approved for primary vaccination of dogs older than 3 months of age. However, it has been shown that most young puppies born from non-vaccinated mothers develop protective antibody titers after a vaccination as early as 4 weeks of age [44, 45].
A high vaccination coverage rate might not necessarily guarantee an effective rabies control. Studies have shown that African dogs experience reduced sero-conversions after rabies vaccination. The study of, for instance, indicated that only 71.9% of the vaccinated dogs in Nigeria developed neutralizing antibodies to rabies virus. The antibody titre depends on the time from vaccination, the technical efficacy of the vaccine used, the nutritional status, sex and age of the dogs [46, 47] also on the quality of the vaccine conservation. Furthermore, vaccination must be repeated more than once to effectively control rabies in a dog population.
A limitation of this study is the exclusion of unpublished reports and non-English articles in the review. Assessment of publications such as country reports of the Southern and Eastern Africa rabies group (SEARG) and the African Rabies Expert Bureau (AfroREB) could have provided additional information. However, as comparable reports are lacking or not accessible due to language barriers it could have created bias if only those country reports were included that were publicly accessible. Therefore, we focused our literature search on objective scientific articles that are internationally accessible.
The language selection criteria also excluded peer reviewed research papers published in French. This is could be a serious limitation of our study due to the fact that French is a very common language in African countries. A comparable systematic search in the French literature resulted in a selection of only 28 French papers (compared to the 1,239 hits resulted from the English search). Based on an evaluation of the English abstracts, only two publications [48, 49] fulfilled the remainder selection criteria as applied in this study. Main findings as presented in the abstracts of these papers were completely in line with the presented results.
This review provided a comprehensive account on the dog rabies parenteral vaccination coverage, dog demography and ownership within the African situation. The main findings of this systematic review indicate that 1) there has been a growing scientific interest in dog rabies control in Africa during the last two decades, which reflects a positive development given the argument that scientific evidence dictates stakeholders in allocation of resources for control and prevention of infectious diseases, 2) only a small proportion of the African dog population is ownerless and, 3) puppies younger than 3 months of age constitute a considerable proportion of the African dog population, 4) male dogs are dominant in the African dog population and 5) the proportion of dogs vaccinated against dog rabies when vaccination is free of charge is closer to WHO recommendations compared to owner charged vaccination schemes. Therefore, as most dogs in Africa are owned and therefore accessible for parenteral vaccination, a high vaccination coverage can be obtained once the necessary financial arrangements are arranged through organized community participation and/or public funding arrangements.
|
10.1371/journal.pgen.1004815 | The Complex I Subunit NDUFA10 Selectively Rescues Drosophila pink1 Mutants through a Mechanism Independent of Mitophagy | Mutations in PINK1, a mitochondrially targeted serine/threonine kinase, cause autosomal recessive Parkinson's disease (PD). Substantial evidence indicates that PINK1 acts with another PD gene, parkin, to regulate mitochondrial morphology and mitophagy. However, loss of PINK1 also causes complex I (CI) deficiency, and has recently been suggested to regulate CI through phosphorylation of NDUFA10/ND42 subunit. To further explore the mechanisms by which PINK1 and Parkin influence mitochondrial integrity, we conducted a screen in Drosophila cells for genes that either phenocopy or suppress mitochondrial hyperfusion caused by pink1 RNAi. Among the genes recovered from this screen was ND42. In Drosophila pink1 mutants, transgenic overexpression of ND42 or its co-chaperone sicily was sufficient to restore CI activity and partially rescue several phenotypes including flight and climbing deficits and mitochondrial disruption in flight muscles. Here, the restoration of CI activity and partial rescue of locomotion does not appear to have a specific requirement for phosphorylation of ND42 at Ser-250. In contrast to pink1 mutants, overexpression of ND42 or sicily failed to rescue any Drosophila parkin mutant phenotypes. We also find that knockdown of the human homologue, NDUFA10, only minimally affecting CCCP-induced mitophagy, and overexpression of NDUFA10 fails to restore Parkin mitochondrial-translocation upon PINK1 loss. These results indicate that the in vivo rescue is due to restoring CI activity rather than promoting mitophagy. Our findings support the emerging view that PINK1 plays a role in regulating CI activity separate from its role with Parkin in mitophagy.
| Two genes linked to heritable forms of the neurodegenerative movement disorder Parkinson's disease (PD), PINK1 and parkin, play important roles in mitochondrial homeostasis through mechanisms which include the degradation of dysfunctional mitochondria, termed mitophagy, and the maintenance of complex I (CI) activity. Here we report the findings of an RNAi based screen in Drosophila cells for genes that may regulate the PINK1-Parkin pathway which identified NDUFA10 (ND42 in Drosophila), a subunit of CI. Using a well-established cellular system and in vivo Drosophila genetics, we demonstrate that while NDUFA10/ND42 only plays a minimal role in mitophagy, restoration of CI activity through overexpression of either ND42 or its co-chaperone sicily is able to substantially rescue behavioral deficits in pink1 mutants but not parkin mutants. Moreover, while parkin overexpression is known to rescue pink1 mutants, it apparently achieves this without restoring CI activity. These results suggest that increasing CI activity or promoting mitophagy can be beneficial in pink1 mutants, and further highlights separable functions of PINK1 and Parkin.
| Parkinson's disease (PD) is the second most prevalent neurodegenerative disorder, the etiology of which remains unknown. Mitochondrial dysfunction, including complex I (CI) deficiency, are frequently observed in pathologic specimens. To elucidate the underlying molecular events, intensive research has focused on identifying the causative gene mutations for inherited forms of PD. An impressive array of disease-causing mutations have been found including autosomal recessive mutations in PARK6 which encodes PTEN-Induced Kinase 1 (PINK1), a mitochondrially targeted serine/threonine kinase, and PARK2, encoding the cytoplasmic E3 ubiquitin ligase Parkin [1], [2]. Previous work has shown that PINK1 and Parkin regulate several cellular processes that impinge on mitochondrial homeostasis [3]–[13], including the fission-fusion dynamics and trafficking of mitochondria [14]–[17], the degradation of damaged mitochondria [18]–[22], inter-organelle communication [23], and activity of CI of the electron transport chain [4], [8], [9], [24], [25]. However, the mechanisms by which PINK1 and Parkin affect these processes is incompletely understood.
In order to identify functional partners of PINK1 we performed a cell based RNAi screen to identify genes that either phenocopy or suppress PINK1 phenotypes. In Drosophila cells lacking pink1 or parkin steady state levels of Marf are increased causing excess fusion of the mitochondrial network [14], [17], [26]. This imbalance in mitochondrial fission-fusion at least partially contributes to the observed organismal phenotypes – enlarged and disrupted mitochondria, muscle degeneration, male sterility, and associated behavioral deficits and neurodegeneration – since they can be partially rescued by genetic interaction with fission (Drp1) or fusion (OPA1 and Marf) genes to restore the balance of normal mitochondrial morphology [14], [16]. We used this phenotype to screen for factors that impact on mitochondrial morphology and which may play a role in PINK1/Parkin function. Our RNAi screen identified multiple hits including a subunit of CI, ND42/NDUFA10. Given the link between CI deficiency and PD, we selected ND42/NDUFA10 for further characterization.
Knockdown of Drosophila ND42 phenocopied pink1 RNAi, causing excessive mitochondrial fusion. Genetic studies in Drosophila reveal that overexpression of ND42 or its co-chaperone sicily is sufficient to rescue behavioral deficits in pink1 mutants through restoration of CI activity. In contrast, overexpression of neither ND42 nor sicily rescued parkin mutant flies, and attenuation of the mammalian homolog, NDUFA10, in HeLa cells only modestly reduced mitophagy. Furthermore, NDUFA10 overexpression cannot restore Parkin translocation to mitochondria in the absence of PINK1, suggesting that ND42 selectively rescues pink1 mutants through a mechanism independent of mitophagy. Our study provides additional evidence in support of a role for PINK1 in CI activity, and further highlights separable functions of PINK1 and Parkin. Future characterization of the other factors from our screen promises to shed additional light on the functional roles of PINK1 and Parkin.
To identify new components in PINK1-Parkin pathway we performed an RNAi screen in Drosophila cells, using a subset library enriched for kinases and phosphatases, for genes that alter mitochondrial morphology (Figure 1A). In particular, we sought to identify genes whose knockdown could either phenocopy or suppress the pink1 RNAi-induced excess fusion. dsRNA treated cells were imaged live using MitoTracker Red to fluorescently label the mitochondrial network. The mitochondrial morphology in each image was scored as primarily falling into one of four categories based on control knockdown phenotypes (Figure 1B). Control (DsRed) dsRNA treated cells showed a typical mix of short-round and long-tubular mitochondria. This category was scored 2. Cells treated with dsRNA against the fly mitofusin homologue, Marf, caused a complete fragmentation of the mitochondrial network as expected and were scored 1. pink1 RNAi resulted in a tubular, highly interconnected network, as previously reported [14], [17], scoring 3. RNAi against the pro-fission factor Drp1 caused mitochondrial hyperfusion and peri-nuclear clumping. Being qualitatively different from the pink1 RNAi phenotype and an extreme result of hyperfusion this category scored 4.
The screen was performed on a library of selected genes mostly comprising kinases and phosphatases but with additional genes of interest. The effect on mitochondrial morphology was assessed in two backgrounds; one in a wild type background to identify manipulations that phenocopy pink1 RNAi tubulation, and another in a pink1 RNAi background to identify manipulations capable of suppressing the pink1 phenotype. The results from the two screens were cross-referenced to further refine the selection of possible hits (Figure 1C). The limits of the regions considered to phenocopy or suppress were defined by the mean ± standard deviation for pink1 or DsRed dsRNAs respectively. The gene targets whose dsRNAs either phenocopy or suppress pink1 RNAi are summarized in Table 1 and Table 2.
While there are several interesting candidates for further investigation in both suppressors and phenocopiers (see Discussion), a specific motivation for this work was to identify additional components of the PINK1-Parkin pathway that regulate mitochondrial dynamics and mitophagy. An attractive candidate that could fulfill this role would be a mitochondrially localized protein. Notably, we identified ND42, which encodes a subunit of CI of the electron transport chain, as a phenocopier; knockdown of ND42 caused excess tubulation similar to pink1 knockdown. Importantly, CI activity has been shown to be specifically affected in various models of pink1 deficiency, supporting ND42 as an attractive target for analysis in pink1 function.
CI is a very large, multi-subunit complex comprising of around 44 subunits, consisting of a hydrophobic portion embedded in the inner membrane and a hydrophilic portion extending into the matrix [27], [28]. In order to assess the specific versus general effect of attenuating CI subunits on mitochondrial morphology, several additional subunits were analyzed. dsRNAs targeting 6 other subunits from different subdomains (α, β or λ) were generated and the effect of knockdown on mitochondrial morphology was assessed.
We confirmed that ND42 knockdown caused excess mitochondrial fusion in a WT background, indistinguishable from pink1 knockdown, but did not further enhance the pink1 phenotype (Figures 2A, 2B, and S1). In contrast none of the other selected subunits induced fusion; 4 subunits had no effect on morphology while 2 subunits caused fragmentation (Figures 2C and 2D). Further analysis of one of these subunits, CG7712, which did not perturb mitochondrial morphology, also did not modify the pink1 phenotype. These results support the specificity of the effect observed with ND42 knockdown.
Since ND42 RNAi phenocopies loss of pink1 in cells we assessed whether loss of ND42 may phenocopy pink1 mutants in vivo. Drosophila mutant for pink1 exhibit characteristic locomotor deficits in climbing and flight, associated with degeneration of the musculature and profound disruption of mitochondria [3], [11]. In agreement with previous observations [29], we found that knockdown of ND42 in all tissues is lethal, consistent with the subunit playing a critical role in this essential metabolic enzyme.
Due to the essential nature of this gene, and lack of pink1 phenocopy, we did not further characterize ND42 loss-of-function. We hypothesized that, since pink1 mutants show CI deficiency, overexpressing ND42 may suppress pink1 phenotypes. While ND42 overexpression of two independent transgenes driven by the strong ubiquitous driver daughterless (da)-GAL4 had no effect on motor performance in a wild type background (Figures 3A and 3B), we found that expression of either transgene was able to significantly restore climbing and flight ability in pink1 mutants (Figures 3C and 3D). However, ND42 overexpression only partially restored flight muscle and mitochondrial integrity (Figure 3E), but was not able to improve the male sterility (Figure S2A).
Genetic interaction studies in Drosophila have linked pink1 and parkin in a common pathway with pink1 acting upstream of parkin [3], [6], [11]. To further characterize the putative action of ND42/NDUFA10 in the PINK1-Parkin pathway, we tested whether ND42 overexpression could also rescue parkin mutants. Surprisingly, we found that overexpression of ND42 was not able to rescue any parkin phenotypes tested, including locomotor behaviors, muscle and mitochondrial integrity, and male sterility (Figures 4A–4C, and S2B).
Recently it was reported that Drosophila sicily acts as a co-chaperone to bind and stabilize ND42 in the cytoplasm, promoting its mitochondrial import and the formation of CI [29]. Supporting a potential role for sicily in pink1 function, we found that knockdown of sicily in Drosophila cells phenocopied pink1 mitochondrial hyperfusion (Figure S3). Since sicily promotes ND42 stability, and overexpression of ND42 can rescue pink1 mutant phenotypes, we hypothesized that sicily overexpression may also rescue pink1 mutants. Indeed, we found that sicily overexpression rescued pink1 mutant locomotor and mitochondrial phenotypes comparable to ND42 overexpression (Figures 3A–3E). Also, overexpression of sicily failed to rescue similar parkin mutant phenotypes (Figures 4A–4C), mirroring the effects of ND42. Together these results demonstrate a genetic interaction of ND42 and sicily with pink1 but not parkin.
The fact that ND42 overexpression can rescue pink1 but not parkin mutants would be consistent with it acting in a common pathway between PINK1 and Parkin. An intensively studied field of PINK1-Parkin function is the autophagic turnover of mitochondria, termed mitophagy [30]. In HeLa cells overexpressing YFP-Parkin, depolarization of the mitochondrial membrane with the protonophore carbonyl cyanide 3-chlorophenylhydrazone (CCCP) causes a rapid stabilization of PINK1 on the outer mitochondrial membrane, which stimulates the re-distribution of cytoplasmic Parkin to co-localize with mitochondria (Figures 5A and 5B). Prolonged exposure to CCCP induces substantial degradation of the mitochondria (Figures 5C, 5D, and S4). These phenomena are almost completely abolished by PINK1 knockdown (Figures 5A–5D) [18]–[20], [22].
We next analyzed the effect of the human homolog of ND42, NDUFA10, on Parkin translocation and mitophagy. We found that NDUFA10 knockdown had a modest but significant effect on Parkin translocation, though clearly not as much as loss of PINK1 (Figures 5A, 5B, S1 and S5). Moreover, loss of NDUFA10 only very minimally reduced the degree of mitophagy (Figures 5C, 5D, and S6). We also found no effect of NDUFA10 knockdown on PINK1 stabilization following mitochondrial depolarization (Figure S7). These data indicate that NDUFA10 does not play a significant role in PINK1/Parkin mediated mitophagy and suggests that the in vivo rescue was unlikely via the mitophagy pathway.
To further exclude a role for mitophagy in the rescue of pink1 mutants, we assessed in vitro whether NDUFA10 overexpression could promote CCCP-induced Parkin translocation in the absence of PINK1. Encouragingly, re-expression of either NDUFA10 or ND42 almost completely restored Parkin translocation reduced by NDUFA10 knockdown (Figures 6A and 6C). However, when Parkin translocation was completely blocked by loss of PINK1, this was not rescued by expression of either NDUFA10 or ND42 (Figures 6A and 6D). These results support the idea that the rescue seen in vivo is unlikely due to activated mitophagy, raising the question of what mechanism is responsible.
As loss of PINK1 has been reported to cause decreased CI activity [24], we reasoned that suppression of pink1 mutants by ND42 overexpression may be due to restoration of CI activity. As previously reported, we observed a CI deficiency in pink1 mutant flies, leading to decreased ATP production (Figures, 7A and 7B). We found that ND42 overexpression was indeed able to completely restore both CI activity and ATP levels in vivo (Figure 7A and 7B). Extending these analyses to parkin mutants, we saw a non-significant decrease in CI activity in parkin mutants that remained unchanged by ND42 overexpression (Figure 7C). Similarly, the significant depletion of ATP evident in parkin mutants was not rescued by ND42 overexpression (Figure 7D), consistent with a lack of phenotypic rescue by ND42 overexpression. Interestingly, we also found that sicily overexpression was able to completely restore CI activity in pink1 mutants (Figure 7A), while the increase in ATP levels was not significant (Figure 7B). Similar to ND42, sicily overexpression had no effect in parkin mutants (Figures 7C and 7D).
While this work was in preparation, Morais et al [25] reported that NDUFA10 lacked phosphorylation at serine-250 in the absence of PINK1, and that expression of phospho-mimetic NDUFA10/ND42 specifically reversed PINK1 deficits in various model systems, including restoring CI activity in mammalian systems and synaptic phenotypes in Drosophila pink1 mutants. Our preceding data concur that overexpression of ND42 can rescue some pink1 mutant phenotypes (and not others), but interestingly we found that this can be achieved with expression of the wild type version without a specific requirement for the phospho-mimetic.
To further explore the potential role of Ser-250 phosphorylation in these assays, we tested in parallel our existing lines (previously reported in [31]) and those of Morais et al [25] and of the yeast equivalent of CI, NDI1 [31]. As before, we found that expression of the previous WT transgene (designated ND42HB) significantly rescued pink1 climbing and flight defects (Figures 8A and 8B). We also found that the Morais et al. transgenes (designated ND42BDS) expressing either WT or phospho-mimetic (SD) also partially rescued climbing, albeit to a lesser extent (Figure 8A). The WT version of these lines did not statistically improve flight ability whereas the SD did provide a modest rescue (Figure 8B). We also found that the non-phosphorylatable version (SA) provided significant rescue of climbing, equivalent to the phospho-mimetic, but again did not rescue flight (Figure 8A and 8B). Notably, in these assays NDI1 expression significantly rescued climbing but not flight ability (Figure 8A and 8B), consistent with previous observations [31].
To better understand the relationship between the behavioral rescue and CI activity, and to assess functional efficacies of the various transgenic lines, we tested the ability of the phospho-variants transgenes to rescue the CI deficiency in pink1 mutants. We found that expression of all phospho-variants were able to fully restore CI activity in pink1 mutants (Fig. 8C). The degree of rescue was similar to that seen with the previous WT transgene (Fig. 7A), consistent with an equivalent level of expression between these lines (Fig. S1G). Interestingly, while we see the highest level of CI activity with the phospho-mimetic (SD), we also see a substantial rescue by the phopho-null (SA) in this in vitro assay.
The ability of multiple transgenes that restore CI activity to at least partially rescue climbing behavior supports the idea that promoting CI activity is differentially beneficial in pink1 mutants but not parkin mutants and may hint at different underlying causes of pathogenicity. However, a puzzling aspect of this is the long-standing observations that parkin overexpression is sufficient to almost completely rescue many pink1 mutant phenotypes (Figures 9A and 9B, and [3], [11], [32], [33]). Since, to our knowledge, it had never been reported, we tested whether the rescue may be due to restoration of CI activity in pink1 mutants. Surprisingly, we found that parkin overexpression mildly improved ATP levels but did not restore CI function (Figures 9C and 9D). Hence, these genetic interactions further support independent and separable functions of PINK1 and Parkin.
PINK1 and Parkin have long been genetically linked in a common pathway that promotes mitochondrial homeostasis at least partly by directing the autophagic degradation of dysfunctional mitochondria as a mechanism of mitochondrial quality control. While this model potentially explains the occurrence of CI deficiency, oxidative stress, calcium dysregulation and elevated mtDNA mutations seen in patient tissues, and the age-related onset of PD [34], other models have been proposed to explain the pathological consequences of PINK1 and Parkin deficiency. Moreover, many mechanistic details by which the PINK1-Parkin pathway functions remain unexplained. To address these matters, we conducted an RNAi screen to identify genes whose loss-of-function either phenocopied or suppressed a pink1 RNAi phenotype. We have identified a number of genes that fulfill these criteria (discussed below) but focused our current investigation on ND42/NDUFA10 given the extensive literature implicating CI deficiency in PD pathogenesis and the fact that CI deficiency has previously been reported in PINK1 mutant models and patient samples.
Loss of ND42/NDUFA10 phenocopies the effect of pink1 loss on mitochondrial morphology in Drosophila cells, and ND42 overexpression rescues the pink1 mutant phenotypes. However, NDUFA10 knockdown has only modest effects on mitophagy, supporting a separate link between CI and PINK1 function. The simplest interpretation of these findings is that PINK1 normally regulates ND42/NDUFA10 abundance or activity through direct phosphorylation. Indeed, it was recently reported that NDUFA10 lacks phosphorylation at Ser-250 in Pink1-/- cells [25], although it remains to be determined whether PINK1 directly or indirectly regulates NDUFA10 phosphorylation. Moreover, it was reported that expression of a phospho-mimetic version of ND42/NDUFA10 specifically rescued phenotypes in multiple PINK1 deficient systems, while an S250A mutant version of ND42/NDUFA10 that is incapable of being phosphorylated was unable to confer rescue. Consistent with this we find that, from equivalent expression levels, the phospho-mimetic (SD) provides a slightly better phenotypic rescue than the other variants, and likewise promotes a higher CI activity. Nevertheless, our results also show that the non-phosphorylatable S250A version is still able to restore CI activity and significantly rescue the climbing deficit in pink1 mutant flies.
