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.ppat.1003817 | Host-Specific Enzyme-Substrate Interactions in SPM-1 Metallo-β-Lactamase Are Modulated by Second Sphere Residues | Pseudomonas aeruginosa is one of the most virulent and resistant non-fermenting Gram-negative pathogens in the clinic. Unfortunately, P. aeruginosa has acquired genes encoding metallo-β-lactamases (MβLs), enzymes able to hydrolyze most β-lactam antibiotics. SPM-1 is an MβL produced only by P. aeruginosa, while other MβLs are found in different bacteria. Despite similar active sites, the resistance profile of MβLs towards β-lactams changes from one enzyme to the other. SPM-1 is unique among pathogen-associated MβLs in that it contains “atypical” second sphere residues (S84, G121). Codon randomization on these positions and further selection of resistance-conferring mutants was performed. MICs, periplasmic enzymatic activity, Zn(II) requirements, and protein stability was assessed. Our results indicated that identity of second sphere residues modulates the substrate preferences and the resistance profile of SPM-1 expressed in P. aeruginosa. The second sphere residues found in wild type SPM-1 give rise to a substrate selectivity that is observed only in the periplasmic environment. These residues also allow SPM-1 to confer resistance in P. aeruginosa under Zn(II)-limiting conditions, such as those expected under infection. By optimizing the catalytic efficiency towards β-lactam antibiotics, the enzyme stability and the Zn(II) binding features, molecular evolution meets the specific needs of a pathogenic bacterial host by means of substitutions outside the active site.
| The presence of Zn(II)-containing metallo-β-lactamases (MβLs) that confer resistance to all penicillins, cephalosporins and carbapenems in Pseudomonas aeruginosa adds significantly to the threat of this pathogen in our health care system. SPM-1 is an MβLs widely distributed in South America and only found in P. aeruginosa. In common with all MβLs, the active site residues are highly conserved. In this work we asked the following question: how would substrate specificity evolve in SPM-1 if the active site residues are highly uniform and do not permit substitutions. To this end, we explored the role of two amino acids (S84 and G121) that are outside the active site (second sphere) and are unique in the SPM-1 β-lactamase. We discovered that replacing these amino acids impacts resistance to cephalosporins and carbapenems and that this resistance profile depends on the enzymatic behavior and the availability of Zn(II) in the environment. This work demonstrates how protein evolution by means of subtle substitutions outside the active site meets the specific needs of a pathogenic bacterial host.
| β-lactam antibiotics (penicillins, cephalosporins, monobactams and carbapenems) are the most dependable and frequently employed chemotherapeutic agents for eradicating bacterial infections [1]. Their safety and efficacy as antimicrobial agents derives from their ability to selectively inhibit cell wall biosynthesis, provoking bacterial cell wall lysis [2]. The newest types of β-lactam antibiotics (e.g. carbapenems) and expanded-spectrum cephalosporins (e.g. cefepime), evade most common mechanisms of resistance against these compounds [3]. These compounds are currently used as “last resort” drugs for treating multi-resistant gram-negative pathogens [1], [3].
The major mechanism of resistance against β-lactam antibiotics is the production of bacterial β-lactamases which catalyze cleavage of the antibiotic β-lactam ring rendering an inactive derivative [2]. β-lactamases fall into four classes (A–D). Classes A, C and D are serine-β-lactamases (SβLs) which employ an active-site serine to catalyze antibiotic hydrolysis, while metallo-β-lactamases (MβLs), or class B β-lactamases, are metallo-enzymes requiring one or two zinc ions for their activity [4].
MβLs gained importance in the 1990s as the principal mechanism of resistance against carbapenems (imipenem, meropenem) [5], [6], [7]. MβLs degrade all classes of β-lactams except monobactams and, unlike most SβLs, these enzymes are not susceptible to therapeutic β-lactamase inhibitors. This fact, together with the facile dissemination of MβL genes among different clinical pathogens, relegates them as a serious clinical threat [5], [7]. Indeed, outbreaks of pathogens producing NDM-1, IMPs, VIMs or SPM-1 MβLs are increasingly common worldwide [8].
Atomic structures reveal that clinically relevant MβLs (subclass B1) possess similar active sites: indeed, residues binding the essential Zn(II) ions (first sphere residues) are strictly conserved (Figure 1A) [6], [7]. Despite being “broad-spectrum” enzymes, MβLs exhibit quite different substrate profiles, which cannot be correlated to different active site structures [9]. Many structural and mechanistic studies have focused on the analysis of active site residues and the role of active-site flanking loops to account for substrate recognition of MβLs [6], [9], [10]. However, the mechanism by which different B1 enzymes are tailored to hydrolyze some antibiotics better than others is not known. The fundamental question remains: how does protein evolution occur among MβLs that are found exclusively and adapted to a particular host? This problem represents a central issue in linking molecular features to organismal behavior. In the clinic this notion may contribute to therapeutic failure.
Pseudomonas aeruginosa is one of the most clinically important non-fermenting Gram-negative pathogens, being well known for its ability to acquire genes encoding resistance determinants, such as the acquired MβLs [11], [12]. In addition, P. aeruginosa harbors a host of virulence factors. Of particular relevance, SPM-1 is an MβL produced only by P. aeruginosa, while other MβLs have been found in many different bacterial hosts [12], [13], [14], [15], [16], [17], [18]. At the present time, the blaSPM-1 gene is associated with a single clone (SP/ST 277) of P. aeruginosa. This clone emerged relatively recently in South America. This unique dissemination suggests either: 1) the blaSPM gene came from another organism and has expanded in SP/ST 277 because of a fitness advantage; or that 2) this genetic determinant may have been optimized to meet the need of its native host. SPM-1 in P. aeruginosa is therefore a unique system to analyze the role of host-specific constraints in molecular evolution.
The structure of SPM-1 has revealed unique features among pathogen-associated MβLs [19]. Spencer and coworkers have shown that clinically relevant B1 enzymes share a hydrogen bonding network spanning below the active site base, generally known as second sphere residues (Figure 1) [19]. This network is disrupted in SPM-1 due to the presence of two atypical second sphere residues: S84 and G121, which replace the conserved D84/R121 couple (Figure 1B) [9].
Here we examine the role of these positions (located outside the enzyme active site “in the second sphere”) and their impact on antibiotic resistance in the native bacterial host, P. aeruginosa. We report that this unique combination of residues is able to provide resistance to anti-pseudomonal β-lactams such as latest generation cephalosporins and carbapenems [11], while sacrificing the catalytic efficiency against other β-lactams. Our findings reveal that second sphere residues are able to modulate the substrate specificity of MβLs according to the requirements of the bacterial host In addition, we show that these second sphere residues optimize the zinc binding affinity of SPM-1 in the bacterial periplasm, providing P. aeruginosa antibiotic resistance under zinc-limiting conditions, such as those prevalent during bacterial infection [20], [21].
E.coli is usually employed as a model bacterial host to compare the ability of the different MβLs to confer resistance, even for enzymes which are not found in Enterobacteriaceae [9]. We designed a system aimed to reproduce the native conditions of expression of the blaSPM-1 gene. The complete blaSPM-1 transcriptional unit from the clinical strain P. aeruginosa 48-1997A [15] (i.e., including the natural promoter, the leader peptide for periplasmic location, the mature protein and the transcriptional terminator) was amplified and subcloned into the broad-spectrum vector pBBR1-MCS5 [22], replicative in P. aeruginosa PAO (Figure 2A). P. aeruginosa PAO cells transformed with this vector (pΔEP-SPM-1) were able to express SPM-1, export and process it properly to the periplasmic space. Western blot analysis showed two SPM-1 forms of 30.6 and 27.5 kDa in whole cell extracts, corresponding to the precursor and mature species respectively [13]. Instead, the periplasmic fraction contained only the mature form of the enzyme (Figure 2B,C). Accordingly, the transformed cells were resistant to imipenem.
In order to assess the “flexibility” of positions 84 and 121 in accommodating residues different from the native ones (S84 and G121), codons 84 and 121 were individually randomized in blaSPM-1 by overlap-extension PCR [23], [24]. The amplifications were targeted to the mature blaSPM-1 coding sequence, and then subcloned into the screening vector, so as to avoid undesired mutations in promoter and terminator sequences. In addition, codons 84 and 121 were randomized together looking for possible synergistic effects between these positions.
Single-codon random libraries gave rise to 103–104 transformants, while the double-codon mutant library elicited >3×104 transformants. According to Poisson distribution, the libraries obtained have a probability of harboring a mutant blaSPM-1 gene with a specific codon at position 84 or 121 (or a specific combination of codons) >99% [23]. Sequences of ten randomly selected mutants from each library indicated no obvious bias.
Active mutants were selected by examining the ability of the different libraries to confer resistance toward different types of β-lactam antibiotics in P. aeruginosa PAO (Figure 3). Paper discs embedded with different antibiotics were applied onto LB-Gm agar plates with P. aeruginosa PAO transformed with the randomized libraries. We employed a penicillin (piperacillin), a third-generation cephalosporin (ceftazidime), a cephamycin (cefoxitin), and a carbapenem (imipenem).
Twenty bacterial clones exhibiting resistance (i.e., located within the halos) were isolated for each library and for each tested antibiotic (a total of 240 clones). Plasmids were extracted and blaSPM-1 was sequenced in each clone. In total, 16 different variants (wild type SPM-1, 10 single mutants and 5 double mutants) were isolated. As expected, wild type clones (with residues S84 and G121) were selected in all cases. Mutants G121A, S84N and S84N/G121S were also selected against all tested antibiotics (Figure 3). On the other hand, some substitutions were isolated depending on the screening antibiotic, implying that positions 84 and 121 modulate the substrate profile of the enzyme. Surprisingly, none of the selected mutants carried mutations G121R or G121H (prevalent in B1 and B3 enzymes, respectively). Substitutions at position 84, instead, displayed typical residues from B1 (S84D), B2 (S84G) and B3 enzymes (S84N), among others [9]. The S84D/G121S combination is present in the B1 enzyme IMP-1, closely related to SPM-1 [13]. We then analyzed the resistance profile of the libraries.
MIC values for P. aeruginosa cells expressing each of the selected SPM-1 mutants were determined against different antibiotics. Cefepime (an antipseudomonal cephalosporin) was added to the initial set of antibiotics. Expression of SPM-1 markedly increased resistance towards antipseudomonas drugs such as ceftazidime and cefepime (200–250 times), while for cefoxitin (an antibiotic to which P. aeruginosa PAO is naturally resistant), the increase in MIC was only 7-fold (Figure 4).
In general, single-codon variants S84G, S84N (naturally present in B2 and B3 enzymes) and G121A (the most conservative substitution in this position) display the highest MIC values after the wild type (WT) enzyme (MIC values equal or up to 2-dilutions lower compared to WT SPM-1). In fact, together with S84N/G121S, these mutants were the most ubiquitous in the antibiotic selection experiments. Synergistic effects between residues 84 and 121 are apparent when comparing double mutants vs. single mutants. For example, while S84G and G121S mutations were detrimental for resistance against piperacillin (MIC values approximately half a dilution lower than for WT), the combination of both mutations generated an enzyme conferring higher levels of resistance than the wild type (MIC value of 16 µg/ml for S84G/G121S vs. 10 µg/ml for WT SPM-1) (Figure 4). Surprisingly, the S84D/G121S combination, naturally occurring in IMP enzymes, was not among the most resistant mutants for any of the antibiotics assayed (MIC values 2–3 dilutions lower compared to WT SPM-1).
Figure 4 summarizes our data showing that mutations had different impact in the bacterial resistance profile depending on the antibiotic (selection criteria). Therefore, second sphere positions 84 and 121 are able to shape the resistance profile, and possibly the substrate specificity. In general, mutants conferred lower levels of resistance than wild type SPM-1 (within a range of 5 dilutions in MIC values). Surprisingly, piperacillin is an exception, since four mutants outperform the wild type variant (S84N, S84N/G121S, S84Q/G121S and S84G/G121S in up to half a dilution in MIC values). In the case of cefoxitin, the range of MIC values spanned by the different variants is smaller than for the rest of the tested antibiotics (3 dilutions vs. 4 or 5 dilutions). The impact of mutations on MIC values is more informative for the case of the antipseudomonas compounds ceftazidime and cefepime, where several single and double mutants provide levels of resistance comparable to the WT enzyme, with MIC values increasing by two orders of magnitude. For imipenem, G121A is the only mutant giving rise to a large MIC value (35 µg/ml vs. 48 µg/ml for WT SPM-1).
Enzymatic studies in vitro of MβLs have been useful to uncover structural and mechanistic aspects of these enzymes. However, these data rarely correlate with the in vivo behavior [9]. We attempted to correlate the MIC values with the hydrolytic profile of the different SPM-1 mutants assayed in periplasmic extracts of P. aeruginosa, i.e., in an environment closer to in vivo conditions. The β-lactamase activity of SPM-1 mutants was assayed in periplasmic extracts of P. aeruginosa PAO (in periplasma) and normalized relative to the amount of enzyme present in the periplasm (quantitated from Western blot gels). Given that SPM-1 is an efficient cephalosporinase in vitro, we focused on these substrates. We employed three substrates with antipseudomonal activity already used in the MIC experiments: ceftazidime, cefepime and the carbapenem drug imipenem, together with two first-generation cephalosporins devoid of antipseudomonal activity (cephalexin and cephalothin).
Hydrolysis rates in periplasma show a very good correlation with MIC values in the case of cefepime and imipenem (Figure 5). For these two substrates, only mutant G121A was competitive with the performance of wild type SPM-1. Instead, in the case of ceftazidime, cephalotin and cephalexin, the hydrolytic performance of the wild type enzyme was surpassed by several mutants. Mutant S84G (present in B2 enzymes), and to a lesser extent S84N (present in B3 enzymes) were the variants eliciting the best performance for first-generation cephalosporins. We conclude that the second coordination sphere modulates the substrate specificity in SPM-1 so that this enzyme is adapted to better hydrolyze the latest antipseudomonal antibiotics (cefepime and imipenem) while the catalytic performance against first-generation drugs is far from being optimized. In all cases, the activity of endogenous AmpC was negligible (as revealed by the lack of activity in the “No SPM-1” control strains).
Ceftazidime shows a different profile: albeit being an antipseudomonal drug, can be better hydrolyzed by several mutants than by native SPM-1. However, P. aeruginosa has developed different resistance mechanisms against ceftazidime which do not affect cefepime and imipenem (hyperproduction of endogenous AmpC, deregulation of efflux pumps or acquisition of ESBLs) [11]. We therefore evaluated the role of SPM-1 in the resistance of the P. aeruginosa clinical strain against these three antibiotics. Disks embedded with ceftazidime, cefepime and imipenem were paired with disks containing dipicolinic acid (DPA, an inhibitor of SPM-1), on an agar plate inoculated with the clinical strain P. aeruginosa 48-1997A (including its native blaSPM-1 gene) and the control P. aeruginosa PAO expressing SPM-1 [25]. While halos of inhibition were similar for all antibiotics in the control strain, ceftazidime exhibited a reduced halo in the clinical strain (Figure 6). When DPA was added to whole cell extracts of the model strain, no residual activity was monitored. Instead, residual ceftazidime hydrolysis was present after addition of DPA to extracts from the clinical strain (Figure 6). We conclude that resistance against ceftazidime in the clinical strain is not exclusively due to expression of SPM-1, and therefore this drug has not elicited a significant evolutionary pressure on this enzyme (or a higher activity against this antibiotic was not necessary a part of the substrate spectrum in order to be acquired by P. aeruginosa). In fact, there is evidence of SPM-1 producing isolates of P. aeruginosa that express AmpC (probably due to an hyperproduction phenotype) and OXA-52 enzymes, supporting our hypothesis [26].
We postulate that SPM-1 in the bacterial host has been exposed to the evolutionary pressure of the administration of newest antibiotics such as cefepime and imipenem, thus adapting to better hydrolyze these compounds by changes in the second coordination sphere. At this point, it is intriguing that mutant G121A, providing high levels of resistance and hydrolysis rates in the periplasm, has not been selected during natural evolution.
MβLs are exported to the bacterial periplasm as unfolded polypeptides [27]. Therefore, in the apo (non-metallated) form, metal site assembly (giving rise to the active variants) takes place in the periplasmic space [27]. We have recently shown that Zn(II) availability is limited in this compartment, and that MβLs with reduced Zn(II) binding capabilities are unable to confer resistance [28]. We determined the MIC values of P. aeruginosa cells with different SPM-1 mutants in media containing excess or limiting concentrations of Zn(II) against cefepime. MIC values were unaffected by Zn(II) supplementation in all cases. However, under metal deprivation conditions (by adding the chelating agent DPA), strikingly distinct effects were observed for the different SPM-1 variants (figure 7).
Mutant G121A, exhibiting a high specific activity in periplasma (Figure 5), was the most sensitive variant to metal deprivation (MIC values are 64-fold lower in 1 mM DPA) followed by variants S84K and G121D. Wild type and mutant G121S, on the other extreme, were almost unaffected by these conditions (MICs values diminished in only one-dilution in 1 mM DPA). The bacterial growth was also unaffected by these conditions as assayed by MIC values for the strain without the SPM-1 expression system. The lack of effect of excess Zn(II) in bacterial resistance likely reflects the action of the CzcABC pump in P. aeruginosa which, by extrusion of excess Zn(II) into the extracellular medium, keeps constant the levels of periplasmic Zn(II).
Thus, the second coordination sphere exquisitely tunes the zinc binding ability of SPM-1 so that the enzyme has been evolved to provide resistance at metal limiting concentrations. As a result, the atypical S84/G121 combination allows SPM-1 to confer antibiotic resistance under these conditions.
P. aeruginosa PAO periplasmic extracts revealed similar levels of periplasmic SPM-1 variants, with the exception of S84P and S84K mutants, which were undetectable. We analyzed the thermal stability of the mutants in periplasma, by studying the temperature dependence of (1) periplasmic β-lactamase activity or (2) SPM-1 solubility for each mutant. Hydrolysis rates were evaluated at 30°C after incubation at different temperatures. Instead, protein solubility was analyzed by Western blot quantitation of the levels of soluble SPM-1 variants after incubation at different temperatures. A plot correlating the hydrolytic activities (or solubilities) with the incubation temperature revealed in all cases a well-behaved sigmoidal behavior, that can be fit to obtain apparent Tm (Tmapp) values for each variant (Figure S1 and Table 1). Tmapp values determined from both strategies display an astonishingly good correlation.
The four most stable periplasmic SPM-1 variants were G121S>S84/G121 (WT)≈S84G/G121S≥S84G, while the combination S84D/G121S (naturally found in IMP-1) was the least stable mutant together with G121D and, to a lesser extent, G121A. At this point we selected some representative mutants for further characterization (S84D, S84G, G121A, G121S, G121D, G121N, S84D/G121S and wild type).
Similar experiments were performed with the apo-derivatives of periplasmic SPM-1 variants, which were obtained by dialyzing the periplasmic fractions against EDTA and DPA metal-chelators, excess NaCl and finally metal-free reaction buffer. Tmapp values of the apo variants were estimated as before by β-lactam activity (in this case supplementing reaction media with 2 µM Zn(II)) and protein solubility (Table 1).
Apo-derivatives exhibited a narrower range of Tmapp values, suggesting that the main differences observed in the stability of holo-derivatives are due to stabilization upon metal binding. Mutant G121S showed the largest metal-induced stabilization (∼30°C difference in Tmapp between holo and apo derivatives). Mutant S84D/G121S, on the other extreme, was marginally stabilized by metal binding. Mutants G121A and G121D precipitated during Zn(II) removal. Mutants showing larger differences in stabilities between the apo and holo form are expected to be those displaying large Zn(II) binding affinities. In agreement with this hypothesis, the variants exhibiting the highest metal-induced stabilization (wild type SPM-1, S84G, G121S and S84G/G121S) were the least sensitive to metal deprivation (Figure 7).
We explored the structural effects of these mutations by molecular dynamics (MD) simulations in wild type SPM-1 and three selected variants: G121A, G121S and S84D/G121S [19]. After 5 ns of MD simulations, a water molecule from the bulk solvent penetrates the active site of holo-wild type SPM-1, occupying the vacant position between T115 and S84 (Figure S2), and reconstructing the hydrogen bond network present in all B1 enzymes [19]. In the apo forms, Zn1 ligands become mobile, mostly due to alterations in second sphere residues. The largest changes were observed in the cavity between residues T115 and S84. In WT SPM-1, the second sphere adopted two different conformations: (a) “holo-like”, in which the cavity was able to accommodate water molecules connecting residues T115 and S84, and (b) “apo-like”, in which water molecules were excluded from this network, and T115 and S84 show a direct interaction (Figures 8 and S3). The holo and apo variants of G121S (the mutant showing the highest metal-induced stabilization) closely resemble the structure of the holo and apo forms of WT SPM-1. Instead, in the case of the S84D/G121S, the second sphere residues are locked into an “apo-like” conformation, disfavoring metal binding. In the case of G121A, the Ala121 side chain avoids contraction of the cavity, locking the second sphere into the “holo-like” form. The low stability of apo G121A suggests that this conformation is not viable in the apo form.
SPM-1 from P. aeruginosa is unique among pathogen-associated MβLs in presenting the singular S84/G121 combination as second sphere residues, instead of the conserved D84/R121 couple [13], [19]. Here we report a thorough study of the impact of mutations in these positions in the antibiotic resistance, specific enzymatic activity, metal binding features and protein stability.
A major novelty in our approach is the extensive use of the native host, P. aeruginosa. This approach allowed us: (1) to perform a medium-throughput screening of activities and stabilities of a series of mutants with a high degree of reproducibility, (2) to correlate enzymatic activities reproducing the native conditions (i.e., within the cell) during bacterial growth, which parallel the resistance profile and (3) to identify additional environmental factors which may not stem out from in vitro studies or by using E.coli as a model bacterial host [9]. In fact, MIC values of imipenem elicited by MβLs in E. coli are markedly lower to those determined in their natural hosts [9], [29]. In the particular case of MβLs, these enzymes are active only when the Zn(II) availability in the periplasm allows proper metal uptake, therefore being much more dependent on the bacterial host than serine lactamases [27], [28].
Expression of WT SPM-1 in P. aeruginosa selectively raises the MIC values against ceftazidime, cefepime and imipenem. These MIC values correlate with the in periplasma specific activities, in particular against cefepime and imipenem. The wild type variant, together with G121A, shows the highest specific activities against these two antibiotics. Instead, many single and double mutants in positions 84 and 121 outperformed wild type SPM-1 versus several antibiotics to which P. aeruginosa is intrinsically resistant. This remarkable substrate selectivity control by second sphere residues shows that the atypical S84/G121 combination present in SPM-1 has been fixed to provide resistance to anti-pseudomonal drugs, while sacrificing the catalytic efficiency against other antibiotics.
Analysis of the resistance profile against cefepime by controlling the Zn(II) availability in the external medium reveals that G121A (the only variant able to compete with wild type SPM-1 in terms of specific activity) is extremely sensitive to metal deprivation. The fact that G121A is not a natural variant of SPM-1 despite the high resistance observed in metal-rich media suggests that evolutionary pressure has been exerted to select MβL variants capable of providing resistance in low Zn(II) environments. Native SPM-1, instead, is able to confer resistance under conditions of Zn(II) deficiency. Indeed, during infection, the immune system produces large amounts of calprotectin, a host-defense protein that prevents bacterial colonization by chelating Mn(II) and Zn(II) [20], [21]. Thus, optimization of the zinc binding capabilities is a crucial evolutionary trait for MβLs in their natural environment. This finding, together with a recent report highlighting the need of proper assembly of a dinuclear site in the active site of MβLs in the periplasm [28], highlights the need to address the periplasmic bacterial mechanism of Zn(II) homeostasis and its role in antibiotic resistance, which have been largely overlooked.
The role of second sphere residues in catalysis is an emerging issue in enzymology [30]. A hydrogen bond network connecting metal binding residues below the active site is meant to preserve the electrostatics and modulate the active site features. Directed evolution experiments on the B1 enzyme BcII enzyme revealed that mutations responsible of enhancing the lactamase activity were located in this hydrogen bond network [31], [32]. As analyzed in detail by Spencer [19] and Oelshlaeger [33], [34] in structural, modeling and mutagenesis studies, this network spans metal ligands His116, Asp120 and Cys221, and the second sphere residues 115, 84, 121, 69, 70 and 262. The D84/R121 combination is the most commonly found in B1 enzymes [35]. Molecular dynamics simulations showed that water molecules can enter into the second sphere hydrogen bond network in SPM-1. These calculations also support how changes in the second sphere can modulate the Zn(II) binding affinity, ultimately impacting in the resistance profile in limiting metal environments.
Most acquired MβLs, such as enzymes from the IMP, VIM and NDM families present many allelic variants, in contrast to SPM-1 [36]. This difference could be due to the fact that SPM-1, as we demonstrate here, is optimized to meet specific Pseudomonas requirements, in contrast to the other MβL genes, present in many different genera of bacteria.
In a broader perspective, our approach allowed us to investigate how resistance determinants adapt to specific host requirements, linking fine details of the structural and biophysical features of the enzymes with bacterial fitness. More studies using this approach are required to account for the versatility and adaptability of MβLs to overcome the challenge imposed by new antibiotics.
Rabbits were housed and treated according to the policies of the Canadian Council on Animal Care guidelines on: antibody production (http://www.ccac.ca/Documents/Standards/Guidelines/Antibody_production.pdf). All efforts were made to minimize suffering and the procedures were approved by the Bioethics Commission for the Management and Use of Laboratory Animals inside the Science and Technical Committee of the University of Rosario, under resolution number 490/2012 (PICT-2008-N°0405).
Escherichia coli DH5α (Gibco- BRL, Gaithersburg, MD, U.S.A.) was used for construction of pΔEP-SPM-1 plasmid. Pseudomonas aeruginosa 48-1997A, originally identified in Brazil, was provided by M. Castanheira and M. Toleman [13], and used as the source of blaSPM-1. Laboratory strain P. aeruginosa PAO was used for transformation of mutant libraries, microbiological and biochemical studies. All strains were grown aerobically at 37°C in lysogeny broth (LB) medium supplemented with antibiotics when necessary.
Molecular biology procedures were done according to Sambrook et al. Transcriptional unit of blaSPM-1 was PCR-amplified from a genomic preparation of P. aeruginosa 48-1997A using primers SPM-1-fw and SPM-1-rv (Table 2), both containing a BamHI restriction site, and subcloned into pBBR1-MCS5 plasmid [22]. The product was digested with XhoI and SmaI enzymes (Promega) to eliminate restriction sites EcoRI and PstI from the MCS of the plasmid. Extremes were made blunt by treatment with Klenow fragment (Promega) and then ligated with T4 DNA ligase (Promega). Restriction sites EcoRI and PstI were introduced at the each edge of SPM-1 coding sequence by mutagenesis using primers EcoRI-fw, EcoRI-rv, PstI-fw and PstI-rv (Table 2). The resultant plasmid, pΔEP-SPM-1, was introduced into P. aeruginosa PAO by electroporation as described [37].
All constructs and amplifications were verified by sequencing at the University of Maine (Orono, USA).
Codons corresponding to positions 84 and 121 (BBL numbering [35]) of SPM-1 were randomized individually by Overlap Extension PCR, as previously described [23], [24]. Mutagenic primers were designed so as to contain random trinucleotides at the desired positions (S84X-fw, S84X-rv, G121X-fw y G121X-rv, Table 2), using pΔEP-SPM-1 as the template [23], [24]. The products were subcloned into pΔEP-SPM-1 through EcoRI and PstI restriction sites (thus avoiding unwanted mutations in promoter or terminator during PCR reactions), and the ligation mixtures electroporated in P. aeruginosa PAO. Electrocompentents from each mutant library (S84X and G121X) were spread in LB-agar plaques containing 30 µg/ml gentamicin, then collected and stored at −80°C. Library of double mutants S84X/G121X was constructed by submitting a plasmid preparation from S84X library to codon randomization of position 121, in the same way as before.
Selection of mutants capable of conferring some degree of resistance towards β-lactam antibiotics was done as follows. LB-agar plaques were inoculated with a bacterial culture (O.D. 0.1) of each mutant library, and disks embedded with 10 µg imipenem, 30 µg ceftazidime, 1000 µg cefoxitin, or 10 µg piperacillin placed on top of the agar. Mutant clones growing in the area of the antibiotic gradients were picked and the sequence of blaSPM-1 further determined [23].
Production and/or resistance levels of SPM-1 in P. aeruginosa PAO pΔEP-SPM-1 or P. aeruginosa 48-1997A were assayed by pairing disks embedded with 1.5 mg dipicolinic acid (DPA) with disks containing 10 µg imipenem, 30 µg ceftazidime or 30 µg cefepime, onto LB-agar plaques inoculated with the corresponding bacterial culture (O.D. 0.1) [25], [38]. Minimal inhibitory concentrations (MICs) were determined on plaque by the dilution method [38].
P. aeruginosa PAO crude extracts were obtained through sonication of cells washed in Tris 10 mM, MgCl2 30 mM pH 7.3 followed by centrifugation at 4°C. Periplasmic preparations of P. aeruginosa PAO were obtained by shock with chloroform as previously described [39]. Contamination of periplasmic extracts with cytoplasmatic proteins was discarded by Western-blot with antibodies against cytoplasmatic DnaK [27].
Levels of periplasmic wild type SPM-1 and mutants were determined by Western-blot of periplasmic extracts with polyclonal antibodies against SPM-1 (obtained after inoculating a rabbit with a mixture of recombinant SPM-1 and Freund's adjuvant) and immunoglobulin G-alkaline phosphatase conjugate. Protein band intensities were quantified with the Gel-Pro Analyzer 4.0 software (Exon-Intron, Inc.) and normalized to a bacterial periplasmic protein arbitrarily chosen.
Initial rates of hydrolysis were measured in a JASCO V550 spectrophotometer at 30°C in 300 µl of reaction media containing 300 µM of substrate and 10 µl of P. aeruginosa PAO periplasmic or crude extract in 10 mM Tris, 30 mM MgCl2 at pH 7.3. For comparison, hydrolytic activities of periplasmic extracts were made relative to the amount of SPM-1 or mutant present in the extract, estimated by Western-blot anti-SPM-1 of the extracts normalized as before.
In order to study the contribution of SPM-1 in whole β-lactam activity, crude extracts (normalized in total protein concentration by Bradford assay [40]) were incubated during 20 minutes at room temperature with and without addition of 25 mM DPA, and initial rates measured and compared.
Aliquots from each periplasmic extract of P. aeruginosa PAO were incubated for 5 minutes at various temperatures in the range 30–90°C, and then placed on ice for (a) determining initial rates of hydrolysis against ceftazidime, or (b) determining the levels of soluble SPM-1 or mutants (as before by Western-blot anti-SPM-1 of normalized extracts) after centrifugation for 10 min at 10,000 rpm and 4°C. Activity curves or soluble protein fraction as a function of temperature was adjusted to the sigmoid curve f = y0+a/(1+exp(−(x−x0)/b)) in Sigma Plot 9.0 program, with x0 the apparent melting temperature.
In order to generate apo-derivatives of periplasmic SPM-1 and mutants, periplasmic fractions of P. aeruginosa PAO were dialyzed in duplicate against 500 mM EDTA, 500 mM DPA, 50 mM Tris pH 8, then 2M NaCl, 50 mM Tris at pH 8, and finally 10 mM Tris, pH 7.3 30 mM MgCl2. The solutions were previously treated with chelating ion exchange resin (Chelex 100, Sigma-Aldrich) and dialysis times were of 6 hours.
All simulations were performed in AMBER [41] starting from the crystal structure of SPM-1 determined with resolution of 1.9 Å (PDB code 2FHX) [19]. As crystallization of SPM-1 was achieved with a vacant Zn2 site, the metal site structure of SPM-1 was reconstructed by aligning it to the geometry of the Zn2 site of the homologous enzyme B. cereus BcII (PDB code 1BC2) [42]. In this way, a starting structure with a complete active site was obtained. Each simulation was performed using monomeric wild type SPM-1, or mutant proteins G121S, G121A, S84D/G121S modified in silico. Furthermore, three crystallographic azide molecules were replaced by water molecules (Wt1, Wt2 and Wt3) in the cavities present at the base of SPM-1 active site.
The systems were immersed in a box of water molecules TIP3P [43] and were simulated using periodic boundary conditions and Ewald sums for treating long-range electrostatic interactions [44]. The SHAKE algorithm was applied to all hydrogen-containing bonds [45]. This allowed us to use a time step of 2 fs for integration of Newton equations. Parm99 and TIP3P force fields implemented in AMBER were used to describe the protein and water, respectively [41]. The force field of the active site (Zn,-OH, Asp, Cys and His) was taken from the literature [46]. The temperature and pressure were controlled by the Berendsen thermostat and barostat respectively, as implemented in AMBER [41]. Cut-off values used for the van der Waals interactions were 10 Å. The systems were first minimized to optimize possible structural crashes and then slowly heated from 0 to 300 K under constant volume conditions, using a time step of 0.1 fs. Finally, a short simulation was conducted at a constant temperature of 300 K and under constant pressure of 1 bar, using a time step of 0.1 fs, to allow the systems reach a suitable density. These balanced structures were the starting points for the 10 ns of molecular dynamics simulations.
|
10.1371/journal.pgen.1000752 | Pparγ2 Is a Key Driver of Longevity in the Mouse | Aging involves a progressive physiological remodeling that is controlled by both genetic and environmental factors. Many of these factors impact also on white adipose tissue (WAT), which has been shown to be a determinant of lifespan. Interrogating a transcriptional network for predicted causal regulatory interactions in a collection of mouse WAT from F2 crosses with a seed set of 60 known longevity genes, we identified a novel transcriptional subnetwork of 742 genes which represent thus-far-unknown longevity genes. Within this subnetwork, one gene was Pparg (Nr1c3), an adipose-enriched nuclear receptor previously not associated with longevity. In silico, both the PPAR signaling pathway and the transcriptional signature of Pparγ agonist rosiglitazone overlapped with the longevity subnetwork, while in vivo, lowered expression of Pparg reduced lifespan in both the lipodystrophic Pparg1/2-hypomorphic and the Pparg2-deficient mice. These results establish Pparγ2 as one of the determinants of longevity and suggest that lifespan may be rather determined by a purposeful genetic program than a random process.
| The progression of aging is controlled by both genetic and environmental factors. Many of these factors are present also in adipose tissue, which itself has been shown to determine lifespan. Applying advanced bioinformatics methods on a large mouse gene expression data set, we identified Pparg (Nr1c3), an important metabolic controller that regulates the expression of many other genes particularly in adipose tissue, to be associated with longevity. This association was verified in experimental mouse models where the lowered expression of Pparg reduced lifespan. In addition to Pparg, our analysis identified >700 potential novel aging genes in mouse adipose tissue. More generally, these findings suggest that lifespan may not be a random process but controlled by a purposeful genetic program.
| Aging is not a disease, but a natural evolution characterized by declining biological function, whose timeline is sensitive to both environmental and genetic factors. Several longevity candidate genes have been identified, including the insulin/IGF1 signaling pathway [1]–[3]. With the use of dietary regimens, such as caloric restriction (CR) and by modulating core body temperature, the control of energy metabolism has been implicated as a critical determinant of the aging phenotype [4]–[6]. A central physiological component of energy metabolism, involved in energy preservation, is the white adipose tissue (WAT), which has also been directly associated with the determination of lifespan [7],[8]. However, it is still uncertain whether WAT modulates aging via its ability to e.g. store fat, sensitize towards insulin, or produce adipocyte hormones. Also unknown is the nature of the involved genetic players and importantly, whether they function in a purposeful program or as random genetic events.
Using a systems approach we identified a novel subnetwork of genes in mouse WAT, which potentially impacts longevity, suggesting that aging is the result of a determined transcriptional network program and not entirely accidental. Furthermore, the most significantly enriched biological pathway revealed within this aging subnetwork was the PPAR signaling pathway. The aging subnetwork also contained the nuclear receptor Pparg (Nr1c3), a transcription factor well associated with adipocyte biology [9],[10], but whose contribution to longevity has not been previously assessed. In this study, we support our network theory of aging by demonstrating a significantly altered lifespan in 2 independent genetic mouse models expressing reduced levels of Pparg. Thus, in addition to providing novel candidate ‘longevity genes’ such as Pparg2, this study also provides further insight into the potential role of WAT biology and genetics as determinants of lifespan.
We hypothesized that the age-dependent physiological remodeling that leads to phenotypic aging is caused by concerted changes in a longevity-determining genetic network rather than by random changes at the level of individual genes. This hypothesis was tested using a mouse transcriptional network that consists of a union of 4 individual Bayesian networks of predicted causal regulatory interactions in the WAT generated from individual F2 crosses. We interrogated this network of 13088 genes with a seed set of 60 genes, derived from public resources, which either increase or reduce lifespan when genetically perturbed in the mouse (Table S1). Out of these 60 ‘known’ longevity genes, 33 were also present within the adipose tissue network (Table S1; Figure 1A). The pair-wise shortest path analysis against 106 randomly selected sets of 33 genes showed that these 33 genes on average were much more tightly connected than expected by chance (p = 0.00149) (Figure 1B). Furthermore, the distribution of the shortest paths within the set of 33 ‘known’ longevity genes was significantly tighter than that for the randomly selected sets as >99% of all Kolmogorov-Smirnov two-sided test p-values were less than 0.05. This tight, non-random interconnection of known aging-linked genes suggests that the associated biological phenomena are deliberate such that other ‘unknown’ age-related genes and/or biological processes may be predicted. This network theory is reminiscent of the transcriptional consequences of single genetic perturbations, such as knock-out mouse models or DNA polymorphisms, which result in concentrations of transcriptional changes in the genes functionally relating to the perturbed gene rather than altering genes diffusely distributed across the whole network [11]. Following the concept of using the ‘known’ to discover the ‘unknown’, we thus expanded the subnetwork beyond the 33 longevity genes to other genes most highly connected to them, and obtained a larger subnetwork, containing 742 genes (Table 1, Table S2). By assigning importance to the closeness of connection with known longevity genes, we were thus able to suggest several hundred additional genes that may influence the aging process. One such example, among the top 20 genes for the closeness of connectivity with the 33 ‘known’ longevity genes (Table 1), was the eukaryotic translation initiation factor 4E (eIF4E) binding protein 1 (Eif4ebp1, or 4E-BP1) which, in the unphosphorylated state, represses mRNA translation by binding to eIF4E. Since it regulates adipogenesis and metabolism [12], and one of the mediators of its phosphorylation is insulin signaling [13], Eif4ebp1 can be linked to the established effects of insulin signaling on longevity. Moreover, in Drosophila 4E-BP plays an important role in lifespan extension upon dietary restriction [14]. Eif4ebp1 has furthermore been identified as a “funnel factor” in cancer, through which several oncogenic pathways converge [15].
Biological pathway enrichment analysis is a powerful tool to uncover functional associations within an a priori selected set of genes. When applied to the aging subnetwork of 742 genes (excluding the 33 ‘known’ longevity genes from the full set of 775 genes to eliminate bias), significant enrichment was revealed in several ontology classes with established links to aging such as complement and coagulation cascade (i.e. inflammation), insulin signaling, and ubiquinone pathway (i.e. oxidative stress) (Table 2). Importantly, however, several pathways lacking previously demonstrated association with longevity also appeared among the significantly enriched ontologies. One of these, the PPAR signaling pathway, was actually ranked the highest for the enrichment of all potential longevity genes. Although direct in vivo evidence linking Ppars to aging are scarce, conceptual evidence does exist [16],[17], including links to age-related changes in inflammatory response, insulin sensitivity, distribution and proportion of body fat, oxidative stress [18], and fatty acid oxidation rate. Of the three actual Ppar family members, the only one that was present within the aging subnetwork was Pparγ (Table S2). Given that signaling through Pparγ is also of vital importance to proper adipose tissue development and function [9], [10], [19]–[21], and that Pparγ is regulated in WAT by one of the best established longevity determinants, mammalian SIR2 orthologue sirtuin 1 (Sirt1) [22], we hypothesized that perturbing Pparγ signaling might affect longevity.
We first tested this hypothesis in silico by using the WAT gene expression signature generated from mice with chemically modulated Pparγ activity through the administration of the Pparγ agonist, rosiglitazone [11]. Notably, 213 out of the 1669 genes whose transcriptional expression was altered by Pparγ activation, overlapped with the genes in the aging subnetwork at a very high significance level (p = 5.2028*10−30) (Table S2). This finding thus validates the association of Pparg with the aging subnetwork and further implicates it as a potential determinant of the aging phenotype.
To put this hypothesis to further rigorous in vivo testing, we investigated the role of Pparg in longevity in two mouse models with genetically altered levels of Pparg expression: the hypomorphic Pparg1/2 knock-out mouse, which lacks Pparg exclusively in WAT (Figure S1A) and is severely lipodystrophic and remains insulin resistant throughout life [19]; and the Pparg2 deficient mouse that lacks Pparγ2 in all tissues (Figure S1B) and shows some features of moderate lipodystrophy and insulin resistance at a young age [23], but which fully compensates upon aging (see below). The nearly complete knockdown of Pparg1 and Pparg2 in the WAT of male Pparghyp/hyp mice resulted in a reduction in lifespan by approximately 16 weeks when compared to the wild type mice (93.7±4.4 vs 109.6±3.4 weeks, p = 0.03) (Figure 2A). In some respects this observation goes against the prediction that reduced fat mass, as seen during CR [4],[5], would increase longevity; however, if the known insulin sensitizing effects of Pparγ were key to mediating the effects of CR, then one would expect reduced longevity in the Pparghyp/hyp mice, where whole body insulin resistance is prominent. However, one potentially confounding factor in this experiment is the profound lipodystrophy exhibited by the Pparghyp/hyp mice, which may not represent ‘normal’ metabolic environment due to the amount of metabolic compensation by the upregulation of other signaling pathways that these mice need for survival [19]. Also, although differences in the amount of gross tumors were not observed upon macroscopic necropsy, we can not exclude the possible contribution of more discrete tumors to the decreased longevity of the Pparghyp/hyp mice. Interestingly though, the males of an equally lipodystrophic A-ZIP/F-1 mouse model have more than 40% mortality rate before 30 weeks of age [8], in comparison to the Pparghyp/hyp mice which survived 85% of the average ∼2 year lifespan of wild type mice. In this sense, Pparghyp/hyp mouse model is one of the longest living severely lipodystrophic models reported.
In order to assess more directly the effects of Pparg on longevity, without the added complication of reduced adiposity or insulin sensitivity, we made use of Pparg2−/− mice that we generated in the laboratory and which lack Pparγ2, the WAT enriched Pparγ isoform, in all tissues. Although young Pparg2−/− mice are lean [23], our ∼2 year old Pparg2−/− mice had the same total and lean body mass, body fat content (Figure S2A and S2B), and caloric intake (12.33±1.53 vs. 14.24±1.53 kcal/day/mouse, p = 0.421) as their age-matched littermate controls. Young Pparg2−/− mice have also been reported to be insulin resistant [23]. Again in contrast, there were no differences in glucose tolerance, the HOMA index for insulin resistance, nor in circulating insulin or adiponectin levels between our Pparg2−/− and Pparg2+/+ mice at ∼2 years of age (Figure S2C, S2D, S2E, S2F). Thus, our aging Pparg2−/− mice represent a very metabolically ‘clean’ model for investigating the role of Pparg2 in longevity.
Consistent with reduced longevity in the Pparghyp/hyp mouse, we noted a significant decrease in lifespan in Pparg2−/− mice. The female Pparg2−/− mice lived, on average, 8.8 weeks less than their wild type controls (p = 0.02 when limiting the analysis to those living no more than 120 weeks), although this difference seemed to disappear towards extreme age (Figure 2B). Gross morphological differences that could contribute to mortality were not observed between the genotype groups, although again the contribution of more discrete tumors can not be excluded. Since the Pparg2−/− mice had reduced longevity, comparable to that in Pparghyp/hyp mice, but were not lipodystrophic or insulin resistant, our observations point more towards a specific role for Pparγ2 and any of its downstream pathways in the regulation of longevity, rather than mere changes in fat content and/or insulin signaling. Together our studies thus reveal another genetic factor, Pparg2, that affects the basic mechanisms of aging, independent of changes in fat mass or insulin sensitivity [1],[2],[7]. Interestingly, a potential molecular mechanism linking aging and Pparγ has recently been suggested to involve a steroid receptor coactivator-1 (SRC-1) as the age-induced loss of PPARγ/SRC-1 interactions increased the binding of PPARγ to the promoter of a model adipogenic gene for fatty acid binding protein 4 (FABP4, also called aP2) [24].
Both our in silico and in vivo results in the mouse tie longevity tightly together with signaling through Pparγ, and especially the Pparγ2 isoform. We have recently shown increased longevity in knock-in mice carrying the Ala12 allele of the common human genetic variant Pro12Ala variant of PPARG2 [25], which associates with leanness and improved insulin sensitivity in both man and mouse [25]–[27]. The species gap between mice and humans for the role of Pparγ2 in longevity is bridged by the observation that lifespan is increased also in human carriers of the Ala12 allele of the Pro12Ala variant of PPARG2 [28]. In the clinical setting, therefore, the links we show between longevity and both Pparg and the rosiglitazone signature suggest that thiazolidinediones [29] (TZDs), like rosiglitazone or pioglitazone which are widely used Pparγ agonists and insulin sensitizers in the treatment of type 2 diabetes mellitus (T2D), could be beneficial for longevity. On the face of it, this may in fact seem paradoxical, considering that impaired insulin signaling through insulin receptor or its substrates increases, rather than decreases lifespan in a number of mouse models [1],[2],[7]. However, this can be reconciled by the fact that these models are primarily protected from the detrimental effects of age-induced increase in plasma insulin levels as TZDs lower circulating insulin levels [30],[31]. Fittingly, low insulin levels and maintained insulin sensitivity characterize human centenarians [32]. In light of the above, the results from ongoing outcome trials evaluating the long-term health benefits of treatments with PPARγ-agonists, i.e. TZDs, are eagerly awaited.
In summary, we have identified a substantial set of potential novel longevity genes in mouse adipose tissue, and demonstrate, as a case study, the significant effects of perturbed Pparγ activity on mouse lifespan. Furthermore, our network analysis suggests that, at least in the context of adipose tissue, the determination of longevity may not be a random process, but governed by a concerted effort of a distinct subnetwork of genetic players.
Animal experiments were approved by the local ethics committee and performed according to governmental guidelines.
To obtain a list of genes with known association to longevity, we used the Phenotypes section of the Mouse Genome Informatics (MGI) resource of The Jackson Laboratory (http://www.informatics.jax.org/) [33], the GenAge Model Organisms pages for mouse within The Human Ageing Genomic Resources (HAGR) [34], and a literature search. The list was compiled in October, 2007.
Detailed description of these methods is given in Text S1. In summary, we obtained male adipose tissue gene expression data from 4 different mouse F2 crosses [35],[36] using Agilent microarrays, and generated a Bayesian network for each cross by integrating genetic and gene expression data [37]–[39]. The combined network, containing 13088 nodes and 22809 edges, was obtained as the union of all these 4 separate Bayesian networks.
To assess the degree of connectivity of the 33 ‘known’ longevity genes that were present in the adipose consensus network, mean shortest paths were computed using Dijkstra's algorithm [40] for our set of 33 nodes ( = genes) as well as 106 randomly selected sets of 33 nodes. Briefly, the algorithm finds the smallest number of edges we have to “walk” in order to “travel” from a source node ( = gene) to another node ( = gene) of interest within the map/network. The probability of finding random sets of 33 nodes with shorter mean paths than with our set was obtained by counting the number of such eventualities within the randomized sets, and amounted to a p-value of 0.00149, demonstrating that indeed our 33 genes are much more connected within the adipose tissue consensus network than expected by change. Kolmogorov-Smirnov (KS) test was used to further assess whether there were any significant differences between the shortest path distribution within our longevity gene-set and those within each of the 106 random sets. The resulting p-value distribution demonstrated that indeed the longevity genes shortest path distribution is not a normal occurrence in the network.
Pparg2−/− mice were generated from Pparghyp/hyp [19] mice by successive matings with transgenic C57Bl/6J mice expressing FLP and Cre recombinases to remove the Pparg2 specific exon B. All mice studied were backcrossed a minimum of 9 generations to achieve an essentially pure C57Bl/6J background.
The original survival cohorts consisted of 38 wild type and 24 Pparghyp/hyp male, and 25 wild type and 26 Pparg2−/− female mice which were maintained on a 12 hour light/dark cycle, fed regular chow, had free access to H2O and received standard animal care. The mice were bred locally and were entered into the survival cohort over the course of 23 weeks for male Pparghyp/hyp mice, amd 19 months for female Pparg2−/− mice. For all groups, deaths were recorded weekly. Mice observed as moribund were euthanized and recorded as dead on that week. All Pparghyp/hyp reached the end-point, but a few Pparg2−/− mice survived at the time of analysis.
Approximately 2 year old wild type (n = 5) and Pparg2−/− (n = 9) mice were subjected to the following analysis according to standardized Eumorphia/EMPReSS (http://empress.har.mrc.ac.uk/) protocols: body composition by quantitative nuclear magnetic resonance on a Minispec analyzer (Bruker Optics, The Woodlands, TX), food intake, intraperitoneal glucose tolerance test (IPGTT), and fasting plasma insulin and adiponectin measurements using Ultrasensitive Mouse Insulin ELISA kit (Mercodia, Uppsala, Sweden) and Quantikine Mouse Adiponectin/Acrp30 Immunoassay (R&D systems Inc., Minneapolis, MN), respectively. HOMA index for insulin resistance was calculated from fasting glucose and insulin values [41].
Pparg1 and Pparg2 gene expression in WAT, BAT, liver and skeletal muscle of Pparghyp/hyp mice was previously reported [19] and is presented for comparative purposes. For Pparg2−/− mice, total RNA was extracted from WAT, BAT, liver and skeletal muscle either with RNeasy for Lipid Tissues Mini Kit (Qiagen, Valencia, CA) or Trizol (Invitrogen, Carlsbad, CA), and reverse transcribed to cDNA using SuperScript II System (Invitrogen) and random hexamer primers. Pparγ1 and Pparγ2 gene expression was quantified by qRT-PCR using isoform-specific primers and SYBR Green chemistry on a LightCycler 480 (Roche, Penzberg, Germany).
Statistical methods pertaining to the network and other associated analysis of gene expression and gene set data were as detailed above. Kaplan-Meier survival analysis, which allows for censored cases, was used to analyze the survival data in SPSS (version 14). Metabolic and molecular data for Pparghyp/hyp and Pparg2−/− mice were analyzed using Student's t-test and are presented as means ± s.e.m.
|
10.1371/journal.pgen.1007264 | MKLN1 splicing defect in dogs with lethal acrodermatitis | Lethal acrodermatitis (LAD) is a genodermatosis with monogenic autosomal recessive inheritance in Bull Terriers and Miniature Bull Terriers. The LAD phenotype is characterized by poor growth, immune deficiency, and skin lesions, especially at the paws. Utilizing a combination of genome wide association study and haplotype analysis, we mapped the LAD locus to a critical interval of ~1.11 Mb on chromosome 14. Whole genome sequencing of an LAD affected dog revealed a splice region variant in the MKLN1 gene that was not present in 191 control genomes (chr14:5,731,405T>G or MKLN1:c.400+3A>C). This variant showed perfect association in a larger combined Bull Terrier/Miniature Bull Terrier cohort of 46 cases and 294 controls. The variant was absent from 462 genetically diverse control dogs of 62 other dog breeds. RT-PCR analysis of skin RNA from an affected and a control dog demonstrated skipping of exon 4 in the MKLN1 transcripts of the LAD affected dog, which leads to a shift in the MKLN1 reading frame. MKLN1 encodes the widely expressed intracellular protein muskelin 1, for which diverse functions in cell adhesion, morphology, spreading, and intracellular transport processes are discussed. While the pathogenesis of LAD remains unclear, our data facilitate genetic testing of Bull Terriers and Miniature Bull Terriers to prevent the unintentional production of LAD affected dogs. This study may provide a starting point to further clarify the elusive physiological role of muskelin 1 in vivo.
| Lethal acrodermatitis (LAD) is an autosomal recessive hereditary disease in dogs. It is characterized by poor growth, immune deficiency and characteristic skin lesions of the paws and of the face. We mapped the LAD locus to a ~1.11 Mb segment on canine chromosome 14. Whole genome sequence data of an LAD affected dog and 191 controls revealed a candidate causative variant in the MKLN1 gene, encoding muskelin 1. The identified variant, a single nucleotide substitution, MKLN1:c.400+3A>C, altered the 5’-splice site at the beginning of intron 4. We experimentally confirmed that this variant leads to complete skipping of exon 4 in the MKLN1 mRNA in skin. Various cellular functions have been postulated for muskelin 1 including roles in intracellular transport processes, cell morphology, cell spreading, and cell adhesion. Our data from dogs reveal a novel in vivo role for muskelin 1 that is related to the immune system and skin. MKLN1 thus represents a novel candidate gene for human patients with unsolved acrodermatitis and/or immune deficiency phenotypes. LAD affected dogs may serve as models to gain more insights into the function of muskelin 1.
| Acrodermatitis enteropathica in humans (OMIM #201100) is an inherited disorder of zinc metabolism. Affected patients display an inflammatory rash, diarrhea and a general failure to thrive [1–3]. This disease is caused by variants in the SLC39A4 gene encoding a zinc transporter that mediates the uptake of dietary zinc in the gut. Clinical signs in patients will ameliorate or even resolve upon oral supplementation with zinc [4]. A similar SLC39A4 associated hereditary zinc deficiency exists in cattle [5].
In Bull Terriers, a related phenotype termed lethal acrodermatitis (LAD) has been reported in the scientific literature as early as 1986 [6]. LAD is inherited as a monogenic autosomal recessive trait. Affected puppies show characteristic skin lesions on the feet and on the face, diarrhea, bronchopneumonia, and a failure to thrive. The skin lesions consist of erythema and tightly adherent scales, erosions or ulcerations with crusts involving primarily the feet, distal limbs, elbows, hocks, and muzzle. Later on, hyperkeratosis of the footpads and deformation of the nails occur. LAD affected dogs also show a coat color dilution in pigmented skin areas. An abnormally arched hard palate impacted with decayed, malodorous food is a characteristic clinical marker for the disease (Fig 1) [6–8].
LAD dogs are immunodeficient with a reduction in serum IgA levels and frequently suffer from skin infections with Malassezia or Candida [9,10]. LAD manifests clinically in the first weeks of life. Affected puppies typically die before they reach an age of two years, either due to infections such as bronchopneumonia or because they are euthanized when their paw pad lesions become very severe and painful. They grow slower than their non-affected littermates and at the age of one year have about half the body weight and size of an unaffected dog [8]. Some, but not all studies found reduced levels of zinc in the serum of LAD affected dogs [6,8,11]. In contrast to acrodermatitis enteropathica in humans, oral or intravenous supplementation of zinc does not lead to an improvement of the clinical signs in LAD affected dogs [6]. A proteomic analysis reported changes related to inflammatory response in the liver of LAD affected puppies [12].
In the present study, we performed a genome-wide association study (GWAS) followed by a whole genome sequencing approach to unravel the causative genetic variant for LAD in Bull Terriers and Miniature Bull Terriers.
We performed a GWAS with genotypes from 78 Bull Terriers and Miniature Bull Terriers. After quality control, the pruned dataset consisted of 22 LAD cases, 48 controls and 76,419 markers. We obtained a single strong association signal with 57 markers exceeding the Bonferroni-corrected genome-wide significance threshold after adjustment for genomic inflation (PBonf. = 6.5 x 10−7). All significantly associated markers were located on chromosome 14 within an interval spanning from 0.9 Mb– 10.6 Mb. The three top-associated markers all had a P-value of 1.4 x 10−9 and were located between 5.2 Mb– 5.9 Mb on chromosome 14 (Fig 2).
To narrow down the identified region, we visually inspected the genotypes of the cases to perform autozygosity mapping. We searched for homozygous regions with allele sharing and found one region of ~1.11 Mb, which was shared between all 22 cases. The critical interval for the causative LAD variant corresponded to the interval between the first flanking heterozygous markers on either side or chr14:5,248,244–6,355,383 (CanFam 3.1 assembly).
We sequenced the genome of an affected Bull Terrier at 24x coverage and called single nucleotide variants (SNVs) and small indel variants with respect to the reference genome (CanFam 3.1). We then compared these variants to whole genome sequence data of 3 wolves and 188 control dogs from genetically diverse breeds. This analysis identified five private homozygous variants in the critical interval in the affected dog (Table 1, S1 Table).
Four of these five variants were intergenic and classified as “modifier” by the SNPeff software. The remaining fifth variant was located within the 5’-splice site of intron 4 of the MKLN1 gene and its SNPeff impact prediction was “low”. The formal designation of this variant is chr14:5,731,405T>G or MKLN1:c.400+3A>C (S1 Fig).
We confirmed the presence of this variant by Sanger sequencing (Fig 3). As MKLN1:c.400+3A>C represented the only plausible candidate causative variant, we genotyped 251 Bull Terriers, 89 Miniature Bull Terriers, and 462 dogs from 62 other breeds for this variant (Table 2). The variant showed perfect association with the LAD phenotype in Bull Terriers and Miniature Bull Terriers (PFisher = 4.8 x 10−58). All 46 available cases were homozygous for the variant, whereas the unaffected dogs were either homozygous wildtype or heterozygous. The test dogs included a subset of unaffected Bull Terriers and Miniature Bull Terriers from Finland, which were not specifically collected for this study and therefore considered representative for the general population. The 166 Finnish dogs contained 37 heterozygous dogs (22%). The variant was not found in any of the tested dogs from other breeds.
To assess the putative impact of the MKLN1 variant on splicing, we analyzed the frequency of the wildtype and mutant sequence motifs in a compilation of 186,630 human 5’-splice sites [13,14]. The canine wildtype sequence TAGgtaagg was identical to the sequence of 276 human 5’-splice sites, while the mutant sequence motif TAGgtcagg occurred in only 3 human 5’-splice sites. The very low frequency of the mutant sequence motif suggested that MKLN1:c.400+3A>C might affect the efficacy of the splicing process. Several other pathogenic A>C transversions at 5’-splice sites’ position +3 with subsequent exon skipping have been described in the literature [15–17].
We experimentally analyzed MKLN1 transcripts in skin RNA from an LAD affected dog with the homozygous mutant C/C genotype in comparison to a healthy control dog (A/A genotype). RT-PCR with primers located at the exon 2/3 and exon 5/6 boundaries yielded a cDNA fragment of the expected size in the control dog, but not in the LAD affected dog. In the LAD affected dog, a very clean cDNA amplicon lacking exon 4 was obtained. This experiment demonstrated a complete skipping of exon 4 in MKLN1 transcripts as consequence of the genomic MKLN1:c.400+3A>C variant (r.312_400del89; Fig 4). If translated, the mutant transcript was predicted to result in a severely truncated protein containing only the first 105 of a total of 735 amino acids of the wildtype protein (p.(Gly105SerfsTer10); S2 Fig).
In the present study we identified a splice defect in the canine MKLN1 gene in Bull Terriers with LAD. The combination of GWAS and haplotype analysis localized the causative variant to a relatively small chromosomal region with only a few characterized genes including MKLN1. The splice region variant in MKLN1 was the only plausible variant within this critical interval that showed the expected genotype concordance with the LAD phenotype in a large cohort of more than 300 Bull Terriers and ~500 dogs from other breeds.
The identified MKLN1:c.400+3A>C variant resulted in exon 4 skipping and a frameshift as 89 nucleotides were missing from the mutant transcripts. It therefore seems likely that mutant transcripts are degraded by nonsense-mediated mRNA decay. Considering the strong genetic association of the variant with the phenotype and the fact that we demonstrated a functional defect on the MKLN1 transcript level, we think that our data strongly suggest the causality of the MKLN1:c.400+3A>C variant for LAD in Bull Terriers and Miniature Bull Terriers.
MKLN1 encodes the widely expressed intracellular protein muskelin 1, also known as TWA2. The function of muskelin 1 is only partially understood. It was originally described as a protein that mediates adhesive and cell-spreading responses to thrombospondin 1, an extracellular matrix adhesion molecule [18]. However, different studies suggested that the function of muskelin 1 goes beyond this pathway and are also supported by the fact that muskelin 1, which has homologs in invertebrates and even fission yeast, evolved earlier than the vertebrate-specific thrombospondin 1 [19,20]. Muskelin 1 is a multidomain protein with an N-terminal discoidin domain, a LisH / CTLH tandem domain, and six C-terminal Kelch repeats, which forms homotetramers [21]. The LisH domain was shown to be crucial for muskelin 1 dimerization and cytoplasmic localization, and, together with the head-to-tail interaction via the discoidin domain, also for the tetramerization of muskelin 1 [20,21].
Consistent with its multidomain structure and ubiquitous expression, diverse binding partners have been reported for muskelin 1. It binds prostaglandin EP3 receptor isoform α [22] and heme-oxidase 1, which counteracts inflammatory and reactive oxygen species induced damage [23]. It is part of the CTLH complex, the homolog of yeast E3 ubiquitin ligase, where it binds to RanBPM and Twa I [24–26] and interacts with the cardiogenic transcription factor TBX-20 [27]. In the rat lens, muskelin 1 is a substrate of Cdk5 and interacts with the Cdk5 activator p39 [28]. Also in lens, it was shown that p39 links muskelin 1 to myosin II and stress fibers [29].
Mkln1-/- knockout mice are viable and do not have skin lesions comparable to those in Bull Terriers with LAD. However, they exhibit a subtle coat color dilution phenotype similar to that seen in LAD affected dogs. In these Mkln1-/- knockout mice, muskelin 1 was identified as a protein required for GABAA receptor endocytosis and trafficking in neurons via direct interaction with the α1 subunit of GABAA receptors and the motor proteins dynein and myosin VI. The dilute coat color of Mkln1-/- knockout mice suggested that muskelin is a trafficking factor involved in several different intracellular transport processes, possibly including melanosome transport [30].
The lacking skin lesions in Mkln1-/- knockout mice raise the questions whether muskelin 1 depletion does not result in disease in mice; whether their clinical signs would only manifest at a (much) older age; or whether the sterile environment of the laboratory animals prevented infections and thus the development of skin lesions. In the latter case, LAD would be a primary immunodeficiency disorder, in agreement with the observation of lower IgA levels and higher susceptibility to microbial infection in LAD affected dogs [8–10, 12]. Given the diverse known protein-protein interactions of muskelin 1, it is however likely that absence of muskelin 1 leads to dysfunctions beyond the immune system.
In humans, an intronic SNV in MKLN1 was associated with urinary potassium excretion in Korean adults and another intronic MKLN1 SNV with early bipolar disorder [31,32]. Furthermore, MKLN1 has been associated with asthma in independent GWASs. A SNV in MKLN1 ranked among the top 100 SNVs associated with childhood asthma in a study sample of 429 affected-offspring trios from a European American population [33]. A different SNV in the 5’-UTR of MKLN1 was associated with asthma in a population including patients with severe or difficult-to-treat asthma [34].
In the ExAC database, only one MKLN1 missense, but no nonsense, frameshift or splice site variants present in a homozygous state were found [35,36]. Furthermore, the probability of loss of function (LoF, specified as nonsense, splice acceptor, and splice donor variants) tolerance was estimated to be 1.00, indicating that the MKLN1 gene is extremely LoF intolerant [37]. Therefore, it is conceivable that loss of function variants on both alleles might lead to severe phenotypes in humans.
To our knowledge, no link between muskelin 1 and zinc or copper metabolism has been reported to date. While acrodermatitis enteropatica in humans and acrodermatitis in cattle clinically resemble LAD in dogs, these diseases may be caused by completely different molecular mechanisms. The fact that findings on zinc levels in the few published studies on LAD affected dogs were contradictory and zinc supplementation did not lead to improvement of lesions [6] support this hypothesis.
In conclusion, we identified the MKLN1:c.400+3A>C variant leading to a splice defect in the MKLN1 gene as candidate causative variant for LAD in Bull Terriers and Miniature Bull Terriers. The molecular pathogenesis of LAD remains unclear. Our data facilitate genetic testing of Bull Terriers and Miniature Bull Terriers to prevent the unintentional breeding of LAD affected dogs. LAD affected dogs may serve as models to further clarify the elusive physiological role of muskelin 1 in vivo.
All animal experiments were performed according to the local regulations. The dogs in this study were examined with the consent of their owners. The study was approved by the “Cantonal Committee For Animal Experiments” (Canton of Bern; permits 22/07, 23/10, and 75/16).
Bull Terriers with their characteristic egg-shaped head were founded as a dog breed in the 1850s in the United Kingdom. Originally, there were no size standards in this breed and smaller dogs were bred as a variety of the regular Bull Terrier. Eventually, two sub-populations formed and the Miniature Bull Terrier with a maximum height of 35.5 cm was recognized as an independent breed in 1991 by the American Kennel Club (AKC) and in 2011 by the European Fédération Cynologique Internationale (FCI). Therefore, Bull Terriers and Miniature Bull Terriers share a common ancestral gene pool, but represent independent closed populations today.
This study included samples from 251 Bull Terriers (41 LAD cases / 210 controls) and 89 Miniature Bull Terriers (5 LAD cases / 84 controls). Case/controls status was based on owners’ reports. We additionally used 462 dogs from 62 breeds, which were assumed to be free of the disease allele (S3 Table). Skin biopsies were taken from two LAD affected Bull Terriers from toe, nose, lip, and forearm and fixed in 10% buffered formalin for 24 hours. Biopsies were processed, embedded in paraffin and sectioned at 4 µm. Skin sections were stained with hematoxylin and eosin. The histopathology was performed by veterinary pathologists (BR, Dipl.-ECVP, and SH). Two further biopsies from comparable sites of the same dogs were submerged in RNAlater solution for subsequent RNA isolation.
We isolated genomic DNA from EDTA blood samples. Seventy-eight dogs were genotyped for either 173,662 or 218,256 SNVs on the illumina canine_HD chip. The raw SNV genotypes are available at https://www.animalgenome.org/repository/pub/BERN2017.1208/.
The initial dataset consisted of 78 dogs and 220,853 markers. Using Plink version 1.9 [38] we excluded markers that were not located on autosomes or the X chromosome (n = 2,327) and markers with a genotyping rate lower than 90% (n = 49,814). Using the R package GenABEL [39] and the command “check.markers”, dogs with a call rate < 90% (n = 3), ibs > 95% (n = 0), high individual heterozygosity (FDR = 0.01) (n = 1, included in dogs with low call rate) as well as markers with a maf < 1% (n = 55,931) and a genotyping rate < 90% (n = 10,951) were excluded. Five outliers in the multidimensional scaling plot based on a genomic distance matrix were also removed. In a second quality control step, markers deviating from Hardy-Weinberg equilibrium (FDR = 0.2) in controls (n = 1,239), markers with a genotyping rate < 90% (n = 0) and maf < 1% (n = 26,215) were excluded, resulting in a final dataset of 70 dogs (22 cases, 48 controls) and 76,419 markers. A polygenic model of the hglm package [40], with a kinship matrix based on autosomal markers in the cleaned dataset as random effect, was estimated and a score test for association using the function “mmscore” was performed. The genomic inflation factor was 1.16. We corrected for multiple testing using Bonferroni correction with a significance level of 0.05. QQ plots were created using qqman version 0.1.4 [41].
We visually inspected plink tped files for the region of interest on chromosome 14 using Excel and searched for homozygous regions with haplotype sharing in cases with a call rate >90%. The first flanking heterozygous markers on either side of the homozygous region in 22 cases defined the borders of the critical interval.
An Illumina PCR-free TruSeq fragment library with 350 bp insert size of an LAD affected Bull Terrier was prepared. We collected 219 million 2 x 150 bp paired-end reads or 24x coverage on a HiSeq3000 instrument. The reads were mapped to the dog reference genome assembly CanFam3.1 and aligned using Burrows-Wheeler Aligner (BWA) version 0.7.5a [42] with default settings. The generated SAM file was converted to a BAM file and the reads were sorted by coordinate using samtools [43]. Picard tools (http://sourceforge.net/projects/picard/) was used to mark PCR duplicates. To perform local realignments and to produce a cleaned BAM file, we used the Genome Analysis Tool Kit (GATK version 2.4.9, 50) [44]. GATK was also used for base quality recalibration with canine dbsnp version 139 data as training set. The sequence data were deposited under the study accession PRJEB16012 and sample accession SAMEA4504844 at the European Nucleotide Archive.
Putative SNVs were identified in each of 192 samples (S2 Table) individually using GATK HaplotypeCaller in gVCF mode [45]. Subsequently all sample gVCF files were joined using Broad GenotypeGVCFs walker (-stand_emit_conf 20.0; -stand_call_conf 30.0). Filtering was performed using the variant filtration module of GATK using the following standard filters: SNVs: Quality by Depth: QD < 2.0; Mapping quality: MQ < 40.0; Strand filter: FS > 60.0; MappingQualityRankSum: MQRankSum < -12.5; ReadPosRankSum < -8.0. INDELs: Quality by Depth: QD < 2.0; Strand filter: FS > 200.0. The functional effects of the called variants were predicted using SnpEFF software [46] together with the NCBI annotation release 104 on CanFam 3.1. For the filtering of candidate causative variants in the case, we used 191 control genomes, which were either publicly available [47] or produced during other projects of our group or contributed by members of the Dog Biomedical Variant Database Consortium. A detailed list of these control genomes is given in S2 Table.
We used the dog CanFam 3.1 reference genome assembly for all analyses. Numbering within the canine MKLN1 gene corresponds to the accessions XM_005628367.3 (mRNA) and XP_005628424.1 (protein). Numbering within the human MKLN1 gene corresponds to the accessions NM_013255.4 (mRNA) and NP_037387.2 (protein).
We used Sanger sequencing to confirm the candidate variant MKLN1:c.400+3A>C and to genotype the dogs in this study. A 797 bp fragment containing the variant was PCR amplified from genomic DNA using AmpliTaq Gold 360 Master Mix (Life Technologies) and the primers CCATGCACTGTAGCCACATC and TGGAAAAGGTTCCACTTGAAAT. After treatment with shrimp alkaline phosphatase and endonuclease I, PCR products were directly sequenced on an ABI 3730 capillary sequencer (Life Technologies). We analyzed the Sanger sequence data using the software Sequencher 5.1 (GeneCodes).
RNA was extracted from skin samples using the RNeasy Fibrous Tissue Mini Kit (Qiagen). The tissue was first finely crushed by mechanical means using TissueLyser (Quiagen), and RNA was extracted by centrifugation following the instructions by the manufacturer. Total mRNA was reverse transcribed into cDNA using the SuperScript IV Reverse Transcriptase kit (Thermo Fisher) with oligo d(T) primers. A PCR on the synthesized cDNA was carried out using primer MKLN1_c_F2, CCTCCCCAGTACTTGATTCTG, located at the boundary of exons 2 and 3, and primer MKLN1_c_R2, TTCCTGTTCACGGTACTTGC, located at the boundary of exons 5 and 6 of the MKLN1 gene. The products were analyzed on a Fragment Analyzer capillary gel electrophoresis instrument (Advanced Analytical). The sequence of the obtained RT-PCR products was confirmed by Sanger sequencing as described above.
|
10.1371/journal.pgen.1005027 | The Nuclear Receptor DAF-12 Regulates Nutrient Metabolism and Reproductive Growth in Nematodes | Appropriate nutrient response is essential for growth and reproduction. Under favorable nutrient conditions, the C. elegans nuclear receptor DAF-12 is activated by dafachronic acids, hormones that commit larvae to reproductive growth. Here, we report that in addition to its well-studied role in controlling developmental gene expression, the DAF-12 endocrine system governs expression of a gene network that stimulates the aerobic catabolism of fatty acids. Thus, activation of the DAF-12 transcriptome coordinately mobilizes energy stores to permit reproductive growth. DAF-12 regulation of this metabolic gene network is conserved in the human parasite, Strongyloides stercoralis, and inhibition of specific steps in this network blocks reproductive growth in both of the nematodes. Our study provides a molecular understanding for metabolic adaptation of nematodes to their environment, and suggests a new therapeutic strategy for treating parasitic diseases.
| Animals adjust their internal biological processes in response to their environments. In this study, we report that in a nutrient rich environment the free-living nematode, Caenorhabditis elegans, induces an energy-generating metabolic pathway to govern its reproductive growth by activating the nuclear receptor, DAF-12. By responding to its endogenous ligands, called dafachronic acids, DAF-12 induces oxidation of lipids to produce the energy necessary to support growth and reproduction; and likewise, in the absence of dafachronic acids, DAF-12 prevents activation of this pathway. Through gene expression analysis, we show that DAF-12 regulates a network of genes involved in energy homeostasis and lipid metabolism. Given that dafachronic acids are produced only in well-fed worms, we conclude that DAF-12 functions as an environmental sensor that coordinately governs energy homeostasis. Through analogous studies in the incurable human parasite, Strongyloides stercoralis, we demonstrate that this pathway is conserved and that blocking it compromises the viability of the parasites. These findings elucidate a molecular mechanism for how nematodes govern their energy needs in response to the environment, and provide a potential new strategy for treating nematode parasitic diseases.
| The evolutionary success of nematodes is derived from their ability to adapt different developmental pathways depending on environmental conditions. The best studied of these pathways exists in the free-living nematode Caenorhabditis elegans (C. elegans). After hatching from eggs when enviromental conditions are favorable, C. elegans larvae continually develop through four stages (L1–L4), which eventually mature into reproductive adults. In contrast, in unfavorable environments, C. elegans larvae interrupt their reproductive growth by arresting at an alternative L3 stage termed dauer, which is characterized by developmental quiescence, stress resistance and a substantial extension of lifespan. Once conditions become favorable, the L3-dauers exit the developmental diapause and rapidly progress into the L4 stage through a series of metabolic and developmental changes that are governed by a coordinated transcriptional network [1]. Through this developmental adjustment, C. elegans is able to maximize its reproductive advantage under diverse environmental conditions [2]. Similar to free-living species like C. elegans, parasitic nematodes also alter their larval development based on environmental conditions [3,4]. Before host infection occurs, larvae of developing parasitic nematodes, such as hookworms and Strongyloides stercoralis (S. stercoralis), arrest their reproductive growth at a dauer-like stage called infectious L3 (iL3). Then, upon infection of their hosts where environmental conditions favor the completion of the parasite’s life cycle, the arrested iL3 larvae resume reproductive growth and develop into fertile egg-laying adults.
At the molecular level, the nematode development program is controlled by a hormonal signaling pathway initiated by insulin/IGF-I and TGF-β, which eventually converges in the activation of a nuclear receptor called DAF-12 [2,5,6]. In C. elegans, favorable environments stimulate insulin/IGF-I and TGF-β pathways that induce expression of DAF-9, a cytochrome P450 enzyme that catalyzes the synthesis of steroid-like hormones, called dafachronic acids [2,5,6]. Dafachronic acids serve as ligands that bind and activate DAF-12 [5–7], which in turn commit the nematode to reproductive growth. Conversely, in unfavorable environments, the insulin/IGF-I and TGF-β pathways remain inactive, preventing dafachronic acid synthesis, which in turn allows DAF-12 to interact with DIN-1, a strong co-repressor that is required for dauer formation [5,6,8]. In C. elegans larvae lacking DAF-12, this repressor activity is absent, causing a dauer-defective phenotype that would be expected to decrease viability in an unfavorable environment [2,9]. In parasitic nematodes, the insulin/IGF-I/DAF-12 signaling pathway controlling development appears to be conserved [10–15]. Similar to C. elegans, in parasitic nematodes ligand-free DAF-12 is required for formation of the dauer-like iL3, whereas ligand-activated DAF-12 is required for reproductive growth [15].
In C. elegans, larvae undergoing reproductive growth or dauer diapause display distinct patterns of energy metabolism. L2–L4 larvae in reproductive growth exhibit aerobic energy metabolism by converting dietary energy sources (carbohydrates and lipids) into acetyl-CoA, which is then fed into the TCA cycle and oxidative phosphorylation [2,16,17]. This aerobic metabolism produces sufficient energy to support rapid, energy-demanding reproductive growth. In contrast, aerobic energy metabolism is greatly reduced in dauer larvae, which instead exhibit a slower rate of anaerobic energy metabolism. Paradoxically, however, anaerobic metabolism also utilizes fat metabolism to meet the nematode’s energy needs for survival during privation [18–21]. The pathways involved in anaerobic energy metabolism are the glyoxylate cycle, a variant of the TCA cycle that converts acetyl-CoA to malate, and malate dismutation, a fermentation pathway that metabolizes malate for energy production [2,16,17]. The tendency towards the lower rate of anaerobic metabolism in dauer larvae prevents premature depletion of energy reserves and facilitates extended lifespan [22]. Thus, two distinct types of metabolism are employed to produce energy during reproductive growth and diapause stages, raising the question as to how the two pathways are differentially regulated.
Although the study of metabolism in parasitic nematodes is hampered by the difficulty in obtaining sufficient numbers, it is known that during reproductive growth in their hosts certain species of parasitic larvae migrate through the circulatory system, lungs and trachea, where aerobic conditions are high [3,4]. Furthermore, iL3 larvae of the hookworm Ancylostoma caninum are suggested to use fat reserves as an energy source [23], and there is an inverse correlation between oxygen consumption and iL3 longevity in parasitic nematodes [24]. These findings suggest that similar mechanisms control developmental energy metabolism in free-living and parasitic nematodes.
In the present study, we show that in addition to governing expression of developmental genes required for entry and exit from dauer diapause, DAF-12 is required for activating a metabolic network that is required for the normal progression to reproductive maturity. Efforts to elucidate the molecular targets of DAF-12 have focused mainly on the identification of heterochronic and microRNA genes that ensure the correct developmental decision is made during entrance and exit from dauer [25–28], and on longevity genes that are repressed in long-lived mutant worms [29–31]. Notably, however, a role for DAF-12 in energy homeostasis has not been well documented. Utilizing a combination of biochemical and genetic approaches, we show that DAF-12 is a key transcriptional regulator of developmental energy metabolism. In C. elegans, DAF-12 induces expression of a gene network that is responsible for aerobic fat utilization during reproductive growth. Further, this DAF-12-dependent metabolic network is conserved in the parasitic nematode, S. stercoralis. This work provides a molecular understanding of how nematodes adjust energy metabolism to assure successful reproduction in wide-ranging environments, and it suggests a therapeutic strategy for treating parasitic diseases by inhibiting fat utilization.
To investigate the potential role of DAF-12 in regulating energy metabolism in C. elegans, we used a dauer defective double-null mutant lacking both the DAF-12 co-repressor din-1 and the dafachronic acid-synthesizing enzyme daf-9 [8]. Employing a mutant that lacks both din-1 and daf-9 commits C. elegans to constitutive reproductive growth even in the absence of dafachronic acids. The advantage of the din-1;daf-9 mutant is that it permits evaluation of the direct effects of DAF-12 activation on metabolism while at the same time minimizing effects due to the developmental switching that would otherwise occur in the single null mutant of daf-9. We found that treating din-1;daf-9 larvae with the high affinity endogenous ligand, Δ7-dafachronic acid (DA) decreased triglyceride levels in a dose dependent manner (Fig. 1A). This decrease was not due to reduced dietary nutrient uptake, since DA treatment had no effect on pharyngeal pumping rates of the larvae (Fig. 1B) but rather slightly increased dietary fatty acid uptake (Fig. 1C). In contrast, DA treatment increased the fatty acid oxidation and oxygen consumption in din-1;daf-9 larvae (Fig. 1D, E). At the same time, DA treatment did not significantly change either triglyceride levels, fatty acid oxidation, or oxygen consumption in mutants that lack DAF-12 (din-1;daf-12) or in wild type N2 larvae (Fig. 1A, D, E), indicating that the effects of DA on metabolism were DAF-12 dependent.
We also examined whether DAF-12 activation affects reproduction. As shown in Fig. 1F, progression from L4 to the young adult stage, when reproductive organs become well-developed [32], occurred earlier in din-1;daf-9 larvae treated with DA compared to vehicle in a DAF-12 specific manner. DA treatment also advanced the onset of egg laying, another marker of reproductive maturity (Fig. 1G). Together, these findings demonstrate that DAF-12 activation induces aerobic energy metabolism and accelerates larval reproductive growth.
To gain insight into the molecular mechanism underling the DAF-12-regulated fatty acid metabolism, we evaluated global changes in C. elegans gene expression by comparing vehicle and DA treated L3 larvae. Microarray analysis identified 796 genes that were up-regulated and 985 genes that were down-regulated (>2-fold change and FDR<5%) in response to DAF-12 activation (S1 Table). The DA-regulated transcriptome was then grouped into several functional categories based on gene ontology (DAVID, ref. [33], S1 Table). In addition to the expected changes in expression of heterochronic and molting genes (e.g., dre-1) that coordinately regulate developmental and reproductive pathways [26], DA governed expression of a distinct cadre of genes involved in the metabolism of lipids (S1 Table). In contrast, no changes were observed in the expression of genes required for metabolizing glucose.
We then compared the microarray data from our DA responsive genes with that of genes that have been shown to be regulated during the exit from dauer [1], which is another process where DAF-12 is activated. As expected, the transcriptome of DA up-regulated genes significantly overlapped the trancriptome of genes up-regulated during dauer recovery (S1A Fig.). These data indicate that DAF-12 engages a gene network that governs metabolism and growth during both reproductive development and dauer recovery. We also compared the DAF-12 transcriptome with genes that are regulated by the transcription factor DAF-16 [34]. Whereas DAF-12 activation suppresses dauer, activation of DAF-16 is known to promote dauer [2]. As expected by this reciprocal regulation of dauer, there was no statistically significant overlap between genes regulated by DAF-12 and DAF-16 (S1B Fig. and S1C Fig.). Although this comparison was between the transcriptomes from different stages of worms (L3 for DAF-12 vs. adult for DAF-16), these results suggest that DAF-12 and DAF-16 regulate distinct gene networks to coordinate entry and exit from dauer diapause, and the initiation of metabolic pathways that promote reproductive development.
To further investigate the metabolic network governed by DAF-12, we used qPCR to confirm changes in expression of fatty acid metabolic genes identified in our microarray study, as well as other candidate genes known to be involved in fatty acid metabolism [19,20]. A total of 69 genes were evaluated (S2 Table). A hallmark of the DA-regulated metabolic pathway that correlates to reproductive growth is the induction of aerobic fatty acid oxidation (Fig. 1). DA increased expression of genes involved in every aspect of aerobic fatty acid utilization, including lipolysis, transport, esterification, and oxidation in both peroxisomes and mitochondria (Fig. 2A-E; S2 Table).
To provide an additional objective assessment of DAF-12’s role in regulating energy metabolism, we compared the DA-regulated lipid metabolic gene profile to changes observed in response to fasting. In addition to reproductive growth, fasting is another physiological process that mobilizes and utilizes fatty acids [19,20]. However, in contrast to reproductive growth, fasting utilizes anaerobic metabolism marked by reduced metabolic rates [21] and activation of the glyoxylate cycle (through icl-1 expression) [19,20]. Of the 69 fatty acid metabolic genes tested above (S2 Table), 20 were increased by DAF-12 and 37 were increased by fasting (Fig. 2F, S2 Table). Importantly, there was no significant overlap (based on hypergeometric distribution) in the number of genes that were either up- or down-regulated under both conditions, demonstrating that DAF-12 and fasting engage distinct gene networks for fatty acid utilization. DA decreased the mRNA levels of icl-1 (Fig. 2G), the bi-functional enzyme with isocitrate lyase and malate synthase activities that is unique to the glyoxlate cycle and thus serves for an indicator of anaerobic fatty acid utilization. Taken together with the biochemical measurements of fatty acid oxidation (Fig. 1), these results suggest that DAF-12 activation selectively governs the pathways that lead to energy mobilization and utilization by increasing expression of genes involved in aerobic lipid metabolism.
We also investigated the signaling pathways that govern metabolism of DA. Interestingly, expression of daf-28 (insulin/IGF-I-like), daf-7 (TGF-β-like) and daf-9 (the DA biosynthesis enzyme) were specifically suppressed by exogenous DA treatment, while expression of strm-1, which quenches DAF-12 ligand synthesis [35], was induced (S2 Fig.). The ability of DAF-12 to repress its own activity is reminiscent of a classic endocrine feedback circuit that functions to maintain homeostasis.
To determine whether the up-regulation of gene transcription by DA was through the direct action of DAF-12, we analyzed the promoters of several of the DA-induced genes that were confirmed by qPCR. In the absence of a DAF-12 antibody to perform chromatin immunoprecipitation experiments, we used bioinformatics to search for the consensus DAF-12 DNA binding element [36] in the promoters/introns of five representative DAF-12-induced genes. Selection of these genes was based on their representation of different metabolic processes, their high levels of expression, their response to DA, and their distinct chromosomal locations (i.e., genes not likely to be in a gene cluster sharing a common promoter). Our analysis revealed 33 putative DAF-12 response elements (S3 Table). We found that DAF-12 bound efficiently to 13 of these elements (Fig. 3A, B; S3 Table) and activated transcription in a standard cell-based reporter assay through four of them (Fig. 3C; S3 Table). These four DAF-12 response elements corresponded to three (K08B12.1, acs-1 and acs-3) of the five genes originally selected for analysis. Consistent with these genes being direct transcriptional targets of DAF-12, K08B12.1 and acs-3 expression was induced rapidly within 30 to 60 minutes after treatment of din-1;daf-9 larvae with DA (Fig. 3D). These data suggest that at least a portion of the genes regulated by DAF-12 are likely to be direct targets.
We next asked whether aerobic fatty acid utilization is required for DAF-12 to promote reproductive growth by inhibiting aerobic fatty acid utilization with etomoxir, a specific inhibitor of the carnitine palmitoyltransferases that mediate fatty acid transport into mitochondria [37]. Etomoxir treatment significantly decreased fatty acid oxidation and increased fat storage in C. elegans (S3 Fig.), demonstrating the effectiveness of the drug in inhibiting this pathway in nematodes. A further consequence of etomoxir treatment was that it completely blocked the earlier onset of egg laying that is dependent on DA, which is a marker for reproductive growth (Fig. 4A). Etomoxir treatment also prevented DA-mediated rescue of reproductive growth in the daf-7 and daf-9 mutants (Fig. 4B, C) and delayed the rescued growth in the daf-2 mutant (Fig. 4D). In this latter mutant, the inability of etomoxir to inhibit growth completely is likely due to compensatory anaerobic fatty acid utilization that is known to occur in the daf-2 mutants [34,38]. Consistent with the data shown in Fig. 1F, DA treatment did not affect reproductive capacity in wild type N2 worms or in mutants lacking DAF-12 expression (din-1;daf-12, daf-9;daf-12, and daf-7;daf-12), regardless of the absence or presence of etomoxir (S4 Fig.). In sum, these data demonstrate that aerobic fatty acid metabolism is required for DAF-12 to promote growth from larvae to reproductive adults in C. elegans.
Like C. elegans, many species of parasitic nematodes such as S. stercoralis also use the conserved insulin/IGF-I and DAF-12 signaling pathways to regulate their development [10–15]. We therefore asked whether the role of DAF-12 in promoting fatty acid utilization is conserved during reproductive growth of S. stercoralis. To test this, we confirmed the presence of fat utilization genes in S. stercoralis and then examined whether they are regulated by DA (Fig. 5A, S4 Table). As in C. elegans, expression of genes encoding a lipase (Ss_F28H7.3), acyl-CoA synthase (Ss_acs-1) and a gene involved in acyl-CoA transport (Ss_acbp-3) was induced, while the key glyoxylate cycle gene (Ss_icl-1) was repressed by DA treatment. Expression of the carnitine palmitoyltransferase gene, Ss_W03F9.4, required for mitochondrial β-oxidation, was also increased by DA (Fig. 5A; S4 Table). In contrast, expression of genes involved in peroxisomal β-oxidation (Ss_acox-2, Ss_acox-3 and Ss_ech-8) was repressed by DA (Fig. 5A; S4 Table). This profile of gene regulation supports the notion that DAF-12 activation induces aerobic fat utilization in S. stercoralis similar to that observed for C. elegans (Fig. 2).