While further studies are needed to clarify the functional relationship between PINK1 and NDUFA10 in the regulation of CI, our findings provide further support mounting evidence that many manipulations that promote CI activity – overexpression of NDUFA10, sicily, heix, Ret, dNK, TRAP1 and NDI1, or treatment with vitamin K, deoxynucleosides or folic acid [25], [31], [35]–[38] – can rescue pink1 mutants, suggesting a more general defect underlies CI deficiency in loss of pink1. We hypothesize that the loss of CI activity in pink1 mutants may be due to a general de-stabilization of CI. Assembly is a particular challenge for such a large, multi-subunit complex and occurs in a stepwise process that is highly regulated by many factors [39]. Even its association with other ETC complexes in supercomplexes affects CI's stability [40]. There is evidence for reduced complex stability in pink1 mutants, though this may not be specific to CI [37], [41], [42]. One possibility is that PINK1 influences CI stability by directly promoting the assembly of CI, which may be regulated by NDUFA10.
The current findings also further support that the mechanism by which PINK1 influences CI activity appears to be separable from its well-characterized role in mitophagy, since, in agreement with some studies [24], [31] but in contrast to others [4], [8], [9], we do not find clear evidence of CI deficiency in parkin mutants flies. Moreover, it was unexpected to find that overexpression of parkin does not rescue the CI deficiency in pink1 mutants, because substantial previous work has shown that parkin overexpression rescues all of the other pink1 phenotypes, and because a prediction of the PINK1-parkin mitophagy pathway is that activation would trigger the selective removal of mitochondria deficient in CI activity. This finding suggests that CI deficiency alone cannot fully account for adult locomotor phenotypes seen in pink1 mutants. Further studies are needed to clarify full spectrum of cellular defects in pink1 and parkin mutants and their relative importance to the pathologic mechanism.
The present screen analyzed the effect of ∼600 genes comprising mostly genes with homology to kinase or phosphatase domains. Other hits from this screen, identifying both phenocopiers and suppressors, could also be attractive candidates as potential new factors of pink1/Parkin function. Notably several hits play a role in lipid biology. This is particularly noteworthy in light of our recent report that another RNAi screen identified the master regulator of lipogenesis, SREBF1, to affect pink1/Parkin-mediated mitophagy [43].
The phenocopier Sphingosine kinase 1 and the suppressor easily shocked (encoding Drosophila Ethanolamine Kinase) are involved in phospholipid metabolic pathways. Sphingosine kinase catalyzes the production of sphingosine-1-phosphate (S1P), a key signaling molecule affecting cell growth and survival [44]. While S1P affects many cellular processes perhaps the most intriguing is its role in calcium mobilization from the endoplasmic reticulum (ER) [45] since Parkin was recently shown to promote ER-mitochondrial calcium transfer [23]. Interestingly, the breakdown of S1P generates phosphoethanolamine, the enzymatic product of ethanolamine kinase and a precursor metabolite of the key phospholipid phosphatidylethanolamine (PE). Loss of mitochondrial PE has been shown to affect mitochondrial morphology, oxidative phosphorylation and even the formation of complex I-containing supercomplexes [46]. The identification of four wheel drive (encoding Drosophila phosphatidylinositol 4-kinase beta), which catalyzes the formation of PI(4)P, is also intriguing since mutations in SYNJ1, which encodes PI(4,5)P2 phosphatase, were identified in families with early onset parkinsonism [47], [48].
Also related to lipid biology is Nocturnin although a direct link to mitochondria biology is less obvious. Nocturnin encodes a circadian deadenylase thought to be involved in the rhythmic regulation of gene expression by removal of polyA tails from mRNAs. Mice lacking Nocturnin are resistant to diet-induced obesity and hepatic steatosis [49], linking its function to lipid metabolism. Further studies will be needed to determine the extent to which lipids in general or specific lipids, and their regulated synthesis, impact on pink1/Parkin biology and regulation of mitochondrial dynamics and quality control. Nevertheless, this screen provides a resource for characterizing novel factors that regulate mitochondrial morphology.
Drosophila S2R+ cells were cultured in Schneider's medium (Gibco) containing 10% (vol/vol) heat-inactivated fetal bovine serum (Sigma), Penicillin 10 units/ml (Sigma) and Streptomycin 10 µg/ml (Sigma). Cells were maintained in a 25°C incubator. HeLa cells were cultured in DMEM GLUTAMAX media (Gibco) containing 10% (vol/vol) heat-inactivated fetal bovine serum (Sigma), Penicillin 10 units/ml (Sigma) and Streptomycin 10 µg/ml (Sigma). Cells were maintained in a 37°C incubator with 5% CO2. A stable transfected HeLa cell line expressing YFP-Parkin in pLVX-puro was cultured in DMEM GLUTAMAX media (Gibco) containing 10% (vol/vol) heat-inactivated fetal bovine serum (Sigma), Penicillin 10 units/ml (Sigma) and Streptomycin 10 µg/ml (Sigma). Cells were maintained in a 37°C incubator with 5% CO2.
The kinome/phosphatome sub-library was generated based upon the second generation Drosophila dsRNA library (Heidelberg 2). Detailed information on amplicon targets is available online (http://rnai-screening-wiki.dkfz.de/signaling/wiki/display/rnaiwiki/Drosophila+RNAi+libraries). This sub-library was designed to contain all known and computed kinases and phosphatases, genes with some homology to these enzyme classes, but also some other genes of general interest. Library dsRNAs were plated at 250 ng in 5 µl of H2O into Perkin Elmer 384 well view plate (Product number: 6007470). Screen plates were arrayed with the inclusion of gaps to allow for the addition of user-specific controls. Here we added dsRNAs targeting Marf, Drp1, OPA1, Fis1, pink1 and parkin. These controls consistently all showed the expected results. For the pink1 RNAi background, 250 ng dsRNA against pink1 was added to each well prior to screening. A ‘double dose’ of dsRNA did not appear to affect mitochondrial morphology (see below). 15,000 Drosophila S2R+ cells were added to each well in 30 µl of Schneider's medium (Gibco) without FBS (Sigma). Plates were incubated at 25°C for 1 hour in which time dsRNAs are taken up by the cells. Following this incubation 30 µl of Schneider's medium (Gibco) containing 20% FBS (v/v) was added. The plates were then sealed and incubated at 25°C incubator for 4 days. Cells were stained with 100 nM Mitotracker Red (Invitrogen, M7512) and 20 µg/ml Hoechst 33342 (Invitrogen, H3570) for 15 minutes. Media was replaced and imaging was performed on live cells. Imaging was performed on an IN Cell Analyzer 1000 (GE Healthcare) automated microscope using a 40× air objective (Nikon, 0.60 NA).
Cells were prepared identically as for the high-throughput screening conditions, except for wild type where background 250 ng dsRNA targeting DsRed was added to each condition to mirror the ‘double dose’ of dsRNA in the pink1 background. Cells were imaged live under ambient conditions on a Deltavision RT deconvolution wide field microscope (Olympus, 100× objective, 1.4 NA) using 8 well µ-Slides (Ibidi), with 10 images taken per condition. Cells were scored for their gross mitochondrial morphology by eye with the scorer blinded to the experimental conditions. Where rotenone (20 µM) was used, cells were treated for 2 hours before mitochondrial morphology was analysed. A score would be assigned for a whole field of view and an average score would be calculated for the 10 images per condition. A score of 1 would be given to a field of view that had mainly fragmented mitochondria. A score of 2 would be given to a field of view that had a mainly wild type mitochondrial network with a mix of short-round and long-tubular mitochondria. A score of 3 would be given to a field of view that had mainly tubular mitochondria. A score of 4 would be given to a field of view that had mainly clumped mitochondria where the mitochondria had formed a single or few large peri-nuclear clusters.
Drosophila gene dsRNAs were generated using the MEGAscript T7 Kit (Ambion), using T7-flanked DNA amplicons from the library as template. Control dsRNA for Drosophila cell qRT-PCR analysis was a 782 bp sequence targeted against C. elegans gene R06F6.2 which has no ∼21mer homology within the Drosophila genome. siRNAs targeting human genes were obtained from the siGENOME SMARTpool collection (Dharmacon) as follows: control siRNA is Non-Targeting siRNA Pool #1 (product code; D-001206-13-20); PINK1 (product code; M-004030-02); NDUFA10 (product code; M-006752-00).
Total RNA from live cells was prepared from three replicates of each dsRNA treatment using RNeasy (Qiagen). RNA concentration was then determined spectrophotometrically. Once treated with DNase, total RNA was reverse-transcribed using RETROscript (Ambion) or Protoscript (NEB) according to the manufacturer's protocol. Quantitative PCR was performed using SYBR Green (Sigma) on a MyIQ real time PCR detection system (Bio-Rad). Each PCR included three technical replicates, which were repeated as three biological replicates. Levels for each transcript were normalized to a 18S rRNA (Drosophila: 18SrRNA; Human: RNA18S5) control by the 2-ΔΔCT method.
For Drosophila genes, primers used were:
18S - Forward: TCTAGCAATATGAGATTGAGCAATAAG
18S - Reverse: AATACACGTTGATACTTTCATTGTAGC
pink1 - Forward: GACGACCCTCGCACATAA
pink1 - Reverse: AACAGTCCGGAGATCCTACAG
ND42 - Forward: CGTTTCGATGTCCCGGAGCT
ND42 - Reverse: GTCTGCATTGTAGCCAGGAC
CG7712 - Forward: CGCAATGTGACCGACATCCG
CG7712 - Reverse: CGCATGATATGGCCTTCTG
For human genes, primers used were:
18S - Forward: CAGCCACCCGAGATTGAGCA
18S - Reverse: TAGTAGCGACGGGCGGTGTG
PINK1 - Forward: GCCGGACGCTGTTCCTCGTT
PINK1 - Reverse: TGGACACCTCTGGGGCCATC
NDUFA10 - Forward: GATCCGAGAAGCAATGATG
NDUFA10 - Reverse: TGGAGCGCTCCAACACAACA
The following primary antibodies were used, mouse anti-ATP5A (MS507, MitoSciences; 1∶2000), rabbit anti-GFP (ab6556, Abcam; 1∶5000). Secondary antibodies used were rabbit polyclonal anti-mouse IgG-H&L (DyLight 594, Invitrogen; 1∶5000) and goat anti-rabbit IgG (Alexa Fluor 488, Invitrogen; 1∶5000).
YFP-Parkin HeLa cells were reverse-transfected with siRNAs using DharmaFECT 1 (Dharmacon). For Parkin translocation, cells were incubated for 4 days then treated with 10 µM CCCP or equivalent volume of the solvent (EtOH) for 4 hours. For mitophagy, cells were incubated for 3 days then treated with 10 µM CCCP or equivalent volume of the solvent (EtOH) for 24 hours. Cells were fixed in ice-cold methanol for 10 minutes and washed in PBS. Mitochondrial staining was achieved by using anti-ATP5A antibody. Imaging was performed on an Olympus FV1000 confocal microscope (Olympus, 60× oil objective, 1.25 NA). For Parkin translocation, cells were scored for the accumulation of YFP-Parkin on mitochondria. For mitophagy, cells were scored for mitochondrial load based on having a normal load of mitochondria, few mitochondria or no mitochondria. At least 20 cells were scored per treatment and 3 biological replicates were performed.
HeLa cells were reverse-transfected with siRNAs using DharmaFECT 1. After 3 days cells were transfected with PINK1-EGFP using Effectene (Qiagen). After a further 1 day cells were treated with 10 µM CCCP or equivalent volume of the solvent (EtOH) for 1 hour before fixation with ice-cold methanol for 10 minutes and washed in PBS. Immunofluorescence was performed using anti-ATP5A and anti-GFP antibodies, and appropriate fluorescent secondary antibodies. Imaging was performed on an Olympus FV1000 confocal microscope. Cells were scored for the accumulation of PINK1-EGFP on mitochondria. At least 20 cells were scored per treatment and 3 biological replicates were performed.
TMRM assay to measure mitochondrial membrane polarity was done as previously described [8]. Briefly, HeLa cells were reverse-transfected with siRNAs using DharmaFECT for 4 days. Cells were then treated with 10 µM CCCP or equivalent volume of the solvent (EtOH) for 1 hour. Cells were incubated with 50 nM TMRM (VTX668, Fisher) in PBS with 10 µM CCCP or equivalent volume of solvent for 30 minutes, then washed in PBS with 10 µM CCCP or equivalent volume of solvent 5 times. TMRM fluorescence was read on a spectrophotometer at 550 nm (Berthold technologies Mithras LB940). Triplicate readings were taken from 3 biological replicates.
Drosophila were raised under standard conditions at 25°C on food consisting of agar, cornmeal and yeast. pink1B9 mutants [11] were provided by J. Chung (KAIST). park25 mutants, fertility tests, flight and climbing assays were performed as previously described [6], [50]. w1118 and da-GAL4 strains were obtained from the Bloomington Drosophila Stock Center (Bloomington, IN). UAS-ND42-RNAi lines (GD: 14444; KK: 101787) were obtained form the Vienna Drosophila Resource Centre [51]. UAS-ND42, UAS-ND42-HA and UAS-sicily have been described previously [29] and were a gift from H. Bellen (Baylor College of Medicine). The additional UAS-ND42 lines (WT, SA and SD) from Morais et al. [25] were provided by Patrik Verstreken.
Mitochondria-enriched fractions were prepared from whole adult male flies (∼3 days old) with the indicated genotype (10 flies were used for each sample). Flies were gently crushed in chilled isolation buffer (250 mM sucrose, 10 mM Tris-HCl, pH 7.4, 0.15 mM MgCl2) using a plastic pestle homogenizer, then centrifuged twice at 500×g for 5 minutes at 4°C to remove debris. The supernatant was centrifuged 5000×g for 5 minutes at 4°C. The resulting pellet containing mitochondria was re-suspended in the isolation buffer or assay buffers. Complex I activity was measured using a modified method from Birch-Machin et al [52]. Briefly, samples were subjected to 3 cycles of freeze-thaw in liquid nitrogen. Complex I activity was determined by following the oxidation of NADH at 340 nm with a reference wavelength of 425 nm (ε = 6.22 mM−1 cm−1) at 30°C using a BMG Labtech FLUOStar plate reader. The assay buffer contained 25 mM KH2PO4, 5 mM MgCl2, (pH 7.2), 3 mM KCN, 2.5 mg per ml BSA, 50 µM ubiquinone, 2 µg/ml antimycin A and mitochondrial extract. The baseline was recorded for 5 minutes and the reaction was started with 125 µM NADH measured for 30 minutes, 15 µg/ml rotenone was added to inhibit the reaction and measured for 15 minutes. The results are expressed as µmol NADH oxidised/min/citrate synthase activity. Citrate synthase was measured by following the production of 5-thio-2-nitrobenzoate at 30°C using a BMG Labtech FLUOStar plate reader after samples had undergone 3 cycles of freeze-thaw in liquid nitrogen. The assay buffer was 100 mM Tris HCl (pH 8.0), 0.1 mM DTNB, 50 µM acetyl Coenzyme A, 0.1% Triton X-100 and mitochondrial extract per well. The baseline was recorded for 5 minutes at 412 nm, then the reaction was started by the addition of 0.5 mM oxaloacetate acid and the rate was recorded for 15 minutes.
Five male flies (3 days old) were homogenized in 100 µl of 6 M guanidine-HCl in extraction buffer (100 mM Tris, 4 mM EDTA, pH 7.8) to inhibit ATPases. Homogenized samples were subjected to rapid freezing in liquid nitrogen, followed by boiling for 3 min. Samples were cleared by centrifugation, and supernatant was diluted (1/100) with extraction buffer and mixed with a luminescent solution (CellTiter-Glo Luminescent Cell Viability Assay, Promega, Fitchburg, WI, USA). Luminescence was measured on a Varioskan™ Flash Multimode Reader (Thermo Scientific, Waltham, MA, USA). The relative ATP levels were calculated by dividing the luminescence by the total protein concentration, which was determined by the Bradford method.
For Parkin translocation, mitophagy and PINK1 stabilization, statistical significance was calculated by using Student's t-test on triplicate experiments comparing against control siRNA/dsRNAs of the equivalent experimental condition. Biochemical and behavioral assays in Drosophila were analyzed by one-way ANOVA with Bonferroni correction. Male fertility was analyzed by Chi-square test.
|
10.1371/journal.ppat.1001183 | The Gag Cleavage Product, p12, is a Functional Constituent of the Murine Leukemia Virus Pre-Integration Complex | The p12 protein is a cleavage product of the Gag precursor of the murine leukemia virus (MLV). Specific mutations in p12 have been described that affect early stages of infection, rendering the virus replication-defective. Such mutants showed normal generation of genomic DNA but no formation of circular forms, which are markers of nuclear entry by the viral DNA. This suggested that p12 may function in early stages of infection but the precise mechanism of p12 action is not known. To address the function and follow the intracellular localization of the wt p12 protein, we generated tagged p12 proteins in the context of a replication-competent virus, which allowed for the detection of p12 at early stages of infection by immunofluorescence. p12 was found to be distributed to discrete puncta, indicative of macromolecular complexes. These complexes were localized to the cytoplasm early after infection, and thereafter accumulated adjacent to mitotic chromosomes. This chromosomal accumulation was impaired for p12 proteins with a mutation that rendered the virus integration-defective. Immunofluorescence demonstrated that intracellular p12 complexes co-localized with capsid, a known constituent of the MLV pre-integration complex (PIC), and immunofluorescence combined with fluorescent in situ hybridization (FISH) revealed co-localization of the p12 proteins with the incoming reverse transcribed viral DNA. Interactions of p12 with the capsid and with the viral DNA were also demonstrated by co-immunoprecipitation. These results imply that p12 proteins are components of the MLV PIC. Furthermore, a large excess of wt PICs did not rescue the defect in integration of PICs derived from mutant p12 particles, demonstrating that p12 exerts its function as part of this complex. Altogether, these results imply that p12 proteins are constituent of the MLV PIC and function in directing the PIC from the cytoplasm towards integration.
| All retroviruses reverse transcribe their RNA genome to a DNA copy in the cytoplasm of the infected cell. To be expressed, the viral genomic DNA has to travel to the cell nucleus and to integrate into the cellular chromosomes. This trafficking is governed by cellular and viral proteins that associate with the viral genome to form a ‘pre-integration complex’ (PIC), yet the full composition of this complex is unknown. Former studies showed that for the murine leukemia virus (MLV), mutations in a viral protein named p12 abrogate MLV infection, after reverse transcription and prior to the integration step, suggesting a role for this protein in early stages of infection. However, the precise mechanism of p12 action is not known. We combined microscopic, genetic and biochemical techniques to provide evidence that the p12 protein is part of the MLV PIC and that it exerts its function from within this complex. These analyses also suggest a role for p12 in the trafficking of the PIC from the cytoplasm to the chromosomes of the infected cell. Altogether, these findings highlight an important ‘building block’ of a complex that is essential for MLV infection.
| Reverse transcription and integration are the hallmarks of the retroviral life cycle. These steps include reverse transcription of the genomic RNA into a linear double-stranded DNA and the subsequent integration of this DNA into the genome of the infected cell. These events are part of the ‘early’ stages of the retroviral life cycle, starting with the binding of the virus to its cellular receptor and ending once the integration step has occurred. Reverse transcription and integration are mediated by the viral enzymes; reverse transcriptase (RT) and integrase (IN), respectively; both are cleavage products of the polyprotein encoded by the viral pol gene. Reverse transcription occurs in a cytoplasmic complex, termed reverse transcription complex (RTC), which transforms to the PIC (reviewed in [1], [2]). The PIC harbors the viral DNA and travels from the cytoplasm to the nucleus, to target the chromatin of the infected cell for integration.
The full composition of the RTC and PIC is not known; this is true not only for the cellular components, but also for the viral constituents of these complexes [2], [3]. Some of the known cellular components identified in RTC/PIC of different retroviruses include: the barrier of auto-integration factor (BAF) [4], [5], high-mobility group proteins (HMGs) [6], [7], Ku [8], lamina-associated polypeptide 2α (LAP2α) [9], and lens epithelium-derived growth factor (LEDGF/p75) [10], [11]. To date, the viral protein components identified in the RTC/PIC of the simple MLV include: RT [12], nucleocapsid (NC) [13], capsid (CA) [12] and IN [7], [12], [13], [14]; while in the complex human immunodeficiency virus type-1 (HIV-1), NC [15], matrix (MA) [16], [17], [18], [19], RT [15], [16], [17], [18], [19], IN [6], [15], [16], [19], [20] and Vpr [15], [19] are present.