We also examined the effect of etomoxir on DA-induced reproductive growth in the post-free-living larvae of S. stercoralis, which typically arrest at the dauer-like iL3 stage. Although DA is less potent as an agonist for DAF-12 in S. stercoralis compared to C. elegans [15], DA treatment was able to induce >85% maturation of S. stercoralis larvae to the L3-L5 stages (Fig. 5B). Notably, co-treatment of etomoxir with DA resulted in a significant decrease in the number of L3-L5 larvae (Fig. 5B). Administration of etomoxir to DA-treated worms also led to a marked increase in lethality of these L3-L5 larvae (Fig. 5C). Importantly, treatment with etomoxir alone did not kill S. stercoralis iL3 larvae (which are in an arrested developmental stage), demonstrating that the lethality observed in L3-L5 worms is due to the specific effects of etomoxir on the reproductive developmental program induced by DA (Fig. 5D).
These results reveal that aerobic fatty acid metabolism is required for DAF-12-dependent reproductive growth of S. stercoralis and suggest that control of this metabolic pathway by DAF-12 might be a promising strategy for regulating development in these important pathogens.
The nuclear receptor DAF-12 plays an essential role in C. elegans in linking nutritional status to developmental programs, including the transition from second- to third-stage larval development and the progression into and out of dauer diapause. In this report, we show that activation of DAF-12 by its hormonal ligand DA directly regulates energy homeostasis by inducing aerobic fatty acid utilization, which is required for reproductive growth. DAF-12 was shown previously to regulate heterochronic genes, which determine the stage-specific timing of cell type differentiation [25–27]. Interestingly, the DAF-12 gene regulatory network is distinct from that of DAF-16, which is active in the absence of DA signaling to promote dauer and inhibit reproductive growth. Thus, DAF-12 controls the balance between reproductive growth and dauer both by controlling the expression of genes directing development, and by regulating the flux of nutrients required to fuel the developmental program. Our findings provide a molecular explanation for the longstanding observation that C. elegans switches to aerobic metabolism once reproductive growth has been initiated [2,16].
Under favorable environmental conditions, which promote the biosynthesis of DA, we showed that DAF-12 stimulates aerobic fatty acid utilization in larvae as evidenced by decreased storage of triglycerides and increased oxygen consumption and fatty acid oxidation (Fig. 1A, D, E). Although a previous study originally suggested DAF-12 might increase rather than decrease fat storage [8], it has been shown since that the lipid staining assay used in this previous study is non-specific and ineffective for determining fat content [39]. DA treatment also stimulates progression from L4 to the young adult stages and advances the onset of egg-laying activity (Fig. 1F, G). Thus, DAF-12 coordinates the release of energy required to support the rapid, energy-intensive growth of larvae to reproductive maturity. At the molecular level, DA induces metabolic genes that regulate fatty acid utilization at multiple steps, including fatty acid mobilization, esterification, and peroxisomal and mitochondrial β-oxidation. The rate-limiting step enzyme in fatty acid oxidation, carnitine palmitoyltransferase, is encoded by several homologs in C. elegans, two of which are DAF-12 targets (Fig. 2E) and inhibition of the enzyme’s activity blocks DA-stimulated reproductive growth (Fig. 4). Concomitant with its regulation of genes involved in oxidative metabolism, DA represses the expression of icl-1, which encodes a bifunctional enzyme in the glyoxylate cycle that is essential for anaerobic fatty acid catabolism (Fig. 2G). These results support the role of DAF-12 in promoting reproductive growth of C. elegans by acting as a key controller of energy homeostasis in response to nutrient supply.
The adaptive response to fasting is a process that also mobilizes fat storage to maintain energy homeostasis. However, in contrast to the metabolic pathway involved in reproductive growth, fasting results in an increase in anaerobic metabolism. Fasting decreases metabolic rate [21] and utilizes the glyoxylate cycle to provide energy from fatty acids [18–20], which is diametric to the action of DA. The alternate role of DA in aerobic metabolism is supported by the finding that the metabolic gene networks induced by DAF-12 and by the fasting response are distinct (Fig. 2G). Interestingly, the metabolic response to fasting is mediated by another nematode nuclear receptor, NHR-49 [19,20]. Analogous to the role of DAF-12 in the dauer diapause, NHR-49 is required for the entry and exit of adult reproductive diapause, a process that preserves reproduction in C. elegans during starvation [40]. Thus, DAF-12 and NHR-49 appear to control two separate gene networks that alternatively use aerobic and anaerobic fatty acid utilization to ensure successful reproduction under varying environmental conditions.
Another key finding of the current study is that DAF-12-mediated regulation of energy homeostasis is conserved in the human parasite, S. stercoralis. As in C. elegans, DAF-12 activation stimulates the expression of genes involved in aerobic fatty acid utilization (Fig. 5A). While genes involved in peroxisomal β-oxidation were induced in C. elegans and repressed in S. stercoralis, we note that peroxisomal β-oxidation can support either aerobic or anaerobic fatty acid catabolism. Importantly, however, blocking aerobic fatty acid utilization with etomoxir inhibited DAF-12-induced reproductive growth in S. stercoralis (Fig. 5B-D). Inhibiting this metabolic pathway in other parasitic species using this type of strategy has recently been suggested [41]. To date, there is a very limited armamentarium of anthelmintic drugs that is effective against S. stercoralis, which can cause disseminated strongyloidiasis and multi-organ failure in infected humans. Although further studies in relevant host species are needed, our results suggest targeting metabolic enzymes may lead to a therapeutic approach for treating diseases caused by S. stercoralis and possibly other parasites [42]. To that end, it is interesting that etomoxir and other drugs that were originally developed to regulate fatty acid metabolism [37] as a means for treating diabetes and metabolic disease might be repurposed for treating parasitism.
In summary, our studies reveal a novel facet of DAF-12 activity in both C. elegans and parasitic nematodes, namely the regulation of fatty acid catabolism and energy homeostasis. In this regard, DAF-12 is similar to the PPAR subfamily of nuclear receptors, which coordinately regulate fatty acid homeostasis and energy balance in vertebrates in response to nutrient availability. Our results provide a molecular explanation for how nematodes adjust energy homeostasis in response to changes in environmental conditions for reproduction. Moreover, they suggest a new strategy for developing new classes of anthelmintic drugs.
Δ7-DA was synthesized as described [6]. C. elegans strains din-1;daf-9(dh127;dh6), din-1;daf-12(dh127;rh61rh411) and daf-9(dh6) were from Dr. Adam Antebi (Max-Planck Institute for Aging); wild type (N2 strain), daf-2(e1368) and daf-7(e1372) worms were from C. elegans Genome Center (University of Minnesota). daf-7;daf-12(e1372;rh61h411) mutant was made by crossing daf-7 hermaphrodites with hemizygous daf-12 males and was screened for dauer defective F2 progenies. The wild type (UPD) strain of S. stercoralis, used for developmental switching studies, and an iso-female line (PV001) of this parasite, used for qPCR studies, were maintained as described [43].
Vehicle or Δ7-DA was mixed with 5× concentrated OP50 bacteria culture and loaded on NGM-agar plates. L1 larvae prepared by egg synchronization were cultured on these plates at 25°C for 22.5 h and the resulting L3 larvae were collected and washed in M9 buffer for the indicated assays. For triglyceride (TG) content, worms were sonicated and the resulting lysates were centrifuged at 13000×g at 4°C. From supernatants, total glyceride (TGs plus free glycerol) and free glycerol were measured by Infinity TG Reagent (Thermo Sci) and Free Glycerol Reagent (Sigma), respectively. TG levels were calculated by subtracting free glycerol from total and were normalized to protein amounts in the lysates.
Fatty acid oxidation was measured as described by the production of H2O from fatty acid [44]. Briefly, the L3 larvae from different treatment groups were incubated with a mix of cold and [3H]-palmitic acid (Perkin Elmer) complexed with fatty acid-free BSA (Sigma), and incubated in a shaker at 25°C for 1 h. The reaction was terminated by adding 10% TCA followed by centrifugation at 13000×g for 5 min to obtain supernatant. Remaining [3H]-palmitic acid was deprotonated by adding 5N NaOH and PBS and removed by ion exchange column (Dowex 1×8 200–400 mesh Cl, strongly basic, Sigma). [3H]2O left in the supernatant was measured by scintillation counting.
For oxygen consumption, worms were mixed with antibiotic-killed OP50 bacteria, and then transferred to Oxygen Biosensor Plates (BD Bioscience) for oxygen measurement. Oxygen consumption was expressed as the increase of fluorescence units (ΔFU) and was normalized by protein amounts of the worms. For fatty acid uptake, worms were mixed with OP50 bacteria and 250 nM of fluorescent tracer (C1-BODIPY-C12 fatty acid, Invitrogen). Following 1 h incubation at 25°C, worms were washed, mounted, and photographed under fluorescence microscopy. Fluorescence density units (FU) of each worm were quantified by the software Image-J.
Pharyngeal pumping rates were measured as described [45]. Briefly, din-1;daf-9 L3 larvae were transferred to a fresh NGM plate with OP50 bacteria lawn and were videotaped through a stereomicroscope. Pumping rates were measured by counting the grinder movements and presented as pumps per minute. For each treatment, 10 L3s were assayed.
Reproductive growth of C. elegans was measured by L4-young adult transition or by egg laying assays. For L4-YA transition assay, synchronized L1 larvae from were cultured on NGM-agar plates pre-loaded with 5× concentrated HT115 bacteria culture. Worms were grown at 20°C and young adults were counted at indicated time points. Data were presented as the percentage of young adults in whole populations. For the egg-laying assay, synchronized din-1;daf-9 L1 larvae were grown at 25°C on NGM-agar plates pre-loaded with 5× concentrated OP50 bacteria culture. At the indicated time points, 10~15 worms were transferred to fresh plates with bacterial lawns for 2.5 h and laid eggs were counted. Data were presented as numbers of the eggs laid by each worm per hour.
For C. elegans experiments, synchronized L1 larvae were treated with vehicle or Δ7-DA for 22.5 h at 25°C, or synchronized wild type L1 larvae were cultured to L4 stage and then harvested as fed worms or were deprived of food for an additional 12 h to obtain fasted worms. For S. stercoralis experiments, iL3s worms were treated with or without Δ7-DA in M9 buffer at 37°C and 5% CO2 in air for 24 h. Total RNA from worms was extracted with TRIzol Reagent (Life Technologies), and analyzed by qPCR. Relative mRNA levels were normalized to expression of reference genes inf-1 or ama-1 (C. elegans) or 18S ribosomal RNA (S. stercoralis). Data were presented as fold changes of relative mRNA levels in DA versus vehicle treated worms or in fasted versus fed worms.
Total RNA was also subjected to the C. elegans Genome Array (Affymetrix) for whole genome gene expression analysis. Briefly, gene expression values were log2 transformed and genes with >10-fold difference between replicates in either of the treatments were removed from our analysis. To identify the differentially expressed genes, we applied Significance Analysis of Microarrays (SAM) analysis using the R package samr [46]. Genes with median false discovery <5% and fold changes >2.0 were considered differentially expressed.
DAF-12 proteins were prepared with TNT Quick-Coupled Transcription/Translation System (Promega) and blocked with poly-[dI-dC] and non-specific single-stranded oligos. The DAF-12 proteins were then incubated with [32P]-end-labeled dsDNA probes (S3 Table) at room temperature for 30 min and binding to DAF-12 was analyzed by 5% PAGE followed by autography. For competitive binding experiments, 20- or 200-fold excesses of unlabeled DNA probes were also included in the binding reaction.
Co-transfection and luciferase reporter assays were performed as described in HEK 293 cells [15]. Eight hours post-transfection, cells were treated with vehicle or 1 μM Δ7-DA, and luciferase and β-galactosidase activities were then measured 16 h later. Relative luciferase units (RLU) were normalized to β-galactosidase activity. Reporter plasmids were constructed by inserting DAF-12REs and their 10-bp genomic flanking sequences into a TK-luc reporter plasmid.
Synchronized din-1;daf-9 L1 larvae were grown in 5×concentrated OP50 in liquid suspension and shaken at 25°C for 15 h. Resulting L2 larvae were treated transiently with vehicle or 500 nM Δ7-DA and harvested in ice-cold M9 buffer. Cell nuclei were extracted and incubated with ATP, CTP, GTP and 5’-Bromo-UTP (BrUTP) at 30°C for labeling of nascent RNAs with BrUTP (BrUTP-RNAs). The BrUTP-RNAs were then enriched with anti-BrUTP agorase beads (Santa Cruz) and quantified by qPCR.
Development switching assays were performed as described [5,15]. Briefly, synchronized L1 larvae (C. elegans) or eggs (S. stercoralis) were grown on NGM-plates pre-loaded with etomoxir and phenotypes were observed after 60 h incubation at 25°C. Data from three independent experiments were pooled and significance was determined by Fisher’s exact test.
S. stercoralis homologs were identified as reported [43] by a TBLASTN (NCBI) search of C. elegans versus S. stercoralis (6 December 2011 draft; ftp://ftp.sanger.ac.uk/pub/pathogens/HGI/) databases, followed by annotation to RNA-seq data (ArrayExpress accession number E-MTAB-1164). Phylogenetic tree analyses were constructed to resolve gene homology. S. stercoralis genes with 1:1 homology to C. elegans genes were identified as homologous genes.
Unless otherwise stated, data were expressed as mean ± SD or SEM and significance tests between vehicle- or DA-treated groups were performed by Student’s t-test. The statistic tests of overlap between two gene sets were based on hypergeometric distribution and calculated by the R function “phyper ()” (https://stat.ethz.ch/R-manual/R-patched/library/stats/html/Hypergeometric.html).
|
10.1371/journal.pbio.2001886 | Simple rules can guide whether land- or ocean-based conservation will best benefit marine ecosystems | Coastal marine ecosystems can be managed by actions undertaken both on the land and in the ocean. Quantifying and comparing the costs and benefits of actions in both realms is therefore necessary for efficient management. Here, we quantify the link between terrestrial sediment runoff and a downstream coastal marine ecosystem and contrast the cost-effectiveness of marine- and land-based conservation actions. We use a dynamic land- and sea-scape model to determine whether limited funds should be directed to 1 of 4 alternative conservation actions—protection on land, protection in the ocean, restoration on land, or restoration in the ocean—to maximise the extent of light-dependent marine benthic habitats across decadal timescales. We apply the model to a case study for a seagrass meadow in Australia. We find that marine restoration is the most cost-effective action over decadal timescales in this system, based on a conservative estimate of the rate at which seagrass can expand into a new habitat. The optimal decision will vary in different social–ecological contexts, but some basic information can guide optimal investments to counteract land- and ocean-based stressors: (1) marine restoration should be prioritised if the rates of marine ecosystem decline and expansion are similar and low; (2) marine protection should take precedence if the rate of marine ecosystem decline is high or if the adjacent catchment is relatively intact and has a low rate of vegetation decline; (3) land-based actions are optimal when the ratio of marine ecosystem expansion to decline is greater than 1:1.4, with terrestrial restoration typically the most cost-effective action; and (4) land protection should be prioritised if the catchment is relatively intact but the rate of vegetation decline is high. These rules of thumb illustrate how cost-effective conservation outcomes for connected land–ocean systems can proceed without complex modelling.
| Many coastal marine ecosystems are threatened by anthropogenic activities, but often, the best way to restore and protect these important ecosystems is unclear. Conventional wisdom suggests that the 2 most effective conservation actions to benefit coastal marine ecosystems are implementation of marine protected areas or, alternatively, reduction of land-based threats. Active marine restoration is typically considered a low-priority option, in part due to high costs and low success rates. But does this conventional wisdom hold up to closer scrutiny? We developed a model to ask: should we restore or protect, on either the land or in the ocean, to maximise the extent of coastal marine ecosystems? We based the model on seagrass meadows and adjacent catchments in Queensland, Australia. Surprisingly, we found that direct, active marine restoration can be the most cost-effective approach to maximising extent of marine ecosystems over longer (decades-long) timescales. There is, however, substantial uncertainty in our understanding of the dynamics of complex linked land–sea ecosystems. Further, geomorphological and ecological conditions vary geographically. Therefore, we also used the model to investigate how uncertainty in key parameters affects decision-making outcomes. Our results can be used to guide investment into coastal marine conservation in the absence of complex, region-specific modelling.
| Widespread degradation and loss of coastal marine ecosystems has occurred over the previous centuries and has accelerated in recent decades [1–5]. These changes compromise the delivery of important ecosystem services to human society [6]. Coastal marine ecosystems pose a particular challenge to environmental managers because they are exposed to threats occurring both in the ocean (e.g., overfishing, direct damage) and on land. The conversion of native terrestrial vegetation for agriculture, urbanization, and industry increases runoff [7], causing degradation and die-offs of coastal ecosystems such as coral reefs [8] and seagrass meadows [2]. These declines threaten the functional integrity of coastal and marine ecosystems and the services they provide, such as food supplies, coastal protection, and climate regulation [9–11]. Consequently, the conservation of coastal species and ecosystems requires a mixture of both marine and terrestrial conservation actions [12–16].
Conservation prioritisation of marine ecosystems and adjacent landscapes traditionally focuses on protecting intact habitats, in either marine and or terrestrial realms, from future degradation [e.g. 17, 18–20] [but see 21]. Ecological restoration is commonly considered a less preferred management strategy than protection [22], particularly in marine environments, where restoration costs are high and success rates are low [23]. However, restoration can deliver better ecological outcomes than protection, depending on existing land uses, conservation intervention costs, and ecosystem expansion rates [24]. Compared to other actions, restoration is rarely considered [25], and trade-offs between restoration and protection actions have never been evaluated across complex land–sea systems.
Comparing the costs of conservation actions, both on land and in the ocean, with the benefits accrued in the marine ecosystem (‘cost-effectiveness’) is at the forefront of conservation planning for land–sea ecosystems [e.g. 17]. Incorporating exchanges across the land–sea interface is challenging, requiring the integration of data and models across the terrestrial, freshwater, and marine realms [12–14, 20]. Recent advances have allowed the benefits of terrestrial actions on marine ecosystems to be estimated [17–20, 26–33], but in practice, land–sea conservation planning has rarely explicitly quantified how the management of terrestrial threats impacts marine ecosystems. For instance, recent implementations of the ‘Reef 2050 Long-term Sustainability Plan’ for the Great Barrier Reef and the ‘Chesapeake Bay Total Maximum Daily Load’ programs aim to minimise sediment, nutrient, and pollutant delivery to the ocean and assume that marine ecosystems will respond positively [34, 35] but do not predict the effect size of the marine ecosystem response. As a result, it is not clear that their terrestrial focus will outperform actions in the marine environment, which have the advantage of directly affecting the management goal. To compare and prioritise actions across the land–sea interface, we need to identify the links between (1) the amount of land-based actions required to reduce a threat on receiving marine environments and (2) the amount of change in the marine ecosystem triggered by such a reduction.
We propose that integrated land–sea planning must compare the cost-effectiveness of 4 broad conservation actions: protect habitat on the land, protect habitat in the ocean, restore habitat on the land, and restore habitat in the ocean. Here, we develop a repeatable and transferable approach to determine which of those 4 actions maximises the extent of intact marine habitat for a given budget and project timeframe (Fig 1). The model extends an existing terrestrial model [24] across the land–sea interface. It is general in structure and could potentially apply to any marine system that is affected by sediment runoff and marine-based threats. In the original terrestrial model [24], the landscape is divided into 4 states describing the condition of the native vegetation—intact and unprotected, intact and protected, cleared, or restoring. The act of restoring or protecting habitat moves it between these different states. In our expanded model, we consider the state of habitat in both a landscape and adjacent seascape, which are connected by sediment runoff from the land into the ocean. Cleared terrestrial habitat increases sediment loads, which reduces water clarity in the adjacent ocean. The resulting decrease in light reaching the seafloor reduces the amount of habitat suitable for light-dependent species [13]. We focus on suspended sediments because they are a key driver of marine ecosystem condition in many inshore areas [2, 8, 36] but acknowledge the importance of the other components of runoff more broadly, including toxicant and nutrient loads. Importantly, our model assumes that the marine ecosystem is sensitive to both sediment runoff [37] and marine-based threats and that there is habitat in appropriate condition for marine restoration; if these conditions are not met, then approaches targeted towards either land- or ocean-based threats would be required. The model is spatially implicit, i.e., it is parameterised by spatial data (see Materials and methods, S1 Text, S1 Table).
We apply this model to a case study of seagrass meadows in Moreton Bay and riparian areas in adjacent catchments in Queensland, Australia (Fig 2). Seagrass meadows are an excellent test system because they provide a suite of ecosystem services and are strongly influenced by both land-based processes and direct local impacts in the ocean [2, 38–40]. The catchments draining into Moreton Bay are heavily modified, with only 20% intact remnant vegetation. Historical and ongoing land-clearing has significantly increased soil erosion, primarily through the process of gully erosion [41–42]. Suspended sediment delivery has negatively impacted marine ecosystems in the region [43–45]. The site is therefore representative of the wider global challenges posed to marine ecosystems by increased sediment runoff [14, 15, 39].
We ultimately aim to identify key factors that determine which broad conservation action is most effective under different circumstances. Therefore, we use the model output and sensitivity analyses to answer 2 questions. One, which of the 4 conservation actions maximises the extent of intact seagrass after 30 years? And two, under which conditions would our decision-making vary? Using the results from this and other studies [24, 37], we propose simple ‘rules of thumb’ that can help decision makers identify whether restoration or protection, either on the land or in the ocean, will be the most cost-effective approach to improving the state of marine ecosystems. These rules of thumb are likely specific to our study system but may be used as guidelines (or, alternatively, viewed as hypotheses) to inform decision-making in other regions until models are parametrised for those sites.
Using our dynamic model of seagrass meadows and riparian areas in adjacent catchments in Southeast Queensland, Australia, we investigated the effect of investment of $50 million (all costs are in 2015 USD unless otherwise stated) per year over 30 years into each of 4 separate conservation actions. The model was used to quantify the area of intact seagrass habitat resulting from marine restoration, marine protection, terrestrial restoration or terrestrial protection (see below, Materials and methods and S1 Text, for detailed descriptions). We found that if the objective is to increase the amount of habitat suitable for (but not necessarily occupied by) light-sensitive species (in this case, seagrass), then restoration of riparian areas on land is the most cost-effective strategy (Fig 3A). However, this will not necessarily maximise the area of occupied (‘intact’) marine habitat immediately, as that depends on how fast the marine ecosystem can recover and expand into habitat which was previously unsuitable due to low light availability; there is substantial uncertainty in this parameter (S2 Table). Controversially, we find that the most cost-effective way to maximise the extent of intact marine habitat over decadal timescales is to directly restore the marine ecosystem, despite the higher cost [23] (Fig 3B). Obviously, this conclusion depends on the availability of suitable, unoccupied habitat. If all marine habitat is unsuitable for marine restoration due to low water clarity, then revegetation of riparian vegetation to minimise sedimentation stress is required [46]. Below, we discuss the costs and benefits of each specific conservation action in turn for our study system and then examine how these decisions may vary for other systems.
Marine restoration was defined as planting seagrass transplants into habitat that is suitable for, but not presently occupied by, seagrass and was the most cost-effective action for achieving the highest coverage of seagrass habitat after 30 years (Fig 3B). Our modelling assumes that: (1) it takes 3 years for seagrass transplants to grow, fill in meadow gaps, and become a ‘healthy’, self-sustaining meadow [47]; (2) it costs $418,000 per ha [23] to source, transplant, and monitor the seagrass; (3) restoration has a high failure rate (62%) [23]; and (4) a maximum of 0.1% of the existing meadow can be in a restoring state in any year. Surprisingly, despite an expected cost of over $1 million per ha for successful restoration, restoring seagrass is a better strategy to maximise seagrass coverage than marine protection or land-based actions. Larger areas of intact seagrass are achieved if the area of seagrass that is in a restoring state in any year is less conservative (e.g., 1% [S1 Fig]). Our results imply that, given sufficient funding, effort, and suitable habitat, large-scale marine restoration projects could achieve significant gains in ecosystem extent, as recently assessed [46].
Marine protection was defined as the installation of environmentally friendly moorings, which avoid seafloor damage caused by traditional moorings and minimise the effects of dragging anchor chains [48]. In other regions, where trawling or dredging are the main threats to seagrass, the implementation of Marine Protected Areas (MPAs), which minimise seafloor damage by excluding destructive activities, may be a more appropriate conservation action [49]. Our study area is a Marine Park where seafloor destructive fishing techniques are forbidden and environmentally friendly moorings are the approach currently used to increase seagrass protection [50]. Our model predicts that marine protection yields the fastest initial increase and greatest total area of protected seagrass habitat (Fig 3C), because it is relatively cheap compared to the other actions. Marine protection increased the overall area of seagrass through time by a small amount (Fig 3B), because seagrass habitat decline rates are proportional to the amount of unprotected habitat. At $131,000 per ha, the 8.8% of seagrass habitat that is suitable for protection in the study region (S1 Text) can be protected in the first year so that, over decadal scales, the impact of marine protection on seagrass habitat area is limited.
Land restoration was defined as using revegetation and other actions in the riparian zone to reduce erosion in riverine locations where native vegetation had been previously cleared, at a cost of $17,310 per ha and with a probability of success of 50% (personal communication, J. O’Mara, SEQ Catchments). The resulting reduction in runoff (Fig 3D) increases the area of suitable marine habitat (Fig 3A), as the increased water clarity improves light availability on the seafloor. Our model advances and operationalises our understanding of the impacts of sediment input on light-dependent benthic marine species by factoring in an ‘action–response curve’ [51] describing the relationship between sediment loads and illuminated seafloor area that is suitable for light-dependent species (S1 Text). This relationship was generated using modelled daily sediment loads, monthly observed water clarity, and a species distribution model of seagrass habitat (S1 Text) [52] and is applicable in geomorphic and ecological contexts where sediment runoff impacts marine ecosystems [37]. Our results show that land restoration only offers small increases in seagrass coverage (Fig 3B) because there is a substantial 10-year time lag between restoration actions and the mitigation of sediment erosion and because we estimate that seagrass colonises newly available areas slowly (1.13% per year, [53]) (see S1 Text). Varying this parameter changes the results substantially, which we explore further below.
Land protection was defined as purchasing privately held land containing intact native vegetation and designating it as a nature reserve, at a relatively low cost of $3,530 per ha (S1 Text). Land protection only provides second-order benefits to marine habitat (Fig 3D): It reduces terrestrial habitat decline rates, leading to relatively less erosion and less sediment within the rivers. It therefore had relatively little impact on any metric of seagrass habitat (Fig 3A, 3B and 3C). Land protection therefore offers little benefit to catchments that are already highly degraded and where riparian habitat decline rates are low, such as in our case study.
While the model presented here can in theory be applied to any sensitive marine ecosystem affected by both land- and ocean-based threats, it is not straightforward to source the data needed for accurate parameterisation. We therefore varied key model parameters to identify contexts where the optimal conservation strategy may differ from our results, including rates of marine ecosystem decline and expansion, rates of terrestrial ecosystem decline, and the magnitude of previous land clearing. For instance, we can find the optimal investment strategy for landscapes with extensive historic and ongoing land clearing, such as parts of Malaysia and Indonesia [54], or high magnitudes of degradation but lower rates of ongoing land clearing, such as our study system [41] and Mediterranean countries including Albania, Algeria, and Bosnia [55]. Similarly, we can identify optimal approaches to marine ecosystems with different rates of habitat decline and expansion. For instance, kelp beds can undergo rapid declines yet can also recover rapidly when conditions are suitable [56]. In contrast, Posidonia oceanica seagrass meadows in some Mediterranean regions, such as Corsica, are declining slowly [57] but also have slow expansion rates [58].
We discovered that the relative rates of decline and expansion in the marine ecosystem, as well as the rate and magnitude of degradation on land, are key factors in our decision-making (See Fig 4, S2, S3 and S4 Figs). When marine habitat can recover more quickly than the rate of marine habitat decline, then land-based conservation can yield optimal results. Specifically, if the ratio of marine habitat expansion to decline rates is greater than approximately 1:1.4 (ratio of x- and y-axis values indicated by red dashed line in Fig 4A), then actions on land typically deliver the greatest cost-effectiveness (Fig 4, A-3, B-3, D-3, but see C-3 below). In that case, land restoration is the best option (Fig 4A and 4B, A-3, B-3), unless the catchment is relatively intact and the rate of land decline is high, in which case land protection is most cost-effective (Fig 4D, D-3). If the catchment is relatively intact with a low rate of loss, then marine protection is optimal (Fig 4C, C-3). Conversely, if marine habitat decline rates are greater than expansion rates, we should act in the ocean, essentially regardless of what occurs on land. Specifically, if the ratio between expansion and decline rates for the marine ecosystem is less than approximately 1:1.4 (Fig 4, A-1, A-2, B-1, B-2, C-1, C-2, D-1, D-2), then we should act in the ocean. If the rates of marine ecosystem decline and expansion are similar and relatively low (less than 1% per year), then restoration in the ocean is the most cost-effective strategy (Fig 4, A-1, B-1, C-1, D-1). Marine protection is the most cost-effective action for our system when rates of seagrass decline outside MPAs are high (Fig 4, A-2, B-2, C-2, D-2).
The findings from our analyses are factored into a generic decision-making protocol for conservation investment in marine ecosystems influenced by both land- and ocean-based threats (Fig 5). A number of conditions must be met for land-based stressors to critically impact marine ecosystems [37]. In addition to the assumptions outlined previously, the nearshore marine region must be within the impact radius of 1 or more rivers and be in an enclosed or shallow region, and land uses within the catchments must have increased the erosion of sediments or nutrients on a large scale [37] (Fig 5). If these criteria are met, then the results from our model can be used as a first step to guide decision-making, without the need for complex modelling.
The conservation dynamics of coastal ecosystems are driven by a combination of terrestrial and marine drivers. Efficient conservation investment will require decision-support tools that are repeatable, transparent, and that quantitatively describe the connections between the land- and sea-scape. The optimisation model we describe here provides a robust and extendable method to support these decisions.
Our study drew an unexpected conclusion: Despite high costs and low success rates, direct restoration of marine ecosystems may be the most cost-effective method to maximise marine habitat extent over decadal timescales. If marine restoration is not likely to succeed due to poor water quality, lack of suitable substrate, or other factors, then this conclusion will clearly not hold true. Nonetheless, we propose that a paradigm shift is occurring, whereby restoration is being recognised in particular contexts as an important option for the recovery of biodiversity [24, 59]. This is supported by recent findings that marine restoration is more likely to succeed if it is conducted on larger spatial scales [46] and if care is taken to select appropriate sites and techniques [23].
Our study highlights several factors that are essential elements in determining the most cost-effective management action but which are not often present in decision-support tools. These include the effects of time lags and the complex, nonlinear relationship between activities on the land and benefits in the sea. Land-based impacts are key drivers of seagrass extent and condition, suggesting that terrestrial protection should be a high-priority conservation action [40]. However, from a management perspective, it is essential to compare the outcomes of actions against a specific objective and timeframe and to factor in economic constraints. Our findings align broadly with those of Klein et al. [60], who report that the cost-effectiveness of marine conservation was almost always higher than that of terrestrial conservation within any ecoregion in the Coral Triangle. Our findings differ from those of Gilby et al. [21], whose findings support the more commonly held view that the most effective actions to benefit inshore coral reefs in Moreton Bay, Australia, would be expansion of the marine reserve network and reductions in sediment inputs from land, without considering variation in the costs of management actions or time lags. Other studies have focused on quantifying the effects of land-based impacts or protection on marine ecosystems but have not compared the results to those obtained from protection from ocean-based threats or have not quantified the effects of restoration in either marine or terrestrial realms [18, 19, 31].
In reality, a combination of approaches—on land and in the ocean—will be required to achieve ecological improvements in many marine regions. For instance, land-based actions would be required first if there was no suitable marine habitat available for restoration (e.g., Fig 5). For our study system, all seagrass that is suitable for protection using environmentally friendly moorings could be achieved using the budget in the first year. Similarly, the budget for marine restoration was not completely allocated in a given year when we assume that only small areas of marine restoration can be achieved at one time. This means that marine protection or restoration could be implemented first and that the budget could be used for other strategies concurrently or in later years. In practice, budget and regulatory agencies often do not span the land–sea interface, which means land- and ocean-based actions will likely proceed independently of one another.
Our model predicts the area of seagrass habitat change resulting from management actions on either side of the land–sea interface. The modelling framework provides a substantial advance in our ability to quantify the costs and benefits associated with conservation actions on land and in the ocean. In the present study, our objective was to maximise the extent of seagrass, but, there are multiple benefits from catchment restoration that are unrelated to marine ecosystems, such as enhanced freshwater biodiversity, reduced drinking water treatment costs, and increased public amenity. If we instead aim to maximise the delivery of ecosystem services provided by both seagrass and riparian habitats, which are worth $26,226 per ha per year and $27,021 per ha per year in 2007 USD, respectively, based on [61], then the optimal decision would always be to restore the catchment (S1 Text, S5 Fig). Although we used habitat area as a proxy for the delivery of ecosystem functions and services by a habitat, metrics of habitat condition (e.g., cover, biomass, species composition) could be explored in the future. The ultimate aim of this approach is to identify the most cost-effective ways of maximising ecosystem functions and services (e.g., food or habitat for other species, water filtration, and wave attenuation).
There are a number of uncertainties and assumptions that affect model outputs and interpretation. First, we quantify the seagrass decline rate from satellite imagery in a clear water region [62], which is likely less impacted than nearshore areas and therefore may underrepresent the decline rate. Second, the rates at which marine organisms can colonise new areas vary among species and regions (S2 Table), are scale-dependent, and the factors influencing expansion rates are not well understood. Low seagrass expansion rates may indicate bistability of seagrass and base substratum systems, where feedback mechanisms hinder the re-establishment of vegetation [63]. High seagrass expansion rates following improvement of environmental conditions can occur in some contexts [64], typically when seagrass seed banks are present [65]. The influence of rates of seagrass decline and expansion are explored in the sensitivity analyses. Predicting the area that is suitable for restoration in marine habitats requires a habitat distribution model, the results of which contain several uncertainties, including whether all relevant environmental variables have been included in the model, whether the species is in equilibrium with environmental variables, and which method is used to select the threshold value delineating species presence and absence [52]. The effectiveness of catchment restoration is also uncertain; erosion may in fact temporarily increase following riparian restoration before eventually diminishing [66], although our uncertainty in the parameter representing time lags in restoration is not likely to impact whether we should be acting on land or in the ocean (S2 Fig). There is large variability in the costs and success rates of restoration [23], yet we do not find these to be the most important factors affecting whether actions should occur on land or in the ocean (S3 Fig). Furthermore, the costs and success rates of conservation actions will vary across the land- and sea-scape; when costs vary spatially, managers can target lower-cost areas preferentially, which reduces the average cost of management activities. Lastly, our analysis does not factor in the impact of nutrient runoff, pesticides and herbicides, climatic variability, extreme events, or climate change, the impacts of which are extremely challenging to predict [67] and therefore beyond the scope of our work but which are important areas of future research.
Further uncertainty lies in the impacts of sediment runoff from the catchment on marine habitat dynamics. Previous studies linking land and the ocean have used simple distance-based relationships between sediment load and marine habitat metrics, such as ‘relative condition’ of coral reefs [e.g., 17, 31]. Advancing this approach, Tulloch et al. [19] used a sediment plume model that accounted for depth, bathymetry, currents, and particle size of modelled sediment runoff [28] to quantify reduction in relative coral condition due to sediment. Here, we apply a model that uses spatial empirical time series data of water clarity to estimate habitat suitability for light-dependent species; it would be interesting to compare how results vary based on the different approaches. Our approach is developed for seagrass but is applicable to other benthic marine habitats that are influenced by light availability, such as algae and coral reefs, although the link between sediment loads and the marine ecosystem would need to be modified to represent the dynamics of other ecosystems. We also tested how sensitive our model was to the functional form of the relationship between sediment loads and suitable marine habitat area by running the model with a separate linear relationship between sediment loads and suitable marine habitat area (S1 Text). Surprisingly, while the amount of seagrass habitat that can be achieved by each conservation action varies depending on which relationship is used, the optimal management action does not (S4 and S6 Figs). This is a relatively well-known finding in environmental decision theory, where uncertainties in the input parameters alter predictions but do not change the relative priority of management options [68, 69]. Finally, reductions in sediment supplies, such as those resulting from the construction of dams, negatively influence marine ecosystems such as mangroves [70]; discrepancies in ecological impacts of increases versus decreases in sediment supplies to coasts is a challenge for managers.
Despite structural and parametric uncertainty in the model, a quantitative optimisation framework that explicitly links conservation across the land–ocean interface provides a major conceptual advance. Specifically, it provides a quantifiable and repeatable structure for understanding the costs and benefits of taking different conservation actions and a transparent justification for acting either on the land or in the ocean. Thus, not only are we able to argue that marine conservation actions deliver the best outcomes for marine ecosystems, but this framework also offers a mechanistic explanation for why land-based management may be inferior: Multiple time lags separate terrestrial restoration projects from marine conservation outcomes. Although there are documented instances of land-based actions delivering measurable improvements in coastal water clarity and marine habitat extent [53], these have only materialised following delays in the effect of land-based management on runoff, further delays in the improvement of coastal water clarity, and final delays in the expansion of those marine habitats. Following decades or centuries of land- and ocean-based impacts on marine ecosystems globally [1], the challenge now is to reverse the resultant declines. Using transparent, transferable, and cost-effective approaches is critical to this process.
The study was parameterised for seagrass meadows in Moreton Bay, Queensland, Australia, and adjacent riparian areas below dams, which are considered the primary sources of sediments to the ocean in the region [42] (Lat: -27.0–28.3; Lon: 151.9–153.4) (Fig 2). Moreton Bay is a shallow coastal embayment adjacent to Brisbane, the capital city of Queensland. It is home to 18,000 ha of seagrass comprised of 7 species, which provide grazing areas for iconic, vulnerable, and threatened species such as green sea turtles, dugongs, and migratory shorebirds. Riparian areas in the catchment have been heavily cleared since European colonisation in the mid-1800s, mainly for agriculture and urbanisation, causing ongoing increased sedimentation in riverways and marine environments [41–42, 45]. Local direct threats to seagrass are mainly from physical damage from anchoring and mooring [71].
We extended the dynamic landscape modelling methodology of [24] to apply to both a seascape and adjacent landscape, which are connected together by sediment runoff from degraded landscapes into the ocean (Fig 1). Cleared terrestrial habitat increases sediment loads, which reduces water clarity in the adjacent ocean. The resulting decrease in light reaching the seafloor reduces the area suitable for light-dependent species. Model parameters were obtained from a variety of sources, including raster or shapefile spatial datasets (see below and S1 Table), but the dynamic landscape model is not spatially explicit. See below for additional details.
Each area of land at time t is classified as being in 1 of 4 states: intact and unprotected, A(t); intact and protected, P(t); degraded or cleared, C(t); or undergoing restoration, R(t), with the total amount in each state described as a proportion of the landscape, and with A(t) + P(t) + C(t) + R(t) = 1 at all times (Fig 1). The landscape area is constrained to riparian habitats, because those are the major determinant of sediment input to the ocean in Queensland and the primary target of current restoration projects [42, 72, 73]. The seascape is split into suitable habitat (sufficient light, soft sediments, and suitable wave energy, based on [52]) and unsuitable (US) habitat. While the total area is constant, the amount of each habitat changes in each time step based on sediment loads. The area that is suitable for seagrass is divided into the same categories as on land, but denoted by the subscript S, with AS(t) + PS(t) + CS(t) + RS(t) = 1 at all times. These categories can be considered available and suitable, protected and suitable, etc. When an increase in sediment causes a decrease in suitable habitat, that decrease is taken in the appropriate proportions from each category of suitable habitat, which then becomes unsuitable. When sediment decreases allow for an increase in suitable habitat, this newly suitable habitat is added to the cleared and suitable state (CS(t)). Transitions between habitat categories are determined by 4 rates (degradation and revegetation on land and in the ocean) and 6 processes (restoration and protection on land and in the ocean, expansion in the ocean, and change in suitable habitat area in the ocean), which are described below.
A major challenge to integrated land–sea planning is quantifying the relationship between actions undertaken on land and their effects on the marine environment. For seagrass meadows in Moreton Bay, this relationship is primarily defined by the effects of terrestrial sediment runoff on the amount of illuminated seafloor available in the ocean. We used an ‘action–response’ curve [sensu 51] describing the relationship between sediment loads and seagrass-suitable habitat area calculated in (S1 Text), which in turn uses the habitat distribution model published in [52], monthly water quality data, and monthly sediment load data (S1 Text, S10 Data). This relationship predicts the area of habitat that is suitable for seagrass in each year based on the sediment loads delivered from the catchment in the previous year. This approach provides a simplification of a complex system, whereby sediment distribution and resuspension are affected by sediment composition, rainfall, and oceanographic processes, among other factors. Factoring in spatially explicit hydrodynamic modelling of sediment distribution would be an important next step to this research.
Transitions in the area of suitable and unsuitable marine habitat are determined by the amount of intact land (A + P) in the previous time step, which affects the quantity of sediment delivered to the ocean and the area of habitat available for light-dependent marine species, like seagrass. If the area of suitable habitat is greater in t than in t−1, then the newly suitable habitat area is added to the Cleared (Cs) fraction, since this habitat would not contain seagrass at the outset. If the area of suitable habitat decreases in t compared to t−1, then habitat is removed proportionally from AS, PS, CS, and RS.
Information on the definitions of the 4 conservation actions (marine restoration, marine protection, land restoration, and land protection), as well as on their costs and probabilities of success, are given in the Results, S1 Text and S1 Table.
Here, we provide a summary of the model parameters describing the rates, processes, and initial conditions. Further information is given in S1 Text and S1 Table.
We run the model for 2 scenarios. For both scenarios, we run 4 allocation simulations, where the budget is allocated to each of the actions in isolation. Each simulation expends a budget of $50 million per year over 30 years in 0.1-year time intervals, for a total of $1.5 billion (not accounting for inflation and discount rates). By comparison, it will cost $5–$10 billion over 10 years to mitigate sedimentation issues in the Great Barrier Reef catchments [84]. The time horizon of 30 years aligns with other management policies aimed at mitigating sediment issues, e.g., the ‘Reef 2050 Long-term Sustainability Plan’ for the Great Barrier Reef [34]. All results are standardised to the outcomes achieved with no investment. For the first scenario, we invest the budget according to the parameters that we believe best describe the system. For the second scenario, we run a sensitivity analysis to examine uncertainty or variability in some of the key parameters identified in the model development process. Specifically, the model is run for rates of marine ecosystem decline and expansion varying by 0%–7% and 0%–10% per year, respectively. This approach is replicated for 4 different ecological and resource use contexts, encompassing historic land clearing extent = 20 or 80% and rate of land clearing = 0.75 or 7% per year. Further sensitivity analyses on the effects of time lags in restoration success, costs of actions, and the maximum area suitable for marine restoration or protection are provided in the Supporting Information and are outlined in S3 Table.
At the beginning of each time step, the area of suitable marine habitat is calculated according to the area of intact (A + P) habitat on land. Next, changes in land and ocean habitat in each time step Δt are modelled as described by the following equations, where B is the annual budget, S is the area of marine habitat, and L is the area of land habitat. Subscripts to S and L describe the proportional areas of different habitat categories transitioning between fractions. A representative model output describing the land–sea dynamics through time is provided in S7 Fig, and additional detail is given in S1 Text.
|
10.1371/journal.ppat.1006877 | A new mechanism of interferon’s antiviral action: Induction of autophagy, essential for paramyxovirus replication, is inhibited by the interferon stimulated gene, TDRD7 | The interferon (IFN) system represents the first line of defense against a wide range of viruses. Virus infection rapidly triggers the transcriptional induction of IFN-β and IFN Stimulated Genes (ISGs), whose protein products act as viral restriction factors by interfering with specific stages of virus life cycle, such as entry, transcription, translation, genome replication, assembly and egress. Here, we report a new mode of action of an ISG, IFN-induced TDRD7 (tudor domain containing 7) inhibited paramyxovirus replication by inhibiting autophagy. TDRD7 was identified as an antiviral gene by a high throughput screen of an ISG shRNA library for blocking IFN’s protective effect against Sendai virus (SeV) replication. The antiviral activity of TDRD7 against SeV, human parainfluenza virus 3 and respiratory syncytial virus was confirmed by its genetic ablation or ectopic expression in several types of mouse and human cells. TDRD7’s antiviral action was mediated by its ability to inhibit autophagy, a cellular catabolic process which was robustly induced by SeV infection and required for its replication. Mechanistic investigation revealed that TDRD7 interfered with the activation of AMP-dependent kinase (AMPK), an enzyme required for initiating autophagy. AMPK activity was required for efficient replication of several paramyxoviruses, as demonstrated by its genetic ablation or inhibition of its activity by TDRD7 or chemical inhibitors. Therefore, our study has identified a new antiviral ISG with a new mode of action.
| The antiviral functions of interferons (IFNs) are mediated by the IFN-induced proteins, encoded by the IFN Stimulated Genes (ISGs). Because ISGs are virus-specific, we performed a high throughput genetic screen to identify novel antiviral ISGs against Sendai virus (SeV), a respirovirus of the Paramyxoviridae family. Our screen isolated a small subset of anti-SeV ISGs, among which we focused on a novel ISG, Tudor domain containing 7 (TDRD7). The antiviral activity of TDRD7 was confirmed by genetic ablation of the endogenous, and the ectopic expression of the exogenous, TDRD7 in human and mouse cell types. Investigation of the mechanism of antiviral action revealed that TDRD7 inhibited ‘virus-induced autophagy’, which was required for the replication of SeV. Autophagy, a cellular catabolic process, was robustly induced by SeV infection, and was inhibited by TDRD7. TDRD7 interfered with the ‘induction’ step of autophagy by inhibiting the activation of AMP-dependent Kinase (AMPK). AMPK is a multifunctional metabolic kinase, which was activated by SeV infection, and its activity was required for virus replication. Genetic ablation and inhibition of AMPK activity by physiological (TDRD7) or chemical (Compound C) inhibitors strongly attenuated SeV replication. The anti-AMPK activity of TDRD7 was capable of inhibiting other members of Paramyxoviridae family, human parainfluenza virus type 3 and respiratory syncytial virus. Therefore, our study uncovered a new antiviral mechanism of IFN by inhibiting the activation of autophagy-inducing kinase AMPK.
| Interferon (IFN) system provides the first line of immune defense against viral infections in vertebrates [1–3]. It is designed to inhibit viral infection by blocking virus replication and eliminating the virus-infected cells. The Pattern Recognition Receptors (PRRs), e.g. Toll Like Receptors (TLRs), RIG-I Like Receptors (RLRs) and cyclic AMP-GMP synthase (cGAS)/stimulator of IFN genes (STING), are located in distinct cellular compartments, to sense specific viral components, such as the viral nucleic acids [4–9]. Upon ligand stimulation, the PRRs trigger rapid downstream signaling pathways via respective adaptor proteins to activate the transcription factors, e.g. Interferon Regulatory Factors (IRFs) and Nuclear Factor-κB (NF-κB). The co-operative action of these transcription factors triggers the synthesis of Type-I interferons e.g. IFN-β, an extensively studied antiviral cytokine. After synthesis in the infected cells, IFN-β is secreted and acts on the infected as well as yet uninfected cells via Janus Kinase (JAK)/Signal Transducer of Transcription (STAT) signaling pathways to trigger the synthesis of a number of antiviral genes.
All biological effects of IFN are executed by the induced proteins, encoded by Interferon Stimulated Genes (ISGs), which are either not present or expressed at a low level in untreated cells, but can be transcriptionally upregulated by IFN-action [3, 10, 11]. Most ISGs can also be induced directly in the virus-infected cells without IFN-action [12]. The ISGs perform all physiological and pathological, including viral and non-viral, functions of IFNs. The ISGs function singly or in combination with other ISGs to inhibit virus replication. The antiviral activities of only a handful of these ISGs have so far been identified. Among them, Protein Kinase R (PKR), 2’5’ Oligoadenylate Synthetase (OAS), Mx1, IFN-induced protein with tetratricopeptide repeats (IFIT), tripartite motif (TRIM) family are most well-known for their antiviral activities against a wide spectrum of viruses in vitro and in vivo [13–20]. PKR, upon binding to viral double-stranded RNA (dsRNA), is activated and phosphorylates eukaryotic initiation factor (eIF2α), leading to the translational inhibition of cellular and viral mRNAs [21]. Mx1 is a broad antiviral ISG that acts at an early stage of virus replication, by sequestering the viral components from the desired destination within the cells [18]. OAS recognizes dsRNA and produces 2’,5’-oligoadenylates, which activate the latent ribonuclease, RNase L that degrades both cellular and viral RNAs [14]. The IFIT family of ISGs recognizes viral mRNAs and thereby inhibiting their translation [17, 19]. IFIT proteins also directly modulate cellular translation machinery by inhibiting eIF3 activities [22]. The TRIM family of proteins, which possesses E3 ubiquitin ligase activity, has diverse cellular functions [20]. In addition to directly interfering with virus life cycle, the ISGs often exert their antiviral actions by amplifying the cellular IFN responses [10]. Many of the ISGs also serve as PRRs or signaling intermediates, which are expressed at low levels and are transcriptionally induced by IFN signaling.
IFN can inhibit many stages of virus replication: viral entry, transcription, replication, translation, assembly or egress. IFN induced transmembrane proteins (IFITM) mediate antiviral resistance to a wide range of viruses [23]. IFITM proteins target the attachment and uncoating, two very early stages of viral entry, of several enveloped viruses [24]. Viperin inhibits Hepatitis C Virus (HCV) replication by localizing to the cellular lipid droplets, the site of viral replication [25]. Viperin also inhibits the budding and release of Influenza A virus by disrupting lipid rafts [26]. Tetherin (BST2) prevents the release of human immunodeficiency virus (HIV-1) by tethering HIV-1 virion particles to the cell surface [27]. In addition to targeting individual steps, multiple ISGs may target different steps of virus replication and elicit a large cumulative antiviral effect [3]. A single ISG can also function in cell type-specific manner to exhibit antiviral defense against multiple viruses. Ifit2 provides protection against a wide range of viruses in specific cells and tissues [17].
Viruses take advantage of cellular machineries or their components to achieve productive replication in the infected cells. Autophagy is an evolutionary conserved cellular degradation pathway, which has numerous physiological functions, including maintenance of cellular homeostasis and host defense [28–30]. Autophagy is induced by various cellular stresses, such as nutrient deprivation or microbial infection. Autophagy generates double-membranous cytoplasmic structure, known as autophagosome, which fuses with the lysosome, leading to the degradation of undesired cellular contents [31]. Adenosine monophosphate (AMP)-dependent Kinase (AMPK) directly phosphorylates the critical Ser/Thr residues of Unc-51 like autophagy activating kinase 1 (ULK1) to initiate the autophagy pathway [32]. Autophagy is regulated by class III PI3 Kinase (PI3K III) and mammalian target of rapamycin (mTOR) [33, 34]. How autophagy regulates virus replication is not clear; however, many viruses utilize autophagy or its components to promote their replication. Autophagy induced by virus infection can have both pro- or anti-viral effects [35–37]. Hepatitis C Virus (HCV), Dengue virus and Poliovirus use autophagosome membranes for their replication [38–40]. Human parainfluenza virus 3 (HPIV3) triggers autophagy in the infected cells and the viral P protein blocks the degradation of autophagosome, to enhance the intracellular virus yield [41]. Measles virus sequesters RIG-I within autophagosome, to evade the antiviral action of IFN [42]. In dendritic cells, Respiratory Syncytial Virus (RSV) induces autophagy, which regulates the adaptive immune responses [43]. How the IFN system regulates ‘virus-induced autophagy’ is unclear.
Paramyxoviruses are strong inducers of IFN and ISGs in the infected cells. To identify the ISGs that can inhibit their replication, we performed a high throughput genetic screen of individual ISGs for their ability to inhibit the replication of Sendai virus (SeV) (family: Paramyxoviridae, sub-family: Paramyxovirinae, genus: Respirovirus). Our screen identified a small subset of antiviral ISGs, including Tudor domain containing 7 (TDRD7), which strongly inhibited the replication of SeV in human and mouse cells. The TDRD family of proteins contains multiple Tudor domains and have roles in cellular RNA metabolism. Tdrd7-/- mice show defects in lens development and spermatogenesis, which are related to the deficiency in Tdrd7-associated mRNAs [44, 45]. Our results demonstrate that TDRD7 inhibits the replication of not only SeV, but also other paramyxoviruses, in multiple cell types. In-depth mechanistic studies revealed that the antiviral effect of TDRD7 is mediated by its ability to inhibit ‘virus-induced autophagy’, which is required for paramyxovirus replication.
To identify the ISGs that block SeV replication, we set up an unbiased genetic screen using a shRNA library against human ISGs (GIPZ lentiviral shRNAmir, Open Biosystems) in HeLa cells. The library is a 96-well formatted commercial system, which packages lentiviruses encoding individual ISG shRNA, a GFP reporter, and a puromycin selection cassette [46]. The GFP expression allowed a flow cytometry-based screen, in which lentivirus-transduced cells were tracked by GFP. The SeV-infected cells were stained with an antibody against the whole virion. In each well of cells, all ISGs were induced by IFN-β-pretreatment but the expression of only one of them was prevented by transduction of the cognate shRNA; the cells were then infected with SeV and the degree of virus replication was measured. Reversal of IFN-mediated inhibition of virus replication in a specific well indicated that the corresponding ISG was responsible for inhibiting SeV replication. We optimized the assay by knocking down the expression of genes required for IFN signaling and, therefore, would exhibit a phenotype. In HeLa cells, the knockdown of IRF9 enhanced the expression of viral protein (SeV C) in comparison to the NT control (Fig 1A, lanes 1, 3, 5). Moreover, IRF9 knockdown cells reversed the IFN-β-mediated inhibition of SeV C protein expression (Fig 1A, lanes 2, 4, 6). We used these cells to develop a flow cytometry-based quantitative assay to measure SeV replication as percent of SeV-positive cells within the GFP-expressing cell population (Fig 1B). The results indicate that knockdown of IRF9 reversed the IFN-β-mediated inhibition of SeV replication (Fig 1B). We quantified the GFP-expressing SeV-antigen positive cells (% infectivity) by flow cytometry (Fig 1C). IRF9 knockdown led to enhanced SeV replication (% infectivity) in both untreated and IFN-β-treated cells (Fig 1D). We screened the shRNA library, which consists of 814 lentiviral constructs (multiple targets for each ISG) expressing shRNA against more than 300 human ISGs. The shRNA library was lentivirally expressed in HeLa cells, using the strategy outlined in Fig 1E. A shRNA against IRF9 was used as an internal control to validate the effect of IFN-β (as in Fig 1D) for each screening experiment. Percent infectivity was quantified for each shRNA (as described in Fig 1C), and was used to calculate z-score for the individual shRNAs and normalized to that of IRF9 (Fig 2A and S1 Table) [47]. A set of 25 primary shRNA hits were shortlisted on the basis of high z-scores (>1.9, Fig 2B), and used for secondary validation. The primary hits consisted of both known, as well as some novel, antiviral ISGs. In order to further validate these primary hits, stable HeLa cell lines expressing the individual ISG shRNAs were generated. In the ISG shRNA-expressing cells, the reversal of IFN-β-mediated inhibition of SeV C protein expression was measured. As a control, we used IRF9-specific shRNA expressing cells, which alleviated the IFN-β-mediated suppression of SeV C protein level (Fig 2C, left). From the secondary validation assay, we narrowed down to a small subset of five ISGs, the knockdown of which reversed IFN-β-mediated inhibition of SeV C protein expression (Fig 2C and 2D). Importantly, individual knockdown of any of these five ISGs, elevated the level of SeV replication in IFN-β-treated cells (S1 Fig).
Among the five anti-SeV ISGs identified by our screen, we focused on TDRD7 (in human and Tdrd7 in mouse) because: (a) it is a cytosolic protein and, therefore, is a potential candidate to inhibit paramyxoviruses, which replicate in the cytosol, (b) it has defined functional domains, which may be required for its antiviral action, (c) it has no known functions as an ISG or a viral restriction factor and (d) our previous microarray results indicated its robust transcriptional induction by virus infection [48]. In HeLa cells, our screening system, TDRD7 was transcriptionally induced by IFN treatment (S2A Fig) and the stable knockdown of TDRD7 (S2A Fig) had no impact on the cell viability (S2B Fig). We further examined the inducibility of Tdrd7 in various mouse cells. In RAW264.7 cells, Tdrd7 mRNA was induced, as expected, by IFN-β; it was also induced by adding poly(I:C) to the medium to activate the TLR3 signaling pathway or transfecting poly(I:C) to activate the RLR pathway (Fig 3A). SeV infection, which activates the RLR pathway, induced Tdrd7 mRNA as well (Fig 3A). Similar induction by two strains of SeV was observed in mouse primary bone marrow-derived macrophages (BMDMs) and mouse embryonic fibroblasts (MEFs) (S2C–S2E Fig). The viral induction of Tdrd7 in MEFs was triggered by IRF3-mediated induction of IFN, because it was not observed in Irf3-/- or Stat1-/- cells (S2D and S2E Fig). Importantly, Tdrd7 mRNA was strongly induced in SeV-infected mouse lungs (Fig 3B). These results clearly demonstrate that Tdrd7, an ISG, is induced in virus-infected cells and tissues.
To investigate the antiviral action of TDRD7, we took two approaches: knockdown or knockout of the endogenous TDRD7 gene and ectopic expression of exogenous TDRD7 in multiple human and mouse cell types. In our test cells, we examined the endogenous levels of TDRD7 and as expected, the protein expression of TDRD7 varied between various cell types (S2F Fig). In mouse lung epithelial cells, LA4, the natural target cells of respiratory viruses, knockdown of endogenous Tdrd7 mRNA by two independent shRNAs enhanced the expression of SeV C protein (Fig 3C, S3A Fig). In human retinal epithelial cells, ARPE19, which express relatively higher levels of endogenous TDRD7 compared to LA4, knockdown of TDRD7 (Fig 3D, lower panel) elevated the expression of SeV C protein (Fig 3D). We confirmed these results in mouse fibroblasts, L929, in which the knockdown of endogenous Tdrd7 (S3B Fig) also enhanced SeV C protein expression (S3C Fig). Similar to HeLa, the stable knockdown of Tdrd7 in L929 cells had no impact on the cell viability (S3D Fig). We further examined the antiviral activity of Tdrd7 in non-transformed immortalized MEFs, in which Tdrd7 was transcriptionally induced by IFN-treatment (S3E Fig) and its stable knockdown (S3E Fig) enhanced the SeV C protein expression (S3F Fig). Using CRISPR/Cas9 system, we generated TDRD7 knockout (TDRD7-/-) human HT1080 cells (Fig 3E, lower panel and S3G Fig). SeV C protein expression was elevated in TDRD7-/- cells (Fig 3E, upper panel). In the reciprocal strategy, we ectopically expressed TDRD7 (untagged or V5-tagged) in cells that express low levels of endogenous TDRD7, a scenario that mimics IFN-β-induced synthesis of ISGs. TDRD7 was ectopically expressed in HEK293T cells and confocal analyses showed its cytoplasmic distribution in uninfected cells (S3H Fig). In these cells, TDRD7 strongly inhibited viral protein (SeV C) (Fig 3F) and mRNA (SeV P mRNA, S3I Fig) expression. Similarly, ectopic expression of Tdrd7 inhibited viral protein (SeV C) expression in mouse L929 (S3J Fig) and LA4 (S3K Fig) cells. We investigated whether the TDRD7-mediated inhibition of viral mRNA and protein leads to the reduction of infectious virion production. Indeed, the production of infectious virus particles was inhibited by ectopic expression of Tdrd7 in L929 cells (Fig 3G). In subsequent experiments, we investigated the mechanism of anti-SeV action of Tdrd7.
Because the role of autophagy in SeV replication was not clear, we examined various stages of the autophagy pathway (Fig 4A) in SeV-infected cells. In L929 cells, SeV infection triggered robust induction of autophagy, which was analyzed by the degradation of p62 (Fig 4B, upper panel) and the generation of LC3-II (Fig 4C), the indicators of two late stages of the autophagy pathway. The induction of autophagy was correlated with the expression of viral protein (SeV C, Fig 4B, middle panel). In another cell type (LA4), SeV infection similarly triggered autophagy pathway, which was examined by LC3-II generation and p62 degradation (S4A Fig). The molecular signatures of the early stages of autophagy, the dephosphorylation of Ser757 and phosphorylation of Ser317 of ULK1, were also detected in SeV-infected cells (Fig 4D). These results clearly indicate that SeV infection triggers different stages of the autophagy pathway in multiple cell types. To investigate whether ‘virus-induced autophagy’ is required for SeV replication, we took pharmacological and genetic approaches. A chemical inhibitor of autophagy, 3-MA, inhibited, whereas an activator of autophagy, rapamycin, promoted SeV mRNA synthesis (Fig 4E). Furthermore, chemical inhibitors of various stages of autophagy also significantly suppressed the expression of SeV C protein (S4B Fig). Similarly, knockdown of ATG5 (Fig 4F, lower panel), a key component of autophagy pathway, significantly reduced viral protein expression in human cells (SeV C, Fig 4F). As expected, the SeV-induced autophagy, examined by p62 degradation, was inhibited in these cells (Fig 4F, p62 levels). Collectively, our results clearly indicate that SeV-induced autophagy pathway is required for its replication.
The above results led to the hypothesis that TDRD7 might interfere with ‘virus-induced autophagy’ to inhibit SeV replication. To test this, we examined various stages of SeV-induced autophagy in cells, in which TDRD7 expression had been modulated either by ectopically expressing exogenous TDRD7 or by ablating the endogenous TDRD7 levels. In L929 cells, ectopic expression of Tdrd7 strongly inhibited the degradation of p62 (Fig 5A), and the accumulation of LC3-II (Fig 5B). As expected, the knockdown of endogenous TDRD7 in ARPE19 cells, triggered increased accumulation of LC3-II (S4C Fig). Similar results were also obtained in Tdrd7-ablated murine macrophages (RAW264.7) (S4D and S4E Fig). In these cells, Tdrd7 knockdown also elevated SeV C expression (S4E Fig, middle panel). We further investigated whether the antiviral action of TDRD7 is mediated by direct inhibition of autophagy or its ability to modulate IFN and ISG induction. The inhibition of autophagy pathway by knockdown of ATG5 in TDRD7-/- cells suppressed SeV C protein expression (Fig 5C). As expected, these cells restored the ability of IFN to inhibit SeV C expression (Fig 5C). Furthermore, Tdrd7 knockdown cells did not exhibit any significant difference in induction of IFN by SeV (S4F Fig) and ISG (Ifit1) by IFN (S4G Fig). These results demonstrated that the antiviral ISG, TDRD7, inhibits ‘virus-induced autophagy’ to control SeV replication.
Because autophagy is required for normal cellular homeostasis, it is also induced in response to many cellular stresses, other than virus infection. In the next series of experiments, we examined whether TDRD7 could inhibit autophagy induced by nutrient deprivation (e.g. serum starvation, SS and Hank’s Balanced Salt Solution, HBSS) or rapamycin, known activators of the autophagy pathway. As indicated by the LC3-II level, SS-induced autophagy was inhibited by ectopic expression of TDRD7 in human HEK293T cells (Fig 5D). Similarly, in mouse L929 cells, ectopic expression of Tdrd7 inhibited degradation of p62 (Fig 5E) and enhancement of LC3-II levels (S5A Fig). Accumulated LC3-II produces cytoplasmic puncta, which we analyzed by expressing a GFP-LC3 fusion protein in L929 cells. The cytoplasmic puncta formation by LC3-II was significantly reduced in Tdrd7-expressing cells (Fig 5F). Similarly, rapamycin-induced LC3-II formation in L929 cells was inhibited by Tdrd7 (Fig 5G). In Tdrd7-knockdown RAW264.7 cells, rapamycin induced a faster degradation of p62 (S5B Fig) and increased accumulation of LC3-II (S5C Fig). These results demonstrated that Tdrd7 can inhibit autophagy, induced by both viral and non-viral agents, in various human and mouse cell types.
Next, to identify the specific target of Tdrd7, we biochemically analyzed the four stages of the autophagy pathway (Fig 4A); for this purpose we used SS of L929 cells to induce autophagy. Because we already knew that the ‘fusion’ (p62 degradation) and the ‘elongation’ (LC3-II levels) steps of the autophagy pathway were inhibited by Tdrd7, we focused our attention to the further upstream pre-elongation steps of the autophagy pathway. As a readout of the ‘nucleation’ step, we analyzed PI3 kinase III activity by direct measurement of PI(3)P produced in virus-infected cells and by the generation of NADPH Oxidase (phox) [49, 50]. In SeV-infected L929 cells, PI3K III was rapidly activated and its activity was inhibited by Tdrd7 expression (Fig 6A, S5D Fig). We further expressed a GFP-conjugated p40 subunit of phox in L929 cells, which produced puncta structures upon autophagy stimulation [49]. Ectopic expression of Tdrd7 decreased the number of phox puncta structures (Fig 6B, left panel), which was quantified by counting the number of puncta structures per GFP-expressing cell (Fig 6B, right panel). These results established that the Tdrd7-mediated block is at the ‘nucleation’ step of the autophagy pathway or further upstream. We examined the effects of TDRD7 on the activation of ULK1, a kinase essential for triggering of the induction stage. ULK1 is activated by phosphorylation of multiple Ser/Thr residues, among which Ser317 is a critical residue that is directly phosphorylated by the upstream kinase AMPK [51]. SS triggered robust phosphorylation of ULK1 on Ser317, which was inhibited by ectopic Tdrd7 expression (Fig 6C). Activated AMPK inhibits the activity of mTOR, another Ser/Thr kinase, which phosphorylates ULK1 on Ser757 to inhibit the autophagy pathway [51]. Therefore, dephosphoryation of Ser757 of ULK1 is a positive trigger of the autophagy pathway [51]. SS-induced dephosphorylation of ULK1 (Ser757) was also strongly inhibited by Tdrd7 (Fig 6D).
These results pointed to AMPK as the target of TDRD7. AMPK was required for SeV replication: in HeLa cells, knockdown of endogenous AMPK suppressed the expression of SeV C protein (Fig 6E) and ectopic expression of AMPK strongly enhanced SeV C protein expression in L929 cells (Fig 6F). To determine whether only the physical presence or the kinase activity of AMPK is required for virus replication, we used a small molecule chemical inhibitor of AMPK kinase activity, Compound C (CC). Treatment of cells with CC inhibited SeV C protein expression in both human and mouse cells (Fig 6G and 6H), demonstrating that AMPK enzyme activity is required for SeV replication. As expected, CC treatment inhibited phosphorylation of AMPK (on Thr172) upon SeV infection (Fig 6G and 6H). We investigated the effect of Tdrd7 on the activation of AMPK, by monitoring the phosphorylation of its Thr172. AMPK was rapidly phosphorylated by SS of cells, but Tdrd7 expression inhibited this phosphorylation (Fig 6I). Similarly, HBSS-induced phosphorylation of AMPK on Thr172 was also inhibited by Tdrd7 (S5E Fig). Importantly, SeV infection strongly activated AMPK and this step was inhibited by Tdrd7 (Fig 6J). Together, our results clearly demonstrated that Tdrd7 is an inhibitor of the autophagy-inducing kinase AMPK and the antiviral action of Tdrd7 is mediated by its ability to inhibit AMPK activation.
Next, we investigated whether TDRD7 can inhibit the replication of other paramyxoviruses by inhibiting autophagy. We chose two clinically important human paramyxoviruses, HPIV3 (family: Paramyxoviridae, sub-family: Paramyxovirinae, genus: Respirovirus) and RSV (family: Paramyxoviridae, sub-family: Pneumovirinae, genus: Pneumovirus), to examine the generality of TDRD7 action. HPIV3 infection triggered robust autophagy in the infected cells, as examined by the increased LC3-II levels and p62 degradation (Fig 7A). In ATG5 knockdown cells, HPIV3 replication was strongly inhibited, as manifested by the expression of virus-encoded GFP (Fig 7B) and the viral structural protein, HN (Fig 7C). Similar to SeV and HPIV3, RSV infection also triggered autophagy (S6A Fig) and its replication was inhibited in ATG5-knockdown human cells (S6B Fig). To investigate whether TDRD7 inhibits the replication of HPIV3 and RSV, we used TDRD7-expressing human cells. Ectopic expression of TDRD7 inhibited HPIV3 replication, which was examined by both GFP and viral HN expression (Fig 7D and 7E); IFN-β-treatment, a known inhibitor of HPIV3 replication, was used as a positive control. Similar to HPIV3, ectopic expression of TDRD7 strongly inhibited RSV replication, examined by the expression of viral proteins in the infected cells (Fig 7F). As anticipated, AMPK was required for the replication of both HPIV3 and RSV. AMPK-knockdown cells expressed reduced levels of HPIV3-encoded viral protein (HN) (Fig 7G) and GFP (Fig 7H). They also exhibited reduced RSV replication, as indicated by virus-encoded red fluorescent protein expression (Fig 7I). Similar to SeV, HPIV3 infection caused phosphorylation of AMPK and as expected, this was inhibited by CC treatment (Fig 7J). The treatment with CC caused strong reduction of HPIV3 HN protein expression (Fig 7J). In addition to the inhibition of viral protein expression, CC strongly inhibited the production of infectious HPIV3 particles (Fig 7K) and infectious RSV particles (Fig 7L). Our results clearly demonstrated that the kinase activity of AMPK was required for the replication of paramyxoviruses and TDRD7 inhibited their replication by inhibiting AMPK activation. Finally, to examine the specificity of TDRD7 action, we chose a member of another virus family, encephalomyocarditis virus (EMCV; family: Picornaviridae, genus: Cardiovirus). EMCV replication, analyzed by the expression of viral RNA polymerase (3DPol), was not inhibited but enhanced by the ectopic expression of TDRD7 (S7 Fig). These results clearly established the specificity of antiviral action of TDRD7 against viruses from different families.
Here, we report a novel mechanism by which the interferon system provides an antiviral response against paramyxoviruses (Fig 8). Using a high throughput genetic screen, we have identified a viral restriction factor, TDRD7, which inhibited paramyxovirus-induced autophagy, a critical step of viral life cycle. Our mechanistic studies revealed that TDRD7 blocks the activation of AMPK, the enzyme that triggers autophagy. As expected, a chemical inhibitor of AMPK’s downstream activities also restricted the replication of paramyxoviruses. Because TDRD7 is a newly discovered viral restriction factor and its antiviral action is novel, we validated the critical results in multiple human and mouse cell types (S2 Table). We, therefore, present a new mechanism by which the IFN system not only provides antiviral protection, but also controls cellular metabolic activity.
In search for a common cellular mechanism that the paramyxoviruses utilize, we uncovered a role of autophagy, which was robustly induced in the early phase of viral life cycle and was required for a stage prior to the transcription of viral mRNA. As a model paramyxovirus, we used SeV, also known as mouse parainfluenza virus type I, because of its wide range of infectivity in vitro. SeV triggered a pro-viral autophagy pathway in the infected cells to facilitate its replication. Chemical inhibitors of autophagy blocked, whereas an activator of autophagy promoted, SeV replication. Genetic deficiency of the autophagy pathway significantly inhibited SeV replication. Many RNA viruses use the autophagy pathway to promote their replication [38–42]. Hepatitis C Virus (HCV), Dengue virus and Poliovirus directly use the autophagosome membranes to facilitate their replication [38–40]. Measles virus triggers autophagy to sequester RIG-I in the autophagosome, to inhibit the antiviral action of IFN [42]. In contrast, autophagy impairs the replication of some DNA viruses. In neurons, autophagy is considered an antiviral response against Herpes Simplex Virus (HSV-1) replication [52]. HSV-1 neurovirulence factor, ICP34.5 inhibits autophagy to support virus replication and pathogenesis [53]. However, HSV-2 and Varicella Zoster Virus (VZV), two other α-herpesviruses require basal autophagy to promote their replication [53–56]. The members of γ-herpesviruses, Kaposi’s Sarcoma-associated Herpesvirus (KSHV), Epstein-Barr virus (EBV) antagonize cellular autophagy using viral homologue of Bcl-2 [57]. It will be interesting to investigate whether these viruses can induce TDRD7 and whether TDRD7 has any effect on their replication.
We demonstrated that AMPK, the initiator kinase of the autophagy pathway, is involved in paramyxovirus replication. AMPK was activated by SeV and HPIV3 infection and its activity was required for virus replication. AMPK has multiple cellular functions in addition to initiating autophagy. These include cell growth, mitochondrial biogenesis, and lipid and glucose metabolism [34]. By activating AMPK, the paramyxoviruses may also activate the autophagy-independent activities whose contributions to virus replication remain to be explored. However, our results clearly established that the autophagy-inducing effect of AMPK was critical for virus replication, because ablation of downstream components of the autophagy pathway had similar inhibitory effects. Because paramyxoviruses utilize the autophagy pathway, the role of AMPK in virus replication may be restricted to the activation of autophagy. Whether a specific component of the autophagy pathway, downstream of AMPK, is utilized by the paramyxoviruses, will be investigated in the future. Previous studies have indicated a role of lipophagy and macropinocytosis in promoting AMPK-induced virus replication. Dengue virus promotes AMPK-mTOR signaling pathway to promote lipophagy, a selective autophagy that targets lipid droplets [58]. Vaccinia virus activates AMPK to promote macropinocytosis and actin dynamics, which are required for viral entry into the cells [59]. A kinome screen has revealed that AMPK activity is required for HCMV replication; however, the exact mechanism is currently unknown [60]. KSHV directly interacts with AMPK via its K1 viral protein to promote cell survival [61]. In contrast to these, the picornavirus EMCV replication was promoted in the presence of TDRD7. The exact mechanism behind this will require further investigation.
How the IFN system regulates the autophagy pathway is largely unexplored; however, IFN treatment of cancer cells triggers autophagy via PI3K/mTOR signaling [62]. Some ISGs, e.g. PKR and RNase L promote autophagy in virus-infected cells as antiviral defense mechanisms [63–65]. Several studies indicate crosstalk between autophagy and innate immune signaling pathways; components of autophagy pathway are required for RIG-I signaling [66]. Autophagy is required for TLR7-induced type I IFN production by VSV-infected plasmacytoid dendritic cells [67]. Our study provides a new mechanism of IFN-mediated control of virus replication via inhibition of the autophagy pathway.
Because TDRD7 prevents AMPK activation by non-viral stresses, such as nutrient deprivation, it may have broader effects on AMPK-dependent cellular processes in uninfected cells. For example, IFN is expressed in a low amount in uninfected immune cells, e.g. dendritic cells. The development and biological functions of these cells may be regulated by TDRD7’s anti-AMPK activity. Autophagy is required for cellular homeostasis and unregulated autophagy may lead to disease conditions [68]. In these scenarios, TDRD7-mediated autophagy inhibition would be beneficial. IFN signaling has been shown to inhibit AMPK activation [69]; however, the mechanisms or the involved ISGs are unknown. Our results indicate that TDRD7 is one of the executioner ISGs for the anti-AMPK activity of IFN. AMPK is involved in multiple cellular functions and is activated when cellular ATP levels are low, a scenario that mimics virus infection, which requires high metabolic activity of the infected cells. IFN signaling also regulates the activity of mTOR, a kinase that is involved in protein synthesis. AMPK directly inhibits the activity of mTOR by interacting with intermediate proteins [51, 70]. Therefore, the IFN-mediated inhibition of AMPK may further activate mTOR signaling to enhance the synthesis of desired proteins in the virus-infected cells. Because TDRD7 is known to interact with cellular RNAs [44], this activity may be required for its anti-AMPK functions. AMPK can be activated by long non-coding RNA (lncRNA) [71], and it is speculative that sequestration of the lncRNA activator by TDRD7 may give rise to its AMPK-inhibitory action. Future investigation will be required to explore this possibility.
How paramyxoviruses activate AMPK is an interesting question. To determine whether paramyxoviruses directly trigger AMPK activation by its interaction with viral proteins or by activation of upstream signaling pathways, will require additional investigation. Answers to these questions will lead to therapeutic potential by targeting the virus-AMPK interaction. Because temporary AMPK inhibition is not toxic, the chemical inhibitor is a potential candidate for antiviral therapy. However, AMPK is involved in modulating the functions of both innate immune cells, e.g. macrophages, and the adaptive immune cells, e.g. T cells [72, 73]. Therefore, the use of AMPK inhibition as an antiviral strategy will require in-depth investigation of both innate and adaptive immune responses in virus-infected host. Because type I IFN is also involved in regulating the functions of these cells, TDRD7 may also contribute to the regulation of immune cell functions. TDRD7 knockout mice have only been examined for its role in lens and germ cell development [44, 45]. Because TDRD7 is not present in high amount in majority of the cell types, the transcriptional induction by viruses or IFN exposure will uncover its new roles in other cell types as well. Whether the previously identified functions of TDRD7 can also be regulated by the IFN system will require further investigation.
Human cell lines HeLa, HT1080, ARPE19, HEK293T, and mouse cell lines L929, LA4, MEFs, RAW264.7 were maintained in DMEM containing 10% FBS, penicillin and streptomycin. All cell lines used in this study were maintained in the laboratory. Expression vectors of human and mouse TDRD7/Tdrd7 gene (untagged) were obtained from Origene and sub-cloned into lentiviral vector (pLVX-IRES-puro, V5-tagged). AMPK expression plasmid was obtained from Addgene and was sub-cloned into lentiviral vector (pLVX-IRES-puro, HA-tagged AMPK) and Flag.VPS34 plasmid was obtained from Addgene. Autophagy inhibitors (3-MA, chloroquine, quinacrine, bafilomycin-1) or activators (rapamycin), AMPK inhibitor (Compound C) were obtained from Sigma-Aldrich, human and mouse IFN-β were obtained from R&D, MTT was obtained from Fisher Scientific, and Lipofectamine 2000 was obtained from Thermo Fisher Scientific. The antibodies against the specific proteins were obtained as indicated below: anti-SeV C: raised in the authors’ laboratory [74], anti-whole SeV antibody was a gift from John Nudrud (Case Western Reserve University), anti-HPIV3 HN: Abcam, anti-RSV: Abcam, anti-3DPol: Santa Cruz, anti-TDRD7: Sigma-Aldrich, anti-LC3: Cell Signaling, anti-p62: Fitzgerald, anti-pULK1(Ser757)/anti-pULK1(Ser317)/anti-ULK1: Cell Signaling, anti-pAMPK(Thr172)/anti-AMPK: Cell Signaling, anti-ATG5: Cell Signaling, anti-Actin: Sigma, anti-V5: Thermo Fisher Scientific, anti-Flag: Sigma-Aldrich, anti-Flag-agarose beads: Sigma-Aldrich.
A custom generated, lentivirus-based shRNAmir library against human ISGs was obtained from Open Biosystems. The seed sequences for shRNA targeting each ISG have been described before [46]. The vector was designed to co-express the shRNA and GFP by a cytomegalovirus (CMV) promoter. The shRNA plasmids were packaged into lentiviral vectors in 96-well plates using the manufacturer’s instructions. HeLa cells were transduced with these lentiviruses for 48 h, when the cells were treated with 1000 U/ml of IFN-β. Twenty four hours later, the cells were infected with SeV (Cantell) at an MOI of 10. After 16 h, cells were harvested, fixed with 1% paraformaldehyde in phosphate-buffered saline (PBS) for 10 min, permeabilized with 0.1% (wt/vol) saponin, and incubated with an anti-SeV polyclonal antibody and an Alexa Fluor 647-conjugated goat anti-rabbit secondary antibody. Cells were analyzed using a BD LSRFortessa flow cytometer (BD Biosciences) and the data were analyzed by FlowJo. Viral infection was determined based on the percentage of SeV-positive cells in shRNA-transduced (GFP) populations (% infectivity, as illustrated in Fig 1C). The relative infectivity in each well was normalized to the wells containing a shRNA against IRF9 sequence to obtain z-scores [47]. Independent lentivirus stocks were used to validate the primary screen results in three independent replicates. Primary hits were defined as those z scores greater than 1.9.
The primary hits obtained from the high throughput screen were validated by stably expressing the respective ISG shRNAs (which were positive in the primary screen) in HeLa cells. We used non-targeting (NT) and IRF9-specific shRNAs as controls. The HeLa cells, stably expressing the shRNAs against the shortlisted ISGs were pre-treated with IFN-β followed by infection with SeV Cantell (moi: 10) and viral protein (SeV C) expression was analyzed by immunoblot.
For generating stable knockdown of TDRD7/Tdrd7 genes in human and mouse cells, the respective shRNAs [from the shRNA library, Open Biosystems (TDRD7: GATCGCACATGTTTATTTA, used in all human cells of the study), (Tdrd7#1: CAGGATTTGCCTCAGATTA, used in all mouse cells of the study) or Sigma (Tdrd7#2, SHCLNG-NM_146142, TRCN0000102515, used only in LA4 cells)] were lentivirally expressed and the transduced cells were selected in puromycin containing medium. The stable knockdown cells were evaluated for levels of TDRD7/Tdrd7 by qRT-PCR in the absence or the presence of IFN-treatment or immunoblot. These and the control (NT) cells were evaluated for viral replication. ATG5-specific shRNAs (Sigma # SHCLNG-NM_004849) were stably expressed lentivirally and the transduced cells were selected in puromycin containing medium. AMPK knockdown cells were generated using lentiviral shRNA plasmids (Sigma # SHCLNG-NM_006251) and the transduced cells were selected in puromycin containing medium. Stable cell lines ectopically expressing epitope-tagged human and mouse TDRD7/Tdrd7 using lentiviral delivery systems (pLVX-IRES-puro) and were selected in puromycin containing medium. The stable cells were used for viral infection and other biochemical analyses. Wherever indicated, the stable cell lines were also generated by transfecting the untagged TDRD7/Tdrd7 plasmids (from Origene) and selecting the transfected cells with puromycin. Cells ectopically expressing AMPK were generated by lentivirally transducing HA.AMPK using pLVX-IRES-puro and selecting the cells in puromycin containing medium.
HT1080 cells were transfected with either control (sc-418922) or TDRD7-specific (sc-407210) CRISPR/Cas9 plasmids. Transfected cells were sorted for high GFP-expressers using flow cytometry, and the GFP-expressing cells were expanded to isolate individual clones. These clones were examined for TDRD7 mRNA levels by qRT-PCR analyses and protein levels by immunoblot.
SeV Cantell (VR-907) and 52 (VR-105) strains were obtained from Charles River, and the infection procedure has been previously described [74, 75]. Briefly, the cells were infected by the viruses (at an MOI specified in the figure legends) in serum-free DMEM for 1.5 h, after which the cells were washed and replaced with normal growth medium. The virus-infected cells were analyzed at the indicated time for viral protein expression or as described in figure legends. For quantification of infectious SeV particles in the culture medium, standard plaque assays were performed, as described previously [74, 75]. Recombinant RSV (rrRSV) and HPIV3 (rgHPIV3) infections were carried out in serum-free DMEM at the indicated MOI. Infectious rrRSV and rgHPIV3 particles were analyzed by quantification of fluorescent foci forming units or the virus-infected cells were photographed using fluorescence microscope. EMCV infection was carried out using previously published procedure [76], and the infected cells were analyzed by the expression of viral protein, as indicated in the figure legends.
For analyses of nutrient starvation induced autophagy, the cells were washed (three times) with and then incubated in serum-free DMEM or HBSS (Lerner Research Institute Cell Culture Core) for the time period indicated in the figure legends. At the end of the incubation period, the cells were harvested and the lysates were analyzed for pAMPK (Thr172), pULK1 (Ser317 and Ser757), LC3, p62. Similarly, for the analysis of puncta formation by GFP-LC3 or GFP.p40.PHOX, the cells were transiently transfected with these plasmids and then serum starvation was carried out. At the end of the exposure, the cells were fixed and confocal microscopy was performed. The puncta structures were manually counted in GFP-expressing cells from multiple fields.
Immunoblot was performed using previously described procedures [74, 75]. Briefly, cells were lysed in 50 mM Tris buffer, pH 7.4 containing 150 mM of NaCl, 0.1% Triton X-100, 1 mM sodium orthovanadate, 10 mM of sodium fluoride, 10 mM of β-glycerophosphate, 5 mM sodium pyrophosphate, protease and phosphatase inhibitors (Roche). Total protein extracts were analyzed by SDS-PAGE followed by immunoblot. The density of protein bands on the immunoblots was quantified using Image J program.
Total RNA was isolated using RNA isolation kit (Roche) and cDNA was prepared using ImProm-II Reverse Transcription Kit (Promega). For qRT-PCR, 0.5 ng of cDNA was analyzed using Applied Biosystem's Power SYBR Green PCR mix in Roche LightCycler. The expression levels of the mRNAs were normalized to 18S rRNA. To investigate in vivo gene expression, lungs were harvested from the SeV-infected mice and quickly frozen on dry ice. Total RNA was isolated from frozen lungs using Trizol extraction and the cDNAs were prepared using ImProm-II Reverse Transcription Kit and then subjected to qRT-PCR analyses as described above [74]. For the qRT-PCR analyses of the respective genes, the following primers were used: TDRD7-fwd: CGAGCTGTTCTGCAGTCTCA, TDRD7-rev: GCCATGGCATAGCAGGTAAT, Tdrd7-fwd: CTAAGGGCTGTCCTGCAGTC, Tdrd7-rev: AGAGTTGCCTTTGGCTTT, SeV P-fwd: CAAAAGTGAGGGCGAAGGAGAA, SeV P-rev: CGCCCAGATCCTGAGATACAGA, Ifnb-fwd: CTTCTCCGTCATCTCCATAGGG, Ifnb-rev: CACAGCCCTCTCCATCAACT, Ifit1-fwd: CAGAAGCACACATTGAAGAA, Ifit1-rev: TGTAAGTAGCCAGAGGAAGG, 18S-fwd: ATTGACGGAAGGGCACCACCAG, 18S-rev: CAAATCGCTCCACCAACTAAGAACG.
HEK293T cells expressing V5.TDRD7 were grown on coverslips, fixed in 4% paraformaldehyde, permeabilized in 0.2% Triton X-100 and subjected to immunostaining by anti-V5 antibody followed by Alexa Fluor-conjugated secondary antibody. The objects were mounted on slides using VectaShield/DAPI and analyzed by confocal microscopy. For GFP.LC3 and GFP.p40.PHOX analyses, the cells expressing these plasmids were analyzed by confocal microscopy. The images were further processed and analyzed using Adobe Photoshop software. Multiple culture fields (at least 100 cells from more than 20 fields) were analyzed to select representative images and for quantification.
L929 cells, transfected with Flag.VPS34 and V5.Tdrd7, were infected with SeV for the indicated time (in figure legends), when the cell lysates were immunoprecipitated with Flag-agarose beads. The immunoprecipitates were analyzed for PI3K III activity by measuring the PI(3)P levels using Class III PI3K ELISA kit (Echelon) following manufacturer’s instructions. The PI(3)P levels in the mock-infected VPS34-expressing cells was expressed as 100 and all other values were normalized to this.
Cells at the density of 10,000/well were cultured in 96 well plates for 2 days followed by addition of 10μl of 5mg/ml MTT solution in PBS and additional culturing for 4 hours. The water insoluble formazan was dissolved in DMSO and the absorbance was measured at 570nm [77]. The absorbance in control (NT) cells was expressed as 100 and all other values were normalized to this.
C57BL/6 Wt mice, obtained from Taconic, were either mock-infected (PBS) or intranasally infected with SeV (52 strain, 120,000 pfu), as described previously [74, 78]. The lungs were harvested after 2 days of infection, total RNA was isolated and analyzed by qRT-PCR.