The trafficking of the PIC towards the chromatin of infected cells might be substantially variable between different retroviruses. This is demonstrated in part, by the fact that the nuclear entry by the MLV PIC is strictly dependent on mitosis [21], whereas nuclear entry by HIV-1 PIC can occur efficiently in nondividing cells [22], [23], [24], [25], [26]. Differences in PIC composition may contribute to such a variance; it has been shown that the MLV PIC contains the viral CA protein [14], while the HIV PIC does not [18]; this may contribute to the inability of wt MLV PICs to enter the nucleus through the nuclear pores [27], and render the MLV PICs dependent on nuclear envelope breakdown during mitosis for nuclear entry [21].
One viral protein thought to influence the trafficking of the MLV PIC, is the p12 protein. This protein is a cleavage product of the Gag precursor, which is the major structural protein in the virion. A few thousand Gag molecules assemble to form one viral particle, and these are cleaved by the viral protease during the virion maturation step. For MLV, this cleavage results in the release of the following viral proteins: MA, p12, CA and NC. The p12 domain is known to act in the late budding process of the Gag precursor, and accordingly mutations in the p12 hamper virion release [28], [29], [30], [31]. However, additional mutations in this protein have specifically affected early stages of MLV infections, revealing a critical role for p12 in these stages [3], [28], [31], [32], [33]. Analysis of a subset of these mutants revealed normal generation of the linear genomic DNA but no generation of circular viral DNA forms. Such circular forms are thought to be formed by nuclear enzymes from a portion of the linear viral DNA and although these circles are not substrates for the IN [34], [35], [36], they serve as a marker for nuclear entry by the viral DNA. Their absence during the infection of the p12 mutants suggested that p12 may function in directing the PIC into the nucleus or alternatively affects unknown nuclear steps needed for integration. Importantly, for at least one of these p12 mutants, the PIC appeared competent for integration in an in vitro assay [3]. A later study provided genetic evidence that p12 may function in concert with CA in early stages of infection: using swap mutants between the different domains of the Gag of MLV and its closely related virus, the spleen necrosis virus (SNV), it was demonstrated that productive infection can be achieved only when the p12 (or p18 from SNV) and CA domains are from the same virus [37]. Although these studies strongly indicate p12 participation in the early events of MLV infection, the precise mechanisms of p12 action is not known. Here we provide evidence that p12 is a part of the PIC and functions from within this complex in early stages of infection.
To study p12 function during the early stages of MLV life cycle, we constructed epitope-tagged p12 to enable its visualization in infected cells. Former mutagenesis studies have demonstrated that MLV tolerates mutations in the central region of p12 rather than in other parts of this protein [28], [31], [33], [38] (Fig. 1). In addition, our phylogenetic analysis revealed that the central region of p12 is relatively less conserved between members of the MLV family, further suggesting that this region may tolerate changes in its sequence (Fig. 1). We then inserted into this region a triple repeat of the Myc epitope (3xMyc; Fig. 1) to enhance detection of tagged p12. The p12 protein of this virus could easily be detected by Western blot analysis (Fig. 2A), but the replication of this mutant was greatly reduced in comparison to wt (Fig. 2B). However, we were able to recover a revertant virus with normal replication kinetics, after repeated passages of the 3xMyc virus on naïve NIH3T3 cells. A BsrGI-XhoI fragment of the revertant, containing the entire p12 domain and flanking Gag sequences, was PCR amplified and sequenced, revealing the existence of only one in-frame copy of the Myc epitope in p12 with no other mutations. This fragment was used to replace the wt sequence between the BsrGI-XhoI restriction sites in a molecular clone of MLV (pNCS, Materials and Methods), and the resulting virus was named 1xMycR (Fig. 1). Further analysis of purified virions of the 1xMycR virus demonstrated that its p12 could be detected with anti-Myc antibodies by Western blot analysis and, as expected, the protein had a reduced molecular weight compared to p12 of the 3xMyc virus (Fig. 2A). Importantly, unlike the 3xMyc virus, the 1xMycR virus replicated with wt-like kinetics in NIH3T3 cells (Fig. 2B), demonstrating that the reduction in the number of repeats of the epitope tag in p12, accounted for the improved replication of the latter virus.
Next, we examined whether p12 proteins can be detected by immunofluorescence in 1xMycR-infected cells, using anti-Myc antibodies. To unambiguously distinguish signals resulting from authentic infection and background signals, we preformed the infections with human cells (osteosarcoma; U2OS) that cannot be normally infected due to the lack of the MLV receptor; or with a cognate, U2OS-derived, cell line (U/R) that stably expresses the mCAT-1 murine receptor for MLV [39], and is susceptible to MLV infection (Materials and Methods). Parental U2OS or U/R cells were infected either with 1xMycR or wt viruses, fixed and immunostained with anti-Myc antibody. Clear punctate fluorescence could be detected in the cytoplasm of 1xMycR infected U/R cells (Fig. 3A). We also observed a punctate staining of p12 that was in close proximity to the chromosomes, in dividing cells that were identified by their condensed chromosomes (Fig. 3B and see below). In the control U/R cells, p12 staining was not observed following infection with the wt virus (Fig. 3C); and very low p12 staining was observed in 1xMycR-infected parental U2OS cells (Fig. 3D), likely representing a low level of adherence of the virus to the cells. In addition, no immunofluorescence was observed in mock-infected U/R cells or in 1xMycR-infected U/R cells that were stained with the secondary, but not with the primary anti-Myc antibody (data not shown). Thus, the majority of the punctate staining represents a signal that is derived from an authentic infection, and not from background staining or non-specific adherence of viral particles or antibodies to the cell membrane. Further support to this idea came from the direct correlation that was found between the number of p12 puncta in infected cells and the amount of 1xMycR virions that were used for infection (Fig. S1), suggesting that these dots represent p12 proteins that originated from physical particles. The punctate pattern of staining as opposed to a diffuse staining, suggests that the p12 molecules are associated in a complex. Similar punctate staining was observed when cells derived from the natural host of MLV (murine NIH3T3 cells) were infected with the 1xMycR virus (Fig. 3E). Of note, whereas clear immunofluorescence signal was detected in 1xMycR-infected cells, no such signal was observed in cells that were infected with the previously described replication-competent MLV that contained a Flag-tagged p12 [33], and that were stained with anti-Flag antibodies (data not shown). This discrepancy may result from the different locations of the tags in p12 and from the fact that the Flag epitope replaced few residues in p12, while the Myc epitope was inserted into the complete sequence of this protein, rendering this tag in the in 1xMycR virus more accessible to the antibodies.
The early detection of p12 at 12 h postinfection (Fig. 3), suggested that the immunofluorescence detected the p12 proteins of the incoming virus and not the p12 domain of Gag precursors that were synthesized in late stages of the infection. To further examine this issue, we arrested the cell cycle of U/R cells by serum starvation and aphidicolin treatment, as this procedure was shown to block nuclear entry and integration of the incoming MLV, averting Gag synthesis in later stages of infection [21]. We then infected the cells with 1xMycR virus and observed the same punctate fluorescent signal for p12 in the aphidicolin-treated cells (Fig. 3F), confirming that p12 staining is of the incoming virus and not of newly synthesized Gag precursors.
The p12 staining in aphidicolin-treated cells (Fig. 3F) was mainly restricted to the cytoplasm, as was observed for cells that were not treated with aphidicolin and that were in the interphase stage of the cell cycle (Fig. 3A). In contrast, p12 was detected in both the cytoplasm and in the vicinity of chromosomes of dividing cells (Fig. 3B). To further study and quantify this observation, cycling non-synchronized U/R cells were infected with 1xMycR virus and 12 h postinfection the cells were fixed and stained for p12 and the cellular DNA as described above. Mitotic cells were identified on the basis of chromosome condensation and the typical configuration of each mitotic stage. In addition, the extent of p12 staining (red) that overlapped the staining of cellular DNA (blue) was determined (see Materials and Methods), to quantify the distribution of p12 proteins between the cytoplasm and the chromosomes. Fig. 4 shows images of 1xMycR-infected cells at different stages of the cell cycle, and the quantification of p12 distribution is represented in Fig. S2. As can be seen, p12 was mainly detected in the cytoplasm of interphasic cells (Fig. 4A) with only 6% overlap between p12 and the chromosomes (Fig. S2), similar to what was shown above (Fig. 3A and F). In contrast, cells at different stages of mitosis showed a much higher overlap between the p12 and the chromosomes (Fig. 4B–E); the percentages of this overlap were quantified to be 49, 75, 68 and 72% at prometaphase, metaphase, as well as early and late anaphase, respectively (Fig. S2). No p12 signal was detected in U/R cells that were infected with wt virus and used to control for the immunofluorescence specificity (Fig. 4F). Overall, complexes of p12 proteins appeared to migrate from the cytoplasm to the nucleus and to accumulate at the vicinity of the chromosomes in dividing cells.
To further study the association of p12 with the chromosomes, optical sections were generated for the mitotic chromosomes in infected cells (inset, Fig. 4E). This microscopy analysis clearly demonstrated that p12 proteins were detected at the same plane of the chromosomes (Fig. 4G). Reconstitution of the serial optical sections into a 3D image further showed the close proximity between p12 proteins and the chromosomes (Fig. 4H). Thus, p12 proteins appear to be closely associated with, and even imbedded in, the mitotic chromosomes of infected cells.
The above results implied that p12 proteins are found in complexes that migrate from the cytoplasm to the chromosomes in dividing cells. These resemble the characteristics of the PIC [21], suggesting that p12 may be a constituent of this complex. If this notion is true then in infected cells p12 should co-localize with components of the PIC but not with virion constituents that are not part of the PIC. To test this, we performed immunofluorescence analysis to detect the spatial relationships between p12 and MA - a virion constituent that is thought to become dispersed in the infected cell upon the uncoating step and is not known to be part of the MLV PIC; and CA - a virion constituent that is part of the MLV PIC [12]. 1xMycR-infected U/R cells were stained with anti-Myc antibodies, together with anti- MA or CA antibodies. This analysis revealed that, as expected for virion constituents, the p12 (red fluorescence) and MA (green fluorescence) proteins co-localized (yellow fluorescence) in particles that adhered to the cover glass in regions that were free of cells (Fig. 5A, B). In contrast, no such overlap could be detected between the p12 and MA in the infected cells. The lack of co-localization between p12 and MA in the infected cells was further emphasized in conditions where p12 proteins migrated towards the mitotic chromosomes (Fig. 5B). When the distribution of CA and p12 was tested, co-localization of the two proteins was clearly detected both in virions that adhered to the glass outside the cell and in the infected cell (Fig. 5C). Altogether, these immunofluorescence analyses demonstrated that in the virions p12 proteins are co-localized with MA and CA, yet in the infected cells p12 is associated with the CA but not with the MA proteins, suggesting that p12 puncta in infected cells are derived from uncoated virions and not from internalized virus particles, and further hinting for p12 association with the PIC.
To further address the possibility that p12 molecules are indeed part of the MLV PIC we set to visualize the PICs in infected cells and to test the possible co-localization of p12 with these PICs. Since the presence of the viral DNA genome is the hallmark of the PIC, its detection by FISH has been used to identify MLV PICs by microscopy [21]. Thus, we aimed to detect the incoming PICs in early stages of infection, using DNA FISH in combination with immunofluorescence to define the time-dependent spatial correlation between the viral genomic DNA and the p12 protein.
For this analysis we used U/R cells, to avoid the problem of cross-hybridization with endogenous MLV-like elements found in mouse cells [40]. In addition, to avoid detection of carry-over of MLV plasmid DNA from transfected cells, we used viruses from chronically infected NIH3T3 cells. In preliminary FISH experiments we could readily detect clear and punctate fluorescence staining in U/R, but not in U2OS, cells that were infected with wt virus and that were hybridized with a MLV-derived, biotin-labeled probe that was detected with a Cy3-conjugated avidin, demonstrating the specificity of the detection method (Fig. S3 and see below). This staining was also reminiscent of the previously reported staining of MLV PICs by FISH in Rat-1 cells [21]. We then established conditions for immunofluorescence combined with DNA FISH (see Materials and Methods). U/R or U2OS cells were infected with 1xMycR or wt viruses, and 12 h postinfection, cells were stained with anti-Myc antibody and a secondary Cy3-conjugated antibody. FISH analysis followed, using a MLV-derived biotinylated probe and FITC-labeled avidin for the detection of the probe. A clear overlap (yellow) between the p12 proteins (red) and the MLV genomic DNA (green) was observed in U/R cells infected with 1xMycR virus (Fig. 6A–E). This overlap could be observed in the cytoplasm of infected cells (Fig. 6A, B, D and E), as well as in the vicinity of chromosomes (Fig. 6C), including condensed chromosomes of mitotic cells (Fig. 6B, E and S4). In contrast, such broad overlapping signals were absent in 1xMycR-infected U/R cells that were processed for immunofluorescence/FISH analysis at two hours postinfection (Fig. 6F). In this setting, only extensive punctate staining of p12 was observed, probably reflecting the lack of complete reverse transcription at this early time point. Quantification of more than 600 fluorescent dots in infected cells (Fig. S5) confirmed the above observations: at 12 h postinfection approximately 60% of the fluorescent puncta showed an overlapping signal between the p12 proteins and the viral genomic DNA, while approximately only 10 and 30 percents of the fluorescent dots showed such an overlap at earlier time points (2 and 6 h postinfection, respectively). In addition, at 12 h post infection about 10% of the total fluorescent dots, showed overlapping p12 and genomic DNA signals that could be located with the chromosomes. In contrast to these results, more than 99% of the extracellular dots (representing extracellular virions attached to the glass in cell-free regions, see Fig. 5), showed only p12 staining both at early (2 h) and late (12 h) time point postinfection (Fig. S5), likely due to the absence of efficient reverse transcription in these particles. Additional negative controls showed no overlapping signals between the p12 and the genomic DNA staining further emphasizing the genuineness of this analysis. These controls included: 1xMycR-infected U/R cells that were processed as above but without the addition of the primary anti-Myc antibody, showing only the green FISH signal (Fig. 6G); U/R cells, infected with the wt virus that lacks the Myc epitope, showing only the green FISH signal (Fig. 6H) and 1xMycR-infected U2OS cells, showing neither red nor green punctate fluorescence (Fig. 6I). Thus, the overlapping staining in 1xMycR-infected U/R cells, which was absent from the extracellular particles and from the negative controls, suggested that p12 proteins associate with viral genomic DNA and hence, p12 is indeed a component of the MLV PIC. Moreover, the overlap between p12 protein and viral DNA that was detected in close proximity to the condensed chromosomes (Fig. 6B, E and S5) suggested that p12 proteins escort the viral DNA until very close to the chromatin of dividing cells.
As was suggested above, the absence of intracellular overlapping immunofluorescence/FISH signals at two hours postinfection (Fig. 6F) likely reflected the lack of complete reverse transcription at this early time point, and may further suggest that reverse-transcription is not a prerequisite for the formation of p12 puncta in infected cells. We further investigated this point by the generation of MLV virus-like particles (VLPs), with or without the genomic RNA, and that their p12 proteins were Myc-tagged as the 1xMycR virus (Materials and Methods). These VLPs were then used to infect U/R cells, which were examined by immunofluorescence for the generation of p12 puncta. This analysis revealed a similar formation of p12 puncta, and similar accumulation of these dots close to mitotic chromosomes, for both conditions (Fig. S6). These results suggest that the presence of the genomic RNA, and its subsequent reverse transcription, are not a prerequisite for the formation and migration of p12 puncta in infected cells.
It should be noted that in 1xMycR-infected U/R cells, we also observed signals for the viral DNA and the p12 proteins that did not overlap (Fig. 6D and E; triangles and ellipses, and Fig. S5). Since only a fraction of the incoming viruses establish a productive infection [2], [41], [42], our microscopic analysis cannot resolve between the following two options: 1) only PICs that include functional p12 proteins have the potential to complete the early steps of infection or; 2) PICs that do not contain p12 are the infectious ones. Since p12 is crucial for MLV infection [3], [31], [33], in the latter scenario, p12 proteins function not as part of the PIC but autonomously, outside of this complex.
To address the question whether p12 modifies PIC function as part of the PIC or by acting separately from the PIC, we designed and applied a genetic ‘rescue’ assay based on complementation of wt and mutant p12 proteins. If p12 functions as a constituent of the PIC, upon co-infection of wt and p12 mutant viruses the wt p12 proteins should not rescue the defect in integration of PICs derived from the p12 mutant virus (Fig. 7A). In contrast, rescue should occur if p12 proteins act autonomously of the PIC, particularly in conditions where wt p12 proteins are present in large excess over mutant p12 proteins (Fig. 7B). Two types of MLV particles were generated for this assay: The first, named MLVIRES-GFP, was expressed from the pNCAIRES-GFP clone, a replication-competent MLV with a GFP marker [43], allowing the accurate determination of its titer. MLVIRES-GFP expresses wt p12 proteins. The second type was made of VLPs that were generated from helper plasmids expressing the MLV Gag with or without the PM14 mutation in p12 (Fig. 1), and the Pol and Env (ecotropic) proteins; these VLPs also encapsidated the pQCXIN retroviral vector (Clontech), expressing the neomycin resistance gene (Neor). Importantly, pQCXIN is a self-inactivating vector due to a deletion in the U3 sequence of the 3′ LTR, which renders the vector compatible for only a single cycle of infection even in the presence of replicating virus such as the MLVIRES-GFP. We chose the PM14 mutation for this analysis since a virus that carries this mutation is capable of normal reverse transcription but is defective in integration [31]; yet, in our hands, VLPs with the PM14 mutation could transduce the pQCXIN vector at low but detectable efficiencies, allowing the quantification of p12 function in this complementation assay.
NIH3T3 cells were infected with dilutions of wt VLPs (MOI of 0.2, 0.1 or 0.01), or PM14 VLPs (with RT activities that matched the ones of the wt VLPs dilutions), in the presence or absence of MLVIRES-GFP (MOI of 10). In these settings, the wt p12 proteins, derived from MLVIRES-GFP, are present in large excess over decreasing amounts of VLP-derived p12 proteins (wt or mutant) in infected cells. Infected cultures were then selected in G418- containing medium for two weeks, after which the number of Neor colonies was determined (Fig. 7C). The results of this experiment showed that in all tested ratios, no increase in the number of Neor colonies was observed for the PM14 VLPs (Fig. 7C, white columns) in the presence of MLVIRES-GFP, indicating that p12 proteins act from within the PIC. In fact, co-infection of MLVIRES-GFP with either wt (Fig. 7C, black columns) or mutant VLPs resulted in a reduction in the number of Neor colonies; the level of this decrease augmented with the increase of MLVIRES-GFP/VLP ratio. This phenomenon likely represents competition between MLVIRES-GFP and the VLPs over cellular factor(s) needed to establish the infection. Of note, the reduction in the number of Neor colonies in the presence of MLVIRES-GFP was greater for the PM14 VLPs, compared to the wt VLPs, at all the tested ratios. For example, when the presence of MLVIRES-GFP reduced the infectivity of wt VLPs by less than two-fold, a reduction of more than eight-fold was observed for PM14 VLPs. These results suggest that the PM14 mutant is more sensitive to the competition exerted by the MLVIRES-GFP virus.
Similar results were obtained when we repeated this experiment using VLPs (wt and PM14) that were pseudotyped with the vesicular stomatitis virus G glycoprotein (VSV-G) instead of the ecotropic MLV envelope protein, to avoid direct competition between the VLPs and the MLVIRES-GFP virus on the mCAT-1receptor, (Fig. 7D). No rescue of the infectivity of the p12 mutants was observed when these experiments were repeated using different conditions that included lower MOIs, the use of wt VLPs (instead of the replicating MLVIRES-GFP virus) as a source for the wt p12 proteins, and PM14 VLPs encapsidating a vector with a different selection marker (data not shown). Overall, these experiments provide genetic evidence that p12 functions as part of the PIC.
As mentioned above, the exact composition of retroviral PIC is not known, however the presence of the viral genomic DNA is the hallmark of the PIC. To further demonstrate that p12 is a component of the PIC we aimed to co-immunoprecipitate the viral genomic DNA with the p12 proteins. To immunoprecipitate p12, we infected NIH3T3 cells with wt or 1xMycR virus, lysed the cells and performed IP with anti-Myc, or control anti-Flag monoclonal antibodies. We then analyzed the cell lysates and the immunoprecipitates for the presence of the viral genomic DNA by PCR. This analysis revealed preferential IP of the genome of the 1xMycR virus when anti-Myc antibodies were used, compared to the controls (Fig. 8A). To better quantify this, we measured the amount of the viral DNA genome in the cell lysates and the immunopellets by quantitative PCR (qPCR) and calculated the relative efficiency of this IP (Fig. 8B and S7A). The average IP efficiency, obtained from three independent experiments, showed that when anti-Myc antibodies were used, the genome of the 1xMycR virus was immunoprecipitated approximately 7 fold higher than the genome of the wt virus. In addition, the average IP efficiency of the 1xMycR genome by the anti-Myc antibodies was approximately 25 fold higher than the one obtained for IP using anti-Flag antibodies. This specific IP of the viral genomic DNA with antibodies against p12 strongly suggested that p12 is indeed part of the PIC.