The statistical analyses were performed using GraphPad Prism 5.02 software. The ‘p’ values were calculated using two-tailed, un-paired Student’s t tests and are shown in the relevant figures. The results presented here are the representatives of at least three biological repeats.
|
10.1371/journal.pgen.1004230 | Dysregulated Estrogen Receptor Signaling in the Hypothalamic-Pituitary-Ovarian Axis Leads to Ovarian Epithelial Tumorigenesis in Mice | The etiology of ovarian epithelial cancer is poorly understood, mainly due to the lack of an appropriate experimental model for studying the onset and progression of this disease. We have created a mutant mouse model in which aberrant estrogen receptor alpha (ERα) signaling in the hypothalamic-pituitary-ovarian axis leads to ovarian epithelial tumorigenesis. In these mice, termed ERαd/d, the ERα gene was conditionally deleted in the anterior pituitary, but remained intact in the hypothalamus and the ovary. The loss of negative-feedback regulation by estrogen (E) at the level of the pituitary led to increased production of luteinizing hormone (LH) by this tissue. Hyperstimulation of the ovarian cells by LH resulted in elevated steroidogenesis, producing high circulating levels of steroid hormones, including E. The ERαd/d mice exhibited formation of palpable ovarian epithelial tumors starting at 5 months of age with 100% penetrance. By 15 months of age, 80% of ERαd/d mice die. Besides proliferating epithelial cells, these tumors also contained an expanded population of luteinized stromal cells, which acquire the ability to express P450 aromatase and synthesize E locally. In response to the elevated levels of E, the ERα signaling was accentuated in the ovarian epithelial cells of ERαd/d mice, triggering increased ERα-dependent gene expression, abnormal cell proliferation, and tumorigenesis. Consistent with these findings, treatment of ERαd/d mice with letrozole, an aromatase inhibitor, markedly reduced circulating E and ovarian tumor volume. We have, therefore, developed a unique animal model, which serves as a useful tool for exploring the involvement of E-dependent signaling pathways in ovarian epithelial tumorigenesis.
| Ovarian cancer is currently the most lethal gynecological cancer in the United States. Multiple epidemiological studies indicate that women who take hormone replacement therapy, estrogen or estrogen with progesterone, peri- or postmenopause will have an increased chance of developing ovarian cancer. Unfortunately, the five-year survival rate after diagnosis is very low indicating that better tools are needed to diagnose and treat ovarian cancer. The models that would allow investigation of this disease are severely limited. In this article we introduce a mouse model that develops epithelial ovarian tumors, and by employing inhibitors of estrogen synthesis, we show that ovarian tumorigenesis in this model is dependent on estrogen production within the ovarian tumor. These studies suggest that estrogen may play a role in promoting ovarian tumor growth.
| Ovarian cancer is the most lethal malignancy of the female reproductive system and the fifth leading cause of cancer-related death among women [1]. Approximately 90% of malignant ovarian tumors are derived from either the ovarian surface epithelium (OSE) or fallopian tube epithelium (FTE) [2]. Due to the absence of specific symptoms and the lack of strategies for early detection of ovarian cancer, the majority (70%) of women with this disease are diagnosed at a late stage when the cancer has spread beyond the confines of the ovary [1]. Despite its clinical significance, the etiology of ovarian cancer is poorly understood, mainly due to the lack of an appropriate experimental model for studying the onset and progression of this disease.
Multiple theories regarding the etiology of ovarian cancer have been proposed, but the precise molecular defects underlying the development of this disease remain elusive [3]. The “gonadotropin hypothesis” proposes that high gonadotropin levels can have a stimulatory effect on OSE cells, promoting their neoplastic transformation [4], [5]. It was reported that the addition of gonadotropins to rodents in which ovarian cancer was induced upon treatment with the chemical carcinogen, 7,12-dimethylbenz(a)anthracene (DMBA) led to increased lesion severity, suggesting that gonadotropins play a role in tumor progression [6]. In humans, epidemiologic evidence, indirectly supporting this hypothesis, includes the well-documented protective effects of oral contraceptives and multiparity, which suppress gonadotropin secretion by the pituitary gland [5], [7]. The majority of women with epithelial ovarian cancer present the disease at a postmenopausal stage where circulating follicle stimulating hormone (FSH) and lutenizing hormone (LH) levels are elevated, indicating a causal relationship between chronically elevated gonadotropin levels and ovarian cancer development [5], [8].
Besides gonadotropins, epidemiological studies have reported altered ovarian cancer risk associated with the use of steroid hormones to ease menopausal symptoms. Estrogen (E) is a well-known mitogenic factor associated with the genesis of many cancers. It has been reported previously that the risk of developing ovarian cancer increases in women who use hormone replacement therapy (HRT) for more than five years or use E-only regimens [9]–[13]. While most of these studies comprise a small number of subjects and fail to control for all of the factors that may influence cancer risk, in patients with ovarian cancers, intratumoral production of E via in situ aromatization has been suggested to promote growth of breast, endometrial and ovarian cancer cells [14]. However, only few animal models have been used to investigate the role of E in ovarian tumorigenesis. Bai et al reported the effects of prolonged E exposure on the morphology of rabbit ovaries and found an increase in both OSE cell proliferation and the number of papillae covering the ovarian surface, but no ovarian tumors [15]. In a recent study, Laviotte et al conditionally activated an oncogene, SV40 TAg, in OSE cells and treated the mice with exogenous E [16]. These investigators reported that E treatment resulted in an earlier onset of ovarian tumors and a significantly decreased survival time [16]. While the results from this animal model underscore the importance of E in the progression of ovarian cancer, it is clear that new animal models independent of specifically directed single oncogenic mutations are needed for assessment of the role of E signaling in ovarian epithelial tumorigenesis.
In this study, we present a novel transgenic mouse model of ovarian tumorigenesis. In this model, termed ERαd/d, the estrogen receptor alpha (ERα) gene is dysregulated in the hypothalamic-pituitary-ovarian axis. Conditional deletion of this gene in the anterior pituitary, but not in the hypothalamus and the ovary, led to elevated circulating LH. Hyperstimulation by LH resulted in luteinization of the ovarian stromal cells, expression of P450 aromatase in these cells, and increased E synthesis in the ovarian microenvironment. Our study suggests that E critically controls ovarian tumor growth, presumably by stimulating the proliferation of OSE cells to drive epithelial tumorigenesis. The ERαd/d mouse, therefore, provides a useful model to study the mechanisms by which dysregulated E signaling promotes the initiation and progression of ovarian epithelial tumors.
ERα conditional knockout mice (ERαd/d) were generated by crossing progesterone receptor cre recombinase (PR-Cre) knock-in mice with ERα floxed (ERαf/f) mice [17], [18]. By five months of age, the ERαd/d mice developed palpable ovarian tumors with 100% penetrance. In contrast, the ERαf/f and the global ERα knockout mice did not develop any tumor (Fig. 1A). The ovarian tumors of ERαd/d mice grew progressively with age and became as large as 11 mm in size with an average weight of 300 mg by eight months of age (Fig. S1A). Because of this large tumor burden, 80% of the ERαd/d mice die by 68 weeks of age (Fig. S1C). Histological analyses of the ERαd/d ovaries showed cystic hemorrhagic follicles at 3 months of age. By 6 months, there was evidence of neoplastic epithelial cells migrating into the ovarian stroma, and by 11 months, extensive cellular proliferation occurred, resulting in the formation of a large tumor mass (Fig. S1B). Immunohistochemical analysis of ovaries of ERαf/f mice at 6 months of age, using cell proliferation markers, revealed that the follicular granulosa cells were proliferative but the OSE cells were quiescent (Fig. 1B, panel a,c). In sharp contrast, both OSE and the tumor cells within ERαd/d ovaries exhibited pronounced proliferative activity (panels b, d; proliferative cells indicated by arrow) (Fig. 1B).
We next assessed the expression of ERα in the key tissues of the hypothalamic-pituitary-ovarian (HPO) axis. As shown in Fig. 2A, ERα expression was detected near the third ventricle of the hypothalamus in ERαf/f mice, and this expression remained intact in ERαd/d mice. Widespread expression of ERα was also observed in the anterior pituitary of ERαf/f mice. However, the pituitary expression of ERα was absent in ERαd/d mice. The ERα expression was evident in OSE of ERαf/f mice and remained intact in ERαd/d OSE (panels e,f). In addition theca cell expression of ERα also remained intact in the ERαd/d ovaries (panel h). Most notably, ERα was present in the tumor cells of ERαd/d ovaries (inset j, Fig. 2A). The Cre-mediated excision of the floxed ERα gene in the anterior pituitary is consistent with earlier reports indicating high levels of progesterone receptor (PR) expression in this tissue. The lack of Cre-mediated excision of the ERα gene in the hypothalamus and OSE, on the other hand, is presumably due to relatively low levels of PR expression in these tissues [18], [19].
Due to selective ablation of pituitary ERα expression, the ERαd/d mice are likely to experience a loss of negative-feedback regulation by E at the level of pituitary. Consistent with this prediction, the serum level of LH was significantly elevated in ERαd/d mice (Table 1). Hyperstimulation of ovarian cells by LH resulted in increased steroidogenesis, leading to high circulating levels of progesterone, testosterone and E in ERαd/d mice (Table 1). In contrast, the level of FSH was not statistically different between ERαd/d and ERαf/f mice. According to previous reports, the levels of LH, progesterone, testosterone, and E are also elevated in ERαKO mice [19], [20]. Consistent with the ERαKO mouse phenotype, ERαd/d mice are infertile. Adult mice fail to ovulate due to chronic high levels of LH. Due to the lack of ERα expression in uterine epithelial and stromal cells, the ERαd/d uteri are unable to receive an implanting embryo. Furthermore, uterine tumors are not found in the ERαd/d mice, presumably because the major uterine cell types do not express ERα. However, in contrast to the ERαKO mice, which lack ERα in all cells, including the ovarian cells, ERα was intact in OSE of ERαd/d mice. This raised the possibility that elevated systemic E levels contribute to tumor initiation by stimulating ER signaling in OSE of ERαd/d mice but fails to do so in OSE of ERαKO mice.
In agreement with this view, we observed marked up-regulation of a transcriptionally active form of ERα, phosphorylated at serine 118, in OSE of ERαd/d mice (Fig. 2B, b). We also examined the status of the phosphoinositide 3-kinase (PI3K)/AKT pathway, which is reported to be activated in response to E treatment of ovarian cancer cell lines [21]–[23]. We noted that the level of AKT phosphorylated at Ser 473 (p-AKT) is elevated in the OSE and tumor cells of ERαd/d ovaries, while p-AKT level is maintained at a low level in ERαf/f ovaries (Fig. 2B, c,d). It is likely that the increased level of phosphorylated AKT is linked to the elevated E signaling in ERαd/d ovaries.
To further characterize the nature of the ovarian tumor in ERαd/d mice, we performed immunohistochemical analyses using epithelial and granulosa cell biomarkers. Anti-mullerian hormone (AMH) is a well-known marker for normal granulosa cells and granulosa cell tumors [24]. While both ERαf/f and ERαd/d ovaries expressed AMH exclusively in the granulosa cells of follicles, ERαd/d ovaries did not express AMH in the tumor cells, indicating that these tumors are not of granulosa cell origin (Fig. 3A). Analysis using anti-cytokeratin 8 (CK8) antibody revealed that ERαf/f mice express this epithelial marker exclusively in a single layer of OSE at 3, 6, and 11 months of age (Fig. 3B, panels a,c,e). In contrast, the OSE of ERαd/d mice at 3 months of age exhibited multiple layers of cytokeratin-positive cells (panel b). At 6 months of age, we observed pronounced cytokeratin 8 expression within the ovaries of ERαd/d mice, indicating the presence of epithelial cells within the tumor mass (panel d). By 11 months of age, widespread cytokeratin 8 immunostaining was observed within the ovarian tumor, highlighting its remarkable epithelial component (Fig. 3B, panel f).
Current literature suggests that the human ovarian epithelial tumors are derived from either OSE or FTE [2], [25]. Although these epithelia are derived from a common embryologic precursor, OSE is thought to retain mesothelial characteristics, while FTE is terminally differentiated [25]–[27]. Recent studies on the serous subtype of ovarian cancer have suggested that either OSE differentiates to resemble FTE or the cancer originates in the fallopian tube and spreads to the ovary [27]. To investigate further the origin of epithelial ovarian tumor cells in ERαd/d mice, we removed the oviducts of these mice prior to tumor formation. Interestingly, removal of the oviducts from pre-pubertal ERαd/d mice did not prevent the onset of ovarian tumor growth in these animals, indicating that the tumor cells originate from the OSE rather than the oviductal epithelium (Fig. 4A). We also examined the epithelia of ERαf/f and ERαd/d ovaries by monitoring the expression of biomarkers specific for either OSE or FTE. As shown in Fig. 4B, we detected prominent expression of calretinin, a mesothelial marker [2], [28], in OSE of ERαf/f ovaries but not in OSE of ERαd/d ovaries. We also noted marked up regulation of tubal-specific makers, including PAX8, WT1, and Ber-EP4 in the ovaries of ERαd/d mice, while the ovaries of ERαf/f mice lacked their expression. Since PAX8, WT1, and Ber-EP4 are normally expressed in FTE and are present in serous epithelial ovarian tumors [2], [28]–[30], it is likely that the OSE cells of ERαd/d ovaries have undergone differentiation to resemble FTE. Furthermore, it has been reported previously that PAX8 is expressed in serous, endometrioid, and mucinous ovarian cancer while expression of WT1 is restricted to the serous subtype of ovarian cancer [29]. Currently there are no available biomarkers that can differentiate between high and low-grade serous ovarian carcinoma. It is clear that the ERαd/d tumors do not grow aggressively.
To investigate the molecular pathways underlying ovarian tumorigenesis in ERαd/d mice, we next performed gene expression profiling, using RNA isolated from the ovaries of ERαf/f and ERαd/d mice. We identified more than 2500 genes that were differentially expressed in the tumor tissue compared to the normal ovaries. The GEO accession number for the microarray data is GSE39402. When we compared the differentially regulated genes to three different datasets of differential gene expression profiles of human serous adenocarcinoma versus control human ovaries that exist in the Oncomine database, we noted that a large number of genes, which are differentially expressed in human serous ovarian cancer specimens, are also present in ERαd/d ovarian tumors (Fig. S2A). Remarkably, the identity of genes expressed in ERαd/d ovaries and human serous ovarian cancer ranged from 25–40%. Prominent among these genes were those encoding platelet derived growth factor receptor alpha (PDGFRα), vascular cell adhesion molecule (VCAM), clusterin, intercellular adhesion molecule 1 (ICAM-1), and serine/threonine phosphatase 1 (Wip1), which are overexpressed in human serous ovarian cancer [31]–[36]. We observed that the levels of PDGFRα, VCAM, ICAM1, and clusterin were markedly elevated in the ovaries of ERαd/d mice compared to those of ERαf/f mice (Fig. S2B). Collectively, the presence of these cancer biomarkers in ERαd/d ovarian tumors underscored the importance of this model in deciphering the pathways involved in genesis and progression of epithelial ovarian tumorigenesis.
Although the elevated systemic levels of E in ERαd/d mice likely contribute to the initiation of ovarian tumors by stimulating ERα signaling in OSE, we considered the possibility that, as the follicles are depleted with tumor progression, intratumoral E biosynthesis becomes a major regulator of tumorigenesis. Studies in postmenopausal women reported significantly increased expression and activity of P450 aromatase in serous ovarian carcinomas, but not in benign adenomas, supporting the view that intratumoral E derived from in situ aromatization could function as an autocrine growth regulator for cancer cells [14], [37]. Previous studies have also reported elevated aromatase activity in tumors and ovarian cancer cell lines [38]–[41]. We observed that ovarian tumors of ERαd/d mice do indeed express high levels of P450 aromatase mRNA (Fig. S3A). To localize aromatase expression we digested ERαd/d ovarian tumors into single-cell suspension, plated both fibroblast stromal and epithelial cells, and completed immunocytochemistry co-localizing both aromatase and a marker indicating the cell type. We observed that ovarian tumor cells isolated from ERαd/d mice express high levels of P450 aromatase protein in luteinized stromal cells of the tumor, suggesting that these cells acquired the ability to synthesize E (Figs. S3B). Furthermore, the activated form of ERα, phosphorylated at Ser-118, is abundantly expressed in the OSC, while aromatase is expressed in ovarian stroma of ERαd/d mice as early as 3 months (Fig. S3C). We postulated that the epithelial ERα signaling remains elevated in response to this locally produced E in ERαd/d ovarian tumors, supporting increased ERα-dependent gene expression, abnormal cell proliferation, and tumorigenesis.
To examine whether E plays a critical role in ovarian tumor progression in ERαd/d mice, we chronically treated these mice at 3 months of age with letrozole, a specific inhibitor of P450 aromatase, by implanting silastic capsules containing this drug. Following three months of letrozole treatment, ovarian tumors of 6-month old ERαd/d mice displayed a remarkable reduction, up to 60%, in tumor volume when compared to sham-treated ERαd/d mice (Fig. 5A). It is important to note that while ERαd/d mice treated with letrozole exhibited significantly lower levels of serum E compared to sham-treated ERαd/d mice, their serum LH levels were not altered in response to this treatment (Fig. 5A). Interestingly, ERαd/d mice treated with the letrozole exhibited significantly lower levels of ovarian expression of PDGFRα and VCAM transcripts compared to sham-treated ERαd/d mice (Fig. 5B). Similarly, the levels of Wip1 mRNA and protein were markedly decreased in ovarian tumors of ERαd/d mice upon letrozole treatment (Fig. 5, B and C). Taken together, these results confirmed that elevated E signaling in the ovarian tumors of ERαd/d mice leads to dysregulated expression of a subset of genes with known links to ovarian epithelial cancer.
Genetically engineered mouse models are considered to be among the most powerful and promising tools presently available for studying the biology of various forms of cancer and for developing therapeutics. Although the creation of mouse models of ovarian cancer has lagged behind models for many other neoplastic diseases, significant advances have been made in the last decade. Orsulic et al have shown that p53-deficient ovarian cells engineered to overexpress multiple oncogenes, c-myc, Kras, and Akt, develop ovarian tumors when injected in mice [42]. Similar mouse models for ovarian epithelial tumors were developed via inactivation of various tumor suppressors, such as Pten, APC, p53 and Rb, through intrabursal administration of adenoviral vectors [43], [44]. Conditional inactivation of multiple genes, such as Pten and Kras or PTEN and Dicer, by expression of Cre recombinase driven by the Amhr2 promoter also led to ovarian cancer in mice [45], [46]. These mouse models have provided compelling evidence that OSE or FTE can be transformed by altering the expression of a variety of oncogenic factors or tumor suppressors. Some of these models display tumor histotypes similar to ovarian cancer subtypes seen in women. However, it is clear that these models typically require multiple genetic changes and are limited by very rapid tumor onset, which limits their usefulness for studying early modulators of ovarian tumorigenesis. In the present study, we report the development of a unique animal model, in which initiation of ovarian tumorigenesis is independent of any oncogenic insult but dependent on elevated E signaling in the ovary. Since the onset and progression of tumorigenesis is relatively slow in ERαd/d mice, this model is potentially useful in providing insights into the factors involved in the initiation and early phases of ovarian epithelial tumorigenesis.
In ERαd/d mice, ERα is conditionally ablated in the pituitary but retained in the hypothalamus and ovary. The loss of negative-feedback regulation by E in the HPO axis led to elevated production of LH by the pituitary. Interestingly, high levels of gonadotropins in women in early postmenopause have been postulated to play a role in the development of epithelial ovarian neoplasms [4], [5]. Consistent with this notion, it has been found that women with polycystic ovary syndrome, which is accompanied by high LH levels, have a greater risk of developing ovarian cancer [47]. Further supporting a role of gonadotropins in ovarian cancer development, gonadotropin levels in cysts and peritoneal fluid from ovarian cancer patients have been shown to be elevated [48]. However, not all studies have led to similar findings and it is clear that elevated gonadotropin levels alone do not cause ovarian cancer. In fact, our studies using the ERαd/d model suggested that hyperstimulation of ovarian cells by LH results in increased steroidogenesis, leading to high levels of circulating E as well as locally produced E in the ovarian tissue. High levels of testosterone coupled with increased expression of aromatase in the ovarian tissue would lead to increased synthesis of local E. We propose that this elevated E is an important factor in epithelial ovarian tumorigenesis as it stimulates signaling in the OSE, promoting its proliferation and phenotypic transformation. These results are supported by epidemiological and clinical studies, which indicate that postmenopausal women with elevated gonadotropin levels and receiving E replacement therapy exhibit an increased incidence of ovarian tumors [9]–[13]. Consistent with a role of E in the genesis of ovarian tumors, recent reports point to the clinical use of anti-estrogen drugs in stabilization of ovarian cancers [49], [50].
Although many previous studies indicated that epithelial ovarian cancer arises from OSE, recent studies have revealed that the fimbriae of the fallopian tube is a possible site of origin of this malignancy, particularly high-grade serous carcinoma [51]. The common embryologic precursor of OSE and FTE is the coelomic epithelium, which gives rise to the epithelial linings of the fallopian tube and the ovary [27]. Unlike FTE, OSE retains mesothelial characteristics and is not terminally differentiated. It has been proposed that either OSE terminally differentiates to resemble FTE, or the cancer originates in fallopian tube and then spreads to the ovary. In support of the latter hypothesis, a recent study showed that conditional deletion of Pten and Dicer, using the Amhr2-Cre, led to tumor development in the fallopian tube, which subsequently metastasized to the ovary [46]. Our studies, on the other hand, appear to indicate that ovarian tumorigenesis in ERαd/d mice is associated with differentiation of OSE to FTE. We observed prominent expression of FTE marker proteins, such as PAX8, WT1, and Ber-EP4, which are not normally expressed in OSE, in the ovaries of ERαd/d mice. Furthermore, removal of oviducts from ERαd/d mice did not prevent the onset of ovarian tumorigenesis, indicating that FTE is not the precursor tissue for tumorigenesis in ERαd/d mice.
Interestingly, we did not observe any intraperitoneal metastatic spread of the ovarian tumor in ERαd/d mice. This could be partly due to the fact that the majority of the mutant mice died by 10 months of age due to the enlarged tumor, making it difficult to follow the progression of tumorigenesis beyond this point. The absence of overt malignancy in our model is not entirely surprising as several recent studies indicate that multiple genetic changes are necessary for metastatic transformation. It is conceivable that additional mutation(s) in tumor suppressor genes, such as p53, is required to drive the tumorigenic pathways in ERαd/d ovaries to rapidly progressing ovarian carcinoma, which will culminate in metastasis. Indeed, recent studies, utilizing genomic sequencing data from human high-grade serous ovarian cancer specimens, have shown that these cancers exhibit genomic instability and harbor genetic mutations in p53, Rb, BRCA1, and/or BRCA2 loci [52]–[55].
The ovarian tumors in ERαd/d mice are composed of cells of both epithelial and stromal origins. These tumors appear to be distinct from the tubular or tubulostromal adenomas, which are reported to occur spontaneously in a number of mutant mouse strains, including the WXWX mice [56], [57]. The adenomas, composed of numerous tube-like structures plus abundant large luteinized stromal cells, arise due to a defect in primordial germ cell proliferation and rapid loss of oocytes at birth, resulting in destruction of graafian follicles [58], [59]. They also arise when mice are irradiated and there is a rapid loss of oocytes after radiation exposure [60]. However, these adenomas are not lethal and administration of E prevents rather than promotes their development [61]. Furthermore, in contrast to these mutant mouse strains, the ERαd/d mice exhibit normal number of oocytes at 3 months of age, which then start to decline when ovarian epithelial and stromal cells expand and form the tumor mass at 6 months. Therefore, the initial stages of tumorigenesis in ERαd/d mice are independent of the oocyte loss. Most importantly, the growth of the ovarian tumors exhibited by the ERαd/d mice is inhibited by letrozole, indicating that these tumors, unlike adenomas, are E-dependent. The ovarian tumors in the ERαd/d mice are presumably dependent on pituitary LH production, which help luteinize the stromal cells. However, the local production of E by these tumors and the resulting estrogenic effects on ovarian surface epithelial expansion and transformation appear to be the two key features that distinguish these tumors from the endocrinologically inactive tubular adenomas or tubulostromal adenomas.
Although the ovarian neoplasm in ERαd/d mice did not show signs of overt malignancy, there was nevertheless clear evidence of tumorigenic transformation. Particularly striking is the finding that a large number of genes, associated with human serous ovarian cancer, are also expressed in ERαd/d ovarian tumors. Specifically, these tumors exhibit dysregulated expression of PDGFRα, VCAM, and Wip1, which were previously reported to be involved in human ovarian cancer. PDGFRα, a cell surface tyrosine kinase receptor for members of the platelet-derived growth factor family, is over-expressed in human serous ovarian tumors and is targeted in clinical trials to treat ovarian cancers [32], [33]. VCAM, a vascular cell adhesion molecule, is found in the blood circulation of cancer patients and has recently been proposed as a marker to detect early stages of ovarian cancer [34], [35]. Wip1, a p53-inducible phosphatase and an oncogene, is of particular interest. Under normal conditions, it restores cellular homeostasis following DNA-damage by cooperating with p53 to induce G2/M cell cycle arrest, thereby allowing ample time for repair of the damaged DNA [62]. However, amplification of Wip1 leads to sustained inhibition of DNA damage response and tumor suppressors, and consequently, its overexpression has been implicated in a variety of human malignancies, including ovarian carcinoma [37], [63]. Recent studies have revealed that Wip1 is regulated by ERα [64]. Consistent with this finding, administration of letrozole to ERαd/d mice, which decreased the ovarian tumor size, also markedly reduced the expression of Wip1 along with PDGFRα, and VCAM. These results are consistent with our hypothesis that accentuated E signaling in the ovarian tissue promotes aberrant expression of genes that participate in tumorigenesis.
In summary, we describe a unique mouse model that allows us to identify hormonal effectors, particularly elevated E signaling, which play an important role in the development of ovarian epithelial tumorigenesis. In the future, the ERαd/d model will serve as a valuable tool for exploring the involvement of E-dependent signaling pathways in the onset and progression of this deadly disease.
Mice (C57BL/6; Jackson Laboratory) were maintained in the designated animal care facility at the University of Illinois College of Veterinary Medicine according to the institutional guidelines for the care and use of laboratory animals. We crossed mice harboring ‘floxed’ ERα gene (Esr1tm1.2Mma), termed ERαf/f, with PR-Cre mice expressing Cre recombinase under the control of progesterone receptor promoter (Pgrtm2(cre)Lyd) to develop mice of genotype Esr1tm1.2Mma/Esr1tm1.2Mma Pgrtm2(cre)Lyd/Pgr+, which we termed ERαd/d. The PR-Cre knock-in mice expression of cre recombinase in pituitary, uterus, oviduct, mammary gland, and corpora lutea of the ovary have been described previously [18]. It has been used extensively to ablate “floxed” genes in tissues expressing PR [17], [18].
Paraffin-embedded ovarian tissue sectioned at 4 µm, mounted on slides and subjected to immunohistochemistry as described previously [65]. Sections were incubated at 4°C with polyclonal antibodies against PCNA (Santa Cruz sc-56), cytokeratin 8 (Developmental Studies Hybridoma Bank, TROMA I), ERα (Novacastra Laboratories), p-Akt1/2/3 serine 473 (Santa Cruz SC-33437), AMH (Santa Cruz Biotechnology SC-6886), WT1 (Santa Cruz Biotechnology), PAX8 (Proteintech group 10336-1-AP), calretinin (Invitrogen 18-0291), Ber-EP4 (Dako), aromatase (Abcam ab35604), vimentin (Sigma Aldrich V5255). Biotinylated secondary antibodies were used followed by incubation with horseradish peroxidase-conjugated streptavidin (Invitrogen). Sections were stained in AEC Solution.
Total RNA was isolated from ovaries by standard Trizol-based protocols and converted to cDNA. The cDNA was amplified by real-time PCR to quantify gene expression using gene-specific primers and SYBR Green (Applied Biosystems). As a loading control, the expression level of RPLP0 (36B4), which encodes a ribosomal protein, was determined. For each treatment, the mean threshold cycle (CT) and standard deviation were calculated from CT values obtained individually from 3 to 4 replicates of that sample. Each sample was subjected to three independent real-time PCR trials. The fold change was derived from the mean CT values. Primer sequences recognizing each gene are located in Table S1.
Hormones were measured by radioimmunoassay (RIA) at the Ligand Core facility, University of Virginia, Charlottesville. Statistical significance was determined on SAS program using the Tukey procedure to control for comparison-wise error rate. Significance cutoff value of p<.05 was determined to be statistically significant.
Ovarian tumors were removed from mice and digested with either 6 g/liter dispase (Invitrogen) and 25 g/liter pancreatin (Sigma Aldrich), or 0.5 g/liter collagenase (Sigma Aldrich) in Hank's balanced salt solution (HBSS). After incubation for 1 h at 37°C, the tubes were vortexed for 10–12 s until the supernatant became turbid with dispersed cells. The contents were then passed through an 80-µm gauze filter (Millipore). Cells were re-suspended in Dulbecco's modified Eagle's F12 medium (DMEM-F12; with 100 unit/liter penicillin, 0.1 g/liter streptomycin, 1.25 mg/liter fungizone) containing 10% heat-inactivated fetal calf serum. Cell culture was continued for 48 h after addition of fresh medium.
Ovarian tumor cells were fixed with 10% formalin solution for 10 m. Cells were treated with 25% Triton X-100 (Sigma Aldrich) in PBS for 10 m and exposed to a blocking serum for 1 h. Cells were treated with primary antibodies and incubated at 4°C and exposed to cy3 or cy5-conjugated secondary antibodies.
Silastic capsules were made by filling silastic laboratory tubing with 0.8 mg of ground Novartis Femara tablets (letrozole) and sealing with medical adhesive silicone type A (Dow Corning). For surgery, mice were first treated with analgesic 1 h prior to surgery and then anesthetized. A small dorsal incision was made just below the neck, and the silastic capsule was inserted underneath the skin. The incision was held together with wound clips until healed. After 3 months of exposure to either empty silastic capsules (sham control) or silastic capsules containing letrozole, mice were euthanized and ovarian tumors were fixed or frozen for analysis.
Statistical analysis was performed by ANOVA or two-tailed student's ttest. Values of P<0.05 were considered significant.
|
10.1371/journal.pcbi.1004821 | Ups and Downs of Poised RNA Polymerase II in B-Cells | Recent genome-wide analyses have uncovered a high accumulation of RNA polymerase II (Pol II) at the 5′ end of genes. This elevated Pol II presence at promoters, referred to here as Poll II poising, is mainly (but not exclusively) attributed to temporal pausing of transcription during early elongation which, in turn, has been proposed to be a regulatory step for processes that need to be activated “on demand”. Yet, the full genome-wide regulatory role of Pol II poising is yet to be delineated. To elucidate the role of Pol II poising in B cell activation, we compared Pol II profiles in resting and activated B cells. We found that while Pol II poised genes generally overlap functionally among different B cell states and correspond to the functional groups previously identified for other cell types, non-poised genes are B cell state specific. Focusing on the changes in transcription activity upon B cell activation, we found that the majority of such changes were from poised to non-poised state. The genes showing this type of transition were functionally enriched in translation, RNA processing and mRNA metabolic process. Interestingly, we also observed a transition from non-poised to poised state. Within this set of genes we identified several Immediate Early Genes (IEG), which were highly expressed in resting B cell and shifted from non-poised to poised state after B cell activation. Thus Pol II poising does not only mark genes for rapid expression in the future, but it is also associated with genes that are silenced after a burst of their expression. Finally, we performed comparative analysis of the presence of G4 motifs in the context of poised versus non-poised but active genes. Interestingly we observed a differential enrichment of these motifs upstream versus downstream of TSS depending on poising status. The enrichment of G4 sequence motifs upstream of TSS of non-poised active genes suggests a potential role of quadruplexes in expression regulation.
| Accumulation of RNA polymerase II (Pol II) in the promoter proximal region has been proposed to be important for rapid gene activation, but the full regulatory role of Pol II dynamics is yet to be delineated. Defining polymerase poising as a significant enrichment of accumulation of Pol II near the promoter, and comparing Pol II profiles in resting and activated B cells, we found that Pol II poised genes generally overlap functionally among different B cell states and correspond to the functional groups identified for poised genes in other cell types. In contrast, non-poised genes are B cell state specific. Interestingly, the transition from poised to non-poised state was not associated exclusively with B cell activation. We found quite a few genes expressed in resting B cells that transitioned from non-poised to poised state after cell activation including several Immediate Early Genes (IEGs)—genes which are known to be activated transiently and rapidly in response to a wide variety of cellular stimuli. In addition, we observed an enrichment of G4 sequence motifs upstream of TSS of non-poised but active genes, including IEG genes. We propose that formation of quadruplexes upstream of TSS can facilitate Pol II engagement by destabilizing the DNA duplex. Taken together, our analysis of resting and activated B cells allowed us to provide novel insight into the dynamics of Pol II poising.
| Transcription of protein-coding genes by RNA polymerase II (Pol II) is a complex, multistep process [1–5]. Potentially each of the transcription steps such as polymerase II recruitment, pre-initiation complex (PIC) assembly, open complex formation, promoter escape, pausing, elongation, and termination provides opportunity for a regulatory action [6]. Until recently it has been assumed that the assembly of the pre-initiation complex and Pol II recruitment are the main regulatory steps [7].
Recent genome-wide chromatin immunoprecipitation (ChIP) studies in human [8–10] and Drosophila melanogaster [11,12] cells have shown a high accumulation of Pol II at the 5′ end of genes. Previously, several genes, including HSP70 and Myc [13–16], have been known to harbor promoter proximal paused polymerase whose release regulates Pol II entrance into the productive elongation step; this mode of regulation was assumed to be rare. In those studies, Pol II pausing has been defined formally as an event in which the forward movement of elongation-competent transcription complexes is temporarily stopped owing to template sequence, regulatory factors, or both [1].
Genome-wide studies demonstrated that in addition to Pol II pausing, in some cell types Poll II accumulates at the promoters due to different reasons, some of which could also be regulatory. In particular, Maxwell et al. demonstrated that during starvation in C.elegans, in addition to pausing occurring at active stress-response genes, an inactive ‘‘docked” Pol II accumulates upstream of inactive growth genes [17]. Kouzine et al. showed that promoter melting is another key regulatory step of gene expression in resting B cells [18] and it also leads to accumulation of Pol II in the promoter region. Thus promoter-proximal accumulation of Pol II surveyed, for example, by a ChIP-seq experiment, is clearly not sufficient to determine the precise transcriptional status of Pol II. Bona fide pausing can be observed by experiments such as permanganate sensitivity assays [3,11,12], scRNA-seq [19] [20], GRO-seq [21], or PRO-seq [22].
To make a distinction between bone fide pausing and accumulation of Pol II in Chip-seq experiments, accumulation of Pol II at promoter, independent of its status, is often referred to as polymerase poising [11,23,24]. Following this practice, here we consider Pol II to be poised if its density at the promoter is significantly higher than in the gene body independent of a specific Pol II status. We caution readers that the term “poised” has also been used in literature in other contexts, specifically to denote Pol II phosphorylated on Ser-5 residues (RNAPIIS5P) [25,26] or to denote genes with promoters that comprise a bivalent chromatin domain containing a histone modification associated with transcriptional activation, histone H3 trimethylated at lysine 4 (H3K4me3), along with another associated with transcriptional repression, H3K27me3 (for a review, see [27]).
Pol II poising allows for pre-recruitment of Pol II ahead of gene expression. In particular, it is now broadly accepted that Pol II poising facilitates a rapid response to stimuli [12,28–33]. In agreement, promoter-proximal Pol II poising was shown to be prevalent at genes involved in response to stimuli, immune response, and development [11,12,34–37]. An important group of genes susceptible to fast induction are Immediate Early Genes (IEG)–genes that are activated within minutes from stimuli. Pol II poising has been described for several mammalian IEGs including JunB, cFos, and cMyc [36,38,39]. Analyzing rat neurons, Saha et al. found that several IEGs, such as Arc (also known as activity-regulated gene 3.1), are poised for near-instantaneous transcription by Pol II poising [40]. Similarly, immediate mediators of the inflammatory response were shown to be poised for gene activation through RNA polymerase II stalling [34]. Pol II poising provides an opportunity not only for fast but also for synchronized response to stimuli [41]. However the exact mechanism regulating such synchronization is yet to be elucidated. In particular, not all poised genes are induced in response to stimuli [42–44] thus such synchronization would have to be conditioned on additional factors. In addition, studies of Pol II poising during Drosophila development indicated that de novo recruitment of poised Pol II does not occur in a tissue-specific manner, necessitating additional tissue-specific regulation [23].
Given the role Pol II poising is proposed to play in the regulation of gene expression, comparative analysis of cells in different stages provides an important tool for understanding this mode of regulation. Recently, such a comparative GRO-Seq based analysis of mouse embryonic stem cells (ESCs) and mouse embryonic fibroblasts (MEFs) [45] has suggested that the transition of Pol II from the poised to the productive elongation stage of transcription is a major regulated step during early differentiation in mouse cells. In Drosophila melanogaster, Gaertner et al. showed that Pol II poised status changes during development consistent with the view that poising prepares genes for future expression [23].
While Pol II poising has been proposed as an important regulator of response to stimuli, no analysis of Pol II poising in one of the most informative settings—mouse resting (RESTB) and activated B cells (ACTB)—has been done. Here, we close this gap by performing a comparative analysis of Pol II poising in these cells. Using a classification of Pol II profiles into three groups, we compared functional enrichment of genes in these classes across cell states. Analysis of resting and activated B cells allowed us to identify and investigate groups of genes whose Pol II profile changes in response to B cell activation and thus to provide novel insight into the role of Pol II poising in gene regulation.
It has been broadly recognized that the distribution of Pol II across gene is not uniform and is often characterized by an overall larger accumulation of Pol II in the promoter region than in the gene body. To examine the Pol II distribution in the analyzed cells, we processed reads from ChIP-seq data of IgG and Pol II as described in Material and Methods section. We considered Pol II to be present at the promoter (Pol II+) if the number of Pol II reads at the promoter was significantly higher (p<0.001, Fisher's exact test) than the number of IgG control reads in the same region. Unless otherwise stated, we focused on the Pol II+ genes. We use Pol II− to denote genes not in the Pol II+ group. Following the well-established practice [11], Poising Index (PI) is defined as the ratio of Pol II density in the promoter to Pol II density in the gene body, as described in Material and Methods section. The distributions of Pol II density in the promoter and gene body regions, and PI values for resting and activated B cells are summarized in Fig 1. The PI distributions differ between two B cell states with the lower PI values in activated B cells (p<1.3e-178, Mann-Whitney U test and Cohen's effect size d = -0.41).
We further classified genes with Pol II+ promoters based on the relative Pol II density in the promoter and gene body regions. Specifically, if Pol II presence in the promoter is significantly greater than in the gene body, we call these genes Pol II poised genes; otherwise we call them Pol II non-poised genes. The classification was done using Fisher's exact test where we assessed a null hypothesis that the Pol II density in the promoter and gene body is equal (see Material and Methods). In the end, we defined three major classes of genes based on Pol II activity across gene:
Fig 2 shows the number of genes in each class across different cell states. Both cell states have a similar number of Pol II+ promoter genes (51–54% overall): 9710 genes in activated and 9290 genes in resting B cells. However, ACTB have higher number of class NP genes (14% overall and 26% among Pol II+ promoter genes).
We next analyzed the functional enrichment of genes in class NP and P in both cell states. Interestingly, considering all genes as a background (Table 1), class P genes are enriched for DNA repair, apoptosis, cell cycle, cellular macromolecule catabolic process, translation, and transcription in both resting and activated B cells. These classes generally correspond to genes that function in response to extracellular or intracellular stimuli [45]. Importantly, the enriched functional categories we identified for B cells largely overlap with categories identified using GRO-seq in mouse embryonic stem cells and mouse embryonic fibroblasts when similarly all genes were taken as the background [45]. This suggests that functional enrichment of poised genes is not only common across different cell stages but also across different cell types. This is consistent with recent results showing that during Drosophila development de novo recruitment of poised Pol II does not occur in a tissue-specific manner [23].
In contrast, class NP genes in ACTB (Table 2) are enriched with regulation of lymphocyte activation—a processes that is specific to lymphocyte cells. Thus while Pol II poised genes correspond to the functional groups previously associated with poised genes in other cell types, non-poised genes are B cell state specific. Additionally, class NP genes in ACTB are enriched in translation and transcription. This is consistent with the previous observation that there is an overall 10 fold transcriptional amplification in ACTB relative to RESTB [18]. Gene Ontology (GO) enrichment analysis of class NP genes in RESTB does not show enrichment in any specific GO processes, however, as discussed in detail later, we found that the top genes in this class includes Immediate Early Genes—genes that are relevant for the cell activation process.
To focus more closely on B cell active genes, we additionally performed enrichment analysis using only Pol II+ genes as the background. Poised RESTB genes remained enriched in translation and metabolic process (Table 3). Non-poised genes in ACTB remained to be enriched in B cell specific processes and both RESTB and ACTB were enriched in immune response (Table 4). The fact that Pol II poised genes functionally overlapped between different cell states suggests that the tendency of certain groups of genes to be poised is largely context independent, while non-poised and transcriptionally active genes include functional groups that are cell state specific.
To gain insight into the role of Pol II poising in B cell activation, we asked whether there are any functionally coherent groups of genes that change Pol II profile class after B cell activation. If poising prepares genes for a rapid and simultaneous expression, we might observe an enrichment of certain functional groups within the genes moving from Pol II profile class P to class NP upon B cell activation. GO enrichment analysis of genes that changed Pol II profile from class P to NP (with the background of all Pol II+ promoter genes in both RESTB or ACTB) showed that these genes are enriched for cellular processes related to cell cycle and transcription such as translation, RNA processing, and mRNA metabolic process. On the other hand, Pol II+ promoter genes non-poised in RESTB (class NP) but poised in ACTB (class P) are not enriched for any specific cellular processes (Fig 3). This observation remains true when Pol II profile change is measured by at least two-fold increase of the PI value rather than by switching profile classes.
We next examined how the change in gene transcriptional activity upon B cell activation relates to the changes in Pol II densities, poising index and expression. The genes that transitioned from poised to non-poised state upon B cell activation where characterized, on average, by increased Pol II density in the gene body and reduced Pol II density at promoters (Fig 4a). In addition, genes with increased density of Pol II in the gene body were characterized by lower poising index and higher mRNA expression (Fig 4b). Given these observations one might expect that there is a correlation between changes in poising index and changes in gene mRNA expression, but there was no such correlation. However, we observed that the genes that transitioned from poised to non-poised state upon B cell activation experienced higher expression increase than the genes that continued to be poised (p<4.2e-7, Mann-Whitney-Wilcoxon test); although the effect size was negligible. Thus we separately examined changes in Pol II density at promoters and gene bodies by asking whether the increase in gene expression upon B cell activation correlates more strongly either with increased Pol II promoter proximal density or with its increased density in gene bodies. We calculated the Spearman and partial Spearman correlations between changes in gene expression and changes in Pol II promoter/body density during B cell activation (Fig 4c). The increase in gene expression upon activation correlated with the increase of Pol II density in both promoter and gene body regions; the correlation was stronger for the increase of Pol II in the gene body even after correcting for the Pol II density change in the promoter. The correlation between the change in gene expression and the change in Pol II promoter density became weak (although statistically significant) after correcting for correlation of mRNA expression change with Pol II change in gene body (see Spearman partial correlation in Fig 4c).
Thus for many genes reduction in poising index value during the transition from RESTB to ACTB can be associated with an increase of Pol II density in the gene body and an increase in gene expression. This is consistent with the enrichment of these genes for cellular processes related to cell cycle and transcription such as translation, RNA processing, and mRNA metabolic process. Indeed activated B cells are much larger and transcriptionally more active than resting B cells [18]. Yet, Fig 4a also indicates that poised to non-poised transition can be a result of reducing of Pol II presence at promoter. Since increased gene expression correlates with increased Pol II promoter density then the increased expression is not the only explanation for reduced poising index. This analysis reveals that while transition from poised to non-poised state is often associated with increased Pol II body density and increased gene expression, change in poising index is not a simple function of expression change.
In resting cells, a majority of genes with Pol II presence at the promoter that are non-poised (class NP) are characterized by low Pol II presence at both the promoter and the gene body. As we showed above, the non-poised genes in resting cells that transition to poised state in activated cells were not enriched in any GO categories. We were interested to see whether such transition can also occurs for genes that are transcribed at significant levels in resting B cells. Sorting them based on Pol II density in the gene body, we found that the top ten of these genes contain known immediate early genes such as Btg1, Fos, Jun, Egr1, Bcl6, and Zfp36. As an illustration, Fig 5 shows the Pol II profile of Jun and Fos in RESTB and ACTB.
We then used measurements of mRNA expression level in resting B cells and in cells activated for 30 minutes, 3, 24, and 72 hours to trace expression dynamics of these genes. We found that their expression increases even further when measured 30 minutes after the B cell activation. Thus despite being highly expressed in resting B cells, these genes have not achieved their full induction level yet. Their expression then gradually drops when measured at 3 and 24 hours after B cell activation and after that remains stable (Fig 6). (We note that our measurements of Pol II in ACTB were done 72h after activation.) Many IEGs have previously been shown to be poised prior to the burst of their expression [39,46,47]. For example, we noted that an immediate early gene, Egr2, is poised and expressed at a very low level in resting B cells and experiences an expression burst 30 minutes after activation. It is possible that the IEGs that we found to be highly expressed in RESTB have been spontaneously activated, which prevented us from observing them in their poised state. Interestingly, our analysis reveals that when the cells are fully activated, the expression of these genes decreases while Pol II accumulates at promoters. Thus after expression burst these IEGs return to poised state.
These results suggest that Pol II poising is not just a regulatory step that prepares genes for sprint. We found that a number of early response genes, including Jun, Fos and Egr1, are highly expressed and non-poised in resting B cells but are poised and have relatively low expression in activated B cells, suggesting that Pol II poising also accompanies attenuation of expression of previously active genes.
Previous studies identified a correlation between the presence of G4 motifs and promoter-proximal poising [48]. Given the observation that for B cell Pol II poising status is cell state dependent, we asked if there is any difference in such promoter proximal sequence motifs between genes with different poising dynamics. G-quadruplexes are non-canonical conformations of DNA molecules. A necessary, but not sufficient, condition for their formation is a particular sequence motif consisting of four runs of G tracks. In this study quadruplex motif occurrences were predicted using QuadParser [49] as described in Material and Methods. Quadruplex motifs are abundant in the mouse genome and are present in the 2kb upstream regions of promoters of up to 50% of the genes. We found that for both RESTB and ACTB cells quadruplexes were enriched downstream of TSS of genes with poised Pol II. Such enrichment was observed before for human cancer NCI-60 cell line and primary T cells [48]. Interestingly we also observed that, as opposed to the poised genes, the non-poised genes in both B cell stages were enriched in these motifs upstream of TSS (Fig 7).
These results suggest that, in the context of Pol II poising, the role of G4 motifs upstream and downstream of TSS is likely to be quite different. Indeed, formation of a quadruplex on template strand can obstruct Pol II movement. In addition, runs of Gs downstream of TSS can be involved in R-loops formation, which was also proposed to facilitate Pol II poising [48], [50,51]. Such R-loops are found to be even more stable if RNA-DNA quadruplexes are involved [52–54]. The enrichment of G4 motifs in downstream region is strongly associated with CpG islands. In contrast the enrichment of G4 motifs in upstream region cannot be explained by the presence of CpG islands (S2 Fig).
Given that the biases with respect to the presence of G4 sequence motifs where similar in both ACTB and RESTB, we asked whether the genes that, similarly to the early response genes, have higher relative expression level shortly after B cell activation as compared to fully activated B cell show any particular bias for the presence of these motifs. We divided the genes into three equal classes: high (HI), medium, and low (LO), depending on the ratio of gene expression in B cells activated for 30 minutes to the expression after 72h of activation. In particular, the genes in the high ratio (HI) group are relatively more actively expressed in the cells shortly after the B cell activation (30m) than in the fully activated cells (72h), as for example the early response genes. Comparing the distribution of G4 motifs in the HI and LO groups we found that the HI group was strongly enriched in G4 motifs upstream of TSS; the enrichment was higher for G4 motifs on the non-template strand than on template strand (Fig 7) suggesting the importance of these motifs for rapid gene expression. Indeed, focusing specifically on early response genes identified above, we found that all of them have G4 motifs upstream of TSS but not downstream.
Overall, our analysis suggests that destabilization of the DNA duplex that accompanies formation of the quadruplexes upstream of TSS can facilitate Pol II engagement and in this way might facilitate rapid progression to elongation for genes that are highly expressed immediately after activation.
Polymerase poising has been defined as a significant enrichment of accumulation of Pol II near the promoter relative to the gene body. This accumulation can be attributed to several mechanisms, including pausing of transcription during early elongation and docking. Such Pol II poising reflects pre-recruitment of Pol II ahead of gene expression. To understand the role of poising for B cell activation, we performed a comparative analysis of Pol II poising in resting and activated B cells. We found that on the genome-wide scale poised genes are consistently enriched for DNA repair, apoptosis, cell cycle, cellular macromolecule catabolic process, translation, and transcription. These functional terms were similar to the previously identified for genes exhibiting Pol II poising in embryonic stem cells and mouse embryonic fibroblasts suggesting that these groups of genes are common across cell types [45]. Also consistently with this view, the genes that change from poised to non-poised as the result of B cell activation have higher relative expression. Focusing on B cell active genes, we additionally performed enrichment analysis using only Pol II + genes as the background. Poised RESTB genes were still enriched in translation and metabolic process. Non-poised genes in ACTB remained to be enriched in lymphocyte activation and both RESTB and ACTB were enriched in immune response—all B cell specific processes. It has been proposed that poising provides a mechanism for synchronized response [1]. Indeed, cellular processes related to cell cycle and transcription such as translation, RNA processing, and mRNA metabolic processing consistently change Pol II profiles from poised to non-poised upon cell activation. Our analysis also supports the contribution of poising to regulation of early response genes. In addition, it provides further evidence to the observation made in the context of Drosophila development that the presence of poised polymerase does not necessarily equate to direct regulation through pause release to productive elongation [55].
Our results indicate that that Pol II poising is not just a regulatory step that prepares genes for sprint. We found that a number of early response genes, including Jun, Fos and Egr1, are highly expressed and non-poised in resting B cells but are poised and have relatively low expression in activated B cells, suggesting that Pol II poising is also associated with previously active genes. A closer analysis of group of genes allowed for identification of interesting relationships between distributions of G4 motifs in poised and non-poised genes. Genes with poised Pol II in both resting and activated B cell where enriched in these motives downstream of TSS while non-poised genes showed upstream enrichment suggesting that the role of G4 motifs in Pol II dynamics might be context dependent. It has been proposed that the runs of Gs downstream of TSS can facilitate Pol II poising by R-loops formation or obstructing Pol II progression [48]. Our results are consistent with such hypothesis. The enrichment in G4 motifs upstream of TSS for non-poised genes has not been observed before. We propose that formation of the quadruplexes upstream of TSS can facilitate Pol II engagement by destabilizing the DNA duplex. Specifically, they could expedite a recruitment of Pol II to the promoter region and accelerate its progression to elongation. Thus, taken together, our analysis of resting and activated B cells allowed us to provide novel insight into the dynamics of Pol II poising.
CD43 negative B cells were isolated from the spleens of 6 to 8 week old C57BL/6J mice (The Jackson Laboratory) by immunomagnetic depletion (Miltenyi Biotech). Activation of the purified B cells was as follows: RPMI-1640 containing FCS, Pen/Strep, glutamine, nonessential amino acids, sodium pyruvate, 2-β-mercapto-ethanol, and HEPES for: 30m, 3h, 24h or 72h in the presence of 25 μg/ml of LPS (E.Coli 0111:B4; Sigma), 5 ng/ml of IL-4 (BioSource) and purified rat anti-mouse CD180 (BD Pharmingen). The cells were incubated at 37 degrees in 5% CO2. At the appropriate aforementioned time points the activated B cells were spun down at 1500 RPM for 5min and resuspended in 1mL of Trizol (Life Technologies) and processed for RNA isolation. Resting B cells were also spun down at 1500 RPM processed in the same manner. Class switching to IgG1 was verified by FACS analysis at 72h.
We downloaded Pol II binding and IgG control data (ChIP-seq), and mRNA sequencing data (RNA-seq) for resting and activated B cells from the Gene Expression Omnibus under accession number GSE24178 and from the Short Read Archive under accession number SRA072844. For each cell line, all replicates were merged for joint analysis.
The Chip-seq data with read length from 36 bps to 50 bps were aligned to the mouse reference genome mm9 using Bowtie 2 (version 2.1.0) [56] allowing no more than 2 mismatches and no gaps. We disregarded reads that have multiple best match aligned loci on the reference genome.
We used the spliced read aligner TopHat (version 1.31) [57] to map all RNA-seq reads to the mouse genome (mm9). In order to estimate mRNA expression, we used Cufflinks (version 2.1.1) [58] with the complete set of mouse RefSeq transcripts.
We used NCBI’s DAVID software [59] to perform GO analysis. To summarize and remove the similar GO terms, we used REVIGO [60] with the similarity parameter set to medium.
We downloaded RefSeq gene coordinates for mouse mm9 assembly from UCSC Genome Browser. Only protein coding genes were considered in our analysis i.e. we kept only transcripts with tag NM. We also filtered out genes that were shorter than 2kb. For each cluster of overlapping genes, we only kept the gene producing the longest transcript and disregard the rest. After these filtering steps, we obtained 18,211 protein coding transcripts.
For each gene, we defined its promoter region as a genome segment from 100 bp upstream to 500 bp downstream of its transcription start site. The body region of a gene starts from 1kb downstream of its transcription start site to its transcription termination site. For each gene, the promoter density of Pol II was calculated as the number of reads aligned to its promoter region normalized to reads per kilobase per million reads mapped (RPKM). Similarly, the gene body density of Pol II was computed using RPKM values. The poising index was computed as the ratio between promoter density and gene body density [59].
In our analysis, we were interested in genes with significant Pol II binding in their promoters. These Pol II+ promoter genes have a significantly higher number of Pol II reads than the average number of IgG reads in the promoter region. Given that the total number of reads is large and the lengths of the promoter/body regions are small compared to the length of the genome, we assume that the number of reads in a segment in the IgG control experiment follows a Poisson distribution. For each gene, we evaluated the null hypothesis that its Pol II promoter density is equal to average IgG promoter density using a one sample Poisson test, with p-values from multiple tests being adjusted using Benjamini and Hochberg's procedure [61]. If p-value< 0.01, the gene has Pol II+ promoter; otherwise it has Pol II- promoter.
Among Pol II+ promoter genes, we further checked whether they have poised Pol II in their promoters by assessing whether the Pol II promoter density is significantly greater than gene body density. First, we estimated the average IgG promoter and gene body density from the control experiments. Similarly, we calculated the average Pol II promoter and gene body density. For each gene, we formed 2x2 contingency with the rows corresponding to the Pol II and IgG experiments and the columns corresponding to promoter and gene body density. Using Fisher's exact test as in [45], we then assessed the null hypothesis that the Pol II in the promoter and gene body is equal. The p-values were corrected using Benjamini and Hochberg's procedure [61] with the significance threshold of 0.001.
Non-B DNA forming sequences were identified within 3.0 kb region upstream of TSS or within 500bp region downstream of TSS. Regions with propensity to form quadruplex were predicted using QuadParser program [49] with at least 3 G bases in each of four runs of G repeat and gap size between 1 and 7 nucleotides; both strands were searched. Overlapping regions were merged into a single region.
For each gene, we computed the numbers of quadruplex forming regions; and for each gene group as in Fig 7 and S2 Fig, we calculated median and median absolute deviation of these numbers. Statistical differences between groups were computed using Mann-Whitney-Wilcoxon test.
|
10.1371/journal.pntd.0002047 | Molecular Epidemiology and Antibiotic Susceptibility of Livestock Brucella melitensis Isolates from Naryn Oblast, Kyrgyzstan | The incidence of human brucellosis in Kyrgyzstan has been increasing in the last years and was identified as a priority disease needing most urgent control measures in the livestock population. The latest species identification of Brucella isolates in Kyrgyzstan was carried out in the 1960s and investigated the circulation of Brucella abortus, B. melitensis, B. ovis, and B. suis. However, supporting data and documentation of that experience are lacking. Therefore, typing of Brucella spp. and identification of the most important host species are necessary for the understanding of the main transmission routes and to adopt an effective brucellosis control policy in Kyrgyzstan. Overall, 17 B. melitensis strains from aborted fetuses of sheep and cattle isolated in the province of Naryn were studied. All strains were susceptible to trimethoprim-sulfamethoxazole, gentamicin, rifampin, ofloxacin, streptomycin, doxycycline, and ciprofloxacin. Multilocus variable number tandem repeat analysis showed low genetic diversity. Kyrgyz strains seem to be genetically associated with the Eastern Mediterranean group of the Brucella global phylogeny. We identified and confirmed transmission of B. melitensis to cattle and a close genetic relationship between B. melitensis strains isolated from sheep sharing the same pasture.
| Brucellosis is a bacterial disease causing abortion in cattle, sheep, and goats. It is transmissible to humans by direct transmission and the consumption of untreated milk. Brucellosis has become more and more frequent in Kyrgyzstan in the last decades, and its control has been made a priority. Knowing the bacterial strain circulating is important for the understanding of the transmission and the selection of interventions. The latest identification of Brucella in Kyrgyzstan dates from the 1960s. We report the molecular characterization 17 strains identified as Brucella melitensis from Naryn oblast. Strains were mainly isolated from sheep but also from cattle. All strains were susceptible to a series of antibiotics. We hence identified and confirmed transmission of B. melitensis among sheep which is likely the most important host species. We found close genetic relationship between B. melitensis strains isolated from cattle sharing the same pasture with sheep. Our results support the strategy of pursuing a mass vaccination of livestock in Kyrgyzstan. Further research is needed to identify the most important circulating strains in humans.
| Agriculture is a key component of Kyrgyzstan's economy and livestock play a major role in the daily lives of the population. Sixty four percents of the population live in rural areas and rely on agriculture for their livelihoods. Up to 76% of the rural population of the country is classified as poor. [1].
Since independence in 1991, veterinary support ceased then largely and the incidence of diseases transmitted from animals to humans (zoonoses) has increased dramatically in many regions in Kyrgyzstan. Brucellosis, anthrax, rabies and echinococcosis are public health concerns and constitute a serious risk to the human and the livestock health. The incidence of brucellosis has increased steadily and Kyrgyzstan has now one of the highest human brucellosis incidences worldwide (annual incidence: 77.5 new cases per 100,000 people in 2007) [2].
Currently, Kyrgyz communities are concerned about the effective reduction of the brucellosis burden in people and livestock.
The latest species identification of Brucella spp. cultures in Kyrgyzstan was done in the 1960ies. Both B. abortus and B. melitensis were isolated from cattle. B. melitensis infections in cattle were thought to be a spill-over from sheep. Smirnov and Tretyakova noted that abortions in cows after immunization with S19 were most often seen in herds that were infected with Brucella spp. B. melitensis was isolated from vaccinated and non-vaccinated sheep [3], [4]. the authors concluded that B. melitentsis steadily adapted to sheep [5].
At present, the circulating genotypes of Brucella spp. are not known. This is true for virtually all Central Asian regions. Bacteriological confirmation of Brucella spp.-induced abortions is almost absent, because owners do not report suspected abortions to the veterinary services. Here we report recently isolated Brucella spp. strains from sheep and cattle, which were collected in addition to a representative national study on brucellosis sero-prevalence in humans and livestock [6]. and to cost of disease studies in Kyrgyzstan (data not shown). The results contribute to the understanding of the main transmission routes and effectively inform brucellosis control policy in Kyrgyzstan.
The study was performed in the province of Naryn oblast, which has the highest human brucellosis incidence in Kyrgyzstan and most of its population has an income through selling of animals and animal products. First primary isolations of Brucella strains from aborted fetuses were done at the veterinary laboratory of the Naryn province in November 2008. All public and private veterinarians were informed about the ongoing project on brucellosis. Farmers were informed beforehand and asked to report abortions through local village veterinarians; leaflets with information were distributed through veterinarians and announcement was published in the province newspaper. Abortions from sheep and cattle were collected during the lambing seasons of 2009 and 2010. In general, the lambing season starts in January and continues until March and April and thus first abortions can occur in late November/December. Veterinarians brought the collected specimens – aborted sheep and cattle fetuses - dissected on site - to the Naryn laboratory. Stomach content was collected in tubes and liver, spleen, kidney, lung, heart and other tissues were collected in plastic bags. Veterinarians collected accompanying basic information on the animals and farms such as geographic position and keeping of other than affected animals. Two weeks after the abortion, a visit to the affected farm allowed blood sampling of farm animals for serology (data not shown) and an interview with the livestock holders to obtain epidemiological data with a questionnaire. Total number of fetuses collected by the veterinarians was 125 from whole district and positive isolates by the Urease and Oxidase were selected for further study.
Primary cultures were done at the Naryn zonal Center for Veterinary Diagnostic and specimens were frozen. When culture was negative, frozen specimens were re-cultured at the Republican Center for Veterinary Diagnostic in Bishkek.
Stomach content and organs of the aborted fetuses were cultured onto Brucella selective agar (bioMérieux, Switzerland) and onto own produced Brucella selective agar (with agar, horse serum and antibiotics from Oxoid, Switzerland). Strains were cultured on Brucella agar at 37°C with 10% CO2 for 2 days [7].
For the investigation of the sensitivity of the cultures to, phenotypic antibiotic resistance to 7 different drugs was assessed by the standard E-tests (bioMérieux, Switzerland) on Mueller-Hinton blood agar (MHS2, bioMérieux SA, France) and their minimum inhibitory concentrations (MIC) were determined additionally. The following antibiotics were tested: trimethoprim-sulfamethoxazole (SXT) (1.25+23.75 µg), gentamicin (GM) (10 µg), rifampicin (RA) (30 µg), ofloxacin (OFX) (1 µg), streptomycin (S), (15 µg), doxycyclin (D), (30 µg), and ciprofloxacin (CIP). (5 µg), Inducible clindamycin resistance test (“D-zone” test) was also carried out for all isolates. Results were interpreted according to the Clinical and Laboratory Standards Institute (CLSI) guidelines; for the purpose of this study, intermediate results were classified as resistant.
DNA was extracted from one loopful of bacterial cells grown for 48 h on chocolate agar, and single colonies were isolated by using the tissue protocol of the QIAamp DNA minikit (Qiagen, Germany). DNA concentrations were measured by UV spectrophotometry (Shimadzu, Japan). Multiple Loci Variable Number of Tandem Repeat Analysis (16 locus MLVA) typing was performed with the 17 isolates according to the protocol initially proposed by Le Flèche et al. [8]. and modified by Al Dahouk [9]. to include 1 additional locus, bruce19. The protocols are available online on the MLVA-NET for Brucella (http://mlva.u-psud.fr). In brief, the assay comprised the typing of eight mini-satellites of the so-called panel 1 (bruce06, bruce08, bruce11, bruce12, bruce42, bruce43, bruce45, and bruce55), three micro-satellites of the panel 2A (bruce18, bruce19, and bruce21), and five micro-satellites of the panel 2B (bruce04, bruce07, bruce09, bruce16, and bruce30).
The 16 published VNTR loci were PCR-amplified in parallel and the numbers of tandem repeats determined after electrophoresis on agarose gel. DNA extracts of B. melitensis 16MT and vaccine strain Rev1 were used as positive controls. The obtained MLVA patterns of each sample were then matched with an online database (http://minisatellites.u-psud.fr/MLVAnet/querypub1.php) for identification.
A small amount of a colony of each pure culture was transferred to a FlexiMass target well using a disposable loop and overlaid with 1.0 µl alpha-cyano matrix solution (CHCA; 40 mg alpha-cyano in 33% acetonitrile, 33% ethanol, 33% ddH2O and 1% trifluoroacetic acid). The spotted solution was air-dehydrated during 1–2 min at room temperature and analysed with MALDI-TOF MS Axima Confidence spectrometer (Shimadzu-Biotech Corp., Kyoto, Japan). The reference strain Escherichia coli K12 (GM48 genotype) was used as a standard for calibration and as reference measurement for quality control. Mass spectrometry (MS) analyses were performed in positive linear mode in the range of 2,000–20,000 mass-to-charge ratio (m/z) with delayed, positive ion extraction (delay time: 104 ns with a scale factor of 800) and an acceleration voltage of 20 kV. For each sample, 2×50 averaged profile spectra were stored and used for analysis. All spectra were processed by the MALDI MS Launchpad 2.8 software (Shimadzu Biotech) with baseline correction, peak filtering and smoothing. A minimum of 20 laser shots per sample were used to generate each ion spectrum. For each bacterial sample, 50 protein mass fingerprints were averaged and processed. Spectra were analyzed using SARAMIS (Spectral Archive and Microbial Identification System, AnagnosTec GmbH) at default settings. Cladistic analysis were based on the peak patterns of all analyzed strains submitted to single-link clustering analysis using SARAMIS with an error of 0.08% and a m/z range of 2,000 to 20,000 Daltons.
Allelic diversity was calculated using the formula below, where xi is therelative frequency of the ith allele at the locus, n the number of isolates in the sample and (n/(n-1) is a correction for bias in small samples [10]. VNTR data was the basis for the phylogenetic analysis using SAS (Statistical Analysis Systems Inc. Cary, USA) proc cluster using the unweighted pair-group method with arithmetic averages, (UPGMA). For the assessment of the phylogenetic place of the Kyrgyz isolates, strains were selected from the online database by Maquart [11], [12]. (1471-2180-9-145-S1.xls; http://www.biomedcentral.com/1471-2180/9/145). Isolates were selected to reflect the diversity of geographical origin and the different biovars. Phylogenetic trees were drawn using SAS proc tree.
The Ethics Committee of the University and the state of Basel has approved this study without restrictions in the meeting of January 11, 2007 (Reference number 02/07). The project conforms with the ethics requirements on animal testing (Published in Schweiz. Ärztezeitung, 2006, Band 87, S. 832–837) by the Swiss Academy of Medical Sciences and the Swiss Academy of Natural Sciences. Animal owners were asked for consent to test aborted fetuses of their livestock for brucellosis.
Livestock systems and management of herds from which B. melitensis were isolated varied between owners. Livestock owners kept cattle and small ruminants together and practiced seasonal transhumance to high-altitude pastures. They sometimes also kept entrusted animals from several owners and traded actively animals. During the lambing seasons 2009 and 2010 in Naryn, 125 aborted fetuses (112 from sheep and 13 from cattle) were collected in the 4 villages and the city of Naryn (Figure 1). The rate of isolation for sheep was 8.9% and for cattle is 15% but the difference is not statistically significant. Urease and oxidase positive cultures were selected and 17 out of 23 isolates were confirmed B. melitensis by MALDI-ToF MS and MLVA-16 (Figure 2). The dendrogram is based on the MLVA-16 genotyping assay showing the relationship of the 15 sheep and two cattle isolates of Brucella melitensis. For each locus showing variability, the number of tandem repeats is presented. Additional information is provided on the type of sample, the local strain designation, the serial number of the animal owner and the name of the village in Naryn oblast Numbers in brackets indicate repeated isolates from the same animal. Isolates not indicated as primary were frozen prior to cultivation. Of the 17 isolates, 15 were isolated from sheep and two from cattle. All strains were susceptible to the tested antibiotics. The allelic diversity of VNTR (h) was low, with only three loci showing variation in the numbers of repeats. For locus 4 it was 0.6, for locus 16 0.16 and 0.49 for locus 30 (Table 1). All other loci did not show any variation. Eight out of 17 strains grouped into 6 different clusters. However, it has to be noted that more than one isolate was obtained from four animals. Isolates of cluster 2 were found in herds of two different owners in sheep and cattle. With regard to the geographical location, the Kyrgyz isolates are closest to strains from Kazakhstan, Israel and Iraq which are all biovar3 (Figure 3) [11].
B. melitensis isolates from Kyrgyzstan appear to be close to the so-called Eastern Mediterranean group (Figure 3) [11], but a more detailed analysis and more isolates are required to conclusively determine the position of Kyrgyz Brucella in the global phylogeny. All B. melitensis isolates from Naryn Oblast were closely related according to VNTR patterns. Isolates belonging to second cluster from the top (Strain No. 3–6) (Figure 2) were found in the herd of two owners of sheep and cattle, indicating that strains circulated between farms and were transmitted between small ruminants and likely to cattle during communal grazing. These two owners live 45–50 km apart. The owner of the cattle lives in the city of Naryn and his cattle graze on a summer pasture with several other animals suggesting rural/urban spill over through sharing of common pasture. The 8 isolates (sixth cluster from the Top in Figure 2) from sheep stem from Jer-Kochku and Lakhol, two villages 10 km apart. The animals from which they originated use the same pasture for grazing, except for the two strains from Kulanak which is located at more than 80 km from Jer-Kochku and Lakhol. This may indicate a contact relationship between Kulanak, Jer-kochku and Lakhol (Figure 1). Owner 1 had sheep in which three B. melitensis genotypes are present. A better understanding of the contact network of each animal owner could possibly further explain genetic diversity.
Multiple strains were isolated from liver, spleen and heart in three animals (Figure 2). Isolates from different organs of the same animal had always the same VNTR pattern, hinting to a likely monoinfection. The isolation of B. melitensis in sheep and cattle is the first recent confirmation by culture since the 1960ies in Kyrgyzstan. It was expected because brucellosis in cattle was not a problem a decade ago and increasing sero-prevalences and brucellosis abortions in cattle were observed during the past years. It was therefore speculated that cattle may be a spill-over host of B. melitensis from small ruminants. More isolates are needed to further consolidate this finding. If confirmed, this may have policy implications for ongoing pilot mass livestock vaccination campaigns, considering cattle vaccination. We found no antibiotic resistance and therefore the standard regimen used in Kyrgyzstan (i.e., Gentamicin plus Doxicycline) is likely to be adequate for humans. However, human isolates should be tested as well. The use of antibiotics in livestock is clearly not recommended.
This study confirms ongoing transmission of B. melitensis in sheep and likely to cattle in the province of Naryn in Kyrgyzstan. The high genetic homogeneity indicates rather clonal expansion and ongoing transmission, confirming serological observations [6]. The role of cattle in the transmission of B. melitensis transmission should be examined more closely. Further studies on human brucellosis strain characteristics are needed to confirm sheep as the suspected principal source of livestock to human transmission [6]. For this purpose more discriminatory methods than VNTR may be needed. Further collection of isolates from aborted fetuses including information on contact networks are needed to monitor the success of the ongoing mass vaccination campaign and to allow calibrating VNTR dynamics in space and time.
We conclude that B. melitensis is endemic in Naryn oblast and sheep are apparently the main host. B. melitensis is also transmitted to cattle. In the study period we observed no abortions in goats and hence consider them less important for brucellosis transmission in Naryn oblast. Our findings confirm an earlier serological study, which related human brucellosis sero-prevalence to sheep but not to goat and cattle [6].
|
10.1371/journal.pgen.1004157 | Chromatin Targeting Signals, Nucleosome Positioning Mechanism and Non-Coding RNA-Mediated Regulation of the Chromatin Remodeling Complex NoRC | Active and repressed ribosomal RNA (rRNA) genes are characterised by specific epigenetic marks and differentially positioned nucleosomes at their promoters. Repression of the rRNA genes requires a non-coding RNA (pRNA) and the presence of the nucleolar remodeling complex (NoRC). ATP-dependent chromatin remodeling enzymes are essential regulators of DNA-dependent processes, and this regulation occurs via the modulation of DNA accessibility in chromatin. We have studied the targeting of NoRC to the rRNA gene promoter; its mechanism of nucleosome positioning, in which a nucleosome is placed over the transcription initiation site; and the functional role of the pRNA. We demonstrate that NoRC is capable of recognising and binding to the nucleosomal rRNA gene promoter on its own and binds with higher affinity the nucleosomes positioned at non-repressive positions. NoRC recognises the promoter nucleosome within a chromatin array and positions the nucleosomes, as observed in vivo. NoRC uses the release mechanism of positioning, which is characterised by a reduced affinity for the remodeled substrate. The pRNA specifically binds to NoRC and regulates the enzyme by switching off its ATPase activity. Given the known role of pRNA in tethering NoRC to the rDNA, we propose that pRNA is a key factor that links the chromatin modification activity and scaffolding function of NoRC.
| Tumour cells overexpress ribosomal RNA (rRNA), which is required for ribosome assembly and cell growth. rRNA gene repression is mediated by the chromatin remodeling complex (NoRC) and a non-coding RNA that binds to this enzyme. This study addresses the mechanism of nucleosome positioning by NoRC and the functional role of the non-coding RNA, which is termed pRNA because it corresponds to the promoter sequence. NoRC recognises the promoter nucleosome in a chromatin array with high affinity and uses a release mechanism to position the nucleosome over the transcription initiation site. The pRNA binds specifically to NoRC and inhibits its ATPase activity. We suggest that the RNA retains NoRC at the gene promoter after remodeling, linking its chromatin modification and scaffolding activity to inactive rDNA copies.
| Nucleosomes present a major obstacle for the binding of sequence-specific DNA-binding factors, the interaction of positively charged histone tails with DNA and the masking of DNA binding sites that face in towards the histone octamer surface [1], [2]. As a result, all DNA-dependent processes, such as transcription, replication, repair and recombination, are affected by the positioning of nucleosomes on regulatory sites. ATP-dependent chromatin remodeling enzymes, which use energy from ATP hydrolysis to slide, evict or replace histones within nucleosomes, are key modulators of chromatin structure and DNA-dependent processes [3]. Thus, it is of particular importance to reveal their molecular mechanism of nucleosome remodeling, how these enzymes are targeted to their genomic loci and their role in defining nucleosome positions in vivo [4]–[10].
In mammalian cells, there are numerous types of remodeling enzymes that associate with different subunits to form remodeling complexes with distinct biological functions. Due to the high combinatorial complexity, it is estimated that several hundred different chromatin remodeling complexes exist in humans. These remodeling enzymes comprise several groups of ATPases classified into the Snf2, ISWI, Mi-2, Chd1, Ino80, ERCC6, ALC1, CHD7, Swr1, RAD54 and Lsh subfamilies [9], [11]. In addition to their diversity, chromatin remodeling enzymes are highly abundant, with approximately one enzyme for every 10 nucleosomes in yeast and human cells [5], [10].
Remodeling enzymes preferentially localise to specific genomic regions, raising the questions of which signals target the enzymes to these locations and what their functions are at these sites [12], [13]. Recently, the continuous sampling model was suggested for the abundant ISWI type remodeling enzymes. According to this model, the enzyme continuously samples all nucleosomes by transiently binding and dissociating without translocation. Only upon introducing additional signals, such as the direct interaction with sequence-specific DNA-binding factors, histone modifications and altered DNA/nucleosome structures, do these nucleosomes become marked as “to be translocated” by converting them to high-affinity substrates [13]. However, there is still a lack of mechanistic proof for the continuous sampling model.
Active rRNA genes cover the promoter-bound nucleosomes from −157 to −2 (relative to the transcription start site), compatible with the binding of the UBF and TIF-IB/SL1 factors required for transcription initiation [8]. On repressed genes, the nucleosome is shifted 24 nt downstream, occluding the TIF-IB binding site [8], [14]. NoRC (nucleolar remodeling complex), which is an ISWI type remodeling enzyme that consists of two subunits, Tip5 (TTFI interacting protein 5) and the Snf2H ATPase, is required to establish the repressed rRNA genes and initiate heterochromatin formation [15], [16]. NoRC is recruited to the rRNA gene by the Transcription Termination Factor-I (TTF-I), which binds upstream of the gene promoter [17]. Recent studies have revealed that NoRC also interacts with the pRNA (promoter-associated RNA), a 150–200 nt long non-coding RNA that is complementary to the rRNA gene promoter sequences and is required for efficient rRNA gene silencing and subsequent DNA methylation [18], [19].
We addressed whether NoRC affects the architecture of the repressed rRNA gene, its mechanism of nucleosome positioning and how the enzyme is targeted to the promoter nucleosome. We demonstrate that, within arrays of nucleosomes, NoRC is capable of recognising the rRNA gene promoter nucleosome with a higher affinity than that for other nucleosomes and that it specifically repositions the nucleosome to the site that was observed in vivo. We show that the mechanism of positioning corresponds to a release model of nucleosome positioning, in which NoRC has a reduced affinity for the remodeled substrate. We further studied the role of the pRNA-NoRC interaction and observed that this RNA serves as a negative regulator of NoRC activity, indicating that tight regulation of these enzymes reduces the wasteful turnover of ATP when maintained within chromatin.
The remodeling complex NoRC, consisting of the Snf2H and Tip5 subunits, was expressed using the baculovirus system and purified to apparent homogeneity (Figure 1A). The activity of NoRC was tested on mononucleosomal substrates reconstituted on the 601 nucleosome positioning sequence in the centre or at the border of the DNA fragment ([20], [21], Figure 1B and S1). The end-positioned nucleosomes were repositioned by NoRC to the central locations in an ATP-dependent remodeling reaction (Figure 1B, upper panel). In contrast, when the nucleosomes were located at the centre of the DNA fragment, only minor ATP-dependent effects were detected (Figure 1B, lower panel). The initial analysis indicated that the recombinant NoRC complex was active but required a specific nucleosomal substrate for its activity.
One of the features of the nucleosomal substrate is the linker DNA. To test whether linker DNA is required for NoRC function, we analysed the ATPase activity of NoRC in the presence of nucleosomal arrays and mononucleosomes with and without linker DNA (Figure 1C and S1D). Interestingly, mononucleosomes lacking linker DNA stimulated the ATPase activity of NoRC significantly less than the linker-containing mononucleosomes or nucleosomal arrays. This experiment suggests that recognition of the nucleoprotein structures in the core nucleosome by NoRC activates its ATPase activity but that linker DNA is required for full stimulation.
Next, to determine the minimal length of DNA required for NoRC binding, we carried out DNA-binding experiments using a mixture of DNA molecules with different lengths (from 10 to 130 bp in 10 bp increments, Figure 1D). Quantification of DNA:NoRC complexes in a competitive assay revealed that the DNA-binding affinity of NoRC strongly decreases with DNA lengths below 60 bp and that the remodeler does not significantly bind to DNA of 40 bp or shorter.
Initial experiments did not demonstrate that Tip5 or NoRC have any sequence-specific DNA binding activity (data not shown). However, NoRC may recognise DNA with a particular structure. Therefore, initial binding of Tip5 to cruciform DNA was analysed. Cruciform DNA and two linear, double-stranded 40 bp DNA fragments (‘DNA sequence controls’) were prepared as described [22]. Increasing amounts of Tip5 were incubated with either the cruciform DNA or the linear DNA and analysed in an electromobility shift assay (EMSA). No binding of Tip5 to either of the linear DNA fragments was visible under the experimental conditions (Figure 1E, panels 1 and 3). In contrast, the incubation of Tip5 with the cruciform DNA resulted in the formation of protein/cruciform DNA complexes (panel 2). The experiment shows preferential binding of NoRC to structured DNA.
To test whether linker DNA is required for a stable interaction of NoRC with the nucleosomes, EMSAs using reconstituted mononucleosomes containing linker DNA of 0 bp (146 bp template), ∼25 bp (171 bp template), ∼50 bp (247 bp template, centrally positioned nucleosome) and ∼100 bp (247 bp template, end-positioned nucleosome) and increasing amounts of NoRC were performed (Figure 1F). NoRC bound with similar affinity to the DNA molecules ranging in length from 146 bp to 247 bp, forming discrete NoRC:DNA complexes as expected from the previous experiment. However, when this DNA was reconstituted into nucleosomes, NoRC failed to form a stable complex with the nucleosomes containing 0 bp and 25 bp of linker DNA but formed discrete NoRC-nucleosome complexes with nucleosomes bearing 50 or 100 bp of linker DNA (Figure 1F). Thus, NoRC has a higher binding affinity for free DNA than nucleosomal cores, which suggests that linker DNA is required for efficient targeting of NoRC to remodeling sites.
To determine the relative orientation of NoRC when bound to the nucleosome, we performed DNase I footprinting experiments. Nucleosomes were reconstituted on the central position of the radioactively end-labelled 247 bp mouse rDNA promoter fragment, a known target site of NoRC [15]. Free DNA, nucleosomes and NoRC-nucleosome complexes were incubated with DNase I, the reaction was stopped by the addition of EDTA and the reaction products were resolved by EMSA (Figure 2A, B). Free DNA, nucleosomes and the corresponding NoRC-nucleosome complexes were gel-purified and further analysed on sequencing gels. When compared to free DNA, DNase I digestion of the nucleosomal DNA resulted in a characteristic cleavage pattern, revealing sites of protection and a repeated pattern of DNase I-sensitive sites with a distance of approximately 10 bp, indicating a nucleosome positioned in the centre of the rDNA fragment (Figure 2C). Because a natural DNA sequence was used in this study, the nucleosome lacked precise positioning and a mixture of rotationally phased nucleosomes broadened the protected region [23]. To avoid the formation of multimeric complexes or template precipitation, NoRC was incubated with the nucleosomal substrates at concentrations that result in 50–70% complex formation. NoRC significantly protected the borders of the nucleosome and the adjacent linker DNA from DNase I digestion (Figure 2D). Our data suggest that the binding of NoRC to the nucleosome is bilateral, interacting with both exit and entry sites of the nucleosome, and confirms that NoRC binds to the linker DNA.
To examine the ability of NoRC to reposition nucleosomes on its target site, we reconstituted mononucleosomes on a DNA fragment containing the rRNA gene promoter sequence in vitro (position −190 to +90, relative to the transcription start site). Nucleosomes reconstituted on the rDNA promoter region occupied multiple positions on the DNA, as demonstrated by native gel electrophoresis (Figure 3A, lane 1). NoRC dependent remodeling establishes a preferential nucleosome position that is located close to the center of the DNA (Figure 3A). This nucleosome position was characterised by Exonuclease III footprinting, showing that it protected the DNA from positions −120 to +27 (Figure S2). This position correlates well with the nucleosome position of the repressed rRNA genes in vivo [8]. This suggests that NoRC recognises specific DNA sequences or structures on the nucleoprotein complex that allow site-specific positioning.
A common feature of ribosomal gene promoters is that they lack sequence homology but retain structural similarity and contain intrinsically distorted regions [24]. The relative DNA curvature of the mouse rDNA promoter was calculated with the Bolshoy algorithm using the ‘bandit’ program (Figure 3B, [25], [26]). The mouse rRNA gene promoter contains a region of high local DNA curvature ([25]; at about position −110) that is specifically bound by Tip5 (Figure 3C). This result agrees with the results of the previous experiment, which demonstrated the preferential binding of Tip5 to cruciform DNA (Figure 1E). Thus, these data indicate the specific recognition of structured DNA by the remodeling enzyme, suggesting a potential mechanism for targeting NoRC to the rRNA gene promoter.
Two kinetic models were proposed to explain how chromatin remodelers are able to direct the nucleosome to a specific position on DNA [9]. The release model implies that remodelers bind with high affinity to nucleosomes positioned at the “wrong” sites and remodel the nucleosome until it reaches the final (correct) position. The nucleosome at the final position exhibits the lowest affinity for the remodeling enzyme and is thus the worst substrate for the remodeling enzyme. In contrast, the arrest model postulates that the nucleosome exhibits a much higher affinity for the remodeling enzyme at the final position, locking it on the nucleosome and reducing the catalytic conversion rate [9], [27]. To assign one of the kinetic models for a particular remodeler, the binding and remodeling of nucleosomes must be compared. Thus, we compared the differential binding affinities of NoRC to the individual nucleosome positions by EMSA. The incubation of rDNA −190/+90 reconstituted into nucleosomes with increasing concentrations of NoRC resulted in a stepwise binding of the different nucleosome species (Figure 3D). Free DNA and most of the nucleosomes were bound with similar affinities and retarded in the gel. However, the nucleosome occupying the −120/+27 position bound with the lowest affinity. This nucleosome position is the final position of the NoRC-dependent remodeling reaction (Figure 3A), revealing that NoRC has the lowest binding affinity for the “remodeled” nucleosome, therefore suggesting that NoRC remodels nucleosomes according to the release model.
Differential local binding affinities are required to position nucleosomes on DNA. However, on a more global scale, differential binding affinities could also serve to target the remodeling enzymes to specific genes and regulatory regions. To test how NoRC and Snf2H select their remodeling targets, we used competitive binding and remodeling assays. Nucleosomes were reconstituted on a fluorescently labelled rRNA gene promoter fragment (Cy5 labelled) and the 601 nucleosome positioning sequence ([20], Cy3 labelled). Nucleosomes were mixed and binding or remodeling reactions were performed with increasing amounts of remodelers. Snf2H bound with similar affinity to both nucleosome substrates, and remodeled them with similar efficiency (Figure 3E, F). In contrast, NoRC showed preferential binding to the nucleosomes reconstituted on the rRNA gene promoter, preferentially binding the DNA and nucleosomes at lower NoRC concentrations when compared to the 601 substrate (Figure 3E, lanes 8 to 12). Binding with higher affinity was mirrored in the remodeling assay where NoRC was remodeling the rRNA gene promoter nucleosomes prior to the 601 nucleosomes (Figure 3F and Figure S3).
As cellular nucleosomes are arranged in arrays, we tested whether NoRC is also capable of selectively recognising and repositioning the rRNA gene promoter nucleosome within nucleosomal arrays. Chromatin was reconstituted using the salt dialysis method on a circular DNA containing the rRNA gene promoter and incubated with NoRC or ACF in the presence of ATP. A partial MNase digestion of the nucleosomal DNA was performed and analysed in a primer extension reaction (Figure 3G). ACF did not qualitatively change the distribution of the nucleosomes within the analysed region of the rRNA gene promoter. However, NoRC induced a specific relocalisation of the promoter nucleosomes, placing the 3′ end of the nucleosome at position +22. NoRC-dependent nucleosome positioning at +22 perfectly corresponds to the cellular nucleosomal configuration of the repressed rRNA gene [8]. The 5 bp discrepancy between the mononucleosome remodeling and array remodeling assay could arise from internucleosomal interactions that influence the remodeling outcome. Our data strongly support the hypothesis that nucleosome remodeling complexes determine nucleosome positioning in vivo, thereby directly affecting gene expression.
Previous studies have revealed a specific interaction between TTF-I and NoRC, suggesting that TTF-I recruits NoRC to the rRNA gene promoter [15], [17]. The results described here reveal an additional targeting signal, encoded by the high affinity of NoRC for nucleosomes positioned at “wrong” sites of the rDNA promoter. TTF-I improves the efficiency of NoRC recruitment to the rRNA gene promoter without affecting the outcome of the NoRC-dependent nucleosome remodeling reaction (Figure S4).
Recent studies have demonstrated that NoRC binds to a non-coding RNA, which is initiated upstream of the rRNA gene promoter and contains promoter sequences in the sense orientation. It was suggested that promoter RNA (pRNA) is required to tether NoRC to inactive rRNA genes, where it establishes repressive epigenetic marks [18], [19], [28]. We studied two pRNA constructs that exhibit strong and weak binding affinities for Tip5, pRNA−143/−39 and pRNA−113/−39, respectively [19]. pRNAs were generated by in vitro transcription, re-natured and added to the remodeling reactions (Figure 4). First, the presence of the pRNAs did not influence the nucleosome positioning behaviour of NoRC. Second, we observed specific inhibition of the NoRC-dependent remodeling reaction with increasing levels of pRNA−143/−39 (Figure 4A). In contrast, Snf2H was similarly inhibited by both pRNAs, suggesting that the Tip5 subunit determines RNA-binding specificity and activity. Moreover, NoRC recognises the secondary structure of the pRNA, as inhibition of its nucleosome-remodeling activity was lost when the stem-loop structure was mutated (Figure 4B). We identified a regulatory role of the pRNA, demonstrating that the non-coding RNA serves as an inhibitor of the remodeling enzyme.
To gain more insight into the inhibitory mechanism of pRNA, we investigated the effect of pRNA on NoRC ATPase activity. The incubation of NoRC with an increasing amount of DNA or pRNA only modestly stimulated the NoRC ATPase activity (Figure S5), whereas the presence of nucleosomes considerably accelerated ATP/ADP exchange. The incubation of NoRC with nucleosomes and increasing amounts of pRNA−143/−39 or pRNA−113/−39 resulted in a RNA concentration-dependent inhibition of the ATPase activity (Figure 5A). As in the remodeling reaction, pRNA−143/−39 inhibited the NoRC-dependent ATPase activity more efficiently than pRNA−113/−39, confirming the higher binding affinity of the remodeling complex for this RNA and explaining the inhibition of the nucleosome remodeling reaction.
To reveal the mode of RNA-dependent inhibition, we studied the binding of NoRC to nucleosomes in the presence of RNA (Figure 5B). A competitive EMSA revealed that the pRNA competes with nucleosomes for NoRC binding, indicating that only exclusive NoRC:pRNA or NoRC:nucleosomes complexes exist. Again, competition of nucleosomes from the NoRC:nucleosome complex required less pRNA−143/−39 than pRNA−113/−39, indicating the higher binding affinity of pRNA−143/−39 for NoRC (Figure 5B). Both RNA species competed similarly with Snf2H, pointing to the specific role of Tip5 in NoRC (lanes 13 to 24). In summary, our data demonstrate that pRNA competes with nucleosomes for NoRC binding and therefore directly interferes with its ATPase activity and the nucleosome remodeling reaction.
NoRC is an ISWI type remodeling enzyme that requires linker DNA for nucleosome binding and efficient activation of its ATPase activity and remodeling. The complex recognises structured and non-structured DNA with a minimal length of 30 bp, and the same length of linker DNA is required for stable interactions with nucleosomes. Our data suggest that the most stable interactions are formed with the linker DNA rather than the nucleosome core, as we were not able to detect interactions between NoRC and the nucleosome core in electromobility shift assays. Reduced binding affinities to the nucleosome core potentially explain the reduced ATPase activity observed with NoRC using nucleosome cores. However, binding to the linker DNA and the orientation of the complex with respect to the nucleosome core are not random, as specific interactions with the DNA entry/exit sites of the nucleosome were visible in DNase I footprinting experiments. NoRC was specifically aligned adjacent to the nucleosome, giving rise to symmetrical DNase I protected and enhanced cleavage sites, a pattern reminiscent of ACF binding to nucleosomes [23], [29].
Ribosomal genes present an ideal model system for studying the dynamics and mechanism of chromatin remodeling, as the epigenetic marks, the chromatin structure of the active and repressed genes and the factors involved are well characterised. Active rRNA genes contain a nucleosome covering the gene promoter from positions −157 to −2, allowing the binding of UBF and TIF-IB/SL1 to their recognition sites at the nucleosomal borders [8]. In contrast, repressed genes have a nucleosome covering the positions from −132 to +22 relative to the transcription start site, masking the binding site of TIF-IB. The repression of rRNA genes is intimately linked with the recruitment of NoRC, which induces nucleosome remodeling, gene repression and the acquisition of heterochromatic marks [16]. We show that the activity of NoRC is sufficient for recognition of the promoter structure and nucleosome positioning in vivo. Nucleosomal arrays are required to establish the cellular nucleosome positioning pattern, suggesting that internucleosomal interactions influence the activity of remodeling enzymes. Our results are in good agreement with data demonstrating that ISWI machines are molecular rulers and potentially act in the context of di-nucleosomes [30], [31]. Although NoRC does not serve as a sequence-independent spacing factor, it is capable of recognising sequence features of the rRNA gene promoter, which serve as positioning signals.
Several studies have demonstrated the importance of positioned nucleosomes in the genome [32]. However, irrespective of the ability of many sequences to position nucleosomes in vitro they fail to do so in vivo [33], [34], suggesting that there are additional mechanisms that structure chromatin. We show that NoRC positions nucleosomes according to the release mechanism [9], [13]. The enzyme binds with high affinity to nucleosomes positioned at “wrong” sites, which is the recruitment signal. The remodeling reaction is highly processive, with ACF moving a nucleosome for approximately 200 bp without leaving the nucleosomal substrate [35]. After initiation of the remodeling reaction, the endpoint of the translocation reaction is determined by a reduced affinity of the remodeler for the nucleosome at this site. As any remodeler with distinct binding affinities to nucleosomes at different but close positions on DNA could position nucleosomes, we suggest that chromatin remodeling enzymes serve to organise chromatin structure with respect to the underlying DNA sequence. The concentration and composition of the remodeling enzymes in combination with the specific targeting of those complexes to chromatin would determine a specific chromatin architecture and specify accessible regulatory sequences that determine the activity of DNA-dependent processes. We suggest that the combinatorial aspect of remodeling enzymes and complex constitution may determine cell types and their responses to the environment.
There are a multitude of signals targeting remodeling enzymes to specific genomic regions, including direct recruitment by proteins, protein modifications, histone variants, coding and non-coding RNAs, as well as nucleosomes at “wrong” positions [13]. The continuous sampling model for chromatin remodeling enzymes suggests that high concentrations of remodeling enzymes and low binding affinities towards the non-signalling nucleosomes allow for efficient screening of the genome for signals that attract remodeling enzymes [10]. Here, we provide evidence for the continuous sampling mechanism of NoRC, where the remodeling enzyme selectively remodels the promoter nucleosome within an array of nucleosomes. Differential binding affinities guide the remodeling enzyme to these sites of action. However, on the genomic scale additional targeting signals help to further increase the local concentration of the remodeling enzymes at their sites of action. In the case of NoRC, interaction with TTF-I directly recruits NoRC and thereby improves the efficiency of the remodeling reaction, but does not influence the remodeling outcome [14], [17].
Previous studies have shown that the TAM domain in the Tip5 subunit interacts with pRNA and that this interaction is a prerequisite for maintaining NoRC in the nucleolus [18]. We show that pRNA competes with nucleosomes for NoRC binding, specifically inhibiting its ATPase activity. Therefore, we suggest that a ternary complex consisting of NoRC, nucleosomes and RNA does not exist, despite the fact that NoRC contains several DNA/nucleosome-binding domains and an RNA-binding TAM domain [36].
We suggest, that the pRNA serves three functions (Figure 5C). First, after replacing TTF-I at the rRNA gene promoter, it serves to maintain NoRC localisation at the promoter. Due to the release mechanism of nucleosome positioning, NoRC has a low affinity for the remodeled chromatin structure and most likely would dissociate from the promoter. Given, that Grummt and colleagues have shown that the 5′-end of the pRNA forms a triplex with the T0 site at the promoter region and that the 3′-end interacts with Tip5, we propose a tethering function for the pRNA. Switching off the ATPase activity of NoRC ensures that the nucleosome is stably maintained in the OFF position and that the enzyme does not waste ATP. The pRNA and NoRC recruit DNA methyltransferases, histone deacetylases and histone methyltransferases to silence the rRNA genes [37]–[39] and recruit the silenced genes to the heterochromatin environment of the nuclear matrix [36].
The proteins were expressed in SF21 cells. N-terminally His tagged Snf2H with or without Tip5 was purified via Ni-NTA (Qiagen) chromatography. Flag tagged Snf2H and Acf1 were purified using M2 beads (Sigma) [14].
Murine rRNA gene promoter fragments of 146 bp (−231 to −86; positions relative to the transcription start site), 171 bp (−231 to −61), 247 bp (−231 to +16) and 280 bp (−190 to +90) were amplified by PCR from a plasmid containing the genomic DNA isolated from the NIH3T3 cell line (genbank access #KC202874.1). To radioactively label the DNA fragments, α-32P dCTP was added to the PCR reaction mix. The 601 DNA and the pRNA were prepared by PCR as described [19], [40]. PCR products were used for nucleosome assembly reactions as described [23].
Nucleosomes were assembled according to Rhodes and Laskey using the salt gradient dialysis technique [41]. A typical assembly reaction (50 µl) contained 5 µg DNA, varying amounts of histone octamer, 200 ng BSA/ml, and 250 ng competitor DNA in high salt buffer (10 mM Tris, pH 7.6, 2 M NaCl, 1 mM EDTA, 0.05% NP-40, 2 mM β-mercaptoethanol). The salt was continuously reduced to 200 mM NaCl during 16–20 h. The quality of the assembly reaction was analysed on a 5% PAA gel in 0.4× TBE followed by ethidium bromide staining. Nucleosomes reconstituted on the 247 bp rDNA promoter fragment display two distinct positions that can be separated by native gel electrophoresis [21].
Nucleosome mobility was assayed as described [42]. Briefly, reactions contained 4 nM Cy5 labelled DNA reconstituted into nucleosomes, 1 mM ATP, 100 ng/µl BSA, 1 mM DTT, 70 mM imidazole in Ex80 buffer (20 mM Tris pH 7.6, 80 mM KCl, 1.5 mM MgCl2, 0.5 mM EGTA, 1 mM β-mercaptoethanol, 10% glycerol, 200 ng/µl BSA) and recombinant remodeling enzymes. Nucleosomes were incubated with NoRC for 45 min at 30°C. The reactions were stopped by the addition of 1700 ng CMV14 plasmid DNA and incubated for 15 min on ice. The nucleosome positions were analysed by electrophoresis on 5% PAA gels in 0.4× TBE and fluorescence scanning.
Tip5 binding to cruciform DNA was performed as described [22]. NoRC binding to the DNA and nucleosomes was studied by electromobility shift assays (EMSA). The substrates used in the assay were either radioactively or fluorescently labelled as indicated in the legends. Reactions were performed in Ex80 buffer and the indicated amounts of NoRC. Reactions were incubated for 45 min at 30°C and then analysed by native PAGE. Competitive titration experiments were performed using identical reaction conditions, containing 25 nM NoRC, 4 nM fluorescently labelled mononucleosomes and the indicated amounts of the indicated pRNA constructs. The reactions were analysed on 5% polyacrylamide gels in 0.4× TBE and subsequent fluorescence scanning.
NoRC/nucleosome and nucleosome DNase I footprinting experiments were performed as described [29]. Essentially, radioactively end-labelled DNA was reconstituted into nucleosomes and incubated with NoRC using the same experimental conditions as in the remodeling reactions. DNase I digestions were stopped by the addition of EDTA to a final concentration of 5 mM. The complexes were resolved on native PAA gels and the DNA, nucleosome and NoRC/nucleosome complexes were excised from the gel. DNA was purified and analysed on 7% sequencing gels. Mapping nucleosomal boundaries on nucleosomal arrays before, or after remodelling with NoRC or ACF was performed as described [14].
An ATPase reaction contained 150 ng of DNA or chromatin in 10 µl of Ex75 buffer, 10 µM ATP and γ32P-ATP (0.1 µl; 3000Ci/mmol, Hartmann Analytic), the indicated amounts of pRNA−143/−39 or pRNA−113/−39 and 10 units RNasin. The reactions were initiated by the addition of the remodeling enzyme and incubated for 60 min at 30°C. Aliquots of 1 µl were spotted on thin layer cellulose chromatography plates (Merck) and air-dried. The hydrolyzed phosphate was separated from unreacted ATP by thin layer chromatography in 0.5 M LiCl/acetic acid buffer. The plates were dried at 65°C for 5 min and exposed to Phospho Imager plates (FujiFilm BAS-1500). ATP and hydrolyzed phosphate spots were quantified using the Multigauge software (Fuji). The percentage of hydrolyzed ATP was calculated according to the following equation: Pi/(ATP+Pi)×100%, where Pi: amount of hydrolyzed radioactive phosphate; ATP: amount of left γ32P-ATP.
Nucleosome positioning on the Cy5 5′ end-labelled mouse rDNA fragment (from positions −190 to +90 relative to the transcription start site) was determined with Exo III mapping. Reactions were carried out in an initial volume of 50 µl with 30 nM nucleosomes and 2 U/µl of Exo III (NEB) in 10 mM Tris, 90 mM KCl,1 mM MgCl2, and 1 mM DTT at 16°C. At different time points 7 µl of the reaction mix were removed and the reaction was stopped by the addition of EDTA (final concentration of 50 mM). Proteins were digested with Proteinase K after the addition of SDS to a final concentration of 1% and the DNA was subsequently purified by ethanol precipitation. DNA samples were analysed on 6% sequencing gels. The DNA ladder was prepared with the DNA Cycle Sequencing Kit (Jena Bioscience) using a Cy5 labelled oligonucleotide and the mouse rDNA promoter fragment (−190 to +90), with either ddTTP or ddCTP in the reaction mix. Results were imaged with a FLA-5000 imager (Fujifilm). As control, we carried out Exo III digestions with naked DNA in order to discriminate nucleosome positions from exonuclease pause sites on free DNA. To map NoRC dependent positions a remodeling reaction was performed prior to Exo III analysis. Remodeling was performed with 7.4 ng/µl of NoRC and Cy5 labelled nucleosomes in the presence or absence of 1 mM ATP for 60 min at 30°C. The reaction was stopped with competitor plasmid DNA and used for native gel analysis and Exo III footprinting.
|
10.1371/journal.pgen.1004115 | How a Retrotransposon Exploits the Plant's Heat Stress Response for Its Activation | Retrotransposons are major components of plant and animal genomes. They amplify by reverse transcription and reintegration into the host genome but their activity is usually epigenetically silenced. In plants, genomic copies of retrotransposons are typically associated with repressive chromatin modifications installed and maintained by RNA-directed DNA methylation. To escape this tight control, retrotransposons employ various strategies to avoid epigenetic silencing. Here we describe the mechanism developed by ONSEN, an LTR-copia type retrotransposon in Arabidopsis thaliana. ONSEN has acquired a heat-responsive element recognized by plant-derived heat stress defense factors, resulting in transcription and production of full length extrachromosomal DNA under elevated temperatures. Further, the ONSEN promoter is free of CG and CHG sites, and the reduction of DNA methylation at the CHH sites is not sufficient to activate the element. Since dividing cells have a more pronounced heat response, the extrachromosomal ONSEN DNA, capable of reintegrating into the genome, accumulates preferentially in the meristematic tissue of the shoot. The recruitment of a major plant heat shock transcription factor in periods of heat stress exploits the plant's heat stress response to achieve the transposon's activation, making it impossible for the host to respond appropriately to stress without losing control over the invader.
| Transposons are programmed to amplify within their host genomes. In defense, hosts have evolved mechanisms to impede transposon activation, often by epigenetic transcriptional silencing. A constant and likely unending arms race between host and invader has brought about different strategies to mutually counteract the tricks of the other. Several such strategies are combined in one transposon in the Arabidopsis genome. Its promoter is devoid of symmetric sites necessary for stable maintenance of repressive DNA methylation, and a reduction of methylation at the remaining cytosines does not activate the element. More sophisticated still: its promoter shares a sequence motif with heat stress-responsive plant genes and is recognized by a heat-induced plant transcription factor. Whenever the plants must activate their heat stress defense under high temperatures, the transposon is able to generate new extrachromosomal DNA copies that can potentially integrate into new sites of the genome. In addition, the heat response is especially strong in tissue with dividing cells, which form consequently the largest amount of extrachromosomal transposon copies. We see this as an example of a “wolf in sheep's clothing” strategy, whereby the transposon becomes visible as such only under specific stress conditions of its host.
| Transposable elements (TEs) and their host organisms depend on each other for better or for worse. New TE insertions can give rise to deleterious mutations [1] or overall genetic instability [2], but they can also make a positive contribution to gene regulation and adaptation [3], [4]. Host organisms have developed mechanisms to reach a balance between both consequences by suppressing TE activity. Plants have evolved a complex regulatory network of epigenetic silencing that is effective for numerous different TEs. Silent elements are typically associated with high levels of DNA methylation at cytosines in every sequence context (mCG, mCHG, mCHH, where H stands for A, T or C), with methylation at lysine 9 of histone H3 (H3K9me2), and with the presence of 24 nt small interfering RNAs (siRNAs) that guide the RNA-directed DNA methylation (RdDM) machinery in a reinforcing loop, reviewed in [5], [6], [7]. Disruption of DNA methylation patterns can activate transposons, as, for instance, a null mutation of the Arabidopsis maintenance METHYLTRANSFERASE 1 (MET1) activates EVADÉ (EVD), a retrotransposon of the ATCOPIA93 family that can amplify during sexual propagation of the mutant [8]. EVADÉ is also transcriptionally active in plants with hypomethylated DNA due to a lack of the chromatin-remodeling factor DECREASE IN DNA METHYLATION 1 (DDM1). DDM1 mutants and many other plants lacking components of the RdDM pathway activate a specific but partially overlapping subset of TEs, including members of ATCOPIA13, ATCOPIA21, ATGP3 retrotransposon families and VANDAL21 and CACTA DNA transposons [9], [10], [11].
In addition to genetic interference, TEs can be activated by stress. In fact, this was already postulated by the discoverer of TEs [12] who also recognized the important role of these elements for gene regulation. Stress-induced transposon activation was later documented by molecular data in many different hosts, for instance the activation of the Tnt1 element by pathogens in tobacco [13], [14], of the ZmMI1 in maize and the PAL/Tam3 in snapdragon by cold [15], [16], [17], or of the CLCoi1 by wounding or salt stress in lemon [18]. More recently, a Ty1/copia-type long terminal repeat (LTR) retrotransposon family (ATCOPIA78) named ONSEN was found activated by heat stress in Arabidopsis [19], [20]. Surprisingly, without heat stress, ONSEN was not expressed in ddm1 mutants [19] or other mutants lacking RdDM components, in contrast to most other known and potentially functional TEs in Arabidopsis. However, new ONSEN insertions were found in the progeny of heat-stressed plants deficient in small RNA production, and this retrotransposition appears to occur during flower development and before gametogenesis [21]. Higher activation in callus compared to vegetative tissue indicates a possible coupling to active cell cycling [22]. Transcription of ONSEN-related sequences after heat exposure was found in most species of the Brassicaceae [23], indicating a conserved mechanism of activation and control of their spreading.
Here, we provide insight into the initiation of ONSEN activation in Arabidopsis, quantifying transcription, and the formation of extrachromosomal DNA upon extended heat stress. We also determine which of the genomic copies become transcriptionally active. Although the LTR sequences representing the ONSEN promoter are methylated, reduction of the methylation is not sufficient to activate the element. Rather, the LTR has acquired a sequence that is recognized by the plant's heat-responsive transcription factors, thereby coupling ONSEN activation to an important stress defense. Even more cunningly, this response is most pronounced in regions comprising the meristematic zone, providing an enhanced chance for new ONSEN copies to enter the next generation.
A. thaliana grows mainly in zones with moderate climates. A temperature of 37°C represents acute and drastic heat stress for this plant, as shown by the quick transcriptional activation of many heat shock factors, [reviewed in 24]. The heat-induced transcriptional activation of retrotransposon ONSEN was also observed above a threshold of 37°C [19], [20], [22], though only after an extended exposure to heat [19], [20]. To determine the kinetics of ONSEN activation in more detail, we monitored the ONSEN expression at several time points during 30 h of heat stress (HS) treatment, keeping the regular long-day light regime (Fig. 1A). This treatment causes substantial growth arrest but is sublethal and allows recovery of the plants upon subsequent transfer to ambient temperature [19]. ONSEN RNA was first quantified by qRT-PCR 6 hours after the onset of heat stress, and its amount continued to increase to the highest level after 24 h, remaining high until 30 h, the end point of the stress treatment (Fig. 1B).
Following transcriptional activation, retrotransposons form linear extrachromosomal DNA copies along with circular by-products of replication; these are capable of reintegrating in the genome [25], [26], [27]. To determine the presence and amount of extrachromosomal ONSEN DNA, we performed Southern blot analysis with non-digested DNA from samples collected at the same time points as for the RNA. The single, high molecular weight band hybridizing to the ONSEN probe in the non-heat stressed samples corresponds to ONSEN copies integrated in the genomic DNA (gDNA). An additional band present in later heat stress samples indeed indicated extrachromosomal ONSEN DNA (exDNA), with a size of 5 kb corresponding to the expected full length of the linear element that is capable of reintegrating in the genome (Fig. 1C). To quantify the relative amount of exDNA, we calculated the relative intensity of hybridization signals between exDNA and gDNA in each sample, postulating a fixed number of integrated ONSEN copies. Small amounts of linear exDNA of ONSEN first appeared after 12 hours of heat stress. The maximum was achieved after 30 hours HS, corresponding to approximately five times more than integrated in the genome (Fig. 1D). Therefore, the formation of the linear extrachromosomal ONSEN DNA follows the transcriptional activation during heat stress after approximately 6 hours, likely reflecting the need for a threshold level of RNA and time to complete the reverse transcription. Once started, the process can produce many more additional copies than templates present in the genome.
The family of ONSEN retrotransposons in the Col-0 reference genome consists of eight full length copies distributed over chromosomes 1, 3, and 5 (Supplementary Fig. S1A). To investigate whether all copies contribute to the pool of extrachromosomal DNA under heat stress, we performed a sequencing analysis of isolated exDNA, scoring for element-specific SNPs in the ONSEN coding region that distinguish seven out of eight genomic templates (Supplementary Fig. S1A, Supplementary Table S1). Out of 55 independent clones, 57% could be assigned to three elements (ONSEN 1: At1g11265; ONSEN 2: At3g61330; and ONSEN 3: At5g13205; Fig. 2) that form a subgroup with 100% identical 5′and 3′ LTRs, indicating their evolutionarily recent transposition [21]. Several other sequences (20%) contained additional and different SNPs, so not allowing an unambiguous assignment to genomic templates and supporting the notion that reverse transcription is an error-prone process [28].
In spite of the prevalence of the elements with perfect LTRs, three other ONSEN elements from the Col-0 genome contributed 23% to the pool of sequences, indicating that ONSEN 4, 5 and 6 (At1g58140, At1g48710 and At3g59720) are still functional and capable of forming extrachromosomal copies. In contrast, ONSEN 7 and 8 (At1g21945 and At3g32415) were not represented. Interestingly, these two copies are shared between most natural accessions of Arabidopsis [23], and they have acquired the largest number of polymorphisms.
RdDM is the major pathway to restrict transposon mobility by installing DNA methylation, leading to a reinforcing loop that creates a heterochromatic environment. Most RdDM targets have multiple CHG and CHH sequences, which become modified by cooperative action of two different DNA methyltransferases. However, the LTR of active ONSEN copies contains only CHH sites (Supplementary Fig. S1B). To analyze the DNA methylation state at the 5′ LTRs of several ONSEN elements, we performed bisulfite conversion and sequencing before and after heat stress, distinguishing three genomic copies by specific primers in the flanking genomic regions (Supplementary Table S2). At ambient temperature, the two most recently transposed elements (ONSEN 1 and 2) showed relatively high CHH methylation profiles across their 5′LTRs (Fig. 3A, Supplementary Fig. S2A, B, and Supplementary Dataset S1). Strikingly, ONSEN 8 that was not activated had substantially less methylation under normal growth conditions (Fig. 3A, Supplementary Fig. S2C). After heat stress, ONSEN 8 was more methylated throughout the 5′LTR, while ONSEN 1 and 2 lost the modification at several positions (Fig. 3A, Supplementary Fig. S2D–F).
To investigate whether this erasure of DNA methylation had any impact on the level of ONSEN activation, we repeated RNA and exDNA quantification as well as DNA methylation analysis in the triple mutant ddc (drm1/drm2/cmt3), lacking RdDM-mediated methylation at CHH and CHG sites [29]. As expected, the levels of CHH methylation were significantly reduced at the ONSEN LTRs even in unstressed mutant plants (Fig. 3B, Supplementary Fig. S2G–I, and Supplementary Dataset S1). However, this was not sufficient to trigger any ONSEN transcription or exDNA formation (Fig. 3C, D). After heat stress, the already reduced CHH methylation in the triple mutant was hardly changed (Fig. 3B, Supplementary Fig. S2J–L and Supplementary Dataset S1), but the levels of ONSEN mRNA, as well as its extrachromosomal DNA, were significantly increased after heat stress compared to Col-0 wild type (Fig. 3C, D). Sequencing 47 clones of ONSEN extrachromosomal DNA from ddc mutant plants after heat stress revealed an even more dominant representation of activated ONSEN 1 and 2 than in the wild type (Supplementary Fig. S3). Collectively, these results indicate that reduction of DNA methylation at the retrotransposon's promoter does not per se activate the element but can favor the activation of ONSEN upon heat exposure.
The strong activation of ONSEN by heat stress conditions, together with the enhanced response in the methylation triple mutant, suggested the dependence on a heat-induced transcription factor whose action could be modulated by presence or absence of the DNA modification. Therefore, we analyzed the ONSEN promoter for cis-regulatory elements and found a heat response element (HRE) with the consensus sequence nTTCnnGAAn in the LTR of all elements (Supplementary Fig. S1C). HREs are bound by heat shock factors (HSFs), which form trimers and thereby induce expression of downstream target genes [30]. Arabidopsis has 21 heat shock factors [31], among which HSFA1 was identified as a main positive regulator in heat-responsive gene expression [32].
To analyze whether HSFA1 mediates the transcriptional activation of the retrotransposon we determined ONSEN RNA and exDNA in hsfa1 mutants. There are 4 genes for HSFA1-type proteins in Arabidopsis [a, b, d, e; 32] with partially redundant functions. We tested all different combinations of triple mutants along with the quadruple mutant (Fig. 4). HSFA1b or d were sufficient to activate ONSEN to comparable levels as in the wild type, and HSFA1a to a lower extent. In contrast, no ONSEN RNA or exDNA was formed after heat stress in the quadruple mutant or in plants with HSFA1e as the only functional HSFA1 factor. Therefore, either HSFA1a, or HSFA1b, or HSFA1d are necessary for heat-induced expression of ONSEN. However, they were unlikely candidates for direct transcriptional activators of ONSEN, since all three genes are constitutively expressed and proposed to initiate a cascade of heat stress-responsive genes only upon additional signals [32].
Transcripts of the Arabidopsis heat stress transcription factor HSFA2 are not detectable at ambient temperatures, but the gene is most strongly and stably expressed under heat stress conditions [33]. The protein can be found in leaves after 3 hours of heat stress and is still present after 21 hours of recovery [34]. We quantified the HSFA2 transcript under our heat stress conditions in wild type, hsfa1 triple and quadruple mutants. Interestingly, HSFA2 transcription showed the same dependence on the individual HSFA1 factors as ONSEN, hardly or not at all transcribed upon lack of HSFA1a,b,d or of all four factors (Fig. 5A). This suggested HSFA2 as a candidate for the heat stress factor necessary to activate ONSEN.
The HSFA2 protein has a highly conserved N-terminal DNA binding domain (DBD) that is required for its binding to HREs, and mutation within the domain or within the HRE block the binding [34] . To investigate whether HSFA2 would bind to the HRE in the ONSEN promoter we performed electrophoretic mobility shift assays (EMSAs) with recombinant Arabidopsis HSFA2 protein and the LTR DNA (Fig. 5B). Indeed, HSFA2 bound to the LTR probe in a concentration-dependent and specific manner: increasing amounts of the protein enhanced the shift (Fig. 5B, lanes 1–5), while pre-incubation with non-labeled LTR fragment inhibited the shift of the labeled probe (Fig. 5B, lane 6). The specificity of HSFA2 binding to the HRE in the LTR could be further supported by successful binding competition with an LTR fragment containing just 51 bp with the complete HRE (Fig. 5B, lane 7).
To further confirm the involvement of HSFA2 in transcriptional activation of ONSEN we quantified RNA and extrachromosomal DNA with and without heat stress in the hsfa2-1 mutant (Fig. 5C, D). ONSEN activation was severely reduced in the mutant, although not to the same low level as in the hsfa1a/b/d triple or the hsfa1a/b/d/e quadruple mutants. This indicates that HSFA2 is a major, but probably not the only heat shock factor involved in the heat-induced activation of ONSEN. However, it is clear that the HRE in its promoter couples the retrotransposon to the heat response pathway of the plant, thereby exploiting an important defense mechanism of its host to prepare for its own amplification.
Transgenic analysis of the HSFA2 promoter activity by fusion with the GUS reporter gene had revealed very low activity under non-stress conditions but high expression under heat stress, in rosette leaves and even more in veins, root tips and root branching points [35]. We asked if the expression pattern of the ONSEN promoter would be similar and constructed a transgene consisting of the full 440 bp LTR upstream of a sequence encoding a GUS-GFP fusion protein (Supplementary Fig. S4A). This construct was introduced into wild type plants, and T4 plants homozygous for a single copy insert of the transgene analyzed for reporter gene expression. Under ambient temperatures, very low GUS activity was detectable in some root cells, while the 30 h heat stress treatment resulted in deep blue staining all over the plants (Fig. 6A–F). Expression starts already shortly after the beginning of the stress, as the first staining in leaf and root tips can be seen as early as 1 h, sometimes also at specific sites such as the leaf tips or emerging root branches, similar to the pattern seen for HSFA2 (Supplementary Fig. S4B, C; [35]). Reporter gene expression later spreads rapidly (Fig. 6G–L). Live imaging of the GFP expression confirms the preferential promoter activity of the LTR in dividing tissues (Supplementary Fig. S4D, E). Correspondingly, the amount of extrachromosomal ONSEN DNA after heat exposure in meristem-enriched tissue exceeds by 8 times that in leaves (Supplementary Fig. S4F), thereby producing a high number of retrotransposon copies ready to reintegrate into the genome of cells that give rise to the next generation. This might represent an especially successful strategy of the element to exploit the plant's own indispensable stress protection mechanism to increase the probability of proliferation.
Transposons employ various strategies to gain replication advantage over their hosts in order to produce additional copies. In this study we show that ONSEN exploits a conserved stress defense response to initiate its amplification, in tissue that ultimately produces the germ line, and avoiding stable maintenance of DNA methylation.
Several mobile elements in many organisms are activated upon different environmental cues. Closest to the example described here is the heat responsiveness of Copia elements in Drosophila. While the data on the rate of transposition and their interpretation are controversial [reviewed in 36], there is unquestioned evidence for a correlation between heat-induced expression of a copia type retrotransposon and the presence of heat shock consensus sequences in the promoter [37]. However, in contrast to ONSEN, the Drosophila element is transcribed even without heat stress, the expression increase is faster but much less pronounced, and its promoter lacks at least three more copies of the nGAAn sequence required for high affinity interactions with HSFs [30]. Accumulation of DNA copies of the Drosophila element was not described and is unlikely, since a substantial amount of transcript is needed to generate sufficient Gag and Pol proteins and to serve as template for reverse transcription. In the case of the tobacco Tto1 retrotransposon studied in a heterologous system, Gag protein and linear DNA molecules are synthetized in a ratio of 1400∶1 [38]. Therefore, ONSEN has been (more) successful in acquiring an efficient and specific heat-responsive regulatory element that allows production of extrachromosomal DNA copies by far exceeding the genomic copies.
Extended heat stress conditions can transiently release epigenetic regulation from several genes and transposons [19], [20], and most of these do not have a heat responsive element in their promoter. Some transposons respond equally well to other types of stress [39]. High copy number repeats like TSI and the GUS transgene are effectively activated also in an hsfa2 mutant [19]. The requirement of the HSFA factors suggests that the pronounced ONSEN activation is not just a consequence of a general loss of epigenetic control as for the other elements. At least one of the three genes HSFA1a, b or d is necessary to generate ONSEN transcript and extrachromosomal DNA, and the triple mutant shows severely impaired resistance and viability under heat stress conditions [40]. Expression of HSFA2, the dominant heat shock factor downstream of the HSFA1 complex in Arabidopsis [33] able to bind to the HRE of several target genes [34], is almost completely eliminated in the hsfa1a/b/d triple mutant upon heat exposure, like ONSEN. The specific binding of the HSFA2 protein to the HRE in the ONSEN LTR in vitro demonstrates that ONSEN can recruit this factor, thereby “masking” itself as a heat shock gene and coupling its transcriptional activation to the plants' heat stress response. Reduced but not completely abolished ONSEN transcription in the hsfa2 mutant indicates that it can also acquire other factors from the partially redundant heat-related transcription factor family.
HSFA2 protein is found in leaves, stem, flowers, green siliques and roots of heat-stressed plants [34], and the GUS reporter with the ONSEN LTR shows a similar ubiquitous expression after longer heat exposure. However, the early and strong expression of the ectopic ONSEN GUS fusion construct in tissue with dividing cells resembles the more pronounced expression in the root meristems and root branch points of the HSFA2 promoter [35]. Although an analysis of microarray data for transposon expression under ambient temperatures in different Arabidopsis tissue did not provide evidence for a preferential expression in dividing tissue [41], ONSEN activation upon heat stress was elevated in undifferentiated callus [22], and several mobile elements in maize are preferentially expressed in the shoot apical meristem [42]. This and other evidence has led to the hypothesis of developmental relaxation of TE silencing (DRTS) and suggested that transposons could amplify preferentially in tissues or cells that undergo epigenetic reprogramming in the course of developmental processes [43]. Accordingly, reintegration of ONSEN into new locations occurred frequently in somatic cells during flower formation, although only in epigenetically compromised mutant plants [21]. Therefore, exploiting the higher activity of the heat stress defense in highly dividing tissue for a preferential accumulation of extrachromosomal ONSEN copies might be another optimization strategy of the element to achieve a higher probability of transmission to the next generation. What appears as developmental relaxation might, in some cases at least, be a similarly sophisticated adaptation.
In addition to coupling itself to the plants' stress response and preferentially producing amplicons where propagation chances are high, the sequence of the ONSEN promoter might represent another refinement. Its LTR has a GC content of 28%, not far from the average of 34.7% in the Arabidopsis genome [44], but it does not contain any C in palindromic arrangement between the strands. This is in contrast to LTRs of similar length and composition of other elements that are under the control of the RdDM pathway: EVADÉ [8] (AtCopia93, 406 bp, 37% GC, 5 CG and 6 CHG) and solo-LTR [45] (376 bp, 27%GC, 2 CG, 3 CHG). The absence of CG in the ONSEN promoter circumvents maintenance methylation by MET1, and the lack of any CHG site also precludes modification by CMT3. This, in consequence, interferes with the reinforcing loop in which the histone methyltransferase KRYPTONITE (KYP) installs H3K9me2 [46], thereby strongly impeding the formation of heterochromatin at the ONSEN LTR. The CHH sites in the promoter can only be modified by CMT2 [47] or RdDM, two synergistic pathways that are responsible for substantial methylation of transposable elements in Arabidopsis. While ONSEN was not amplified in non-stressed ddc mutant plants, the activation after heat stress was more pronounced compared to wild type. Since the lack of the RdDM-related methyltransferases DRM1 and DRM2 reduced, but did not totally erase DNA methylation at ONSEN LTRs. the residual CHH methylation by CMT2 could potentially also contribute to the responsiveness of the promoter. However, lack of CMT2 is also not sufficient to completely eliminate CHH methylation at the ONSEN LTRs (Assaf Zemach and Daniel Zilberman, personal communication). This is only obtained in plants lacking both chromatin remodelers DDM1/DRD1 that support CMT2 and RdDM-mediated methylation [47]. However, even upon complete loss of CHH methylation in the double mutant ddm1/drd1, there is very weak activation of ONSEN in non-stressed plants (Assaf Zemach and Daniel Zilberman, personal communication). A more pronounced response of ONSEN to heat stress after partial demethylation may simply indicate better accessibility, as suggested for other regions in Arabidopsis [48].
In spite of the multiple levels at which ONSEN has optimized its transcription and production of extrachromosomal copies, it is an element with only a limited copy number in the Arabidopsis thaliana reference genome, and although other accessions have variable copy numbers [23], we do not know of any ecotype with a substantially larger number of inserts. Therefore, integration of extrachromosomal DNA copies into the plant genome seems to be a limiting factor, although rapid propagation of ONSEN was observed in the progeny of plants compromised in the RdDM pathway [21]. Restricting the chances for successful retrotransposon integration might indicate one efficient counter defense of the plants, the rapid mutation of already integrated copies another. Since retrotransposons usually remain in the genome and LTRs are identical upon insertion, an evolutionary history of degeneration like this can in fact be read from their sequence. The extrachromosomal ONSEN DNA copies were mainly produced from elements with 100% identical LTRs. These evolutionarily young elements are found in very few of the 95 Arabidopsis accessions tested [23], whereas elements not represented in the extrachromosomal fraction have acquired a large number of polymorphisms and are shared between most accessions. In addition to the eight recognizable full length copies, the reference genome has more than 10 sequences annotated as incomplete Copia78 = ONSEN (TAIR database), remnants that indicate previous invasion but sequence degeneration of the element.
The tandem repeat array of the heat-responsive LTRs at the 5′ and 3′ end of ONSEN render any downstream sequence potentially transcribed upon heat stress, and a new insertion indeed conferred heat responsiveness to a neighboring gene [21]. Our finding that this is exerted by integration of the plants' target site for a heat shock transcription factor into the LTR lets us speculate that this is another example of co-evolution between the genomes of host and TE: activation of the element in times of stress might allow the element to propagate but provides at the same time some benefit by rendering additional plant genes stress-responsive, thereby generating genetic diversity as a premise for selection. It resembles other copia-like retrotransposon insertions in Citrus that convey cold induction of the transcription factor Ruby controlling anthocyanin production. This results in the intensive coloration of blood oranges [49] and is just one example that stress-induced transposon-mediated control of plant genes might also be of interest for plant breeders and agriculture.
On the one hand, the acquisition of the plants' heat responsive element by ONSEN masks the retrotransposon as familiar, corresponding to the “wolf-in-sheep's-clothing” strategy [50], and represents a particularly intriguing molecular parasitism in which it is impossible for the plant to respond appropriately to heat stress without the risk of retrotransposon amplification. On the other hand, plants might occasionally benefit from this strategy if their own regulatory element gets moved around upon stress into new insertion sites where it might prove useful. Evolution has the last word.
The hsfa2-1 [51] and drm1 drm2 cmt3 triple mutant (ddc) [29], [52] in Col-0 background were previously described and obtained from the NASC stock center. hsfa2 has a T-DNA insertion in At2g26150 (SALK_008978). cmt3-11, drm1-2 and drm2-2 are T-DNA insertion mutants in AT1G69770 (SALK 148381), At5g15380 (SALK 021316), and At5g14620 (SALK 150863), respectively. Also the hsfa1 mutants were previously described [32], [53]. hsfa1a (former hsf1-tt1) has an insertion in At4g17750, hsfa1b (former hsf3-tt1) in At5g16820, hsfa1d-1 in At1g32330 (SAIL_410_E01) and hsfa1e in At3g02990 (SALK_094943). Triple (aTK, bTK, dTK, eTK) and quadruple (QK) mutants of HSFA1 generated in Col-0 and Wassilewskija background were a kind gift from Yee-yung Charng.
Plants were sown on germination medium and grown in vitro at 21°C under long-day conditions (16 h light/8 h dark) for 21 days prior to heat stress. For heat stress treatments (HS), plants were transferred to 37°C for 30 h (standard HS conditions) or shorter periods as indicated, starting 3 h after beginning of the light period. GFP and GUS images were taken on 7 day-old seedlings after standard heat stress conditions. Meristem-enriched tissue was prepared from 21 day-old seedlings by manual dissection of shoot tips smaller than 2 mm.
Total RNA was extracted from whole seedlings using the RNeasy Plant Mini Kit (Qiagen). An additional DNase treatment was performed to remove all extrachromosomal ONSEN DNA in the heat stress samples, prior to cDNA synthesis with random hexamer primers and the RevertAid M-MuLV Reverse Transcriptase (Thermo Scientific).
All qRT-PCR analyses were performed using the Sensi Mix SYBR & Fluorescin Kit (Bioline) and iQ5 equipment (Biorad). Each data point is based on nine PCR reactions deriving from three biological replicates. Expression values were normalized to AtSAND (At2g28390), documented to have equal expression levels under all tested conditions [54], [55].
Genomic DNA was isolated from whole seedlings, leaves or meristem-enriched tissue using the Illustra DNA Extraction Kit Phytopure (GE Healthcare). Total DNA was separated on 0.8% agarose gel, depurinated for 10 min in 250 mM HCl, denatured in 0.5 M NaOH and 1.5 M NaCl for 30 min, and neutralized in 0.5 M Tris, 1.5 M NaCl and 1 mM EDTA at pH 7.2 for 2×15 min. DNA was blotted onto Hybond N+ membranes (Amersham) with 20× SSC, washed and UV-crosslinked with a Stratalinker (Stratagene).
Hybridization was performed as described [56]. An ONSEN-specific probe (see Supplementary Table 2) was radioactively labelled with 50 µCi of dCT-α-32P (Amersham) using the Rediprime II Random Prime Labelling System (GE Healthcare) and purified via G50 Probequant (Amersham) columns. Signals were detected using Phosphorimager screens (Amersham) and scanned by a Molecular Imager FX (Biorad). Densitometric quantification was performed using Image Lab software on three biological replicates.
Undigested DNA prepared from whole seedlings was run on a 0.8% agarose gel, and a region around the size range of 5 kb, corresponding to the full length of linear ONSEN copies, was cut out. DNA was purified using the QIAquick Gel Extraction Kit (Qiagen). After a PCR amplification step, using ONSEN-SNPs F and R primers (Supplementary Table S2), the PCR product was purified with the MinElute Gel Extraction Kit (Qiagen) and ligated using the InsTAclone PCR Cloning Kit (Thermo Scientific). Individual clones were Sanger-sequenced.
Bisulfite conversion of BamHI digested genomic DNA was performed using the Epitect Kit (Qiagen). Purified DNA after conversion was amplified and cloned for Sanger sequencing. Primers are listed in Supplementary Table S2. For each analysis, at least 13 independent clones were Sanger-sequenced. Sequencing data were visualized and the methylation percentage was calculated for each cytosine position using CyMate [57].
The full length of the HSFA2 ORF was amplified by PCR from cDNA (see Supplementary Table S2), cloned in frame with the C-terminal 6× His-tag using NdeI and XhoI sites in the pET-24B vector. The construct was transformed into E. coli strain BL21(RIL) (Novagene) for expression. The bacterially expressed HSFA2-His fusion protein was purified with HisTrap 5 ml columns (GE Healthcare) and ÄKTA purifier (GE Healthcare) and used for EMSA. EMSA was performed as described previously [58]. The LTR sequence was amplified by PCR (see Supplementary Table S2) and the fragment gel-purified prior to labeling. For the competitor assay, complementary oligonucleotides (Supplementary Table S2) were annealed to generate double-stranded DNA. All probes were end-labeled using T4 PNK and (γ-32P)ATP and purified with G50 Probequant columns (Amersham). The binding assay was carried out in a buffer containing 10 mM Tris (pH 7.5), 1 mM EDTA, 0.1 M KCl, 0.1 mM DTT, 5% vol/vol glycerol and 0.01 mg/ml BSA. The binding reaction was incubated at RT for 30 min. The reaction mixtures were separated on 5% non-denaturating polyacrylamide gels in 0.5× TBE buffer at 140 V for 2 h, and the gels were exposed to Phosphoimager screens (Amersham) and scanned by a Molecular Imager FX (Biorad).
The 440 bp LTR sequence of ONSEN 1 was cloned into the pENTR2B vector (Invitrogen) using BamHI and EcoRV restriction sites and subsequently into the pBGWFS7 binary vector [59] using the Gateway LR Clonase II Enzyme Mix (Invitrogen). The transgene (Supplementary Figure S4) was introduced into Agrobacterium strain AGL1, which was then used to transform Col-0 plants by floral dipping [60]. After selection of primary transformants by selection with Basta, plants were grown and allowed to self-pollinate. Plants homozygous for a single copy of the transgene (confirmed by Southern blot analysis) were selected and amplified. All experiments were done with plants of the T4 generation.
GUS histochemical staining was performed as described [61].
|
10.1371/journal.pgen.1003000 | The Contribution of RNA Decay Quantitative Trait Loci to Inter-Individual Variation in Steady-State Gene Expression Levels | Recent gene expression QTL (eQTL) mapping studies have provided considerable insight into the genetic basis for inter-individual regulatory variation. However, a limitation of all eQTL studies to date, which have used measurements of steady-state gene expression levels, is the inability to directly distinguish between variation in transcription and decay rates. To address this gap, we performed a genome-wide study of variation in gene-specific mRNA decay rates across individuals. Using a time-course study design, we estimated mRNA decay rates for over 16,000 genes in 70 Yoruban HapMap lymphoblastoid cell lines (LCLs), for which extensive genotyping data are available. Considering mRNA decay rates across genes, we found that: (i) as expected, highly expressed genes are generally associated with lower mRNA decay rates, (ii) genes with rapid mRNA decay rates are enriched with putative binding sites for miRNA and RNA binding proteins, and (iii) genes with similar functional roles tend to exhibit correlated rates of mRNA decay. Focusing on variation in mRNA decay across individuals, we estimate that steady-state expression levels are significantly correlated with variation in decay rates in 10% of genes. Somewhat counter-intuitively, for about half of these genes, higher expression is associated with faster decay rates, possibly due to a coupling of mRNA decay with transcriptional processes in genes involved in rapid cellular responses. Finally, we used these data to map genetic variation that is specifically associated with variation in mRNA decay rates across individuals. We found 195 such loci, which we named RNA decay quantitative trait loci (“rdQTLs”). All the observed rdQTLs are located near the regulated genes and therefore are assumed to act in cis. By analyzing our data within the context of known steady-state eQTLs, we estimate that a substantial fraction of eQTLs are associated with inter-individual variation in mRNA decay rates.
| Recent studies of functional genetic variation in humans have identified numerous loci that are associated with variation in gene expression levels, called expression quantitative trait loci (eQTLs). The mechanisms by which these loci affect gene expression, however, are still largely unknown. Specifically, since most studies rely on measures of steady-state gene expression levels, they are unable to distinguish between the relative influences of either transcriptional- or decay-related processes. To address this gap, we examined the specific impact of mRNA decay processes on steady-state gene expression levels for over 16,000 genes in human lymphoblastoid cell lines. By characterizing decay rates in 70 individuals, we show that steady-state expression levels are significantly influenced by variation in decay rates for 10% of genes. Yet, for roughly half of these genes, we find that individuals with higher expression levels also have faster decay rates. This pattern points to a non-simple mechanistic interplay between transcriptional and decay processes, especially for genes involved in rapid cellular responses. Finally, we identify 195 genetic variants that are significantly associated with both gene expression variation and variation in mRNA decay rates. Using these data, we estimate that that a substantial fraction of eQTLs are associated with inter-individual variation in mRNA decay rates.
| Substantial variation in gene expression levels exists in natural populations [1]–[5]. Over the past decade, we have learned that much of this inter-individual regulatory variation is associated with specific genetic polymorphisms, which can be identified by mapping expression quantitative trait loci (eQTLs) [6]–[10]. Expression QTL mapping studies in different organisms have led to important insights into the genetic basis for gene regulation and, in a number of cases, into the mechanistic basis for complex phenotypes. In particular, recent eQTL mapping studies in humans have identified thousands of genetic variants affecting gene expression levels [11]–[14], some of which are loci that are also associated with complex diseases [15]–[18]. Nearly all human eQTLs, regardless of the tissue in which they were found, have been identified near the regulated genes and hence are assumed to act in cis. A partial explanation for the relatively small number of trans eQTLs that have been identified is the low power to map such loci compared to cis acting eQTLs (due to the stringent significance criteria required to avoid false positives when mapping across the entire genome, and generally small effect sizes of trans-QTLs [8], [19]–[25]).
Despite the recent success in mapping gene expression phenotypes, we still know little about the specific regulatory mechanisms that underlie eQTLs [26]–[29]. Partly, this gap is being addressed by a growing number of large-scale mapping studies of inter-individual variation in genetic and epigenetic regulatory mechanisms (which complement studies of gene expression variation [13], [30]–[34]). Yet, even by incorporating such studies, the processes underlying regulatory variation and their relative importance remain difficult to infer, because all eQTL studies to date – regardless of the model system or species - have relied on measures of steady-state gene expression levels.
Steady-state gene expression levels are generally the result of two opposing biological processes: mRNA transcription, which includes transcript initiation, elongation, and processing, and mRNA decay, which includes spontaneous and targeted degradation of transcripts, as well as dilution [35], [36]. Using only measurements of steady-state gene expression levels, it is impossible to determine the relative contribution of variation in transcription rates and mRNA decay rates to overall regulatory variation. In other words, without additional data, the particular mechanisms underlying steady-state expression level QTLs cannot be inferred with confidence.
To better understand the basis for variation in steady-state gene expression levels requires data on specific aspects of gene regulatory mechanisms. Most recent studies that have done so (though only rarely in the context of QTL mapping), have focused on understanding transcriptional processes contributing to gene expression variation, such as splicing, DNA methylation, histone modification, chromatin accessibility, and transcription factor binding. Results from this emerging body of work indicate that although transcriptional processes contribute substantially to steady-state measurements of gene expression, neither the independent or combinatorial effects of these mechanisms can completely account for variation in steady-state gene expression levels [28], [29], [37], [38]. It is likely that a better account of regulatory variation can be obtained once transcription initiation and RNA decay mechanisms are considered together.
While the details of transcriptional regulation are becoming increasingly understood, the mechanisms influencing variation in mRNA decay rates have thus far received less attention, particularly in mammalian systems [11], [37]–[39]. This bias may reflect the prevalent assumption that transcription initiation rates are the major determinants of overall gene expression levels [40]–[43]. Yet, a few recent studies of mRNA decay mechanisms have challenged this historical view [3], [33], [44]–[47]. In particular, it has been argued that the regulation of mRNA decay processes might be a key determinant of the expression patterns of a large subset of genes. Recent studies in eukaryotic cells have revealed a wide variability of mRNA decay rates across transcripts – with individual mRNA half-lives ranging from a few minutes to several hours – which can often be tied to differences in the functional role of the regulated genes [44], [48]–[50]. For example, studies in yeast, worms, plants, and human primary cells have all found that genes involved in the regulation of transcription tend to produce mRNA that decays faster than mRNA from genes involved in cell cycle or metabolic pathways [1], [3], [41], [48], [51], [52]. Furthermore, the steady-state mRNA levels of the lowest or highest expressed genes are strongly correlated with mRNA decay rates [41], [44], [49], [50], suggesting that in these cases, regulation of mRNA decay is likely an important determinant of gene expression levels.
A number of mechanisms are known to contribute to variation in mRNA decay rates among genes. These include the roles of certain RNA-binding factors such as small RNAs, RNA-binding proteins, and larger RNA-binding complexes, all of which have been shown to bind to both general (such as the AU-rich 3′ untranslated region elements; AREs [15], [26], [53]) and specific RNA motifs [11], [37], [54]. For example, many RNA-binding small RNAs, including miRNAs, have been shown to expedite decay of specific transcripts by creating double stranded RNA that is targeted for degradation by endonuclease enzymes [11], [38], [47]. Similarly, certain interactions between RNA binding proteins and mRNA have been shown to contribute to either higher (“destabilizing proteins”) or lower decay rates (“stabilizing proteins”), though the mechanisms by which they act are not yet fully understood [11], [54], [55]. More generally, we now appreciate that, much like transcription rates, mRNA decay rates are regulated by a combination of trans elements (such as proteins, complexes, or small RNAs) binding to a collection of cis binding motifs (typically included within the transcript itself) [6], [56], [57]. However, despite increasing understanding about mechanistic details of mRNA decay processes, we still know little about inter-individual variation in mRNA decay rates, in any species.
We characterized mRNA decay in 70 Yoruba lymphoblastoid cell lines (LCLs) from the HapMap project [36], [58]. These cell lines have been extensively genotyped and/or sequenced at high-depth [40], [59], [60], making them ideal for genetic mapping studies. To determine decay rates, we measured changes in mRNA abundance levels in each cell line at different times after treatment with the RNA elongation complex inhibitor Actinomycin D (ActD), which arrests transcriptional processes. We measured mRNA abundance before treatment (time point 0) and at four time points after treatment (at 0.5 hours, 1 hour, 2 hours, and 4 hours). To account for the decrease in total RNA caused by the ActD treatment over the timecourse experiment, we increased the number of cells from which we extracted RNA as the experiment progressed (Figure S1). We thus were able to hybridize the same amount of mRNA from each time point to an Illumina HT-12 expression microarray. We processed a total of 350 samples over the five time points and seventy cell lines (see Table S1). Our experimental design allowed us to normalize transcript abundance across all 350 arrays using standard approaches (see Methods for more details).
To estimate mRNA decay rates, we fit an exponential decay model to the normalized expression data to obtain estimated gene-specific decay rates for each cell line. Due to our choice of hybridization study design and normalization procedure, all estimated decay rates are relative to the mean cellular mRNA decay rate in the sample, which itself can be estimated by taking into account the number of cells used to extract RNA across the time points (see Methods for more details). We excluded from all further analyses genes that were not detected as expressed even before the arrest of transcription (time point zero) in at least 80% of individuals (see Methods). Overall, we obtained individual-specific estimates of mRNA decay rates for 16,823 Ensembl genes (see Table S1).
As a first step of our analysis, we characterized the genome-wide distribution of mRNA decay rates. To do so, for each gene we used the median decay rate across individuals as a measure of the gene-specific mRNA decay rate. We observed a wide range of mRNA decay rates across genes (Figure 1A), consistent with findings of previous studies. We also observed a substantial amount of variation in decay rates across individuals within each gene (Figure 1B), consistent with expectations from previous studies in human cells [1], [35], [40]. We classified genes as either consistently slow or fast decaying when their decay rates in at least 80% of individuals in our study were classified as slow or fast relative to the individual-mean decay rate (see Methods). We thus identified 146 genes that consistently decayed slower than average across individuals and 716 genes that consistently decayed faster than average.
In agreement with previous observations, we found that genes with related biological functions often decayed at similar rates [1], [52,52]. Genes with slower decay rates tend to be involved in cellular and organelle-related housekeeping processes, such as cytoplasmic and mitochondrial processes (Table S2). Genes with faster decay rates are enriched for gene regulatory functions that might require rapid mRNA decay to ensure rapid turnover of expression levels in response to changing cellular conditions (Table S3). This includes enrichments for functional annotations such as metabolic processes, regulation of gene expression, and regulation of transcription.
We next investigated possible mechanisms that could account for variation in mRNA decay rates across genes. Previous studies have suggested that increased transcript length [3], [41], and specifically 3′UTR length [1], [3], might significantly influence mRNA decay rates. Indeed, we find that both are slightly but significantly positively correlated with decay rates across genes (Spearman ρ = 0.17, P<10−16 for gene length and Spearman ρ = 0.09, P<10−16 for 3′UTR length). This association is also evident when we limit this analysis only to genes classified as decaying slower or faster than the mean decay rate (Figure 2A; Figure S4; Spearman ρ = 0.15; P<10−16 for gene length and Spearman ρ = 0.09; P<10−8 for 3′UTR length). The increased 3′ UTR length in faster decaying genes is thought to indicate an increase in potential regulatory space that could harbor RNA-decay regulatory elements (reviewed in [6]).
Studies of mRNA decay of individual genes have previously identified two main classes of cis regulatory elements that might play roles in decay processes: microRNA (miRNA) binding sites [11] and AU-rich elements [15], [17]. To determine the possible influence of miRNA binding on decay rates in the LCLs, we curated several miRNA databases [19], [20], [22]–[25] to create a list of confident miRNA target binding sites (see Methods S1). To account for the confounding effect of transcript length (more binding sites in longer 3′UTRs), we standardized the number of miRNA target binding sites by the 3′UTR length (see Methods). Using this approach, we found a slightly positive correlation between the density of miRNA target sites and decay rates. Again, when we focused exclusively on the genes classified as decaying slower or faster than the mean decay rate, we observed a stronger association (Figure 2B, Spearman ρ = 0.16; P<0.003). We then considered the presence of AU-rich elements (AREs) in slower versus faster decaying genes. To do so, we used the AREScore algorithm [26], which searches within 3′UTRs for features associated with typical type-II AREs, to assign an AREScore to each gene. A larger AREScore essentially implies increased potential for binding by an ARE-recognizing RNA binding protein to regulate the decay processes of the gene. We found that there is a significantly increased median AREScore in faster decaying genes compared to slower decaying genes (Figure 2C, Spearman ρ = 0.14; P<10−16).
As our findings support the general notion that cis regulatory elements, such as miRNA bindings sites or AU-rich elements, are important determinants of mRNA decay rates, we next searched for additional sequence motifs that might represent novel binding sites for specific decay factors in LCLs. To do so, we used the FIRE algorithm [30] to search for motifs in the 146 slow decaying genes and 716 fast decaying genes. We identified three significantly enriched motifs – one enriched in the fast decaying genes, and two enriched in the slow decaying genes (Figure 2D). We performed the motif search across the entire promoter and transcript region for each gene, yet all three enriched motifs are located in 3′UTRs. The motif enriched in fast decaying genes closely resembles a typical AU-rich element sequence. The two motifs enriched in slow decaying genes could not be linked to any currently known miRNA seed sequence or RNA-binding protein motif and hence might be novel regulatory elements.
We are specifically interested in the effect that mRNA decay has on steady-state expression levels (in these analyses, we defined “steady-state expression” as the mean expression across all time points so that our estimates of steady-state expression levels would be statistically independent of the estimated decay rates when the null hypothesis of no association between steady-state levels and decay rates is true; see Methods). Considering this relationship across all genes (Figure 3A), we found little or no correlation between decay rates and gene expression levels. However, we observed a significant difference in expression levels between genes classified as decaying significantly slower or faster than the mean decay rate (as defined above; P<6×10−6, Figure 3B; Figure S5). This difference in expression levels is in the expected direction – that is, genes with slower decay rates have higher steady-state expression levels than genes with faster decay rates.
We also observed a small number of cases in which genes with faster decay rates are highly expressed (we refer to this as a ‘discordant’ relationship between gene expression levels and decay rates). One example is the BTG1 gene, which is involved in regulating the glucocorticoid receptor autoregulatory pathway [35], and has both a significantly increased decay rate and a high expression level (Figure S5). Interestingly, seven of the top nine genes with discordant patterns (both the expression levels and decay rates of these nine genes are within the top 5% of the genome-wide distributions of gene expression and decay rates respectively; Figure 3C; see Methods) have been experimentally shown to be involved in auto-regulatory or regulatory feedback pathways (Table 1) [61]–[69]. More broadly, the top 49 genes with discordant patterns (constituting the top 10% of both the genome-wide distributions of gene expression levels and decay rates; Figure 3C) are enriched for genes with functions related to signaling pathways, stress response, and immune function (when genes expressed in LCLs are used as the background for the analysis; Table S4).
We next examined the extent to which variation in decay rates might contribute to overall variation in steady-state expression levels across individuals. For each gene, we calculated the correlation between gene expression levels and mRNA decay rates across individuals and focused on genes with a significant (FDR = 10%) correlation between the two measurements (Figure 4A). We found a significant negative correlation between expression levels and decay rates for 695 genes. It is reasonable to assume that inter-individual variation in steady-state expression levels of these 695 genes is driven by corresponding variation in decay rates. Based on gene ontology functional annotations, these 695 genes are enriched for genes involved in endopeptidase inhibitor and regulator activity (Table S5).
On the other hand, we also found a discordant relationship between gene expression levels and decay rates across individuals for 989 genes (10% FDR; Figure 4A). This finding may seem counter-intuitive as it contradicts our expectation that higher decay rates should result in lower steady-state gene expression levels. However, genes with a discordant relationship between expression and decay are enriched for processes involved in the regulation of cellular, metabolic, and transcriptional activities (Table S6). A similar observation of discordant relationships between decay rates and expression levels that are enriched for genes in the same functional categories (metabolic, and transcriptional activities) has been previously reported in yeast [37], [38]. Put together, these results suggest a role for mRNA decay in complex regulatory circuits that have the property of fast response time, for instance auto-regulation by negative feedback loops.
Studies across yeast species [37], [38] have further suggested that positive correlations between gene expression levels and decay rates are often coupled with correspondingly increased transcription rates – presumably to increase response speed [40]. To test this notion in our system, we used a combination of previously published [33] and newly generated PolII occupancy ChIP-seq data from seven of the same Yoruba LCLs as a proxy measurement of gene-specific transcription rate (Table S7). Our hypothesis, based on the observations from the yeast studies, was that transcription and decay rates are often positively correlated in genes with discordant relationship between expression levels and RNA decay rates across individuals. Indeed, we found a significant increase in positive correlations between transcription and mRNA decay rates for genes with discordant compared to genes with a concordant relationship between expression and decay (P<10−3; Figure 4B; Figure S6) and compared to the distribution of correlations between transcription and mRNA decay rates of all genes in the data set (P<10−16).
Finally, we investigated the genetic basis for inter-individual variation in mRNA decay rates. To do so, we treated the mRNA decay rates as a quantitative trait and mapped genetic loci influencing variation in this trait. We tested for association between individual-specific estimates of mRNA decay rates and genotypes in a cis candidate region of 25 kb centered around the target transcript boundaries. Using this procedure, we found 31 genes with significant RNA decay quantitative trait loci (rdQTLs) at a 15% FDR (Figure 5A). Expanding our mapping procedure to include genome-wide polymorphisms, we found no evidence for significant trans-acting rdQTLs, likely because our experiment is underpowered to detect trans effects (see Methods S1).
Given the observed significant correlation between steady-state gene expression levels and decay rates across individuals, we hypothesized that we might have better power to detect more rdQTLs at a given FDR if we focused on SNPs already identified as steady-state expression QTLs. To do so, we first mapped eQTLs using the mean expression data across time points. We identified 1,257 eQTLs (at 15% FDR; see Methods), most of which were previously observed in these cell lines. Within this set, 195 (16%) of the eQTLs were also significantly (at 15% FDR) associated with variation in mRNA decay rates (Figure 5B, Table S8). In other words, 195 of the steady-state gene expression QTLs are also classified as rdQTLs using our approach; a significant enrichment of decay effects compared to that expected by chance (P<0.001). Using the method of Storey et al. to conservatively estimate the proportion of tests where the null hypothesis is false (while accounting for incomplete power [48]), we estimate that 35% of the most significant eQTL SNPs are also associated with decay rates (Figure S7).
We asked whether SNPs that are identified as rdQTLs are enriched in particular genomic annotations, especially when compared to eQTL SNPs. Since our mapping approach does not allow us to identify with confidence the causal site, we proceeded by considering and comparing the strength of association with decay rates across SNPs in different genomic annotations. Using this approach we found that, in general, the same functional annotations that were previously found to be enriched for eQTLs are also enriched for rdQTLs (e.g., exons, UTRs, and promoter regions; Figure S8A). Yet, while eQTL are generally enriched in 3′ UTRs (Figure S8B), rdQTLs are specifically enriched in predicted miRNA binding sites within 3′ UTRs (Figure 6). This observation is consistent with the hypothesized importance of miRNA-mediated regulation of mRNA decay.
We next examined the relationship between eQTLs and rdQTLs in more detail. We found that in the majority of the joint QTLs (55%), the allele that is associated with lower steady-state expression level is also associated with faster mRNA decay rate, as expected if differences in decay rates drive differences in expression levels across individuals (Figure 5C). However, in the remaining 45% of cases, the allele that is associated with lower gene expression levels is associated with slower mRNA decay rates (Figure 5D). This implies a more complicated regulatory mechanism, which counters the effect of decay at these loci to drive opposite patterns of gene expression across individuals (see Discussion). We thus focused only on the 55% of eQTLs-rdQTL sites with concordant genotypic effects, for which a more intuitive and simple mechanistic explanation is likely. We again used the method of Storey et al. [48] and estimated that as many as 19% (95% CI by bootstrapping: 15%–21%) of eQTLs might be regulated, at least in part, by differences in decay rates. We acknowledge that (as with any comparison and combination of results from genome-wide mapping studies) any factor that affects the power to find associations may result in a biased estimate of the proportion of eQTLs that are also classified as rdQTLs. It is unclear how one could identify and test for all possible relevant factors. In our analysis, we have taken into account the possible effect of overall gene expression levels on eQTL/rdQTL mapping (see Methods), and confirmed that the distributions of the number of SNPs in the proximal window are similar whether one considers sites classified as either eQTLs only or as eQTLs/rdQTLs (Figure S9). On the other hand, we did find a difference in the distribution of minor allele frequency, and the distributions of the number of individuals that are homozygote to the minor allele, between eQTLs and eQTLs/rdQTLs (Figure S9), but this would be conservative with respect to the estimated proportion of eQTLs that are also rdQTLs (namely, the true overlap might be higher than 19%).
Using a similar approach, we have previously found that up to 55% of eQTLs might be explained by variation in DNase sensitivity (these eQTLs were also classified as dsQTLs [32]). We expected that the combination of RNA decay data and DNase sensitivity profiles might explain a larger proportion of inter-individual variation in gene expression levels. To test this using LCLs from the 66 individuals used in both the DNase sensitivity [32] and the current study, we first examined the overlap between SNPs identified as either eQTLs, rdQTLs or dsQTLs. In order to standardize the analyses, we re-mapped eQTLs, rdQTLs, and dsQTLs using only the set of 66 YRI LCLs used in both our study and Degner et al. [32]. We identified 1,147 eQTLs (15% FDR), of which 171 were also classified as rdQTLs (15% FDR) and 168 as dsQTLs (15% FDR; Figure S10). There is a slight enrichment in the overlap of eQTLs classified as both rdQTLs and dsQTLs (33 SNPs; 25 are expected by chance along; P = 0.03). This might reflect variation that affects gene expression levels through coupled transcription and decay processes.
Put together, 26.7% eQTLs are also classified as either rdQTLs and/or dsQTLs. Combining all three annotations (see Methods; Figure S11) we estimated (by using the Storey method [48])that up to 63% of eQTLs could be driven, at least in part, by either decay or chromatin accessibility-related mechanisms. We note that for this comparison we are including both concordant and discordant rdQTLs, since both patterns could be representative of either simpler or complex mechanisms underlying gene expression variation.
We conducted a genome-wide study of inter-individual variation in mRNA decay levels in 70 human LCLs to investigate the extent to which variation in mRNA decay might account for overall gene expression variation. Our observations, both across genes as well as across individuals, lend support to the notion that regulation by decay processes is a significant mechanism by which steady-state transcript levels are modulated.
Consistent with previous studies, we found substantial variation in mRNA decay rates across genes [44], [49], [50]. We caution that the experiments to obtain decay rates involve treatment with an antibiotic (ActD), which is toxic to cells and may therefore be associated with certain artifacts. That said, ActD is a well-established reagent for studies of this type and the conditions we used here closely reflect those of earlier studies of mammalian mRNA decay. One inherent limitation of our study design is the inability to calculate absolute decay rates and thus mRNA half-lives. Instead, we were only able to estimate decay rates relative to the mean cellular mRNA decay rate. Using data collected using commercial microarrays (rather than, for example, RNA sequencing data), this was the only way we were able to normalize the data across time points without making explicit assumptions regarding the distribution of decay rates. Our normalization approach allowed us to maintain the relative order of genome-wide decay rates across genes and individuals. Yet, it also likely resulted in a limited range of the estimated variance of decay rates across genes compared to the true underlying distribution of absolute decay rates. Thus, the results and analyses presented here may underestimate the magnitude of variation in mRNA decay rates across genes.
In many cases, our observations across genes were consistent with the intuitive model whereby faster mRNA decay rates are associated with lower steady-state gene expression levels. Accordingly, we observed lower and higher steady-state gene expression levels for the most rapidly and slowly decaying genes, respectively. Focusing only on these intuitively simple regulatory interactions across QTLs, we estimated that up to 19% of eQTLs might influence gene expression variation through an effect on mRNA decay rates. Incorporating rdQTLs with data on DNase sensitivity QTLs (dsQTLs), we estimated that a combination of variation in RNA decay rates and chromatin accessibility might explain the majority (63%) of eQTL effects. In addition, we find that SNPs within miRNA binding sites show an enrichment for association with variation in decay rates compared to all 3′UTR SNPs, leading to a hypothesis that variation in miRNA binding plays a particularly important role in regulating decay rate variation.
Interestingly, however, we observed many instances of the opposite (discordant) relationship between mRNA decay rates and steady-state gene expression levels. Overall, 59% of genes with a significant correlation between decay rates and expression levels across individuals show a discordant relationship (though only 45% of eQTL/rdQTL pairs). The frequency of this phenomenon seems somewhat unexpected especially given the stronger overall concordant relationship between decay and expression when all genes are considered. It may also cast doubt on the mechanistic explanation we offered for the more intuitive – concordant – relationship between RNA decay and gene expression levels. On the other hand, prevalent discordant decay rates and expression levels across genes have been previously observed in yeast. We speculate that these discordant patterns are the result of complex regulatory circuits, which have evolved to address the need for shorter response time or to stabilize steady-state gene expression levels within the cell. Indeed, the majority of genes with discordant decay and expression patterns are known to be involved in biological processes that require fast response time (Table S3). In a subset of these cases, an auto-regulatory or regulatory feedback circuit has been demonstrated (Table 1). Since many stress and immune response pathways are activated (namely, these genes are highly expressed [53]) in LCLs due to the EBV infection which causes immortalization, we hypothesize that we were able to identify discordant patterns of decay and gene expression at a higher frequency than otherwise expected in resting cells.
Discordant differences in the rates of transcription and mRNA decay could be achieved by a coupling of decay and transcriptional regulatory mechanisms. Dori-Bachash and colleagues suggested that discordant patterns between two closely related yeast species might be due to such coupling whereby the same cis elements may regulate both processes [37]. Supporting these findings, Shalem et al. found that PolII binding in yeast could regulate coordinated mRNA synthesis and degradation processes [38], building on work from Harel-Sharvit et al. that implicated PolII as a factor linking both transcription and mRNA decay to translation in yeast [55]. Additional evidence has pointed to an intrinsic role for the same promoter binding elements promoting both mRNA synthesis in the nucleus and mRNA degradation in the cytoplasm [56], [57]. Our observations also lend support to an explanation based on coupling of the transcription and RNA decay processes.
Such mechanistic coordination implies complex regulatory circuitry, which suggests that decay processes might be playing an important role in maintaining an upper limit of steady-state gene expression, while allowing for rapid transcriptional response - a classical auto-regulatory feedback loop motif [36]. Coupling different regulatory mechanisms to cause such regulatory motifs has been suggested as a way by which cells optimize systems-level functionalities [40]. This is especially important in the context of transcriptional responses to external stimuli or stress. In these situations, coupling of transcription and mRNA decay might be an efficient strategy that allows rapid and precise control of cellular response to external perturbations [40].
Previous studies provided evidence for the important role of mRNA decay in regulating cellular response. For instance, Raghavan et al. found that activation-induced genes in human T-lymphocytes cells, which are enriched for transcriptional regulatory functions, tend to have fast decay rates [52]. Shalem and colleagues evaluated changes in mRNA decay and transcription rates in yeast subjected to either transient or enduring stresses [70]. Yeast subjected to the enduring stress displayed an expected behavior whereby most induced genes were stabilized, while under the transient stress, most induced genes exhibited faster decay rates regardless of their increased steady-state expression levels [70]. Our rdQTL data suggest that variation in regulatory elements that affect mRNA decay rates may play an important role in the individual-specific efficiency of response regulatory circuitry.
We have taken some of the first steps towards characterizing the impact of variation in mRNA decay rates on variation in gene expression levels. Our results indicate that decay processes might play a crucial role in fine-tuned genome-wide regulation of gene expression variation in humans. In particular, we found that a moderate proportion of eQTLs might be due to variation in decay rates, and that negative feedback regulatory circuits involving mRNA decay processes may be common in humans. Further study of the mechanisms underlying variation in mRNA decay rates is needed to increase our understanding of the genetic basis of steady-state gene expression levels and the underlying regulatory circuits.
Cell lines were grown using standard procedures (as recommended by Coriell) by culturing cells in RPMI 1640 (supplemented with 2 mM L-glutamine and 15% fetal bovine serum). Each of the cell lines was treated with Actinomycin D (ActD) to inhibit transcription, with one biological replicate from each cell line. Because ActD terminates active transcript elongation by binding directly to DNA in a reversible manner [12]–[14], [71], [72], it is generally thought to be the most effective transcriptional inhibitor [16], [18], [72]–[74]. ActD treatment was performed by culturing cells at a concentration of 750,000 cell/ml with 7.5 ug/ml of ActD.
Based on the results from a pilot experiment (see Methods S1, Figure S1, Figure S2, Figure S3), we extracted RNA at a total of five timepoints: before the treatment with ActD (0 hours) and after treatment (0.5 hours, 1 hour, 2 hours, and 4 hours). To account for the decrease in total RNA resulting from the treatment and to obtain enough RNA from each timepoint for subsequent microarray hybridization, we increased the number of cells from which we extracted RNA over the timecourse (Figure S1). Total RNA was extracted using an RNeasy Mini Kit (Qiagen) and RNA quality was assessed using an Agilent Bioanalyzer.
We estimated gene expression levels in all samples (350 total samples across all 5 time points and 70 cell lines) by hybridizing RNA to the Illumina HT-12 v4. Expression BeadChip arrays. As RNA yield is expected to change across samples from different time points (due to RNA decay), previous microarray based studies of RNA decay have typically normalized their data using spiked-in samples [3], [8], [21]. The Illumina HT-12 arrays, however, do not include non-human probes that would allow us to use spike-ins. Instead, we hybridized the same quantity of RNA from each time point to the microarrays using standard Illumina hybridization protocols. Subsequently, we normalized the array data using standard approaches across all the arrays [27]–[29], [75], [76].
All low-level microarray analyses were performed in R using the Bioconductor software package lumi [13], [31]–[34], [77]. Specifically, we performed a log2 variance stabilizing transformation and robust spline normalization (RSN). Following normalization, we removed probes with intensities indistinguishable from background noise in either the 0 and/or 4 hour time points on the array (as measured by the negative controls present on each array). In addition, we mapped the Illumina 50 bp probe sequences using BWA v.0.4.6 [36], [78] and retained only probes that mapped uniquely with 100% identity to an exon within an annotated gene from the Ensembl database (2009-12-31 version). Following filtering based on detection and probe mapping (see Supplemental Materials), data from 23,065 probes corresponding to 16,823 genes were used for all further analyses. For gene-based analyses, we considered the mean expression across the set of probes corresponding to a single gene as the expression level of that gene. For all genotype analyses, SNPs located within probes could bias probe hybridization and downstream measures of steady-state gene expression across individuals. For the 3,327 probes overlapping one or more SNPs, we aimed to remove the effect of SNPs on probe hybridization by regressing steady-state expression levels on the genotype of the SNP located within the probe. In cases where this regression was significant (P<0.05), we used the residual of the regression as the steady-state expression measurement [28], [29], [79].
After all normalization and filtering steps, genes whose transcripts decayed at an “average” rate appeared to be expressed at a constant level through the timecourse measurements (Figure S2). For ease of visualization, the expression levels across time points in all decay profiles plotted throughout this manuscript have been standardized by the total number of cells from which RNA was extracted (Figure S1).
Because mRNA decay has been shown to exhibit properties of first-order decay [11], [39], [80], [81], we estimated gene-specific RNA decay rates in each cell line by using a regression equation of the form (a linear transform of the first-order exponential decay model):(1)where y(t) is the mRNA abundance at time t, B0 is the mRNA abundance at the untreated time point (time point ‘0’), k is a gene-specific decay rate constant, and variance ε∼N(0,σ2). For subsequent analyses, we used the gene-specific decay rate constant k as an estimate of a decay rate. Under these conditions, genes with decay rates close or equivalent to the mean cellular decay rate are represented by k = 0. To identify genes that decay significantly faster or significantly slower than the mean mRNA decay rate in LCLs, we identified genes for which k significantly differed from zero (mean decay rate). We fit gene-wise decay rates for each cell line and identified genes for which least 80% of individuals had estimated values of k that differed significantly from 0 (P<0.1) in the same direction (either faster or slower decay than the mean decay rate).
To rank genes by their combined gene expression and decay values, we examined the genome-wide distributions. For example, genes with discordant patterns are those with high (or low) expression levels and whose mRNA decays rapidly (or slowly). To classify such patterns, we independently identified genes within the top 5% and 10% tails of the decay rate and steady-state gene expression distributions and then considered the overlaps across the two data sets (Figure S5B). We identified 9 and 49 genes at the top 5% and 10%, respectively, of both the gene expression and decay rate distributions.
To determine the effect of gene length and 3′UTR length on mRNA decay rates, gene lengths and 3′UTR lengths were calculated using information extracted from the Ensembl gene database (2009-12-31 version). [41]–[43], [82]. Total gene length was defined as the distance between the upstream most TSS and the downstream most transcription end site (inclusive of both exons and introns). Total 3′UTR length was calculated as the number of bases annotated as being within a 3′UTR in any isoform of the given gene.
In order to create a comprehensive set of microRNA (miRNA) binding site predictions, we downloaded the miRNA binding predictions from three databases: microRNA.org, PicTar, and targetScan [3], [19], [20], [22]–[25], [44]–[47]. By parsing the predictions for all miRNAs in these three databases, we obtained a combined set of miRNA predictions that were present in one, two, or all three databases. Because each of these databases uses different sets of annotations and identifiers, we applied a series of conversion and filtering steps for each database (see Methods S1 for details). We used the AREScore algorithm (http://arescore.dkfz.de/arescore.pl) [26], [44], [49], [50] to calculate an AREScore as a proxy for the number of AU-rich elements present in 3′UTRs. The program was run with default parameters on RefSeq defined 3′UTR regions for all genes in our dataset [1], [3], [41], [51], [52], [83].
To identify significantly over- or under-represented motifs in either fast or slow decaying genes, we used the FIRE algorithm (https://tavazoielab.c2b2.columbia.edu/FIRE/) [30], [41]. We tested for motif enrichment in promoter regions and full gene bodies of both fast and slow decaying genes, using default FIRE parameters. In all tests, we compared against a background set of all genes that were present in our study.
We used GeneTrail (http://genetrail.bioinfo.uni-sb.de) [15], [26], [84] to test for enrichments of functional annotations among different classes of genes: (a) genes consistently decaying faster or slower than the mean cellular decay rate, (b) genes at the top 10% of both the gene expression and decay rate genome-wide distributions, and (c) genes showing either concordant or discordant relationships between decay rates and gene expression levels. In all tests, we used a background set of all genes that were present in our study and detected as expressed in either the zero or four hour timepoints. The tests were performed using all GO categories and KEGG pathways. We calculated p-values using a hyper-geometric distribution and report false discovery rates for each p-value.
To investigate the contribution of variation in decay rates to overall variation in steady-state gene expression levels across individuals, we identified genes whose expression levels and decay rates were significantly correlated. Specifically, for each gene, if yi denotes the steady-state expression level (defined here as the mean of the expression levels across all time points in order to increase statistical independence from the estimated decay rates) for individual i and ri denotes the corresponding decay rate estimate, we fit a linear model of the form:(2)where the coefficient, β, measures the strength of the association between decay rate and steady-state gene expression levels. In order to identify genes where the coefficient, β, represents a significant association, we repeated the analyses with 3 sets of permuted decay rates, recorded the significance of β from each permutation, and used these permuted p-values as an empirical null distribution. We estimated the FDR by comparing the true distribution of p-values of β to this null distribution.
PolII ChIP-seq data on six YRI LCLs (GM18505, GM18522, GM19141, GM19193, GM19204, and GM19238) were collected within the context of another study within the lab. ChIP-seq libraries were prepared as described previously [11], [54], [85], using the non PolII antibody H-224 (Santa Cruz Biotechnology, sc-9001x). In addition, raw PolII ChIP-seq reads from a seventh YRI LCL, GM19099, was obtained from a previously published study [11], [33], [47] and analyzed in a similar fashion to the PolII ChIP-seq data generated in-house.
Raw PolII ChIP-seq reads were mapped back to human genome (hg18) using BWA v.0.4.6 [11], [54], [78] and reads from multiple lanes from the same individual were combined into a single mapped file. For each individual, we used Samtools [6], [86] to isolate reads in genic regions (as defined in the Genomic Annotations section above) and promoter regions (defined as 1 kb upstream and 1 kb downstream of the transcription start site). For genic regions, read counts were normalized by the total length of the genic region to be able to compare across genes with varying length. For individual-specific measures of PolII occupancy for each gene, read counts were normalized by the total number of mapped reads per individual.
For all QTL mapping analyses, we used close to full genotype information for each of the 70 YRI individuals, achieved by combining available datasets and imputing missing genotypes with the BimBam software [58], [87], [88] as described previously [32], [59], [60]. Briefly, we built a reference panel consisting of the largest set of all 210 YRI HapMap individuals and gathered genotypes for any SNP or short insertion/deletion (indel) called in either HapMap (Release 28; October 2010, [1], [35], [59]) or 1000 Genomes (May 2011 interim phase 1 release, [1], [52], [60]) datasets. Missing genotypes in the individuals in this study were imputed using this reference panel, resulting in a total of approximately 15.8 million variants genome-wide.
All associations between genotypes and either decay rates or gene expression were examined using a linear regression model in which each phenotype was regressed against genotype. For all analyses, we only tested association under the assumption that SNPs affected the resulting phenotype in an additive manner (i.e. heterozygote phenotypic mean equals the average of the two homozygote means). For each gene, we tested for association of the phenotype with the genotypes of SNPs and indels within a cis-candidate region of 25 kb around the gene (25 kb upstream of the TSS and 25 kb downstream of the TES). We chose this definition of a cis-candidate region to map variation in mRNA decay rates in an unbiased manner by including SNPs outside of transcript regions. Indeed, recent reports have indicated that elements in intergenic promoter elements [56] and RNA binding proteins binding intronic regions [89] could regulate mRNA decay mechanisms. To evaluate genotypic effects on decay variation for a given gene, we tested associations with SNPs or indels with a minimum allele frequency greater than 10%, using the following model:(3)where ri is defined as in model (2) and gij corresponds to the genotype of individual i at variant j, coded as 0, 1, or 2 copies of the minor allele. In this model, the coefficient γ indicates the strength of association between the mRNA decay rate of the gene and genotypes at variant j. To estimate the false discovery rate, we permuted phenotypes three times, re-performed the linear regressions, and recorded the minimum p-value (across SNPs/indels) for each gene for each permutation. These sets of minimum p-values were used as our empirical null distribution. We estimated the FDR by comparing the true distribution of the minimum p-values to this null distribution, as previously described. Previous studies mapping cis-associations have found that statistical power to detect associations can be dramatically increased by accounting for unmeasured confounders within quantitative measure of the phenotype [3], [12], [13], [31], [32], [41], [90], [91]. When considering decay as the phenotype, we did so by performing principal components analysis (PCA) on the (70 by 70) correlation matrix of decay rate estimates. We found the strongest rdQTL signal (largest number of findings at a fixed FDR) when 13 principal components (PCs) were regressed out.
When considering steady-state gene expression as the phenotype, we performed all analyses on mean expression levels across all time points per individual in order to reduce the variance of expression measurements and increase the statistical independence between the eQTL estimates and the estimates of decay rates. We quantile normalized these measurements and performed PCA to account for unmeasured confounders. For the eQTL analyses, we again found the most QTL signal when 13 PCs were regressed out. The eQTL analyses were performed by testing for association between mean expression levels and SNPs or indels with a minimum allele frequency greater than 10%, using the following model:(4)where yi is defined as in model (2) and gij corresponds to the genotype of individual i at variant j. In this model, the coefficient γ indicates the strength of association between the mean steady-state expression level of the gene and genotypes at variant j. FDR calculations were performed as described above.
To assess whether the enrichment of significant mRNA decay effects among eQTL SNPs could occur by random chance, we performed a permutation based significance test. Specifically, we evaluated the effect of genotype on mRNA decay variation using the most significant cis-eQTL SNP for all genes in our dataset (regardless of the genome-wide significance of the SNP). Then, we randomly chose 1,257 SNPs from this full set (representing the number of genome wide significant eQTLs identified) and calculated the number that showed significant association with mRNA decay variation among this set. We also ensured that the distribution of gene expression levels associated with the randomly sampled SNPs matched the distribution of expression levels for genes with significant eQTLs. By repeating this 1,000 times, we were able to arrive at a permutation-based expectation for the enrichment of significant mRNA decay effects among eQTL SNPs.
In order to look at overlaps between the set of identified rdQTLs and previously identified dsQTLs, we focused on the set of 66 YRI LCLs that were used in both studies. Using mean gene expression measures from this study, we re-mapped eQTLs as described above in this set of 66 LCLs and identified 1,147 steady-state eQTLs (15% FDR). Using these 1,147 eQTL SNPs, we tested for association between each SNP and DNaseI sensitivity as described previously [32] and between each SNP and RNA decay rates (as described above). To obtain an estimate of the total proportion of eQTLs we might be able to account for by either RNA decay variation or variation in DNaseI sensitivity, we assessed, for each SNP, the evidence for association with either data type. We then chose the minimum p-value for the association with decay rates or DNaseI sensitivity and compared the resulting distribution to the following analytical transformation:We then applied the Storey et al. qvalue approach to account for incomplete power [48] to this transformed distribution of p-values.
All raw data and tables of all rdQTLs are available under GEO accession number GSE37451.
|
10.1371/journal.pntd.0007540 | The use of chicken and insect infection models to assess the virulence of African Salmonella Typhimurium ST313 | Over recent decades, Salmonella infection research has predominantly relied on murine infection models. However, in many cases the infection phenotypes of Salmonella pathovars in mice do not recapitulate human disease. For example, Salmonella Typhimurium ST313 is associated with enhanced invasive infection of immunocompromised people in Africa, but infection of mice and other animal models with ST313 have not consistently reproduced this invasive phenotype. The introduction of alternative infection models could help to improve the quality and reproducibility of pathogenesis research by facilitating larger-scale experiments. To investigate the virulence of S. Typhimurium ST313 in comparison with ST19, a combination of avian and insect disease models were used. We performed experimental infections in five lines of inbred and one line of outbred chickens, as well as in the alternative chick embryo and Galleria mellonella wax moth larvae models. This extensive set of experiments identified broadly similar patterns of disease caused by the African and global pathovariants of Salmonella Typhimurium in the chicken, the chicken embryo and insect models. A comprehensive analysis of all the chicken infection experiments revealed that the African ST313 isolate D23580 had a subtle phenotype of reduced levels of organ colonisation in inbred chickens, relative to ST19 strain 4/74. ST313 isolate D23580 also caused reduced mortality in chicken embryos and insect larvae, when compared with ST19 4/74. We conclude that these three infection models do not reproduce the characteristics of the systemic disease caused by S. Typhimurium ST313 in humans.
| Salmonella Typhimurium ST313 is associated with systemic infection in human populations in sub-Saharan Africa, and contrasts with the related pathovariant ST19 which causes gastrointestinal disease worldwide. Although the systemic pathology associated with ST313 infection in humans has been comprehensively documented in clinical and epidemiological studies, such pathology has been inconsistently reproduced in animal models of infection. Animal models that reliably recapitulate ST313 infection in humans are needed to study the biological mechanisms underpinning the systemic disease caused by ST313. In this study we performed extensive infection experiments, using several defined and alternative animal infection models to look for robust phenotypes that differentiate infection by S. Typhimurium ST313 from ST19. Large sample sizes and multivariate statistical analysis of infection data for inbred chicken lines allowed us to detect small but consistent differences between the strains. Overall, S. Typhimurium ST313 was associated with a reduced infection burden and pathology relative to ST19. This subtle phenotype may reflect the limitation of animal models to accurately represent infection by pathogens that have adapted to specific host phenotypes, for example, immunodeficiency in humans. Our study demonstrates the challenge of using animal models to differentiate closely-related bacterial pathovariants, and shows that inter-pathovar differences detected in animal models of infection often do not reflect clinical differences in humans at the level of disease mechanism.
| Salmonella Typhimurium sequence type (ST) 313 is a novel pathovariant circulating in sub-Saharan Africa that causes invasive non-typhoidal Salmonella (iNTS) disease [1]. The most common S. Typhimurium ST present throughout the rest of the world is ST19, which generally causes gastroenteritis. Recently, we reported the use of a comparative transcriptomic approach to identify genes that were differentially expressed between ST313 and ST19 in a variety of environmental conditions and during macrophage infection [2].
Previously, we used the chicken model to study the infection biology of ST313 and ST19 S. Typhimurium strains. The data showed that S. Typhimurium ST313 strains caused invasive disease in Lohmann brown chickens and caused a more invasive phenotype than ST19 strains at early stages of infection [3]. However, the research was limited by small sample sizes and the exclusive use of outbred chicken lines.