Since CA was identified as a component of the MLV PIC [14] we also tested if this protein co-immunoprecipitates with the p12 proteins. Co-IP experiments were performed similarly to what was described above for the IP of the viral genomic DNA. The presence of CA in cell lysates and in immunopellets was determined by Western blot analysis with polyclonal antibodies against CA (Fig. 8C). This analysis revealed that CA could be detected in the precipitate only when the precipitation included the lysates of the 1xMycR-infected cells and the anti-Myc antibodies; CA protein was absent from control precipitates obtained from wt-infected cell lysates that were reacted with anti-Myc antibodies, or from lysates of 1xMycR-infected cells that were reacted with the control anti-Flag antibodies.
To verify that the detected p12-CA interaction reflects an authentic, intracellular interaction and not an interaction present in internalized virus particles, we carried out an analogous pull-down experiment on both extracellular virions and infected cells. The results of this experiment clearly demonstrate that CA of the 1xMycR virus was immunoprecipitated by anti-Myc antibodies only from lysates of infected cells and not from lysates of extracellular virions (Fig. S7B and C). These results indicate co-association of CA - a known component of the MLV PIC - and p12 proteins, further providing evidence that p12 is indeed a component of the PIC.
Our data provide strong evidence that p12 is a functional component of the PIC. Yet, the exact function of p12 is currently unknown. In principle, p12 may influence one or more steps that include: migration of the PIC along the cytoplasm, nuclear entry and targeting the chromatin for integration. Elaborate biochemical analysis of cells infected with wt MLV or with an integration-defective p12 mutant - the PM14 virus - demonstrated that the two viruses have the same distribution of the viral genomic DNA in cytoplasmic and nuclear fractions; yet no circular genomic DNA was detected for the mutant virus, suggesting that PM14 virus is defective in nuclear steps required for productive infection [3]. However, as the authors suggested, association of the PIC of the PM14 virus to the external side of the nuclear envelope and/or nuclear retention could not be dismissed. We tested if the immunofluorescence procedure that we developed to investigate 1xMycR infection could be applied to the analysis of mutant p12 proteins. For this we introduced the PM14 mutation (Fig. 1) into p12 of the 1xMycR clone to generate the 1xMycR/PM14 virus. We then separately transfected the 1xMycR and the 1xMycR/PM14 clones into 293T cells, harvested the virion-containing supernatants and infected sub-confluent U/R cultures with equal amounts of virions (normalized by an exogenous RT assay). The infected cells were processed for immunofluorescence analysis 12 h postinfection with anti-Myc antibodies to detect the p12 proteins. This analysis revealed that both viruses showed similar intracellular distribution in interphasic cells, where the majority of the p12 proteins were cytoplasmic (Fig. 9). However, a remarkable difference between the wt and the mutant virus was observed in mitotic cells, where the p12 proteins of the 1xMycR/PM14 virus were localized mostly to the cytoplasm, in contrast to the substantial association of the p12 proteins of the 1xMycR virus with the chromosomes (Fig. 9). Further quantification of this phenomenon revealed that in interpahsic cells about 8% of the p12 proteins that were derived from the 1xMycR/PM14 virus overlapped the chromosomes, similar to the value (6%) obtained for the 1xMycR virus (Fig. S8). In mitotic cells, however, where almost 70% of the p12 proteins of the 1xMycR virus overlapped the chromosomes, the p12 proteins with the PM14 mutation showed only 11% overlap - a level that was similar to the one obtained in interphasic cells (Fig. S8). These results demonstrated that the PM14 mutation hindered the movement of the p12 proteins from the cytoplasm towards the chromosomes and strongly suggest a role for p12 in PIC trafficking.
In this work we provide microscopic, genetic and biochemical evidence that the p12 protein - a Gag cleavage product - is a functional constituent of the MLV PIC.
We used fluorescent microscopy analysis to detect Myc-tagged p12 proteins of a replication-competent virus (1xMycR), in infected cells. The comparison between the epitope-tagged virus and the untagged wt virus, and between infected cells that express, or do not express, the mCAT-1 receptor for MLV, allowed us to distinguish unambiguously between background signals and immunofluorescent staining that was related to authentic infection of the incoming virus. This is important since MLV particles with the ecotropic envelope may bind human cells independently of specific receptor-envelope interactions, and such binding can be detected by immunofluorescence [42].
When susceptible U/R and NIH3T3 cells were infected with the 1xMycR virus, our immunofluorescence analysis revealed punctate p12 staining at an early time of infection. This staining was attributed to the incoming virus, and not to newly synthesized Gag proteins since it could be detected as early as two hours postinfection, and in aphidicolin-treated cells, where integration and the subsequent Gag expression are inhibited. Of note, throughout the immunofluorescence experiments, the cells were infected with MOI of approximately 3, yet many of the infected cells exhibited hundreds to thousands of p12-positive dots This likely reflects the fact that for retroviruses only a minute fraction of the virions is infectious [41], [42]. The number of these dots was in a direct correlation with the amount of the virus used for infection. Thus, it is likely that the majority of these dots represent p12 proteins that originated from physical particles that are defective in their infectivity.
The punctate, rather than diffuse, pattern of the staining of p12 suggested that p12 proteins are found in a complex in early stages of the infection cycle. Furthermore, this staining was mainly cytoplasmic in aphidicolin-treated cells or in mitotic cells early after infection, in contrast to its localization adjacent to mitotic chromosomes 12 h postinfection. Thus, the pattern and the distribution of p12 proteins in dividing, and in cell-cycle arrested cells, resemble the one of the MLV genomic DNA [21], suggesting that p12 is part of the MLV PIC. In support of this idea are the findings that whereas p12 staining overlapped the staining of MA or CA in virions, in the infected cells p12 staining co-localized only with that of CA, which unlike the MA protein is thought to be part of the MLV PIC. However, in principle, p12 proteins can migrate towards the cellular chromatin as PIC-independent complexes. Therefore, to better evaluate p12-PIC interaction we combined immunofluorescence with FISH analyses, and detected co-localization of p12 proteins with the viral genomic DNA, both in the cytoplasm and in close proximity to the cellular chromatin of infected cells. Since the viral genomic DNA is the hallmark of the PIC, this co-localization suggested that p12 is indeed part of the PIC. The early detection of punctate p12 staining at two hours postinfection, when almost no FISH signal for the genomic DNA could be detected, hints for organization of p12 proteins in a complex that precede the PIC, most likely in the RTC. Moreover, the appearance of p12 puncta in cells that were infected by MLV VLPs with no encapsidated genomic RNA suggests that the presence of the genomic RNA is not necessary for the early organization of p12 in complexes and their trafficking towards the chromosomes. In this scenario, the protein components of the RTC and the PIC may have intrinsic assembly ability which is independent of the presence of the genomic RNA and/or cDNA.
Our analysis revealed, however, that not all the signals of the p12 proteins and the viral DNA genome overlap, raising the possibility that p12 proteins may function in early stages of infection not as constituents of the PIC but as elements that are separated from this complex; in any scenario, such variation in the association of p12 with the viral genomic DNA may account in part for the fact that only a fraction of MLV particles are infectious. Since our microscopic examination could not unambiguously distinguish between these two scenarios describing p12-PIC interactions, we performed a genetic ‘rescue’ assay to evaluate if p12 proteins exert their activity in trans, in respect to the PIC, or as constituents of this complex. In this assay, VLPs that were defective in early steps of infection - between reverse transcription and integration - due to a mutation in p12 (PM14 mutation; [31]) were used together with wt virus (pNCAIRES-GFP) to co-infect NIH3T3 cells. The presence of wt p12 proteins in the infected cells did not improve the infectivity of the mutant VLPs and this was true for all wt-to-mutant ratios tested, including conditions where wt p12 proteins were present in a very large excess over the mutant p12 proteins. These results strongly suggest that p12 proteins function as part of the PIC. The co-infection experiments also revealed that excess of pNCAIRES-GFP particles repressed the infection of the VLPs regardless if the latter harbored wt or mutant p12 proteins, and that the level of this inhibition rose with the increase in the pNCAIRES-GFP-to-VLPs ratio. This probably reflects a competition between the pNCAIRES-GFP virus and the VLPs on cellular factor(s) needed to establish infection and that are found in limiting amounts. The ecotropic receptor may serve as such a factor since more Neor colonies were obtained when the pNCAIRES-GFP virus and the VLPs (transducing the Neor vector) had different (the ecotropic and the VSV-G glycoprotein, respectively), instead of the same, envelopes, resulting in the use of different receptors to initiate infection. However, even in such settings, excess of pNCAIRES-GFP over the VLPs resulted in a dose-dependent reduction in the infectivity of the latter, suggesting that saturable cellular factors other than the ecotropic receptor exist. Notably, the PM14 VLPs were more sensitive to the pNCAIRES-GFP-mediated suppression, compared to wt VLPs. This phenomenon may indicate that PM14 VLPs are kinetically slower than wt VLPs in transducing the vector. In this scenario, wt PICs may traffic towards integration faster than the PICs with mutant p12 proteins and accordingly, saturate cellular factors needed for infection, before such factors will be available for the mutant PICs.
The above results suggest that p12 is a constituent of the PIC and thus, it is anticipated that the viral genomic DNA should be co-immunoprecipitated with p12 proteins. Indeed, preferential IP of the genome of the 1xMycR virus with anti-Myc antibodies from infected cells was identified, compared to the controls (made of the wt virus or anti-Flag antibodies), providing strong evidence that p12 proteins are found in a complex with the viral genomic DNA. Further biochemical support to the interaction of p12 with the PIC came from the detection of CA, another component of the MLV PIC [14], in the p12 immunoprecipitates, as CA proteins were immunoprecipitated from lysates of cells that were infected with the 1xMycR virus, using anti-Myc antibodies. It should be noted that we failed to directly detect the Myc-tagged p12 proteins in the pellets where CA proteins were found. This may be the result of the use of a monoclonal antibody (anti-Myc) to detect p12, unlike CA detection that was performed using polyclonal sera that may enhance the sensitivity of the detection. Support for this explanation comes from the fact that Western blot analysis of virions showed an intense signal for CA but a relatively weak signal for p12, although the two proteins are present in equimolar quantities inside the viral particles; the difference in detection sensitivity by Western blot analysis was also observed when only CA, but not p12, was detected in extracts of the 1xMycR-infected cells prior to the IP step (data not shown). Importantly, although this analysis was not sensitive enough to detect the tagged p12 proteins, CA IP was specific since it was observed only when the 1xMycR virus and the anti-Myc antibody were used, and not when the co-IP was performed with controls that included an isotype-matched control antibody (IgG1, anti-Flag) or the untagged, wt virus. In the co-IP experiments, only low amounts of CA were detected in the pellets, compared to its level in the cell lysates. This may reflect a labile p12-CA interaction and/or the possibility that not all of the CA molecules interact with p12 proteins. At this stage, it is not clear whether the p12 and CA proteins are complexed together through direct or indirect interactions; yet, our success to detect this interaction in infected cells but not in extracellular virions implies that at least some of the p12 and CA molecules in the virion form new mutual interactions during or after the uncoating step. In any case, these co-IP experiments provide a biochemical support to the notion that p12 and CA proteins have a cooperative effect in early stages of infection, as was concluded from the analysis of swap mutants between the different domains of the Gag of MLV and its closely related virus - the SNV [37]. Because CA is a component of the MLV PIC and since our analysis suggests that p12 is also a functional constituents of the same complex, the two proteins may act in concert to direct the PIC towards nuclear entry and integration. In line with this idea is the observation, made by Yuan et al. [3], and Lee et al. [37], regarding the phenotypic similarity that exists between wt MLVs that are restricted by the Fv1 restriction factor and mutant MLVs harboring either specific mutations in p12, or p12/CA chimeras derived from MLV and SNV. In each of these cases virus infection is characterized by normal levels of reverse transcription but defective production of genomic circular forms and integration.
If p12 proteins act in directing the PIC towards integration, their function(s) may be required during PIC migration through the cytoplasm, nuclear entry and/or nuclear trafficking. Remarkably, we demonstrated here a clear accumulation of the p12 proteins adjacent to mitotic chromosomes, and this accumulation appeared to increase along the steps of mitosis; ranging from 6% of p12 proteins that overlapped the chromatin during interphase, to about 70% overlap during mitosis. Moreover, the PM14 mutation in p12 that renders the virus integration-defective in vivo [3], also hindered the accumulation of p12 proteins with the mitotic chromosomes, as the majority (∼90%) of the puncta of the p12 mutant proteins remained cytoplasmic in dividing cells. These results are in agreement with the postulation, discussed above, that wt PICs are kinetically faster in their trafficking towards integration, than PICs that are derived from the infection of the PM14 virus. Yet, it should be noted that Yuan et al. [3], observed that the distribution of the viral genomic DNA between the cytoplasm and the nuclear fractions was similar for both the wt and the PM14 mutant viruses. While our microscopic analysis revealed such similarity for the p12 content in the nucleus of interphasic cells, our results also showed differential nuclear-cytoplasmic distribution of p12 in mitotic cells, for the two viruses. This difference can be explained by the different methodologies used: whereas our microscopy analysis monitored the infection in individual mitotic cells, in which the nuclear envelope is broken, the biochemical fractionation used by Yuan et al., relayed on the lysis of unsynchronized infected cultures and the subsequent separation of intact nuclei from the cytoplasm. In such conditions, mitotic cells with no intact nuclei can be overlooked, especially if these cells present only a fraction of the unsynchronized culture.
Altogether, our results suggest that p12 proteins are associated with the viral genomic DNA and indicate for a role(s) of p12 along the course of trafficking of this DNA from the cytoplasm to the chromosomes. One possibility of such a role is that p12 interacts with cytoplasmic factors that are required for trafficking towards the nucleus. Such specific factors have not been identified for MLV, but may involve cytoskeleton proteins, as was described for HIV and other retroviruses [3]. In addition, the extensive accumulation of p12 in close proximity to the chromosomes hints for a role of p12 very close to the integration step and may even indicate for a direct function in the integration process itself. However, Yuan et al. [3], have demonstrated that PICs derived from a MLV with a mutation in p12 that render the virus integration-defective in vivo, were integration-competent in an in vitro assay; demonstrating that a defect in p12 function does not necessarily hamper PIC-mediated integration. Thus, p12 may be needed at the vicinity of the chromatin for steps that precede the integration reaction itself. What could be such a role? For HIV-1, it has been demonstrated that the cellular protein LEDGF is a component of the PIC [10], [11] and may serve to tether the IN proteins to the chromatin [10], [44], [45], [46], [47]. LEDGF interacts with IN proteins derived from several lentiviruses but not with MLV IN [10], [48], and no factor with such tethering activity has been described for MLV. In light of the similarity of the N-termini of p12 and the histone H5 protein [49], and because our microscopic data showed an intimate association of p12 with the chromatin it is tempting to assume that p12 also functions in tethering the MLV PIC to the chromatin. Such a scenario may explain why specific mutations in p12 block MLV integration in live cells but do not interfere with integration into naked DNA, in vitro [3].
The ability to detect MLV PICs by immunofluorescence through the detection of p12 proteins provides a new tool that may assist the analysis of several issues concerning MLV infection: (i) such microscopic analysis should complement the biochemical approach used to study p12 mutants and resolve, for example, the question whether PICs of specific p12 mutants are capable of entering the nucleus or are trapped on the external side of the nuclear envelope [3]; (ii) p12 tagging may also be applied for the analysis of the recently discovered human retrovirus, the XMRV, which expresses p12 protein with a high degree of similarity to the MLV p12; suggesting a way to monitor XMRV infection in human cells; (iii) Means were developed to visualize HIV-1 RTC/PIC, which were also exploited to study the interaction of these complexes with the TRIM5α restriction factor [50]. Similarly, tagging p12 proteins of N or B -tropic MLVs with the Myc epitope, as was described here for the 1xMycR virus, should allow the analysis of the restriction of these MLVs by Fv1 [51], and TRIM5α-mediated restriction of N-tropic MLV [52], [53].
In summary, the combination of microscopic, genetic and biochemical assays described here provide strong evidence that p12 is a component of the MLV PIC and this interaction is crucial for the progression of the PIC towards integration.
The pNCS plasmid contains an infectious molecular clone of the Moloney MLV [54], and a simian virus 40 origin of replication in the plasmid backbone. Overlapping PCR was used to insert three tandem repeats of the Myc epitope (EQKLISEEDL), between amino acids 45 and 46 of p12 in pNCS, generating the 3xMyc clone. Virions of a faster replicating revertant, which was derived from the 3xMyc virus, were collected and the genomic RNA was extracted and reverse transcribed with MLV RT and random primers (Promega). The cDNA was amplified by PCR using Ex-Taq (Takara Bio Inc.) and a forward primer (5′CCCAGGTTAAGATCAAGG3′, derived from the matrix sequence), together with a reverse primer (5′CTTGGCCAAATTGGTGGG3′, derived from the capsid sequence). The resulting 875 bp fragment was cloned into pTZ57R (Fermentas) and sequenced, revealing that the revertant virus retained only a single, in-frame copy of the Myc epitope in p12. The cloned PCR fragment was digested with BsrGI and XhoI and the resulting 640 bp fragment, containing the entire p12 and the Myc epitope sequence, was used to replace the BsrGI-XhoI wt sequence in pNCS, to generate the 1xMycR clone. Myc-tagged proteins were detected by Western blot analysis using mouse monoclonal anti-Myc antibody (supernatant of 9E10 hybridoma), and a secondary donkey anti-goat horseradish peroxidase (HRP)-conjugated antibody (Jackson Immunoresearch Laboratories, product no. 705-035-147). The PM5 or PM14 mutations in p12 [31] were introduced by overlapping PCR into the BsrGI-XhoI fragment of pNCS to generate PM5 or PM14 clones, respectively. Overlapping PCR was also used to combine the Myc epitope of the 1xMycR virus with the PM14 mutation (Fig. 1) in the above BsrGI-XhoI fragment to generate the 1xMycR/PM14 clone. pNCAIRES-GFP encodes for a replication-competent MLV, which expresses the green fluorescent protein (GFP) under the translational control of the encephalomyocarditis virus internal ribosome entry site (IRES) [43]. This clone was generously provided by Jeremy Luban (University of Geneva).
To generate MLV VLPs with Myc-tagged p12 proteins, containing or lacking the genomic RNA, the following plasmids and procedure were used: the helper plasmid pVSV.G expresses the VSV-G glycoprotein. The pGag-PolGpt.p12 1xMycR, is a derivative of the pGag-PolGpt helper plasmid that expresses the MLV Gag and Pol proteins, and its p12 sequences was modified to include the Myc epitope tag as in 1xMycR virus (Fig. 1). The pQCXIP plasmid encodes an MLV-based vector (Clontech). The pQCXIPΔ5′ encodes for a defective vector that was generated by deleting an internal BsrGI fragment from pQCXIP, resulting in the removal of the 5′ LTR and the packaging signal. VLPs were generated by transfecting subconfluent 293T cells in 60 mm plates with 10 µg of pQCXIP or pQCXIPΔ5′ together with 7.5 µg of pGag-PolGpt.p12 1xMycR and 2.5 µg of pVSV.G DNAs, using the calcium phosphate procedure. Culture supernatants were harvested two days posttransfection and used for infection.
Human embryonic kidney 293T cells, human osteosarcoma U2OS cells, and the mouse NIH3T3 fibroblasts were cultured in Dulbecco's Modified Eagle Medium (DMEM), supplemented with 10% heat-inactivated fetal calf serum (FCS), 2 mM L-glutamine, penicillin (20 U/ml), streptomycin (20 µg/ml) and nystatin (2.5 U/ml), in a humidified incubator at 37v and 5% CO2. All tissue culture products were purchased from Biological Industries (Beit Haemek, Israel). To generate human cells stably expressing the murine receptor for MLV, U2OS cells were transfected with 7.5 µg of pCDNA 3.1 Zeo(+) that encode for the murine receptor for MLV (mCAT-1; [39]), using the calcium phosphate precipitation method. Transfected cells were selected in the presence of 100 µg/ml zeocin for two weeks and zeocin-resistant colonies were expanded and screened for their susceptibility to MLV infection. One of the clones (clone #12, named hereafter U/R), was chosen for further experiments since it could efficiently be infected with MLV particles encapsidating a MLV vector that expresses the GFP reporter gene (pQCXIP-gfp-C1 vector [55]; data not shown).
To determine the kinetics of spreading of various MLV clones in NIH3T3 cultures, the cells were either infected or transfected with these clones as indicated. The transfection was carried out using the DEAE-dextran method, as described before [56].
When required, cell synchronization was achieved by arresting cell division before S phase by serum starvation and aphidicolin treatment, as previously described [21] and briefly explained here. On day 1, cells were plated on 13 mm round cover-slips in a 24-well plate at approximately 5% confluency per well. On day 2, cells were serum-starved by removing the media, washing the dish with serum-free media and then adding DMEM containing 0.25% FCS. On day 4, serum was added to a concentration of 10%. 6 h after the addition of serum, aphidicolin (2 µg/ml) was added. After 6 h, the medium was removed and fresh medium containing the virus, 10% serum, polybrene (hexadimethrine bromide; 8 µg/ml) and aphidicolin (2 µg/ml) was added. After 2 h, the virus-containing medium was removed and replaced with 10% serum-containing medium and aphidicolin (2 µg/ml). After an additional 12 h, cells were fixed and subjected to immunofluorescence analysis.