In Salmonella research, the majority of infection studies have been carried out with inbred mouse lines such as BALB/c and C57BL/6 [4,5]. Other animal models such as pigs, rhesus macaques and calves have also been used, but results with S. Typhimurium ST313 have been inconsistent [6–8]. In some models, like the chicken, mouse, bovine and non-human primate models, ST313 strains were capable of invading the intestinal mucosa and eliciting intestinal inflammation [3,6]. In other models, such as in streptomycin-treated mice and bovine ligated ileal loops, ST313 strains had a slightly reduced ability to cause intestinal inflammation in comparison to ST19 [9]. A recent review article suggested that these differences may be attributed to variations between the animal models that were involved and the different ST19 strains that were used as the comparator [10].
Fewer studies have been done in chickens, and outbred lines (which represent a genetically heterologous population) such as Lohmann Brown laying chickens, have most commonly been used. However, several inbred chicken lines (representing populations of genetically-identical animals) have been used for Salmonella infection research [11–13]. Different inbred lines vary in their resistance to systemic salmonellosis caused by S. Typhimurium ST19, largely due to the salmonellosis resistance locus SAL1 which is associated with an increased pro-inflammatory response [14]. Inbred lines previously used in Salmonella studies include the White Leghorn-derived inbred lines W and 61, which show a “Salmonella-resistant” phenotype; and lines 72, Cb4 and 15, which show a “Salmonella-susceptible” phenotype.
This Salmonella resistance classification was based on infection studies performed in 1-day-old chicks, infected with S. Typhimurium ST19 strain F98, and in older birds with the avian-adapted serovar Salmonella Gallinarum [11]. The resistant phenotype manifests as reduced mortality, tolerance to a higher infectious dose and reduced systemic pathology. However, the same birds showed a weak resistance phenotype after oral infection with S. Typhimurium ST19 at 2 weeks old, whilst retaining a strong resistant phenotype to S. Gallinarum [11].
Reduced biological bird-to-bird variance has been observed with inbred lines of chickens compared with outbred populations, improving the quality of the experimental outcomes [15]. Other less commonly-used infection models such as the chicken embryo [16] and the wax moth larvae (G. mellonella) models [17] are particularly experimentally-tractable and reliable, and therefore show promise for use in Salmonella pathogenesis research [18].
We investigated the relative virulence of S. Typhimurium ST313 isolate D23580 and ST19 isolate 4/74 in outbred and inbred chicken lines, chick embryos and G. mellonella larvae.
All the procedures were performed in accordance with UK legislation governing experimental animals under project licence PPL40/3652 and the experimental protocols were approved by the University of Liverpool ethical review process.
S. Typhimurium strain 4/74, a representative strain of nontyphoidal Salmonella sequence type 19 (ST19), D23580, a representative strain of nontyphoidal Salmonella sequence type 313 (ST313), and derivates from these strains (S1 and S2 Tables) were used in this study. Strain D23580 was isolated from an HIV− child from Malawi with bloodstream infection and use of this strain has been approved by the Malawian College of Medicine (COMREC ethics no. P.08/14/1614). All isolates were maintained as frozen stocks at −80°C. Prior to infection, isolates were grown in LB (Lennox) broth at 37°C in a shaking incubator at 220 rpm for 18 h.
One-day-old Lohmann Brown outbred chickens of mixed sex were obtained from a commercial hatchery. 48 birds were randomly divided into two groups, one was inoculated with S. Typhimurium strain 4/74 and the other group was inoculated with S. Typhimurium strain D23580.
One-day-old specific pathogen-free White Leghorn-derived inbred lines W and 61 chickens (“Salmonella-resistant”), and lines 72, Cb4 and 15 chickens (“Salmonella-susceptible”) were supplied by the National Avian Research Facility (Roslin Institute, Edinburgh). Between 28 and 40 birds of mixed sex were used per line. For each line, birds were randomly separated in to two groups as above.
Both inbred and outbred chicks were housed in the University of Liverpool high-biosecurity poultry unit. Birds were housed in accommodation meeting the UK legislation requirements and were given ad libitum access to water and a vegetable protein-based pelleted diet (Special Diet Services, Witham, Essex, UK). Chicks were housed on wood shavings in floor pens at a temperature of 30°C. All housing and environmental conditions were identical between groups. The birds were tagged with metal wing bands to allow identification of individuals. All animals were checked a minimum of twice daily to ensure their health and welfare. Prior to experimental infection, all birds were confirmed as Salmonella-free by cloacal swabs, which were streaked onto selective XLT4 (Lab M) and grown for 24 h at 37°C.
For both sets of birds (outbred and inbred lines), individual groups were inoculated by oral gavage with 108 CFU of either S. Typhimurium 4/74 or D23580 in 0.2 ml LB (Lennox) broth (overnight cultures) when the birds were two weeks old. For inbred birds, one further group of birds was mock-infected with sterile Phosphate Buffered Saline (PBS) as a control for histopathological studies.
At 3, 7 and 12 days post-challenge, 4–7 birds from each group were culled by cervical dislocation and post-mortem examinations were carried out. Assessment of Salmonella colonisation levels and histopathological studies were carried out.
Samples from spleen, liver and caecal content were removed aseptically from each bird, weighed and diluted 1:5 (wt/vol) in sterile PBS. Tissues were then homogenized in a Colworth 80 microstomacher (A.J. Seward & Co. Ltd). Serial 10-fold dilutions were made of each sample in PBS, and according to the method of Miles and Misra [19], triplicate 20 μl spots were plated onto Harlequin™ Salmonella ABC Medium (Lab M). The plates were incubated at 37°C for 24 h, and colonies were enumerated to give CFU/g of sample.
Liver, spleen and caecal tissue samples from inoculated birds were taken at each time point and fixed in 4% paraformaldehyde in PBS. Tissues were embedded in paraffin wax, cut and stained with haematoxylin and eosin, and then examined by a board certified veterinary pathologist at the University of Liverpool. All sections were examined blind and were scored individually for each tissue from each animal based on the scoring system detailed in Parsons et al, 2013 [3], to make our data comparable to the data previously published.
Fertilised White Leghorn chicken eggs were obtained from Lees Lane Poultry, Wirral, or Tom Barron, Preston, UK. Fertilised eggs were incubated at 37°C with a relative humidity of 50–60% (Octagon 40 egg turner incubator, Brinsea, Weston Super Mare, UK) until E11 (embryonic day 11) or E14 and all animal work followed UK regulations. On the 10th day of embryonic development the eggs were candled to detect and discard those that were not fertilized or had dead embryos. On day E11 a hole was aseptically made in the egg shell and 102 CFU bacterial suspensions from an overnight culture diluted in PBS were inoculated intra-allantoically using a sterile 1 ml syringe with a 27G needle attached. The hole was then covered with tape and the eggs returned to the incubator. Injections with PBS and S. Typhimurium 4/74 and D23580 ΔrpoE::frt mutants were used as negative controls. At the end of the incubation period, embryos were separated from the eggs and washed with sterile PBS several times. Embryos were dissected and their livers were extracted, washed several times with PBS and placed on ice. The number of bacteria in the liver was determined using the same protocol used in the chicken model. Yolk samples were taken prior to extraction of the embryos.
For the competition assay, 11-day-old embryos were inoculated via the allantoic cavity with a mix of equal numbers of S. Typhimurium D23580 TcR (JH3950, S1 Table) and 4/74 KmR (JH4284, S1 Table) in 100 μl of PBS. The combined bacterial inoculum was ~ 4 × 102 CFU. The numbers of each strain in the inoculum and in the liver and yolk of inoculated embryos were calculated by plating onto LB (Lennox) containing 25 μg/ml tetracycline or 50 μg/ml kanamycin. The competitive index was calculated as the ratio of D23580 and 4/74 at 16 hours post-inoculation, divided by the same ratio in the inoculum.
The S. Typhimurium 4/74 ΔSL1483::aph mutant (JH4284) (S1 and S2 Tables) was constructed with the sole purpose of having a wild-type strain with a selection marker, in this case resistance to kanamycin. The resistance marker was introduced by λ Red recombination in a pseudogene (SL1483) that does not code for a functional protein. Briefly, the kanamycin resistance cassette (aph) was amplified from plasmid pKD4 using the oligonucleotides ins_STnc230_rev and del_SL1483_for. The resulting PCR product was transformed into 4/74 carrying the recombineering plasmid pSIM5-tet [20] by electroporation. The ΔSL1483::aph was transduced into 4/74 WT using the high-frequency-transducing bacteriophage P22 HT 105/1 int-201 [21].
For D23580, a tetracycline resistant version of D23580 (D23580 TcR), that carries the tetRA genes between the pseudogene STMMW_41451 and the gene STMMW_41461 (between coordinates 4441510 and 4441511 in D23580) was constructed by λ Red recombination as follows: the tetRA resistance cassette was amplified by PCR from the Tn10 transposon with primers NW_202 and NW_203. The resulting PCR fragment was electroporated into D23580 wild-type carrying the λ Red recombination plasmid pKD46-Gm, as described previously [22], and tetracycline resistant recombinants were selected on LB agar plates supplemented with tetracycline hydrochloride (25 μg/ ml). The tetRA cassette was subsequently transduced into D23580 wild-type, using phage P22 HT 105/1 int-201 as previously described [23]. Both mutants were confirmed by sequencing with external primers.
Groups of 10 eggs/strain were used. Eggs were inoculated at embryonic day 11 as above and incubated at 37°C for 24 hours. Embryo viability was recorded by candling and the mortality rate was calculated as the mean percentage of embryonic deaths at 24 hours post-infection.
Genetically homogeneous final-instar greater wax moth G. mellonella larvae were purchased from BioSystems Technology UK. Groups of 10 larvae were contained in a petri dish lined with a circle of Whatman filter paper #1 and injected with 10 μl of bacterial suspension using a Hamilton 701LT syringe fitted with a 27G mm needle. All injections were administered via the last left pro-leg into the haemocoel. As controls, a group of larvae were injected with PBS and another group was not injected. The plates containing the larvae were incubated at 37°C for 24 hours at which point mortality rate was recorded. Larvae that did not show any movement in response to touch were considered dead.
For the chicken model, a Mann-Whitney test was used to determine the statistical significance of differences between the strains in individual timepoints and tissue types, as the distribution was not normal.
To detect broader trends in the colonisation level and histopathological score data, we used R version 3.5.2 to construct seven linear models to predict tissue burden or histopathology as a function of strain, tissue type, chicken line and timepoint. For the outbred chicken data, the model contained 141 CFU/g tissue (log10) counts from three tissue types (caeca, liver and spleen) and 3 timepoints post infection (1, 3 and 5 days) (raw data available in S3 Table). For the inbred chicken data, the models contained 534 CFU/g tissue (log10) counts and 521 histopathological scores respectively, as well as both tissue count and histopathology score datasets split into the Salmonella-susceptible (72, Cb4 and 15I) and Salmonella-resistant (W and 61) chicken lines, The models included a total of 9 variables representing strain (n = 1), tissue (n = 2), chicken line (n = 4) and timepoint (n = 2). We used the car [24] package to determine significance of model parameters using type III sums of squares [25]. Linear model and Anova summary tables are shown in S4–S6 and S9–S11 Tables. For the chick embryo and G. mellonella larvae models, a t-test was used to determine the statistical significance of strain differences. The raw data from all experiments undertaken in this study can be found in S1 Dataset.
To study the ability of isolates D23580 and 4/74 to colonise chickens, we inoculated 14-day-old outbred Lohmann Brown Layers with 108 bacteria by oral gavage. We sampled the caeca, liver and spleen of infected animals at 1, 3 and 5 days post-infection (dpi). The experiment was comparable to Parsons et al. 2013 [3], but used four times more animals per group. Fig 1 shows the colonisation levels at 3 dpi, indicating that the colonisation levels for the two strains were not significantly different when the experiment involved a large sample size (P > 0.05). Colonisation levels at 1 and 5 dpi were also assessed and showed no significant differences between strains (S1 Fig).
The abundance of bacteria in the liver, spleen and caeca of birds infected with strain D23580 were similar to those published by Parsons et al., 2013 [3] at all timepoints. However, the abundance of bacteria found in birds infected with 4/74 were somewhat different here, with lower levels found in caeca and higher levels found in spleen at 3 dpi compared to the previous study [3]. Linear modelling of the total dataset from the infection of outbred chickens with strains D23580 and 4/74 did not find bacterial strain to be a significant variable that affected the burden of infection (S3 Table).
Animal models have been used in biomedical research for centuries. Beginning in the Pasteur era, infection studies involved outbred animals (genetically heterogeneous). However, large animal-to-animal variations prompted researchers to gradually move to inbred lines, as it became evident that genetic factors determined many animal characteristics. For example, genetically-homogeneous inbred mice yield more reproducible infections than outbred mice [26].
Since the results from our experimental infection of outbred birds (S1 Fig and S3 Table) did not show a differential phenotype for ST313 and ST19 strains [3], we hypothesized that this finding could reflect the bird-to-bird variation associated with outbred animals. Therefore, we compared the virulence of S. Typhimurium D23580 and 4/74 in inbred chicken lines.
We first compared the two isolates in the inbred chicken lines 72, 15 and Cb4 that had been reported to be “Salmonella-susceptible”. In all cases, the same infection procedure was used: oral-gavage of 14-day-old birds with 108 cells and quantification of bacterial burden in the spleen, liver and caeca at 3, 7 and 12 dpi. In all three “Salmonella-susceptible” lines, some statistically significant differences (p-value < 0.05) were seen in pairwise comparisons between strains D23580 and 4/74 at 3, 7 and 12 dpi. Specifically, higher levels of colonisation of caeca, liver and spleen were seen in birds inoculated with 4/74 compared to birds inoculated with D23580. Specifically more 4/74 than D23580 bacteria were isolated from spleen at 3 dpi (P = 0.047) and liver at 12 dpi (P = 0.007) for Line 72; from liver at 7 dpi (P = 0.018) for Line 15; and from the caeca of Line Cb4 at 12 dpi (P = 0.010, Fig 2).
To detect broader trends in infectivity between the strains, we used a linear model to predict log10CFU/g tissue as a function of time point, chicken line, strain and tissue. In the susceptible chicken lines, strain 4/74 was associated with a small, but consistent increase in tissue burden relative to D23580 (average increase of 0.9 log10 CFU/g tissue, P < 0.001, S5 Table).
Next, we compared the isolates in “Salmonella-resistant” inbred chickens from Line W and Line 61. In these lines, differences between S. Typhimurium D23580 and 4/74 were more subtle, but showed a similar trend to the “Salmonella-susceptible” line data, with 4/74 causing a slightly higher tissue burden than D23580 (Fig 3). In general, the data show only small differences between colonisation levels of the 14-day-old “Salmonella-susceptible” and “Salmonella-resistant” chicken lines, consistent with reports that the SAL1 resistance trait was only evident when one day-old chicks were challenged with S. Typhimurium [11].
Linear modelling of the “Salmonella-resistant” inbred chicken infection data showed that 4/74 was associated with an average increased tissue burden of 0.6 log10 CFU/g tissue, compared with infection by D23580 (P < 0.001, S6 Table).
When tissue burden data from 4/74 and D23580 in all five inbred lines were analysed with a linear model, strain 4/74 was associated with an overall average increased burden of 0.8 log10 CFU/g tissue relative to strain D23580 (P < 0.001, S4 Table). The model showed that the chicken line, tissue and timepoint were also significant variables that affected the tissue burden (S4 Table). Our findings show that statistical analyses involving multivariate comparison of the two bacterial strains across different chicken lines, tissue types and timepoints enabled the detection of subtle inter-strain variations in virulence.
Severity of infection may not necessarily correlate with bacterial infection burden, for example if one strain possesses enhanced virulence factors, a lower bacterial burden could cause the same severity of disease. Therefore, to qualitatively compare the infection biology of D23580 and 4/74, we assessed histopathological changes associated with the infection by these two isolates in caeca, liver and spleen of infected birds at 3, 7 and 12 dpi. Similar tissue changes were observed in the five inbred chicken lines; there was no difference in the nature of the histopathological changes caused by bacterial strain D23580 or 4/74 in the different chicken lines. However, tissue changes did vary in severity with 4/74 causing more pronounced pathology than D23580 (Fig 4). Changes in the caecum include variable degrees of heterophilic inflammatory infiltrates within the lamina propia, extending to intra-epithelial populations. Lesions in the liver varied from small clusters of often perivascular macrophages or heterophils, to multifocally scattered larger accumulations, sometimes exhibiting the morphology of small dense aggregates of macrophages, reminiscent of a granuloma. Meanwhile, splenic lesions consisted of increased numbers of inflammatory cells, particularly macrophages, occupying the red pulp and in more severe cases, also including the white pulp.
The data indicated that both isolates D23580 and 4/74 caused similar tissue changes during chicken infection, though 4/74 was associated with consistently higher histopathological scores than D23580 (as illustrated by the dark blue-shift in the heatmaps shown in Fig 4, and data in S7 and S8 Tables). The observed pathological changes were consistent with the bacterial colonisation levels present in the tissues of individual infected birds, with a higher pathology score being associated with a greater bacterial load. Statistical modelling showed that strain 4/74 was associated with an average increase of 0.22 in the tissue histopathology scoring system relative to D23580 (P < 0.001, S9–S11 Tables), supporting the trend shown in Fig 4.
Overall, the colonisation levels and histopathological analysis of infected inbred and outbred chickens in this study differed from previous findings with outbred birds [3]. The key difference between the two studies is that here we infected a total of 113 birds per strain representing one outbred line and five inbred lines, whereas Parsons et al. (2013) [3] used 20 outbred birds per strain. Here we did not observe the statistically-significant difference that had been reported previously in outbred birds [3], where ST313 strain D23580 was associated with increased virulence relative to ST19 strain 4/74. Further, when genetically homogeneous inbred chickens were used, ST313 strain D23580 showed a reduced virulence phenotype relative to ST19 strain 4/74, in terms of both tissue burden and pathology.
Our data highlight the challenge of achieving reproducibility in animal experiments, and the importance of sample size in infection research. We conclude that there are limitations to the use of animal models when investigating differences in virulence between closely-related strains. Additionally, we note that the small but consistent inter-strain differences detected here were clear only when utilising linear modelling analysis, which allowed global comparison of strains across different tissue types and timepoints. Simple pairwise comparisons between strains in specific tissue types and time points rarely detected differences between the strains (Fig 2).
To build on our findings, we studied 4/74 and D23580 in other infection models. For this, we used alternative infection models that did not require a UK Home Office licence or ethical approval. The first of these was the chick embryo infection model [16]. The embryos used in this model are classified as non-protected under the UK Animals (Scientific Procedures) Act 1986 since the experiments are carried out at embryonic day 11 and the embryos were euthanised before embryonic day 14; only animals at embryonic age 14 or older are covered by the act [26]. The chick embryo infection model involves the injection of bacteria into the allantoic cavity of the egg, in this model only virulent bacteria are able to migrate from this site to reach other compartments of the egg, including the embryo and yolk. Meanwhile, attenuated bacteria remain confined to the allantoic cavity [16]. Virulence is assessed by determining the colonisation levels in the embryo liver after a specific incubation period. We used an incubation period of 16 hours post-inoculation. Longer incubation times led to mortality, rapid tissue destruction, and necrosis which prevented meaningful measurements (S1 Data set). At 16 hours post-inoculation, similar levels of colonisation by D23580 and 4/74 were found in the embryos, around 105 CFU/g of liver (Fig 5).
To confirm the embryo liver-colonisation level data, we performed a competitive index assay by simultaneously inoculating embryonated eggs with equal quantities of both isolates. At 16 hours post-inoculation, in both liver and yolk, the D23580 isolate exhibited CI values < 1.0, indicating that fitness of D23580 was reduced compared to 4/74 at a statistically significant level (Fig 6). The CI of D23580 vs 4/74 was 0.38 (P = 0.001) in egg yolk and 0.34 (P = 0.002) in liver.
Next, we assessed the ability of the bacteria to kill chick embryos, and found that isolate 4/74 caused higher mortality than D23580 at 24 hours post inoculation (P = 0.0018; Fig 7A).
As Salmonella infection of chick embryos is not an established model, we tested the ability of a number of mutants that lacked important Salmonella virulence factors to infect the embryos. Our data showed that mutants deficient in LPS and flagella biosynthesis were attenuated in the chick embryo model, but SPI-1, SPI-2 and the PhoPQ regulon were not required for infection (S12 Table and S2 Fig); further investigation is needed to confirm this finding.
Lastly, we compared the levels of mortality induced by D23580 and 4/74 in the wax moth G. mellonella larvae model [17]. The lymph of G. mellonella larvae contains haemocytes (phagocytic cells that share many properties with mammalian phagocytes), representing a good model for the innate immunity response [27]. The G. mellonella larvae were infected by inoculation of bacteria via injection into the haemocoel, and percentage mortality caused by each of the two isolates was assessed at 24 hours post-inoculation. The 4/74 isolate caused a higher mortality rate than D23580 (P<0.0001; Fig 7B), consistent with our finding that 4/74 was more fit than D23580 in the chick embryo model.
Our work highlights the importance of sample size, choice of host and sampling protocol in animal infection research. Previous findings that S. Typhimurium ST313 strain D23580 had a more invasive phenotype than ST19 strain 4/74 in the chicken model [3] were not reproduced in either outbred or inbred chicken lines in this study.
In their review, Casadevall and Fang [28] suggested that sample size in infection research is a crucial variable, and that the use of too few animals can obscure significant differences that would be evident in larger populations. Although increasing the number of animals reduces the probability of false negative or false positive results, experimentation with vertebrate animals is limited by the Principles of the 3Rs that advocate the reduction of the number of animals used in experiments [26,29]. An unintended consequence of the Principles of the 3R’s is that different laboratories rarely repeat experiments to verify key virulence differences that have been reported to distinguish bacterial isolates, or that differentiate wild-type from mutant bacteria. We note that experimental chicken infections have been reported to have their own limitations, even when genetically homogeneous inbred lines are used [14]. Variations in the levels of colonisation of infected chickens could reflect stochasticity in oral infections, natural differences in the microbiota, maternally-derived immunity or simple variation in size and fitness of the experimental animals.
The use of alternative infection models such as the chicken embryo or the wax moth G. mellonella larvae model allow large sample sizes without compromising ethical standards, whilst maintaining genetic homogeneity. Through the use of multiple alternative infection models and large sample sizes, we detected consistent significant differences between ST313 and ST19 strains of S. Typhimurium, suggesting alternative infection models represent promising alternatives to vertebrate infection models where differences between strains in question may be nuanced.
Overall, our data suggest that the qualitative nature of disease caused by S. Typhimurium ST313 and ST19 isolates (D23580 and 4/74) in avian and insect models is broadly similar. However, ST313 shows reduced levels of colonisation of inbred chickens, relative to ST19. The use of alternative infection models showed that S. Typhimurium ST313 exhibited reduced virulence in chick embryos and G. mellonella larvae. It remains likely that African S. Typhimurium ST313 has a host-specific adaptation that is responsible for systemic infection in humans but was not apparent in the animal models of infection tested here.
Human tissue culture models of infection have failed to identify consistent differences in infection biology between ST313 and ST19 strains [10]. Furthermore, the variety of animal models that have been used to study African Salmonella and to compare ST313 with ST19 did not find a consistent ST313-associated phenotype. The animal infection data presented here do confirm that ST313 does not share the host-restricted lifestyle of S. Typhi, S. Gallinarum or S. Abortusequi, but the chicken and insect-based models only revealed subtle phenotypic differences between ST313 from ST19.
One model, the streptomycin-treated mouse, has been used effectively to demonstrate the importance of pseudogenisation of the sseI effector gene of S. Typhimurium ST313 for dendritic cell-mediated dissemination [30]. One explanation for the efficacy of this murine model is that the effect of streptomycin treatment in some ways mimics the immune-compromised human population to which S. Typhimurium ST313 is niche-adapted. It is possible that the clinical usage of broad-spectrum antibiotics amongst the immunocompromised S. Typhimurium ST313-susceptible African populations is a significant factor. The host immune-deficiency, which occurs in African populations suffering from HIV, malaria and malnutrition, or an antibiotic-induced gut microbiota disturbance, may be critical for studying the infection biology of ST313 strains. Immune-deficient infection models are in development, including HIV-infected macrophages [31], and malaria-infected mice [32], and could prove critical for understanding the genetic mechanism of adaption of African S. Typhimurium ST313 to an invasive lifestyle.
|
10.1371/journal.pgen.1003611 | Histone Methyltransferase DOT1L Drives Recovery of Gene Expression after a Genotoxic Attack | UV-induced DNA damage causes repression of RNA synthesis. Following the removal of DNA lesions, transcription recovery operates through a process that is not understood yet. Here we show that knocking-out of the histone methyltransferase DOT1L in mouse embryonic fibroblasts (MEFDOT1L) leads to a UV hypersensitivity coupled to a deficient recovery of transcription initiation after UV irradiation. However, DOT1L is not implicated in the removal of the UV-induced DNA damage by the nucleotide excision repair pathway. Using FRAP and ChIP experiments we established that DOT1L promotes the formation of the pre-initiation complex on the promoters of UV-repressed genes and the appearance of transcriptionally active chromatin marks. Treatment with Trichostatin A, relaxing chromatin, recovers both transcription initiation and UV-survival. Our data suggest that DOT1L secures an open chromatin structure in order to reactivate RNA Pol II transcription initiation after a genotoxic attack.
| Through the deformation of the genomic DNA structure, UV-induced DNA lesions have repressive effect on various nuclear processes including replication and transcription. As a matter of fact, the removal of these lesions is a priority for the cell and takes place at the expense of fundamental cellular processes that are paused to circumvent the risks of mutations that may lead to cancer. The molecular mechanism underlying transcription inhibition and recovery is not clearly understood and appears more complicated than anticipated. Here we analyzed the process of transcription recovery after UV-irradiation and found that it depends on DOT1L, a histone methyltransferase that promotes the reformation of the transcription machinery at the promoters of UV-repressed genes. Our discovery shows that transcription recovery after a genotoxic attack is an active process under the control of chromatin remodelling enzymes.
| Short-wave UV light is a significant source of mutagenic and cytotoxic DNA damage. UV irradiation induces two major types of DNA lesions; the cis-syn cyclobutane-pyrimidine dimers (CPD) and the pyrimidine (6-4) pyrimidone photoproducts (6-4PP) [1]. Through the deformation of the DNA structure, these lesions have repressive effect on various nuclear processes including replication and transcription. As a matter of fact, the removal of these lesions is a priority for the cell and takes place at the expense of fundamental cellular processes that are paused to circumvent the risks of mutations that may lead to cancer. The molecular mechanism underlying transcription inhibition and recovery is not understood yet but it includes proteins such as CSB, a member of the SWI2/SNF2 family of chromatin remodeling proteins, which promote transcription re-initiation at the promoters of UV-repressed genes [2], [3].
UV lesions are removed from DNA by the nucleotide excision repair (NER) mechanism through two sub-pathways. The general global genome repair (GG-NER) removes DNA damage from the entire genome, while the transcription-coupled repair (TC-NER) corrects lesions located on actively transcribed genes [4]. In TC-NER, an elongating RNA polymerase II (RNA Pol II) stalled by a lesion triggers efficient repair of the cytotoxic damage that blocks transcription, while lesion elsewhere in the genome are detected by the XPC/hHR23B complex for GG-NER [5]. Then, both sub-pathways funnel into a common process involving XPA, RPA, TFIIH, XPG and XPF-ERCC1 to excise damaged oligonucleotides from DNA.
Post-translational histone modifications modulate promoter activity. Histone acetylation, phosphorylation, ubiquitination, and methylation dictate the transcriptional fate of any given locus [6]. Inactive heterochromatin is associated with high levels of methylation at H3K9, H3K27, and H4K20 residues and low levels of acetylation, while actively transcribed euchromatin shows a high level of acetylation of H4K16 and H4K20 and methylation of H3K4, H3K36, and H3K79 residues [7], [8].
The dot1 gene (disruptor of telomeric silencing-1), also called kmt4 (lysine methyltransferase-4), encodes a protein that exclusively methylates lysine 79 of histone H3 (H3K79) [9], [10], [11]. Unlike most modified histone residues that are located within the N-terminal tail, H3K79 is found within the globular core of the histone octamer [12]. The Dot1 protein is the only histone lysine methyltransferase that does not contain the conserved SET domain but exhibits a methyltransferase fold that is responsible for its activity [13], [14]. In mammals, several studies have shown that DOT1L (the Dot1 homolog) exists in a complex that trimethylates H3K79 and that contains several myeloid/lymphoid or mix-lineage leukemia fusion partners, such as MLLT1, 2, 3 or 10 [15]. More recently, DOT1L was shown to be involved in cell cycle progression, the control of the differentiation of pluripotent cells [16] and leukemogenesis [17].
In addition to these roles in fundamental cellular processes, several lines of evidence suggest that DOT1L plays an important role in genomic stability. DOT1L has been reported to favor the recruitment of double strand breaks sensor 53BP1 to DNA lesions [18]. Dot1 is known to be required for the activation of the RAD9/RAD53 checkpoint function by UV and γ-radiation [19], [20], [21]. Dot1 also plays a role in the yeast cellular response to UV damage but its specific function in this mechanism is unclear [22].
Here, we provide evidence that the mammalian DOT1L protein is required to re-initiate transcription after UV irradiation. Knocking-out of DOT1L results in high sensitivity to UV irradiation in mouse embryonic fibroblasts (MEF), but preserves an accurate repair of (6-4)PP lesions. Instead, MEFDOT1L are unable to recover transcription of constitutively expressed genes after UV irradiation. Using fluorescence recovery after photobleaching (FRAP), we have shown that DOT1L regulates the recruitment of RNA Pol II to chromatin after UV irradiation. Applying chromatin immunoprecipitation assay, we additionally revealed that DOT1L triggers the formation of the pre-initiation complex (PIC) to the promoters of UV-repressed genes, as well as the appearance of active transcriptional marks on histones. Altogether, our results highlight a new role for DOT1L in the maintenance of an open chromatin structure in order to reactivate the formation of the transcription machinery on the promoter of constitutive genes after a genotoxic attack.
To investigate the role of DOT1L in the repair of UV-induced DNA damage, we used knocked-down MEFDOT1L cells carrying a homozygous gene trap insertion in Dot1l [23]. Together with an absence of Dot1l protein expression, the mono- and tri-methylation of H3K79 were strongly reduced in MEFDOT1L (Figure 1A). In a UV-C survival assay, MEFDOT1L cells were more sensitive to irradiation as compared to control MEFWT but less than MEFXPG, knocked-out for the NER factor XPG and deficient both for GG- and TC-NER [24] (Figure 1B). However, MEFDOT1L showed similar UV-C sensitivity than the MEFCSB cells, knocked-out for the CSB protein involved only in TC-NER. Note that tri-methylation of H3K79 was similar in MEFWT, MEFXPG and MEFCSB (Supplemental Figure S1A). Knocking-down of DOT1L expression in HeLa cells using siRNA (Supplemental Figure S1B) recapitulated the mild UV-sensitivity of the MEF cells compared to the high UV-sensitivity induced by the knocked-down of the NER factor XPA (Supplemental Figures S1B–C). These data indicate that DOT1L deficiency induces UV-sensitivity in mammalian cells.
We further investigated whether DOT1L affected UV-C survival by sustaining the repair of UV-induced DNA damage. We performed unscheduled DNA synthesis assay (UDS), which is mainly a measure of the GG-NER efficiency [25]. The UDS of MEFDOT1L was identical to that of the MEFWT (Figure 2A). We also used an assay based on immunofluorescence coupled to quantification of DNA lesions directly in cell nucleus to measure the removal of the (6-4)PPs, the main UV-induced DNA lesions (see experimental procedures). MEFXPG cells showed low removal of (6-4)PP along the time course of repair, compared to MEFWT or MEFCSB in which lesions were rapidly removed (Figures 2B). When measured in MEFDOT1L cells, the removal rate of (6-4)PP lesions was identical to that of MEFWT (Figures 2B), which implied that the absence of DOT1L does not impair GG-NER.
To determine whether MEFDOT1L were able to perform TC-NER, we performed two sets of experiments. First we measured cell survival following incubation with Ecteinascidin 743 (et743), an anti-tumor drug that shows cytotoxicity effect only on human TC-NER proficient cells [26]. In our experimental conditions, knocking-down of the TC-NER factor CSB in MEF cells resulted in et743 resistance, compared to MEFWT cells (Figure 2C). In contrast, MEFDOT1L cells showed sensitivity to et743 to a level equivalent to that of MEFWT cells (Figure 2C). Next, we performed a host cell reactivation assay [27]. We employed a dual GFP/RFP readout with a UV-damaged plasmid (600 J/m2, 3 Kb) containing a GFP-tagged reporter and an undamaged plasmid containing an RFP-tagged reporter transiently transfected into MEF cells. Recovery of GFP-reporter expression, 12 hours post-transfection, was efficient in both MEFWT (Figure 3, panels a–f) and MEFDOT1L (Figure 3, panels g–l) cells but not in the TC-NER deficient MEFCSB cells (Figure 3, m–r). Overall, these results suggest that the absence of DOT1L induces sensitivity to UV irradiation that is not the consequence of a defect in GG- or TC-NER.
Next we aimed to determine the global RNA synthesis of MEF cells after irradiation using the recovery of RNA synthesis (RRS) assay [25]. Cells were UV-irradiated with 10 or 20 J/m2 and incubated with radioactive [3H]uridine during a 30 minutes pulse performed before and 24 hours after irradiation. Mock-treated MEFWT and MEFDOT1L cells showed similar levels of RNA synthesis, visualized by equivalent number of black dots in their nuclei (around 100 dots/nucleus, Figure 4A and 4B, compare panels a and c). In contrast, we observed a marked deficiency in RNA synthesis in MEFDOT1L cells, 24 hours after UV-C treatment, as compared to MEFWT (Figure 4A and 4B, compare panels b and d). We estimated the residual transcription activity in the MEFDOT1L cells to 30% of that of the mock-treated cells, 24 hours after irradiation with 20 J/m2.
To unveil the molecular mechanisms that led to the inhibition of transcription in MEFDOT1L cells after UV irradiation, we examined live-cell protein mobility of RNA Pol II by fluorescence recovery after photobleaching (FRAP). In brief, a small region in the middle of the nucleus was bleached and the subsequent fluorescence recovery was measured in time (Figure 5A). For that purpose, the largest RNA Pol II subunit RPB1 was fused with GFP and expressed either in MEFWT or MEFDOT1L. In these conditions, we found an equivalent mobility of RNA Pol II in mock-treated and UV-irradiated (UV-C, 16 J/m2) MEFWT (Figure 5B and Supplemental Data S1) (T-test = 3.4E-2). In marked contrast, FRAP analysis of UV-damaged MEFDOT1L cells revealed a significant increase in fluorescence recovery when compared to mock-treated cells (Figure 5C and Supplemental Data S1) (T-test = 1.4E-5), indicating that a fraction of RNA Pol II became mobile in the absence of DOT1L, after UV irradiation.
We next wondered whether DOT1L was required to mobilize RNA pol II either during the initiation or elongation steps of transcription. For this purpose, we examined the step of transition from initiation to elongation by the RNA Pol II in vivo on an endogenous mouse gene. Briefly, we reversibly blocked gene transcription by incubating cells with the P-TEFb inhibitor DRB (5,6-dichloro-1-β-D-ribobenzimidazole), which inhibited the transition from initiation to elongation but did not block elongation of ongoing mRNA transcripts [28] (Figure 6A). Following the removal of DRB, RNA Pol II was released from promoter-proximal regions and the level of newly synthesized pre-mRNA had been measured owing to the presence of exons and introns. We measured the transition from initiation to elongation on the Utrophin gene that possesses a very short Exon1 (170 bp). We estimated that an average of one transcription blocking lesion (CPD or (6-4)PP) was created per 5 kb of DNA at 20 J/m2 [29], indicating that less than 5% of cells harbor a UV-lesion in this exon following a dose of 15 J/m2 of UV-C. Therefore, any significant inhibition of transcription initiation cannot be explained by the blockage of RNA Pol II in front of a lesion in Exon 1.
We treated MEF cells for 3 hours with DRB and extracted RNA at 10 minutes intervals after the removal of the drug (Figure 6B). Next, we performed RT-PCR using primers spanning Exon1-Intron1 junctions to detect newly synthesized pre-mRNA of Utrophin gene. In the absence of genotoxic stress, MEFWT and MEFDOT1L were both able to recover transcription of the Exon1 region within 10 to 20 minutes after DRB removal (Figure 6C), indicating that the transcriptional initiation rate in MEFWT and MEFDOT1L was identical, in absence of a genotoxic attack. Then, we irradiated MEF cells with UV-C (15 J/m2) after the DRB treatment (Figure 6B). In these experimental conditions, MEFWT were able to recover transcription of the Exon1 within 60 minutes after removal of DRB and UV-treatment, while MEFDOT1L showed no recovery of Exon1 transcription even after 80 minutes (Figure 6D).
Since DOT1L was shown to be involved in chromatin remodeling, we next tested whether chromatin relaxation could overcome the absence of DOT1L. For that purpose, we treated MEFDOT1L with Trichostatin A (TSA, 20 nM), a class I histone deacetylase (HDAC) inhibitor that relaxed chromatin (Figure 6B). Following TSA treatment, we observed a recovery of Exon1 pre-mRNA expression in MEFDOT1L, which peaked between 30 to 40 minutes after UV irradiation (Figure 6E) and paralleled the transcription of Exon1 in MEFWT cells. Together with this recovery, pre-treatment of MEFDOT1L with TSA (10 nM) induced a potent recovery of UV survival (Figure 6F) suggesting that transcription inhibition in MEFDOT1L was indeed responsible for the UV-sensitivity of these cells. Pre-treatment of the TC-NER deficient MEFCSB with TSA did not modify their UV-sensitivity (Figure 6F). Altogether, these data suggested that DOT1L allowed RNA Pol II transcription re-initiation after UV irradiation.
The above data prompted us to perform a detailed analysis of the PIC formation on the promoter of UV-repressed genes throughout the time, after irradiation. We studied the promoter of the several housekeeping genes such as DHFR (dihydrofolate reductase), B2M (beta-2-microglobulin) or KLF7 (Kruppel-like factor 7) that we used as models for assembly/disassembly of the PIC after UV-irradiation. Using chromatin immunoprecipitation (ChIP), we observed a slight decrease in both RNA Pol II (Figures 7B and Supplemental Figure S2) and basal transcription factors (Figure 7C) occupancy at these promoters in UV-irradiated MEFs, 2 hours post-UV irradiation. The transcription machinery started to re-assemble on the promoter between 6 and 10 hours after UV irradiation and the steady state level of mRNA did not vary significantly with time in the wild-type situation (Figure 7A). In contrast, the basal transcription machinery did not re-form on the UV-repressed gene promoter throughout the entire time course in MEFDOT1L cells (Figures 7B, 7C and Supplemental Figure S2). In line with these observations, the steady state level of mRNA decreased progressively after UV irradiation in MEFDOT1L cells (Figure 7A). This observation is in agreement with results obtained in Figure 4 in which we measured global transcription.
We then analyzed chromatin modifications associated with active or inactive chromatin in the promoter region of the DHFR gene. We observed an acetylated histone H4 pattern that followed that of the RNA Pol II, with a first phase of decrease 2 hours post-UV, and a phase of increase from 6 to 24 hours (Figure 7D). In contrast, no phase of recovery was observed in MEFDOT1L cells (Figure 7D). Di-methylation of H3K9 residue, a mark of transcription silencing, was stable in MEFWT throughout the time course after UV-treatment, while it increased just after irradiation in MEFDOT1L (Figure 7E). Finally, di-methylation of H3K79, performed by DOT1L, was observed transiently in the promoter region of DHFR and peaked 6 hours after UV irradiation (Figure 7F), when RNA Pol II and TFIIB started to come back to the DHFR promoter in MEFWT cells. Altogether these data suggested that DOT1L was required to drive recovery of PIC formation at the promoters of UV-repressed genes.
Transcription inhibition and the subsequent recovery that operate after a genotoxic attack are thought to limit the risks that lesions represent for the genome. The molecular mechanisms that are responsible for the turn-off/turn-on of transcription after DNA damage are not well understood. Here we show that the methyltransferase DOT1L is required for the re-initiation of transcription by triggering re-formation of the transcription machinery at the promoter of UV-repressed genes. Our data supports the hypothesis that transcription recovery after a genotoxic attack is an active process involving specific actors insuring not only the repair of the DNA but also the recovery of fundamental cellular processes such as transcription.
Our study demonstrated that disruption of DOT1L caused hypersensitivity to UV irradiation in mammalian cells. There are several potential mechanisms that could explain the increased sensitivity to UV irradiation conferred by DOT1L depletion. If it was directly involved in DNA repair, its absence may results in increased levels of DNA damage, leading to cellular apoptosis. A function for yeast DOT1 in GG-NER has been described recently [30]. However, we did not find any DNA repair defect associated with the absence of DOT1L in mammalian cells. Indeed, MEF cells depleted of DOT1L exhibited normal UDS level, an assay that mainly measured GG-NER. Furthermore, these cells repaired (6-4)PP, the best UV-induced NER substrate, at the same rate than wild-type cells. To show that MEFDOT1L cells were also proficient in TC-NER, we used two strategies. First we made use of the sensitivity of cells to the drug et743, a natural marine product isolated from the Caribbean see squirt. Antiproliferative activity of et743 was shown to be dependent on active TC-NER [26]. In our experimental conditions, we observed that MEFs knockdown for CSB, one of the two TC-NER specific factors, exhibited a higher resistance to et743 treatment than MEFWT, confirming previous observation obtained with patient cells [26]. Using et743 on MEFDOT1L cells, we observed that these cells were as sensitive to treatment as MEFWT. In a second set of experiments, we measured the capacity of MEFDOT1L to drive the recovery of a reporter expression construct previously exposed to UV-irradiation. In mouse-derived cells, the reactivation depends both on efficient GG- and TC-NER activities as illustrated by the absence of transcription recovery in MEFCSB cells observed both in our study and in other reports [31]. However, MEFDOT1L showed a full recovery of reporter expression. Finally, the treatment of MEFDOT1L with TSA induced a recovery of UV-survival while treatment of MEFCSB did not. We concluded from these cellular observations summarized in Table 1 that it was unlikely that UV-irradiation sensitivity in DOT1L-deficient mammalian cells was due to a GG-NER or TC-NER defect.
Alternatively, DOT1L may serve to reactivate global mRNA synthesis after UV irradiation. Transcriptional arrest has been shown to lead to a highly cytotoxic cellular response to stress [5]. This response has multiple causes and is likely not only the result of DNA lesions that block RNA Pol II in elongation. Previous studies have challenged the relationship between efficient repair of a lesion in the transcribed strand of active genes and the restoration of DNA damage inhibited transcription. For instance, cells carrying mutations in CSB were unable to recover lesion-inhibited transcription while they efficiently repaired acetylaminofluorene lesions in transcriptionally active genes [32]. In addition, CSB was shown to accumulate at the promoters of UV-repressed genes, where it stimulated the recovery of transcription independently of the presence of lesions [3]. This finding led to the hypothesis that removal of transcription blocking lesions was insufficient to restore transcription after DNA damage and that in addition, chromatin changes in the promoters of UV-repressed genes may be required. We performed RNA-sequencing on MEFWT and MEFDOT1L after UV-irradiation (unpublished Data). However, RNA-sequencing measures steady-state level of individual mRNA but does not provide information on the de novo RNA activity. To measured the global level of newly synthetized RNA we used the RRS assay and we observed a general inhibition of de novo RNA synthesis in cells depleted of DOT1L; 30% of residual transcription activity was detected 24 hours post irradiation (20 J/m2). This inhibition was confirmed at the single gene level since DHFR showed a similar level of inhibition.
Early steps of mRNA expression include formation of PIC, transcription initiation and escape of RNA Pol II from the promoter to the elongation step. Using an assay that measured the transcription of the newly synthetized first exon of Utrophin gene model in vivo, we demonstrated that the transition from initiation to elongation was deficient in the absence of DOT1L after UV. We further analysed both PIC occupancy and the chromatin modifications on the promoter of a UV-repressed gene after irradiation and have shown that this promoter was temporally depleted of basal transcription factors in the first hours after irradiation in wild-type cells. Six to ten hours after UV irradiation, when DNA repair took place, the PIC occupancy recovered completely in the wild-type situation. In the absence of DOT1L, RNA Pol II and TFIIB did not get back to the promoter of our UV-repressed gene model, and heterochromatin marks such as methylation of the H3K9 residue appeared. FRAP experiment using GFP-tagged RPB1 subunit of RNA Pol II showed that a fraction of RNA Pol II became mobile after UV-irradiation in the absence of DOT1L, indicating that dissociation of RNA Pol II from chromatin of UV-repressed genes after irradiation in the absence of DOT1L is a broad phenomenon.
Altogether, these data suggest that DOT1L favors an opened chromatin structure around the promoter of UV-repressed genes to allow transcription re-initiation. In line with this hypothesis, H3K79me2, the mark of DOT1L activity, was accumulating transiently on the promoter in wild-type cells after UV irradiation. In addition, the absence of DOT1L was circumvented by the class I HDAC inhibitor TSA that relaxes chromatin. TSA restored both transcription initiation and UV-survival of MEFDOT1L cells, creating thus a link between these two events.
Among the sites of histone methylation, H3K79 is unique as it is not located within the H3 N-terminal tail domain but in the core region. Specifically, this methylation occurs on the surface of the nucleosome and may serve as a platform to recruit additional chromatin modifiers and DNA damage response factors [12]. On the other hand, regions of chromatin where transcription is repressed are depleted of H3K79 methylation, indicating that silencing of chromatin probably requires hypomethylation of H3K79 [11], [33]. The mechanism that links euchromatin to H3K79 methylation is not fully understood but it is believed that in addition to recruiting chromatin modifiers, this histone mark plays an important role in confining the Sir proteins to heterochromatic regions [11], [34]. In yeast, Sir3 binds to nucleosomes containing deacetylated histone H4K16 and promotes spreading of heterochromatin along the chromatin [35]. Based on these observations and our data, we propose that RNA Pol II re-accumulation at promoters after UV irradiation depends on the chromatin changes orchestrated by DOT1L, including the emergence of active chromatin transcription marks. In the absence of DOT1L, facultative heterochromatin marks such as H3K9me2 appear and RNA Pol II does not get back to the promoters. Through the recruitment of chromatin modifiers and subsequent histone modifications, DOT1L serves to limit the spreading of heterochromatin to UV-repressed genes immediately after irradiation and to allow re-association of the basal transcription machinery on the promoters of these genes to re-activate their transcription. Finally, one can wonder how specific is the requirement of DOT1L for the transcription recovery? We tested MYST2 and G9a, an H4 acetyl-transferase and an H3K4 methyltransferase respectively, but knocking down these chromatin remodelers did not induce any increase in UV sensitivity in Hela cells (Data not shown). Although this result shows that all transcriptional chromatin remodelers are not required to reactive transcription after DNA repair we believe that our observation will lead to the identification of additional factors relevant for the regulation and timing of this crucial step that appears more complicated than anticipated.
MEF cells were cultured at 37°C in the presence of 5% CO2 in Dulbecco's modified medium supplemented with 10% FCS.
Ten thousand cells were plated on a 3.5 cm Petri dish, cultured overnight and UV-irradiated with UV-C light (254 nm) at various doses (0.5 J/m2/s). After 4 days, cells were dried and stained by crystal violet, then lysed and quantified by spectrometry at 570 nm wavelength. When indicated, MEF cells were incubated with TSA (10 nM final concentration) for 3 hours before UV irradiation and during the 4 days of the post-irradiation period.
Five thousand MEF cells were plated in 96 well plates (OptiPlates-96, Perkin Elmer). Twenty-four hours later, cells were UV-irradiated with UV-C lamp (10 J/m2) and recovered in fresh medium for the indicated period of time at 37°C, 5% CO2. Immuno-labeling of (6-4)PP was performed using mouse 64M-2 antibodies. DNA was denatured with 2 M HCl for 30 minutes at RT and blocked in 10% BSA in PBS for 15 minutes prior to labeling. (6-4)PP lesions were quantified using an IN Cell Analyser 1000 imaging system (GE Healthcare) and the percentage of (6-4)PP removal was determined (100% represents the amount of lesions determined just after UV irradiation).
UDS was determined by counting the number of grains on at least 50 non-S-phase cells in autoradiographic preparations of cultures incubated for 3 h after UV-irradiation in medium containing 3H-thymidine (3H-TdR, Amersham, specific activity 25 Ci/mmol) [25]. In RRS, mock-treated or UV-irradiated cells (10 or 20 J/m2) were pulse labeled with 5 µCi/ml of [3H]uridine (PerkinElmer Life and Analytical Sciences, Boston MA 02118 USA) for 30 minutes, 24 h post-irradiation. Cells were fixed and auto-radiographied.
Cell lines stably expressing GFP-RPB1α-amaR were generated by transfection of MEFWT or MEFDOT1L with 1 ug of pAT7h1α-amaR [36] using FuGENE6 (Roche Diagnostics, Mannheim, Germany). One day after transfection, cells expressing GFP-RPB1α-amaR were selected by overnight incubation with 20 ug/ml of alpha-amanitin. Three days prior to microscopy experiments, cells were seeded onto 24 mm diameter coverslips. Imaging and FRAP were performed on a Zeiss LSM 710 meta confocal laser scanning microscope (Zeiss, Oberkochen, FRG).
FRAP analysis was performed at high time resolution as previously described [37]. Briefly, a strip spanning the nucleus was photo-bleached for 20 ms at 100% laser intensity (laser current set at 6.1 Å). We monitored the recovery of fluorescence in the strip every 20 msec for 20 sec at 0.5% of laser intensity. Twenty independent measurements were performed and the average values were used for every mobility curve. Mobility curves show relative fluorescence (fluorescence post-bleach divided by fluorescence pre-bleach) plotted against time. Error bars included in all the plotted FRAP data represent the SEM. Whenever two distinct FRAP curves were not easily dissociable, the statistical significance of their difference was checked by using Student's t-test (two-sample, two-tailed) within an appropriate time window: right after the photobleaching when evaluating mobility differences or after complete recovery when immobile fractions were being evaluated.
We grew cells overnight on 60 mm plates to 70–80% confluency and treated them with 100 µM of 5,6-Dichlorobenzimidazole 1-β-D-ribofuranoside (DRB) (Sigma) in culture medium for 3 hours. The cells were washed twice with PBS and incubated in fresh medium for various periods of time and RNA was extracted. Trichostatin A (TSA) (Sigma) was used at a concentration of 20 nM and was added 12 hours before treatment with DRB and maintained during the time of the experiment. When indicated, cells were UV-irradiated (15 J/m2) after the DRB treatment.
cDNA synthesis was performed by using hexamere and AMV reverse transcriptase (Sigma; St. Louis, MO). Real-time quantitative PCR was done with the FastStart DNA Master SYBR Green kit and the Lightcycler apparatus (Roche Diagnostic; Basel, Switzerland). Primer sequences are available upon request.
Cells were crosslinked with a 1% formaldehyde solution for 10 minutes at RT. Crosslinking was stopped by addition of glycine at 125 mM final concentration. Samples were sonicated to generate 500 bp DNA fragments. For immunoprecipitations, 100 µg of chromatin extract was pre-cleared for 2 hours with 50 µl of protein G-sepharose before addition of the indicated antibodies. Then, 2 µg of antibody was added to the reactions and incubated over night at 4°C in the presence of 50 µl of protein A/G beads. After serial washings, the immunocomplexes were eluted twice for 10 minutes at 65°C and crosslinking was reversed by adjusting to 200 mM NaCl and incubating 5 hours at 65°C. Further proteinase-K digestion was performed for 2 hours at 42°C. DNA was purified using Quiagen columns (QIAquick PCR Purification Kit). Immunoprecipitated DNA was quantified by real-time PCR. Primer sequences are available upon request.
The pEGFP-reporter construct was purchased from Clontech. The pEGFP-reporter vector was UV-irradiated (254 nm, 600 J/m2) at a concentration of 1 µg/ml in 10 mM Tris-HCl (pH 8.0) and 1 mM EDTA. MEF cells were transfected with 3 µg of pEGFP-reporter and 1 µg of unirradiated pRED-reporter in a six-well plate at a confluence of 95% using X-tremeGENE 9 DNA Transfection Reagent (Roche). A 3/1 ratio was used to ensure that every cell expressing RFP protein expresses GFP in non-irradiated condition. After 12 hours of incubation, GFP and RFP were detected by reverse microscopy.
Primary antibodies (the final dilutions are indicated in parentheses) used were anti-6–4PP (Cosmobio; 64M-2, dilution 1/500), anti-TFIIB C-18 (Santa Cruz, sc-225), anti-H3K79me1 (Abcam, ab2886) (1/1.000), anti-H3K79me3 (Abcam, ab2621) (1/1.000), anti Histone H3 (Abcam, ab1791), anti-GFP (Clinisciences, TP-401) and anti-DOT1L (Novus Biologicals, NB100-40845). Anti-Acetylated H4 is a mouse monoclonal antibodies produced at the IGBMC.
When required, the trapezoid rule was used to estimate the area under the curve (AUC). Student's t-test was used to assess whether the mean AUC (graph curves) or mean values from triplicate (bar graphs) were statistically significant. A P-value of 0.05 or less was considered as significant.
|
10.1371/journal.pbio.1001478 | Impairment of TrkB-PSD-95 Signaling in Angelman Syndrome | Angelman syndrome (AS) is a neurodevelopment disorder characterized by severe cognitive impairment and a high rate of autism. AS is caused by disrupted neuronal expression of the maternally inherited Ube3A ubiquitin protein ligase, required for the proteasomal degradation of proteins implicated in synaptic plasticity, such as the activity-regulated cytoskeletal-associated protein (Arc/Arg3.1). Mice deficient in maternal Ube3A express elevated levels of Arc in response to synaptic activity, which coincides with severely impaired long-term potentiation (LTP) in the hippocampus and deficits in learning behaviors. In this study, we sought to test whether elevated levels of Arc interfere with brain-derived neurotrophic factor (BDNF) TrkB receptor signaling, which is known to be essential for both the induction and maintenance of LTP. We report that TrkB signaling in the AS mouse is defective, and show that reduction of Arc expression to control levels rescues the signaling deficits. Moreover, the association of the postsynaptic density protein PSD-95 with TrkB is critical for intact BDNF signaling, and elevated levels of Arc were found to impede PSD-95/TrkB association. In Ube3A deficient mice, the BDNF-induced recruitment of PSD-95, as well as PLCγ and Grb2-associated binder 1 (Gab1) with TrkB receptors was attenuated, resulting in reduced activation of PLCγ-α-calcium/calmodulin-dependent protein kinase II (CaMKII) and PI3K-Akt, but leaving the extracellular signal-regulated kinase (Erk) pathway intact. A bridged cyclic peptide (CN2097), shown by nuclear magnetic resonance (NMR) studies to uniquely bind the PDZ1 domain of PSD-95 with high affinity, decreased the interaction of Arc with PSD-95 to restore BDNF-induced TrkB/PSD-95 complex formation, signaling, and facilitate the induction of LTP in AS mice. We propose that the failure of TrkB receptor signaling at synapses in AS is directly linked to elevated levels of Arc associated with PSD-95 and PSD-95 PDZ-ligands may represent a promising approach to reverse cognitive dysfunction.
| Angelman syndrome (AS) is a debilitating neurological disorder caused by a dysfunctional Ube3A gene. Most children with AS exhibit developmental delay, movement disorders, speech impairment, and often autistic features. The Ube3A enzyme normally regulates the degradation of the synaptic protein Arc, and in its absence the resulting elevated levels of Arc weaken synaptic contacts, making it difficult to generate long-term potentiation (LTP) and to process and store memory. In this study, we show that increased levels of Arc disrupt brain-derived neurotrophic factor (BDNF) signaling through the TrkB receptor (which is important for both the induction and maintenance of LTP). We find that the association of the postsynaptic density protein PSD-95 with TrkB is critical for intact BDNF signaling, and that the high levels of Arc in AS interfere with BDNF-induced recruitment of postsynaptic density protein-95 (PSD-95) and other effectors to TrkB. By disrupting the interaction between Arc and PSD-95 with the novel cyclic peptidomimetic compound CN2097, we were able to restore BDNF signaling and improve the induction of LTP in a mouse model of AS. We propose that the disruption of TrkB receptor signaling at synapses contributes to the cognitive dysfunction that occurs in Angelman syndrome.
| Angelman syndrome (AS) is a severe cognitive disorder caused by loss of expression of the maternally inherited allele of the Ube3A ubiquitin ligase gene [1],[2]. As a result of imprinting, the paternal Ube3A gene is silenced, such that the maternal allele is exclusively active [3],[4]. Prominent clinical characteristics include seizures, ataxia, and mental retardation [5]. A mouse model null for maternal Ube3a [6] showed impairment in long-term potentiation (LTP) and learning [6]. Biochemically, the mice exhibited dysregulation of α-calcium/calmodulin-dependent protein kinase II (CaMKII) activity [7],[8], required for certain forms of learning [9].
Ube3A ubiquitinates and degrades the immediate-early gene Arc (activity-regulated cytoskeletal-associated protein) [10], whose expression is required for LTP consolidation [11],[12] and experience-dependent plasticity [13]–[15]. Arc promotes AMPA receptor (AMPAR) internalization [16] to reduce AMPAR-mediated synaptic transmission [17], and mediates AMPAR clearance at weaker synapses [18]. Arc has been reported to associate with postsynaptic density protein-95 (PSD-95) [19], the prototypical PDZ (PSD-95/Discs large/zona occludens-1) postsynaptic protein [20],[21], known to play a key role in the endocytosis of synaptic AMPARs [22]–[24] and to regulate AMPAR incorporation at synapses [25]–[28]. The PDZ domains of PSD-95 bind the cytoplasmic tails of select NMDA and Kainate receptor subunits [29],[30] to assemble cell-signaling scaffolds [31],[32]. To investigate the function of PSD-95, we synthesized a high affinity PSD-95 PDZ-domain peptidomimetic ligand, CN2097. The design of CN2097 (R7-CC-YK[KTE(β-Ala)]V), incorporates a lactam ring and a β-alanine linker that form unique contacts outside the canonical PDZ binding pocket [33],[34].
In the present study, we sought to test if the deficit in LTP-induction in AS mice might be the result of defective brain-derived neurotrophic factor (BDNF) signaling. BDNF binding to the TrkB receptor has been shown to promote the induction and maintenance of LTP [35]–[39], and BDNF or TrkB deficient mice exhibit a marked reduction in LTP [40]–[42]. We report that PSD-95 association with TrkB is critical for intact BDNF signaling. In AS hippocampal slices, the BDNF-induced association of PSD-95 with TrkB was reduced compared to wild type (WT), resulting in attenuated PLCγ (CaMKII and CREB) and PI3K (Akt-mTOR) signaling, whereas MAPK (extracellular signal-regulated kinase [Erk]) signaling was intact. In AS mice the elevated association of Arc with PSD-95 is shown to interfere with the recruitment of PSD-95 to TrkB. Treatment of AS hippocampal slices with CN2097, the PSD-95 PDZ-domain peptidomimetic ligand, increases the association of PSD-95 with TrkB to restore PLCγ and PI3K-Akt-mTOR signaling in AS mice, and to facilitate LTP. Together, these data suggest drugs based on enhancing Trk-PSD-95 interactions, such as CN2097, may provide a novel approach for the treatment of AS and autism spectrum disorders.
PSD-95 has been shown to regulate synaptic strength and is proposed play a key role in LTP [25],[26],[43]–[45]. Previously, we developed a bridged cyclic peptide (CN2097; Figure S1A), that binds with high affinity to the PDZ1-PDZ2 domains of PSD-95 [33],[34], and incorporates a polyarginine peptide to enable uptake by neurons (Figure S1B). To examine the effects of CN2097 on LTP, field excitatory postsynaptic potentials (fEPSPs) were elicited from hippocampal slices of AS and WT littermate male mice, 2–4 mo of age, by stimulating Schaffer collaterals and recording from the stratum radiatum of the CA1 area. LTP was considered successful when the average EPSP slope showed an increase of at least 20% lasting 55–60 min after induction. In WT mice, we found that CN2097 significantly increased the LTP induction rate under subthreshold conditions.
Thus, a single high frequency stimulation (1× HFS, 1 s at 100 Hz) did not induce LTP in the majority of WT slices tested; with a mean fEPSP slope as a percentage of baseline of 107.9±4.8% (n = 11 of 14; Figure 1A). In contrast, applying two HFS trains (2× HFS; 2×1 s at 100 Hz separated by 15 s) significantly increased the induction rate of LTP (mean fEPSP slope of 178.5±15.8%, n = 6 of 10, p<0.001; Figure 1A). Importantly, in the presence of CN2097 (2 µM), the LTP induction rate of a single HFS was significantly increased, resulting in LTP in 69% of the slices tested with a mean fEPSP slope of 164.8±25.0% (n = 11 of 16, p<0.05 compared to control 1× HFS; Figure 1A). This LTP was shown to be NMDA receptor dependent, as it was blocked by APV, a competitive NMDA receptor antagonist (n = 8, unpublished data). CN2097 alone did not significantly affect baseline synaptic transmission (% fEPSP slope: baseline 100±0.002, CN2097 102.5±2.66, n = 7, p = 0.19; Figure 1C, top). CN5135, a negative control compound where the 0 and −2 ring positions of CN2097 were substituted with alanine residues to disrupt PDZ binding (Figure S1A), did not significantly increase the LTP induction rate of a single HFS (108.1±5.0%, n = 6 of 8, p = 0.5). These results suggest that in WT littermates of AS mice CN2097 can reduce the threshold for LTP induction.
In AS hippocampal slices, deficits in LTP were reported to be due to an alteration in the induction threshold, and LTP could be rescued by increasing synaptic stimulation at 32°C [8]. Consistent with previous studies [6]–[8],[15], we were unable to induce LTP in AS mice, using either a single (n = 5) or two sets of HFS (n = 7). Similarly, applying three sets of HFS also did not induce LTP when recorded at 30°C (3× HFS: 97±4.6%, n = 8 of 10; Figure 1B). However, in the presence of CN2097 (for 30 min before LTP induction) the 3× HFS protocol significantly increased the LTP induction rate, with LTP being observed in 67% of AS slices recorded with a mean fEPSP slope of 167.4±7% (n = 8 of 12, p<0.0001; Figure 1B). CN2097 had no significant effect on fEPSP slopes in AS slices (% fEPSP slope: baseline 100±0.005, CN2097 100.8±5.39, n = 5, p = 0.45, Figure 1C, bottom) of evoked synaptic responses.
The defects in LTP induction observed in AS mice are reminiscent of those observed in BDNF and TrkB mutant mice [40]–[42],[46], which led to examine whether BDNF signaling was compromised in AS mice. NMDA receptor activity promotes the release of BDNF [47] to stimulate TrkB-PLCγ-CaMKII/CaMKIV-CREB and PI3K-Akt signaling pathways [48],[49]. BDNF-TrkB signaling increases the delivery of PSD-95 [50] and AMPAR subunits to synapses [51],[52], and is reported to play a role in transsynaptic coordination [53],[54].
To determine whether AS mice exhibit defects in BDNF signaling, we performed western blot analysis probed with phospho-specific antibodies to assay Erk, Akt, and CaMKII activity, in age-matched WT- and AS-coronal brain slices comprising the hippocampus and cortex [36]. In terms of Erk signaling, both WT and AS slices responded similarly (Figure 2A), with phosphorylated forms of Erk (p-Erk1/2) being significantly elevated within 15 min following the application of BDNF (25 ng ml−1), and remaining elevated for at least 60 min (AS and WT; n = 4). These results suggest that BDNF-induced Erk signaling is normal in AS mice, and infer that TrkB receptor activation is intact. Confirming that TrkB activation is unaffected, we saw no significant difference in the level of TrkB expression or TrkB phosphorylation in lysates prepared from freshly isolated WT and AS hippocampal tissue (Figure 3B), or hippocampal slices (p>0.05; Figure 4A, input blot). In contrast to undetectable changes in BDNF-induced Erk activation in the AS-mouse, severe signaling defects were observed in both the PI3K and PLCγ1 downstream signaling cascades (Figure 2A). As a measure of BDNF-induced PI3K activity, we examined the phosphorylation state of its downstream effector, the serine/threonine kinase Akt [49]. The level of p-Akt (Ser473) in AS mice, at 30 min post-BDNF, was 18%±4.9% of WT (p<0.001, n = 4; Figure 2A, top panel, and quantitative analysis, lower panel). Similarly, peak BDNF-induced phosphorylation of CaMKIIα (Thr286) and CREB (Ser133) was lower, reaching levels of only 14%±3.5% and 41%±5.5%, respectively, compared to WT (p<0.001, n = 4; Figure 2A, top panel, and quantitative analysis, lower panel). As reported previously [8], we found that the level of basal p-CaMKIIα (Thr286) in AS mice was greater than WT (201%±22% at 0′, # in Figures 2A and S2A), while total CaMKII remained unchanged (Figure 2A). To confirm that the reduced signaling was not a result of slice preparation or the incubation period, we repeated the experiments using lysates made from freshly dissected AS hippocampus, and found that basal p-Akt and p-CREB were similarly diminished (p<0.01, n = 3; Figure 2B). It is also worth noting that basal levels of Arc in AS mice were approximately 2-fold higher than in WT (224%±27%, p<0.01; Figure 2B).
As CN2097 facilitated LTP in AS and WT mice, we examined whether CN2097 altered BDNF-TrkB signaling. In both WT and AS brain slices CN2097 increased BDNF-induced downstream signaling of PLCγ1 (p-CaMKIIα and p-CREB) and PI3K (p-Akt) pathways, but did not alter Erk phosphorylation (Figure 2A, top panel, and quantitative analysis, lower panel). Specifically, in WT slices, CN2097 elicited increases in BDNF-induced p-CaMKII of 55%±13.5%, (30′, p<0.01, n = 4) and p-Akt of 93%±32%, (30′, p<0.05, n = 4), whereas the control compound CN5135 had no effect (p>0.1, n = 4; Figures 2A and S2A), demonstrating the specificity of CN2097. Significantly, in AS-slices, CN2097 increased p-CaMKII (86%±11% at 30′, p = 0.21, n = 4) and p-Akt (89%±22% at 30′, p = 0.35, n = 4) to levels comparable to WT slices (Figure 2A). In the absence of BDNF stimulation CN2097 had no effect in both WT and AS brain slices (Figures 2 and S2A).
These results suggest that CN2097 acts downstream of TrkB to facilitate activation of the PI3K and PLCγ pathways. However, in brain slices we cannot rule out that CN2097 enhances synaptic activity to indirectly potentiate TrkB signaling [55]. To study TrkB signaling in isolation of synaptic influences, we heterologously expressed TrkB in SH-SY5Y neuroblastoma cells (SH-SY5Y-TrkB). Although this cell line does not express endogenous TrkB, it expresses PSD-95, Ube3A, and Arc, making it a useful in vitro system to explore the role of these proteins in BDNF signaling. Upon BDNF stimulation of SH-SY5Y-TrkB cells, CN2097 enhanced phosphorylation of Akt (Ser473 and Thr308), GSK3α/β, and the mTORC1 downstream targets 4E-BP1 and ribosomal protein p70S6K (Figure 2C; n = 5), demonstrating that CN2097 directly targets downstream TrkB signaling.
The observed deficits in BDNF signaling in Ube3A knockout mice could also reflect indirect or developmental abnormalities. To directly assess whether Ube3A influences TrkB signaling, we used RNAi to acutely deplete Ube3A in primary cultures of mouse cerebellar granule neurons (CGN), where BDNF-TrkB signaling is known to play a critical role in CGN development and survival [56]–[58]. Recapitulating the AS hippocampal slice results, acute knockdown of Ube3A in WT CGNs (to 18%±10% of scrambled RNAi protein levels), had no effect on p-Erk signaling but severely reduced BDNF-mediated p-CaMKIIβ the isoform of CaMKII that is highly expressed in mouse CG neurons [59],[60], and p-Akt (S473) signaling (p<0.005; Figure 2D). Similar to the responses documented in the AS slice studies shown above, p-Akt and p-CaMKII levels were restored to WT levels with CN2097 pretreatment (Figure 2D; p>0.1). In addition, knockdown of Ube3A in SH-SY5Y-TrkB cells (to 10%±3.0% of untransfected protein levels), produced identical deficits in BDNF signaling (Figure S2B, upper panel, and quantitative analysis, lower right panel), and enhanced expression of Arc, a previously identified Ube3A substrate (Figure S2B, middle right panel) [10]. The specificity of the RNAi was demonstrated by coexpression with recombinant Ube3A, which rescued signaling (Figure S2C, upper panel).
Defective recruitment of signaling intermediates to TrkB could explain the deficits in Akt and CaMKII signaling observed in AS mice [49]. Using coimmunoprecipitation (co-IP) studies, we examined the association of Grb2-associated binder 1 (Gab1), a major adaptor protein required for BDNF activation of the PI3K-Akt cascade [48],[49], and PSD-95, which has been shown to coimmunoprecipitate with TrkB [50],[61]. In contrast to WT brain slices where BDNF increased PLCγ, Gab1 (Figure 3A, upper panel) and PSD-95 (Figure 3A, lower panel, and Figure 4A, upper panel) recruitment to TrkB, these interactions were diminished in AS slices (PLCγ, 28%±4%; Gab1, 23%±5.5%; and PSD-95, 15%±4%; p<0.001 compared to WT; Figures 3A and 4A). CN2097 improved the association of PLCγ (71%±4%), Gab1 (62%±14%), and PSD-95 (53%±9%), with TrkB in AS slices, and enhanced these interactions in WT slices (∼50%, p<0.001; Figures 3A and 4A). Consistent with these results, co-IP studies performed on freshly dissected hippocampal and cerebellar AS tissue also showed reduced association of PSD-95 and Gab1 with TrkB (p<0.01; Figure 3B). The similar deficits in both the hippocampus and cerebellum are consistent with a lack of Ube3A expression in these brain regions [62],[63]. Overall, the observed deficits in TrkB signaling in AS mice are consistent with the inability of TrkB to associate with specific signaling adaptors (Figure 4D).
We next tested whether the presence of PSD-95 is required for BDNF-induced PLCγ1 and PI3K signaling. RNAi knockdown of PSD-95 using a synthetic siRNA duplex (sh95A) was highly effective in primary CGN cultures (18.7%±8.9% compared to WT), and significantly reduced BDNF-induced phosphorylation of Akt and CaMKII (p<0.01; Figure 3C), identifying a unique role for PSD-95 in neurotrophin signaling. In contrast, BDNF-induced Erk signaling was normal, inferring that PSD-95 knockdown did not affect TrkB receptor activation. In support of this, PSD-95 knockdown did not alter the levels of TrkB expression or activation in CGNs or SH-SY5Y cells (Figure S3C). PSD-95 knockdown in SH-SY5Y-TrkB cells also blocked Akt and CaMKII, but not Erk activation (Figure S3C). The specificity of the PSD-95 RNAi was demonstrated by coexpressing RNAi resistant PSD-95 along with a pRS plasmid containing the shRNA to PSD-95 (sh95A) in SH-SY5Y-TrkB cells, which restored Akt signaling (Figure S3F). Further supporting that the signaling defects are a result of reduced PSD-95 expression, a second previously characterized PSD-95 shRNA construct (sh95B) [44], similarly decreased Akt and S6 activation (Figure S3F). Importantly, the knockdown of PSD-95 abrogated CN2097 rescue of p-Akt and p-CaMKII signaling (p<0.001; Figures 3C and S3C), lending support to the hypothesis that the mechanism by which CN2097 rescues signaling is through its interaction with PSD-95.
If PSD-95 acts as an intermediate in TrkB activation of PLCγ1 and PI3K signaling, then increasing PSD-95 levels would be predicted to enhance BDNF signaling. The exogenous expression of PSD-95 in neurons could generate ambiguous results, as PSD-95 promotes synaptic maturation [26], which could indirectly potentiate TrkB signaling [55]. To examine whether increased levels of PSD-95 enhance TrkB signaling, we again utilized the simplified SH-SY5Y cell line system, where TrkB and PSD-95 were exogenously expressed. We found that overexpressing PSD-95 in SH-SY5Y-TrkB cells increased BDNF-induced phosphorylation of Akt (S473, T308), GSK3β, CaMKIIα (T286), and the mTORC1 downstream target p70S6K, while leaving p-Erk and p-TrkB unaffected (Figures 3D and S3D for quantitative analysis). Thus, increasing the levels of PSD-95 mimic the effect of CN2097 on BDNF signaling, leading to the hypothesis that even though PSD-95 expression in AS mice is similar to WT, PSD-95 function is compromised. All indications from this study are that CN2097 increases the amount of PSD-95 available to bind to TrkB, which in turn facilitates signaling. Extending this rational, increasing the levels of PSD-95 would be predicted to overcome the effects of Ube3A knockdown. Testing this hypothesis, we found that overexpression of PSD-95 in SH-SY5Y-TrkB cells, in which Ube3A was knocked down, partially restored BDNF-induced Akt (S473 and T308) and CaMKII phosphorylation (Figures 3E and S3E for quantitative analysis). Co-transfection with other related MAGUKS (SAP97, SAP102, and Chapsyn110) did not facilitate TrkB-induced p-Akt or p-S6 signaling (Figure S3G), which is consistent with the higher PDZ binding affinity of CN2097 for PSD-95 and not these MAGUKs [33]. We also would anticipate that if CN2097 is specific for PSD-95, it should not affect signaling pathways that do not involve PSD-95. In agreement, neither CN2097 nor PSD-95 knockdown had any effect on altering insulin-like growth factor-1 (IGF) (Figure S3H) or epidermal growth factor induced signaling (unpublished data). Overall, these results show that PSD-95 plays a key role in BDNF-induced PLCγ1 and PI3K signaling, and suggest that in Ube3A-deficient neurons the association of PSD-95 with TrkB is decreased, resulting in defective signaling.
Loss of Ube3A expression in AS mice result in increased levels of Arc (Figures 2B and 4A, lower panel), which has been shown to deleteriously affect synaptic plasticity [10],[16]. Arc has been reported to bind the SH3-GK domain of PSD-95 [64], and proteomic studies show that Arc is one of the proteins to co-purify with PSD-95 [19], raising the possibility that Arc's association with PSD-95 influences BDNF signaling. Co-IP studies performed from hippocampal slices are consistent with this hypothesis, showing that higher amounts of Arc co-IP with PSD-95 in AS compared to WT slices (Figures 4A, upper blot, and S4A, upper panel for quantification). Conversely, higher amounts of PSD-95 co-IP with Arc in AS compared to WT slices (Figures 4A, lower panel, and S4A, lower panel showing co-IP performed with normalized Arc levels). Interestingly, Arc appears to associate with PSD-95 independently of BDNF stimulation (Figure 4A), which was also observed in SH-SY5Y-TrkB cells co-expressing Arc-myc (Figure S4B, upper and quantitation in middle panel). Treatment of the brain slices with CN2097 reduced the Arc-PSD-95 interaction, which was coincident with increased association of TrkB receptors with PSD-95 (Figure 4A).
To directly test the hypothesis that Arc expression can interfere with BDNF signaling, we expressed Arc in SH-SY5Y-TrkB cells and examined TrkB signaling. Arc overexpression decreased the levels of BDNF-induced p-Akt, p-CaMKII, p-4EBP1, p-S6 (Figures 4B and S4B, lower panel quantitation), but not p-Erk or p-TrkB, identical to knockdown of PSD-95 (Figure 3C). Significantly, CN2097 restored the defective CaMKII, Akt, 4EBP1, and S6 signaling resulting from Arc overexpression to control levels (Figures 4B and S4B, lower panel quantitation), and disrupted the interaction between Arc and PSD-95, with a concomitant increase in BDNF-induced TrkB-PSD-95 association (Figure S4B, middle panel). In a different approach to increase Arc levels we expressed a dominant negative Ube3A (E6AP C833A) [65] in the SH-SY5Y-TrkB cells. We also observed significantly reduced BDNF-induced p-Akt signaling when compared to SH-SY5Y-TrkB control cells (Figure S4C), which was rescued with an Arc-shRNA (Figure S4C) [10].
If Arc is a critical inhibitor of PSD-95-TrkB interaction and downstream signaling in AS mice, then decreasing Arc expression in AS neurons would be predicted to rescue TrkB downstream signaling. Using a previously described Arc siRNA [10], we knocked down Arc in primary cultures of CGNs prepared from the cerebellum of individual postnatal (day 5) WT and AS mice. Ube3A expression is extremely low in the cerebellum of AS mice (Figure S3B), and as shown in Figure 4C, AS granule neuron cultures had greatly reduced expression of Ube3A compared to WT, resulting in increased expression of Arc. These AS cultures exhibited reduced BDNF-induced Akt signaling (comparing sc-RNAi transfected WT and AS, p<0.01; Figure 4C), consistent with the Arc overexpression studies performed in SH-SY5Y-TrkB cells (Figure 4B). Arc-RNAi reduced Arc expression, and in AS neurons restored BDNF-induced Akt phosphorylation to WT control levels (p>0.1). We also observed that knockdown of Arc in WT neurons decreased BDNF-induced Akt signaling (Figure 4C; p<0.01), suggesting that basal levels of Arc are needed for TrkB signaling. This observation is not unexpected as the expression of Arc appears to be finely tuned to regulate LTP [11],[12],[66], endocytosis [16], and LTD [12],[17].
TrkB signaling has been reported to be required for synaptic localization of PSD-95 [50] and excitatory synapse formation [67]. This finding, together with the observation that AS mice exhibit abnormal synapse formation in the CA1 region of the hippocampus [62], suggests that the synaptic localization of PSD-95 may be reduced in AS neurons. To investigate this we performed a quantitative analysis of synaptic puncta density in the hippocampal CA1, CA3, and dentate gyrus (DG) subregions, using antibodies to synaptophysin (SYN) and PSD-95 to visualize presynaptic and postsynaptic specializations, respectively.
To first examine for differences between WT and AS mice we used Image J software to compare the relative numbers and intensity of staining of discrete PSD-95 and SYN immunolabeled puncta in the CA1, CA3, and DG subregions (Figure S5). Importantly, in the CA1 stratum radiatum (CA1-SR) region, AS mice exhibited a significant reduction in the number of PSD-95-positive puncta (0.710±0.026 puncta/µm2, p<0.05, 16.2%) compared to control mice (0.847±0.055 PSD-95 puncta/µm2; Figure S5A), with no significant reduction in CA3 or DG regions (Figure S5E and S5I). Examination of presynaptic puncta revealed a significant reduction in the number of SYN puncta in both the CA1 (Figure S5B) and CA3 subregions of the AS mouse (Figure S5F), with the greatest reduction occurring in CA1-SR region (28.0%) (Figure S5B). However, no difference in SYN puncta was detected in DG (Figure S5J) between AS and WT mice. In addition to a decline in SYN-puncta in the AS mouse, the total puncta intensity of CA1 SYN-IR was significantly reduced in CA1-SR (Figures 5E, insert, and S5D) compared with WT (Figures 5B, insert, and S5D). A significant reduction of SYN-staining intensity was also shown in the neuropil of CA3 (Figure S5F), but not in the DG (Figure S5J). Overall, these data indicate that the CA1 region in the AS mouse has a significant reduction in both PSD-95 and SYN puncta density and intensity. Independent conformation of SYN-decline in the hippocampus of the AS mouse was demonstrated by Western blot (Figure 5J).
We next addressed whether the number of PSD-95 contacts with SYN-IR puncta (synaptic density) was affected. Although the use of image analysis software was able to show significant relative differences between AS and WT mice in the distinct regions of the hippocampus, these data were limited by the minimal thresholding capabilities of the software, whereby faintly stained puncta, particularly in the AS mouse, were not recorded. In order to get an accurate assessment of the synaptic density in CA1, manual counting the number of PSD-95 (Figure 5A and 5D), SYN (Figure 5B and 5E), and their synaptic contacts (Figure 5C and 5F) was preformed (See Methods for description of manual counting). In agreement with previous reports [67], the distribution of PSD-95-IR puncta averaged 1.29 puncta/µm2±0.03, and the synaptic density averaged 0.95±0.03 in the WT mouse (Figure 5G and 5I). In the AS-CA1 region, we found that there was a significant reduction in both the number of PSD-95-IR puncta (Figure 5G; 1.15 puncta/µm2±0.02: p<0.01, two-tailed t test, with normality tested and confirmed using the method of Kolmogorov and Smirnov), as well as the number of synaptic contacts between PSD-95 and SYN (Figure 5I; 0.82 contacts/µm2±0.02, p<0.01), which translates to a 10.9% reduction of PSD-95-IR puncta and a 13.7% decline in synaptic density in the AS-mouse within the neuropil region of CA1. These data confirm that in the AS CA1 region there is a significant reduction in the number of synapses.
Confirming that the PSD-95-IR staining was in agreement with previous reports in the WT-mouse [67], we next sought to determine if there is a shift in the distribution of staining intensity of individual PSD-95-puncta making contact with SYN in the AS-mouse. Mean intensity for each synaptic PSD-95 puncta was measured by placing a 34-pixel circle over the punctum using ImageJ (see Methods). Results showed that the mean intensity of puncta in AS CA1 was lower than WT (Figure 5K), and plotting the distribution of all measured puncta as a function of mean intensity showed that there is a shift to the left in the AS mouse, indicating that a significant number of synaptic PSD-95-IR puncta in the AS mouse are less intensely stained than in WT. Also, despite this shift to less-intensely stained synaptic puncta in the AS mouse, there was no difference noted between AS and WT CA1-SR in the average intensity distribution of all identified PSD-95 stained puncta (Figure S5C), indicating that synaptic PSD-95 puncta are preferentially compromised. A similar loss in synaptic PSD-95 staining intensity was observed in the DG region (Figure 5M) but not in the CA3 (Figure 5L). These results are consistent with deficits in TrkB signaling [50].
Numerous studies show that BDNF plays a critical role in synaptic plasticity and memory [68],[69]. Significantly, the deficits in synaptic plasticity observed in AS mice are similar to those in BDNF or TrkB mutant mice [40]–[42]. TrkB-PLCγ signaling on both sides of the synapse has been reported to be required for LTP induction [70], and synaptic activity is believed to trigger the secretion of BDNF either pre- [71] or post-synaptically [47], to activate postsynaptic TrkB-PI3K signaling and increase PSD-95 transport to synapses [50],[72]. Suppression of TrkB signaling leads to a reduction in the number and intensity of PSD-95 puncta along dendrites [72] and TrkB knock-out mice show decreased numbers of hippocampal Schaffer collateral excitatory synapses [67],[73]. Our data show that in the hippocampal CA1 region of AS mice there is a significant reduction in the number of synapses, and a decrease in the intensity of PSD-95 puncta colocalized with SYN (Figure 5I and 5K). These results are compatible with the reported abnormalities in spine morphology, number, and length in CA1 pyramidal neurons of Ube3a maternal-deficient mice [62], and are consistent with the impaired LTP in this region (Figure 1B), and compromised cell signaling mediated by BDNF/TrkB in CA1 pyramidal cells.
We have identified PSD-95 as a novel TrkB-associated protein, critical for full activation of downstream PI3K-Akt and PLCγ signaling. We found that BDNF-induced association of PSD-95 with TrkB is impaired in Ube3A m−\p+ mice, resulting in reduced activation of signaling molecules downstream of TrkB, including Akt-mTORC1 and PLCγ-CaMKII, whereas the Erk pathway appeared intact. The in vivo defects in BDNF signaling observed in AS mice were recapitulated in vitro by depletion of Ube3A or increasing the expression of Arc, while elevating the expression of PSD-95 restored defective signaling. Moreover, a bridged cyclic peptide (CN2097) shown by nuclear magnetic resonance (NMR) studies to uniquely bind the PDZ1 domain of PSD-95 with high affinity [33],[34] enhanced BDNF-induced TrkB-PSD-95 complex formation, to improve signaling in AS brain slices. Depletion studies show that PSD-95 is essential for the effectiveness of CN2097 restoration of signaling in Ube3 deficient neurons. Rescue of BDNF signaling by CN2097 was shown to strongly correlate with the ability of CN2097 to decrease PSD-95 association with Arc. Overall, these studies demonstrate PSD-95 enhances TrkB-induced PLCγ and PI3K signaling pathways (Figure 4D), which could further facilitate PSD-95 trafficking to synapses [50] and modulate NMDA receptor-dependent synaptic plasticity [69],[74].
In addition to postsynaptic signaling deficits in the AS mouse, we also observed deficiencies in the presynaptic terminals, where immunostaining for the synaptic vesicle protein SYN was shown to be significantly less in CA1-SR in the AS mouse (Figure 5E, inset), as compared with a WT littermate (Figure 5B, inset). In agreement with the reduction of the average intensity of SYN puncta (∼53%; Figure S5D), the expression level of SYN in the hippocampus of Ube3A m−\p+ mice, by Western blotting, was greatly reduced (∼30%; Figure 5J), as reported previously [75]. In addition, there was a significant reduction in the number of SYN puncta in this region of the AS mouse (Figure 5H). It is possible that loss of Ube3A, which has been shown to also localize to the presynaptic compartment [62], leads to an impairment in presynaptic function. In support of this, Philpot and colleagues found a decrease in the number of synaptic vesicles at both excitatory and inhibitory synapses of Ube3am−/p+ mice, resulting in a severe deficit in inhibitory drive to neocortical pyramidal neurons [76]. Alternatively, the reduction in the number of SYN puncta in AS mice could be the indirect result of defective transsynaptic BDNF signaling [53],[54]. BDNF and TrkB knockout mice also show a reduction in SYN levels in synaptosomes, leading to a reduction in docked synaptic vesicles and synaptic fatigue [46],[73],[77]. Prolonged BDNF treatment reversed the synaptic fatigue observed in BDNF knockout mice, suggesting a role for BDNF in the mobilization and/or docking of synaptic vesicles [35],[46].
Mechanistically, the specific function of PSD-95 in BDNF-mediated signaling remains to be elucidated. Increasing evidence indicates that Trk activation of differential signaling pathways is tied to their internalization and trafficking to endosomes [78],[79]. Lipid rafts also appear to be essential for Trk-activation of the PLCγ pathway [80]. Interestingly, in Fyn-deficient mice TrkB translocation to lipid rafts is prevented and mice exhibit compromised BDNF-induced PLCγ, PI3K-Akt, but not Erk signaling [80], identical to our findings in AS mice. PSD-95, which associates with lipid rafts by palmitoylation [81], and has been shown to bind Fyn [82], may be required for the intracellular sorting or retention of TrkB to lipid rafts [83].
Our data are consistent with the hypothesis that, in AS mice, the increased levels of Arc and its association with PSD-95 interfere with BDNF/TrkB-dependent recruitment of PSD-95 (Figures 3A and 4A). However, it still remains to be determined if Arc directly binds to PSD-95 or if linking proteins are required for its association. One possibility is that Arc could bind indirectly to PSD-95 via dynamin or endophilin to regulate TrkB endocytic trafficking and signaling from early endosomes. Trk signaling is dependent on internalization via a dynamin-dependent process and endophilins regulate BDNF-TrkA early endocytic trafficking and signaling [84]. Arc directly binds and recruits endophilin and dynamin to early/recycling endosomes [16], and recent studies indicate that Arc plays a role in the postsynaptic trafficking of amyloid precursor protein [85]. Furthermore, Arc and PSD-95 are both critical for AMPAR trafficking [16],[22],[25],[27],[86], raising the possibility that the association of Arc with PSD-95 may also provide a mechanism to control AMPAR trafficking.
A number of recent studies aimed at restoring Ube3A activity in the AS mouse demonstrate that it may be possible to improve learning and memory in adult AS patients [87],[88]. The discovery that CN2097 can reinstate BDNF-signaling pathways in AS mice to facilitate LTP suggests that the development of drugs targeting PSD-95 may be beneficial in treating the behavioral deficits in AS. Therapeutic strategies that enhance neurotrophin signaling may be additionally advantageous as signaling through these receptors appears to be compromised in other neurological diseases [89],[90].
2–4-mo-old male WT and AS mutant (Ube3Am−/p+) mice were bred on a 129S7 background strain, supplied by Jackson laboratories (stock number 00447) [6], were deeply anesthetized with isoflurane, then decapitated. Brains were rapidly removed and placed in 4°C dissecting solution (containing 60 mM NaCl, 3 mM KCl, 1.25 mM NaH2PO4, 28 mM NaHCO3, 110 mM sucrose, 0.6 mM ascorbic acid, 5 mM Dextrose, 7 mM MgCl2, and 0.5 mM CaCl2.H2O, [pH 7.25–7.35]). Brains were then sectioned and glued to the stage of a vibrating blade microtome (Vibratome) and coronal “brain slices” (450 µm) containing the dorsal hippocampus and adjoining cortex were cut and incubated in a humidified interface chamber containing oxygenated artificial cerebral spinal fluid (ACSF) (95% O2/5% CO2) containing 119 mM NaCl, 2.5 mM KCl, 1 mM NaH2PO4, 26 mM NaHCO3, 11 mM glucose, 1.3 mM MgSO4.7H2O, and 2.5 mM CaCl2 with pH (7.25–7.35) at room temperature for >1–2 h prior to use. For Western blot analyses, several brain slices were transferred to room temperature 20-ml submersion chambers containing continually oxygenated ACSF. Test reagents were added directly to the chamber.
For electrophysiological recordings, slices were transferred to a 1-ml submersion-type recording chamber perfused with 30°C, oxygenated ACSF at 2 ml/min−1. Borosilicate glass microelectrodes (resistance <1 MΩ) were placed in CA1 stratum radiatum for extracellular recordings. Synaptic responses were elicited by stimulation of the Schaffer Collaterals with 0.3-ms square wave pulses with a concentric bipolar electrode. Stimulation intensity was adjusted to record stable (<5% drift) fEPSPs at 50% of maximum amplitudes (>2 mV minimum). fEPSPs were recorded (AxoClamp2B amplifier, Axon instruments), Bessel filtered at 1 Hz and 1 kHz (Dagan, EX1 Differential Amplifier), digitized at 10 kHz (NI BNC2010A), and stored for analysis (Igor pro, Neuromatic and nClamp, www.neuromatic.thinkrandom.com). HFS trains consisted of one, two, or three 1-s 100-Hz, 0.2-ms pulse duration, over 30 s. Effects were presented as average ± SEM. Significance was determined using paired t tests.
To obtain heterozygous AS mice missing the maternally Ube3Am−/p+ gene, we crossed a heterozygous female mouse with a WT male mouse. In all experiments we used male mice aged between 2–4 mo. Control mice were age-matched, male, WT littermates. Mice were raised on a 12-h light/dark cycle, with food and water available ad libitum and were housed in groups of two to three per cage. Electrophysiological recordings were obtained from at least three mice for each condition. All animal procedures were performed in compliance with the US Department of Health and Human Services and the IACUC animal care guidelines at Brown University.
Methods for isolation and purification of murine and rat CGNs were as previously described [91],[92]. CGNs were purified from P5 mouse pups, resuspended in serum-free medium (SFM) and plated at a density of 2.5×106 cells/well, six-well plate) with poly-L-ornithine (0.01%; Sigma) and mouse laminin (20 µg/ml; Invitrogen). SFM was composed of Eagle's basal medium with Earle's salts (BME; Gibco) supplemented B27, bovine serum albumin (10 mg/ml), all from Sigma, and glutamine (2 mM; Gibco), glucose (0.5%) and penicillin/streptomycin (20 U/µl; Gibco).
Brains were fixed by perfusion of the animals under deep anesthesia (Ketamine, 75 mg/kg bw and Medetomidine, 0.5 mg/Kg bw, i.p.) through the ascending aorta with 4% (w/v) paraformaldehyde in 0.1 M sodium phosphate buffer (PBS, pH 7.2); brains were removed and immersed overnight at 4°C in the same solution. Brains were then cryoprotected by soaking in 20% (w/v) sucrose in PBS (12 h, 4°C) and 30% (w/v) sucrose in PBS (8 h, 4°C) and frozen in OCT by immersion in dry ice-cooled isopenthane. Brains were sectioned into a series of 30-µm-thick frontal or coronal slices; slices were post fixed, before the immunohistochemical staining, with the fixative solution for 2 h at 37°C. Immunohistochemistry was performed by the free-floating technique. After a 15 min treatment with 0.3% (v/v) Triton X-100 (in PBS), slices were immersed and gently shaken in a PBS-diluted (10%, v/v) horse serum (HS), containing 5% (w/v) BSA for 4–5 h at room temperature (“blocking” step). The immunostaining step was conducted by immersion and gentle shaking of the slices with the primary antibodies, diluted with blocking medium (anti-PDS-95 rabbit polyclonal antibody, Cell Signal, 1∶100 and anti-SYN mouse monoclonal antibody, SY38, Abcam, 1∶30), for 24–36 h at 4°C. Primary antibody specificity for PSD-95 and was confirmed using Western blot, After washing (10 min in blocking medium, three times), sections were incubated with secondary antibodies diluted in blocking medium, as biotinylated anti-rabbit antibody (Vector Laboratories,1∶300) and anti-mouse Texas Red-conjugated antibody (Vector Laboratories, 1∶300), for 4–5 h at room temperature. After antibody removal (blocking solution, 10 min, once; PBS, 10 min, three times), slices were then incubated with fluoroscein-conjugated avidin (Vector Laboratories, diluted 1∶500 with PBS), for 1 h at room temperature. After washing (PBS, 10 min, four times), the wet sections were placed on slides and coverslipped with anti-fluorescence fading fluid (Vectashield, Vector Laboratories). Control sections in which primary antibodies were omitted showed no labeled cells. Images of equivalent regions, 512×512 pixels, were made on a Leica TCS SP2 AOBS spectral confocal microscope using a 100×, 1.4 numerical aperture oil-immersion objective at a 4× zoom. A single optical section (0.8-µm thick) was taken from each tissue section, and at least five sections per region of interest were analyzed for each animal. All microscope settings were unchanged from section to section.
Due to the reduced intensity of PSD-95 and SYN immunostained puncta in the AS mouse, a significant number of the weaker stained opposing puncta could not be resolved using digital-quantification software; therefore, we chose to manually quantify the synaptic contacts to obtain a more accurate assessment of PSD-95/SYN contacts. PSD-95 and SYN paired confocal images were adjusted for contrast/brightness to optimize detection of the faintly stained PSD-immunostained puncta and then magnified to fit the dimensions of a high resolution flat-screen monitor (179 mm×179 mm). PSD-95-immunostained puncta were first mapped onto an acetate sheet overlay and the counted to obtain total distribution of PSD-95 puncta/µm2. The corresponding SYN-immunostained image was then superimposed under the PSD-95 acetate template, whereby all PSD-95/SYN contacts were then circled and quantified. Quantification of synaptic contacts was preformed blindly by at least three separate individuals. Data represent a minimum of seven sets of non-overlapping images/region/strain quantified blind. Only direct SYN/PSD-95 contacts were counted, i.e., PSD-95 puncta in apposition to SYN-immunoreactive puncta, were then assessed with Image J for their corresponding staining intensities by placement of a circle 32 pixels2 over each of the thresholded puncta. It is important to note that in Image J, thresholding an image does not change the value of the puncta intensity from the non-thresholded image. Groups of animals were compared using a two-tailed, two-sample equal variance t test.
WT human Myc-DDK-tagged Arc cloned into the pCMV6 vector was obtained from OriGene and HA-E6AP cloned into pCMV4 was from Addgene. The WT-TrkB cDNA was a gift from Luc De Vries and Myc–PSD-95 cloned into GW1-CMV (British Biotechnology) was a gift from Morgan Sheng. Knockdown was performed using HUSH technology (Origene). A pRS plasmid vector (Origene) containing shRNA that is effective against human or mouse Ube3A (5′AGGTTACCTACATCTCATACTTGCTTAA); human and mouse PSD-95 (5′-GG AGA CAA GAT CCT GGC GGT CAA CAG TGT, sh95A; and 5′-AAT GGA GAA GGA CAT TCA GGC GCA CAA GT, sh95B) were purchased from Origene. RNA duplex oligonucleotides (RNAi) against Ube3A (5′-GUUACCUACAUCUCAUACUUGCUUUAA) or PSD-95 (5′-AGA CAA GAU CCU GGC GGU CAA CAG UGU, sh95A; and AAT GGA GAA GGA CAU CCA GGC ACA CAA GT, sh95B) were purchased from Integrated DNA Technologies (IDT). Arc shRNA was generated using the previously described sequence [10]. Plasmid transfection into neuroblastoma SH-SY5Y cells was performed with Lipofectamine 2000 (Life Technologies) in six-well plates. A pEGFP cDNA plasmid was added to determine the efficiency of transfection which was at least 30%–40% for each experiment.
For knockdown of Ube3A or PSD-95 primary CGNs, dense cultures were grown for 3 d in vitro and then switched to DMEM (no FCS) medium 1 h before transfection. 2 µl of Ube3A- or PSD-95-specific RNAi duplexes (100 µM, diluted in siRNA dilution buffer [Santa Cruz Biotechnology] and 3.0 µl of Lipofectamine PLUS Reagent [Invitrogen]) were diluted in 90 µl of siRNA dilution buffer. To this was added 2.8 µl of Lipofectamine LTX and incubated for 30 min at room temperature. The complex was added to the well containing 1 ml of DMEM for 2 h, with a final siRNA concentration of 100 nM. 10% FCS medium was then added back to the CGN and cultured for 24 h before DMEM starvation and further treatment.
Antibodies against Gab1 (sc-9049), PLCγ1 (sc-166938), Arc (sc-15325), CaMKIIα (sc-13141), Erk1/2 (sc-93), and Akt1-2 (sc-8312), goat antibody against rabbit immunoglobulin G (IgG) conjugated to horseradish peroxidase (HRP) (sc-2030), and goat antibody against mouse IgG-HRP (sc-2031) were purchased from Santa Cruz Biotechnology. Mouse monoclonal antibody (mAb) against PSD-95 (clone K28/43) (MABN68) was purchased from Millipore, rabbit monoclonal antibody against PSD-95 (cs-3450) was purchased from Cell Signaling Tech. Mouse mAb aAnti-Ube3A (E6AP) (clone Ex-8) was obtained from Enzo Life Sciences (BML-PW0535). P-CaMKII (Thr286/287, clone 22B1) was obtained from Cayman Chemicals and β-actin (mouse mAb) was purchased from Sigma. P-Akt (Ser473) antibody (9271) was used unless otherwise indicated as p-Akt (Thr308) antibody (9275); p-GSK-3α/β (Ser9/21) antibody (9336), p-S6 ribosomal protein (Ser235/236) antibody (2211), p-p70 S6 Kinase (Thr398) antibody (9209), p-4E-BP1 (Ser65) antibody (9451), p-p44/42 MAPK (Erk1/2) antibody (9102), p- CREB (Ser133) antibody (9191), TrkB antibody (4606), and p-Trk (C35G9) Rabbit mAb (4619) were obtained from Cell Signaling Technology.
CGNs, SH-SY5Y cells or brain slices treated with the appropriate stimuli were lysed with lysis buffer (200 mM NaCl [pH 7.4], 1% Triton X-100, 10% glycerol, 0.3 mM EDTA, 0.2 mM Na3VO4, and protease inhibitor cocktail [Roche Diagnostics]). Aliquots of 40 µg of protein from each treatment were separated by 10% SDS-PAGE and transferred onto a PVDF membrane (Millipore). After blocking with 10% instant nonfat dry milk for 1 h, membranes were incubated with specific antibodies overnight at 4°C followed by incubation with secondary antibodies (HRP-conjugated anti-rabbit or anti-mouse IgG at the appropriate dilutions) for 1 h at room temperature. Antibody binding was detected with the enhanced chemiluminescence (ECL) detection system (Amersham Biosciences). Western blot results were quantified by using Image J software (NIH) after normalization to their individual loading controls. For IP, aliquots of 700 µg of proteins from each sample were precleared by incubation with 20 µl of protein A/G Sepharose (beads) (Amersham) for 1 h at 4°C. Pre-cleared samples were incubated with specific antibodies in lysis buffer overnight at 4°C. 30 µl of protein A/G beads were added and the samples were incubated for 2 h at 4°C. The beads were washed five times with phosphate-buffered saline (PBS) (4°C) and once with lysis buffer, boiled, separated by 10% SDS-PAGE, and transferred onto a PVDF membrane followed by Western blotting analysis as described above. 40 µg of protein from each treatment was also utilized for Western blot as input controls. All blots are representative of at least three independent experiments.
The peptide KNYKKTEV, was cyclized between the valine (V) and threonine (T) residues via a β-lactam alanine and linked to a seven member linear poly-arginine tail (R7) by a disulfide bond (CN2097: R7-CC- KNYKKT[βA]EV, MW2376) to enhance its diffusion and uptake capacity by neurons in intact tissues. The cyclic and poly-arginine moieties were synthesized and purified separately and subsequently coupled. A negative control peptide was prepared by introducing two disruptive alanine residues at the critical 0 and −2 ring positions that comprise the critical binding region (CN 5135). Standard fmoc-based protocols were used to synthesize all peptides. The chemical structures were determined using a high-resolution time-of-flight electrospray mass spectrometer. Reagents were sourced from Fisher, BA Chemicals, Berry, Wikem, Sigma, Novabiochem, Chempep, Quanta, and RSP amino-acids.
A two-tailed Student's t test was used to test for statistical significance in electrophysiological results and for measuring statistical significance of quantitative Western blot and imaging analyses. GraphPad Prism data analysis software was used for graph production and statistical analysis. Values are presented as mean ± SEM. Results were deemed significant where *p<0.05, **p<0.01, and ***p<0.001.
|
10.1371/journal.pcbi.1006661 | A mathematical model of calcium dynamics: Obesity and mitochondria-associated ER membranes | Multiple cellular organelles tightly orchestrate intracellular calcium (Ca2+) dynamics to regulate cellular activities and maintain homeostasis. The interplay between the endoplasmic reticulum (ER), a major store of intracellular Ca2+, and mitochondria, an important source of adenosine triphosphate (ATP), has been the subject of much research, as their dysfunction has been linked with metabolic diseases. Interestingly, throughout the cell’s cytosolic domain, these two organelles share common microdomains called mitochondria-associated ER membranes (MAMs), where their membranes are in close apposition. The role of MAMs is critical for intracellular Ca2+ dynamics as they provide hubs for direct Ca2+ exchange between the organelles. A recent experimental study reported correlation between obesity and MAM formation in mouse liver cells, and obesity-related cellular changes that are closely associated with the regulation of Ca2+ dynamics. We constructed a mathematical model to study the effects of MAM Ca2+ dynamics on global Ca2+ activities. Through a series of model simulations, we investigated cellular mechanisms underlying the altered Ca2+ dynamics in the cells under obesity. We predict that, as the dosage of stimulus gradually increases, liver cells from obese mice will reach the state of saturated cytosolic Ca2+ concentration at a lower stimulus concentration, compared to cells from healthy mice.
| It is well known that intracellular Ca2+ oscillations carry encoded signals in their amplitude and frequency to regulate various cellular processes, and accumulating evidence supports the importance of the interplay between the ER and mitochondria in cellular Ca2+ homeostasis. Miscommunications between the organelles may be involved in the development of metabolic diseases. Based on a recent experimental study that spotlighted a correlation between obesity and physical interactions of the ER and mitochondria in mouse hepatic cells, we constructed a mathematical model as a tool to probe the effects of the cellular changes linked with obesity on global cellular Ca2+ dynamics. Our model successfully reproduced the experimental study that observed a positive correlation between an increase in ER-mitochondrial junctions and the magnitude of mitochondrial Ca2+ responses. We postulate that hepatic cells from lean animals exhibit Ca2+ oscillations that are more robust under higher concentrations of stimulus, compared to cells from obese animals.
| In most multicellular organisms, calcium (Ca2+) is a ubiquitous second messenger that controls a vast array of cellular activities spanning from cell birth to apoptosis [1]. The endoplasmic/sarcoplasmic reticulum (ER/SR) and mitochondria have been the center of attention in the study of intracellular Ca2+ dynamics, due to their role as internal Ca2+ stores. The SR is mostly found in muscle cells, which are not the subject of this paper, so we only refer to the ER. It has been suggested that dysfunction of Ca2+ regulation in the ER and/or mitochondria leads to disrupted cellular homeostasis, and is associated with pathological processes, including metabolic diseases and neurodegenerative diseases [2–7].
Upon agonist stimulation, almost all types of cells exhibit fluctuations in cytosolic Ca2+ concentration, phenomena often referred to as Ca2+ oscillations, with signals encoded in oscillation frequencies and amplitudes. Among many cellular compartments, the ER, whose internal Ca2+ concentration is three to four orders of magnitude larger than that of the cytosol in resting condition, is considered as the main contributor to the generation of Ca2+ oscillations. The ER has several types of Ca2+ channels on the membrane that release Ca2+ once activated. The most well-studied Ca2+ release channels are inositol trisphosphate receptors (IPRs) and ryanodine receptors (RyRs). As a high cytosolic Ca2+ concentration is toxic and often leads to cell death, released Ca2+ is quickly pumped back into the ER lumen through sarco/endoplasmic reticulum Ca2+ ATPase (SERCA) pumps, which consume energy to sequester Ca2+ against its concentration gradient. Some Ca2+ released from the ER can be taken up by mitochondria through the mitochondrial Ca2+ uniporter (MCU), and then released back to the cytosol via the sodium/calcium exchanger (NCX). Thus, it is generally accepted that mitochondria have the ability to modulate oscillation frequencies and amplitudes, and consequently, affect the progression of cellular activities [4].
Having a spatially extended membrane network, the ER is often positioned in close proximity with other cellular organelles and forms membrane contact sites. Such sites between the ER and mitochondria are called mitochondria-associated ER membranes (MAMs), and it has been suggested that they play a critical role in Ca2+ exchange between the organelles [4, 5]. Since mitochondrial Ca2+ regulation is closely linked with adenosine triphosphate (ATP) synthesis and reactive oxygen species (ROS) production [8], understanding the mechanisms underlying the ER-mitochondrial Ca2+ crosstalk is of great scientific and physiological interest. A major advantage of MAM formation is that due to its minuscule size, even a small Ca2+ flux into the domain would be amplified, which is important for the MCUs as they have a low Ca2+ affinity, i.e., they require a high concentration of Ca2+ in order to activate.
Arruda et al. [9] reported a positive correlation between obesity and the degree of MAM formation. They also found different expression levels of Ca2+ channels between liver cells of lean and obese mice. These findings indicate the possibility of obesity-induced changes in Ca2+ dynamics in MAMs, and consequently, in the ER as well as mitochondria. Indeed, liver cells from obese animals showed higher baselines of cytosolic Ca2+ concentration and mitochondrial Ca2+ concentration, compared to cells from lean mice. Furthermore, Ca2+ transients generated from ATP stimulation led to higher concentration peaks in obese mouse mitochondria. Interestingly, this observation was not accompanied by higher peaks in cytosolic Ca2+ concentration, i.e., cells from obese and lean mice exhibited similar ATP-induced rises in cytosolic Ca2+ concentration.
Computational models of experimental data have been a valuable tool for understanding the dynamics of intracellular Ca2+. Most models have focused on either the ER Ca2+ handling [10–13] or mitochondrial Ca2+ dynamics [14–17] and only a handful of them have integrated the dynamics from both organelles [18, 19]. Recently, there have been an increasing number of studies that combined both experimental and theoretical approaches to probe the cellular mechanisms underlying Ca2+ crosstalk between the ER and mitochondria. The model proposed by Szopa et al. [20] assumes that due to the minuscule volume of MAMs, the MCUs in MAMs sense Ca2+ concentration in the ER. Thus, the MCU Ca2+ flux in their model is essentially direct Ca2+ flow from the ER. Using numerical methods, they investigated the effects of this flow on the shape (bursting) and period of Ca2+ oscillations, and observed that mitochondrial Ca2+ concentrations tend to a high level in some regions of parameter space. Another recent model by Qi et al. [21] considers a range of possible distances between the IPRs and MCUs in MAMs, and expresses Ca2+ concentration in MAMs as a solution to a linearized reaction-diffusion equation. In this model, the concentration of Ca2+ that is sensed by the MCUs in MAMs depends on the distance of the MCUs from the point source (a cluster of IPRs) and how fast Ca2+ diffuses in MAMs. The authors showed that Ca2+ signals can be significantly modulated by this distance, and determined an optimal distance between the IPRs and MCUs for effective Ca2+ exchange for the generation of Ca2+ oscillations. On the other hand, Wacquier et al. [22] published a model that associates Ca2+ oscillations with mitochondrial metabolism, and investigated the role of mitochondrial Ca2+ fluxes on the oscillation frequency. Their model modified one of the parameters that describes the Ca2+ concentration for the activation of the MCUs to a lower concentration than the one originally suggested by Magnus and Keizer [16]. By doing so, they implicitly included MAMs, with the following assumption: MCUs are activated at the average concentration of Ca2+ in the whole cytosol (including MAMs). They found that mitochondrial Ca2+ fluxes can modulate the frequency of Ca2+ oscillations.
Here, we construct a mathematical model to investigate the cellular mechanisms underlying the altered mitochondrial Ca2+ dynamics observed in obese mice. The model extends the model of Wacquier et al. [22], and explicitly includes Ca2+ dynamics in MAMs. We incorporated the model structure proposed by Penny et al. [23], wherein the cytosol is compartmentalized to two separate domains: the bulk cytosol and membrane contact sites between organelles. Rather than expressing the Ca2+ concentration in MAMs as an algebraic function of that of the cytosol, we model it as a dynamic variable that is determined by influxes and effluxes of the domain. We investigated how Ca2+ signals are affected by the obesity-related changes in Ca2+ channel expression levels.
The full model is a fifteen-dimensional system of ordinary differential equations (ODEs). Using the quasi-steady state approximation, a standard reduction technique for systems with multiple timescales, we reduce the model dimension to eleven. The non-dimensionalization of the full model is explained in S1 Appendix.
A better way of understanding the model structure is to consider the model as two sub-models that are coupled by a common factor, Ca2+. One of the sub-models describes intracellular Ca2+ dynamics, while the other models mitochondrial metabolic pathways and membrane potential.
We first compartmentalized the cellular domain into four separate regions: the ER, the bulk cytosol, a mitochondrion, and MAM, and assumed that Ca2+ concentration within each region, denoted by CER, Ccyt, Cmito, and CMAM, respectively, is homogeneous and is determined by Ca2+ influxes and effluxes going in and out of that region. Fig 1 shows a schematic diagram of the compartments and Ca2+ fluxes in the model. The total intracellular Ca2+ concentration, Ct, is governed by an interplay of Ca2+ fluxes across the plasma membrane. We assumed Ct to be the sum of all compartments’ concentrations. Then, CER can be written as:
C E R = R V 2 f c ( C t - 1 R V 1 f c C M A M - 1 f c C c y t - 1 R V 3 f m C m i t o ) , (1)
where the f’s represent the fraction of free Ca2+ that is not bound by buffers. The Rv’s account for compartment volume differences and are defined as:
R V 1 = total cytosolic volume total MAM volume , R V 2 = total cytosolic volume total ER volume , R V 3 = total cytosolic volume total mitochondrial volume .
Given that the mitochondrial outer membrane is freely permeable to small molecules, such as Ca2+, through the voltage dependent anion channel (VDAC), we assume that Ca2+ concentration in the inter-membrane space is equivalent to that of the cytosol. Thus, there is one effective layer of impermeable boundary, the mitochondrial inner membrane, that separates cytosolic Ca2+ from mitochondrial Ca2+.
The Ca2+ dynamics part of the model consists of the following ordinary differential equations (ODEs):
d d t C c y t = f c [ ( 1 − R S1 ) ( J IPR − J SERCA ) + 1 − R S2 R V3 ( J NCX − J MCU ) + J diff + J in − J pm ] (2)
d d t C M A M = f c [ R V 1 R S 1 ( J nIPR - J nSERCA ) + R V 1 R S 2 R V 3 ( J nNCX - J nMCU ) - R V 1 J diff ] (3)
d d t C m i t o = f m [ R S 2 ( J nMCU - J nNCX )+ ( 1 - R S 2 ) ( J MCU - J NCX ) ] (4)
d d t C t = J in - J pm (5)
d d t P = τ p ( P s - P ) + pulse (6)
d d t h 42 = λ h 42 ( h 42 ∞ - h 42 ) (7)
d d t h n 42 = λ h n 42 ( h n 42 ∞ - h n 42 ) (8) P represents the homogeneous concentration of IP3 in the bulk cytosol and the MAM. h42 and hn42 denote the activation variables of the IPRs in the bulk cytosol and the MAM, respectively. The RS’s are surface ratios, defined as
R S 1=surface area of the ER that adjoins the MAM the total ER surface area , (9)
R S 2=surface area of mitochondrion that adjoins the MAM the total mitochondrion surface area ; (10)
see Table 1 for their values. Short descriptions for the J* Ca2+ fluxes are given below.
The main function of mitochondria is to create ATP by oxidative phosphorylation. Due to this particular role, mitochondria are the powerhouse of the cell. The mitochondrial metabolic pathway is initiated by the uptake of pyruvate, which is the end product of cytosolic glycolysis. Pyruvate in the mitochondrial matrix then enters the tricarboxylic acid (TCA) cycle, also known as the citric acid cycle or the Krebs cycle, to generate the reducing agent NADH that has electrons with a high transfer potential. The concentration of mitochondrial NADH can also be increased by the activity of the malate-aspartate shuttle (MAS). NADH then goes through the electron transport chain (ETC), where the electrons are separated and used to drive protons (H+) across the inner membrane and generate a proton gradient between the intermembrane space and mitochondrial matrix. As protons accumulate in the intermembrane space, the potential gradient across the inner membrane is used by the F1FO-ATPase to convert mitochondrial adenosine diphosphate (ADP) to ATP via phosphorylation. The produced ATP is then transported to the cytosol by adenine nucleotide translocases (ANT), which carry out the exchange of cytosolic ADP and mitochondrial ATP across the inner mitochondrial membrane.
Ca2+ is an important component in mitochondrial metabolism, as it promotes the production of NADH. An increase in mitochondrial Ca2+ concentration upregulates the TCA cycle, and an increase in cytosolic Ca2+ concentration stimulates the aspartate-glutamate carrier (AGC), a protein involved in the MAS. We combined the calcium model with a model of mitochondrial metabolic pathways proposed by Wacquier et al. [22]:
d d t A D P c= I hyd - I ant R V 3 (41)
d d t A D P m= I ant - I F 1 FO (42)
d d t N= I pdh - I o + I agc (43)
d d t V m = 1 C p ( a 1 I o - a 2 I F 1 FO - I ant - I Hleak - ( 1 - R S 2 ) ( J NCX + 2 J MCU ) - R S 2 ( J nNCX + 2 J nMCU ) - I agc ) (44)
The variables ADPc and ADPm measure ADP concentrations in the cytosol and mitochondrion, while N is the concentration of mitochondrial NADH. Vm models the voltage difference across the inner mitochondrial membrane. The I* rates are:
The model suggests the conservation of the following ion concentrations: total NADH (oxidized and reduced), mitochondrial di- and triphosphorylated adenine nucleotides, and cytosolic di- and triphosphorylated adenine nucleotides. Mathematically speaking,
N m i t o tot=N + N A D , (45)
A m i t o tot=A D P m + A T P m , (46)
A c y t tot=A D P c + A T P c . (47)
Other functions of the mitochondrial model in Wacquier et al. [22] are reproduced below for convenience.
I hyd=( 1 - R S 1 ) J SERCA + R S 1 J nSERCA 2 + V hyd A T P c A T P c + K hyd (48)
I ant=V ant 1 - α c A T P c A D P m α m A D P c A T P m e F V m R T ( 1 + α c A T P c A D P m e - 0 . 5 F V m R T ) ( 1 + A D P m α m A T P m ) (49)
I F 1 FO=V F 1 FO ( q 6 q 6 + A T P m ) ( 1 + e q 7 - V m q 8 ) - 1 (50)
I pdh=k gly 1 q 1 + N N A D C m i t o q 2 + C m i t o (51)
I o=k o N q 3 + N ( 1 + e V m - q 4 q 5 ) - 1 (52)
I agc=V agc C c y t K agc + C c y t q 2 q 2 + C m i t o e p 4 V m (53)
I Hleak=q 9 V m + q 10 (54)
The model parameters are in Table 1. For modeling purposes, some of the parameters are modified from their original values as in Wacquier et al. [22]. We find these modifications justifiable, as the original values were chosen by the authors to reproduce their experimental data, and hence were not based on any direct physiological evidence.
All the numerical simulations presented in this paper were computed with XPPAUT [33].
This section consists of two parts: model verification and model prediction. In the first part, model behaviors are compared with the experimental data presented in Arruda et al. [9], to show the validity of the model. In the second part, the model is used to simulate Ca2+ oscillations in hepatocytes from wild type and genetically obese (ob/ob) mice, to predict the effects of obesity-related cellular changes on Ca2+ dynamics.
The panels in Fig 8 show Ca2+ oscillations generated from the control and obesity models with the same magnitude of stimulation. The simulations suggest that hepatocytes from obese mice may exhibit faster Ca2+ oscillations compared to cells from lean mice. To quantify the frequency difference between the model, we simulated the solutions for two hours of simulation time, then computed the average frequency for the comparison. The average frequency of the oscillations in the obesity model was about 19% higher than that in the control model. Moreover, mitochondrial Ca2+ oscillations in the obesity model showed about 86% higher average level than those in the control model. In the following sections, we scrutinize how each cellular change associated with obesity modulates Ca2+ oscillations.
We have presented a mathematical model for Ca2+ dynamics in mouse hepatocytes. To our knowledge, this model is the first mathematical model to explicitly express Ca2+ concentration in MAMs as a dynamical variable and also show MAM Ca2+ levels to be within reasonable proportions of the Ca2+ levels in the other domains. The first aim of the model was to reproduce the data reported by Arruda et al. where hepatocytes with more MAMs exhibited ATP-induced Ca2+ transient with higher peaks in mitochondrial Ca2+ concentration, and hepatocytes from obese animals generated higher peaks of mitochondrial Ca2+ transient, compared to cells from lean animals, while having no significant difference in the peaks of cytosolic Ca2+ transient [9].
Arruda et al. also compared hepatic cellular characteristics between different groups of mice. They had a group of lean mice as their control, and two groups of mice, one that had been under high fat diet (HFD) and the other genetically obese (ob/ob) mice, for mouse models of obesity. The ob/ob mouse cells showed higher expression levels of IPR and MCU, as well as a higher degree of MAM formation.
We used the model to study how Ca2+ signals are altered by the cellular change associated with obesity. According to model simulations, hepatocytes from obese animals exhibit faster Ca2+ oscillations than those from healthy animals. Moreover, the average mitochondrial Ca2+ concentration is higher under obesity.
Metabolic flexibility refers to the ability of the organism to adapt its fuel source, depending on availability and need [41], and emerging evidence suggests the involvement of MAMs in this adaptation [42]. Interestingly, Rieusset et al. [43] reported a link between the disruption of ER-mitochondrial Ca2+ exchange and hepatic insulin resistance in their mouse model. As we have shown in this paper, hepatic cellular changes associated with obesity affect Ca2+ oscillation frequencies and amplitudes. However, the question of whether the altered Ca2+ dynamics plays a causal role in the development of hepatic insulin resistance and metabolic diseases remains to be explored. We hope that our model can help in addressing such puzzles.
Our model is an open-cell type that exhibits dynamic total Ca2+ concentration (Ct), which is determined by Ca2+ fluxes across the plasma membrane. Though we have not introduced any changes to these fluxes for model simulations of the obesity condition, there is experimental evidence that suggests otherwise. Arruda et al. [44] showed diminished protein-protein interaction between the ER membrane and the plasma membrane that, upon ER Ca2+ depletion, facilitates store-operated Ca2+ entry (SOCE). Such protein-protein interactions form another type of Ca2+ microdomain, and they are not explicitly included in the model as they are not the main subject of this paper. However, a follow-up theoretical study of these ER-plasma membrane junctions in a whole cell context, similar to our ER-mitochondria contact model, would be relevant and potentially useful in gaining a complete understanding.
The model is deterministic, which assumes synchronized behaviors of the activated IPRs, within each compartment. This allows relatively fast simulations of Ca2+ oscillations (periodic solutions). Furthermore, as in many previous mathematical models of Ca2+ dynamics [18, 22, 26, 28], the model assumes a homogeneous Ca2+ profile throughout the bulk cytosol. These underlying assumptions are physiologically inaccurate, as many experimental studies show the stochastic nature of intracellular Ca2+ activities and clustered spatial distributions of Ca2+ channels that regulate such activities. Nevertheless, Cao et al. [24] demonstrated that IPR stochasticity is not pivotal for qualitative predictions of oscillation traits, such as frequency. Furthermore, Voorsluijs et al. [45] presented a heuristic model that can explain the coexistence of subcellular Ca2+ signals that occur due to the intrinsically stochastic nature of the IPR activities, and the cell level global Ca2+ spikes that are more likely to arise from a deterministic mechanism. The authors showed that the oscillation periods generated from a deterministic version of their model lie within the interspike interval distributions simulated from a stochastic version of the model. Thus, we find it reasonable to utilize a deterministic model, such as ours, to simulate and probe the attributes of Ca2+ activities, at least for those at the cell level.
The model is not applicable for studying the effects of obesity and obesity-related changes in Ca2+ dynamics on mitochondrial metabolism. For the obesity model simulations, we only modified the parameters that are directly associated with Ca2+ dynamics. This is not to diminish the criticality of mitochondrial metabolic mechanisms nor to suggest that there is no effect of obesity on the mechanisms. There certainly is a plethora of experimental evidence that suggests correlations between obesity and mitochondrial dysfunction that range from structural functions such as fusion and fission to biochemical functions. What we do not fully comprehend yet is, which parts of the pathways are linked with obesity, and how they communicate with each other. Furthermore, the model is not comprehensive enough to describe mitochondrial metabolic pathways under different cell conditions. For instance, the process of glycolysis is represented by a single parameter, kgly, which is an over simplification of the process that is highly correlated with the nutrient availability, i.e., the cell’s metabolic state. Admittedly, the current state of the model and assumptions is not suitable for investigating possible mechanisms underlying the adverse effects of obesity on mitochondrial metabolic dynamics.
The best mathematical model for describing cellular Ca2+ dynamics would be a system of stochastic partial differential equations with spatial and temporal discretization. However, there is a trade-off between sufficient accuracy, computational efficiency, and data availability.
|
10.1371/journal.pntd.0005973 | Developing a dengue forecast model using machine learning: A case study in China | In China, dengue remains an important public health issue with expanded areas and increased incidence recently. Accurate and timely forecasts of dengue incidence in China are still lacking. We aimed to use the state-of-the-art machine learning algorithms to develop an accurate predictive model of dengue.
Weekly dengue cases, Baidu search queries and climate factors (mean temperature, relative humidity and rainfall) during 2011–2014 in Guangdong were gathered. A dengue search index was constructed for developing the predictive models in combination with climate factors. The observed year and week were also included in the models to control for the long-term trend and seasonality. Several machine learning algorithms, including the support vector regression (SVR) algorithm, step-down linear regression model, gradient boosted regression tree algorithm (GBM), negative binomial regression model (NBM), least absolute shrinkage and selection operator (LASSO) linear regression model and generalized additive model (GAM), were used as candidate models to predict dengue incidence. Performance and goodness of fit of the models were assessed using the root-mean-square error (RMSE) and R-squared measures. The residuals of the models were examined using the autocorrelation and partial autocorrelation function analyses to check the validity of the models. The models were further validated using dengue surveillance data from five other provinces. The epidemics during the last 12 weeks and the peak of the 2014 large outbreak were accurately forecasted by the SVR model selected by a cross-validation technique. Moreover, the SVR model had the consistently smallest prediction error rates for tracking the dynamics of dengue and forecasting the outbreaks in other areas in China.
The proposed SVR model achieved a superior performance in comparison with other forecasting techniques assessed in this study. The findings can help the government and community respond early to dengue epidemics.
| Dengue epidemics have posed a great burden expanding of disease, with areas expanding and incidence increasing in China recently. It has remained challenging to develop a robust and accurate forecast model and enhance predictability of dengue incidence. Several state-of-the-art machine learning algorithms, including the support vector regression algorithm, step-down linear regression model, gradient boosted regression tree algorithm, negative binomial regression model, least absolute shrinkage and selection operator linear regression model and generalized additive model, were compared and evaluated to forecast dengue incidence in this study. The SVR model, based on selection by a cross-validation technique, was superior to other models assessed using weekly dengue surveillance data, Baidu search query data and meteorological data during 2011–2014 in Guangdong province. The high accuracy and robustness of the proposed SVR model to predict the occurrence of an outbreak was also validated using data from other provinces, including Yunnan, Guangxi, Hunan, Fujian and Zhejiang, spanning southern China. To the best of our knowledge, this is the first attempt to thoroughly evaluate different algorithms for dengue incidence prediction. Our identification of the optimal model will help to precisely track dengue dynamics in the country.
| Dengue is a serious infectious disease and remains rampant across tropical and subtropical regions [1]. Primary dengue infection in humans often leads to a variety of clinical symptoms, from mild fever to potentially fatal dengue shock syndrome, and effective antiviral agents capable of treating dengue infection are not available at present [1]. Aedes mosquitoes, including Aedes aegypti and Aedes albopictus, serve as the main transmission vector of dengue viruses [2]. The impacts of variability in climate conditions such as temperature and precipitation on development rates and habitat availability for Aedes aegypti and Aedes albopictus larvae and pupae have been identified [3]. By affecting agent development and transmission vector dynamics, climate factors influences the spread of dengue. According to a recent analysis of the global distribution and burden of dengue virus, the number of dengue infections per year is estimated to be 390 million, of which nearly 96 million are symptomatic [4]. The estimated number of dengue infections has sharply increased over the past 50 years, resulting in a huge impact on human health around the world.
In China, dengue is a notifiable disease, and in recent years the area affected by dengue has expanded and the incidence has steadily increased [5]. According to the China Center for Disease Control and Prevention (CDC), the range of dengue incidence is from 0.0091 to 3.4581 per 100,000 people, with a total of 52,749 new cases of dengue having been reported during 2009–2014 [6]. In particular, a succession of dengue outbreaks occurred in several provinces including Guangdong, Yunnan, Fujian, and Guangxi during 2014 (S1 Fig) [6]. All of these provinces are located close to Southeast Asian countries including Laos, Vietnam, Thailand, Singapore and Malaysia, where dengue has been hyperendemic for decades and poses a large burden of disease [7–10]. However, dengue is still characterized as an imported disease in China due to localized transmission sparked by regular virus importations from returned travelers or visitors, rather than endemic transmission [5].
Guangdong, the most developed province located in southern China, experienced an unprecedented outbreak in 2014, and the number of cases reached the highest level over the past 25 years [5]. Our previous study showed that most of indigenous dengue cases occurred in the autumn of 2014, and the Pearl River Delta Region accounted for the majority of cases [11]. In addition to this remarkable spatial heterogeneity of cases, we observed a wide temporal variation of weekly dengue incidence ranging from 0 to 9,660 cases, which makes predicting dengue incidence difficult [11].
In the absence of an effective vaccine against dengue in China, accurate and early forecasts of dengue epidemics might allow for more effective targeting of control measures for the government. Since 2008, the China CDC has introduced the China Infectious Disease Automated-alert and Response System (CIDARS), which uses a time series moving percentile method based on historical data, for detecting dengue outbreaks in China [12]. This traditional method is overly dependent on the numbers of the routine surveillance data [12]. However, routine surveillance data is typically available with a 1- to 2-week lag [13]. Recently, several studies have explored the application of internet search terms to timely monitor disease outbreak and verify the usefulness and effectiveness of the approach [13–16]. The idea of applying internet search query data may contribute to enhancing predictability for dengue in Guangdong where dengue poses a great temporal cycling of incidence.
For dengue surveillance, several attempts have been made to develop robust predictive models for dengue incidence worldwide. Althouse et al. comprehensively assessed three regression models including step-down linear regression, gradient boosted regression tree model (GBM) and negative binomial regression model (NBM) for dengue incidence prediction in Singapore, and suggested the linear model selected by AIC step-down was superior to other models compared [16]. A more recent study achieved good performance by applying the least absolute shrinkage and selection operator (LASSO) algorithm to develop a real-time model to forecast dengue in Singapore [17]. In addition, generalized additive models (GAMs) were also used as valuable tools of risk assessment for dengue dynamics in previous studies [18, 19]. Furthermore, as a kind of the state-of-the-art and powerful machine learning algorithm, support vector regression (SVR) [20] displayed excellent performances in time series prediction. However, thorough comparisons of different predictive models and thus identifying an optimal model in China are still lacking. We aimed to construct an accurate forecast model to track the epidemic trajectory of dengue by comparing different prediction algorithms. This work addressed the gap by a) rigorously evaluating predictive performance of a variety of state-of-the-art algorithms using different assessment strategies and determining the optimal model, and b) combining dengue surveillance data, meteorological and internet query information with the proposed model for dengue incidence prediction in China.
Temporal characteristics of dengue cases, DSI, mean temperature, rainfall and relative humidity for each city in Guangdong province during 2011–2014 are presented in S3–S12 Figs. There was a sharp increase in dengue cases in the autumn of 2014 for each city. In particular, the Pearl River Delta cities had the most obvious increase in the number of the notified dengue cases in September and October, and most areas in Guangdong have hotter temperatures and more rain during the summer season. The fluctuating trend in DSI was fairly consistent with the epidemic activity of dengue.
In 2014, Guangdong accounted for about 96.3% of all notified dengue cases nationwide (S1 Fig). Spatiotemporal dynamics of dengue infections and DSIs during 2011–2014 in Guangdong is presented in Fig 1. Most of the dengue cases occurred in the Pearl River Delta region of Guangdong, especially for Guangzhou, Foshan, Zhongshan, Zhuhai and Shenzhen (Fig 1A). There was a close correlation between the number of dengue cases and the DSI in Guangdong (Fig 1 and S13 Fig).
The relative predictive accuracy of dengue incidence and goodness-of-fit assessment for each model are shown in Table 1. The standardized RMSE and R-squared values for each city in Guangdong are shown in Fig 2. According to the model performance for the two prediction periods, the SVR model had the smallest RMSE values, irrespective of city. The results suggested that the SVR model outperformed other compared models and was chosen as the optimal model in this study. Results of goodness-of-fit suggested that the discrepancy between observed incidence and the incidence expected under the SVR model was smallest. Forecasts of the SVR model for the last 12 weeks and the outbreak period of dengue incidence in 2014, including 95% prediction intervals, for Foshan are presented in Fig 3. The epidemic during the last 12 weeks and the peak of the large 2014 outbreak were accurately forecasted by the SVR model. SVR model forecasts for the other four cities including Guangzhou, Zhongshan, Zhuhai and Shenzhen with a high risk of dengue infection are displayed in S14–S17 Figs, respectively. The ACF and PACF plots revealed that there was no autocorrelation in the residuals from the SVR approach established, and thus ensured the validity of the models (Fig 3 and S18 Fig). SVR algorithm consistently yielded the smallest prediction error rates for all the studied cities among the models compared, supporting the use of SVR to perform the forecasts. Additionally, the forecast accuracy of the SVR model increased as the value of parameter C got larger, and then quickly converged to a stable level, indicating the model had a good stability predictive ability (S19 Fig).
Predictions of dengue incidence in 2014 using an out-of-sample forecasting approach (1-week-ahead prediction for each forecast window) for the best fitted SVR model are shown in Fig 4. We observed an outstanding performance of the SVR model for detecting the peak of the large 2014 outbreak for the cities with a high risk of dengue infection (Fig 4A). Dynamic forecasts of dengue incidence for the five cities are presented in S1–S5 Videos. The estimated map of dengue incidence in 2014 for Guangdong province by the SVR model well described the truly epidemic proportions of this disease (Fig 4B). The ACF and PACF plots of the residuals from the fitted SVR models also revealed that there was no any autocorrelation in the residuals and the models had captured the patterns in the data quite well (S20 Fig).
To further validate the established models, we used dengue data from five other provinces, Yunnan, Guangxi, Hunan, Fujian and Zhejiang (S1 Fig), with a high risk of dengue infection in southern China. There was a high correlation between the epidemic activity of dengue infection and the trend in DSI in these areas (Fig 5A–5F). The assessment of predictions for single observations that were left out of the data set used to fit the model is presented in Fig 6. The results demonstrated a more competitive prediction by the SVR model relative to the other models, because the RMSE values of the SVR model were consistently smallest for the 1-month-ahead predictions in 2014, irrespective of the region investigated (Fig 6). The proposed SVR model had satisfactory prediction performance with large R-squared values for Yunnan (R-squared = 0.976), Guangxi (R-squared = 0.970), Hunan (R-squared = 0.997), Fujian (R-squared = 0.981) and Zhejiang (R-squared = 0.985) (Fig 6). It shows that the SVR model is a practical method to predict dengue dynamics in the five provinces.
This study demonstrates an efficient tool using a SVR algorithm to predict dengue outbreaks and track the epidemic trajectory in China. To the best of our knowledge, it is the first attempt to thoroughly evaluate the state-of-the-art algorithms for dengue prediction, and identify an optimal model that may help to complement the traditional surveillance for dengue dynamics.
Located in southern China, Guangdong has a subtropical humid monsoon climate and has frequent economic and cultural communication with the nations of Southeast Asia where dengue poses a great burden of disease. The climate, combined with Guangdong’s highly urbanized environment, favors the presence of Aedes mosquitoes and the transmission of dengue virus, thus making the area highly vulnerable to dengue outbreaks. In the absence of an effective vaccine against dengue in China, conducting a rapid survey on mosquito vector density and suppressing the vector population comprise the core of dengue-control programs at present [38]. Though a community-based integrated intervention strategy has been carried out to control dengue outbreaks in Guangdong [39], it is still important to enhance the predictability of dengue outbreaks that exhibit strong temporal cycling.
Although the China CDC has introduced the CIDARS for detection of dengue outbreaks, this method is overly dependent on numbers of notified dengue cases, and there is room to improve the predictive performance [12]. Moreover, due to an inherent defect in the routine surveillance approach, reports of the spread of dengue are delayed [13]. This may slow the quickly public health response to an impending outbreak of infectious disease to some degree. Taking these points into account, we believe that a statistical model holds the promise of being able to provide near real-time quantitative predictions of the occurrence and evolution of an outbreak of dengue, and may be used to efficiently guide the deployment of vector-control operations. Recent studies have exploited digital surveillance based on internet search behavior to timely monitor infectious diseases that have substantial seasonal and geographic variation [13–16]. Due to the increased availability and use of internet over the last decade, the behavior of people seeking information about health has been greatly changed by the availability of health-related information on the internet [40]. In China, according to the 39th Statistical Report on Internet Development, there are 73.1 million internet users in China until 2016, accounting for about 53.2% of the national population [41]. The remarkable increase in the internet use and search trends data of people is the basis for us being able to detect and track dengue outbreaks in the country.
However, evidence for a working statistical model that exhibits robust ability in the practice of dengue dynamics forecasting is still not available in China, especially for near real-time estimates of dengue epidemic activity in Guangdong, where the risk of dengue infections is high. Our study aimed to develop an accurate prediction tool for dengue outbreaks using machine learning in conjunction with internet search queries and meteorological data in China. Marcel et al. recently discussed the importance of internet-based disease surveillance for rapid disease outbreak detection, and proposed it as a powerful tool to complement traditional disease surveillance [42]. Our analysis found that specific search terms from Baidu are highly correlated with dengue incidence in China. Particularly, for Guangdong, the included search keywords showed a correlation of 0.91 with observed dengue incidence, which is basically consistent with previous studies [16].
We further demonstrate the feasibility of applying SVR in dengue incidence forecasting and show that the established SVR model is superior to the other models compared according to the results of the empirical analysis of this study. Our results, based on dengue surveillance data from five other high risk provinces of Yunnan, Guangxi, Hunan, Fujian and Zhejiang also demonstrate a more competitive performance by the SVR model. Our proposed method exhibited itself as a highly efficient tool to predict dengue incidence, and should have predictable positive impacts on the development of an early forecasting system for dengue outbreaks in China. Previous studies also show that a support vector machine-based model has high generalization performance and outperforms classical models in terms of prediction accuracy in Malaysia and Thailand, where the incidence of dengue outbreaks is also high [43, 44]. Our proposed SVR model further supports the support vector machine-based model as a highly efficient tool to predict dengue incidence.
The proposed SVR is a machine learning algorithm implementing the structural risk minimization inductive principle to minimize the generalized error bound and achieve good generalization in complex and noisy data [45]. In comparison to the considered models including step-down linear regression, GBM, NBM, LASSO and GAM, one of the main features of the SVR model is that it performs linear regression in the high-dimension feature space using ε-insensitive loss and tries to reduce model complexity, and handle different types of data sets with high prediction accuracy [46].
Although good generalization performance with SVR has been presented in this study when compared with other five models considered, this model can be abysmally slow in large-scale tasks since it has the extensive memory requirements [47]. Also, another important practical question of SVR lies in choice of the kernel [47]. Regarding the establishment of the SVR model herein, the most suitable kernel function for the dengue data should be considered. It has been suggested that linear kernel function is more robust to multicollinearity, and using the linear kernel function could achieve better performance than the RBF kernel function in case where the number of predictors is relatively large [48]. Additionally, the linear kernel has less complexity than other kernel functions because it has fewer hyperparameters and will be easier to understand. Therefore, the linear kernel function in SVR was used because it could effectively handle many variables in this analysis. Carefully tuning the cost parameter C for the established SVR model and selecting the most suitable value was also an important practical question to avoid overfitting and enhance predictive performance. In practice, the cost parameter C was varied through a wide range of values and the optimal performance assessed using cross-validation for verifying performance [49]. In this study, we applied a cross-validation technique to search the optimized value for the parameter C. By training several SVR models for different values of the parameter C, we chose the best model with the smallest RMSE.
Baidu is the most popular search engine in China, making it the most representative data source for tracking online behavior of Chinese people. However, several limitations related to internet search query based surveillance for infectious diseases should be mentioned. First, according to the 39th Statistical Report on Internet Development, the percentage of internet users in the rural areas has steadily increased and is responsible for 27.4% until 2016 [41]. Although the availability and popularity of the internet has grown greatly in the rural areas in recent years, the differences in the internet penetration between the rural and urban areas still exist and may influence the internet search queries based surveillance for dengue. Second, internet searching behavior is susceptible to the impact of media reports, which may affect the performance of the internet search term-based predictive model [50]. For example, due to a loss of resolution occurring as a result of media-driven interest that change search behavior, Google Flu Trends was reported to over-estimate the seasonal influenza [40]. In this study, we retrospectively assessed the performance of the proposed SVR model for dengue prediction. Prospective studies should be conducted to evaluate the impacts of media-driven interest or other events that change search behavior of people on the model in the future. In addition, although the variables of dengue case data, internet search surveillance data, meteorological data, and human population data were integrated and analyzed in this work, other sources of information on relevant indicators of risk, particularly evidence on mosquito density and herd immunity [16], may subsequently be incorporated in future studies. Furthermore, since annual population data in Guangdong province during the study period could not be obtained, the latest data of the 6th population census in 2010 was used to calculate the observed and predicted dengue incidence. The variation of population during the study period might affect the estimates of dengue incidence in this study.
In conclusion, the present study demonstrates the utility of using SVR model to track dynamics of dengue outbreaks in China. The proposed SVR model achieves a superior performance in comparison with other forecasting techniques we assessed. The findings of this study will be useful for the government in identifying initiatives needed to strengthen dengue control.
|
10.1371/journal.ppat.1004670 | Revealing the Sequence and Resulting Cellular Morphology of Receptor-Ligand Interactions during Plasmodium falciparum Invasion of Erythrocytes | During blood stage Plasmodium falciparum infection, merozoites invade uninfected erythrocytes via a complex, multistep process involving a series of distinct receptor-ligand binding events. Understanding each element in this process increases the potential to block the parasite’s life cycle via drugs or vaccines. To investigate specific receptor-ligand interactions, they were systematically blocked using a combination of genetic deletion, enzymatic receptor cleavage and inhibition of binding via antibodies, peptides and small molecules, and the resulting temporal changes in invasion and morphological effects on erythrocytes were filmed using live cell imaging. Analysis of the videos have shown receptor-ligand interactions occur in the following sequence with the following cellular morphologies; 1) an early heparin-blockable interaction which weakly deforms the erythrocyte, 2) EBA and PfRh ligands which strongly deform the erythrocyte, a process dependant on the merozoite’s actin-myosin motor, 3) a PfRh5-basigin binding step which results in a pore or opening between parasite and host through which it appears small molecules and possibly invasion components can flow and 4) an AMA1–RON2 interaction that mediates tight junction formation, which acts as an anchor point for internalization. In addition to enhancing general knowledge of apicomplexan biology, this work provides a rational basis to combine sequentially acting merozoite vaccine candidates in a single multi-receptor-blocking vaccine.
| The development of an effective malaria vaccine is a world health priority and would be a critical step toward the control and eventual elimination of this disease. In addition, new pharmacological solutions are necessary as Plasmodium falciparum, the deadliest of the malaria-causing parasites, has developed resistance to every drug currently approved for treatment. Understanding the interactions required for the parasite to invade its erythrocyte host, as well as being valuable to our basic knowledge of parasite biology, is important for the development of drug-based therapies and vaccines. In this study we have, for the first time, filmed P. falciparum parasites invading erythrocytes while systematically blocking several specific interactions between the parasite and the erythrocyte. We have shown there is a sequential progression of specific interactions that occur in at least four distinct steps leading up to invasion. Previous vaccine attempts have targeted one or two of these steps, however, if a single vaccine were designed to block interactions at all four steps, the combined effect might so reduce invasion that parasite growth and disease progression would be arrested. A better understanding of each interaction during invasion, their role and order, can also inform the development of new anti-malarial drugs.
| Malaria is caused by protozoan Plasmodium parasites and Plasmodium falciparum (Pf) is the most pathogenic of the five species known to infect humans, accounting for the majority of mortality from malaria. Recent clinical trials involving a pre-erythrocytic Pf vaccine, known as RTS,S, demonstrate partial efficacy [1,2], however, there remains a need to explore other vaccine options, especially those which have the potential of controlling blood stage infection. To prevent malaria caused by blood stage infection, pre-erythrocytic vaccines need to be capable of preventing virtually all parasites from exiting the liver to infect the blood. To date this has not been achieved, so pre-erythrocytic vaccines should therefore be paired with a blood stage vaccine to eliminate breakthrough parasites, thereby providing better protection from both clinical malaria and more severe sequelae. Vaccines targeting merozoites, the stage of the parasite that infects erythrocytes, have long shown promise, but their development has been hampered by limited functional knowledge of the molecular targets. In particular, while many receptor-ligand associations have been characterised, their distinct functions and relative contributions to invasion are not well established [3].
To improve our understanding of merozoite invasion, we filmed invasion of Pf merozoites and analysed the kinetics and morphology of its distinct steps [4,5]. We categorised these into three stages; pre-invasion, internalisation and echinocytosis, as was first described in P. knowlesi (Dvorak et al., 1975). The approximately 10 second pre-invasion step, is characterised by dramatic deformation of the target erythrocyte. Internalisation then ensues and 20–60 seconds later the newly infected erythrocyte takes on a stellate appearance, a phenomenon known as echinocytosis. The erythrocyte remains like this for 5–10 minutes before returning to its pre-invasion biconcave shape. The morphology and kinetics of these invasion steps are remarkably conserved across evolutionarily divergent Plasmodium species [4,5,6].
Despite its formidable technical challenges [7], live cell microscopy is a powerful tool for examining the behaviour of parasites and can reveal much about pathogenesis. Most studies of pharmacological or biological (i.e. antibodies) growth inhibitors of Pf consist of adding the inhibitor to parasite culture and measuring the parasitemia after a few days. This approach often provides little data on whether the inhibitor blocks growth, egress or invasion, and how quickly this occurs. While the effects of invasion inhibitors have been examined in great detail using fluorescent antibody probes or electron microscopy, it has usually been done with fixed cells, and therefore provides only a snapshot of a single moment in time during a rapid and highly dynamic process. To complement the considerable body of work on merozoite invasion using traditional microscopy methods with fixed cells, we have used live cell microscopy to provide an unprecedented examination of the process of invasion through systematically blocking ligands known to be involved. Here we have given new definition to these interactions, elucidating the resulting cellular morphology and temporal sequence.
A number of merozoite invasion ligands have been described. One group of these, merozoite surface proteins (MSPs), form a major component of the merozoite surface coat [3,8]. While there are many MSPs, the GPI-anchored merozoite surface protein 1 (MSP1) is both the largest, a dimer of >500kDa, and the most abundant [8,9,10,11]. Although the function of MSP1 remains unclear, the binding of exogenous heparin sulphate to MSP1 blocks invasion, suggesting it interacts with an as yet unknown erythrocyte receptor [12].
Another group of ligands involved in invasion are the ‘alternative-pathway’ ligands, so-called because the individual ligands appear to be functionally redundant. It is most likely, however, that these proteins have slight variation in their roles and work together with a combination of overlapping function and cooperation. In Pf, these ligands comprise the erythrocyte binding antigens (EBA-175, EBA-140, EBA-181 and EBL1), and the reticulocyte-binding like homologs (PfRh2a, PfRh2b and PfRh4) (reviewed in [13,14]). It appears that two PfRh proteins, PfRh1 and PfRh5, have distinct functions which are different from the others [15,16]. PfRh1 likely has a role immediately upstream of the alternative-pathway ligands, in signalling the release of micronemes containing EBA-175 [16], a process which is dependent on calcium [17]. While the alternative-pathway EBA and PfRh ligands bind to different erythrocyte receptors, studies with gene knockout parasites, in combination with mutant erythrocytes deficient in particular receptors, have shown that increased expression of some EBAs and PfRhs can functionally compensate for the lack of another [18,19,20]. This redundancy has most likely evolved to counter erythrocyte receptor polymorphism, although varying the expression of these ligands may also help the parasite circumvent the host’s antibody responses [21]. The other PfRh protein known to have a distinct function is PfRh5. Both the essentiality of PfRh5 [22], and the erythrocyte protein which PfRh5 interacts with, basigin [15,23], have been recently identified. Unlike other alternative ligands, PfRh5 lacks a transmembrane anchor, localises to the tight junction and appears anchored to the merozoite via the Ripr protein [15,24]. The PfRh5–basigin interaction, and hence invasion, can be blocked by antibodies to either protein but the precise role of this interaction in the process of invasion is unknown [15,22,25,26].
Once the merozoite binds to the erythrocyte, it reorientates its apical end onto the host cell surface and forms a connective ring between it and the erythrocyte. This ring is called the ‘tight junction’ or ‘moving junction’, and the merozoite passes through it to enter the erythrocyte powered by the parasite’s actin-myosin motor [27,28,29]. An important component of the tight junction is AMA1, a type 1 transmembrane protein, secreted onto the merozoite surface during egress from the old host cell. AMA1 binds to RON2, a member of the RON complex, which is translocated from the merozoite into the erythrocyte surface [27,30,31,32]. RON2 has an exposed loop that acts as the AMA1 binding site and in this way Plasmodium parasites encode their own ligand and host-embedded receptor. [29,33]. Both a synthetic peptide called R1 [34,35], and a peptide derived from the RON2 exposed loop, can block native RON2 binding by competing for the RON2-interacting groove of AMA1 and thus inhibit invasion [29,32,33,36]. Video microscopy of parasites treated with R1, or which have had their ama1 gene deleted, shows that merozoites still bind to and deform erythrocytes, but they cannot properly invade and fail to progress to a ring stage [32,37,38]. This appears to be due to a defective tight junction that fails, either in assisting internalization, or in resealing of the host membrane behind the merozoite, in the few cases where internalization does occur [38].
At least a dozen receptor-ligand interactions are known to play roles in the invasion of erythrocytes by Pf merozoites and these are summarised in Fig. 1A. Here we have inhibited the following five receptor-ligand interactions using; 1) heparin sulphate to inhibit MSP142 binding to unknown erythrocyte glycoproteins, 2) treatment with neuraminidase (NM) to remove sialic acids on glycophorin A (GYPA) and thus prevent binding of EBAs, 3) genetic deletion of EBA175 and NM treatment in addition to complement receptor 1 (CR1) fragments to block PfRh4 binding to CR1, 4) anti-PfRh5 and anti-basigin IgGs to block the binding of PfRh5 to basigin, and 5) R1 and RON2 peptides to inhibit AMA1–RON2 interactions.
To study these invasion pathways we have used live cell imaging while specifically ablating these interactions to characterise the role played by each receptor-ligand event in the morphological and physiological events that typify invasion. In addition to revealing the order of these events, we show that after initial contact, vigorous deformation associates with successful invasion. This deformation is mediated by the parasite’s actin-myosin motor and alternative-pathway receptor-ligand interactions. We further define the role of PfRh5, and provide evidence that an open connection forms between the apical tip of the parasite and the host cell immediately prior to invasion. We hypothesize that this open connection is mediated by PfRh5 to basigin binding and may act as a conduit for invasion proteins which establish the tight junction.
As a prelude to these invasion inhibition studies we first determined whether the four Pf strains, 3D7, D10, W2mef and ΔEBA175 (W2mef with its EBA175 gene deleted), used here, had similar invasion kinetics under permissive conditions. We began by examining the proportion of contacts between merozoites and erythrocytes that culminated in successful invasion. This method was chosen rather than total invasions per schizont rupture since it would remove some of the elements of chance such as differences in the density of target erythrocytes surrounding rupturing schizonts or the directions in which merozoites happened to be released. Putative brief contacts between merozoites and erythrocytes that produced no definitive deformation or adhesion period were discounted. For 3D7, D10 and W2mef, we observed that approximately half of parasite-host contacts progressed to invasion (Fig. 1B, S1 Table). In contrast, only 25% of the ΔEBA175 contacts resulted in invasion, however, this difference only reached statistical significance relative to 3D7 and D10.
Next, in a parallel approach, we counted the number of erythrocytes contacted before invasion occurred including the invaded erythrocyte. Both unique and total contacts (where the same erythrocyte may be contacted more than once) were counted (Fig. 1C). This indicated that 3D7 was the most efficient, since it contacted fewer erythrocytes before invading than did other strains (p≤0.04 for all comparisons), with 85% of 3D7 merozoites invading the first erythrocyte they encountered. On the other hand, ΔEBA175 was the least efficient tending to contact 2 erythrocytes on average before invasion although this was only significant relative to 3D7 and W2mef (Fig. 1C, S1 Table). Because ΔEBA175 does eventually invade, after more contacts, the number of invasions per rupture was comparable across all strains (Fig. 1D). Multiple schizont ruptures (n = 15–23) were observed for each strain to minimize discrepancies from rupture to rupture.
We also determined the periods of pre-invasion (primary contact, deformation and resting [5]), internalisation and the time to the commencement of echinocytosis and noted that, with the exception of 3D7 pre-invasion as compared to D10 (p = 0.025), there were no significant differences between the four strains tested (Fig. 1E-G). The mean lengths of pre-invasion ranged from 9–13 seconds, followed by 10–11 seconds for merozoite internalisation, which was the most tightly regulated phase of invasion with the narrowest range (Fig. 1E, F, S1 Table). From the completion of merozoite internalization to the beginning of erythrocyte echinocytosis, average times ranged from 31–38 seconds (Fig. 1G). Erythrocyte echinocytosis occurred following most, but not all, invasions (3D7: 76.7%; D10: 72.7%; W2mef: 91.9%; ΔEBA175: 83.8%). Examples of invasions for each strain are shown in S1–S4 Videos. As shown, the four strains used here have very similar invasion kinetics making it probable that the role of an invasion protein studied in one strain is conserved across all strains.
To further dissect the initial contact period we noted the degree to which merozoites deformed individual erythrocytes. Deformation is a complex occurrence that is difficult to quantify because it varies in intensity and time, lasting from a fraction of a second to several seconds and can occur in multiple waves. A simplified four-point deformation scale (0, 1, 2 and 3) was therefore devised, based on the most extreme degree of deformation achieved (Fig. 2A, B, and S1, S5–S7 Videos). When assessed using this scale there were no differences between the four parasite strains when all contacts were taken into account (Fig. 2C, D). However, there was a significant difference when comparing the deformation scores caused by merozoites which were invaders vs. non-invaders. The majority of merozoites that invaded deformed strongly (scores 2 and 3), while the majority of merozoites which did not invade deformed weakly or not at all (scores 0 and 1) (Fig. 2C, D). Thus, the degree of deformation correlated with the success of subsequent invasion.
To investigate the underlying causes of deformation we treated purified schizonts with cytochalasin D (cytD), an inhibitor of actin polymerisation, to block the merozoite’s actin-myosin invasion motor. The cytD not only blocked invasion as expected but also inhibited deformation with 88.4% of treated merozoites deforming their erythrocytes weakly or not at all (Def score 0–1), and only 11.6% deforming their erythrocytes strongly (W2m cytD, Lane 3 vs. 5, Fig. 3A,B, S2 Table, S8 Video). This data suggests that a functioning actin-myosin motor is required for deformation of the erythrocyte. To determine if there was any contribution made by the erythrocyte’s actin, we added untreated schizonts to erythrocytes treated with cytD and observed no differences as compared to untreated erythrocytes (S2 Table).
It has previously been shown that heparin is capable of inhibiting an early step in invasion [12]. To quantify the effects we performed video microscopy of D10 parasite invasion in the presence of heparin and observed a 17-fold reduction in the number of invasions per schizont rupture (from 1.7 to 0.1, D10 vs. Heparin or Lane 2 vs. 6, Fig. 3A, S2 Table). In addition, heparin markedly reduced the capacity of merozoites to deform erythrocytes following contact and deformation scores of greater than 0 were rarely observed (Lane 6, Fig. 3B, S5 Video). In spite of dramatically reduced invasion and deformation, heparin treated merozoites maintained contact with erythrocytes for extended periods. Thus, a heparin-blocked protein is likely not the initial protein binding merozoites to erythrocytes, but is involved in an early binding event that mediates weak deformation, and blocking with heparin prevents both this weak deformation and progression to stronger levels of deformation. While heparin binds to several parasite proteins, the most obvious known candidate is the MSP142 fragment of MSP1, which is already present on the merozoite surface at the time of egress and is thought to play a role early in invasion [12]. We hypothesize, therefore, that MSP1 may be responsible for mediating weak deformation of the erythrocyte, allowing progression to stronger deformation and subsequent invasion, and that blocking MSP1 via heparin prevents both this weak deformation and the progression to stronger deformation via downstream receptors.
To examine the effect of blocking the alternative-pathway we used W2mef, a parasite strain that preferentially expresses EBA175 over other alternative-pathway ligands, and relies on it during invasion. The target erythrocytes were treated with NM to cleave sialic acid from erythrocyte glycoproteins and thus inhibit the EBA proteins from binding. With this block in place the invasion rate declined 9 fold from 2.6 to 0.3 invasions per rupture (W2m vs. W2m NM or Lane 3 vs. 7, Fig. 3A, S2 Table). While the majority of merozoites weakly deformed their erythrocytes, scoring 1, 2.4% of merozoite deformed more strongly, scoring 2 (Lane 7, Fig. 3B, C).
We further tested invasion in alternative pathway null conditions by filming ΔEBA175 merozoites which primarily use PfRh4, invading NM treated erythrocytes in the presence of soluble CR1 [39]. Under these alternative-pathway null conditions, the invasion rate declined 9 fold from 2.6 to 0.3 invasions per rupture, which is in line with the degree of inhibition previously published (W2m vs Δ175, NM, CR1 or Lane 4 vs. 8, Fig. 3A, S2 Table, S6 Video [20]). In terms of invasions per erythrocyte contact, this represents a decline from 22% to 4% (S2 Table). As the inhibited ΔEBA175 merozoites failed to deform strongly and the ratios of deformations scoring 0 and 1 are similar to W2mef under NM treatment, this suggests PfRh4 and EBA175 are performing comparable roles (Lane 7 vs. 8, Fig. 3B). Overall the data indicate that EBA/PfRh protein interactions most likely cause the strong deformation observed in the pre-invasion stage.
To examine the role of PfRh5 binding to basigin we inhibited the interaction using rabbit anti-PfRh5 polyclonal IgG, which reduced invasion of 3D7 by ~90% [25]. For all parasite strains tested, a total of 42 schizont ruptures were filmed in the presence of anti-PfRh5 IgG with no successful invasions (3D7 vs 3D7 α-Rh5 or Lane 1 vs. 10, D10 vs D10 α-Rh5 or Lane 2 vs. 11 and W2m vs W2m α-Rh5 or Lane 3 vs. 12, Fig. 3A, S2 Table). Despite this, and in sharp contrast to the receptor-ligand events described above, the merozoites were able to bind and vigorously deform the erythrocytes to a similar degree as normal parasites (Lane 1–3 vs. 10–12, Fig. 3B, C, S9 Video). To determine if inhibitory anti-basigin antibodies would produce a similar effect to anti-PfRh5 antibodies, we observed invasions in the presence of the anti-basigin mAb MEM-M6/6 [22]. Similar to blocking with anti-PfRh5, in four schizont ruptures, no invasions were filmed compared to an average of 1.7 invasions per rupture without antibody (S2 Table). Likewise, although invasion was blocked, deformation with anti-basigin treatment was similar to that seen with anti-PfRh5 treatment, and deformation in both conditions was comparable to no treatment (Fig. 3C, S7 Video). This indicates that PfRh5 has a distinctive role from the other PfRhs and the EBAs, that appears to be downstream as the PfRh5-blocked merozoites can progress beyond weak to strong erythrocyte deformation. This putative role is further explored and substantiated by experiments shown below.
With PfRh5 appearing to function downstream of the EBA/PfRh ligands we next examined a step we hypothesized to be even further downstream, the AMA1 to RON complex interaction that forms the tight junction [29,30,32,33,36]. We filmed invading merozoites in the presence of RON2 peptide, which has been shown to block invasion with approximately 99% efficiency [36,40]. In our studies, treatment with RON2 peptide reduced the invasion rate of D10 merozoites approximately 8 fold, from 1.7 to 0.2 invasions per rupture (D10 vs RON2 or Lane 2 vs. 13, Fig. 3A, S2 Table, S10 Video). Some of the peptide treated merozoites appeared to embed themselves into the erythrocyte surface, however, they failed to transform into intracellular rings indicating invasion was unsuccessful. Deformation scores were similar to those of untreated parasites (Lane 2 vs. 13, Fig. 3B, C). Although the invasion-blocking effects of RON2 peptide appeared similar to that previously described for R1 peptide [40] we repeated this analysis here (Lane 14, Fig. 3A, S2 Table). We found the ratio of deformation scores in R1 peptide was similar to untreated parasites (Lane 1 vs. 14, Fig. 3B, C, S11 Video). Our data are consistent with AMA1 having a major function downstream of the EBA/PfRh’s, namely at the tight junction.
Apart from merozoite invasion and deformation we also assessed whether the erythrocytes underwent echinocytosis after being contacted following the various inhibitory treatments. Merozoite contacts in heparin or following inhibition of the alternative pathways did not result in erythrocyte echinocytosis which is unsurprising since they were blocked quite early in the invasion sequence (Fig. 3C). Interestingly, cytD treated merozoites, despite being unable to deform and invade erythrocytes, were able to reorientate and trigger echinocytosis in the absence of invasion. A comparison between untreated and cytD treated W2mef parasites, revealed that per schizont rupture, there was no difference in the number of resulting invasions or echinocytes, respectively (Fig. 4A). However, cytD treated merozoites were less efficient at triggering echinocytosis in the first cell they contacted and tended to detach and contact additional erythrocytes before triggering echinocytosis in one of them (Fig. 4B). However, once an erythrocyte was selected, cytD-treated parasites caused echinocytosis within a similar time period as untreated parasites, irrespective of invasion (Fig. 4C). This indicates that the parasite’s actin-myosin motor is not required for reorientation and echinocytosis, but rather might be important for rapid host cell selection though deformation which possibly helps embed the merozoite in the erythrocyte surface, leading to subsequent downstream events.
In cultures treated to block the PfRh5-basigin interaction, vigorous deformation did not result in echinocytosis (Fig. 3C, S7 and S9 Videos). When we blocked invasion by preventing the AMA1-RON2 tight junction interaction with RON2 or R1 peptides, echinocytosis occurred (Fig. 3C, S10 and S11 Videos). This places the AMA1-RON2 interaction downstream of PfRh5-basigin and suggests that echinocytosis is not caused by tight junction formation but rather an event upstream, possibly the PfRh5-basigin interaction. Rhoptry release from apically reorientated merozoites has previously been shown to occur even when the AMA1-RON2 interaction is blocked [27]. We thus hypothesize that the PfRh5-basigin interaction is responsible for triggering rhoptry release from apically reorientated merozoites, which in turn is responsible for stimulating echinocytosis, an event that can be separated from invasion.
If erythrocyte echinocytosis is triggered by rhoptry release or some other perturbation to the erythrocyte membrane, what specific factor(s) is (are) causing echinocytosis? It had previously been hypothesised that entry of Ca2+ into the erythrocyte during invasion may elicit echinocytic shape changes by direct effects of elevated intracellular Ca2+ concentration [4] possibly by acting on the cytoskeletal mesh[41]. Since erythrocytes do not store Ca2+, an obvious Ca2+ source was the growth media and live cell imaging of invasion was therefore performed in Ca2+ free media to determine if echinocytosis still occurred. Prior to this, invasion assays were carried out with mature schizont cultures exposed to calcium containing RPMI and DMEM media and Ca2+ free DMEM media with or without EGTA over a 90 minute invasion window (S1A Fig.). These experiments indicated invasion was reduced several fold in Ca2+ free media in agreement with previous studies [42,43,44]. We then performed live cell imaging in Ca2+ free DMEM with EGTA and found a 13 fold decrease in the average number of invasions per schizont rupture relative to RPMI (S2 Table). Interestingly, without external Ca2+ the merozoites were still able to attach to and deform their target cells normally, but there was no echinocytosis, and thus the inhibition of invasion in Ca2+ free media occurred at the same point as it did when the Rh5-basigin interaction was blocked (Fig. 3, S2 Table, S1 Fig., S12 Video).
Having established that external Ca2+ was needed for invasion, we sought to determine if the external Ca2+ was entering the erythrocyte during invasion by visualising a potential flux with live cell imaging. We did this by purifying mature schizonts and adding them to erythrocytes labelled with the membrane permeable calcium-sensitive dye Fluo-4 AM. The majority of Ca2+ signals were punctate and confined to the invasion site (Fig. 5A, S13 Video). The punctate Ca2+ signal was detected 112 times in 248 invasions (45.2%). While there was approximately a half-second time lapse between brightfield and Fluo-4 images, allowing slight movement, it appeared that the punctate Ca2+ signal was located at the apical end of the merozoite (Fig. 5A). The Ca2+ signal first appeared near the end of deformation, on average 3.5 seconds prior to the initiation of invasion, and continued for an average of 11.3 seconds (Fig. 5B, C). Since rhoptry release into the erythrocyte surface has to precede invasion, the Ca2+ signal was being observed at around the same time we would expect the rhoptries to discharge. Occasionally, simultaneous to or immediately after the punctate Ca2+ signal, we observed a strong Ca2+ flux spreading into the erythrocyte from the invasion site prior to the start of echinocytosis (5.2% of invasions) (S14 Video). There were also occasional instances where a strong Ca2+ flux in the erythrocyte occurred during echinocytosis (3.4% of invasions). The intensity and magnitude of the signal suggested a large influx of Ca2+ was possibly coming from the media.
The punctate Fluo-4 fluxes could indicate that Ca2+ is a component of the rhoptry contents and that we are observing the ion being discharged into the host cell. The highly punctate and often confined nature of the signal could also indicate that Fluo-4 is entering the rhoptry compartment from the erythrocyte due to an opening forming in the erythrocyte membrane during rhoptry discharge (Fig. 5D). The punctate Ca2+ signal was also strongly associated with echinocytosis and invasion, with the Ca2+ signal observed in 59.2% of the cases where echinocytosis occurred. In addition, in 94.7% of cases where a Ca2+ signal was observed, echinocytosis followed, strongly linking the Ca2+ signal, and putative rhoptry discharge, to echinocytosis (Fig. 5E). However, the question remains; is it the Ca2+ or is it other components being discharged from the rhoptries such as proteins and lipids that cause echinocytosis?
To investigate this further we performed live cell imaging of merozoites invading erythrocytes treated with BAPTA-AM to chelate rhoptry Ca2+ or even external Ca2+ leaking into the erythrocyte during invasion. Relative to untreated erythrocytes, no inhibition of deformation, invasion or echinocytosis was observed (S2 Table, S1B Fig.). The timing of all events leading up to invasion was also normal (S1C Fig.). Collectively our observations indicate that Ca2+ is probably not required for echinocytosis, and we hypothesize that upon rhoptry discharge other rhoptry contents enter the erythrocyte membrane and trigger echinocytosis independent of invasion (Figs 5 and 6).
In the absence of being able to directly visualise rhoptries discharging their contents we decided to use the punctate Ca2+ signal as marker for the release of these organelles and explore their putative role in triggering echinocytosis. It follows that if invasion was blocked under conditions where putative rhoptry Ca2+ signals were still observable, that echinocytosis should occur in the majority of cases. Conversely, when the rhoptry discharges were blocked, and no Ca2+ signals were observed, no echinocytosis should take place. The first invasion blocking condition tested where echinocytosis still occurred was with cytD-treated parasites. In this condition, apical orientation without deformation occurs, followed by tight junction formation but no erythrocyte internalization. In these parasites, we observed a punctate Ca2+ signal and where the alignment of merozoite and erythrocyte was side on, the Ca2+ signal localised to the region where the parasite and host made contact (Fig. 5A, right panel set). In 75% of cases where there was a Ca2+ signal, echinocytosis occurred (Fig. 5D), again reinforcing the link between the Ca2+ signal, echinocytosis and probable rhoptry release.
Similarly when we blocked with R1, which allows rhoptry release and echinocytosis, but not tight-junction formation or invasion, we observed the punctate Ca2+ signal which appeared at the same the time after merozoite contact as without R1, with the average signal starting at 16.3 seconds after contact in Fluo-4 alone and at 12.6 seconds after contact for Fluo-4 with R1. These R1 Ca2+ signals were observed in 39% of contacts that triggered echinocytosis (28 echinocytosis events and 11 Ca2+ signals, S11 Video). This further supports the use of the punctate Ca2+ signal as an indicator of rhoptry discharge which leads to echinocytosis.
Next, we attempted to verify that blockage of rhoptry release and subsequent echinocytosis would coincide with no observable punctate Ca2+ signals. Live imaging was performed on untreated schizonts with Fluo-4 AM treated erythrocytes in Ca2+ free media with EGTA, and of 11 schizont ruptures with 127 merozoite contacts we observed no Ca2+ signals (S15 Video, S2 Table). Next we imaged untreated schizonts with Fluo-4 AM treated erythrocytes in normal media containing anti-basigin IgG at 20, 10, and 2.5 μg/mL. In 20 μg/mL, from 22 schizont ruptures we observed no echinocytosis and 3 punctate Ca2+ signals in 128 merozoites that failed to invade their target erythrocytes but remained attached until the end of filming. In 10 μg/mL, from 19 ruptures we observed 2 incidents of echinocytosis (2.5%) and 9 punctate Ca2+ signals in 81 merozoites that failed to invade their target erythrocytes but remained attached until the end of filming. In 2.5 μg/mL, in 16 ruptures we observed 5 incidents of echinocytosis (5.9%) and 21 punctate Ca2+ signals in 85 merozoites that failed to invade their target erythrocytes but remained attached until the end of filming. This corresponds to punctate Ca2+ signals in 2.3%, 11.1%, and 24.7% of cases, respectively, and in comparison to 45.2% with no block, we see a dose-related increase in punctate Ca2+ signals with a decrease in anti-basigin antibody.
This study represents the first attempt to comprehensively overlay known receptor-ligand interactions with the morphological effects on erythrocytes and physiological events that occur in the seconds preceding, during and immediately after merozoite invasion of erythrocytes. By linking receptors to morphological effects on erythrocytes, as well as physiological and kinetic features, we show the probable order of sequentially acting receptor-ligand interactions leading up to invasion. In order, these are a heparin-blocked interaction (possibly MSP1 binding an unknown erythrocyte receptor [12]), the alternative-pathway PfRh/EBA ligands binding a range of known and unknown host receptors, PfRh5 ligand binding the erythrocyte basigin receptor and finally the AMA1 ligand binding to another parasite protein, RON2, that has been translocated into the erythrocyte membrane (Fig. 6) [22,30,39].
To investigate the functional order of invasion ligands, three parasite strains and one mutant derivative were characterized. Strain 3D7 was chosen because inhibition of invasion in this strain has previously been characterized with a variety of laboratory reagents including R1 peptide. Recently, the D10 strain has gained favour because its 48-hour cell cycle permits ease of use relative to other lines. To study EBA/PfRh function, the EBA175 dominant line, W2mef, was chosen since this ligand is easily blocked by NM treatment. A derivative of this line, the ΔEBA175 mutant was also assayed because its switch to dominant use of the CR1-dependent PfRh4 ligand can be blocked with soluble CR1 subunits. Before commencing our study of inhibition of invasion it was important to ensure the lines invaded with similar frequencies and kinetics, which they did. The exception was in the number of invasions per erythrocyte contact, where merozoites lacking EBA175 were only half as efficient as other strains tested. ΔEBA175 parasites also interacted with more erythrocytes prior to invasion than did other strains, thus overall invasion rates were comparable. In most parasites EBA175 is a dominant ligand with about a million copies of its receptor, glycophorin A, available for binding (reviewed in [45]). In contrast, there are only about a thousand molecules per erythrocyte of CR1, the receptor for PfRh4 and the dominant ligand in ΔEBA175 parasites. The relatively low number of CR1 molecules could result in lower avidity, with the parasites detaching more readily, both preventing invasion and allowing merozoites which are still viable to interact with other erythrocytes. Thus, both the reduction in efficiency of invasions per contact in ΔEBA175 parasites, and the increased number of erythrocytes contacted prior to invasion could be due to the abundance of the respective erythrocyte receptors. The difference in receptor abundance, however, does not translate into different growth rates between the ΔEBA175 mutant and its W2mef parent strain, both in our data presented here and in unpublished growth rate data (Lopatiki and Cowman, see Acknowledgments).
Of the receptor-ligand interactions that we inhibited, MSP1 and the EBAs/PfRhs reduced the ability of merozoites to deform their target erythrocytes and invade. In addition, cytD treatment of parasites, but not erythrocytes, inhibited deformation and invasion, indicating that function of the actin-myosin motor is also required to deform the erythrocyte during apical reorientation prior to the motor’s role in host cell internalization. MSP142 is a target of heparin [12], which almost completely ablated deformation. This block of progression to stronger deformation suggests both that MSP142 binding is responsible for weak deformation and that this leads to stronger deformation, reorientation and invasion. Although heparin is known to bind to other invasion ligands [46], because it inhibits an early step in invasion, MSP142 could be the main invasion blocking target of heparin [12]. Specific inhibition of MSP1 using an antibody could help validate its role in early deformation, but unfortunately, the only known MSP1 invasion inhibitory antibody blocks surface shedding and therefore would be expected to function further downstream as the merozoite passes through the tight junction [47]. Heparin did not stop merozoites binding to erythrocytes suggesting other ligands are responsible for primary attachment.
The weak deformation (scoring 1) mediated by MSP1 could be passive in nature, as GPI-anchored proteins such as MSP1, have fluid movement in membranes and, as they aggregate at the site of merozoite contact, could form a depression in the erythrocyte membrane. This “cap” of aggregated GPI-anchored ligands could hold the merozoite to the erythrocyte membrane but still allow the merozoite to rotate and move until the EBAs and PfRhs bind their respective partners. Blocking EBA175 and PfRh4 strongly inhibited both deformation and invasion. The observation that both EBA175 and PfRh4 function similarly, supports a long held view that these and most of the other EBA/PfRh ligands have redundant functions [48]. This, however, likely does not apply to all of these ligands since PfRh1 binding was recently shown to trigger parasite calcium signalling and EBA175 release [16], and a separate role for PfRh5, late in invasion, has been indicated by data presented here.
Strong deformation (scoring 2 and 3) that culminated in apical reorientation could be produced passively by a gradient of high affinity EBAs/PfRhs emanating from the merozoite apex or could be actively powered by the parasite’s actin-myosin motor interacting with the cytoplasmic tails of the EBAs/PfRhs. To discriminate between these we blocked the actin-myosin motor with cytD and the merozoites no longer deformed strongly, suggesting a role for the motor in powering deformation in addition to its role in internalization. Unexpectedly, the cytD treated merozoites could reorientate their apical ends onto the erythrocyte surface, release their rhoptries and presumably form a tight junction and trigger echinocytosis in the same time frame as untreated parasites. This indicates a gradient of EBA and PfRh proteins are passively responsible for reorientation and so what then is the function of strong deformation? We observed in cytD treated merozoites that they contacted more erythrocytes before finally selecting one with which to form a tight junction than did untreated parasites. Could the actin-myosin powered movement of EBA and PfRh molecules be causing deformation to help to embed the merozoite into the erythrocyte surface so it is less easily detached and more rapidly commits to invasion? Within an in vivo setting, strong binding, embedding into the host cell, and rapid host cell selection could be important to out-pace host immune responses and the shear forces associated with circulation.
Whilst attempting to determine the sequence of invasion events, some key new findings have emerged. We show that successful invasion correlates with preceding strong deformation suggesting that vigorous surface contacts might have a function in promoting invasion or triggering downstream signalling required for subsequent invasion steps. Variability in the time and amplitude of deformation may correspond to where the merozoite makes initial contact with its target erythrocyte. For example, if the merozoite contacts near its apical end, then shorter and shallower deformation may occur prior to apical reorientation than if the merozoite makes contact with its basal end. To validate this would require the apical end to be visibly discernible from the rest of the merozoite during reorientation, but unfortunately we could not distinguish this using brightfield illumination alone. The apical end was, however, evident with Ca2+ imaging after reorientation where an apical Ca2+ signal was observed.
When Fluo-4 AM dye was loaded into erythrocytes and added to unstained parasites, we observed a Fluo-4 flux at the apical end of the merozoite immediately prior to invasion. As erythrocytes have no known Ca2+ stores and the parasites were not labelled with Fluo-4 AM, this suggests that a Ca2+-containing organelle in the parasite is coming into contact with contents of the erythrocyte or that Ca2+ from the media is entering possibly via a pore. Previous observations indicate that rhoptry release still occurs when the AMA1–RON2 interaction is blocked [27] and here we show that when the AMA1–RON2 interaction is inhibited, the Ca2+ signal in the parasite still occurs. Further, when the parasite’s actin-myosin motor was blocked with cytD, the Ca2+ signal was still observed. Presence of the Ca2+ signal associates strongly with echinocytosis in untreated parasites as well as cytD and RON2–treated parasites. Together, this implicates that the rhoptries release factors that result in permeabilization of the host cell, Ca2+ entry and the triggering of echinocytosis.
That Ca2+ entry might be triggering echinocytosis by changing the cytoskeletal mesh was tested by performing invasions in BAPTA-AM treated erythrocytes to chelate the ion upon entry. Chelation of Ca2+ had little effect upon any aspect of invasion apart from slightly decreasing the time to echinocytosis contrary to expectations. Since the secretion of lipid-rich rhoptry contents have been observed to penetrate the host membrane and form whorls within its cytoplasm, even when invasion is blocked with cytD [49,50], these lipids could be responsible for echinocytosis. It has been proposed that materials which insert into the outer leaflet of a cell membrane will cause outward bending, leading to crenation of the cell [51]. The released rhoptry lipids would first make contact with the outer leaflet of the erythrocyte before penetrating as whorls into the cytoplasm. The excess material in the outer leaflet could cause transient crenation and echinocytosis of the host cell before flippases correct the asymmetry and return the echinocyte to its biconcave shape after several minutes, which is what we observe [38]. In RON2 and R1 blocked invasions we often observe that echinocytes fail to quickly return to their pre-invasion shape, even after 30 minutes which was the length of the filming period. It has been noted previously that in the presence of R1 peptide, rhoptry material was distributed over the outside of the erythrocyte membrane rather the penetrating it, since it was not confined within the boundary of the tight junction [27]. The deposition of the rhoptry’s full contents into the outer leaflet of the erythrocyte membrane could trigger great imbalance in bilayer components, which is why it could take longer for the erythrocyte to recover its normal shape.