Cells were grown on 13 mm round cover-slips in a 24-well plate to ∼1.5×105 cells/well and were infected with the indicated virus with multiplicity of infection (MOI) of approximately 3 [virus preparations were estimated to have ∼5×106 infectious units (IU)/ml, based on comparisons of their RT activity, determined by RT exogenous assay [57], to a standard MLV stock with known IU concentration. This stock was made of pNCAIRES-GFP, a replication-competent MLV that expresses the GFP marker [43], allowing the determination of its titer by measuring the number of GFP-positive cells in infected cultures, by fluorescence-activated cell sorting (FACS) analysis]. Cells were incubated with the virus for the indicated time, after which they were washed three times with PBS and fixed with 4% paraformaldehyde for 20 min. After three washes with 50 mM glycine in PBS, the cells were permeabilized with 0.1% Triton in PBS for 2 min and immediately washed three times with PBS. The cells were then incubated with blocking solution [1∶10 normal goat serum in Tris-Buffered Saline (TBS; 50 mM Tris-HCl, 150 mM NaCl pH 7.5)] for 30 min, followed by one hour incubation with a mouse monoclonal anti-Myc antibody (supernatant of 9E10 hybridoma, diluted 1∶6 in TBS), and washed once with TBS and twice with PBS (5 min each wash). The cells were then incubated for one hour with a Cy-3-conjugated goat anti-mouse antibody (Jackson Immunoresearch Laboratories, product no. 115-166-072) diluted 1∶500 in TBS and washed once with TBS and twice with PBS (5 min each wash). The nuclei were stained for 20 min at room temperature with DAPI (1 µg/ml in PBS), followed by two washes with PBS. The cover-slips were glued to glass slides with aqueous mounting media containing anti-fading agent (BIOMEDA). All the above steps were carried out at room temperature.
For co-localization experiments of p12 and MA or CA proteins, the immunofluorescence procedure was performed as described above with the following modifications: the blocking solution was made of 3% bovine serum albumin (BSA) in PBS and the supernatant of the 9E10 hybridoma was diluted 1∶6 in PBS. Goat polyclonal anti-MA antiserum (American National Cancer Institute, product no. 78S-282) or goat polyclonal anti-CA antiserum (American National Cancer Institute, product no. 81S-263), were used at a 1∶1000 dilution in PBS containing 3% BSA. These antisera were raised against the MA or the CA proteins of the Rauscher MLV, but cross-react with the MA and CA proteins of the Moloney MLV, respectively. Secondary antibodies included the FITC-conjugated F(ab')2 fragment donkey anti-mouse IgG (H+L) (Jackson Immunoresearch Laboratories, product no. 715-096-150), or Red-X- conjugated F(ab')2 fragment donkey anti-goat IgG (H+L) (Jackson Immunoresearch Laboratories, product no. 705-296-147), both antibodies were used at a 1∶200 dilution in PBS containing 3% BSA.
BX50 microscope (Olympus), LSM 510 META confocal microscope (Zeiss) or spinning disk confocal (Yokogawa CSU-22 Confocal Head) microscope (Axiovert 200 M, Carl Zeiss MicroImaging) were used in this study where indicated. Quantification of the overlap between p12 and chromatin signals was achieved through the following procedure: Cells, infected with the 1xMycR virus were processed for immunofluorescence analysis as described above. Then, the entire cell volume was imaged by confocal microscopy and the picture was deconvolved with the Nearest Neighbors deconvolution algorithm of SlideBook. Subsequently, three dimensional acquisitions were projected on a two dimensional plane. After this, the specific signals of the p12-based and DAPI-based staining were identified through intensity based segmentation, the total signal intensity was calculated for each signal and the percentage of overlapping signal was deduced by subtraction of the DAPI region of interest (ROI) from the p12 ROI. Approximately 400 dots of p12 signal per image were analyzed. All the above steps were performed employing the SlideBook software (Intelligent Imaging Innovations).
Cells were grown on 20×20 mm cover-slips in 6-well plates to 30% confluency and were infected with 1 ml media, containing equal amounts of 1xMycR or wt viruses, normalized by RT activity using the exogenous RT assay [57]. Unless otherwise indicated, all the following steps were performed at room temperature. At the indicated time postinfection, slides were washed with PBS, fixed with 4% paraformaldehyde in PBS for 10 min, permeabilized with 0.1% Triton in PBS for 10 min, washed for 10 min in 0.1 M Tris pH 7.4, incubated for 20 min in 20% glycerol in PBS, freeze-thawed three times in liquid nitrogen, washed once with PBS and once with 0.1% Triton in PBS, blocked for 30 min with normal goat serum that was diluted 1∶10 with PBS, incubated for 60 min with mouse anti-Myc monoclonal antibody (supernatant of hybridoma 9E10, diluted 1∶6 in TBS), and washed once with TBS for 5 min and twice for 5 min in PBS. Slides were then incubated for 60 min with a Cy-3-conjugated goat anti-mouse antibody (Jackson Immunoresearch Laboratories, product no. 115-166-072, diluted 1∶500) and washed once with TBS and twice with PBS (each wash for 5 min). The slides were re-fixed in 4% paraformaldehyde in PBS for one min, rinsed with PBS and incubated in 70% ethanol overnight. To detect the viral DNA in situ, DNA FISH was performed as described previously [58] with the following modifications: The following day, the cover-slips were dried and glued to glass slides with the cells facing up. The slides were then immersed for 10 min in 0.1M HCl, 10 min in 0.5% Triton X-100 in PBS at 37°C and washed three times in PBS. Slides were dehydrated by a series of ethanol washes (70%, 90% and 100%; 5 min each), and incubated at 37°C for 1 hour, followed by denaturation in 70% (v/v) deionized formamide (F9037, Sigma) in 2xSSC, at 75°C, for 5 min. The slides were then briefly washed with 70% ethanol and dehydrated once again in a series of ice-cold ethanol washes (70%, 90% and 100%, 5 min each), air dried and warmed to 37°C. Each slide was spotted with 10 µl of biotin-labeled probe in hybridization solution (see below), sealed with glass cover-slips and rubber cement and incubated in a moist chamber overnight at 37°C. Following hybridization, the sealing was gently removed and the slides were washed three times with 50% (v/v) formamide in 2xSSC (prewarmed to 42°C, 5 min each wash), followed by three washes with 0.1XSSC (prewarmed to 60°C, 5 min each wash). The slides were then incubated with blocking solution (3% BSA in 4xSSC, 30–60 min at 37°C). All stages following blocking were carried out in the dark. The hybridized biotin-labeled probe was detected with FITC-conjugated avidin (A-2011, Vector Laboratories,1∶400 dilution), which was incubated with the slides for 30 min at 37°C in 1% BSA/4xSSC and 0.1% Tween 20. Slides were then washed three times with 4xSSC and 0.1% Tween 20 (prewarmed to 42°C, 5 min each wash), and covered with antifade (VECTASHIELD, Vector Laboratories) containing DAPI (200 ng/ml) under glass cover-slips.
The biotin-labeled probe was prepared in a nick-translation reaction (100 µl, 2 h at 16°C), containing nick-translation buffer (50 mM Tris-HCl pH 7.8, 5 mM MgCl2,50 ng/ml BSA), plasmid DNA template (pNCS, 2 µg), dATP, dGTP, dCTP (50 nM of each nucleotide, Sigma), and biotin-11-dUTP (50 nM, Roche), β-mercaptoethanol (10 mM), DNaseI (30 ng/ml, freshly diluted), Klenow polymerase (20 U, New England Biolabs). A sample (8 µl) of this reaction was separated in a 2% agarose gel to verify the generation of a smear, made of approximately 150–500 bp-long DNA products. The rest of the reaction was kept frozen at −20°C until used. On the day of the hybridization the probe was ethanol precipitated with 10 µg of salmon sperm DNA, resuspended in 50 µl of 100% deionized formamide (Sigma, F9037), thoroughly mixed with 50 µl of 20% dextran sulfate in 2xSSC, denatured (75°C, 5 min), and immediately applied to the denatured slides (10 µl/slide).
For Co-IP of the viral genomic DNA, supernatants of sub-confluent NIH3T3 cultures, chronically infected with the wt or the 1xMycR virus, were harvested, diluted 1∶1 with growth medium and complemented with polybrene (8 µg/ml final concentration). 2 ml of the virus-containing media were used to infect 5×106 NIH3T3 cells in 10 cm plates for 2 h, after which the media were replaced with fresh growth media, followed by additional 5 h incubation. For each IP, 2 plates were trypsinized and the cells were washed once with PBS. Cell pellets were resuspended in 1 ml of IP lysis buffer [50 mM Tris pH 7.5, 150 mM NaCl, 0.5% NP-40, 10 mM MgCl2, 1x protease inhibitor cocktail (Roche, product no. 1183614500)] and incubated for 30 min at 4°C with constant agitation. Lysed samples were centrifuged at 20,800 g for 10 min. 5 µl sample of each of the cleared lysates was diluted 1∶10, and triplicates (5 µl each) were analyzed by qPCR. The remaining lysates were then incubated for 30 minutes at 4°C with 50 µl of magnetic polystyrene Dynabeads Protein G (Invitrogen, product no. 100.03D), pre-conjugated to specific antibodies [the magnetic beads were pre-incubated with 200 µl of supernatant of the anti-Myc hybridoma (9E10), or with 0.2 µl (in 200 µl PBS) of anti-Flag monoclonal antibody (Sigma, F1804), for 10 minutes at room temperature with a constant agitation and washed once with PBS]. The samples were placed on a magnet and the beads-free lysates were discarded. The magnetic beads were washed once with IP washing buffer (50 mM Tris pH 7.5, 150 mM NaCl, 0.1% NP-40, 10 mM MgCl2), twice with PBS, and then transferred into a new tube. The beads were resuspended in 50 µl TE pH 7.4 (10 mM Tris-Cl pH 7.4, 1 mM EDTA) and 5 µl of the bead slurry were analyzed by PCR with MLV-specific primers (forward primer 5′CCCAGGTTAAGATCAAGG3′, and reverse primer 5′CTTGGCCAAATTGGTGGG3′). For qPCR, triplicates (5 µl each) of the bead slurry were analyzed; real-time PCR reactions were performed with MLV-specific primers (forward primer 5′-AGCCCTTTGTACACCCTAAGC-3′ and reverse primer 5′-GAGGTTCAAGGGGGAGAGAC-3′) and Fast SYBR Green Master Mix (Applied Biosystems, product no. 4385612), and analyzed with StepOnePlus Real-Time PCR System (Applied Biosystems). Standard curves were used to determine the absolute DNA quantity in the samples. To calculate the relative efficiency of the immunoprecipitation of the viral genomic DNA in the different experimental settings, the levels of the viral genomic DNA in the cell extracts and in the IP pellets were quantified by qPCR, and the background signal obtained in the ‘mock-infected’ sample was subtracted. The level of the immunoprecipitated viral genomic DNA (IP sample) was divided by the level of this DNA in the cell lysate (input sample), to give normalized IP value. The normalized value for the genome of the 1xMycR virus that was immunoprecipitated with anti-Myc antibodies was set as 100% and was compared to the normalized values obtained for the genome of the 1xMycR virus that was immunoprecipitated with anti-Flag antibodies, and for the genome of the wt virus that was immunoprecipitated with anti-Myc antibodies. The average value of this comparison, obtained from three independent Co-IP experiments, gave the ‘Relative IP Efficiency’ index.
For Co-IP of CA, infections and lysis of the infected cells were carried out as described above. To reduce non-specific binding to the agarose beads that were used for this procedure, the cell lysates were first incubated for 30 minutes at 4°C with a mixture of 7.5 µl of protein A-agarose beads (Roche, product no. 11 134 515 001) and 7.5 µl of protein G-agarose beads (Roche, product no. 11 243 233 001), Then, the lysate-bead slurries were centrifuged at 20,800 g for 2 minutes, the beads were discarded and the cleared supernatants were incubated overnight at 4°C with 15 µl protein A and 15 µl G beads, pre-conjugated to specific antibodies [beads were pre-incubated with 0.5 ml of supernatant of the anti-Myc hybridoma (9E10) or with 0.5 µl (in 0.5 ml of PBS) of anti-Flag monoclonal antibody (Sigma, F1804)]. Samples were centrifuged at 10,000 rpm for 2 min at 4°C and the pelleted beads were washed once with IP washing buffer, twice with PBS, and directly boiled in 2x Sample Buffer. The CA protein was detected by Western blot, using goat anti-MLV CA polyclonal antibody (National Cancer Institute, product no. 81S-263) and a secondary donkey anti-goat HRP-conjugated antibody (Jackson Immunoresearch Laboratories, product no. 705-035-147). For Co-IP of CA from extracellular virions, supernatants of sub-confluent NIH3T3 cultures, chronically infected with the wt or the 1xMycR virus, were harvested and virions were purified from 2 ml of undiluted medium by ultracentrifugation (107,000 g for 2 h) through 25% sucrose cushions. Lysis of the virion pellets and CA Co-IP were performed as described above for the infected cells.
The p12 domain of the Gag polyprotein of the Moloney MLV (SWISS-PROT: P03332) was aligned with 12 homologous sequences that were retrieved from the SWISS-PROT database, using a PSI-BLAST search. The sequences were then multiple aligned (MSA) using the CLUSTALW program, and a phylogenetic tree consistent with the MSA was constructed. Calculation of the conservation scores for each residue was carried out by the Rate4Site algorithm [59], which is based on the Maximum Likelihood method. All these stages were curried out automatically by the ConSeq server (http://conseq.tau.ac.il/), as described in [60].
|
10.1371/journal.pcbi.1004033 | Computer Simulations Suggest a Key Role of Membranous Nanodomains in Biliary Lipid Secretion | The bile fluid contains various lipids that are secreted at the canalicular membrane of hepatocytes. As the secretion mechanism is still a matter of debate and a direct experimental observation of the secretion process is not possible so far, we used a mathematical model to simulate the extraction of the major bile lipids cholesterol, phosphatidylcholine and sphingomyelin from the outer leaflet of the canalicular membrane. Lipid diffusion was modeled as random movement on a triangular lattice governed by next-neighbor interaction energies. Phase separation in liquid-ordered and liquid-disordered domains was modeled by assigning two alternative ordering states to each lipid species and minimization of next-neighbor ordering energies. Parameterization of the model was performed such that experimentally determined diffusion rates and phases in ternary lipid mixtures of model membranes were correctly recapitulated. The model describes the spontaneous formation of nanodomains in the external leaflet of the canalicular membrane in a time window between 0.1 ms to 10 ms at varying lipid proportions. The extraction of lipid patches from the bile salt soluble nanodomain into the bile reproduced observed biliary phospholipid compositions for a physiologi-cal membrane composition. Comparing the outcome of model simulations with available experi-mental observations clearly favors the extraction of tiny membrane patches composed of about 100–400 lipids as the likely mechanism of biliary lipid secretion.
| Formation of the bile is one of the central functions of the liver. The bile fluid aids in the digestion of edible fats and removal of drugs and toxins from the body. The bile fluid is mainly composed of bile salts (BS), phosphatidylcholine (PC) and cholesterol (CH) in a fairly fixed proportion that prevents liver impairment by gallstone formation or cholestasis. During bile formation, BS are actively pumped out of the hepatocyte into the extracellular space where they extract PC and CH from the canalicular membrane. This extraction process bears the risk for the canalicular membrane to be destructed. Hence, only a certain fraction of the membrane should be accessible to the solubilizing activity of BS. We have developed a mathematical model that describes the temporal formation of CH-enriched ordered and PC-enriched disordered nanodomains in the canalicular membrane. Model simulations reveal that the disordered nanodomains exhibit a composition of PC and CH similar to that also found in the bile. From this finding and the good concordance of model simulations with experimental data we conclude that PC and CH are mainly secreted into the bile from the disordered nanodomain. Our work adds a new layer of physiological importance to the spontaneous formation of lipid domains in biological membranes.
| One central function of the liver is the production of the bile which is indispensable for the efficient digestion of dietary lipids, elimination of hydrophobic xenobiotics and removal of cholesterol from the body. The bile is formed in the biliary canaliculi, i.e. the extracellular space that faces the canalicular membrane of hepatocytes. About 80% of the bile content is bile salts (BS), while the other components are phospholipids (≈ 15%) and cholesterol (≈ 5%). The secretion rate and composition of the bile are important factors in metabolic regulation, malfunctioning leading to hepatocyte damage, hypercholesterolemia, advanced levels of high-density lipoproteins or formation of atherosclerotic plaques [1].
Different models have been proposed to account for the canalicular secretion of hepatocyte-born lipids into the bile [2–6]. One possible mechanism of lipid secretion consists in the BS driven extraction of single lipids into the lumen of the canaliculus followed by assembly of these lipids and BS in mixed-micelles (‘single-lipid extraction’). An alternative model developed on the basis of ultrastructural investigations suggests that lipid vesicles highly enriched in phosphatidylcholine are formed via BS facilitated exo-vesiculation of microdomains present in the exoplasmic hemi-leaflet of the canalicular membrane [7, 8]. These BS extractable microdomains appear to coexist with sphingolipid-enriched microdomains which cannot be solubilized by BS and thus represent potential localized target areas for the clustering of proteins involved in bile formation as, for example, ABC transporters, aquaporins and P-type ATPases [9, 10]. One may hypothesize a mechanism of lipid secretion where both BS extractable and BS non-extractable microdomains of different lipid composition coexist in the exoplasmic leaflet of the canalicular membrane. The first ones represent membrane patches that can be easily intercalated by BS, protruded into the canalicular space and finally be released, the latter ones containing most of the proteins required to enrich the outer leaflet with specific lipids like phosphatidylcholine and conferring stability of the membrane against detergents. In both models the solubilization of membrane lipids is mediated by BS that are actively pumped into the canalicular lumen. They rely on the coexistence of BS soluble and BS insoluble membrane areas as the canalicular membrane has to meet two opposing conditions: (i) Maintenance of integrity in the presence of high concentrations of BS which are powerful detergents and (ii) at the same time allowing the secretion of membrane lipids.
The formation of microdomains goes along with the segregation into liquid ordered and liquid disordered phases. The formation of different lipid phases is largely determined by the saturation profile of the fatty acid tails of the phospholipids and of sphingomyelin [11]. The tight interaction between cholesterol and sphingomyelin side chains usually representing long and saturated fatty acids allows for the formation of liquid ordered phases [12, 13], whereas the high content of unsaturated fatty acids of phospholipids usually leads to a lower interaction between these membrane lipids and thus entails a liquid-disordered fraction of the membrane. Phosphatidylcholine species found in the rat liver canalicular membrane and bile have a fairly hydrophilic fatty acid composition with palmitate (16:0) in sn1-position and an unsaturated fatty acid (mostly 18:1, 18:2, 20:4) in the sn2-position [14].
Most findings on microdomains have been obtained in in vitro experiments with lipid mixtures [15, 16] and giant unilamellar vesicles [17]. Depending on the number, type and relative abundance of lipids used to constitute these artificial membrane-like systems, one may observe tiny unstable nanodomains (≈ 10 nm) [18–20] or much larger (> 200 nm) and stable microdomains with life-times of several seconds containing up to 100,000 lipids. The problem is that microdomains of this size and life-time have never been observed in the membranes of living cells. Instead, the existence of much smaller (≈ 10 nm) and very short-lived (< 0.1 ms) microdomains has been inferred from donor-quenching FRET analysis [21, 22], atomic-force microscopy [23] or deuterium-based nuclear magnetic resonance [24]. Possible reasons for this remarkable instability of microdomains in membranes of living cells are thermodynamic variations and mechanical perturbations exerted by cell-cell and cell-matrix interactions. In order to distinguish these tiny and very unstable microdomains from their counterparts present in model membranes we will refer to them later on as nanodomains [11].
As the processes of nanodomain formation and lipid transfer from the external leaflet of the canalicular membrane into the lumen of the bile canaliculus cannot be directly monitored by experimental means we have developed a mathematical model to simulate the formation of nanodomains in the canalicular membrane and the extraction of the major bile lipids cholesterol (CH), phosphatidylcholine (PC) and sphingomyelin (SM) from BS soluble nanodomains. The model enables dynamic simulations of the transit from an initially random to a highly structured lipid distribution on a time scale of several milliseconds. With the model we address the problems of membrane self-organization, biliary phospholipid secretion and composition as well as membrane integrity.
We simulated the lipid dynamics by using a Potts model approach [25]. The membrane is represented by a triangular lattice with periodic boundary conditions. Each lattice site is occupied by exactly one lipid. The model comprises the membrane lipids CH, PC and SM which represent the prevailing membrane lipids in the canalicular membrane of hepatocytes. Note that the model variable PC actually represents also other phospholipids of the exoplasmic leaflet (e.g. ≈ 11% phosphatidylethanolamine) as well as different fatty acid tail compositions. This is a necessary simplification of the model as experimental information on the impact of glycerophospholipids other than PC on domain formation and lipid diffusion is not available.