One of our most interesting new findings was that when PfRh5–basigin was blocked, or when invasion was inhibited by absence of extracellular calcium, incidences of echinocytosis virtually stopped, suggesting that the rhoptries were not released. The fact that an apical Ca2+ signal was rarely observed in Fluo-4 labelled erythrocytes when the PfRh5-basigin interaction was blocked corroborates this and suggests that PfRh5–basigin binding mediates a pre-tight junction in a calcium-dependant manner that triggers rhoptry release. Calcium does not appear to be required for PfRh5 binding to basigin [52] and it may be required for some upstream or downstream function. As external calcium is required for proteases namely PfSUB2, that shed the merozoite surface coat [53] [54], we did not initially suspect rhoptry release would be affected. Discharge of the rhoptries releases the RON complex that then crosses to the erythrocyte surface in order to form a tight junction with AMA1 to allow subsequent invasion of the erythrocyte.
The Fluo-4 signal observed prior to and during invasion could be due to release of Ca2+ from the rhoptries and/or diffusion of erythrocyte Fluo-4 into the rhoptry neck region via a pore. Alternatively, media Ca2+ could be entering the erythrocyte at the invasion site. To discriminate between these, we tested for Fluo-4 signals in Ca2+ free media with the expectation that if a signal was observed it must be derived from the merozoite, or if absent, then the Ca2+ must be from media. The result however proved to be inconclusive for although no Fluo-4 signal was observed there appeared to be no release of rhoptries (because there was no echinocytosis) that would have provided a putative pore for entry of Ca2+ into the erythrocyte.
Recently it was reported that P. berghei parasites with their ama1 gene deleted grow at about a third the rate of wildtype parasites and can still penetrate their target erythrocytes and form a parasitophorous vacuole membrane (PVM) [55]. We concur with the latter because we have observed here that a few RON2 peptide treated merozoites still penetrated their host cells. Previously, we also observed penetration in AMA1 knock down Pf merozoites, however in both cases we did not observe merozoites transforming into rings after internalization, and there was a delay in the recovery of normal erythrocyte shape after echinocytosis [38]. We suspect, therefore, that there was a defect in erythrocyte resealing at the invasion site, which could be one of the additional functions of the AMA1-RON2 tight junction. Normal PVM formation in P. berghei Δama1 parasites is compatible with our own observations and those of others, namely that AMA1 function is not needed for the release of rhoptry contents to form the PVM [27,55].
Live cell imaging can reveal much about parasite behaviour and pathogenesis. Our work here sheds light on the specific roles and order of receptor-ligand interactions leading up to invasion, expanding our knowledge of the biology of Plasmodium. This work also lays the foundation for the development of a sequential-step invasion-blocking vaccine, which might provide a significant biological advantage by targeting proteins functioning at multiple sequential steps of invasion, rather than targeting multiple proteins that function at the same step.
Pf strains 3D7, D10 PfM3’ [56], W2mef and W2mefΔEBA175 were maintained in continuous culture as per [57]. Where indicated, parasites were also cultured in Dulbecco’s Modified Eagle Medium (DMEM, Gibco®) supplemented with L-glutamine (Sigma) and Albumax II (Invitrogen). Parasites were synchronised using sorbitol and heparin treatments as described previously [58].
Highly synchronous parasite cultures at 4% hematocrit were diluted to 0.16% in RPMI media or washed in Ca2+ free DMEM+EGTA media (see below) and 2 mL of this was allowed to settle to produce a monolayer onto a 35 mm Fluorodish (World Precision Instruments). Custom dishes holding smaller 25 or 50 μL volumes were used for experiments containing invasion inhibitory antibodies. All live-cell experiments were performed at 37°C on a Zeiss AxioObserver Z1 fluorescence microscope equipped with humidified gas chamber (90% N2, 1% O2, and 5% CO2). Late stage schizonts were observed until they looked ready to rupture (described in [7]) and time-lapse videos were recorded with a AxioCam MRm camera usually at 4 frames per second. ImageJ and Prism (Graphpad) were used to perform image and statistical analyses. For data sets with normal distribution (Figs 1E-G, 4C, and S14C) an unpaired t test was used. For data sets without normal distribution (Figs 1B-D, 3A, 4A&B, S6 and S14A&B) the Mann-Whitney test was used. For comparison of deformation scores between groups (Figs 2, 3B&C and 5E) a Chi-square analysis was performed. A value of p≤0.05 was used as the determinant of statistical significance for all tests.
To score the degree of merozoite deformation of the erythrocyte surface we initially attempted to develop software to automate the process but were unable to. Instead, we resorted to scoring deformation by eye using a simplified deformation score and multiple trained scorers. We defined contact between the merozoite and erythrocyte as interactions where the erythrocyte and merozoite maintained physical contact with each other for two frames (0.25 seconds) or longer. The vast majority of contacts which we observed lasted considerably longer than 0.25 seconds. Interactions where the merozoite and erythrocyte appeared to touch for a single frame were not counted unless there was evident deformation, indicating clear contact, however these were very rare events. A deformation score of: 0 = sustained contact but no deformation, 1 = weak deformation at the point of contact, 2 = strong deformation with the erythrocyte membrane extending up sides of the embedded merozoite, and effects of deformation no longer strictly local to the merozoite, and 3 = extreme deformation with the deeply embedded merozoite partially covered by the erythrocyte membrane and extremely strong deformation distant from the merozoite to the point of distortion of erythrocyte borders (Fig. 2A, B).
Invasion-blocking reagents were used as follows; Heparin 100 μg/ml (Sigma), 100 μg/mL R1 peptide (Mimotopes) [35], 10 μg/mL RON2 peptide (LifeTein LLC) [36], 1 μg/mL cytochalasin D (Sigma), 10 μg/mL anti-basigin mAb MEM-M6/6 [22], and 10 mg/mL rabbit anti-PfRh5 [25]. Calcium within erythrocytes was chelated using BAPTA-AM (Sigma). The stock BAPTA-AM prepared in DMSO was added to erythrocytes at 0.16% hematocrit to achieve the final concentration of 50 μg/mL. The treated erythrocytes were incubated at 37° for 30 min, and then were washed three times. To inhibit merozoite invasion in W2mef ΔEBA175 parasites 0.2U/mL neuriminidase (Sigma) was added directly to the parasitized erythrocytes for 30 mins at 37°C and 10 μg of CR1 CCP1–3 was added prior to microscopy [20,39].
Control invasion assays were performed in complete RPMI or DMEM media (Life Technologies). For Ca2+ free experiments Ca2+ free DMEM was used (Life Technologies), both in its standard form and modified with 2.5 mM EGTA since Ca2+ free RPMI was not commercially available. The media were supplemented with 25 mM HEPES and dialysed calcium free Albumax II to 0.5% and prior to use all the media were supplemented with 0.2% NaHCO3. The invasion experiments were conducted with two transfected 3D7 parasites lines expressing secreted or exported Nanoluciferase [59]. Late stage magnet purified schizonts (~2 × 108) were split into four equal portions and each was washed in 1000 volumes of either RPMI, DMEM, Ca2+ free DMEM or Ca2+ free DMEM+EGTA. In parallel, uninfected erythrocytes were washed the same media types and then each batch of parasites and erythrocytes of the same media type were mixed together to final hematocrit of 2%. Each of the fractions was then split into two and heparin (Sigma) at a final concentration of 100 μg/mL was added to one half of the parasites to block invasion (Boyle et al 2010). The parasites were then incubated at 37°C for 90 mins to permit some invasion to occur. The infected erythrocytes were then pelleted at 3000g and a small amount of media retained to measure luciferase activity to ensure that egress and Nanoluciferase release had occurred. The parasites treated with 5% sorbitol to lyse the remaining schizonts and merozoites. The sorbitol was then replaced with complete RPMI media and the new ring stage parasites were grown in triplicate 100 μL cultures in a 96 well plate for 24 and 72 hrs to amplify the invasion signal. To assay parasite growth at 2, 24 and 72 hrs post invasion, 5 μL of the culture was lysed in 50 μL of water containing 1/1000 NanoGlo substrate (Promega). Relative light units (RLU) were measured in a FLUOstar Omega Luminometer (BMG Labtech) for 2s with the gain set to the brightest well and adjusted to 10% below saturation. To derive the invasion signal, the RLU from the heparin blocked invasion samples were deduced from their corresponding heparin free samples.
Fluo-4 AM (Life Technologies) was added to a final concentration of 10 μM to erythrocytes diluted in incomplete RPMI with no albumax or serum at 2% hematacrit. The erythrocytes were incubated at 37° for 1 hr and washed 3x in 10 volumes in complete RPMI or Ca2+ free DMEM+EGTA before adding to a 35 mm Fluorodish as described above. Magnet purified late schizont stage parasites was added to the dish. The parasites were imaged with low power brightfield conditions (2–3 volts) until rupture occurred whereupon alternative brightfield and fluorescence images with the GFP filter set were recorded. To limit photodamage each 35 mm dish was only imaged four times, once per quarter, and the 50 μL dishes were prepared fresh for each movie made.
|
10.1371/journal.ppat.1002831 | Structural Analysis of Specific Metal Chelating Inhibitor Binding to the Endonuclease Domain of Influenza pH1N1 (2009) Polymerase | It is generally recognised that novel antiviral drugs, less prone to resistance, would be a desirable alternative to current drug options in order to be able to treat potentially serious influenza infections. The viral polymerase, which performs transcription and replication of the RNA genome, is an attractive target for antiviral drugs since potent polymerase inhibitors could directly stop viral replication at an early stage. Recent structural studies on functional domains of the heterotrimeric polymerase, which comprises subunits PA, PB1 and PB2, open the way to a structure based approach to optimise inhibitors of viral replication. In particular, the unique cap-snatching mechanism of viral transcription can be inhibited by targeting either the PB2 cap-binding or PA endonuclease domains. Here we describe high resolution X-ray co-crystal structures of the 2009 pandemic H1N1 (pH1N1) PA endonuclease domain with a series of specific inhibitors, including four diketo compounds and a green tea catechin, all of which chelate the two critical manganese ions in the active site of the enzyme. Comparison of the binding mode of the different compounds and that of a mononucleotide phosphate highlights, firstly, how different substituent groups on the basic metal binding scaffold can be orientated to bind in distinct sub-pockets within the active site cavity, and secondly, the plasticity of certain structural elements of the active site cavity, which result in induced fit binding. These results will be important in optimising the design of more potent inhibitors targeting the cap-snatching endonuclease activity of influenza virus polymerase.
| The 2009 influenza pandemic, the on-going potential threat of highly pathogenic H5N1 avian strains and the widespread occurrence of resistance to current anti-influenza drugs targeting the neuraminidase or the M2 ion channel, all highlight the need for alternative therapeutic options to treat serious influenza infections in the absence of protection by vaccination. The viral polymerase, which performs transcription and replication of the RNA genome, is an attractive target for novel antiviral drugs since potent polymerase inhibitors will directly stall replication. The heterotrimeric polymerase performs transcription by a unique cap-snatching mechanism, which involves host pre-mRNA cap-binding and endonucleolytic cleavage by the PB2 and PA subunits respectively. Crystal structures of both the PB2 cap-binding and PA nuclease domains are now available allowing structure-guided optimisation of cap-snatching inhibitors. Here we present a series of co-crystal structures of the 2009 pandemic H1N1 PA endonuclease domain that reveal the binding mode of several known endonuclease inhibitors. All inhibitors chelate the two manganese ions in the active site of the nuclease but different extensions to the metal binding scaffold bind in distinct sub-pockets of the active site cavity. These results highlight the value of structure-based approaches to the development of more potent influenza polymerase inhibitors.
| Influenza virus replicates in the nucleus of infected cells where the heterotrimeric viral RNA-dependent RNA polymerase, with subunits PA, PB1 and PB2, is responsible for replication and transcription of the single-stranded viral RNA genome (vRNA). Transcription of viral mRNAs occurs through an unusual ‘cap-snatching’ mechanism [1] which has only been reported for negative strand, segmented RNA viruses, including orthomyxoviruses (notably influenza), bunyaviruses and arenaviruses. For influenza, cap-snatching involves the binding of host cell pre-mRNAs via their 5′ cap structure to the PB2 subunit of the polymerase followed by cleavage at nucleotides 10–13 by an endonuclease activity which resides in the PA subunit of the polymerase. The short capped oligomers then serve as primers for transcription of the viral mRNAs by the PB1 subunit of the polymerase. The viral transcripts are poly-adenylated by a stuttering mechanism at a conserved U-rich region of the template vRNA [2]; thus the viral mRNAs have both the 5′ and 3′ signals to be competent for translation after nucleo-cytoplasmic export. In the last few years, crystal structures of the two functional domains involved in cap-snatching have been determined (reviewed in [3]). The cap-binding domain resides in the central region of the PB2 subunit and has a unique fold while still binding the m7G ligand by means of an aromatic sandwich, similar to other cap-binding proteins [4]. The endonuclease domain is at the N-terminus of the PA subunit and has a core fold similar to other two-metal dependent nucleases of the PD…D/E…K superfamily [5]–[6]. Indeed, the isolated, recombinant endonuclease domain has divalent cation dependent, in vitro nuclease activity with a strong preference for manganese ions, consistent with the much tighter binding of manganese than magnesium [7]. Since transcription by cap-snatching is essential for virus replication, inhibition of either the cap-binding, endonuclease or polymerase activities are all potential means of anti-viral therapy and indeed each of these targets have been or are being actively pursued [8]–[10]. Indeed combination therapy targeting more than one of the polymerase active sites is an attractive possibility. Here, we exploit the availability of the endonuclease crystal structure to provide the first detailed structural information on specific inhibitor binding to the influenza polymerase.
The need for new therapeutic options targeting influenza virus is now widely recognised. This follows recent developments, such as the on-going circulation of highly pathogenic avian H5N1 strains, which could potentially adapt for human-to-human transmission [11], the unexpected emergence of the 2009 H1N1 pandemic strain [12], which was highly contagious and thus spread rapidly around the world, but was fortunately not so virulent, and the development of resistance in wild-type strains to currently available anti-viral drugs targeting the neuraminidase or M2 ion channel [13]. These have all highlighted the vulnerability of the world population to novel influenza strains for which there may be no vaccine for several months and a limited variety of resistance prone anti-viral drugs [14]. The cap-snatching endonuclease of influenza virus polymerase has been targeted for anti-influenza drug development since the 1990s because its inhibition would directly stall viral transcription and hence replication. Firstly, a number of 4-substituted 2,4-dioxobutanoic acid compounds that specifically inhibit influenza polymerase endonuclease activity with IC50s in the range 0.2 to 29 µM were identified by Merck [15]–[17]. Subsequently, a substituted 2,6-diketopiperazine natural compound (Flutimide) from the fungus Delitschia confertaspora, and its derivatives, were shown to inhibit endonuclease activity and influenza A and B virus replication in cell culture [18]. Bristol-Myers Squibb identified N-hydroxamic acid and N-hydroxyimide compounds that inhibit the endonuclease [19]. Roche discovered a new class of endonuclease inhibitors, with IC50s down to 3 µM based on considerations of the likely divalent cation binding properties of the enzyme [20]. More recently other compounds have been shown to inhibit the endonuclease, for instance green tea catechins [21], phenethylphenylphthalimide analogues derived from thalidomide [22] and macrocyclic bisbibenzyls [23].
The X-ray structure of the A/Victoria/3/1975(H3N2) endonuclease [5] was obtained with a crystal form in which crystal contacts blocked access to the active site. Despite many trials, no crystal structures have been obtained of H3N2 endonuclease with bound inhibitors, substrate or product analogues, severely limiting the possibility for structure-based inhibitor optimisation. To overcome this problem, we investigated whether the endonuclease from other influenza strains having slightly different amino acid sequences might yield more useful crystal forms. We found that a construct comprising residues 1–198 of A/California/04/2009(H1N1) (2009 pandemic strain, pH1N1), expressed from a synthetic gene, readily crystallised with and without relevant ligands. In the residue range 1–198 of the construct, the pH1N1 sequence differs in 12 positions from that of H3N2; however the active site and close vicinity are identical between the two strains. Here we report crystal structures of the pH1N1 influenza endonuclease complexed with two divalent metal ions and four 4-substituted 2,4-dioxobutanoic acid inhibitors [15]. The compounds are 2,4-dioxo-4-phenylbutanoic acid (DPBA), 4-[3-[(4-chlorophenyl)methyl]-1-(phenylmethylsulpho)-3-piperidinyl]-2-hydroxy-4-oxo-2-butenoic acid (denoted R05-01), 4-[1-cyclohexylmethyl-4-(p-chlorobenzyl)piperidin-4-yl]2,4-dioxobutanoic acid (denoted R05-02) and 4-[N-benzyl-3-(4-chlorobenzyl)piperidin-3-yl]2,4-dioxobutanoic acid (Merck L-735,882; here denoted R05-03) (Table 1). These compounds are reported to have IC50s of 21.3, 0.19, 0.33 and 1.1 µM respectively in an in vitro cap-dependent transcriptase assay [15]. A structure is also presented with bound (-)-epigallocatechin gallate (denoted EGCG, Table 1) from green tea. Two additional structures are presented of the pH1N1 endonuclease with bound mononucleotides rUMP and dTMP, components of the RNA (and DNA) endonuclease substrate. In an accompanying paper, structures of a complementary set of inhibitors bound to the endonuclease of strain A/Vietnam/1203/2004 (H5N1) are reported [24].
All these structures show in detail how these compounds bind directly to the metal ions as well as interacting with a number of residues in the active site, some of which change conformation upon ligand binding. This three-dimensional knowledge of the ligand interacting residues and the regions of plasticity of the active site is critical for the optimised design of modifications to existing inhibitors to improve their potency or for structure based design and optimisation of novel inhibitors that effectively block endonuclease activity.
Structures are now known of the PA-Nter from H3N2 [5], H5N1 [6] and pH1N1 (this work). There are respectively 5 and 12 differences between the pH1N1 sequence compared to the avian H5N1 and human H3N2 strain sequences, at a total of 15 positions (Figure 1A). It is not known whether any of these differences play a role in inter-species transmission or virulence. The most structurally variable region, which is also a hotspot for sequence variation (Figure 1A), is between residues 55–66. This forms part of a mobile inserted element (residues 53–73) of unknown function that is solvent exposed and usually disordered. In the H3N2 and pH1N1 structures, this element is well defined in some chains in the asymmetric unit and shows a significantly different orientation between the two strains (Figure 1B). However this could be due to different crystal contacts since in the H3N2 structure, loop residue Glu59 interacts with a divalent cation in the active site of a neighbouring molecule, thus blocking access to the active site in this crystal form (Figure S4 in [5]). To avoid the problems associated with this flexible loop, two of the six structures described below (those of R05-1 and EGCG) were determined using a truncated form of the protein in which residues 52–64 were replaced by a single glycine (Δ52–64:Gly). In an accompanying paper it is shown that an even larger loop deletion (Δ51–72:Gly-Gly-Ser) does not significantly affect enzymatic activity [24]. Apart from this, the three PA-Nter structures from different strains are overall very similar, with only small differences in helix orientation and loop conformation (Figure 1B).
Previous structural studies on H3N2 endonuclease showed that there are two divalent cation binding sites in PA-Nter: Site 1 is coordinated by His41, Asp108, Glu119 and the carbonyl oxygen of Ile120 and site 2 by Glu80 and Asp108 [5]. However as mentioned above, the active site structure of H3N2 PA-Nter could have been influenced by crystal contacts. Subsequent in vitro studies showed that manganese ions bind strongly and exclusively to site 1 and that both manganese and magnesium ions can bind at site 2 albeit with significantly lower affinity [7]. A similar situation occurs in the related La Crosse bunyavirus endonuclease [25]. As the exact in vivo situation is unknown, all in vitro work reported here, notably crystallization and in vitro activity assays, was performed in buffer containing both 2 mM MnCl2 and 2 mM MgCl2. In the unliganded pH1N1 structure at 2.1 Å resolution (Tables 1), there is excellent definition of the solvated bi-metal binding site in each of the four crystallographically independent active sites. Both metal ions have octahedral co-ordination: site 1 by His41, Asp108, Glu119, Ile120 and two water molecules (W4 and W5) and site 2 by Glu80 and Asp108 and four water molecules (W1, W2, W3 and W4) (Figure 1C). Whereas site 1 refines as a fully occupied manganese ion, the metal in site 2 has weaker electron density and thus could correspond to a magnesium ion or partially occupied manganese. Indeed, refinement with a fully occupied magnesium ion in site 2 gives B-factors similar to that of the manganese ion in site 1. Since the anomalous signal in the original dataset was poor, we attempted to clarify this assignment by measuring additional crystallographic data at a longer X-ray wavelength of 1.55 Å, where the putative manganese anomalous signal is higher. In this 2.6 Å resolution dataset, a strong anomalous signal is observed in site 1 in each of the four active sites (between 7.4 and 10.1σ, data not shown), whereas the electron density is much weaker or even absent for site 2 and has no significant anomalous scattering. These results provide structural confirmation of our previous biochemical studies that indicated strong binding of manganese to site 1 and weaker binding of either magnesium or manganese to site 2 [7].
All diketo compounds and ECGC were tested for their in vitro nuclease inhibitor and in cellulo antiviral activities by a fluorescence resonance energy transfer (FRET) assay and cell viability assay, respectively (Table 2). An alternative in vitro fluorescence polarization assay for nuclease inhibition has recently been described elsewhere [26]. For the FRET assay, a single stranded RNA oligonucleotide labeled with an emitter and quencher fluorophore at opposite ends was incubated with A/Victoria/3/1975(H3N2) PA-Nter with and without inhibitors. RNA cleavage was monitored by the increase in fluorescence when the quencher is released from the emitter (Figure S1A). IC50 values of 2.7 µM for DBPA, 1.9 µM for EGCG, 1.1 µM for R05-03, 0.13 µM for R05-01, and 0.06 µM for R05-02 were obtained with this method (Figure S1B). In the cell viability assay, in which MDCK cells were infected with influenza virus and treated with endonuclease inhibitor at the same time, most compounds could inhibit virus replication thereby preventing virus induced cytopathicity and restoring cell viability compared to a virus infected control sample. DBPA did not show any inhibitory effect whereas IC50 values of 19.9 µM for R05-03, 20.4 µM for R05-01, 15.9 µM for R05-02 and 1.1 µM for EGCG were obtained (Table 2). EGCG and R05-3 showed some meaurable cytotoxicity (Table 2). The 15–20-fold lower IC50 for EGCG in this assay compared to the diketo compounds R05-01 and R05-02, whereas in the nuclease inhibition assay, the reverse was true, might be due to the different physicochemical properties, and hence cell availability, of the substances tested. It should be noted that the systematically lower IC50 values quoted in the introduction for the diketo compounds referred to an in vitro transcriptase assay, not an anti-viral assay [15]. In addition to the functional assays, the effect of the diketo inhibitors on the thermal stability of the endonuclease was tested by a Thermofluor assay in which a hydrophobic fluorophore has little affinity for native proteins but binds to denatured proteins, leading to an increase of fluorescence [27]. The apparent melting temperature (Tm) of denaturation can be obtained from the temperature dependence of the fluorescence (Figure S1C). It has previously been shown that PA-Nter is significantly thermally stabilized by divalent cation binding and, even more so, by DPBA binding [5], [7]. Here we report Tm values of 53.5, 65, 69, 71 and 69°C for no ligand, DPBA, R05-01, R05-02 and R05-03 respectively (Table 2). This confirms that R05-01, R05-02 and R05-03 all chelate the cations in the active site of the endonuclease and enhance the thermal stability even more than for DPDA, presumably by making additional stabilizing interactions with the protein.
The DPBA bound structure was determined at 2.3 Å resolution in the same C2 space-group as the native protein (Tables 1 and 3) and is shown in Figure 1D (electron density in Figure S2A). As with the native structure, anomalous scattering confirms the presence of manganese in site 1 (Figure S2A), whereas the metal in site 2 has a higher B-factor and no significant anomalous scattering. As expected, DPBA binds directly to the two cations bound in the active site. Compared to the unbound state, metal ion coordinating water molecules W3, W4 and W5 are replaced by oxygens from the ligand (Figures 1C,D). An identical configuration was observed for DPBA binding to the active site of bunyavirus cap-snatching endonuclease [25]. The catalytic Lys134 makes an electrostatic interaction with the carboxyl-group of DPBA, which also interacts with the hydroxyl of Tyr130 via a bridging water molecule. Ligand stabilisation of the metal binding to the active site almost certainly explains why DPBA supershifts the endonuclease thermal stability [5]. The phenyl ring of the DPBA is less well-defined than the rest of the molecule as its rotational conformation is only weakly stabilized by partial stacking with the side-chain of Arg84. An identical conformation of DPBA bound to A/Vietnam/1203/2004 (H5N1) PA-Nter is reported in the accompanying paper although this structure contains a second DPBA molecule in the active site stacking against the first, probably as a result of the high concentration used [24].
As summarised in Table 3, co-crystals with compounds R05-2 and R05-3 were obtained with native protein in the P212121 space-group with four molecules in the asymmetric unit. For R05-01, crystallisation was successful with the Δ52–64:Gly truncation mutant in the P6222 space-group. All diketo inhibitors co-ordinate the two metal ions in the same manner as described for DPBA and refinement is consistent with the presence of two bound manganese ions as confirmed by anomalous scattering (Figures S2B, S2D and S2E). In general, for each of the diketo compounds, the ‘arms’ have higher B-factors than the metal binding, diketo moiety of the ligand, suggesting flexibility due to sub-optimal interactions of the arms. Indeed, bound R05-03 is observed in two different conformations, corresponding to different rotamers of the chlorobenzene, in respectively chains A, B (denoted conformation 3A, Figure S2B) and C, D (denoted conformation 3D, Figure S2C) in the asymmetric unit. Conformation 3A has stronger electron density than 3D. For R05-02 electron density for the arms is not well-defined in all four copies in the asymmetric unit, although the configuration of the compound is unambiguous and consistent in each copy (Figure S2D). As discussed in more detail below, the two ‘arms’ of each diketo compound sample different sub-pockets of the active site cavity.
Figures 2A–D show a comparison of the binding mode of R05-01, R05-02, R05-03A and R05-03D in the active site indicating nearby residues. Most interactions of the compound arms are with residues in the range 24–38 (notably Tyr24, Glu26, Lys34, Ala37 and Ile38), which comprise a flexible loop leading from the C-terminal end of helix α2 into the N-terminal half of helix α3, but Arg84 and Phe105 are also involved in some cases. Due to the hydrophobic and aromatic nature of the arms, most interactions are van der Waals or stacking and there are no polar interactions.
Epigallocatechin 3-gallate (EGCG) is the ester of epigallocatechin and gallic acid and is the most abundant catechin in green tea. EGCG, a polyphenol with antioxidant properties, has been extensively investigated as a possible antiviral or anticancer compound [28]–[29]. It has recently been reported that EGCG inhibits the influenza endonuclease [21]. Co-crystallisation of the pH1N1 PA-Nter Δ52–64:Gly truncation mutant with EGCG gave a new crystal form diffracting to 2.6 Å resolution (Table 1,3). The compound was clearly observed in the active site as well as anomalous scattering peaks corresponding to the two manganese ions (Figure S3A). Strong extra density also exists around a 2-fold crystallographic axis and represents another EGCG molecule non-specifically trapped by crystal packing. The conformation and placement of the EGCG in the active site is shown in Figure 2E with more details of the interactions shown in Figure S3B. The two manganese ions are co-ordinated by two of the hydroxyls of the gallo-group, whilst the galloyl-group is orientated towards helix α3, stacking on Ala37 and Ile38 and hydrogen bonding to the carbonyl oxygen of Val122. The planes of the gallo- and galloyl-phenyl groups are parallel but not significantly overlapped. The double ring of EGCG is orientated towards the preceding loop, with notably the resorcinol moiety stacking on Tyr24 and making a hydrogen bond to Glu26. However three of the eight hydroxyl groups of EGCG do not make direct interactions with the protein. The configuration of the EGCG in the active site is quite different from that previously proposed by docking studies [21], [23].
Co-crystallisation trials were attempted with all four deoxy- and oxy- mononucleotides, to mimic putative substrate binding by the endonuclease. The only compounds that resulted in structures were dTMP and rUMP, both of which gave large, well-ordered crystals in a new orthorhombic space-group (Tables 1 and 3). Apart from the obvious differences in the ribose and base, the two structures are essentially the same. In both cases, clear anomalous scattering exists for the two manganese ions (Figure 3A) and the nucleotides bind with two oxygens of the phosphate completing the co-ordination sphere of Mn1, one of them also coordinating Mn2 (Figure 2F, 3A, 3B). The base is well stacked on Tyr24 and Lys34 makes a hydrogen bond to the O2 position. The ribose is stacked on Ala37 and Ile38 of helix α3 and the hydroxyl groups do not make hydrogen bonds to the protein. This is consistent with the fact that the protein is a DNAase as much as an RNAase [5], [30]. The conformation we observe for rUMP is quite different from that previously published (PDB entry 3hw3 [31]). The latter structure was obtained by soaking nucleotides into existing crystals of the endonuclease in the absence of manganese and the electron density is very poor. In this structure, a water molecule replaces Mn1 and a magnesium ion replaces Mn2. This difference in metal ligation is reflected in the altered conformation of Glu119. The ribose and base positions are quite different from in our structure and unable to interact with Lys34 or Tyr24 (for comparison of the structures see Figure S4). We suspect that the differences between the two structures reflect firstly the lack of manganese and secondly the fact that soaking pre-grown crystals does not allow the active site to adapt to the ligand as is more likely the case for co-crystallisation.
Unlike some of the diketo inhibitors, dTMP/rUMP exhibits a very well defined, full occupancy binding mode. The apparent optimisation of this binding might reflect its biological significance as representing part of the natural nucleic acid substrate binding site. It has previously been shown that superposing the active sites of PA-Nter with EcoRV restriction enzyme closely overlaps the metal binding centre and catalytic lysine (see [5]). To examine this further, we superposed various complexes of EcoRV with bound substrate or product dsDNA complexes. As seen in Figure 3C, the bound rUMP most closely mimics the position of the post-cleavage nucleotide as observed in EcoRV PDB structure 1STX [32]. A preference for uridine in the natural substrate at the post-cleavage position has not been reported before, although it has been proposed that in infected cells cleavage of donor pre-mRNA preferentially occurs after Cyt-Ade [33] or, alternatively, Gua-Cyt [34]. Further biochemical and structural work is clearly required on PA-Nter substrate or product complexes to advance understanding of any intrinsic sequences preferences of PA-Nter and the exact mechanism of cleavage.
The active site cavity of the endonuclease is quite voluminous, presumably because it has to accommodate at least two nucleotides either side of the cleavage site, with the manganese ions at its back (Figure 4A). As shown in the superposition of Figure 5A, the metal chelating moiety of the three diketo and EGCG inhibitors binds in a similar orientation to the manganese ions (although compound R05-2 is slightly tilted) but the two ‘arms’ of each compound are inserted into combinations of different active site ‘pockets’, denoted pockets 1 to 4 (see also Figure 4B–D). R05-1 has a similar configuration to R05-3D, with the two arms occupying pockets 2 and 3. R05-3A occupies pockets 2 and 4. Compound R05-2, which differs from R05-1 and R05-3 in the point of substitution on the piperidinyl ring (Figure 2) occupies pocket 3 and, uniquely, pocket 1. The green tea compound EGCG and the mononucleotides occupy pockets 3 and 4. Pocket 1 allows stacking on Phe105 and is uniquely observed with the chlorobenzene of R05-2. Pocket 2 is characterised by stacking on the side-chain of Arg84 (e.g. benzene of R05-3A, R05-3D and R05-1). Pocket 3 is characterised by stacking on Tyr24 (chlorobenzene of R05-3D, R05-1, cyclohexane of R05-2, resorcinol moiety of ECGC, base of rUMP/dTMP). Occupation of pocket 4 is characterised by stacking on Ile34 (e.g. chlorobenzene of R05-3A, galloyl-group of EGCG, rUMP/dTMP ribose). These various stacking options may reflect the need for similar interactions with the bases of the RNA substrate, as also suggested by the observed stacking of rUMP on Tyr24.
Figure 5B illustrates the conformational changes and ordering that occur upon R05-03 binding, in particularly of the loop around Tyr24, which is poorly ordered in the native structure. Indeed, Tyr24 is observed to be in three rotamers depending on which of the four pockets are occupied (Figure 5C). Side-chains of other residues (e.g. Arg84, Lys34, Glu26, Phe105) also change depending on which pockets the ligand is occupying. This indicates a plasticity of the active site and an induced fit mode of ligand binding, which is a complication that needs to be taken into account in any in silico screening for putative inhibitors. A second important conclusion for designing more potent inhibitors is to ensure that the extensions (‘arms’) to any ion-binding scaffold optimise interactions in one or more pockets. Imperfect stacking and lack of polar interactions will lead to residual flexibility and sub-optimal potency. This seems to be the case for the three diketo compounds, which do not exhibit very well ordered, full occupancy binding modes. On the other hand the binding mode of EGCG is well-defined although possible interactions with some of the hydroxyls in the compound are not fully exploited.
The question of likelihood of resistance is of critical importance in any consideration of anti-viral compounds. This is especially true of influenza viral protein targets which exhibit extensive host and strain dependent sequence variations. For example, in PA-Nter, natural sequence variants are reported in at least 20 of the 200 positions (Figure 1A, see also http://www.ncbi.nlm.nih.gov/genomes/FLU/Database/nph-select.cgi?go=database). On the other hand we note that all the principle residues interacting with the various compounds described here (Tyr24, Glu26, Lys34, Ala37, Ile38, Arg84, Phe105, Tyr130 and Lys134) are highly conserved amongst all influenza A strains, although four of them are substituted in influenza B strains (Phe24, Met34, Asn37, Tyr105), two of them (positions 34 and 37) non-conservatively (Figure 1A). Thus it is likely that there are severe constraints against mutation of these residues whilst retaining carefully tuned enzymatic activity. However, this needs to be explored further given the experience with neuraminidase inhibitors, which also target a conserved enzymatic active site, and to which resistance has naturally emerged.
There is currently very little information available on resistance mutants to polymerase inhibitors partly because they have not been in clinical use. However, it is interesting to note that a mutant was selected by serial passage of influenza strain A/PR/8/34 that is 2–3 times less susceptible to an inhibitor denoted L-742,001 that is very similar to R05-2 but with the cyclohexane group replaced by a benzene group [17]). L-742,001 is reported to have an IC50 of ∼4 µg/mL in a plaque assay. A/PR/8/34 has a threonine at position 20 (Figure 1A) and this became an alanine in the partly resistant mutant. In fact, position 20 is an alanine in the pH1N1 strain used in this work and in most recently circulating A strains. These observations can be explained by the fact that the cyclohexane group of R05-2 that stacks on Tyr24 in pocket 3 is also in close proximity to the side-chain of Ala20 which precedes Tyr24 by one turn in helix α2 (Figure 2C). The methyl-group of a threonine at position 20 would likely enhance the van der Waals contact with the equivalent benzene moiety of L-742,001 by prolonging the hydrophobic platform formed by Tyr24, thus slightly increasing the affinity for the inhibitor. The absence of such additional stabilising interactions probably explains why the electron density for this arm of R05-2 is generally weak. In the accompanying paper the crystal structure of the complex of L-742,001 with loop-deleted A/Vietnam/1203/2004 (H5N1) PA-Nter is reported [24]. Surprisingly, the configuration of L-742,001 is quite different from that we observe for R05-2. The chlorobenzene arm of L-742,001 is rotated 180° to coincide with cyclohexane arm of R05-2 and the benzene arm of L-742,001 enters a different pocket (denoted pocket 5). This orientation of L-742,001 is incompatible with the electron density of R05-2. Despite this difference, which is reminiscent of the promiscuity observed for R05-3, similar conclusions about the role of position 20 in modulating affinity for L-742,001 have been drawn [24], since both compounds have arms entering pocket 3. This example illustrates how detailed structural knowledge about the mode of inhibitor binding, combined with the extensive database of variation in influenza viral proteins, will be extremely useful in designing new inhibitors minimally susceptible to resistance at least from natural mutants known to be viable.
The 2009 H1N1 influenza pandemic [12], the on-going threat to humans of highly pathogenic H5N1 avian influenza viruses [11] and the widespread occurrence of resistance to current anti-influenza drugs [13] has highlighted the need for alternative therapeutic options to treat influenza infections when vaccines are unavailable [14]. Influenza virus can potentially be targeted by antiviral drugs at numerous points in its infectious life cycle [8]. The unique and essential cap-snatching mechanism of transcription by influenza virus polymerase, and in particular, the endonuclease activity, has long been recognised as a good target for antiviral drug development since, firstly, its inhibition could directly stall viral replication at the primary transcription level, secondly, the relevant active sites are likely to be highly conserved across strains and thirdly, the mechanism has no host cell counterpart [15]. Over a nearly twenty year period a number of specific inhibitors of the endonuclease activity have been published (see introduction) although, apparently, none have sufficiently potent anti-viral activity to have entered clinical development. It is interesting to note, however, that the HIV integrase inhibitor Raltegravir, now in clinical use, is also a diketobutanoic acid derivative and targets a two cation containing active site with some similarities to the influenza endonuclease [35]–[37].
The results presented here and in the accompanying paper [24] provide the first high-resolution structural information showing the different binding modes of distinct small molecule metal chelating scaffolds to the active site of the PA endonuclease domain of influenza polymerase. The active site cavity provides multiple distinct pockets capable of accommodating specific extensions to basic metal binding scaffold. However the endonuclease inhibitors analysed here each demonstrated sub-optimal utilisation of the available binding pockets and no one inhibitor sampled all available binding pockets. Furthermore the plasticity of certain regions of the active site cavity, notably the loop containing Tyr24 resulting in induced fit binding by most of the inhibitors, for instance to promote stacking of Tyr24 on aromatic moieties of the compounds. These considerations will be important in guiding modelling and medicinal chemistry approaches to optimization of lead compounds for more efficient inhibition of PA endonuclease. Together with additional endonuclease-inhibitor crystal structures and taking into account known sequence variations that could cause resistance, this will significantly advance the goal of developing novel and efficacious influenza therapeutics that directly target viral replication.
The DNA coding for PA-N-ter (residues 1–198) from A/California/04/2009-pH1N1 was synthesized and sub-cloned in the expression vector pESPRIT002 (EMBL) by GeneArt, (Regensburg, Germany). The sequence was designed to contain an MGSGMA polypeptide linker between the tobacco etch virus (TEV) cleavage site and the N-terminus to obtain 100% cleavage by TEV protease, as used previously [5]. The sequence of A/California/04/2009-H1N1 used is:
mgsgma(1)MEDFVRQCFNPMIVELAEKAMKEYGEDPKIETNKFAAICTHLEVCFMYSDFHFIDERGESIIVESGDPNALLKHRFEIIEGRDRIMAWTVVNSICNTTGVEKPKFLPDLYDYKENRFIEIGVTRREVHIYYLEKANKIKSEKTHIHIFSFTGEEMATKADYTLDEESRARIKTRLFTIRQEMASRSLWDSFRQSERGE -198
To potentially improve crystallisation properties, a deletion of part of the flexible loop (52–73) was engineered by site directed mutagenesis. For this, a PCR amplification of the whole vector containing the wild type gene was performed using two primers flanking the mutation site, one of them phosphorylated, and TurboPfu polymerase (Stratagene). Subsequently, template vector was digested with DpnI (New England Biolabs) and the mutated vector was re-ligated. In the mutant amino acid sequence 52–64 (HFIDERGESIIVE) was replaced by a single glycine.
Wild type and mutant plasmids were transformed to E. coli BL21(DE3) (Stratagene) and the protein was expressed in LB medium overnight at 20°C after induction at an OD 0.8–1.0 with 0.2 mM isopropyl-β-thiogalactopyranoside (IPTG). The protein was purified by an immobilized metal affinity column (IMAC). A second IMAC step was performed after cleavage by His-tagged TEV protease, followed by gel filtration on a Superdex 75 column (GE Healthcare). Finally, the protein was concentrated to 10–15 mg.mL−1.
Compounds used for co-crystallisation are given in Table 1. DPBA was purchased from Interchim and rUMP and dTMP from Sigma. Compounds R05-01/02/03 (first described in [15]) were custom re-synthesised by Shanghai ChemPartner. EGCG was purchased from Sigma (E4143).
For the fluorescence resonance energy transfer (FRET) assay, influenza A virus A/Victoria/3/1975(H3N2) PA-Nter fragment was purified as described [5] and stored in aliquots at −20°C in buffer containing 20 mM Tris pH 8.0, 100 mM NaCl and 10 mM β-mercaptoethanol. A 20 base dual-labelled RNA oligonucleotide with 5′-FAM (5′carboxyfluorescein) fluorophore and a 3′-BHQ1 quencher (3′-Black Hole Quencher 1) (Sigma) was used as a substrate for the endonuclease. Cleavage of the RNA liberates the fluorophore from the quencher resulting in an increase of the fluorescent signal. All assay components were diluted in assay buffer containing 20 mM Tris-HCl pH 8.0, 100 mM NaCl, 1 mM MnCl2, 10 mM MgCl2 and 10 mM β-mercaptoethanol. The test compounds were dissolved in DMSO and dilution series were prepared in assay buffer resulting in a final plate well DMSO concentration of 0.5%. Each dilution was tested in quadruplicates. Five µl of each compound dilution was provided in the wells of white 384-well microtiter plates (PerkinElmer). After addition of PA-Nter (1 µM final) the plates were sealed and incubated for 30 min at room temperature prior to the addition of 1.6 µM RNA substrate. Then, the increasing fluorescence signal due to RNA cleavage was measured for 50 min in a microplate reader (Synergy HT, Biotek) at 485 nm excitation and 535 nm emission wavelength. The kinetic read interval was 35 sec at a sensitivity of 35. Fluorescence signal data over a period of 20 min were used to calculate the initial velocity (v0) of substrate cleavage for each compound concentration. The IC50 value was determined using a 4-parameter equation (using Graphpad Prism) whereby positive and negative controls were included to define the top and bottom of the curve.
Influenza A virus was obtained from the American Tissue Culture Collection (A/Aichi/2/68 (H3N2); VR-547). Virus stocks were prepared by propagation of virus on Mardin-Darby canine kidney cells (MDCK; ATCC CCL-34) and infectious titres were determined by 50% tissue culture infective dose (TCID50) analysis. MDCK cells were seeded in 96-well plates at 2×10 E4 cells/well using DMEM/Ham's F-12 (1∶1) medium containing 10% foetal bovine serum (FBS), 2 mM L-Glutamine and 1% antibiotic-antimycotic solution (10.000 Units/ml penicillin, 10 mg/ml streptomycin sulphate, 25 µg/ml amphotericin B) (all from PAA). Until infection the cells were incubated for 5 h at 37°C and 5.0% CO2 to form an 80% confluent monolayer. Test compounds were dissolved in DMSO and dilution series were prepared in infection medium (DMEM/Ham's F-12 (1∶1) containing 5 µg/ml trypsin, and 1% antibiotic-antimycotic solution) resulting in a final plate well DMSO concentration of 1%. The virus stock was generally diluted in infection medium (DMEM/Ham's F-12 (1∶1) containing 5 µg/ml Trypsin, 1% DMSO, and 1% antibiotics) to a theoretical multiplicity of infection (MOI) of 0.05. After removal of the culture medium and a washing step with PBS, virus and compound were added together to the cells. In the wells used for cytotoxicity determination (uninfected cells), the virus suspension was replaced by infection medium. Each treatment was conducted in two replicates. After incubation at 37°C and 5% CO2 for 48 h, each well was observed in the microscope for apparent cytotoxicity, precipitate formation, or other notable abnormalities. Then, cell viability was determined using CellTiter-Glo luminescent cell viability assay (Promega). The supernatant was removed carefully and 65 µl of the reconstituted reagent were added to each well and plates were incubated for 15 min at room temperature under gentle shaking. Then, 60 µl of the solution was transferred to an opaque plate and luminescence (RLU) was measured using Synergy HT plate reader (Biotek). The compounds were titrated on virus infected MDCK cells and the response (RLU) was used to determine the IC50 value using a 4-paramenter equation whereby top and bottom of the curve were defined by the RLU of untreated uninfected cells and untreated infected cells, respectively. The CC50 value was obtained by titrating the compounds on uninfected MDCK cells and similarly fitting the response but with the top of the curve being defined by the RLU of untreated uninfected cells.
The thermal stabilization of the protein in the presence of different inhibitors was performed as described [5], [27]. Briefly, assays were performed with 5 µM H1N1 PA-Nter in 100 mM Hepes pH 7.5, 100 mM NaCl, 1 mM MnCl2, 1 mM MgCl2, 1 mM DTT in the presence or absence of 500 µM of the indicated inhibitors and a 5× dilution of SYPRO Orange dye (Invitrogen. The dye was excited at 490 nm and the emission light was recorded at 575 nm while the temperature was increased by increments of 1°C per minute from 45–93°C (25 to 73°C for no ligand). The fluorescence versus temperature was graphed in Excel and the inflection point taken as the Tm.
Initial sitting drop screening was carried out at 20°C mixing 100 nL of protein solution (15 mg.mL−1) with 100 nL of reservoir solution using a Cartesian robot. Subsequently, larger crystals were obtained at 20°C by the hanging drop method mixing protein and reservoir solutions in a ratio of 1∶1. The protein solution contained 10–15 mg.mL−1 of PA-Nter in 20 mM HEPES pH 7.5, 150 mM NaCl, 2 mM MnCl2, 2 mM MgCl2. The refined reservoir compositions for native crystals and co-crystallization with different ligands are listed in Table 1. Native crystals and those co-crystallized with DBPA, R05-3 and EGCG were flash frozen in liquid nitrogen after cryo-protection in their reservoir solution containing 25% glycerol. Co-crystals with dTMP or rUMP were frozen in their reservoir solution containing 20% glycerol and 10 mM dTMP or rUMP, respectively. Structures of R05-2 and R05-1 were obtained by soaking co-crystals of PA-N-ter and dTMP or rUMP for 2 h with reservoir solution containing the inhibitor followed by cryo protection in reservoir solution containing 20% glycerol and the inhibitor.
Diffraction data were collected at 100 K on various beamlines at the European Synchrotron Radiation Facility (Table 3). Datasets were integrated with XDS [38] and scaled with XSCALE. Subsequent data analysis was performed with the CCP4i programme suite. The initial pH1N1 PA N-ter structure was solved by molecular replacement with PHASER [39] using the previously determined H3N2 PA N-ter structure (PDB code 1W69) [5]. Subsequent co-crystal structures were determined with PHASER using the pH1N1 structure. Refinement was carried out with REFMAC [40] and model building with COOT. In the C2 and P212121 crystal forms, there are four molecules per asymmetric unit. However because of structural variations between the molecules due to plasticity, in particular the 53–73 region, and the generally good resolution, NCS restraints were not applied. In virtually all structures, residues 139–141 and 196–198 are poorly ordered.
Anomalous scattering from manganese was readily observed for the P212121 and P6222 crystal forms at the X-ray energies of normal data collection (∼0.9 Å, Table 3). The signal was much weaker for the C2 crystal form (native and DPDA data), probably due to the lower symmetry and hence less redundant data. A separate dataset of the native C2 crystals was collected at a wavelength of 1.55 Å to enhance the anomalous scattering from manganese (K-edge at 1.90 Å) (see main text).
The sequence alignment in Figure 1 was drawn with ESPript (http://espript.ibcp.fr/ESPript/cgi-bin/ESPript.cgi) [41]. Structure figures were drawn with MOLSCRIPT [42] or BOBSCRIPT [43] and rendered with Raster3D [44].
Structure factors and co-ordinates have been deposited in the PDB as follows: Native (4AVQ), DPBA (4AWF), R05-1 (4AWK), R05-2 (4AVG), R05-3 (4AWG), EGCG (4AWM), dTMP (4AVL), rUMP (4AWH).
|
10.1371/journal.pgen.1003567 | Breakage-fusion-bridge Cycles and Large Insertions Contribute to the Rapid Evolution of Accessory Chromosomes in a Fungal Pathogen | Chromosomal rearrangements are a major driver of eukaryotic genome evolution, affecting speciation, pathogenicity and cancer progression. Changes in chromosome structure are often initiated by mis-repair of double-strand breaks in the DNA. Mis-repair is particularly likely when telomeres are lost or when dispersed repeats misalign during crossing-over. Fungi carry highly polymorphic chromosomal complements showing substantial variation in chromosome length and number. The mechanisms driving chromosome polymorphism in fungi are poorly understood. We aimed to identify mechanisms of chromosomal rearrangements in the fungal wheat pathogen Zymoseptoria tritici. We combined population genomic resequencing and chromosomal segment PCR assays with electrophoretic karyotyping and resequencing of parents and offspring from experimental crosses to show that this pathogen harbors a highly diverse complement of accessory chromosomes that exhibits strong global geographic differentiation in numbers and lengths of chromosomes. Homologous chromosomes carried highly differentiated gene contents due to numerous insertions and deletions. The largest accessory chromosome recently doubled in length through insertions totaling 380 kb. Based on comparative genomics, we identified the precise breakpoint locations of these insertions. Nondisjunction during meiosis led to chromosome losses in progeny of three different crosses. We showed that a new accessory chromosome emerged in two viable offspring through a fusion between sister chromatids. Such chromosome fusion is likely to initiate a breakage-fusion-bridge (BFB) cycle that can rapidly degenerate chromosomal structure. We suggest that the accessory chromosomes of Z. tritici originated mainly from ancient core chromosomes through a degeneration process that included BFB cycles, nondisjunction and mutational decay of duplicated sequences. The rapidly evolving accessory chromosome complement may serve as a cradle for adaptive evolution in this and other fungal pathogens.
| Chromosomal rearrangements are a hallmark of genetic differences between species. But changes in chromosome structure can also occur spontaneously within species, within populations, or even within individuals. The causes and consequences of chromosomal rearrangements affecting natural populations are poorly understood. We investigated a class of fungal chromosomes called accessory chromosomes that are not shared among all individuals within a species. Using a fungal pathogen possessing numerous accessory chromosomes as a model, we assessed chromosome diversity based on whole-genome sequencing and a PCR assay of chromosomal segments that included a global collection of isolates. We show that the accessory chromosomes are highly variable in their gene content and that geographic differences correlate with the number and the structure of the chromosomes. We applied the same approach to document chromosomal rearrangements occurring during sexual reproduction. We identified viable offspring carrying a novel chromosome that originated from a large duplication affecting the majority of the chromosome. Our study showed that chromosomal structure can evolve rapidly within a species to generate a highly diverse set of accessory chromosomes. This chromosomal diversity may contribute significantly to the adaptive potential of fungal pathogens.
| Chromosomal rearrangements are major drivers of genome evolution. Dobzhansky [1] realized that chromosomal polymorphism would “supply the raw materials for evolution”, providing some of the earliest support for Darwin's theory of evolution. Since Dobzhansky's work on Drosophila, cytogenetic studies have revealed a large number of chromosomal rearrangements in the genomes of plant and animal species [2], including humans [3]. Chromosomal rearrangements were shown to contribute to sex chromosome differentiation [4], [5], reproductive isolation [6], speciation [7]–[10] and complex adaptive phenotypes [11].
Chromosomal rearrangements involve deletions, duplications, inversions and translocations within and among chromosomes. In most cases, the molecular mechanisms that generated the observed rearrangements are not known, but a common explanation is mis-repair of double-stranded DNA breaks [12], [13]. Repetitive DNA has been strongly associated with chromosome rearrangements in plant and animal genomes and is thought to promote non-allelic homologous recombination during meiosis due to the misalignment of dispersed repeats [14]–[16]. Telomeres play a major role in maintaining chromosome stability [17], [18]. Although chromosomes lacking a telomere are particularly susceptible to chromosomal fusion, subtelomeric double-strand breaks may also cause chromosomal fusion [19]. McClintock's classic cytogenetic work on maize in the 1930s and 1940s showed that mis-repair of damaged chromosomal ends could generate cycles of chromosomal degeneration termed breakage-fusion-bridge (BFB) cycles [20], [21]. BFB cycles begin when a telomere breaks off a chromosome. When the damaged chromosome replicates, its sister chromatids fuse and form a bridge during anaphase, with the two centromeres of the fused sister chromatids pulled into opposite poles of the dividing cell. After the bridge breaks, the resulting daughter cells receive defective chromosomes that lack telomeres and can initiate new BFB cycles. BFB cycles have also been identified in animals [22], [23] and yeast [24], [25]. In humans, BFB cycles play a significant role in cancer progression [18], [26], [27].
Fungal chromosomes are generally too small for traditional cytogenetic analyses based on chromosome staining and microscopic examination. But fungi were found to show extensive chromosomal polymorphisms following the invention of pulsed-field gel electrophoresis (PFGE). Application of PFGE revealed that many fungal species exhibit a high variability in chromosome number and size, even among individuals drawn from the same random mating population [28]–[30]. Mechanisms generating the differences in chromosome length and number remained largely elusive, although chromosome breakage and non-allelic homologous recombination among repetitive elements during meiosis were suggested to play a role [28], [31]. High chromosomal variability in pathogenic fungi may play an important adaptive role [32]. For example, dramatic changes in copy numbers of an arsenite efflux transporter in Cryptococcus neoformans occurred during experimental evolution favoring arsenite tolerance [33]. Chromosomal disomy was associated with increased antifungal drug resistance in several human pathogens including C. neoformans and Candida albicans [34], [35]. Copy-number variation and aneuploidy were frequently found in clinical and environmental isolates of the same species [36]–[39].
Some of the most polymorphic chromosomal complements were found in plant pathogenic fungi. Several species carry chromosomes that are not shared among all members of the species [40]. Chromosomes exhibiting a presence/absence polymorphism within a species have been referred to as B, dispensable, supernumerary or accessory chromosomes to differentiate them from the “core” chromosomes that are shared among all members of a species [29], [32], [40]. We refer to the chromosomes not shared among all individuals as accessory chromosomes because many of these chromosomes play an adaptive role in pathogen evolution, hence these chromosomes are not truly dispensable [32]. Nor do they fit the classic definition of B chromosomes, because they can carry many coding genes and may be necessary for survival in some environments. One of the best studied fungal accessory chromosomes was found in isolates of the pathogen Nectria haematococca and contains a gene cluster important for virulence on peas [41], [42]. The tomato pathogen Fusarium oxysporum f. sp. lycopersici contains several accessory chromosomes that carry a series of genes important for virulence [43]. In the rice blast fungus Magnaporthe oryzae and related species, a major effector called AVR-Pita that confers virulence on rice was frequently translocated between subtelomeric regions of different chromosomes including accessory chromosomes [44]. Flanking retrotransposons likely contributed to the extreme mobility of the AVR-Pita gene within and among closely related species.
The largest known complement of accessory chromosomes is found in the wheat pathogen Zymoseptoria tritici (syn. Mycosphaerella graminicola [45]). The eight smallest chromosomes of the reference genome of Z. tritici, ranging in size from 409–773 kb, were identified as accessory chromosomes [46]. The core chromosomes of the reference genome range in size from 1,186–6,089 kb [46]. In contrast to accessory chromosomes found in other pathogenic fungi, Z. tritici accessory chromosomes contain over six hundred annotated genes, however the function of these genes is poorly understood [46]. The fungus shows extensive chromosomal length and number polymorphisms within random mating field populations [30], [47]. Some of the chromosomal diversity appears to be generated through meiosis because progeny populations exhibited frequent chromosome loss and disomy as a result of nondisjunction of accessory chromosomes [48]. The origin of the accessory chromosomes of Z. tritici is not known, though both horizontal chromosome transfer from an unknown donor and degeneration of core chromosomes have been proposed [46]. Comparative genomics of closely related species suggested that several accessory chromosomes originated prior to the emergence of Z. tritici [49]. Several lines of evidence suggest that accessory chromosomes may be important for virulence, including the finding that genes on accessory chromosomes are under accelerated evolution and that these are more likely to show a protein signature consistent with a role in pathogenicity [46], [49].
The large set of accessory chromosomes in Z. tritici and its close relatives provides a powerful model system to elucidate the mechanisms underlying fungal chromosomal polymorphisms and the origins of accessory chromosomes. We combined population genomic resequencing and PCR-based chromosome segment genotyping to measure the diversity in chromosomal structure at a global scale. We then performed controlled sexual crosses to trace the fate of accessory chromosomes through meiosis and to identify structural rearrangements in chromosomes among the progeny. We confirmed the findings from our resequencing data with electrophoretic karyotyping that enabled chromosomal separation, isolation and visualization by Southern blotting. Our study provides the most comprehensive view to date of mechanisms underlying chromosomal polymorphisms in evolving fungal populations.
Z. tritici is distributed globally and exhibits high genetic diversity for neutral markers [50]–[52] as well as high phenotypic diversity for quantitative traits, including virulence and thermal adaptation [51]–[53]. To assess the composition and frequency of accessory chromosomes across global populations, we designed 57 PCR assays covering all 8 known accessory chromosomes found in the reference strain IPO323 (chromosomes 14–21; [46]). Amplicons ranging in size from 400–600 bp were targeted to coding regions and primer sites were chosen in conserved regions of each gene (Table S1). The genes comprised in the PCR assay were evenly distributed along the accessory chromosomes and were located mostly in GC-rich regions interspersed by regions of higher repeat content (Figure 1C and 1D). Gene density varies along accessory chromosomes and the PCR assays covered the entire range of known gene locations for each chromosome (Figure 1E). The function of most genes included in the PCR assay is unknown and only 7 out of 57 genes were characterized by gene ontology (Figure 1F; [46]). As a control we designed 15 additional PCR assays covering core chromosomes 10 and 13. Known microsatellite loci were included in each PCR as a positive control. In total, we surveyed 98 isolates sampled from a global collection of four wheat fields at 72 evenly spaced chromosome positions (Table S2): Oregon, United States (n = 19), Israel (n = 23), Australia (n = 30) and Switzerland (n = 26).
The PCR assays on the core chromosomes 10 and 13 showed that 10 chromosomal segments were present in all 98 isolates (Figure 1A). Three chromosomal segments were missing in 1–3 isolates distributed at random across the populations. One segment on each chromosome was missing in a large fraction of the isolates, but was at approximately the same frequency across all populations (Figure 1A). None of the isolates was missing an entire core chromosome. In contrast, chromosomal segments on accessory chromosomes showed large frequency variations among populations and different accessory chromosomes showed different patterns of segmental presence/absence (Figure 1A). With the exception of chromosome 18, all accessory chromosomes were found at a frequency higher than 50% in the four field populations. Chromosome 16 was present at the highest frequency with several chromosomal segments being fixed within populations.
Individual accessory chromosomes showed substantial differences compared to the chromosomes of the Dutch reference strain IPO323. Central chromosomal segments located on chromosome 14 were almost entirely missing in isolates from Australia, the United States and Israel. Swiss isolates showed a central chromosomal segment at approximately half the frequency as segments closer to the telomeric ends of the chromosome. The haplotypic diversity for the presence or absence of individual chromosomal segments was substantial among isolates (Figure S1). Nearly every isolate showed a unique combination of presence or absence of individual accessory chromosome segments. We assessed the population differentiation for presence or absence of chromosomal segments among populations using Wright's FST statistic. Frequencies of several accessory chromosome segments were strongly differentiated among populations. The central segments of chromosome 14 showed FST ranging from 0.15–0.55 (Figure 1B). High levels of differentiation were also found for the second segment of chromosome 15 and the first segment of chromosome 17. Chromosome 18 showed elevated levels of differentiation across the chromosome, largely because this chromosome was almost entirely missing from the Australian and USA populations (Figure 1A). In contrast, previous data on neutral genetic markers on core chromosomes showed little differentiation among these and other populations [50].
We found substantial karyotypic diversity in accessory chromosomes among isolates from Switzerland (Figure 2). In order to obtain a fine-scale map of structural variation in accessory chromosomes among isolates, we performed Illumina resequencing on 9 of the Swiss isolates (mapping coverage 10–23×; Table 1). We identified genomic divergence between the reference isolate IPO323 and the resequenced isolates by mapping all sequence reads to the finished reference genome. To avoid spurious read mapping in repetitive regions of the chromosomes, we restricted our comparison to the coding regions of the accessory chromosomes. Furthermore, we considered exons of multi-exon genes separately to avoid biases introduced by gene length. In summary, we mapped reads to 1763 exons corresponding to 654 unique genes. The average exon length on accessory chromosomes is 314 bp compared to 517 bp on core chromosomes.
The read depth from the resequencing data of 9 Swiss isolates mapped against the reference genome did not suggest any disomic chromosomes (i.e. doubled read depth for a particular chromosome). However, the different isolates varied greatly in gene content on accessory chromosomes. Four isolates (3C4, 3D1, 3F5 and 1A5; Figure 3) showed a nearly complete set of coding sequences compared to the reference genome, with a substantial number of coding sequences present on all 8 accessory chromosomes. Isolate 3D7 contained the smallest complement of accessory chromosome genes, as only four chromosomes showed a substantial proportion of coding sequences to be present. The read mapping to coding regions indicated that accessory chromosomes 14, 19 and 21 likely differ in length among homologous chromosomes (Figure 3). Chromosome 16 was found in all isolates except one. However, chromosome 16 likely differs substantially among isolates due to a large number of deletions compared to the chromosome 16 of the reference genome. Nearly all surveyed 20 kb segments along chromosome 16 showed missing genes in at least some of the resequenced isolates. The strongest variation in coding sequence complements was found among variants of chromosome 14. Isolate 3D1 lacked 149 out of 292 coding sequences, while smaller segments of missing coding sequences were found in six isolates (9G4C, 3B8, 3C4, 3F5, 1E4 and 1A5). The number of missing coding sequences ranged from 18–45 among these six isolates. At one end of chromosome 19, isolate 3B8 showed 46 missing coding sequences out of 220 coding sequences. Similarly, isolates 1A5 and 1E4 showed 31 missing coding sequences out of 155 coding sequences on chromosome 21.
To investigate the nature of large missing chromosomal segments, we performed chromosome-length dotplots of the reference strain chromosome sequence against assemblies of the resequenced isolates. In particular we were interested in whether the large missing segments of chromosome 14 found in isolate 3D1 were due to a single deletion event. The comparison of the reference chromosome 14 of IPO323 with genomic scaffolds of resequenced isolates showed that both the Swiss isolate 3D1 and a previously sequenced Iranian isolate A26b carried one large deletion spanning nearly 400 kb (Figure 4). In addition, we identified two shorter deletions at homologous locations in both isolates (at 210–250 kb and 690–720 kb) compared to the reference chromosome 14. Interestingly, isolate 9G4C was lacking the large central deletion, however, this isolate shared the two peripheral deletions with isolates 3D1 and A26b (Figure 4). A fourth isolate (1E4) shared only the 690–720 kb deletion (Figure 4).
In order to determine the sequence of events leading to the large length polymorphism of chromosome 14 segregating within Z. tritici populations, we performed dotplots with genomic assemblies of three closely related species. We identified significant matching scaffold sequences from isolates of the closest relative Z. pseudotritici spanning the central deletions found in 3D1 and A26b (Figure 4). In the more ancestral species Z. ardabiliae we did not identify any significant matches for chromosome 14. However, in the more distantly related species Z. passerinii, a genomic scaffold spanned the entire central region. The deletion matched the regions identified in 3D1 and A26b, as well as Z. pseudotritici (Figure 4). This suggests that the ancestral chromosome 14 was significantly shorter than the chromosome 14 found in the reference strain IPO323. Furthermore, this finding indicates that the missing sequences in 3D1 and A26b actually represent large insertions into chromosome 14 of the reference strain.
We aimed to ascertain whether the predicted length variants of chromosome 14 are reflected in the karyotypic profiles of the different isolates. For this, we used chromosome-specific probes to identify chromosome 14 in different Z. tritici isolates and Z. passerinii. Hybridization with two chromosome-specific probes (see Table 2) located at opposite ends of the chromosomes showed that the reference isolate IPO323 carried a chromosome 14 in the size range of 780 kb (Figures 5A and 5B; data shown for probe 2) as expected for the isolate [46]. Isolates 3D1 and A26b both carried a substantially shorter chromosome 14 in the range of 400–450 kb, as predicted from the genomic scaffold alignments. The outgroup species Z. passerinii also carried a chromosome 14 that is substantially shorter than in IPO323 (Figures 5A and 5B). Isolate 9G4C was predicted to be of intermediate size between the variants found in IPO323 and 3D1 and A26b. Hybridization with chromosome-specific probes indeed identified a chromosome 14 variant of about 530 kb (Figures 5A and 5B; data shown for probe 2).
To better understand mechanisms leading to the sequence insertions, we identified the precise locations of the breakpoints by performing multiple sequence alignments of the reference chromosome 14 and the scaffold sequences of 3D1, A26b, 9G4C and Z. passerinii. Interestingly, the four sequence breakpoints characterizing the central section of chromosome 14 are at exactly homologous positions in Z. tritici and Z. passerinii (Figure 5D). The first set of breakpoints is located at 213,639 bp and 250,917 bp (breakpoints A and B on Figure 5D) in the IPO323 genome. The second set of breakpoints is located at 256,832 bp and 609,754 bp (breakpoints C and D on Figure 5D).
We aimed to identify the nature of the novel sequences inserted into chromosome 14 of the reference strain. The overall GC-content of chromosome 14 was 48.5% and corresponds to the lowest chromosomal GC content of the Z. tritici reference genome [46]. The two sequences located between breakpoints A-B and E-F showed a consistently lower GC-content than neighboring sequences (Figure 6). The largest sequence, located between breakpoints C and D, showed a heterogeneous GC-content. The density of repeat sequences increased sharply near the breakpoints of the shorter sequences located between breakpoints A-B and E-F (Figure 6). Furthermore, no genes were located between breakpoints A-B and only a single gene was found between breakpoints E-F (Figure 6). The large sequence inserted between breakpoints C-D contained several gene-poor regions. However, the overall gene density of this large sequence is similar to other regions of chromosome 14 (Figure 6). The large inserted sequence contained 16 genes with predicted functions related to a wide variety of metabolic, signaling and transcription factor activities (Figure 7B). By performing a self-alignment of the reference strain chromosome 14 sequence, we identified a substantial number of repeated sequences distributed along the chromosome. In particular, we found a large palindromic sequence located between 500–550 kb that showed high sequence similarity on both sequence strands (Figure 5C).
Chromosome 14 of the reference strain contains a series of transposable element (TE) remnants distributed along the chromosome (Figure 7A). Several of the inserted sequences contain TE remnants near the flanking regions. In particular, a non-long terminal repeat (non-LTR) element is found near both flanking regions of the insertion between alignment breakpoints E and F. The same element is found at flanking regions of the two other insertions (alignment breakpoints A and D). The large palindromic sequence is flanked by outwards facing LTR Copia element remnants.
A major contribution to polymorphisms in accessory chromosomes may arise through meiotic recombination [48]. We performed controlled crosses involving three pairs of isolates from the Swiss population and analyzed 48 progeny from each cross. We applied the same PCR assays targeting 15 chromosomal segments on two core chromosomes and 57 chromosomal segments on the accessory chromosomes. Chromosomal segments on core chromosomes that were missing in either of the two parents were found to be segregating in approximately equal proportions in all three progeny sets (Figure 8). Patterns of segregation were different for several accessory chromosomes. In Cross 1 (9B8B×9G4C), we found a loss of chromosome 16 in one offspring despite the fact that both parental isolates were carrying a near full-length chromosome 16 (Figure 8E). In Cross 2 (1A5×1E4), we found that 8 progeny were missing all chromosome 14 segments, although both parental isolates carried the corresponding chromosome segments (Figure 8C). Similarly, chromosomes 16, 18, 20 and 21 were entirely missing from one offspring though both parents carried these chromosomes. Cross 3 (1A5×3D7) showed the strongest segregation distortions. Parental isolate 3D7 was missing four accessory chromosomes (Figure 8A; chromosomes 14, 15, 18 and 21). Two of these four chromosomes (15 and 21) were inherited in significantly higher proportion than expected under random segregation (X2 test, p<0.0007 multiple comparisons corrected, Figure 8B). Interestingly, in Cross 1 the parental strains similarly differed in their presence of chromosomes 15 and 21, however we did not detect any significant segregation distortion in this cross (Figure 8E). Furthermore, two progeny of Cross 3 lost accessory chromosomes 17, 19 and 20 entirely, although both parental strains carried these chromosomes.
We randomly selected 24 and 34 offspring from Cross 1 and Cross 2, respectively, in order to identify changes in electrophoretic karyotype profiles among progeny. Progeny of both crosses showed substantial karyotypic diversity. Through hybridization with chromosome-specific probes, we found that parental isolates of Cross 2 showed length variation for chromosome 19 of approximately 0.3 Mb (data not shown). Chromosomes 15 and 21 showed nearly identical chromosome lengths among the parental isolates. Progeny of Cross 2 segregated the two length variants of chromosome 19 in approximately equal proportions (data not shown). Larger chromosomes (1.0–3.0 Mb) of parents and progeny of Cross 2 showed similarly diverse electrophoretic karyotypes (Figure 9A). In Cross 2, we identified two progeny (A2.2 and A66.2) out of 34 tested with a chromosomal band estimated to be around 0.9 Mb. However, neither of the two parents were found to have a chromosomal band in the range of 0.7–1.2 Mb, as shown by different PFGE gels optimized to separate either the smallest (<1.0 Mb) or medium-sized chromosomal bands (1.0–3.0 Mb) (Figures 9A and B).
In order to elucidate the origin of the novel chromosome found in two offspring of isolates 1A5 and 1E4, we performed whole genome resequencing of these progeny. The sequencing reads were mapped to all coding sequences of the reference genome, identically to the procedure used for the resequencing of the Swiss population. The parental isolates 1A5 and 1E4 both carried an almost complete set of accessory chromosomes except that 1E4 lacked chromosome 17 (Figure 3). Progeny A2.2 and A66.2 both showed a complete set of accessory chromosomes. However, in contrast to parental isolate 1A5, we did not find any mapping reads for coding sequences spanning the terminal portion of chromosome 17 (ranging from 481–558 kb on the reference chromosome 17; Figure 10A). This missing chromosome segment would result in a reduced length of approximately 100 kb compared to the length of chromosome 17 in the reference strain (full length 584 kb; [46]).
To test for potential duplication events, we used read depth on chromosome 17 as a proxy for duplicated sequences. The parental isolate 1A5 showed a homogeneous distribution of read depth along the chromosome. The parental isolate is suggested to be missing a large chromosomal segment between 1–85 kb compared to the reference genome (Figure 10B). The two progeny A2.2 and A66.2 also lacked the region between 1–85 kb compared to the reference genome (Figure 10B). The central region of chromosome 17 was divided into two sharply distinct regions based on read depth. A region of high read depth between 85–350 kb and a region of low read depth between 350–481 kb (Figure 10B).
We tested whether the increased read depth on chromosome 17 was distinct from the read depth on other chromosomes of the progeny. We calculated the average read density on coding sequences across all 13 core chromosomes as a reference baseline. The average read densities of the parental isolate 1A5 and the two progeny A2.2 and A66.2 were respectively 11.96, 30.93 and 36.34 reads per base pair of coding sequence. We compared these average values to read densities on accessory chromosomes (Figure 10C). In order to mitigate biases introduced by large missing segments on the various accessory chromosomes, we calculated the average read density using only mapped positions for each isolate. Accessory chromosomes of the parental isolate 1A5 showed read densities ranging from 69.4–94.4% of the average read density on core chromosomes, with chromosome 17 showing the lowest read density. Both progeny showed on average slightly higher read densities ranging from 69.1–122% and 76.1–133% for A2.2 and A66.2, respectively. Among all accessory chromosomes, chromosome 17 showed the largest increase (1.77–1.92 fold) in relative read density compared to the parental isolate 1A5 in both progeny (Figure 10C). We hypothesized that this nearly two-fold increase in read density reflected a large duplication event occurring on chromosome 17.
To determine the genomic content of the novel chromosomal band found in the two offspring, we excised the new chromosomal band found at 0.9 Mb from the PFGE gel of progeny A2.2. After purification and whole-genome amplification, we performed Illumina sequencing on the resulting amplified DNA. The sequencing reads were mapped to all coding sequences of the reference genome. The average read density per chromosome was highly variable, with most chromosomes showing an average read density of 0.78–17.4 reads per base pair (Figure 10D). By far the highest read density was found for coding sequences on chromosome 17 with 175 reads per base pair.
We designed two genomic probes specific to chromosome 17 and hybridized the probes to chromosomal bands separated by PFGE. The probes showed that parental isolate 1A5 was carrying a chromosome 17 of the expected length as compared to the reference isolate IPO323 (Figure 9B). We found no hybridization signal on any chromosomal band for parental isolate 1E4. Both progeny A2.2 and A66.2 showed a specific hybridization signal for chromosome 17 on the novel chromosomal band at 0.9 Mb (Figures 9B and 9C; probe 4 see Table 2). A second chromosome-specific probe for chromosome 17 gave identical results (data not shown). Taken together, this strongly suggests that the novel chromosome band is either entirely or almost entirely constituted by sequences belonging to chromosome 17.
We showed that the accessory chromosomes of Z. tritici underwent significant structural rearrangements including fusions and large insertions. The chromosomal complement is highly plastic with substantial variation both in the number of accessory chromosomes carried by each isolate and variation in gene content among homologous chromosomes. We located the exact breakpoints of multiple insertions in chromosome 14 that led to a drastic chromosome length polymorphism within a population. Meiosis played a significant role in shaping accessory chromosome complements. Segregation of some accessory chromosomes was distorted, with nondisjunction during meiosis leading to frequent losses of accessory chromosomes. We found evidence for the fusion of sister chromatids of chromosome 17 in two offspring from the same cross. These progeny carried a nearly doubled chromosome 17 generated through a chromosomal fusion in a subtelomeric region that was likely initiating a breakage-fusion-bridge cycle.
The global survey of chromosomal segments revealed highly diverse accessory chromosome complements. We found that isolates not only differed in the number of accessory chromosomes as expected, but that homologous chromosomes showed markedly different gene contents due to numerous insertions and deletions. Several accessory chromosomes such as chromosome 16 were found near fixation in some populations, such as Australia and Israel. In contrast we found that chromosome 18 was almost entirely missing from the sampled Australian population. The near fixation or losses of accessory chromosomes in some populations may be due to stochastic processes such as founder events during the establishment of the pathogen in previously unaffected geographical regions. Populations also differed strongly in the diversity of chromosomal haplotypes detected by the PCR assays. The Swiss population had a much higher number of unique haplotypes for chromosomes 14, 16 and 18 than the Australian population. Founder effects were hypothesized to explain the low genetic diversity found for neutral markers in Australian Z. tritici populations that were introduced along with wheat not later than ∼200 years ago [50]. In agreement with this earlier finding, accessory chromosomal segments of the Australian population showed the strongest deviation from global frequencies. However, large variations in accessory chromosome complements were also found in other populations. Hence, the diversity in chromosomal complements reflects a previously uncharacterized form of genetic differentiation in this pathogen. Frequency differences in accessory chromosomes among populations may also result from selection operating on chromosomes carrying genes that confer a selective advantage or disadvantage in particular environments. For example, gene products such as effectors that contribute to host virulence in a gene-for-gene interaction may be strongly disfavored in some wheat fields due to the presence of matching resistance genes [54]. If virulence factors such as effectors are located on accessory chromosomes, this may enable rapid adaptation in an arms race to overcome detection by the host immune system. The rapid loss of non-essential virulence factors located on accessory chromosomes may provide a significant selective advantage to a fungal pathogen [32].
The resequencing of Swiss isolates revealed extensive variation in gene content among homologous accessory chromosomes. In comparison to the chromosome sequence of the reference strain, accessory chromosomes of the resequenced isolates carried deletions ranging from a few genes to large sections affecting several dozens of genes. Surprisingly, missing segments were rarely contiguous as would be expected from single deletion and insertion events generating a chromosomal length polymorphism. Accessory chromosome 16 showed numerous short deletions spanning only a few coding sequences in the Swiss population compared to the reference chromosome. Our resequencing analysis (Figure 3) suggests that several chromosomes may be missing chromosomal ends including telomeres. However, our resequencing data was not informative on the integrity of telomeric repeats and we could not be certain that telomeres were missing in these isolates. In the reference genome, one telomere sequence on chromosome 21 could not be sequenced and may be missing [46]. Intact telomeres play a crucial role in chromosomal stability by ensuring homologous chromosomal pairing and disjunction during meiosis [17]. Defective telomeres are thought to initiate the development of breakage-fusion-bridge cycles leading to major chromosomal anomalies [17], [18]. If some accessory chromosomes are indeed defective for telomeres, this may play a major role in generating the observed chromosome polymorphisms.
The most dramatic chromosomal length polymorphism segregating within a population was found for chromosome 14, with the shortest identified chromosome variant approximately half the length of the longest known chromosome variant. In a related pathogen found on barley (Z. passerinii) that is ancestral to the more closely related pathogens found on wild grasses [55] we identified a homologous chromosome. The ancestral form of chromosome 14 is largely identical to the shortest variant found in Iranian and Swiss populations. Several lines of evidence suggest that the large insertions leading to the chromosome 14 variant found in the reference strain occurred recently. First, the longest chromosome variant is found almost exclusively in the Swiss population, which is closest to the location where the Dutch reference isolate IPO323 was isolated. The A26b isolate from Iran was sampled close to the center of origin of Z. tritici and this isolate carried the shortest known chromosomal variant. Second, sequences immediately adjacent to the insertion breakpoint locations showed only a single nucleotide polymorphism compared to the reference chromosome. Third, sequences near the breakpoint location were highly similar even when compared with the phylogenetically distant Z. passerinii.
A major open question is the source of the inserted sequences. We did not find closely related sequences either at a different location in the reference genome or in any resequenced strain. The largest insertion in chromosome 14 contains several dozen genes and may have functional consequences for isolates carrying the large chromosome variant, because several of these genes were predicted to encode transcription factors or other functions. All three inserted sequences are flanked on at least one end by remnants of the same class of transposable elements. The presence of these elements near the flanking regions suggests that non-allelic homologous recombination with an unknown chromosome may have played a role in the insertion of these sequences into chromosome 14. Interestingly, the two shorter insertions in chromosome 14 showed a markedly lower GC-content than surrounding regions and these inserts were virtually devoid of genes. These isochores may be regions of reduced recombination, as the inserted regions may lack homologous sequences necessary for meiotic crossing-over. The largest inserted sequence also contains a very large palindromic sequence similar in extent to palindromes on the human Y chromosome [56], [57]. The palindrome is flanked by the remnants of two copies of a transposable element, similar to the inserted sequences. In yeast, palindromes were shown to mediate gene amplification and intra-chromosomal recombination and may lead to genomic instability [58], [59]. Goodwin et al. [46] hypothesized that accessory chromosomes originated through an ancient horizontal transfer from an unknown donor species. Our analyses show that the accessory chromosome 14 was maintained through multiple speciation events and hence may be a remnant of an ancient core chromosome. The large insertions observed in extant Z. tritici populations suggest that chromosome 14 is undergoing a degeneration process. The insertions do not seem to have a severe effect on the fitness of the organism, as many different length variants of chromosome 14 were found segregating within the Swiss population. The tolerance to large sequence rearrangements may be a hallmark of the degeneration process affecting accessory chromosomes.
Meiosis is thought to play a major role in genomic instability in fungi [28], [48], [60]. Non-allelic homologous recombination among dispersed repeats was hypothesized to be the main source of chromosomal length polymorphism in fungi [28], [60]. In Z. tritici, aberrations during meiosis were suggested to lead to the loss of accessory chromosomes and hence contribute to chromosomal number polymorphism among isolates [48]. Our analyses of progeny from three different crosses showed that chromosomal loss affected nearly all accessory chromosomes. We detected low levels of chromosomal losses for all accessory chromosomes except chromosome 15. The loss of a chromosome may be due to nondisjunction of sister chromatids during meiosis. This defect during meiosis would create progeny carrying a duplicated (i.e. disomic) chromosome. Frequent loss of accessory chromosomes during meiosis poses an apparent paradox in Z. tritici. Populations would be expected to gradually lose all accessory chromosomes over generations in the absence of mechanisms to maintain the accessory chromosomes. Interestingly, our data on inheritance of accessory chromosomes revealed a mechanism that may maintain accessory chromosome complements in populations. Analyses of segregation frequencies revealed that chromosomes 15 and 21, if present in only one of the two parental strains, were inherited significantly more frequently than expected under random segregation of the chromosomes. Distorted segregation was restricted to one cross and no distortion was detected in the second cross differing in the presence of chromosomes 15 and 21. Segregation distortion was found to be a key characteristic of numerous animal B chromosomes [61]. We hypothesize that segregation distortion is one of the mechanisms that maintains accessory chromosomes in Z. tritici populations.
The most striking example of chromosomal plasticity was the fusion of sister chromatids to generate a much longer chromosome 17 in two progeny of Cross 2 (Figure 10). This meiotic abnormality occurred in a cross between a strain carrying a chromosome 17 similar to the reference isolate and a strain lacking the entire chromosome. One mechanism for creating the new chromosome is non-allelic homologous recombination between inverted repeats on sister chromatids ([62], Figure 10E). The recombination event could create an isodicentric chromosome 17 carrying duplicated and non-duplicated regions, consistent with the striking difference in read depth observed across chromosome 17 (Figure 10B). If the fused chromosome contains two centromeres, it is expected to form a bridge at anaphase and undergo BFB cycles [20], [21], [26]. The rejoining of broken ends during new rounds of cell division will create new chromosomal arrangements including deletions and duplications. The lack of a homologous chromosome 17 during meiosis may have contributed to the initiation of a BFB cycle. The fate of the novel chromosome over subsequent generations is currently under investigation. The meiotic pairing of the large duplicated chromosome 17 with the parental chromosome variant is likely to generate further highly unstable chromosomal variants. Z. tritici possesses a genomic defense mechanism known as RIP [46] that is common to a large number of ascomycete fungi [63]. RIP rapidly degenerates highly similar genomic regions through the introduction of point mutations. We predict that the novel chromosome 17 variant generated by duplicating a large fraction of the original chromosome would be subjected to rapid degeneration as a result of RIP. We propose that BFB cycles coupled with RIP played a major role in creating the degenerated accessory chromosomes of Z. tritici.
Our study revealed extensive yet viable chromosomal rearrangements generated by meiosis. Genomic instability and insertion of exogenous sequences led to highly diversified sets of homologous chromosomes affecting hundreds of genes. The large number of insertions and deletions found among accessory chromosomes suggests that these chromosomes underwent an extensive degeneration process. The chromosomal degeneration process may well have been initiated in an ancestor of Z. tritici. The shorter gene length and lower gene density on accessory chromosomes compared to core chromosomes suggests that degeneration processes affected accessory chromosomes over long evolutionary time scales. We identified large insertions and the initiation of breakage-fusion-bridge cycles as two major contributors to chromosomal abnormalities. Surprisingly, isolates of Z. tritici appear to be highly tolerant of these abnormalities, which may contribute to the maintenance of extensive karyotypic diversity in populations. The extensive degeneration, distorted segregation and frequent loss of accessory chromosomes highlight a central question surrounding fungal accessory chromosomes: How and when do these chromosomes originate? We showed that chromosome 14 is ancient, as its origin predates several speciation events prior to the emergence of Z. tritici in the Fertile Crescent [55]. We postulate that the accessory chromosomes found in extant Z. tritici populations likely originated from the core chromosomes through a degeneration process. The initiation of chromosome degeneration is particularly likely in isolates that carry disomic chromosomes due to nondisjunction. Disomy would provide redundancy in gene content and, hence, relax selection pressure to maintain chromosomal integrity. Chromosomal degeneration may then proceed rapidly through BFB cycles, nondisjunction and RIP of duplicated regions. The emergence of a highly diverse and rapidly evolving set of accessory chromosomes in Z. tritici illustrates how an accessory genome can be created to serve as a cradle for adaptive evolution in this and other fungal pathogens.
We assessed the diversity in chromosomal structure in a global collection of Z. tritici. We included field populations from Israel (n = 23), Oregon, USA (n = 19), Switzerland (n = 26) and Australia (n = 30) (Table S2). These populations were previously assayed for neutral genetic diversity and variation in quantitative traits [53]. These isolates showed substantial variation for several quantitative characters, including virulence, fungicide resistance and thermal adaptation, among and within populations [51]–[53], [64].
Sexual crosses were performed between three pairs of isolates from the Swiss population (see Table S2) using the established protocol for Z. tritici [65]. The crosses were between isolates ST99CH9B8B and ST99CH9G4C (Cross 1), ST99CH1A5 and ST99CH1E4 (Cross 2) and ST99CH1A5 and ST99CH3D7 (Cross 3).
In order to survey presence-absence polymorphism among accessory chromosomes, we designed PCR assays to amplify approximately 500 bp of coding sequences at regular intervals of approximately 100 kb along the chromosomes of reference strain IPO323. For detailed information on the targeted genes and chromosomal locations see Table S1. Primers for PCR amplification were designed on conserved sections of the targeted coding sequence. Sequence conservation was assessed using the reference assembly of nine resequenced Swiss isolates and two resequenced Iranian isolates (for details see below). We used Primer 3.0 for primer design [66]. In order to control for successful PCR, we included a primer pair of a microsatellite locus in each PCR mix [67]. Successful PCRs produced a band at approximately 250 bp that was clearly distinguishable from the PCR product associated with each chromosomal segment. PCR reactions were performed in 20 µl volumes containing approximately 5–10 ng genomic DNA, 0.5 µM of each primer, 0.25 mM dNTP, 0.6 U Taq polymerase (DreamTaq, Thermo Fisher, Inc.) and the corresponding PCR buffer. PCR products were visualized on agarose gels.
We used the R graphics package ggplot2 to plot the raw datasets and analyses [68], [69]. Measures of genetic differentiation among populations (FST) were calculated with the function var.comp in the R package hierfstat [70]. The presence-absence data generated by the PCR assays were considered as two possible alleles at haploid loci. We tested for segregation distortion of chromosomal segments among progeny by testing for deviations from the expected 1∶1 segregation ratio of presence-absence polymorphism among progeny with a χ2 contingency table. We accounted for non-independence of chromosomal segments and multiple testing with a conservative Bonferroni correction. We calculated the repeat content on accessory chromosomes by identifying direct repeats with a repeat motif between 2–50 bp [71]. For each repeat, we calculated the total length of the repeat and subtracted the number of mismatches in the repeat motif, as a proxy for the extent and purity of the repeat element.
We used the previously published genome assemblies of two Iranian isolates of Z. tritici ST01IRA26b and ST01IRA48b. In addition, we included five genomes of Z. pseudotritici (STIR04_3.11.1, STIR04_2.2.1, STIR04_4.3.1, STIR04_5.3, STIR04_5.9.1) four genomes of Z. ardabiliae (STIR04_3.3.2, STIR04_3.13.1, STIR04_1.1.1, STIR04_1.1.2) and one genome of the outgroup species Z. passerinii (P63) [72], [73]. All genome assemblies are available under the NCBI BioProject [PRJNA63131] on GenBank. We resequenced nine Z. tritici isolates from Switzerland (full isolate names: ST99CH1A5, ST99CH1E4, ST99CH3B8, ST99CH3C4, ST99CH3D1, ST99CH3D7, ST99CH3F5, ST99CH9B8B and ST99CH9G4C) and two progeny from Cross 2 (A2.2 and A66.2). We performed Illumina paired-end sequencing on 500–700 bp insert libraries to generate between 1–2 Gb of quality-trimmed sequence data per isolate (theoretical coverage of 25–50×). The read length was either 82 bp or 90 bp. Illumina sequence data are available from the NCBI Short Read Archive (see Table 1 for accessions).