We introduced next-neighbor interaction energies w for each pair of membrane lipids. Furthermore, similar as in [26] we assumed that each lipid may be either in a low or high ordered internal state. In the low-ordered state the bulky conformation of the fatty acid chains allows high flexibility and thus rapid movement while in the high-ordered state the fatty acid chains are arranged in a way that the hydrophobic interactions with adjacent lipids are strong and thus restricting lipid mobility. We model the property of lipids to cooperatively synchronize their ordering states by assigning a coupling strength j ( X i σ , X k σ ) between the ordering states σ of the lipid species X resident in neighboring lattice sites i and k. The ordering states ho and lo are represented by σ = −1 and σ = +1 respectively. The coupling j ( X i σ , X k σ ) of lipid pairs contributes to the total ordering energy of the membrane with a positive sign if the two neighboring lipids possess identical ordering states and with negative sign if the ordering states are different:
J
=
−
∑
i
=
1
N
∑
k
∈
n
(
i
)
j
(
X
i
σ
,
X
k
σ
)
σ
i
σ
k
.
The first sum runs over all lattice sites N and the second over the 6 neighbors of lattice site i.
Self-organization of lipid domains comprising lipids that are predominantly present in the lo or ho state is achieved by minimizing this total ordering energy.
Our approach to enforce a phase separation of lipids is similar to the well-known Ising model of ferromagnetism in statistical mechanics describing the formation of phases with different atomic spin states (–1, +1) across a 2-dimensional lattice [27]. Furthermore, we introduce next-neighbor interaction energies for each pair of membrane lipids and describe the dynamics of membrane lipids as a process driven by minimization of the total interaction energy of the lattice. As the lipids may occur in two different ordering states, we assign different interaction energies, who(Xi, Xk) and wlo(Xi, Xk), to adjacent lipids of species X at lattice site i and k depending on whether these two lipids are both in the ho state or lo state, respectively. The total interaction energy of the membrane is
W
=
∑
i
=
1
N
∑
k
∈
n
(
i
)
w
σ
(
X
i
,
X
k
)
.
In cases where the ordering states σ at sites i and k are different we assign to this lipid pair the mean of the two interaction energies. Movement of lipids on the lattice is restricted to their pair-wise interchange of next neighbors. The dynamics of membrane lipids and the distribution of their mobility states are governed by the minimization of the total energy E which is a linear combination of the ordering energy J and the interaction energy W:
E
=
W
+
γ
⋅
J
=
∑
i
=
1
N
∑
k
∈
n
(
i
)
(
w
σ
(
X
i
,
X
k
)
+
γ
⋅
j
(
X
i
σ
,
X
k
σ
)
σ
i
σ
k
)
.
The scaling factor γ relates the ordering energy J to the interaction energy W.
The values of wσ(Xi, Xk) and j ( X i σ , X k σ ) represent Gibb’s free energies which arise from a multitude of electrostatic, Van der Waals and hydrophobic interactions between head groups and fatty acid tails of neighboring membrane lipids [28]. In our thermodynamic-based approach the behavior of the lattice in the equilibrium state is fully determined by the changes of the free energy associated with either switching of neighbored lipids or changing the ordering state of lipids.
For the stochastic simulations of lipid movement we applied the Gillespie algorithm [29]. The core of this algorithm consists in assigning to each possible elementary process pij of lipid switch between i and j, a rate r(i, j) that depends upon the local interaction energies at positions i and j where the process pij is executed: r(i, j) = exp(β(wi +wj)). Here, wi = Σk∈n(i) wσ(Xi, Xk) is the interaction energy of a lipid of species Xi at site i with all its neighbors with species Xk and β = 1/kBT the inverse temperature. The probability P(Δt) that during the time span Δt no elementary process occurs is related to the elementary rates by P(Δt) = exp(–rtotΔt), where rtot = ΣPij r(i, j) is the total rate, defined as the sum of the elementary rates. This relation is used to randomly generate elementary time steps, Δt = -ln η/rtot, where η are equally distributed random numbers between 0 and 1. After having randomly chosen the time step for the next elementary process to occur, the elementary process to be executed has to be specified. This is done by randomly choosing an elementary process, i.e. lipid switch, with its corresponding probability.
So far we neglected the process of changes in the ordering state. From the molecular-dynamics point of view, a change in the conformation of the fatty acid tail should occur much more frequently compared with a switch of neighbored lipids. As a consequence, the elementary process “change the ordering state of a selected lipid” occurs much more frequently than the elementary process “switch adjacent lipids”. Hence, it is reasonable to assume that a large number of changes in the ordering states of lipids will occur between two subsequent lipid switches so that at any time point the conformational energy J becomes minimal.
The rate with which a membrane lipid flips between two alternative ordering states is not known. Considering that such a flip requires only a change in the conformation of the fatty acid tails it is reasonable to assume that flips of the ordering state occur with much higher rates than changes in the spatial position of the membrane lipid.
We thus refrained from executing explicitly the elementary process “change the ordering state of a selected lipid”. Instead, we make the steady-state (or equivalently: partial fast-equilibrium) assumption that the ordering energy is minimal and thus the ordering states of the lipids are in equilibrium at any time point. This also alleviates us from knowing the exact value of the scaling factor γ, as it now does not occur in the algorithm itself.
Adopting the basic concept of multi-scale stochastic simulation of systems comprising a fast-equilibrium subsystem [30, 31] we split the simulation into two alternating steps: (1) A dynamic simulation governed by the interaction energy W and carried out over a critical time which is determined by the condition that each lipid has to change its position on the lattice NW times on average. During this dynamic simulation, the ordering states of the lattice sites are not changed, i.e. the ordering state is a fixed property of the lattice site and not of the lipid just occupying the site. The dynamic simulation step is followed by (2) the minimization of the total ordering energy J. This minimization step is carried out by the Metropolis algorithm [32] whereby the ordering states of all lattice sites are NJ times updated. A change of the ordering state occurs with probability
P ( lo ↔ ho ) = { 1 exp ( − β Δ J ) if Δ J ≤ 0 if Δ J > 0
depending on the difference ΔJ between new and old ordering energy. The updated ordering states of the lipids are assigned to the harboring lattice sites and the simulation is continued with step (1) as depicted in Fig. 1A.
The whole simulation is either stopped at a fixed time point (this termination of the simulation was applied in the simulations of bile formation) or if the statistical properties of the lattice defined through frequency of lipid-lipid contacts, the relative share of lipids in lo states and ho states and the numerical value of the diffusion coefficient do not change over a sufficiently long time interval (this termination of the simulation was applied in the parameterization of the model based on experimental data with ternary lipid mixtures).
To ensure independence of the simulation results from the combination of the two linked optimization algorithms (see flowchart in Fig. 1A) the number of spin updates NJ has to be high enough to minimize J. Likewise the number of lipid switches NW should be sufficiently low to ensure that the spin updates occur frequently enough. On the other hand it is desirable to keep NJ low and NW high to minimize the computational effort. To find optimal values satisfying both requirements the control parameters NJ and NW were varied and the dependence of the statistical properties of the model membranes was monitored. Fig. 1B shows an example demonstrating how the statistical properties of the simulated model configurations are influenced by the control parameters NJ and NW. According to these results, we put NJ = 100 and NW = 10 in the stochastic simulations.
The next step was to determine the interaction energies and the spin energies used in the model. With the three membrane lipids CH, PC and SM the symmetric matrices wlo and who of the interaction energies are 3×3 matrices comprising six independent unknown parameters wσ(CH, CH), wσ(CH, PC), wσ(CH, SM), wσ(PC, PC), wσ(PC, SM), and wσ(SM, SM), with σ = lo, ho state.
The matrix j of the ordering energies is a 6×6 matrix comprising 36 elements. The entries in the ordering matrix j describe the ordering energy of a given lipid (CH in the first column, PC in the second column, SM in the third column) with a neighboring lipid being in a given ordering state σ. The ordering state of the lipid itself is not of importance since in the used metropolis algorithm only energy differences are of importance, i.e. the calculation of ΔJ leads to the same result if one puts j ( X i lo , X k lo ) = j ( X i lo , X k ho ) and j ( X i ho , X k lo ) = j ( X i ho , X k ho ). Therefore the dimension of the matrix is reduced to 3×6. The difference in the ordering energies for a given lipid with a neighboring lipid in either phase determines the tendency of the lipid to adopt either ordering state, with the lower value representing the favored and the higher value representing the disfavored ordering state. These pairs of values are block by block symmetric in the matrix which further reduced the free parameters from 18 to 6.
Thus, in total our model comprises 2·6+6 = 18 parameters with unknown numerical values.
We calibrated the interaction between the different lipids in different ordering states such that the diffusion coefficients would match those from experiment for the reported phases.
Our model allows to track the stochastic movement of individual lipids along with their ordering state and thus to determine mean squared displacements (MSDs) for the different phases. MSDs were calculated by first letting the simulation run until the resulting membrane configuration had reached an equilibrium state. At this point, we tagged the position of all N phospholipids i0 and let the simulation continue for a certain number of steps z. Next we calculated the individual displacement after the z steps for each lipid:
Δ
x
(
z
,
X
i
,
σ
i
)
=
|
i
z
−
i
0
|
For the calculation of the displacements of the different lipid species X in the different phases only those random walks were taken into account for which the tracked lipid had not passed a phase border, i.e. had not changed its ordering state σi.
Given the displacements Δx and choosing an arbitrary timescale t corresponding to the z steps the simulated diffusion coefficient D m , σ sim for the phase σ of the m-th model membrane can be derived from the MSDs with the corresponding ordering state in the respective lipid composition:
〈
Δ
x
(
t
)
2
〉
=
4
D
m
,
σ
sim
t
.
The MSDs are only determined up to an overall scaling factor α:
α
=
∑
m
D
m
,
σ
exp
D
m
,
σ
sim
∑
m
(
D
m
,
σ
exp
)
2
.
Here the index m designates the different lipid compositions of the GUVs, D m , σ exp denotes the measured and D m , σ sim the simulated diffusion coefficient for lipid composition m.
The numerical values of the unknown elements of the matrices wlo and who were estimated by minimizing the difference ε between model-based lateral lipid diffusion rates and experimental values determined in giant unilamellar vesicles (GUVs) containing CH, PC and SM in 25 different proportions [33]:
ε = ∑ m = 1 25 ( D m , σ exp − α D m , σ sim ) 2 → minimum .
calculated for a lattice having the same lipid composition m as the GUVs. Simulated diffusion coefficients that differ from the experimental ones by a global constant factor can be transformed by rescaling of α or alternatively by a rescaling of the time t. Therefor α relates the time scale used in the model simulations to the time scale of the in vitro experiments. The same value of the scaling factor α was used for fitting the unknown elements of the two matrices wlo and who.
To solve this minimization problem, the numerical values of the 12 unknown parameters were varied on a discrete hypercube under the additional constraint of reproducing known interactions between the different lipid species. The condition for mixing of lipid species X and Y can be formulated as 2wσ(X, Y) − wσ(X, X) − wσ(Y, Y) < 0 while the condition for de-mixing is 2wσ(X, Y) − wσ(X, X) − wσ(Y, Y) > 0 [34]. The only constraints applied were that the lipid species CH and SM would mix in the ho state [35]. We determined the points of this hypercube that fulfilled the condition and where ε attained its minimal value for the ho or the lo state respectively. Around these points we again varied the unknown parameters on a finer hypercube and determined ε. The procedure was repeated with successively decreasing sizes of hypercubes until no further significant reduction of ε was possible. This corresponds to a discretized version of a downhill simplex method. The calibration yielded the following numerical values for the lipid—lipid interactions
w ho = ( w ho ( CH,CH ) w ho ( CH,PC ) w ho ( CH,SM ) w ho ( PC,PC ) w ho ( PC,SM ) w ho ( SM,SM ) ) = ( − 0 .70 + 0 .55 − 1 .70 − 0 .60 + 0 .45 − 0 .80 )
in the high ordered state and
w lo = ( w lo ( CH,CH ) w lo ( CH,PC ) w lo ( CH,SM ) w lo ( PC,PC ) w lo ( PC,SM ) w lo ( SM,SM ) ) = ( − 0 .55 − 0 .23 + 0. 43 − 0 .22 − 0 .23 − 0 .55 )
in the low ordered state.
For the interpretation of the numerical values of the interaction matrices we apply them to the mixing and de-mixing properties of the previous section and compare the results to known membrane properties. In the ho state the interaction matrix features a strong tendency for CH and SM to mix and a strong tendency for PC to de-mix from CH and from SM. This is in agreement with works from Silvius [35], van Duyl [36] and Frazier [37].
In the lo state PC has only a weak tendency to de-mix from CH and from SM, which is in agreement with work from Silvius [35] and Tsamaloukas [38]. Contrary to the ho state CH and SM have a tendency to de-mix in the lo state. For this pairing no reliable data could be found since the two species occur in the ho state rather than in the lo state.
The unknown parameters of the ordering matrix j were chosen such that the occurrence of monophasic and biphasic lipid distributions matched those observed in the 25 different variants of GUVs. To this end we defined lipid distributions with more than 90% of all lipids resident in the ho state as monophasic liquid-ordered (Lo), with more than 90% of all lipids resident in the lo state as monophasic liquid-disordered (Ld) and with more than 10% of all lipids resident in both ordering states as biphasic. For the minimization of J the lattice was initialized with all lipids in lo state. A searching of the parameter space yielded parameter values
j = ( j ( CH lo ,CH ) j ( CH lo ,PC ) j ( CH lo ,SM ) j ( CH ho ,CH ) j ( CH ho ,PC ) j ( CH ho ,SM ) j ( PC lo ,CH ) j ( PC lo ,PC ) j ( PC lo ,SM ) j ( PC ho ,CH ) j ( PC ho ,PC ) j ( PC ho ,SM ) j ( SM lo ,CH ) j ( SM lo ,PC ) j ( SM lo ,SM ) j ( SM ho ,CH ) j ( SM ho ,PC ) j ( SM ho ,SM ) ) = ( 0.50 0.90 0.50 1.00 0.50 1.90 0.90 0.90 0.90 0.50 0.50 0.50 0.50 0.90 0.50 1.90 0.50 0.55 )
that allowed to match 21 phases of the 25 model membranes.
The values thus obtained are by no means unique but this is not to be expected since the data described by them are only semi-quantative and the relative size of the ordering pairs for a given lipid to adopt either state is more important than the precise values.
With these values the calculated diffusion coefficients and the occurrence of monophasic and biphasic lipid distributions of our model simulations were in agreement with experimental data obtained with GUVs [33] (see Fig. 2).
The calibration of the model with measured diffusion coefficients allows the definition of an absolute time scale and thus offers the possibility to use the model for real-time dynamic simulations of domain formation. We used the parameterized model to compute lipid distributions for the whole possible range of lipid compositions (Fig. 2B).
We used the calibrated model to simulate the self-organization of nanodomains and the release of lipids from the outer leaflet of the canalicular hepatocyte membrane into the bile canaliculus. We presupposed a situation where the amount and composition of lipids in the outer leaflet is on the average kept constant (quasi steady state), i.e. the transport rate of lipids to the canalicular membrane equals the net transport rate from the inner to the outer leaflet and the release rate into the canalicular lumen.
As the life-time of nanodomain structures, i.e. the simulation time during which an initially random distribution of lipids self-organizes into a domain structure that is representative for the outer leaflet of the canalicular membrane, we chose τ = 0.1, 1 and 10 ms to cover a range around the estimated nanodomain life-time of ≈ 1 ms [39].
We tested two alternative mechanisms of lipid secretion: extraction of single lipids or extraction of lipid patches. An extractable lipid patch is defined as a hexagonal piece of membrane that is fully embedded in a Ld nanodomain. With this definition, extraction of single lipids is identical with extraction of patches with size of 1 lipid. We defined the extractable membrane fraction (EMF) as the fraction of the membrane that can be covered by extractable patches. The flow of lipids into the bile is determined by the detachment rate of patches. This rate depends on the concentration and chemical properties of the BS species present in the canalicular lumen and the size and number of patches present in the outer leaflet. Hence, at fixed concentration of BS, the lipid secretion rate (LSR) is up to a constant unknown factor proportional to the number of the extractable patches npatch(r) with radius r multiplied by their lipid content nlipids(r) and divided by the time span τ required for the formation of nanodomains:
LSR
(
r
)
=
n
patch
(
r
)
n
lipids
(
r
)
/
τ
.
The lipid composition of the bile is given by the average lipid composition of all extractable patches.
First, we carried out model simulations with a physiologically normal lipid composition of the canalicular membrane [8] with CH = 37.8%, PC = 46.5% and SM = 15.7%. The simulation was started with a fixed lipid composition but random distribution of lipids across the lattice with their ordering states calculated for the given random lipid distribution.
Model simulations persistently resulted in a bi-phasic quasi-stationary lipid distribution comprising Ld nanodomains rich in PC and Lo nanodomains enriched in CH and SM. Typical lipid patterns obtained for three different life-times are depicted in Fig. 3. With increasing life-time, the nanodomains tend to merge to form larger domains. Accordingly, the maximal size of patches that can be extracted from the Ld nanodomains increases as well. The predicted lipid composition of the patches and thus of the bile micelles was ≈ 13.5 CH, ≈ 85.5% PC and ≈ 1% SM in good agreement with reported experimental values of ≈ 15% CH, ≈ 85% PC and < 1% SM [40]. The lipid composition of patches was remarkably invariant against variations of the patch sizes and life-times (see Table 1).
As seen in Fig. 3D–F, the calculated LSRs are non-monotone with respect to the patch size. This is due to the fact that the number of lipids that are simultaneously extracted from the leaflet increases with the patch size whereas the number of patches fitting into Ld nanodomains decreases with increasing patch size. The largest LSR was attained for patches containing 37, 169 and 631 lipids for the three different simulation times τ = 0.1, 1 and 10 ms. This has to be compared with the size of sandwich-like micelles which primarily derive from mixed dispersions of egg PC and the BS deoxycholate [41] and quasi-elastic light scattering studies of native bile from the dog [42] consistently comprising about 100–400 lipids. Since this is close to the calculated average patch size at τ = 1 ms and as a life-time of 1 ms is also in good agreement with several measurements on nanodomains we choose τ = 1 ms as the average life-time in further simulations.
The lipid composition of the canalicular membrane has been shown to strongly influence the relative share of membrane lipids in the bile [43]. Thus, we performed simulations where the relative fraction of either CH or PC was varied (see Fig. 4).
A decrease in the relative fraction of CH at otherwise constant PC:SM ratio increased the share of lipids in the Lo state and the average size of Ld nanodomains. As a consequence, the average size of membrane patches that can be solubilized from Ld nanodomains and the LSR increased with decreasing CH content of the external leaflet. At a low fraction of CH = 18% the model simulations predict a steep increase of the average patch size to about the 10-fold of the reference state. Such an increase of the BS-solubilizable membrane area should result in a BS-induced rupture of the membrane. Increasing the CH content of the exoplasmic leaflet promoted the transition from the Ld to the Lo phase and thus reduced the LSR. On the other hand, the CH content of the extractable membrane patches became successively larger. The net effect of these two opposing tendencies was an increase of the extractable fraction of CH up to a critical CH content which amounts to about 36% for extraction of patches with sizes of 100–400 lipids and to 56% for the mechanism of single-lipid extraction.
In a further series of simulations we studied the impact of varying concentrations of PC on the lipid flow. With decreasing PC content at constant CH:SM ratio the simulations revealed a non-linear decline of the average size of extractable patches resulting in reduced LSRs of all lipid species. The composition of the patches remained almost constant although there was a slight shift towards a higher share of CH and a lower share of PC.
Intriguingly, the lower limit of the membranous PC content below which the flow of PC into the bile practically ceased depended strongly on the size of membrane patches supposed to carry the lipid flow into the bile. For example, lipid extraction stopped at a lowered membranous PC content of 32% for patches with a size of 631 lipids, at 10% for 169 lipid patches while the extraction of single lipids did not stop until a PC content of 0%. Studies on mice with a homozygous knock-out of the PC transporter mdr2 found a complete abolishment of PC release into the bile [44]. Compared with our model simulation this finding again suggests lipid secretion to precede via extraction of patches rather than of single lipids. Increasing the PC content resulted in an increased EMF and strongly increased LSRs.
In this work we developed a kinetic model of lateral lipid diffusion and phase separation in monolayers composed of the three membrane lipids CH, DOPC and SM assumed to contain saturated fatty acid tails, even though in biological membranes the diversity of phospholipids is higher than in our three component system (e.g. phosphatidylserine, phosphoinositol, phosphatidylethanolamine) and the fatty acid tail composition is much more complex (hybrid, multiple unsaturated, different chain length). However studies by Marsh and Konyakhina show that the phase distribution between PC with one or two unsaturated fatty acid tails is rather similar [45–47] and PC is the dominant phospholipid species [8, 48]. Naturally mathematical modelling relies on justified simplification of real biological structures. In our model the influence of membrane proteins was neglected and DOPC is used in the model as a substitute for a number of different phospholipids (e.g. PE, PS) with the hybrid lipid 1-saturated, 2-unsaturated-phosphatidylcholine being the most abundant phospholipid. The differences between phase diagrams for DOPC/SM/CH and POPC/SM/CH are not necessarily higher than between different experiments of same the same system [49, 50]. In the end justification comes from the agreement between experiments and theory.
Therefore to keep the parameter space small and to allow for the comparison with experiment, the model was calibrated based on data from in vitro experiments with GUVs. We then used the model to study the dynamics of domain formation in the exoplasmic leaflet of the canalicular membrane of hepatocytes. The central goal of these simulations was to explore how the lipid composition of the leaflet influences the relative size of Ld and Lo nanodomains and thus the biliary phospholipid composition and the likelihood to solubilize membrane segments (patches) from the Ld nanodomains.
Importantly, at physiologically realistic lipid composition of the outer leaflet, our simulation consistently predicted the formation of two distinct types of nanodomains differing significantly in their lipid composition and phase behavior (Fig. 3). The Ld nanodomain was strongly enriched in PC while the Lo nanodomain was rich in CH and SM. These model-based findings are in good agreement with experiments clearly indicating a compartmentalization of lipids within the canalicular membrane [51]. Using either Triton X-100 or Lubrol WX as detergents, Slimane et al. [52] extracted two different membrane fractions. The Triton insoluble fraction was highly enriched in sphingolipids and CH [9]. The proteins mediating the trans-membrane lipid transport and the release of BS into the canalicular lumen (BS export pump, multidrug resistance protein 2, multidrug resistance associated protein 2, Abcg5) were found to be predominantly located in the Triton X-100 soluble fraction [51].
Experiments with model membranes devoid of proteins have provided evidence for the spontaneous formation of nano-scale lipid domains [19]. Whether spontaneous lipid de-mixing is also the primary mechanism for domain formation in biological membranes is a matter on ongoing debate (reviewed, for example, in [53]) as lipid—protein interactions may contribute not only to the stabilization of such domains but may even induce their formation. The fact that our lipid-based model of nanodomain formation in the canalicular membrane of hepatocytes indeed recapitulates a number of experimental findings on the size and lipid composition of bile micelles lends support to the view that spontaneous formation of pure Lo and Ld lipid domains without further assistance of membrane proteins is sufficient to enable the extraction of tiny membrane fragments from Ld domains in the presence of solubilizing agents. Whereas Ld domains are disrupted by the process of bile micelle formation and thus have to be permanently recreated, it is likely that the Lo lipid domains once formed are stabilized by the insertion of membrane proteins involved in the active transport for the various bile components and asymmetric distribution of lipids between the internal and external leaflet.
Intriguingly, Ld nanodomains obtained in our simulations contained PC, SM and CH in relative fractions that perfectly matched their relative abundance in the bile. This finding suggests that the three major lipids of the bile are not independently solubilized from the membrane but secreted in a concerted manner just in proportions present in the Ld nanodomains.
In our simulations the size of nanodomains increased with increasing life-times in a sub-linear fashion. Spectroscopic measurements suggest the life-times of nanodomains in biological membranes to be not longer than a few milliseconds. For this time window our simulations predict a maximal size of Ld nanodomains of only a few hundred lipids.
Our lattice model allows calculating the size and geometry of nanodomains which represent important factors determining the rate with which membrane patches can be solubilized and extracted from the external leaflet. These simulations suggest that the LSR has a maximum for critical patch sizes of 100–400 lipids. Remarkably, the predicted optimal size of lipid patches is in good agreement with the size of micelles that are usually extracted from artificial membranes [28, 48]. Of note, even the largest patches that in our simulations were found to be extractable in a time window of a few microseconds were at least one order of magnitude smaller than bile vesicles observed by means of ultra-rapid cryofixation [7]. We suppose that these vesicles may derive from a fusion and rearrangement of smaller nano-micelles as those suggested by our simulations (micelle-to-vesicle transition, [54]). Such a mechanism would also better explain the formation of bilayered vesicles by pinch-off from a monolayer. The fact that some of the larger bile vesicles were found in direct contact with the canalicular membrane does not necessarily imply that they have originated from exocytosis, as suggested in [7]. Rather, they may represent vesicles that after their formation from micelles back-fuse with the canalicular membrane to be endocytosed [55, 56].
Since altered phospholipid composition of the canalicular membrane can result in impaired bile formation and cellular damage we carried model simulations where the CH and PC content of the canalicular membrane was varied over a wide range (Fig. 4). A reduced PC content of the outer leaflet of the canalicular membrane can result from impaired mdr2 activity. mdr2 selectively transports PC to the outer membrane depending on BS concentration. Experimental studies on changes of bile flow and bile lipid composition in mice being homozygous for a disruption of the MDR2 gene, the analog of the human MDR3 [57, 58] revealed an only moderate decline of the PC flow into the bile in the heterozygous animals whereas in the homozygous mice the flow of PC and CH was almost completely abolished (lower than 5% of the normal). In our simulations a PC content of less than 30% completely prevented the extraction of patches with a size of about 600 lipids. In contrast, extraction of single lipids is predicted to continue—albeit with decreasing activity—if the PC content goes to zero. Hence, these simulations lend further support to the existence of a patch-extraction mechanism of membrane lipids. Unfortunately the residual PC content of the canalicular membrane has not been determined in the mdr2 knockout experiments so that a comparison with the predicted threshold value of about 30% PC content is not possible. However, considering that PC is an indispensable phospholipid of the plasma membrane and that besides mdr2 other transporters of PC exist, for example, the relatively unspecific MDR1 encoded P-glycoprotein [59], a residual PC content of 30% is not unlikely.
Finally we examined the dependence of the biliary phospholipid composition on the CH content of the outer canalicular membrane. Experimental data implicate that the heterodimer of the two half-transporters ABCG5 and ABCG8 translocates CH to the outer leaflet of the canalicular membrane [60, 61]. Mice with knockout of both transporters (Abcg5+/g8+) displayed strongly reduced biliary CH excretion [62]. This is reproduced by our simulations. The fraction of biliary CH shows a strong correlation with the CH content of the membrane. On the other hand the solubilizable fraction of the canalicular membrane decreases with increasing CH concentration demonstrating the ordering and stabilizing effect of CH [43]. Sufficient CH is required for stability of the canalicular membrane and protects cells from BS induced damage.
Taken together, our model-based calculations provide further evidence for the emergence of short-lived nanodomains in the external leaflet of the canalicular hepatocyte membrane. The best overall agreement between model simulations and experimental facts is achieved if we assume that the lipid transfer into the bile is mediated by small patches of 100–400 lipids which are extracted from the Ld nanodomains of the external leaflet by BS and which contain the main lipids CH, PC and SM in proportions as also found in the bile. Likely, the primary nano-micelles further maturate to larger lipid vesicles (see Fig. 5).
|
10.1371/journal.ppat.1003648 | Absence of Siglec-H in MCMV Infection Elevates Interferon Alpha Production but Does Not Enhance Viral Clearance | Plasmacytoid dendritic cells (pDCs) express the I-type lectin receptor Siglec-H and produce interferon α (IFNα), a critical anti-viral cytokine during the acute phase of murine cytomegalovirus (MCMV) infection. The ligands and biological functions of Siglec-H still remain incompletely defined in vivo. Thus, we generated a novel bacterial artificial chromosome (BAC)-transgenic “pDCre” mouse which expresses Cre recombinase under the control of the Siglec-H promoter. By crossing these mice with a Rosa26 reporter strain, a representative fraction of Siglec-H+ pDCs is terminally labeled with red fluorescent protein (RFP). Interestingly, systemic MCMV infection of these mice causes the downregulation of Siglec-H surface expression. This decline occurs in a TLR9- and MyD88-dependent manner. To elucidate the functional role of Siglec-H during MCMV infection, we utilized a novel Siglec-H deficient mouse strain. In the absence of Siglec-H, the low infection rate of pDCs with MCMV remained unchanged, and pDC activation was still intact. Strikingly, Siglec-H deficiency induced a significant increase in serum IFNα levels following systemic MCMV infection. Although Siglec-H modulates anti-viral IFNα production, the control of viral replication was unchanged in vivo. The novel mouse models will be valuable to shed further light on pDC biology in future studies.
| Plasmacytoid dendritic cells (pDCs) represent a minor but functionally important subset of dendritic cells. Siglec-H, a surface receptor expressed on these cells, was shown to modulate IFNα production, which in turn could influence anti-viral functions in vivo. A potential role for Siglec-H as a pathogen uptake receptor has also been postulated. Yet, the precise in vivo function of this molecule in viral replication remained unresolved. In this study, we adopt two novel genetic mouse models to investigate Siglec-H properties and ensuing function in vivo during murine cytomegalovirus (MCMV) infection. By using novel reporter mice which harbour permanently labeled Siglec-H+ pDCs, we show that pDCs downregulate Siglec-H upon infection. In an additional experimental system, in which pDCs lack Siglec-H function, we demonstrate that this molecule is not important for the regulation of MCMV pathogenicity. In contrast, in the absence of Siglec-H more IFNα was detectable in the serum. Importantly, this in vivo increase in IFNα production does not influence viral replication. The biological function of Siglec-H downregulation, also in the context of other infections, requires further investigation.
| Dendritic cells (DCs) are a diverse population of professional antigen-presenting cells that exhibit differences in both their developmental origins and their functional properties. They are distributed in different anatomical compartments such as the skin, intestine, lung and lymphoid organs [1], [2] where pathogen access is prevalent. Two prominent murine DC sub-classes exist, conventional DCs (cDCs) and plasmacytoid DCs (pDCs). The latter subset of DCs expresses the I-type lectin receptor Siglec-H. pDCs are implicated in immune tolerance [3]–[8], but are also known to secrete large amounts of type I interferons (IFNs) in response to viral infections [9], [10].
Double stranded DNA viruses like cytomegalovirus (CMV) are sensed by pDCs via the endosomal Toll-like receptor 9 (TLR9) [11] and trigger strong IFNα responses which are critically required for early viral control during MCMV infection [12], [13]. To this end, Zucchini et al. have shown that pDCs are the main early source of intracellular IFNα/β at 30–36 h post MCMV infection [14]. Moreover, Swiecki et al. demonstrated that specific pDC depletion in blood dendritic cell antigen 2 (BDCA2)-diphtheria toxin receptor (DTR) transgenic mice (where ablation of pDCs is mediated by diphtheria toxin) at 36 h post infection (p.i.) resulted in impaired MCMV clearance with unhindered NK cell expansion and function at later timepoints during infection [13]. However, the precise receptors contributing to pDC-mediated anti-viral defence remain incompletely defined.
Amongst the numerous receptors expressed by pDCs, antibody-mediated crosslinking of Siglec-H was shown to negatively influence IFNα production in response to CpG stimulation [15], [16]. This inhibitory regulation of IFNα by Siglec-H was attributed to its association with the ITAM bearing adaptor molecule DAP-12 [16]–[19]. DAP-12 was postulated to recruit inhibitory signalling mediators to dampen TLR-mediated activation. To address this hypothesis, Takagi et al. employed Siglec-HDTR/DTR mice, where an IRES-DTR-EGFP cassette disrupts the Siglec-H open reading frame, harbouring Siglec-H deficient pDCs. HSV-1 infection of these mice induced a strong increase in IFNα levels 6 h p.i. in vivo [20]. Additionally, Siglec-H was shown to mediate endocytosis and cross-presentation of antigens suggesting that it may contribute to capturing and delivering of viruses and other pathogens to endosomal TLRs [18]. However, the ligands of Siglec-H remain uncharacterized. Thus, Siglec-H presents as an interesting immunomodulatory molecule pre-dominantly expressed on pDCs, yet its in vivo function remains incompletely understood.
To investigate this query, we generated a novel BAC-transgenic Siglec-H reporter mouse (pDCre x RFP) and utilized Siglec-H−/− mice as valuable genetic tools for studying Siglec-H and pDC functions in MCMV infection in vivo. Using these novel mouse strains we were able to show downregulation of Siglec-H on pDCs upon MCMV inoculation. Siglec-H is, however, neither required for MCMV infection nor for activation of pDCs. In contrast, Siglec-H deficiency enhances IFNα production without influencing viral clearance.
Since pDCs account for a minor fraction (<0.5%) of immune cells in the spleen [21], we set out to develop a novel transgenic mouse model where Cre recombinase is expressed under the control of the Siglec-H promoter. Transgenic mice were crossed to floxed RFP reporter mice [22], thereby terminally labeling Siglec-H+ cells with RFP expression driven from the ubiquitous Rosa26 promoter (Figure S1). The resulting reporter mice are termed pDCre x RFP from here on.
In order to determine the distribution of the reporter expression in pDCre x RFP mice, we characterized cells from bone marrow (BM) and spleen. Gating on individual cell populations revealed that Siglec H+ pDCs were targeted most efficiently (Figure 1a, c). Notably, aside of pDCs, we also observed RFP expression by a minor fraction of Siglec-H− cells which consisted of B-, T-, NK-, and NK-T cells from both tested organs as well as splenic cDC and CD11cint BM cells (Figure 1a–c). These results suggest that a small fraction of early lymphoid progenitors actively transcribes the Siglec-H locus and is fate-mapped in pDCre x RFP mice. As previously described, the fluorescence intensity of the RFP reporter differs between cell types and differentiation states and is not a reflection of incomplete recombination [22]. We also found RFP expression by Siglec-H− and Siglec-H+ common dendritic cell precursors (CDP) from in vitro BM cultures of pDCre x RFP mice (Figure S2a).
Interestingly, we observed RFP expression only by 23±5% (BM) or 30±8% (spleen) of pDCs (Figure 1a, c). Phenotypically and functionally distinct pDC subsets have been reported in humans and mice [23]–[27]. Thus, we performed a pDC surface marker expression analysis including CCR9, Clec9A, B220, CD4, CD8α, PDCA-1, and Ly-6C. When comparing RFP+ and RFP− BM pDCs from pDCre x RFP mice, we observed minor alterations in the MFIs of CCR9, Clec9A, B220, PDCA-1 and Ly-6C and more RFP+ pDCs expressed CD4 (Figure 1d, top panel). These differences were however not observed in peripheral splenic pDCs with the exception of minor differences in the MFI for CCR9 (Figure 1d, bottom panel). Recently a CD9+ pDC subset with potent IFNα production capacities has been described in the BM of mice [27]. We observed comparable targeting of both CD9+ and CD9− pDCs in the pDCre x RFP mice (Figure S2b). As we did not identify major phenotypic differences in the surface marker expression of RFP+ versus RFP− peripheral pDCs we also wanted to confirm that they possess comparable cytokine production capacities. To this end, we stimulated ex vivo purified splenic RFP+ and RFP− pDCs from pDCre x RFP mice with different multiplicities of infection (MOI) of MCMV in vitro and analyzed their TNFα and IFNα production. As shown in Figure 1e, we did not observe significant differences in the production of the key pro-inflammatory cytokines tested. Hence, given comparable phenotypic and functional characteristics of peripheral RFP+ and RFP− pDCs, we report targeting of a representative fraction of pDCs in the pDCre x RFP mice.
While studying RFP+ and RFP− pDCs during MCMV infection, we surprisingly observed downregulation of Siglec-H expression by all pDCs upon MCMV infection in vivo (Figure 2a). Thus, defining pDCs by Siglec-H surface expression is hampered. To investigate the extent of Siglec-H downregulation on a per cell basis we used our pDCre x RFP model. The advantage of this model is that pDCs are still fate mapped by RFP expression even after Siglec-H downregulation. Thus, we can accurately gate on the targeted pDC fraction to investigate the consequences of MCMV infection on pDCs. To exclude the non-pDCs targeted in the pDCre x RFP mice a default channel for CD19, CD3ε, and NK1.1 was used. Furthermore, also CD11chi cDCs and MHCII− CDP precursors [28], [29] were gated out. Thus, in the final gate only RFP+ pDCs remained. When comparing RFP+ pDCs from mock treated and infected pDCre x RFP mice, we observed Siglec-H downregulation upon infection (Figure 2b) with some cells becoming even Siglec-H−. Therefore, Siglec-H surface expression does not faithfully identify pDCs during an ongoing MCMV infection because of its downregulation in vivo.
As we observed Siglec-H downregulation in vivo, we next sought to study the mechanism. For this purpose we infected Flt3-L differentiated bone marrow dendritic cells (BMDCs) with MCMV-GFP and tracked Siglec-H expression in vitro over time (Figure 3a). CD11c+Siglec-H− DCs express virus encoded GFP starting from 3 h p.i., while both the frequencies of infected DCs and the GFP expression levels progressively increase until 24 h p.i. Interestingly, Siglec-H was downregulated on CD11c+ DCs starting from 6 h p.i. (Figure 3a, upper panel), whereas B220 expression remained unaltered at different timepoints p.i. (Figure 3a, lower panel). Interestingly, Siglec-H downmodulation precedes CD86 upregulation (Figure S3b). We next used sorted pDCs from BMDC cultures and found that they also downregulated Siglec-H upon MCMV infection or CpG-A treatment, underlining that accessory cells are dispensable for this effect (Figure 3b). Since pDCs sense MCMV mainly via TLR9 [11], we assessed whether Siglec-H downregulation requires TLR9- and MyD88-dependent signalling. Indeed, Siglec-H downregulation upon CpG-A or MCMV treatment only occurred in wildtype (wt) but not in TLR9−/− or MyD88−/− pDCs (Figure 3b). To test whether the absence of Siglec-H downregulation in TLR9−/− and MyD88−/− pDCs was a consequence of impaired IFNα production, we stimulated IFNAR−/− pDCs. Although we found differences in Siglec-H expression between mock treated wt and IFNAR−/− pDCs (Figure 3c), pDCs still clearly downregulated Siglec-H in the absence of type I IFN signalling upon MCMV infection or CpG-A treatment (Figure 3c, d). Notably, we also observed Siglec-H downregulation upon treatment with the TLR7 agonist R837 (data not shown). Thus, downregulation of Siglec-H is linked to early intracellular activation signals which are dependent on TLR9 and MyD88 signalling with only minor contribution of type I IFN signalling.
Siglec-H has been postulated as a pathogen uptake receptor [16] and downmodulation of Siglec-H could prevent further infection of pDCs. Thus, we wanted to test for the requirement of Siglec-H for pDC infection/activation. To address this aim, we compared differences in MCMV-GFP infection between sorted Siglec-H−/− and wt pDCs. We observed similiar low frequencies of GFP+ wt and Siglec-H−/− pDCs at 24 h p.i. (Figure 4a). Next we quantified the viral titers from MCMV infected sorted pDCs and cDCs from wt and Siglec-H−/− BMDC cultures. Infected cDCs from both groups showed high viral titers. As an additional positive control we included mouse embryonic fibroblasts (MEFs) infected with the same MOI. Wt and Siglec-H−/− pDCs showed basal levels of infection close to the detection limit (Figure 4b) which is consistent with the literature describing pDCs as being resistant to productive infection [12], [30]. Additionally, MCMV infected Siglec-H−/− pDCs showed upregulation of the activation marker CD86 albeit to a slightly lower extent than wt pDCs (Figure 4c). Overall, Siglec-H deficiency on pDCs has no influence on MCMV infection and pDC activation occurs in its absence.
Siglec-H has a modulatory role in IFN signalling [16]. Thus, we queried whether the absence of Siglec-H would influence anti-viral immunity to MCMV infection in vivo. To address this aim, we used BM from wt and the recently described Siglec-H−/− mice [31] to generate chimeric mice and confirmed efficient reconstitution prior to infection (Figure 5a). Subsequently, mice were infected with a low dose of 5×104 plaque forming units (PFU) MCMV and a kinetics of IFNα serum levels was performed. Our results show that at both 1.5 and 3 days p.i. Siglec-H−/− mice had significantly higher IFNα serum levels compared with wt mice (Figure 5b). At day 6 p.i., IFNα levels declined to wt levels in the Siglec-H−/− mice (Figure 5b). Thus, we show that MCMV infected Siglec-H−/− mice produce significantly more IFNα during the acute phase of MCMV infection compared with wt mice.
Ablation of pDCs after low dose MCMV infection was shown to enhance the viral load in the spleen, liver and salivary glands of infected mice [13]. However, the functional contribution of Siglec-H in viral clearance remained unknown. As NK cells and CD8+ T cells are the major effector cells for MCMV clearance [32], we first tested whether the increase in serum IFNα in infected Siglec-H−/− mice influenced NK cell activation during the early phase of infection. Both serum IFNγ levels and CD69 expression on blood NK cells were comparable between Siglec-H−/− and wt mice (Figure 6a, b). Furthermore we checked for KLRG-1 expression on MCMV-specific Ly49H+ NK cells at day 8 p.i. and found a similar activation status when comparing cells from wt and Siglec-H−/− mice (Figure 6c, d).
We next analyzed MCMV-specific CD8+ T cells by H-2Db M45 tetramer staining on day 8 p.i. Infected Siglec-H−/− and wt mice showed comparable frequencies (Figure 6e, f) and numbers (Figure 6g) of tetramer+CD8+ T cells in the spleens of infected mice. Also the IFNγ production by CD8+ T cells from both groups was comparable upon restimulation with M45 peptide (Figure 6h).
We finally examined the viral load in the spleen, liver and salivary glands at day 3, 6 and 8 p.i. Importantly, viral clearance in primary and secondary organs showed similar kinetics between the two groups over time (Figure 6i). Infection peaked in the spleen at day 3 and in the liver at day 6 with no significant differences due to Siglec-H deficiency. Also infection in the salivary glands reached the same level on day 6 and 8 p.i. Overall, our data show that absence of Siglec-H does not alter the NK cell activation or expansion of IFNγ+ MCMV-specific CD8+ T cell generated in response to infection. Additionally, viral clearance in the infected organs remains unchanged irrespective of the presence or absence of Siglec-H.
We have shown that Siglec-H expression is downregulated by pDCs upon CpG-A or MCMV treatment in a MyD88- and TLR9-dependent mechanism. This phenotypic change is observed in vitro and in vivo. Since a downregulation of Siglec-H expression complicates the identification of pDCs, we generated a novel mouse model for fate mapping Siglec-H expression. In these mice Siglec-H+ cells are terminally labeled irrespective of Siglec-H downregulation at a later timepoint. The reporter expression in the pDCre x RFP mice is the highest in pDCs (23–30%) when gating on individual cell populations, as expected. Interestingly, we also observed RFP expression in a minor fraction of B-, T-, NK-, NK-T cells, as well as cDCs and common dendritic cell precursors (CDPs). This might be explained by targeting of a Siglec-H+ precursor population as precursor-derived cells would be labeled in our fate mapping approach even in the absence of Siglec-H expression after terminal differentiation. In line with this hypothesis, Siglec-H mRNA expression has been reported recently for B cell progenitors, CDPs as well as common lymphoid progenitors (CLPs) [33]. Furthermore, Satpathy et al. have shown that a fraction of CDPs and particularly a subset of pre-cDCs express Siglec-H at the surface [34]. However, Siglec-H expression by these precursors does not necessarily correlate with pDC differentiation potential and can be transient as a proportion of Zbtb46+ Siglec-H+ precursors retains cDC differentiation potential [34]. In addition, Swiecki et al. have demonstrated that a fraction of cDCs are targeted in a Siglec-H GFP knock in approach [13]. Supporting these findings, we also report minor targeting of pro- and pre-DCs in vitro using Flt3-L cultures of pDCre x RFP BM. Interestingly, a substantial proportion of Siglec-H− precursors is fate mapped, which might indeed indicate that RFP reporter expression precedes Siglec-H surface expression. It will be interesting to test pDCre x RFP mice for possible targeting of CLPs which can give rise to T-, NK-, NK-T and B cells [35]–[38]. To this end, we observed minor RFP reporter expression in early CD3ε− and DN thymocytes (data not shown). It is unlikely that BAC-related positional effects [39] are responsible for the observed expression pattern in non-pDCs as we found similar results in different pDCre founderlines generated from the same BAC (data not shown). Thus, minor cDC, B cell, CDP, and potential CLP targeting observed in the pDCre x RFP mice is likely a reflection of early Siglec-H promoter activity in committed precursors.
Interestingly, not all Siglec-H+ pDCs express the reporter in our BAC transgenic model. For both humans and mice, phenotypically and functionally distinct pDC subsets have been described [23]–[27]. However, the expression levels of the pDC surface markers B220, Ly-6C, PDCA-1, CD8α, Clec9A, Siglec-H and CD11c greatly overlapped between RFP+ and RFP− pDCs in the spleen. Cisse et al. demonstrated when comparing BMDC cultures from wt and E2-2−/− mice, where pDC differentiation is blocked at an early step, that Ly-6C is upregulated before PDCA-1, B220, and Siglec-H [40]. Furthermore, CCR9 expression is also induced late during pDC development. [24]. Thus, it appeared that RFP+ pDCs showed a slightly more mature phenotype in the BM of pDCre x RFP mice. This might indicate that we are preferentially targeting terminally differentiated CCR9+ pDCs and fewer Siglec-H+ CCR9− B220low pDC precursors [24], [41]. pDCs mainly develop from Lin−c-Kitint/loFlt3+M-CSFR+ CDPs [28], [29]. Recently a progenitor with preferential pDC developmental potential has been identified [42]. Yet, no pDC precursor with a sole pDC differentiation potential has been found. Along this line, it would be interesting to test whether RFP+ and RFP− pDCs from the pDCre x RFP mice originate from the same precursor population. Despite the minor phenotypic differences observed in the BM, we did not find a significant disparity in the IFNα or TNFα production upon MCMV infection by peripheral RFP+ and RFP− pDCs indicating that we are targeting a representative fraction of cytokine-competent pDCs rather than a distinct pDC subset. We conclude that the pDCre x RFP mice are a suitable model to study cytokine responses by pDCs upon virus challenge.
The injected BAC was chosen to cover approximately 100 kb up- and downstream of the open reading frame to minimize positional effects which are often observed with classical transgenic mice [39], [43]. However, missing regulatory elements required for Siglec-H expression by all pDCs cannot be completely excluded. It should be noted that many helpful transgenic mouse models display a reporter activity that is lower than 100%. ([44], [45] and own unpublished data). This is a commonly observed phenomenon published for different promoters.
We found that Siglec-H downregulation on pDCs is an early response to PAMP recognition in vitro and in vivo. Productive infection of pDCs with MCMV is, however, low [30], possibly as a consequence of the strong anti-viral response in these cells [12]. Thus, Siglec-H downregulation might be the consequence of pDC activation primarily by means of viral PAMP transfer rather than productive infection as previously shown for other viruses [46], [47]. Because TLR7/9 ligation of pDCs induces a massive type I IFN release, it was important to investigate whether type I IFN signalling alone can induce Siglec-H downregulation. Interestingly, Siglec-H expression by mock treated IFNAR−/− BM-derived pDCs was in general lower as compared with wt pDCs, suggesting a link between IFNAR signalling and Siglec-H expression. Yet, Siglec-H downregulation also occurred in the absence of IFNAR upon MCMV infection. Thus, Siglec-H downregulation is not a direct consequence of type I IFN signalling. Notably, also other soluble factors released by infected wt DCs did not induce Siglec-H downregulation on MyD88−/− pDCs in a transwell experiment (data not shown), arguing for a pDC-intrinsic effect. We conclude that the downregulation of Siglec-H is coupled to early intracellular MyD88-dependent signalling events with minor contributions of the type I IFN signalling pathway.
Siglec-H was previously hypothesized to act as a pathogen receptor although its natural ligands remain enigmatic [16]. However, our data suggest that MCMV infection of pDCs is not affected by the absence of Siglec-H. Moreover, Siglec-H-deficient mice are not protected from MCMV infection. Although Siglec-H seems to be dispensable for infection, this does not exclude that it might participate in virus uptake together with other endocytic receptors with redundant functions.
We show that although Siglec-H is neither required for MCMV replication nor for pDC activation, the absence of this endocytic receptor leads to strongly increased IFNα serum levels in infected mice. It has been previously shown that Siglec-H couples to the adaptor protein DAP12, which negatively regulates type I signalling [19], [48]. DAP12−/− pDCs lack Siglec-H expression and mount higher IFNα levels compared with wt pDCs [19], [48] similar to our findings for the Siglec-H−/− mice. Interestingly, it was demonstrated that DAP12-deficient pDCs also produce increased amounts of IL-12 [19], suggesting that DAP12 may regulate additional DC functions. Whether the absence of Siglec-H or its downmodulation also affects DAP12 expression or function warrants further investigation.
Siglec-H−/− mice showed increased levels of serum IFNα already 36 h p.i. with MCMV. pDCs are the main source of this cytokine at this timepoint [13] and thus the absence of Siglec-H on pDCs has direct consequences for the negative regulation of IFNα signalling by these cells. This finding is consistent with published data from Takagi et al. [20] who demonstrated a similar increase in IFNα serum levels in Siglec-H deficient Siglec-H DTR/DTR mice in the HSV-1 infection model. In contrast, we did not observe a direct effect of these enhanced cytokine levels on the primary CD8 T cell response or NK cell activation and no differences in viral clearance in spleen, liver or salivary glands during acute infection. This is surprising given that Takagi et al. have shown that in the absence of Siglec-H the HSV-1 specific CD8+ T cell response is reduced and viral clearance is diminished [20]. The disparities could be explained by the different mouse or infection model. Alternatively, the wt IFNα serum levels upon MCMV infection might already be saturating.
It is well established that type I IFN promotes anti-viral and immunostimulatory responses [49]–[51]. However, the increased IFNα levels in Siglec-H−/− chimeric mice could also play a dual role and induce negative immune regulation as recently described [52], [53]. Along this line, it would also be interesting to test how Siglec-H−/− mice control reactivation of the virus during chronic infection.
In conclusion, we employed valuable novel mouse models to terminally label a fraction of pDCs and to assess the functional role of Siglec-H in vivo. Although Siglec-H is a well defined pDC marker upon steady state conditions, our data suggest that caution should be taken when using Siglec-H as a single marker to identify pDCs upon MCMV infection. This might also apply for other infection models. The mechanisms of TLR-induced Siglec-H downregulation and increased IFNα secretion deserve future investigations.
All animal experiments were performed in compliance with the German animal protection law (TierSchG BGBl. I S. 1105; 25.05.1998). The mice were housed and handled in accordance with good animal practice as defined by FELASA and the national animal welfare body GV-SOLAS. All animal experiments were approved by the Lower Saxony Committee on the Ethics of Animal Experiments as well as the responsible state office (Lower Saxony State Office of Consumer Protection and Food Safety) under the permit numbers 33.9-42502-04-12/1020 and 33.9-42502-04-09/1785. All surgery was performed after mice were euthanized and all efforts were made to minimize suffering.
pDCre mice were generated using BAC technology [39], [54] with some modifications. We obtained a BAC encoding the complete mouse Siglec-H gene locus (RP24-396N13) from the BACPAC Resources Center at Children's Hospital Oakland Research Institute. As transgene, we created a Cre-IRES-cherry construct containing a 3′ polyA fragment (Figure S1). In contrast to the published overlap PCR strategy [39], we used a strategy based on homologous recombination: 1000 bp long regions homologous to the 5′ end before the ATG of exon 1 and 3′ of exon 1 of Siglec-H were ligated to the transgene construct via PCR-inserted AscI (5′) or PmeI sites (3′). The cherry sequence was a kind gift from Dr. R. Tsien (UC San Diego). The polyA fragment was amplified from the TOPO Tools SV40 pA 3′ element kit (Promega) using primers adding a SpeI site 5′ and a PmeI site 3′. The Siglec-H containing BAC was recombined using the pLD53.SC1 shuttle vector provided by Dr. N. Heintz (The Rockefeller University, New York, NY), gel purified and injected into the pronuclei of fertilized C57BL/6 oocytes. Multiple transgenic pDCre founder lines were established and the pDCre x RFP line with the highest RFP expression of 23±5% in BM- and 30±8% in splenic pDCs was utilized in this paper. Within the RFP+ live cell population pDCs constitute 69±8% (BM) and 9±6% (spleen) of total cells. Frequencies are given as mean ± SD. pDCre mice were genotyped by PCR using the following primers: (5′-ttccatggcatgagagaaca-3′) and (5′-agtccagaagcccaaaggat-3′). To analyze for Cre-recombinase activity, the transgenic mice were bred to reporter mice which carry a floxed td-RFP cassette downstream of the ubiquitous Rosa26 promoter [22].
Siglec-H−/− BM was kindly provided by Dr. Lars Nitschke (Erlangen, Germany). Siglec-H−/− mice were generated by the consortium for functional glycomics (http://www.functionalglycomics.org/) [31] and provided by The Scripps Research Institute (TSRI) (La Jolla, CA, USA). The construct consisted of an FRT-flanked selection vector containing a neomycin resistance gene and a single 3′ loxP site was used in the generation of this construct. The selection vector was inserted downstream of exon 2 of the Siglec-H gene and a second loxP site was inserted upstream of exon 1 via homologous recombination. The mice carrying the construct were then crossed with a Cre-recombinase transgenic mouse thereby removing the majority of the coding sequences of Siglec-H including those for the Ig-like domains and alternatively spliced exon 2a.
For generating Siglec-H−/− chimeric mice, CD45.1 mice were obtained from Jackson laboratory, Maine, USA. CD45.1+ recipient mice were irradiated with 10 Gy and rested for 24 h after radiation treatment. Irradiated mice were injected intra-venously with 4×106 BM cells from Siglec-H−/− or C57BL/6J control mice (both CD45.2 background). Following injection of BM cells, mice were carefully monitored and reconstitution was carried out for 6–8 weeks. Reconstitution efficiency was examined by staining blood cells for CD45.1 vs. 45.2 after 6–8 weeks. Siglec-H−/− chimeric mice did not show any signs of kidney dysfunction as assessed by serum creatinine measurements.
pDCre x RFP reporter and C57BL/6 wildtype (wt) mice (originally obtained from Jackson laboratories) were bred at the animal facility of Twincore (Hannover, Germany) and at the Helmholtz Centre for Infection Research (HZI, Braunschweig, Germany). 8–12 week old mice were used for animal experiments. All animals were housed under specific pathogen-free conditions. Mice were sacrificed by CO2 asphyxiation as approved by German animal welfare law. Every effort was made to minimize any sort of suffering to the animals.
The MCMV laboratory strains used were the BAC-derived wt Smith Strain [55] and the recombinant Smith-based GFP strain [56]. Both strains were kindly provided by Dr. Martin Messerle from the Institute of Virology, Medical School, Hanover, Germany. Virus strains were propagated in mouse embryonic fibroblast (MEF) as described previously [57].
1×106 BMDCs/well were seeded in a 24well flat bottom plate. FACS sorted Flt3-L derived pDCs were seeded at 1×105 cells/well in 96well flat bottom plate. DCs were incubated with MCMV Smith-GFP or MCMV Smith (MOI 2), 1 µM CpG-A (ODN 2216) or PBS for 0, 3, 6 and 24 h. Infection of DCs was performed by centrifugation at 1500 rpm for 30 min at 37°C. Following spin infection, DCs were incubated at 37°C, 5% CO2 till harvested and used for subsequent stainings.
30,000 sorted SiglecH+RFP+, SiglecH+RFP− pDCs were seeded in a 96well U bottom plate and incubated with MCMV (MOI 2 or 5), CpGA-2216 (0.1 µM, 1 µM) or mock infected with PBS. Cells were incubated with appropriate stimuli for 24 h at 37°C, 5% CO2. Supernatants were collected and cytokines were quantified as described below.
Mice were infected intra-peritoneally with 5×104 PFUs of wt MCMV-Smith strain or mock infected with PBS. Mice were carefully monitored over the course of infection and were asymptomatic.
Spleen, liver and salivary glands from sacrificed mice or cell pellets from infected DCs/MEFs (modified protocol from Mathys et al. [56]) were resuspended in MEM medium (Gibco Life Technologies, Darmstadt, Germany). Homogenization was performed by using a handheld homogenizer POLYTRON PT1200 (Fischer-Scientific GmbH, Schwerte, Germany). Organ homogenates were then plated on a confluent layer of murine embryonic fibroblasts (MEFs) and a serial dilution of 10−1–10−6 was performed for each organ in duplicates. Infections of MEFs were carried out in 24well flat bottom plates for 2 h at 37°C. After 2 h of incubation the organ homogenates were decanted carefully and a layer of methyl cellulose was overlayed to avoid infection by residual free floating virus particles and permit only cell to cell transfer of virus infection. After methyl cellulose overlay, the plates were left at 37°C for 6 days until plaques developed.
Spleens were cut into small pieces and digested in RPMI 1640 Glutamax medium (Gibco Life Technologies, Darmstadt, Germany) supplemented with 10% FCS, 5 µM β-mercaptoethanol (Gibco Life Technologies, Darmstadt, Germany), 100 U/ml Penicilin/Streptomycin (Biochrom, Berlin, Germany), 1 mg/ml Collagenase D (Roche diagnostics GmbH, Mannheim, Germany) and 100 µg/ml DNase I (Roche diagnostics GmbH, Mannheim, Germany). Digestion was stopped by addition of 10 mM EDTA (pH 7.2). A single cell solution was prepared and red blood cells were lyzed in RBC lysis buffer (150 mM NH4Cl, 10 mM KHCO3, 0.1 mM EDTA). The isolated cells were counted by trypan blue exclusion and adjusted to the same cell number for FACS staining.
BM cells were isolated from femurs and tibiae. In vitro generation of fms-like tyrosine kinase 3 ligand (Flt3-L) driven DCs from BM has been described previously [58]. 15×106 BM cells were seeded in 10 ml supplemented RPMI 1640 Glutamax medium (Gibco Life Technologies, Darmstadt, Germany) containing Flt3-L (self-made from CHO Flt3-L FLAG cells [58]). Cells were cultured for 9 days at 37°C, 5% CO2 before stimulation experiments were performed. Flt3-L producing CHO cells were generated by Dr. Nicos Nicola and kindly provided by Dr. Karen Murphy, WEHI, Melbourne, Australia.
Cells were washed in PBS and stained with the live/dead fixable aqua dead cell stain kit (Invitrogen, Life Technologies GmbH, Darmstadt, Germany) to exclude dead cells. Cells were washed with PBS and incubated in FACS buffer (0.25% BSA/2 mM EDTA in PBS) containing Fc-block (CD16/32, 2.4G2) for 10 min on ice. Cells were stained with fluorescently labeled antibodies (Ab) for cell surface markers. Stainings were performed for 20 to 30 min on ice. Cells were fixed with 2% PFA in PBS for 20 min on ice. Where indicated, cells were washed with permeabilisation buffer (0.25% BSA/2 mM EDTA/0.5% saponin in PBS) and stained intracellularly for IFNγ 20 min on ice. Samples were stored at +4°C, acquired on a LSRII flow cytometer (BD Bioscience GmbH, Heidelberg, Germany), and analyzed using FlowJo software (Tree Star, Inc. Ashland, USA). Single stains and fluorescence minus one controls were used for accurate gating and compensation. Non-specific binding was estimated by isotype controls and cellular aggregates were excluded by SSC-W.
All antibodies were purchased from eBioscience (Frankfurt, Germany) if not stated otherwise.
The following fluorochrome labeled anti-mouse antibodies were used:
For pDCs after in vitro infection: B220 (RA3-6B2), CD11c (N418), CD86 (GL1), MHCII (AF6-120.1) and Siglec-H (440c). For pDC markers: CD11c (N418), B220 (RA3-6B2), PDCA-1 (eBio927), Ly-6C (HK1.4), CD9 (MZ3), CCR9 (242503, R&D), Clec9A (7H11, Miltenyi), Siglec-H (440c), CD4 (GK1.5) and CD8α (53–6.7). For characterization stainings of spleen and BM from pDCre x RFP mice: NK1.1 (PK136), CD3ε (145-2C11), CD19 (eBio1D3), CD11c (N418), Siglec-H (440c), MHCII (M5/114.15.2).
For blood NK cells: NK 1.1 (PK136), Ly49H (3D10), CD69 (H1.2F3), and CD3ε (145-2C11). For peptide restimulated splenocytes: CD8α (53–6.7), CD62L (MEL-14), CD44 (IM7), KLRG1 (2F1) extracellularly and IFNγ (XMG1.2) intracellularly. H-2Db M45 tetramer (PE-conjugated, a kind gift from Dr. Luka Cicin-Sain) stainings were performed in a 96well U bottom plate at 2×106 splenocytes/well after surface staining for 1 h on ice.
Cell sorting was carried out at the Cell Sorting Core facility of the Hannover Medical School using the FACSAria (BD) or XDP MoFlow (Beckman Coulter) FACS sorting machines. pDCs were sorted as CD11cint Siglec-H+ NK1.1− CD19− CD3ε− after enrichment from spleens of B16-Flt3L treated pDCre x RFP mice using Optiprep gradient (Axis Shield, Oslo, Norway). Cells were sorted with a 100 µm nozzle, collected in filtered FCS and kept on ice till use. The purity of the sorted cells was determined for each sample and was >95–98%.
T cell restimulation assay was performed in a 96well U bottom plate. 2×106 splenocytes from mock treated and day 8 MCMV infected wiltype and Siglec-H−/− chimeric mice were incubated in the presence of 3 µg/ml brefeldin A (eBioscience, Frankfurt, Germany) and M45 H-2Db peptide (HGIRNASFI, Pro-Immune, Oxford) at 1 µg/ml, 1 ng/ml and 1 pg/ml final concentration or without peptide at 37 C, 5% CO2 for 4 h.
Supernatants from stimulated pDCs and serum samples from infected mice were used to determine cytokine concentration by ELISA or by FlowCytomix bead assay. For the IFNα ELISA, maxisorp plates (Nunc, Denmark) were coated with anti-mouse IFNα (clone: RMMA-1) and blocked using 1% BSA in PBS. Samples were incubated overnight at 4°C. A poly-clonal anti-IFNα antibody from rabbit (both IFNα antibodies from PBL Interferon Source, NJ, USA) and a HRP conjugated anti-rabbit IgG (H+L) antibody (Dianova, Hamburg, Germany) were used as secondary reagent. HRP activity was detected by BD OptEIA solutionA/B (BD Bioscience GmbH, Heidelberg, Germany) and quantified at 570/450 nm. TNFα and IFNγ were quantified by a FlowCytomix bead assay (eBioscience, Frankfurt, Germany).
The Students t-test was used to calculate the statistically significant differences between samples. A value for p<0.05 was considered significant indicated by an asterisk sign: (*) for P<0.05, (**) for P<0.01 and (***) for P<0.001.
|
Subsets and Splits