We used SOAPdenovo v. 1.5 [74] to generate de novo assemblies, including scaffolding and gap closing. De novo assemblies yielded a scaffold N50 ranging from 79,920–121,161 bp depending on the resequenced isolate. Total assembly space (scaffolds and singletons) ranged from 35.57–38.33 Mb (see Table 1). All genome assemblies are available on GenBank under BioProject [PRJNA178194] (see Table 1). The comparison with the total finished genome size for the reference isolate IPO323 (39.7 Mb) shows that the genomic assemblies account for a very large proportion of the genome of the sequenced isolates. The assembly statistics were similar to the assemblies reported earlier for the same species [72].
We mapped the Illumina reads of each resequenced isolate and offspring to the finished genome of IPO323 [46]. We used Bowtie 2.1.0 [75] to perform the mapping, allowing only reads that were mapped as paired-ends. We assessed the read coverage on the reference genome by filtering all reads based on their mapping quality (minimum mapping quality of 20) with GATK version 2.3-9-ge5ebf34 [76]. Coverage of coding sequences was extracted using the BEDtools utilities [77]. We scored the absence of coding sequences conservatively, requiring that less than 10 bp of a coding sequence should be covered and that the average read density on the coding sequence would be below 2×.
Structural changes among chromosomes of different isolates were analyzed using Nucmer [78]. We used the –mum option requiring unique anchor matches that are unique in both the query and the reference genome. Genome assemblies were compared in pairwise comparisons between the finished reference genome of IPO323 and the draft assemblies of the different isolates of Z. tritici, Z. pseudotritici, Z. ardabiliae and Z. passerinii. In order to visualize synteny among different variants of chromosome 14, we extracted all scaffolds matching the reference chromosome 14. We discarded scaffolds that were shorter than 10 kb and that showed a match identity with the reference chromosome of less than 80%. Scaffold alignments were plotted with the R package ggplot2 [69]. Repetitive and palindromic sequences of the reference chromosome 14 of IPO323 were visualized by performing a self-alignment with LASTZ (http://www.bx.psu.edu/~rsharris/lastz).
The finished chromosome 14 sequences were analyzed for short and medium length tandem repeats with the software Tandem Repeat Finder v. 4.04 [79]. We set the matching weight to 2, the mismatching and indel penalty to 10 and the match and indel probability to 80 and 10, respectively. The minimum alignment score was required to be 10 and the maximum period size of repeats was set to 50 bp. The occurrence of repeats was visualized along a 5 kb sliding window (with increments of 1 kb). The gene density on each chromosome was reported as the occurrence of start codons according to the latest annotation [46]. GC content of each chromosome was reported in 5 kb sliding windows with increments of 1 kb. We identified transposable element remnants on chromosome 14 by querying the annotated repeat libraries provided by Repbase Update [80].
High molecular weight chromosomal DNA (Ch-DNA) was prepared by in situ digestion of cell walls of agarose-embedded conidia. We used a slightly modified non-protoplasting method according to McCluskey et al. [81]. The following Z. tritici isolates were used: ST01IRA26b, ST99CH9B8B (parental isolate of Cross 1), ST99CH9G4C (parental isolate of Cross 1), ST99CH1A5 (parental isolate of Cross 2 and 3), ST99CH1E4 (parental isolate of Cross 2), ST99CH3B8, ST99CH3C4, ST99CH3D1, ST99CH3D7, ST99CH3F5 and IPO323. In addition, we included the isolate P63 of Z. passerinii [55]. To screen progeny of sexual crosses, we randomly selected 24 and 34 confirmed progeny from Cross 1 and Cross 2, respectively.
All isolates were transferred from stocks maintained in glycerol at −80°C to Yeast Malt Agar (YMA) plates and were grown for 3 to 4 days in the dark at 18°C. After incubation, conidia were washed off the plates with sterile water and 600–800 µl of suspended conidia were transferred to 2 to 3 fresh YMA plates. The plates were incubated for 2 to 3 days as described above. Conidia were harvested using sterile distilled water and filtered through sterile Miracloth (Calbiochem, La Jolla CA, USA) into 50 ml screw-cap Falcon tubes. The tubes were filled with distilled water up to 50 ml total volume. The suspension was centrifuged at 3750 rpm at room temperature for 15 min with a clinical centrifuge (Allegra X-12R, Beckman Coulter, Brea CA, USA). The resulting pellets were resuspended in 1–3 ml TE buffer (10 mM Tris-HCL, pH 7.5; 1 mM EDTA, pH 8.0) and gently vortexed. The spore concentration of the solution was determined using a Thoma haematocytometer cell counter. An aliquot of 1.5 ml spore suspension with a concentration between 8×107 to 2×108 spores/ml was transferred to a fresh 50 ml screw-cap tube and incubated at 55°C in a water bath for several minutes. To each tube, 1.5 ml pre-warmed (55°C) low-melting-point agarose prepared in TE Buffer was added (2% w/v; molecular biology grade, Biofinex, Switzerland). The solution was thoroughly mixed by gentle pipetting. An aliquot of 500 µl was solidified on ice for approximately 10 min in a precooled plug casting mold (BioRad Laboratories, Switzerland). A total of five agarose plugs per isolate were incubated in 15 ml screw-top tubes containing 5 ml of a lysing solution containing 0.25 M EDTA, pH 8.0, 1.5 mg/mL protease XIV (Sigma, St. Louis MO, USA), 1.0% sodium dodecyl sulfate (Fluka, Switzerland). The incubation was performed for 28 h at 55°C. During the incubation the lysing solution was changed once after 18 h and gently mixed every 2–3 h. Chromosomal plugs were washed three times for 15–20 min in 5–6 ml of a 0.1 M EDTA (pH 9.0) solution and then stored in the same solution at 4°C until they were used.
Pulsed-field gel electrophoresis (PFGE) was carried out using a BioRad CHEF II apparatus (BioRad Laboratories, Hercules CA, USA). Chromosomal plugs were inserted into the wells of a 1.2% and 1.0% (wt/vol) agarose gel (Invitrogen, Switzerland) to separate small chromosome (<1 Mb) and medium-sized chromosomes (1.0 Mb–3.0 Mb), respectively. Small chromosomes (i.e. accessory chromosomes) were separated at 13°C in 0.5× Tris-borate-EDTA Buffer (Sambrook & Russell 2001) at 200 V with a 60–120 s pulse time gradient for 24–26 h. Medium-sized chromosomes were separated at 100 V with a 250–900 s pulse time gradient for 48–50 h using the same buffer and running temperature as above. Gels were stained in ethidium bromide (0.5 µg/ml) for 30 min immediately after the run. Destaining was performed in water for 5–10 min. Photographs were taken under ultraviolet light with a Molecular Imager (Gel Doc XR+, BioRad, Switzerland). As size standards, we used chromosome preparations of Saccharomyces cerevisiae (BioRad, Switzerland) and Hansenula wingei (BioRad, Switzerland).
Southern blotting and hybridization were performed according to standard protocols [82]. In summary, hydrolysis was performed in 0.25 M HCl for 30 min and DNA was blotted onto Amersham HybondTM-N+ membranes (GE Healthcare, Switzerland) overnight under alkaline conditions [82]. DNA was fixed onto the membranes at 80°C for 2 h. Membranes were prehybridized overnight with 25 ml of a buffer containing 20% (w/v) SDS, 10% BSA, 0.5 M EDTA (pH 8.0), 1 M sodium phosphate (pH 7.2) and 0.5 ml of sonicated fish sperm solution (Roche Diagnostics, Switzerland). Probes were labeled with 32P by nick translation (New England Biolabs, Inc.) following the manufacturer's instructions. Hybridization was performed overnight at 65°C. Blots were subjected to stringent wash conditions with a first wash in 1× SSC and 0.1% SDS and a second wash with 0.2× SSC and 0.1% SDS. Both washes were performed at 60°C. Membranes were exposed to X-ray film (Kodak BioMax MS) for 2 to 3 days at −80°C. All hybridization probes used to identify specific chromosomes are listed in Table 2.
Chromosomal DNA was separated with CHEF gel electrophoresis as previously described for the separation of small chromosomes except that a 1.0% agarose gel was used. The novel 0.9 Mb chromosomal band from isolate A2.2 was excised and DNA was recovered using the Wizard SV Gel and PCR Clean-up System kit (Promega, Switzerland) with the following modifications to the manufacturer's recommendations: during the incubation at 65°C the gel slice was vortexed two times for 5 minutes, sonication was for 3 min and followed by a final incubation for 1 min. The resulting purified DNA was amplified using a whole genome amplification kit (REPLI-g Mini Kit, Qiagen, Germany). Amplified DNA was subjected to whole genome sequencing with an Illumina HiSeq 2000 as described above.
|
10.1371/journal.pcbi.1005492 | The decay of motor adaptation to novel movement dynamics reveals an asymmetry in the stability of motion state-dependent learning | Motor adaptation paradigms provide a quantitative method to study short-term modification of motor commands. Despite the growing understanding of the role motion states (e.g., velocity) play in this form of motor learning, there is little information on the relative stability of memories based on these movement characteristics, especially in comparison to the initial adaptation. Here, we trained subjects to make reaching movements perturbed by force patterns dependent upon either limb position or velocity. Following training, subjects were exposed to a series of error-clamp trials to measure the temporal characteristics of the feedforward motor output during the decay of learning. The compensatory force patterns were largely based on the perturbation kinematic (e.g., velocity), but also showed a small contribution from the other motion kinematic (e.g., position). However, the velocity contribution in response to the position-based perturbation decayed at a slower rate than the position contribution to velocity-based training, suggesting a difference in stability. Next, we modified a previous model of motor adaptation to reflect this difference and simulated the behavior for different learning goals. We were interested in the stability of learning when the perturbations were based on different combinations of limb position or velocity that subsequently resulted in biased amounts of motion-based learning. We trained additional subjects on these combined motion-state perturbations and confirmed the predictions of the model. Specifically, we show that (1) there is a significant separation between the observed gain-space trajectories for the learning and decay of adaptation and (2) for combined motion-state perturbations, the gain associated to changes in limb position decayed at a faster rate than the velocity-dependent gain, even when the position-dependent gain at the end of training was significantly greater. Collectively, these results suggest that the state-dependent adaptation associated with movement velocity is relatively more stable than that based on position.
| Human motor adaptation of limb movement in response to force perturbations has been shown to be motion-state dependent. That is, the compensatory response to these disturbances is correlated and proportional to the temporal changes in the position, velocity, and acceleration during the motion. Despite a growing understanding of this adaptation process, there is little information on the relative stability of this learning when based on these different temporal features of movement. Here we modified a previous computational model of motor adaptation to predict the decay of the compensatory response associated to different motion states, specifically learning based on temporal variations in limb position and velocity. We confirmed the simulated behavior by examining the decay of the temporal force output after subjects were trained to compensate for movement disturbances based on different combinations and magnitudes of these two motion states. Both simulation and behavioral results show that velocity-based learning decays at a slower rate than position-based, even when learning is significantly biased towards the latter at the end of training. Collectively, these results suggest that motion-state learning based on movement velocity is more stable than that based on limb position.
| The motor system adapts to movement perturbations, a process largely driven by the error between the executed movement and the predicted consequences of that movement [1–3]. This short-term form of motor learning is a gradual updating of the motor commands required to counteract the movement perturbation. Similar to learning, the decay of adaptation following the removal of the perturbation is typically a gradual process as the motor commands revert back to the state prior to exposure [4,5]. Thus, examining and comparing the progression and decay of motor adaptation provides insight into the stability of these updates to the issued motor commands.
The decay of motor adaptation has been studied for various behavioral paradigms involving limb movement: prism displacement [6,7], locomotion [8,9], visuomotor alterations [5,10,11] and force-field perturbations [12–15] In the last case, subjects make reaching movements while interacting with a robotic manipulandum and are exposed to a force perturbation typically dependent upon either a single motion kinematic parameter (e.g., changes in position, velocity or acceleration during the movement) or the combination of these motion states [16]. In response to the movement disturbance, subjects apply an adaptive response based on the temporal characteristics of the limb state. Although previous investigations of force-field adaptation have examined the time course and the factors that influence the stability and retention of these state-dependent compensatory responses [4,12,14,17–23] the relative stability of the different state-dependent components that drive adaptation is not well understood, especially in direct comparison to the initial learning process.
Here, we applied a framework developed by Sing and colleagues [16] to compare the progression and decay of state-dependent adaptation in response to different types of novel movement dynamics. Based on this framework, the feedforward motor output in response to the applied force perturbation is the weighted sum of gains assigned to the kinematic parameters of the reaching motion (changes in limb position and velocity, [16,23–25]). The model predicted that the changes in these gains during adaptation would follow a different time course than during the adaptation decay, but the authors did not explicitly test this prediction nor the relative stability of the changes based on the kinematic parameters. To assess the difference in the relative stability of the motor memory based on changes in limb position or velocity we first modified the original model proposed by Sing et al. [16] based on observed differences in the retention of adaptation in response to purely velocity- or position-dependent disturbances. The resulting model simulations predicted that when motor learning is based on the combination of position and velocity the decay of adaptation is biased towards velocity, independent of the final adaptation level. That is, the model predicted that the decay of position-based adaptation would occur at a faster rate, even when this learning was significantly higher at the end of training. We tested two additional groups of subjects and found that the behavioral results were in agreement with the predictions of the model. Collectively, our behavioral and simulation results suggest that (1) the decay of motor adaptation is not merely the reversal of the learning process, but at least a partially distinct process likely involving separate mechanisms, (2) the velocity-based contribution to updating motor commands is more stable than that based on position, and (3) a model with asymmetrical retention factors for position- and velocity-based motor primitives can predict the time course of adaptation decay for various state-dependent motor learning goals.
We first trained subjects to make reaching movements in either a position- or velocity-dependent force-field (pFF and vFF) (Fig 1). Each subject experienced only one type of perturbation after an initial baseline period, during which error-clamp trials were used to quantify the feedforward adaptive changes to the motor output (see Materials and Methods). Based on the forces subjects applied during the error-clamp trials, we were able to determine the adaptation coefficient (the linear regression of the applied lateral force profile onto the ideal compensatory force profile) and the respective gain of the position-dependent and velocity-dependent force components to the overall force profile (see Materials and Methods). Fig 2A plots the adaptation coefficient as a function of trial number for pFF and vFF training. Similar to previous studies [16,21,26,27], we observed a fast progression of adaptation early on (within the first 15 trials) that plateaued after approximately 75 trials for both force-field types (Fig 2A). An exponential fit of the adaptation curve showed a faster overall adaptation for pFF training (time constant of 7.2 ± 0.8 for pFF compared to 12.3 ± 5.5 trials for vFF. See S1A Fig). The amount of adaptation at the end of training was significantly greater than at the beginning, but the adaptation levels were not significantly different between perturbation types (2-way ANOVA, P < 0.001 for the main effect of training period and P = 0.22 for the main effect of perturbation type). Specifically, early adaptation levels were not significantly different between pFF and vFF training (0.38 ± 0.03 compared to 0.30 ± 0.05, mean ± SEM, P = 0.18, two-tailed t-test). There was also no difference in the adaptation level between vFF and pFF late in training where the behavior asymptotes (0.72 ± 0.02 for pFF and 0.69 ± 0.04 for vFF, P = 0.58, two-tailed t-test). (Early adaptation period was determined over trials 1–15, while the late/asymptotic adaptation period was trials 150–160. See Materials and Methods for justification of these ranges.)
Immediately following training, subjects experienced a sequence of consecutive error-clamp trials to determine the decay of feedforward changes to motor output. Following the start of the consecutive error clamps the adaptation coefficient began to decay and reached asymptote by the end of the period. An exponential fit of the adaptation decay curve showed a faster decrease in adaptation for vFF over pFF training (time constant of 11.0 ± 2.2 trials for pFF compared to 7.1 ± 1.0 trials for vFF. See S1B Fig). The adaptation coefficient levels at the end of the decay period remained significantly greater than baseline levels (vFF: 0.006 ± 0.006 vs. 0.26 ± 0.05, pFF:-0.004 ± 0.004 vs. 0.19 ± 0.03, paired two tailed t-test, P < 0.001 for both cases). Although the pattern of decay was similar for both pFF and vFF training (Fig 2A), there was a slight, but insignificant difference in the adaptation level before the start of decay as noted above. In order to examine the decay with respect to the final adaptation levels, we normalized the decay of adaptation by the initial value of the adaptation coefficient at the beginning of the decay period (Fig 2B). Starting at an adaptation level of 1.0, we analyzed the decay of adaptation in early and late epochs during the decay period. The percentage of adaptation at the beginning of the decay period was significantly greater than at the end, but the adaptation levels were not significantly different between perturbation types (2-way ANOVA, P < 0.001 for the main effect of decay period and P = 0.97 for the main effect of perturbation type). In the early epoch, there was not a significant difference in the percentage of adaptation that remained for pFF and vFF training (pFF: 42.3 ± 5.2%, vs vFF: 36.0 ± 6.2%; P = 0.41, post hoc comparisons using Bonferroni correction). This was also true for the late epoch (pFF: 18.8 ± 3.4%, vs vFF: 25.5 ± 5.3%; P = 0.35, Bonferroni correction). (The early epoch of decay was over trials 11–20, while the late epoch was trials 50–60. See Materials and Methods for justification of these ranges.)
Although the one dimensional adaptation coefficient suggested similar behavior for vFF and pFF training, we were interested in the temporal characteristics of the corresponding force profiles during the adaptation and the decay periods (see Materials and Methods). As in previous studies [16,23,25] we compared the temporal shape of the force profiles with changes in limb position and velocity (Fig 2C and 2D). As shown previously [16], early in adaptation the force pattern was dependent on both the position and velocity changes during the movement (pFF: 21.5 ± 2.2% for velocity and 78.5 ± 2.2% for position, vFF: 67.8 ± 7.3% for velocity and 32.2 ± 7.3% for position) (top panels in Fig 2C and 2D). Notably, late in the adaptation period the force pattern was mostly aligned with the appropriate movement parameter for the adaptation. In other words, the force exerted by subjects in the late phase of pFF adaptation was largely aligned with changes in limb position (95.9 ± 0.6% compared to 4.1 ± 0.6% for velocity), and late adaptation to vFF was mostly aligned with changes in movement velocity (93.2 ± 1.6% compared to 6.8 ± 1.6% for position).
We also examined the temporal force patterns during the decay period of the respective force-field perturbations. Interestingly, the force profiles remained aligned to the appropriate motion state required to compensate for the perturbation in both the early and late stages of decay (bottom panels in Fig 2C and 2D). In the early phase of the decay of pFF learning (Fig 2C bottom panel), the force profiles mainly consisted of a position-dependent component with a minimal velocity-dependent component (90.1 ± 7.0% compared to 9.9 ± 7.0%). In the late decay phase of pFF learning, the position-dependent component continued to contribute the most to the exerted force while the velocity contribution remained small (80.4 ± 8.0% compared to 19.6 ± 8.0%). Similarly, the force profiles in both the early and late decay phases of vFF learning were mostly dependent on movement velocity, with less contribution of limb position (early: 76.4 ± 7.0% compared to 23.6 ± 7.0%, late: 76.0 ± 8.2% compared to 24.0 ± 8.2%) (bottom panels in Fig 2D). Thus, the comparison of the temporal force profiles suggests that the proportional contributions of limb position and velocity to the overall motor output achieved at the end of training were largely maintained during the decay of the motor learning.
Differences in the force profile described above suggest that the gain associated to the respective motion states is different not only between the two types of force-field adaptations, but also between the learning and decay periods. In order to visualize these differences, we examined the changes in the respective gain associated to the motion states for adaptation and decay in a two dimensional gain-space (see Materials and Methods). We parsed the position-dependent and velocity-dependent force components and found a clear separation between adaptation and decay paths for both pFF and vFF training (Fig 3A and 3B). For both types of perturbations, we identified a goal-aligned and a goal-misaligned component. The goal-aligned component for pFF training is parallel to abscissa in gain space and represents the position-dependent force component, whereas the goal-misaligned component is parallel to the ordinate and represents the velocity-dependent force component. These relationships are reversed for vFF training with the goal-aligned and goal-misaligned components represented by the velocity- and position-dependent axes, respectively. In both pFF and vFF training, the goal-aligned force component had a significantly greater contribution to the initial adaptation than the goal-misaligned components (pFF: goal-aligned (0.33 ± 0.03) vs goal-misaligned (0.21 ± 0.02); vFF: goal-aligned (0.25 ± 0.03) vs goal-misaligned (0.09 ± 0.04), P < 0.05 in both cases). (Note there is a larger goal-misaligned component for initial pFF adaptation than for initial vFF adaptation. That is, the velocity-dependent contribution to the adaptive response for pFF training was larger than the position-dependent contribution for vFF training. This asymmetry supports an initial adaptation bias towards velocity-dependent learning as discussed below). As subjects continued to experience the force-field, contributions from the goal-aligned component increased whereas the goal-misaligned component decreased (Fig 3A and 3B). We examined two time points during training (points 1 and 2) which represent early (training trials 1–15) and late adaptation (training trials 150–160) respectively. By the end of the training period (the point labeled 2 in Fig 3A and 3B), the majority of the compensatory force was due to the contribution of the goal-aligned force component (vFF: goal-aligned (0.66 ± 0.04) vs goal-misaligned (0.05 ± 0.008); pFF: goal-aligned (0.69 ± 0.02) vs goal-misaligned (0.1 ± 0.01), P < 0.05 in both cases). This decrease in the goal-misaligned component resulted in a curvature in the learning trajectories (note the difference between points labeled 1 and 2). However, the magnitude of this curvature was not the same for pFF and vFF training; the contribution of the velocity-dependent force for pFF training from early (point 1 in Fig 3A) to late adaptation (point 2 in Fig 3A) was significantly different in magnitude (1st point: 0.21 ± 0.02 vs. 2nd point: 0.10 ± 0.01, P < 0.05). Although there was a similar decrease in the contribution of the position-dependent forces from early to late adaptation for vFF training, this decrease in magnitude was not significant (1st point: 0.09 ± 0.04 vs. 2nd point: 0.05 ± 0.01, P = 0.32).
The gain-space trajectories diverge from the initial adaptation path with the start of the decay period (Fig 3A and 3B gray lines). In both cases, the direction of change in gain is toward the origin of the gain-space. However, the gain-space trajectories never return completely to the origin, indicating only partial decay of the force-field adaptation within the period examined. This is in agreement to the asymptotic behavior seen at the end of the decay period for the adaptation coefficient (Fig 2A and 2B). Separation of the adaptation and decay gain-space trajectories for both pFF and vFF training demonstrate a difference in the behavior of the motor system during the decay of adaptation. The change in the goal-misaligned component between adaptation and decay dictates the shape of this separation. Fig 3C shows the gains for both the aligned and misaligned components for pFF training during the adaptation and decay period as a function of trial. The gain applied to the aligned component at the end of the decay period remained significantly greater than baseline levels, but this was not the case for the misaligned gain (aligned: -0.006 ± 0.004 vs. 0.12 ± 0.02, paired two tailed t-test, P < 0.001; misaligned: 0.009 ± 0.005 vs. 0.03 ± 0.02, paired two tailed t-test, P = 0.27). In order to capture the changes in the goal-misaligned component, we defined a third point in the gain-space trajectory. This 3rd point was the trial range during the decay period at which the gain of the aligned component was not significantly different from the respective gain during initial learning (trials 12–14 and 16–18 of the decay period for pFF and vFF, respectively). For example, for pFF training, there was no significant difference in the gain for the goal-aligned component between the 1st and 3rd points (0.33 ± 0.03 compared to 0.27 ± 0.03, P = 0.16, two-tailed t-test). We determined this point in order to isolate changes in the gain of the goal-misaligned component between adaptation and decay, and quantify the trajectory separation. For pFF training the goal-misaligned component was significantly different between the three different points (ANOVA, P < 0.001 for the main effect of period). The value of the goal-misaligned component at the 1st point was significantly greater than the respective gain at the 2nd and 3rd points (0.21 ± 0.02 compared to 0.09 ± 0.01 and 0.08 ± 0.02, P < 0.05 for both cases, multiple comparisons corrected) (Fig 3C). The difference between the 1st and 2nd points shows that the early adaptation level is less specific to the goal in comparison to late adaptation. The difference between the 1st and 3rd points further shows that for similar values of the goal-aligned component, adaptation and decay gain-space trajectories are significantly distinct. The goal-misaligned component was significantly different from zero for all 3 points, indicating that both adaptation and decay are confined in the 1st quadrant of the gain-space (P < 0.05, two-tailed t-test).
The behavior of the goal-misaligned component was slightly different for vFF training but the overall effect was the same. As above for pFF training, we compared the adaptation and decay gain-space trajectories at points where there was no significant difference in the gain of the goal-aligned component (between the 1st and 3rd points in Fig 3B, 0.24 ± 0.03 compared to 0.21 ± 0.03, P = 0.16, two-tailed t-test). Again, the goal-misaligned component was significantly different between the three different points (ANOVA, P < 0.05 for the main effect of period). Unlike pFF training, there was no significant difference in the gain of the goal-misaligned component between the 1st and 2nd points (0.09 ± 0.04 compared to 0.05 ± 0.01, P = 0.52, multiple comparisons corrected). Additionally, the value of goal-misaligned component at the 3rd point was significantly less than the respective gain at the 1st and 2nd points (-0.02 ± 0.01 compared to 0.09 ± 0.04 and 0.05 ± 0.01, P < 0.05 for both cases, multiple comparisons corrected) (Fig 3D). The goal-misaligned component was significantly different from zero for both early and late adaptation indicating that adaptation was confined to the 1st quadrant of the gain-space, but late decay showed a nominal, but negative gain for the position-dependent component. Similar to pFF training, the gain for the goal-aligned component at the end of the decay period remained significantly greater than baseline levels (0.008 ± 0.005 vs. 0.16 ± 0.03, paired two tailed t-test, P < 0.001 for both cases), but the misaligned component was not (-0.003 ± 0.003 vs. 0.002 ± 0.01, paired two tailed t-test, P = 0.75).
Although a separation between adaptation and decay was present in both vFF and pFF training gain-space trajectories, the shapes of the trajectories were not the same. We identified three differences between the gain-space trajectories. First, the initial learning for pFF training was less specific compared to vFF adaptation. That is, vFF adaptation was more aligned with the goal (parallel to the ordinate) compared to pFF training (parallel to the abscissa). Another way to quantify this difference is to determine the angle between the learning gain-space trajectory and the ideal (straight) trajectory to the adaptation goal. For early training (1st point) this angle was significantly greater for pFF training compared to vFF adaptation (pFF: 31.9° ± 2.6° vs. vFF: 18.6° ± 6.9°, P < 0.05, one-tail t-test). In other words, initial vFF training was more aligned with the learning goal (parallel to the velocity-dependent axis) than initial pFF adaptation (parallel to the position-dependent axis). Second, there was greater change in the learning gain-space trajectory for pFF training—a significantly larger difference was observed along the goal-misaligned gain axis between early and late adaptation for pFF adaptation (a difference in gain of 0.1 ± 0.02 for pFF compared to a difference of 0.03 ± 0.04 for vFF, two-tailed t-test, P < 0.05). Finally, although the decay of the goal-aligned component was slightly faster for vFF training (time constant of 10.4 ± 2.2 trials for pFF compared to 7.9 ± 0.9 trials for vFF. See S2A Fig), there was a much larger difference in the decay of the goal-misaligned component, with levels for pFF training significantly greater than vFF adaptation throughout the decay period (S2B Fig). In other words, velocity-based learning persisted at a nonzero value during the decay of pFF training. However, any subsequent position-based learning quickly decreased to zero for vFF training.
We hypothesized that these asymmetries for pFF and vFF training represent a possible intrinsic bias of the motor system to (1) associate the imposing perturbation with the kinematics of the movement and (2) retain the motion based learning. If the association between movement kinematics and the force-field perturbation is biased toward the velocity changes during the movement then adaptation to a force-field perturbation that is equally dependent on both position and velocity should be biased toward the velocity-dependent axis. This effect should also persist during the decay of the adaptation if velocity-dependent learning is more stable than that based on position. Sing et al. [16] previously studied adaptation to different force-field perturbations that were dependent on a combination of changes in limb position and velocity. However, the decay of this adaptation to different learning goals was not examined. Moreover, in their viscoelastic primitive model to describe the adaptation there was the basic assumption that motor learning based on changes in movement position and velocity is symmetric—an assumption challenged by the results described above. We therefore modified this model in order to make predictions about adaptation behavior and decay to novel movement dynamics dependent on different combinations of changes in limb position and velocity.
The viscoelastic primitive model proposed by Sing et al. [16] captured the changes in the temporal pattern of force during adaptation to different types of force-field perturbations. The application of this model to the vFF and pFF behavioral data are shown in Fig 4A. In this model the force pattern in each trial is a weighted sum of motor primitives that are differentially tuned to changes in position and velocity during the movement. On each trial the error between the current motor output in the 2D gain-space and the learning goal, combined with a gradient descent rule, determines how the weights of the respective primitives are updated. This model captures the initially similar motor output in response to the vFF and pFF perturbations, as well as the late-learning rotation of the gain-space trajectory toward the relevant motor learning goal (velocity for vFF training or position for pFF training). However, as mentioned above, this model assumes that the decay of the adaptation is the same for both types of motion-based learning (a symmetric primitive model). This similarity in retention results in a decay trajectory that travels directly back to the baseline value towards the origin. Interestingly, this decay structure makes testable predictions for force-field perturbations that combine velocity- and position-based learning (Fig 4A). First, utilizing the parameters determined from the simultaneous fit to the vFF and pFF behavioral data, for an unbiased combination (ucFF, equally dependent on both motion states) the learning and decay gain-space trajectories will be similar, with the decay closely following the reverse of the adaptation path. Second, using the same model parameters, the decay for adaptation to a position biased force-field (pcFF, a greater position and smaller velocity dependence) will be biased towards the position axis due to the greater representation of position-based learning at the end of training.
In our modification to this model we assume, based on the behavioral results above (see S2 Fig and S3 Fig), that the retention of learning based on changes in movement velocity is greater than the retention of learning based on changes in limb position. That is, during adaptation, the portion of the primitive population that encodes velocity information maintains a larger representation of this motion-based learning. We modeled this asymmetry by imposing that each primitive has two decay rates, one for position-based learning and one for velocity (see Materials and Methods). In this case, on each trial the amount of adaptation is scaled with different non-unity factors for position and velocity. We refer to this implementation as the asymmetric primitive model and, similar to the symmetric model, we applied this model to the vFF and pFF behavioral data and made predictions for force-field perturbations that combine velocity- and position-based learning (see Materials and Methods). When determining the values of the respective retention factors, we did not put any constraint on the relationship. Thus, the retention asymmetry could be in either direction, allowing a direct assessment of any difference.
This asymmetric model makes similar predictions as the symmetric model for the time course of adaptation for pFF and vFF training (Fig 4B). However, only the asymmetric model captures the small, but distinct separation in the decay of pFF and vFF training (see S3 Fig). In addition, the two models make distinct predictions about the decay for ucFF and pcFF training. As described above, the symmetric primitive model predicts equal adaptation to position and velocity, and decay along a similar trajectory for ucFF learning (Fig 4A and 4B). However, the asymmetric model (whose parameters are based solely on the vFF and pFF behavioral data) predicts that the final adaptation to this perturbation is biased toward velocity-based learning, and that the decay lies completely in the portion of the primitive space with more velocity contribution (αK = 0.942 vs αB = 0.951). When the two models simulate adaptation to the pcFF perturbation, the symmetric model predicts a decay that remains biased toward position. In contrast, the asymmetric model predicts a shift towards the velocity axis during decay. This is important in the sense that the adaptation endpoint in the primitive gain space imposes distinct decay characteristics under the two models that can directly be tested.
In order to further visualize the differences between the decay trajectories under the two models, we normalized the trajectories with respect to the end point of adaptation (Fig 4C). We did this to remove the effect of the adaptation endpoint for each force-field type, and more importantly reveal the difference between the decay rates. Under both ucFF and pcFF, the symmetric primitive model predicts a decay that follows the unity, x = y line. This is expected due to the same decay rates for position- and velocity-based learning. In contrast, the asymmetric model (whose parameters in this case are based only on the vFF and pFF data) predicts that the decay will be biased towards the velocity axis. The bias in the decay is predicted by the larger retention rate for velocity state compared to position (S6 Fig).
Although both models fit the pFF and vFF data qualitatively, it is important to note that there are aspects of the simulated adaptation that both models fail to capture (e.g., differences in the initial adaptation trajectory (magnitude and direction) between pFF and vFF, Fig 3A and 3B). For additional insight into these differences we focused on the predictions of the asymmetric model simulation, fitting the model separately to the vFF and pFF data in S4 Fig. Note that as in Fig 4, the values of the respective retention factors were not constrained. Consistent with Fig 4, in all cases the normalized decay trajectory is above the unity line demonstrating that velocity-based learning is decaying slower than position-based adaptation. Additionally, in S5 Fig and S6 Fig we show the influence of the primitive distribution on the learning trajectory and the influence of the retention rates on the decay trajectory. Finally, based on the same parameters in Fig 4, we also simulated the decay for adaptation to a velocity biased force-field (vcFF, a greater velocity and smaller position dependence, S7 Fig). Although the learning trajectory mirrors the pcFF simulation, the relative stability of the motion-based adaptation are consistent with Fig 4C.
To test the predictions of the symmetric and asymmetric primitive models we trained two additional groups of subjects in force-field perturbations that were unbiased (ucFF) and position biased (pcFF) combinations of the two motion states in order to further characterize the stability of velocity- and position-dependent learning. Previous studies have shown that the adaptation rate to a force-field with a positive correlated dependence on limb position and velocity is faster than adaptation to a purely position or velocity-dependent force-field [16]. Here, we examined how the ratio of position and velocity dependence influenced the motor adaptation and stability during the decay period. As described for the simulations above, we first examined adaptation and decay in response to a force-field equally dependent on both motion states (ucFF). Following this, we examined learning and the subsequent decay for movements made within a combination force-field with a greater position and smaller velocity dependence (pcFF).
We observed that the adaptation to an unbiased combination force-field (equally dependent on the state of the position and velocity during the movement) was generally closer to the learning goal by the end of the adaptation period (Fig 5A). The applied gain was significantly different between the late periods of adaptation and decay, and between the two types of motion states (2-way ANOVA, P < 0.001 for both the main effect of period and the main effect of motion state). Subjects initially adapted to the force-field by applying similar state-dependent gains for changes in position and velocity (position: 0.20 ± 0.03, velocity: 0.28 ± 0.04, P = 0.09, paired two-tailed t-test). However, by the end of the adaptation period, the velocity-dependent gain was significantly greater than the position-dependent gain (position: 0.54 ± 0.03, velocity: 0.68 ± 0.04, P < 0.05, paired two-tailed t-test). This resulted in a gain-space learning trajectory that was above the unity line and clearly biased towards the velocity-dependent gain axis (ordinate in Fig 5A).
To examine the characteristics of this bias, we projected the gain-space trajectory onto the position and velocity gain axes at each point during learning and decay (Fig 5C). As described above, early in adaptation the position- and velocity-dependent gains had similar magnitudes, but the velocity-dependent gain was significantly greater than the position-dependent gain by the end of the adaptation period. This significant difference between the velocity- and position-dependent gains extended throughout the decay period (Late in decay: position: 0.07 ± 0.02 compared to velocity: 0.18 ± 0.03, P < 0.05 for both cases, paired two-tailed t-test, Fig 5C bar graph). An exponential fit to the decay of the velocity and position components showed a larger time constant for velocity (time constant of 9.4 ± 1.8 trials for position compared to 12.1 ± 2.0 trials for velocity. See S8A Fig). Additionally, the applied gains based on velocity and position at the end of the decay period remained significantly greater than baseline levels (velocity: 3.8 x 10−4 ± 0.007 vs. 0.18 ± 0.03, position: 0.003 ± 0.007 vs 0.07 ± 0.02, paired two tailed t-test, P < 0.001 for both cases). This clearly shows that when the force-field is equally dependent on changes in movement position and velocity, the gain of the velocity-dependent force contributed more in both the adaptation and decay periods. This is in agreement with the predictions of asymmetric primitive models, which suggest a bias late in adaptation toward velocity continuing throughout the decay period (Fig 4B).
One might suspect that the observed bias in the decay trajectory for the unbiased combination force-field is the consequence of the unbalanced adaptation levels; the final adaptation has a significantly greater velocity-dependent gain compared to position. In order to remove this confound, we normalized the gain-space trajectory during the decay by the position and velocity-dependent gains by the respective values at the beginning of the decay period. Thus, the rescaled initial point of decay in gain-space is located at [1.0, 1.0]. If the shape of the decay gain-space trajectory in Fig 5A was the result of unequal learning at the end of adaptation, then the normalized decay should be aligned with the equality line in gain space. However, the normalized trajectory clearly shows that the velocity-dependent gain was always greater than the respective position-dependent gain throughout the decay period (Fig 5E). When we examined the temporal changes of the normalized gains during decay (Fig 5E) by projecting the trajectory onto the position and velocity-dependent gain axes, we observed the same effect (Fig 5G). The percentage of adaptation at the beginning of the decay period was significantly greater than at the end, and the percentage of adaptation based on velocity and position was significantly different (2-way ANOVA, P < 0.001 for the main effect of period and the main effect of motion-based learning). For the early and late epochs of decay, the normalized velocity-dependent gain was significantly greater than position (early epoch: position: 37 ± 7% vs. velocity: 52 ± 6%; Late epoch: position: 12 ± 5% vs. velocity: 25 ± 5%; P < 0.05 for all cases, paired two-tailed t-test). This is in line with the decay predicted by the asymmetric primitive model and suggests that there is an asymmetry in the retention rates between position- and velocity-based motion-state learning (Fig 4C).
As stated previously, a potential confound for the unbiased combination force-field is that the velocity-dependent gain at the end of adaptation was significantly greater than position. This may have influenced the decay and resulted in the velocity contribution being more stable throughout the decay period. We therefore conducted an additional experiment using a position-biased combination force-field (pcFF). As predicted by both models, at the end of training the adaptation gain-space trajectory for this force-field is biased toward the position axis (abscissa) as shown in Fig 4A and 4B. However, as the decay period starts, the asymmetric model predicts that the gain-space trajectory will move toward the velocity-dependent gain axis (ordinate) and remain above the unity line for the remainder of the decay period (Fig 4B). In contrast, the symmetric model predicts that the adaptation will decay towards the position axis (Fig 4A)
Fig 5B shows the behavioral results for subjects trained on this combination force field. The decay of adaptation is clearly biased towards the velocity axis, consistent with the predictions of the asymmetric model. This can be seen in the trial-by-trial changes of both gains during adaptation and decay period (Fig 5D). The applied gain was significantly different between the late periods of adaptation and decay, but there was no main effect of motion state (2-way ANOVA, P < 0.001 for the main effect of period and P = 0.27 for the main effect of motion state). (Note that the non-significant effect of motion state is due to significant effects in opposite directions in the late periods of adaptation and decay. See below.) Similar to the ucFF results, an exponential fit to the decay of the velocity and position components showed a larger time constant for velocity (time constant of 5.7 ± 0.6 trials for position compared to 9.3 ± 0.9 trials for velocity. See S8B Fig). In addition, the final gains applied to velocity and position at the end of the decay period remained significantly greater than baseline levels (velocity: 0.001 ± 0.007 vs. 0.10 ± 0.03, position: 8.9 x10-4 ± 0.007 vs. 0.06 ± 0.02, paired two tailed t-test, P < 0.001 for both cases). Although the adaptation starts with equal contribution of both motion components, late adaptation is significantly dominated by the position-dependent learning (Early adaptation: position: 0.26 ± 0.05 vs. velocity: 0.26 ± 0.03, P = 0.99, paired two-tailed t-test; Late adaptation: position: 0.61 ± 0.03 vs. velocity: 0.53 ± 0.02, P < 0.05, paired two-tailed t-test). At the start of the decay period there is a rapid drop in the position-dependent gain. However, the decay of the velocity-dependent gain is much slower, resulting in the gain-space trajectory remaining above the unity line throughout much of the decay period (Late in decay: position: 0.06 ± 0.02 vs. velocity: 0.10 ± 0.03, P < 0.05, multiple comparisons corrected).
Due to the significant difference in the gain magnitudes at the end of the training, we also examined the normalized decay for pcFF training. Similar to the results for the unbiased combination force-field, we observed that the normalized decay gain-space trajectory was above the unity line for the entire decay period, indicating that the velocity-dependent gain decayed at a slower rate than the position-dependent gain. In addition, the percentage of adaptation at the beginning of the decay period was significantly greater than at the end, and the percentage of adaptation based on velocity and position was significantly different (2-way ANOVA, P < 0.001 for the main effect of period and the main effect of motion-based learning). This effect is strongly present in both early and late epochs of the decay period (Early, position: 24 ± 5% vs. velocity: 42 ± 6%; Late position: 11 ± 4% vs. velocity: 20 ± 6%; P < 0.05 for both cases, paired two-tailed t-test) (Fig 5F and 5H). This again is in agreement with the simulations from the asymmetric primitive model (Fig 4B and 4C, and insets in Fig 5A, 5B, 5E and 5F). When this model was applied simultaneously to the ucFF and pcFF behavioral data the simulations (insets in Fig 5A, 5B, 5E and 5F) predicted adaptation during training to be biased toward the position axis, but with the start of the decay, the gain-space trajectory is biased towards the velocity axis due to an asymmetry in the stability of the motion-state learning (αK = 0.9424 vs αB = 0.9654).
We designed a series of experiments to examine the stability of motion-state based updates to motor commands in response to the introduction of novel dynamics during reaching movements. In our first experiments these dynamics were dependent either solely on changes in movement position or velocity. We directly measured the temporal force patterns subjects applied via error clamp trials, during which the robotic manipulandum constrained the movement to a straight trajectory between targets by counteracting any perpendicular motion. Based on the force patterns subjects applied to counter the perturbation, we determined the gain associated to changes in movement position and velocity. We determined the applied gains in response to different types of dynamics and examined the stability of these modifications to motor commands when the perturbation was removed and subjects only made error clamp movements. When the respective gains were represented in a two-dimension gain space we observed a separation between the gain-space trajectories during the learning and the decay periods. Based on the observed behavioral differences in the retention of the learning between pFF and vFF training, we modified a previous model of motor adaptation and made several predictions on the decay of learning following training in force-field perturbations that were a combination of both motion states. Interestingly, the simulations predicted that when the learning goal had partial dependence on both motion states the position-dependent gain would decay at a faster rate relative to the velocity-dependent gain. This was the case even when the gain associated to position at the end of training was significantly greater than that applied to movement velocity. These simulations were confirmed by a second set of experiments in which we examined the learning and decay to these combination motion-state perturbations. Together, our simulation and behavioral results suggest that (1) overlapping, but distinct processes underlie motor adaptation and its decay and (2) the adjustment of motor commands based on movement velocity is relatively more stable than that based on position.
The gradual decay of newly formed motor memories has been studied for different contexts and tasks, including: prism [6,7], locomotion [8,9], visuomotor [10,11] and force-field perturbations [4,12,15]. The decay of adaptation in these studies is often at a different rate compared to initial learning suggesting at least partially separate mechanisms [23,27]. Recently, Kitago and colleagues [5] examined the decay of visuomotor adaptation for different types of assessments. For all methods examined, there was a decay in the adaptation, but the rate was the fastest when the perturbation was removed and slowest when the errors orthogonal to the ideal movement trajectory were visually clamped to zero (similar to the error clamps used in the current study). (Note the minimum decrease in adaptation level occurred with the passage of time, but the decay rate is difficult to quantify and compare for this context.) This suggests that the method in which the decay rate is assessed (i.e., the context of the decay period) potentially has a considerable influence on the decay rate of the motor memory [13,14,28].
In order to evaluate modifications to the feedforward changes in the motor output, it was necessary to utilize error-clamp trials. Although it is possible that assessing adaptation decay in a different manner (e.g., decay trials with perturbation removal) could influence the reductions in motor output we report, our main interest was how these changes in motor output compared to the initial learning and varied for different learning goals. Our results show clear separation in the applied gains to changes in position and velocity between the initial learning of the novel dynamics and the subsequent decay of the motor adaptation. In all cases, the applied gain did not return to baseline levels. This was true for the goal aligned motion state (Fig 3C and 3D) when the perturbation was based on a single state, and when the perturbation had a codependence on both motion sates (Fig 5C and 5D). This behavioral difference mirrors the neuronal retention of learning reported throughout the sensorimotor system following motor adaptation. The activity of a population of premotor, supplementary motor, and primary motor cortex cells is modified during force-field adaptation and these correlated modifications are retained during the decay of the learning, serving as a memory trace of the training [29–33]. The behavioral difference between learning and decay we report may reflect these persistent neural changes throughout the sensorimotor system specifically tuned to the motion kinematics required for force-field compensation.
Although the focus in our study was the examination of the decay of the velocity- and position-dependent learning, there are some aspects of the adaptation trajectories that are not captured by the primitive model (e.g., differences in the initial adaptation trajectories in Figs 3 and 5, compared to the simulations in Fig 4). These inconsistencies may be due to several interesting factors resulting from perturbation-dependent changes in the primitive distribution. The supplemental simulations (S5 Fig and S6 Fig) suggest that there are possible changes in the primitive distribution (at least within this computational framework) occurring during the two types of training (vFF and pFF) that influence the learning trajectories; it is possible that the primitive space may rotate during training, but the extent towards a particular axis may have different rates. We plan to address this possibility with future, systematic experiments.
Finally, the sensory adaptation that occurs with motor learning may provide an additional measure to assess differences in adaptation retention. Ostry and colleagues [34] demonstrated that following the exposure to a force-field movement perturbation there was an accompanying modification in the perception of limb position. Specifically, the perceptual shift was in the direction of the movement disturbance and learning-dependent; there was no observed sensory modification when the limb was moved passively through the same trajectories experienced during the motor perturbation. Thus, another possible assessment of any asymmetry in the retention of the velocity and position components could be to compare the magnitude of the accompanying perceptual shifts in limb estimation and the degree to which these perceptual modifications persists throughout the decay period.
Several studies have suggested that how the perturbation is introduced (e.g., abruptly vs. gradually) and the duration of exposure (e.g., long vs. short) influence the stability and subsequent properties (transfer, long-term retention, etc.) of motor adaptation [21,35–37]. For example, Huang and Shadmehr [22] showed that when the force-field perturbation was applied for a short duration, the decay of the adaptation was much more rapid than for longer training periods, suggesting less relative stability in the modifications to the motor commands. This is in agreement with recordings in motor cortex [38]; the activity of a subset of neurons is modified dependent on the rate of movement perturbations experienced, indicating that the neural representation of adaptation is influenced by the training schedule.
In addition to training schedule, previous studies have examined factors that influence the stability of adaptation retention [12,17–21,39,40]. However, an important distinction of the current study is that we examined the stability of the components of the motor adaptation (the motion-state based learning) rather than the long-term stability of the adaptation or stability in competition with the formation of other motor memories. As in Sing et al. [16], our current observations show that the motor memory in the late stages of training is more specific to the task goal compared to the initial stages due to modification in the gains applied to the goal-aligned and goal-misaligned motion parameters. This specificity in the motor output remains throughout the decay period; there is no reemergence of the initial goal-misaligned learning as the acquired goal-aligned learning gradually decays. Based on the collective work described above, it would be interesting to examine the influence of the (1) training duration, (2) introduction rate and (3) passage of time on the stability of these adjustments to the motion state gains. For example, there is recent evidence that performance becomes more task specific with sufficient breaks after training, suggesting that the passage of time may influence the ability to perform more task-relevant actions [41]. We hypothesize that the well-known savings following a break after initial training (faster re-adaptation with exposure) will reflect more goal-aligned movements [10]. That is, savings over multiple days of training should result in adaptation gain-space trajectories closer to the goal-aligned axis (the motion kinematic of the experienced force-field) than on the first day of exposure.
As demonstrated previously [16,24,25], the initial adaptive responses that we observe when learning novel movement dynamics are consistent with motor primitives with correlated position and velocity tuning. This theoretical framework is based on the codependent encoding of these motion states observed throughout the sensorimotor system [42–47]. Our current results suggest that this codependence does not necessarily result in an equal representation of the two motion states, but rather codependent processing biased towards velocity. For example, similar to previous studies [16,24,25], initial position-dependent learning is biased towards the velocity-dependent gain; the gain-space trajectory is typically closer to the middle of the gain space than that observed for velocity-dependent learning (Fig 3A and 3B). In addition to initial learning, the decay of the position-dependent gain was relatively faster than the reduction of the velocity-dependent gain for both combination force-fields, suggesting an asymmetry in relative stability (Fig 5).
Why should learning based on movement velocity be more stable than that based on position? A possible answer may be found in the encoding asymmetries in motor cortex [47]. Velocity tuning among primary motor cortex neurons is more abundant compared to position. Another reason for a velocity bias could be that movement velocity provides substantially more motion information compared to position. For example, during point-to-point movements, there can be a significantly larger variance in the temporal changes in movement velocity for similar movement trajectories [48]. Take for example the force-field perturbations used here; the peak force experienced by the subject can vary broadly when based on movement velocity, whereas this peak is restricted when based on positional changes. If such a coding bias exists throughout the sensorimotor system, this imbalance would support a preference towards velocity-based learning during initial force-field adaptation and an asymmetry during the subsequent decay. Possible support for this bias may be found in a recent study by Rotella and colleagues [49]. The authors asked subjects to produce isometric hand forces which were then mapped to the position or velocity of a virtual cursor. Under these different mappings, they then tested the generalization of adaptation when a visuomotor rotation was applied to the cursor motion. Interestingly, the generalization of adaptation under the velocity mapping was broader, which is aligned with the current implications that movement velocity is a more stable basis for motor learning than changes in position.
We investigated the decay of short-term adaptation to motion-dependent perturbations applied to reaching movements. We observed a clear separation between the initial learning and subsequent decay when the motor output was represented as the respective gains subjects applied to changes in position or velocity during movement. When the perturbation was only based on one motion state (position or velocity), this separation was a direct effect of a sustained decrease in the gain of the goal-aligned motion parameter, with no reemergence of the goal-misaligned parameter during the decay period. When exposed to novel dynamics that required a combination of position- and velocity-dependent learning, the applied velocity-dependent gain was relatively more stable during the decay period, even when the gain applied to changes in position was significantly greater at the end of training. This difference in the relative state-dependent learning stability suggests that the motor system has an inherent preference towards adjusting and retaining modifications to motor commands based on movement velocity. A modified model of adaptation that accounts for greater retention of velocity-based learning captures these behavioral results, and importantly predicts the decay behavior for training with novel force-fields that are jointly dependent on the two motion states. Overall our results show that the decay of motor adaptation is not exactly unlearning—the complete reversal of the learning process. Rather, in agreement with previous physiological and behavioral studies, our results suggest that the decay of adaptation likely shares overlapping mechanisms with the learning process, but is a distinct process that reduces the motor memory traces formed over the training period.
The study protocol was approved by the George Mason University Institutional Review Board, and all participants gave informed written consent.
Fifty-six healthy subjects (37 male and 19 female) without known neurological impairment were recruited from the George Mason University community to participate in the study. All participants were right-handed and performed the task using their right hand. Each individual participated in only one of the experimental sessions and experienced only one type of force-field (14—Velocity-dependent Force-field, 14—Position-dependent Force-field, 14—Unbiased Combination Force-field, and 14—Position-biased Combination Force-field).
The experimental paradigm was based on the standard force-field adaptation paradigm [26]. The subjects were instructed to move a cursor between two targets located on a screen in the sagittal axis of their body while grasping a robot manipulandum (KINARM End-Point Lab, Fig 1A). The manipulandum measured hand position, velocity, and the force applied by subjects, and its motors were used to apply forces to the hand, all at a sampling rate of 1000 Hz. A semi-transparent mirror was used to project the location of hand and visual targets to the plane of movement while occluding the subject’s view of the hand (refresh rate of 60Hz). During the experiment the subjects reached to circular targets 0.6 cm in diameter that were spaced 10 cm apart on the sagittal axis of the body. The subjects were instructed to ‘‘make quick reaching movements to the targets in both the forward and backward directions.” At the end of each trial, subjects received visual and auditory feedback about the completed movement. If the peak movement velocity was between 0.25–0.35 m/s and the movement duration was shorter than 750 ms, the reach target (green target in Fig 1A) filled green with an auditory reward indicating a movement within the required criteria. If the peak movement speed was below 0.25 m/s, the reach target filled yellow to indicate that the movement was too slow. If the peak movement speed was above 0.35 m/s, the reach target filled red to indicate the movement was too fast. In both of the latter cases no auditory feedback was given. The endpoint of each movement was used as the start point for the following trial, and movements were made only in these two directions. The subjects received a performance score at the end of each block of movements that indicated the percentage of correct trials only in the trained 270° movement direction. Subjects were asked to maintain the score above 50% throughout the experiment. Only 270° movements with a peak velocity between 0.2–0.4 m/s were used in the subsequent data analysis. In addition, subjects had to initiate their movement within 75–2000 ms after the reach target appeared on the screen. Otherwise all targets were extinguished and the trial was immediately repeated.
Three trial types were used during the experiment: null trials, force-field (FF) trials, and error-clamp (EC) trials (Fig 1B). Null trials were used for initial practice, during which the motors of the robot manipulandum did not apply any force to the hand. During FF trials, the robot applied a force at the hand that was dependent either on movement position (with respect to the start location), velocity, or a positive combination of limb position and velocity. The force that the robot applied to the hand was always orthogonal to the direction of movement, and had the general form of:
[FxFy]=cK.[0−KK0].[xy]+cB.[0−BB0].[x˙y˙],K=45N.sm,B=15Nm
(1)
For a position-dependent force-field trial (pFF), cK = ±1 and cB = 0, where cK = ±1 and cK = −1 correspond to clockwise and counterclockwise direction of the force-field, respectively (a clockwise force-field is shown in Fig 1B). For a velocity-dependent force-field trial (vFF), cK = 0 and cB = ±1. Unbiased combination force-field trials (ucFF) had a force pattern dependent on both the position and velocity, with cK = ±0.71 and cB = ±0.71 for clockwise and counterclockwise directions [16,23,24]. Lastly, the Position-biased combination force-field trials (pcFF) had a motion dependent force pattern similar to the ucFF. However, the contribution of the position-dependent component was 20% greater and the velocity-dependent component was 25% less, with cK = ±0.85 and cB = ±0.53. As in Sing et al. [16] the values for K and B were chosen in order to have approximately equal peak perturbing force for vFF and pFF. Each subject experienced only one type of force-field throughout the experimental session.
During error-clamp trials, the robot motors constrained movements in a straight line toward the reach target by counteracting any motion perpendicular to the target direction [21,50]. This was achieved by applying a stiff one-dimensional spring (6 kN/m) and a damper (150 Ns/m) in the axis perpendicular to the reach direction. In these trials, perpendicular displacement from a straight line to the reach target was held to less than 0.6 mm and averaged about 0.2 mm in magnitude.
Each subject experienced the same basic experimental paradigm shown in Fig 1C. Subjects performed sets of 90° and 270° movements. Each experiment started with a baseline period, during which subjects completed 360 null trials (180 movements in the trained 270° direction). These null trials were divided into 4 blocks. The first two blocks had 80 movements each and the last two blocks each required 100 movements. During the last 2 blocks of trials 12 error-clamp trials were pseudo-randomly interspersed for the 270° movement direction in order to measure the baseline levels of forces for each subject. The average lateral forces during these trials were then subtracted from the forces applied on error-clamp trials during the adaptation and decay periods.
Following the baseline period, subjects experienced the adaptation transition block (124 total movements, 62 in the trained 270° direction), during which the force-field environment was suddenly introduced after an initial 30 null movement trials (15 in the trained 270° direction). We designed the adaptation transition block to capture the immediate changes in the applied force due to initial exposure to the force-field. Once the perturbation was introduced all 90° movements were made under the error-clamp condition, and the force-field was only applied to 270° movements. For the first 10 training trials, the ratio of force-field (FF) to error clamp (EC) trials was 3 FF: 2 EC which was then reduced to 5 FF: 1 EC for the last 84 trials (42 in the trained direction). The adaptation transition period was followed by 2 blocks of training (96 total trials each) in which the subjects experienced only one of the four force-field environments (vFF, pFF, ucFF, or pcFF). Similar to baseline period, we pseudo-randomly inserted 16 error-clamp trials in the 270° movement direction in order to measure the adaptation level at different points in training. The ratio of 5 FF: 1 EC was maintained throughout this training period. Only the 270° direction movements were used for analysis of adaptation and decay. The sign (direction) of the FF remained constant for each subject, but was counterbalanced between subjects.
After the adaptation period, subjects experienced the decay transition block of 146 total trials. This block started with 26 training trials. For the first 12 trials, there was a ratio of 5 FF: 1 EC which increased to 4 FF: 3 EC for the last 14 trials in order to obtain an accurate measure of final adaptation levels. These 26 trials were then followed by 120 consecutive error-clamp trials (60 in the trained 270° direction). We refer to these 120 error-clamp trials as the decay period, during which the adaptation decayed to the baseline levels prior to experiencing the force-field. Inclusion of the decay period within the transition block effectively masked any possible context dependent changes in the behavior of the subject due to the removal of the force-field [14,28]. We used 60 consecutive error clamp trials to measure adaptation decay in order to keep this critical experimental block within a reasonable duration, and avoid breaks and possible cognitive influences during the transition to the decay period. In addition, based on the exponential decay time constants (S1 Fig, S2 Fig, S8 Fig), this number of error clamp trials proved sufficient to observe asymptotic levels of decay.
As described previously [16,21,37,50,51] we used error-clamp trials to measure the change in feedforward motor output during the adaptation and decay periods. The use of error-clamp trials reduces the lateral errors experienced during the movement that elicit online feedback correction. Given that the lateral force during error-clamp trials reflects the predictive feedforward adaptive response to the force-fields, we limited our analysis to these force patterns. Based on Eq 1, subjects fully compensate for the force-field when they produce a countering force that is proportional to the movement velocity, position, or the positive combination of the two. We first computed the ideal force pattern by examining the longitudinal movement kinematics (position, velocity) during the error-clamp trial movement. The movement and force signals were analyzed within a temporal window of 1500 ms centered on the peak velocity (±750). Next we defined the adaptation coefficient by determining the linear regression coefficient between the ideal force and the lateral force applied by the subject during the error-clamp trials [16,21,37,50,52]. We computed the adaptation coefficient for each subject during both the adaptation and decay periods and averaged the values over all subjects. In all cases we provide the SEM of this average value.
We further characterized the adaptation and decay behavior by projecting the lateral force during each error-clamp trial onto a two-dimensional space that parsed the position-dependent and velocity-dependent components of the applied force [16]. We refer to this two-dimensional space as the gain-space. This gain-space represents complete adaptation to a vFF by the point [0,1], pFF by the point [1,0], ucFF by the point [0.71, 0.71], and a pcFF by the point [0.85, 0.53]. Additionally, the abscissa and ordinate of each point in this gain-space corresponds to the position-dependent and velocity-dependent components of the applied force. In order to depict adaptation and decay in gain-space, we first calculated a multiple regression between the lateral force during the error-clamp trials, and both the changes in position and velocity during the movement. We then rescaled the coefficients for the position and velocity components by the 45N/m and 15Ns/m factors, respectively, and projected these coefficients onto the gain-space. For each subject, we performed this analysis and calculated the average gains over all subjects [16,23–25].
Similar to Sing et al. [16], the characterization of position and velocity contributions in the force output produce excellent fits (R2 values ranging from 0.91 to 0.99, see Fig 2). As in this prior study, the inclusion of an acceleration term resulted in highly significant but relatively small improvements in the representation of these force profiles; in the majority of cases for the different types of perturbations (vFF, pFF, ucFF, and pcFF; early and late) the acceleration signal’s contribution was significant (P < 0.001 in all cases except early pFF training, P = 0.38), but the overall force profile variance accounted for only improved by at most 3%. We therefore elected to focus only on the contributions of the position and velocity state variables.
We operationally defined early and late/asymptotic adaptation as the first 15 (1–15) and last 10 (150–160) trials of training. Thus, the mean and standard error values for these periods are plotted as a function of the mean trial number within these windows for the adaptation period (Figs 2, 3 and 5). The data during the decay periods were normalized by dividing all the subject data by the mean (across subjects) of the first decay trial. Due to the increased frequency of EC trials during the decay period, we used a smaller window to assess early and late levels (trials 11–20 and 50–60 respectively). Here we excluded the first 10 trials in the analysis for the early epoch in order to remove the effect of normalization of adaptation gains (Fig 5G and 5H). We initially tested the main effect of group condition on the different epochs of interest with a repeated measures ANOVA and subsequently determined the epoch in which these conditions were significantly different with post-hoc analysis. For example, two-tailed t-tests were performed between different force-field groups to compare the behavior within each epoch. For all tests the significance level was 0.05.
In S1 Fig, S2 Fig and S8 Fig we applied a standard exponential model with rate and offset parameters to determine the time constants of learning and decay for the different types of perturbations (vFF, pFF, ucFF and pcFF) and learning components (goal-aligned and goal misaligned, velocity- and position-based). We computed the standard deviation of the best-fit parameter values for these model fits by bootstrapping the fits to the data. We made 500 different bootstrap estimates of the fit parameters, each by averaging data from 14 randomly generated choices made from the 14 subject data pool with replacement. We fit the model to each of these bootstrap estimates and determined the standard deviation of each parameter.
The viscoelastic primitive model first proposed by Sing et al. [16] consists of N motor primitives, Si = [Ki Bi]T = Rn×2, which collectively generate the motor output. The primitives are jointly distributed as
[KiBi]~N(μ,Σ),
μ=[00];
Σ=[σK2ρσKσBρσKσBσB2]
In this model these primitives have a similar dependency on position and velocity via σK = σB. Moreover, the correlation between the primitives is determined by ρ.
The motor output on each trial is determined by a weighted combination of motor primitives. Each primitive receives input from the changes in position and velocity during the movement and creates a force output:
FSi=[KiBi]T[PV]
Given that the vector [P V]T is shared between all primitives, we can factor out this vector and simplify the calculation. The final force is a weighted linear combination of the primitive forces:
Foutput=∑i=1nwiFSi
[KoutputBouput]T[PV]=∑i=1nwi[KiBi]T[PV]
[KoutputBouput]=∑i=1nwi[KiBi]T=STW
In this equation the W ∈ Rn×1 is a weight vector that drives the learning in the model. The output vector [Koutput Boutput]T represents the gain in position-velocity (p-v) primitive gain-space, which we refer to as y, or the current motor adaptation state. The goal of adaptation can be defined as a vector y* ∈ R2×1. On each trial of adaptation we can project the error vector between the goal and the motor output and use a gradient descent rule to compute the weight change for each primitive Si as
dwin=ηSi(y*−yn−1)
The weight change can be used to create a new gain state in the p-v primitive gain-space yn+1
yn+1=[αK00αB]yn+STdW
For the symmetric model, the retention value for αK and αB are the same. When we applied this model to the vFF and pFF behavioral data (Fig 3) we estimated these parameters to be: αK = αB = 0.951, σK = σB = 0.401, η = 1.5 x 10−4, ρ = 0.51. As with the exponential fits, we made 500 different bootstrap estimates of the fit parameters, each by averaging data from 14 randomly generated choices made from the 14 subject data pool with replacement.
For the asymmetric model, the retention values were not constrained and can result in an asymmetry in either direction. When we applied this model to the vFF and pFF behavioral data we estimated these parameters to be: αK = 0.942, αB = 0.951, σK = 0.464, σB = 0.379, η = 1.5 x 10−4, ρ = 0.47. Note that the retention is biased towards velocity primitives, αK < αB. In addition, when asymmetric model is applied simultaneously to the ucFF and pcFF behavioral data these parameters were estimated to be: αK = 0.914, αB = 0.958, σK = 0.546, σB = 0.565, η = 1.5 x 10−4, ρ = 0.48.
|
10.1371/journal.pgen.1005073 | Sex Ratio Meiotic Drive as a Plausible Evolutionary Mechanism for Hybrid Male Sterility | Biological diversity on Earth depends on the multiplication of species or speciation, which is the evolution of reproductive isolation such as hybrid sterility between two new species. An unsolved puzzle is the exact mechanism(s) that causes two genomes to diverge from their common ancestor so that some divergent genes no longer function properly in the hybrids. Here we report genetic analyses of divergent genes controlling male fertility and sex ratio in two very young fruitfly species, Drosophila albomicans and D. nasuta. A majority of the genetic divergence for both traits is mapped to the same regions by quantitative trait loci mappings. With introgressions, six major loci are found to contribute to both traits. This genetic colocalization implicates that genes for hybrid male sterility have evolved primarily for controlling sex ratio. We propose that genetic conflicts over sex ratio may operate as a perpetual dynamo for genome divergence. This particular evolutionary mechanism may largely contribute to the rapid evolution of hybrid male sterility and the disproportionate enrichment of its underlying genes on the X chromosome – two patterns widely observed across animals.
| Millions of species live on Earth, thanks to an evolutionary process that splits one species to two or more new species. The formation of new species is benchmarked by the evolution of reproductive isolation (RI) such as hybrid sterility between new species. The fundamental question of how RI evolves, however, remains largely unknown. In a pair of very young fruitfly species, we localized six loci expressing dual functions of hybrid male sterility (HMS) and sex ratio distortion, implicating an evolutionary causal link between these two traits. The rapid evolution of HMS widely observed across animal taxa can be attributed to the rapid evolution of genes controlling sex chromosome segregation. All genes in a genome are not equal. This study suggests that conflicts among various parts of a genome might confer strong evolutionary pressure—a mechanism that has hitherto been regarded as rare and could actually be more ubiquitous than currently appreciated.
| Intrinsic reproductive isolations (RI) between two newly evolved species can take the forms of hybrid male sterility (HMS), hybrid female sterility (HFS) and hybrid inviability (HI), all manifestations of genetic incompatibilities between two genomes [1]. Speciation genetics studies typically start with genetic analysis of divergent reproductive traits between two species. Numerous genes underlying interspecific divergence have been identified [2,3], but they cannot be automatically qualified as “speciation genes” because some interspecific divergence may have evolved only after speciation was complete. The identification of genes underlying HMS, HFS and HI—also called Dobzhansky-Muller incompatibility (DMI) genes—by themselves, even with their biological functions well understood, can rarely answer which DMI genes are involved in establishing the initial RI, and what adaptive phenotypes of these genes are responsible for their fixations in one but not the other lineage [2]. Thus the evolutionary mechanism(s) for evolving DMI at the initial stage of speciation still remains a mystery.
Nevertheless, two patterns have emerged from extensive speciation genetic studies in the last three decades. The first is the “faster male” evolution in that HMS evolves at a rate an order of magnitude higher than HFS and HI [4], presumably caused by sexual selection [5]. The second is the “large X” evolution in that HMS genes are enriched on the X chromosomes [6–8], presumably caused by more efficient fixation of mutations on the X than on autosomes [9]. However, sexual selection would also make hybrid ZZ males more likely to be sterile than hybrid ZW females, but this prediction is not supported by empirical observations [4]. Similarly, efficient selection of X-linked genes would also predict the “large X” pattern for the HI genes but no empirical support has been garnered either [10]. Thus, neither the “faster male” nor the “large X” pattern has been sufficiently accounted for by any evolutionary theories as well as the associated empirical evidence.
The above two patterns can be better explained by the “conflict theory” in that genomic divergence is driven by selfish genes, prominently by sex ratio distortion (SRD), also called sex chromosome meiotic drive [11–13]. Meiotic drive distorter breaches Mendel’s first law of genetics by gaining more than 50% transmission while quenching its homolog’s share in the gene pool of next generation. The distorter, however, does not commit suicide because of the tightly linked insensitive responder, while its homolog is linked to the sensitive responder. Meiotic drive is generally harmful to a genome, thus suppressors to silence the distorter are under strong selection to evolve and make the meiotic drive cryptic [14]. A tight linkage between the distorter and the responder is a key requirement for a meiotic drive system to evolve [15]. This prerequisite is readily satisfied on the two heteromorphic sex chromosomes, between which recombination is generally absent. Sex chromosome meiotic drive manifests as unequal sex ratio. For a typical XY male, the optimum sex ratio is all females for the X-linked genes but all males for the Y-linked genes, and 50% females for all autosomal genes. Therefore, the optimum sex ratios are at odds from the perspectives of various portions within a genome [16]. If SRD arises repeatedly on the X chromosome, counter evolution on the Y and the autosomes is anticipated, so much so that the SRD operates as a perpetual dynamo for genome evolution and bouts of this distortion-suppression process eventually lead to speciation [13]. The “conflict theory” can readily account for the “faster male” evolution because SRD occurs in XY male, and the “larger X” evolution because this chromosome contributes about half of the genetic changes in the evolution caused by SRD [13]. The “conflict theory” also predicts “faster female” in ZW females [4], and a faster pace of RI evolution in taxa with heteromorphic sex chromosomes than those without.
The best evidence for the “conflict theory” comes from two HMS genes with dual functions of SRD and HMS: tmy mapped between D. simulans and D. mauritiana [17], and Ovd identified between D. pseudoobscura USA and D. p. Bogota [18]. However, these SRD systems could have evolved after speciation. Many other HMS genes are also mapped in these species but they do not have the SRD phenotype [19,20], so are almost all the other known HMS genes across all taxa. Therefore HMS seems to have evolved by mechanisms generally unrelated to SRD. On the other hand, absence of SRD phenotypes in hybrids can be explained by the absence of idiosyncratic genetic background required for SRD expression, gene silencing and loss of function in cryptic SRD systems, or sterility of hybrids. Indeed, there might be an intrinsic difficulty to test the “conflict theory” because the SRD expression is usually transient.
We reasoned that an ideal empirical system for identifying the bona fide “speciation genes”, to test the “conflict theory” or any other theories of speciation for that matter, would be a pair of species at the very incipient stage of speciation, when the HMS just starts to evolve and is directly responsible for establishing the initial RI. Two Drosophila species, D. albomicans and D. nasuta, appear to be such a system because of their young age of ∼120 kyrs [21]. D. albomicans is distributed from Okinawa of Japan through South China, Indochina to Northeast India, while D. nasuta is found in East Africa, Madagascar, Seychelles, Mauritius, Sri Lanka, and the India subcontinent [22]. These two species are not distinguishable in morphology but have distinct karyotypes. D. nasuta has the ancestral karyotype (2n = 8), but the acrocentric 3rd chromosomes are fused to the X and Y to form a pair of new sex chromosomes (X-3/Y-3) in D. albomicans (2n = 6) (Fig. 1). There is almost no pre-mating isolation between these two species [23], and only weak hybrid breakdown was observed in the hybrids of advanced generations [24,25]. SRD is expressed in the F1 males produced by females of certain strains of D. albomicans crossed to D. nasuta males [24–27]. The sex ratio (k, proportion of female) is skewed (k = ∼0.90) if the D. albomicans strains are from Okinawa but normal (k = ∼0.50) if the strains are from Southeast Asia. There is an apparently increasing cline of SRD strength from SE Asia to Japan [24,26].
The “conflict theory” will be strongly supported if most HMS genes have contemporary or historical functions of SRD. The incipient species pair D. albomicans and D. nasuta qualifies as an excellent empirical system for testing the “conflict theory” because both SRD and HMS are expressed in their hybrids. For that, we mapped the genes of HMS and SRD simultaneously through three QTL mappings and multiple lines of introgressions. These genes are polymorphic within D. albomicans. A majority of the genes controlling both traits are colocalized to the same six regions. These findings implicate a contemporarily active SRD system that may have an evolutionary causal link to hybrid male sterility, thus lending strong support to the “conflict theory” of speciation.
For the genetic dissection of SRD and HMS in the species pair D. albomicans and D. nasuta, we first constructed three inbred lines, two from D. albomicans (alb2—Okinawa; shl2—NE India) and one from D. nasuta (nas3—Mauritius) (Materials and Methods). We then surveyed male and female fertilities of these stocks and various F1 genotypes with standard methods, in which single males or females were mated to three virgin testers for 7 days and the progeny size was regarded as the fertility of the tested males or females (see Materials and Methods for details). By the standard methods, all interspecific F1 hybrids appeared to have normal or nearly normal fertility. As expected and consistent with previous studies [24–27], SRD was expressed in the F1 males from alb2♀ × nas3♂ (k = ∼0.9) but not in the F1 males from shl2♀ × nas3♂ and most of the other crosses (k = ∼0.5) (S1 Fig).
Unfortunately, all three inbred stocks are still polymorphic for chromosomal inversions, thus are not ideal for genetic mapping, a major goal of this study. Two true-bred stocks, alb267 and alb215, were then extracted from alb2 with the help of molecular markers (S1 Dataset), so were nas314 and nas384 from nas3. However, we failed to construct inversion-free stocks from shl2, presumably due to recessive sterile mutations locked in the inversions on the two haplotypes (shl2-hap1 and shl2-hap2) (Fig. 1; Materials and Methods). Chromosomes from alb2 or shl2 are not homosequential to those of nas3, thus regions in and around the inversions are not accessible to genetic mapping. However, alb267 and shl2-hap1 are homosequential and have the same standard polytene sequence (S2 Dataset).
The standard test is not powerful enough to detect HMS between these two species. Some subtle abnormalities in spermatogenesis can be revealed by cytological methods. We thus used transmission electron microscopy (TEM) to examine spermatogenesis in the F1 males from the interspecific crosses alb2♀ × nas3♂, nas3♀ × alb2♂ and shl2♀ × nas3♂, as well as that from the intraspecific cross shl2♀ × alb2♂ (Figs. 2, S2). Sperm head development was normal in all the F1 males examined, even those expressing SRD. In contrast, sperm head condensation during spermatogenesis is disrupted in two well studied meiotic drive systems in Drosophila [28,29]. However, pairs of sperm tails were often fused as a characteristic abnormality after the stage of sperm head condensation in many of these males examined. These twin tail fusions were more frequent in the F1 males from alb2♀ × nas3♂ (77% of tails) than those from shl2♀ × nas3♂ (28%), suggesting severer HMS effects contributed by alb2 than shl2 alleles. Unexpectedly, frequent twin fusions (10%) were also seen in the F1 males from the intraspecific cross shl2♀ × alb2♂, tentatively suggesting that HMS has also evolved between these two strains of D. albomicans and possibly also a collateral effect of SRD evolution within the same species. But the intraspecific divergence needs further study with multiple strains of D. albomicans, a species with a very wide geographic distribution. On the other hand, no fusions were found in the hybrid F1 males from nas3♀ × alb2♂ in contrast to the F1 males from the reciprocal cross (alb2♀ × nas3♂), suggesting a lack of HMS loci residing on the X chromosome of nas3 and/or an enrichment of HMS loci on the X-3 chromosome of alb2. Thus, the TEM studies provide evidence that slight HMS has evolved between these two species.
In order to further quantify the weak HMS observed above, we developed a novel, exhaustive mating protocol with the assumption that all of functional gametes can fertilize eggs, so the sperm can be “counted” as progeny size (S3 Fig, Materials and Methods). The results are summarized in Fig. 3. To interpret the data, we posit that there were three antagonistic effects working simultaneously in the tested flies: inbreeding depression, hybrid vigor and outbreeding depression. Inbreeding depression caused much lower fertility of both sexes of the inbred stocks alb2, shl2 and nas3, while hybrid vigor increased the fertility of the F1 males from both reciprocal crosses of shl2 × alb2 (Fig. 3A, C); outbreeding depression, i.e., DMI including HMS, brought down fertility in the F1 males from alb2♀ × nas3♂ and shl2♀ × nas3♂, but not in the F1 males from their respective reciprocal crosses (Fig. 3A). Somewhat consistent with the TEM studies, the fertility of the F1 males from alb2♀ × nas3♂ (mean ± s.e.m = 352 ± 46 offspring per male) was marginally worse than that from shl2♀ × nas3♂ (515 ± 91, 1-tail t-test, P = 0.058). Unlike males, fertility in hybrid females was largely not affected (Fig. 3C). The latter contrast is expected because HMS evolves much faster than HFS [4,30], and there might be only negligible HFS evolution between this pair of species at the very beginning of speciation.
Strong SRD (k = ∼0.92) was expressed in the F1 males but not the F1 females from alb2♀ × nas3♂, consistent with previous interpretation that the observed sex ratio skew is caused by SRD rather than by other mechanisms such as male killing [27]. But unexpectedly, weak SRD (k = ∼0.63) was also detected in the F1 males from shl2♀ × nas3♂ by the exhaustive mating protocol (Fig. 3B). Like HMS, the SRD of this genotype was not detected by standard method (S1C Fig). SRD expression might be affected by sperm storage or competition that must differ between these two mating test protocols.
In sum, both HMS and SRD genes are polymorphic within D. albomicans and their effects are often subtle and difficult to assay. The HMS effects are slight and roughly amount to inbreeding depression suffered in the inbred parental lines. The asymmetry of HMS effects in the F1 males from reciprocal crosses suggests that only a few HMS loci are present [31]. Genotypes with stronger SRD appear to have severer HMS, suggesting a possible connection between these two traits. The HMS and SRD genes might have been enriched on the X-3 chromosome, consistent with a general prediction of the “conflict theory” of speciation. In contrast, D. nasuta might have barely evolved any HMS effects on its X chromosome. The last inference might suggest a lack of SRD activity in D. nasuta since it was split from D. albomicans.
We took a quantitative trait loci (QTL) mapping approach to localize both the HMS and SRD loci divergent among the three chromosomal complements of alb267, shl2-hap1 and nas314 in three separate experiments (Exp1-3) (Materials and Methods). The mapping population of males in Exp1 was produced from crossing the F1 females from alb267 ♀ × shl2 ♂ to nas314 males. Although these males were F2, they actually had interspecific F1-like genetic constitution. The mapping is for genetic variations between these two strains of D. albomicans contributing to SRD and HMS between this species and D. nasuta. In contrast, the mapping populations in Exp2 and Exp3 were generated from backcrossing the F1 females from alb267 ♀ × nas3 ♂ and from shl2 ♀ × nas3 ♂, respectively, to the parental nas314 males. The latter two mapping populations had the backcross 1 (BC1) genetic constitution, and the mappings are for SRD and HMS genes divergent between D. albomicans and D. nasuta. All males of the three mapping populations were mating tested for fertility and sex ratio with standard method (S4 Fig; Materials and Methods).
In QTL analyses, we measured male fertility simply as the raw offspring count (T). In addition, we also transformed T by log10(T+1) or treated it as a binary variable (1 for fertile and 0 for sterile). These three treatments have different biological implications (See Materials and Methods). Consistent with polytene evidence, the third chromosome was almost totally refractory from recombination between alb267 and nas314, as well as between shl2-hap1 and nas314 (S1 Table). Not surprisingly, genetic divergence is low for HMS between alb267 and shl2-hap1 (H2 = ∼20%, Exp1), so is it for SRD between shl2-hap1 and nas314 (H2 = ∼13%, Exp3) (Table 1; S1 and S4 Tables). In contrast, genetic divergence is much higher for HMS between alb267 and nas314, as well as between shl2-hap1 and nas314 (H2 = ∼42–92%, Exp2 and Exp3), and so is it for SRD between alb267 and nas314, as well as between alb267 and shl2-hap1 (H2 = ∼78% in Exp2 and H2 = ∼48% in Exp1, respectively) (Table 1; S2 and S3 Tables). For all mappings, a majority (∼66–100%) of the H2 is additive (h2) (Table 2). This is somewhat unexpected because the interactions between distorters and suppressors would suggest otherwise. The mapping results are incongruent for a few “tentative” QTL where the statistic inferences are not robust (S5 Fig). On the other hand, many QTL are “good” because they are stable with various data transformations and analytical methods (Materials and Methods). All QTL of total offspring (T) and sex ratio with their positions and phenotypic contributions are synopsized in Fig. 4. The nomenclatures of QTL imply their functions: distorter (D), suppressor (S) and hybrid male sterility (HMS).
Exp1 maps intraspecific genetic variations of SRD and HMS genes between the two A. albomicans complements alb267 and shl2-hap1. We found five “good” (D1–D4, S1) and one “tentative” QTL (S2) for sex ratio and two “good” QTL for male fertility (HMS1, HMS2). Except D1 and HMS1 that are colocalized, all the other QTL have only one phenotype. Exp3 maps SRD and HMS between shl2-hap1 and nas314 with two “good” (D8, S5) and three “tentative” QTL (D7, S6, S7) for sex ratio, and two “good” (HMS9, HMS12) and four “tentative” QTL (HMS10, HMS11, HMS13, HMS14) for HMS. Three pairs of loci (D7/HMS9, D8/HMS11, S7/HMS13) are colocalized, but the other five loci have only one phenotype. In these two experiments, the loci with only one phenotype might not genuinely have the alternate phenotype; or, more likely, the other phenotype falls short of detection because of the low H2 and thus low power in QTL mapping. The latter interpretation is supported by introgression studies described in next section.
Since both SRD and HMS have high H2 in Exp2, we expect that the mapping power would be more balanced between these two phenotypes. Indeed, all three “good” (D5, D6 and S4) and one “tentative” (S3) sex ratio QTL from Exp2 are located to regions also harboring HMS QTL. The only exception is the tentative HMS4 without SRD locus nearby (Fig. 4). QTL mapping is known for its lack of resolution, finer mapping is needed as we will show in the next section.
All together, four regions (R1-R4) harbor >90% of additive genetic variance (h2) of SRD and HMS across the three QTL mappings with the only exception of SRD mapping in Exp3, where the “good” S5 and “tentative” S7 outside these four regions contribute 24.4% and 12.2% of h2, respectively (Fig. 4, S4 Table). Because of the low H2 for SRD in Exp3 (13%, Table 1), the robustness of detecting the SRD QTL from Exp3 is questionable; even worse in the case of S5 because of the sparse markers nearby. Nevertheless, the overall colocalization of SRD and HMS suggests that these two traits have evolutionary connection. Because of the limited power and resolution of QTL mapping, more definite evidence can be reached by introgression studies as presented below.
We used a marker-assisted introgression approach to further increase mapping resolution of both SRD and HMS loci by testing the phenotypes of alb267 alleles in the nas314 background (Materials and Methods). Because SRD and HMS are oligogenic systems, the penetrance of the constituent elements depends on appropriate genetic context. Therefore the phenotypes of individual QTL can be best assayed by contrasting two introgression genotypes with and without the focal alb267 alleles (S6 Fig). In addition to the regions of R1 (D5/HMS3 = D1), R3 (S3/HMS6 = S1) and R4 (S4), individual loci of D2, D3 and D4 in the R2 region were also assayed after rare recombinants had been obtained on the 3rd chromosome between alb267 and nas314 (Table 3). All the introgressed alleles are either hemizygous (D1 and D2) or heterozygous (all the other loci). To avoid unnecessary complexities, we ignore the background nas314 allele in genotype nomenclatures throughout. When the introgressed alleles are made homozygous in some genotypes, both copies are included in the genotype nomenclature (e.g., S7 Fig).
We assayed the functions of D1 by contrasting several genotype pairs (Table 4; S8 Fig). In some genetic backgrounds, D1 had strong sterilizing effects while in others it had the dual functions of SRD and HMS. Similar contrasts were made for D2, D3, D4, S3 and S4, as well as S3 and S4 together (S3S4). Each of these loci expressed both HMS and SRD from at least one contrast. The phenotypes of all these loci are obviously sensitive to genetic background. One illustrative example is the R3 region. The shl2-hap1 allele at the R3 region (S1) is a stronger SRD suppressor in Exp1 (Fig. 4), but its SRD suppressing effect was not detected in Exp3 (HMS12), apparently caused by the lack of strong distorter in the hybrids. Similarly, the HMS functions of D2, D3 and D4 were not detected in Exp1 but they were readily detected in some introgressions when alb267 alleles were put into a largely nas314 background. The varying penetrance might have reduced the power of QTL mapping as we noticed earlier. Colocalization of D1 (D5) and HMS1 (HMS3) was more evident by introgressions than QTL mappings (Table 4; cf. Fig. 4), so were the dual functions of the S3/HMS6, S4/HMS7 and HMS12 regions with additional introgressions. When we collected additional mapping data from introgression of shl2-hap1 into nas314 background, even the HMS12 locus was readily detected to express SRD suppressing effect (S9 Fig)
The dual functions might be contributed by separate SRD and HMS genes, but the probability of the four HMS genes with negative effects are each colocalized with one distorter is only 4!∏i=14Ui = 0.0002, where Ui is the 95% confidence interval for the ith QTL, as measured in a fraction of the X-3 chromosome (Materials and Methods). Similarly, the probability is 0.02 for the colocalizations of HMS and SRD genes in the S3 and S4 regions. The overall probability is only 4 × 10–6 if distinct genes control the dual functions in all six QTL intervals. This calculation, albeit rough, strongly argues that the dual functions are most likely to be pleiotropic effects of the same genes.
Notably, some escapers of sterile males sired nearly all female offspring (e.g., 13♀: 1♂ or k = 93%, Table 3), leading us to speculate that many sterile hybrid males may have been potentiated to express extreme SRD. We also assayed the dominance of S3 by crossing D2D3D4;S3 males and females to generate four types of offspring (S7 Fig). When S3 had two copies in the background, the sex ratio was reduced to 0.601 of D2D3D4;S3/S3 from 0.695 of D2D3D4;S3. When S3 was absent, D2D3D4 males can only sire an average of ∼3 males with sex ratio 0.864. Thus S3 is a semidominant SRD suppressor (cf. ref. 27). This explains why this currently silenced SRD system can be reactivated if one complement of suppressors is absent, as clearly shown by SRD (k = 0.628) expressed in one BC1 genotype that differs from alb267 males by only one 2nd chromosome (S10 Fig). It can also be inferred that the Y-3 chromosome of D. albomicans still hosts sensitive responder.
We have uncovered a cryptic SRD system within D. albomicans that appears to have a direct causal link to HMS and thus to speciation. This conclusion is based on the increasing genetic association of SRD and HMS with increasing mapping resolution. We reached the highest resolution in this study with a large collection of introgressions, yet we were unable to separate the SRD and HMS functions to distinct genes at all six major loci. It is much more likely that the dual functions of these loci are pleiotropic effects of the same genes.
Importantly, the phenotypes of distorters (D1—D4) and suppressors (S3, S4) follow the “conflict theory”: all distorters are X-linked and reduce male fertility while all suppressors are autosomal and increase male fertility. This is not the case between an “older” species pair D. mauritiana and D. simulans where the introgressed heterospecific alleles always augment male sterility regardless of their locations [7,19,32]. We interpret the male fertility functions of D1—D4 vis-à-vis S3 and S4 as the former have primarily evolved as SRD distorters while the latter as suppressors. In chronological order, distorters might most likely evolve earlier than suppressors.
Under the above evolutionary scenario, the intraspecific variation of SRD and HMS genes between alb267 and shl2-hap1 can be interpreted in the following way: because Exp1 clearly shows that alb267 has stronger distorter (positive effects of the D1-4 loci) but weaker suppressor (positive effect of the S1-2 loci) (Exp1 of Fig. 4; S2 Table), the SRD system might have become cryptic in alb2 more recently than in shl2, while in the latter the distorter function has been degraded to become residual but the evolution of SRD left permanent footprint on spermatogenesis, so the HMS function such as that of the loci HMS9 and HMS12 stays. In light of this interpretation, the positive effect of HMS1 in Exp1 no longer appear contradictory to the negative effects of HMS3 in Exp2 and HMS9 in Exp3, because the shl2-hap1 allele of D1 expresses weaker SRD but stronger HMS than the alb267 allele. We must point out, however, comparing the magnitudes of HMS effects across HMS1, HMS3 and HMS9 might not be justified because they were measured in very different genetic backgrounds. Furthermore, the shl2-hap1 allele at the R3 region has a stronger suppressing power on SRD than the alb267 allele (Exp1 in Fig. 4). These two different SRD suppressor alleles might also differ in their power to rescue male fertility. Taken together, the difference in HMS effects between alb267 and shl2 as observed earlier (Figs. 2 and 3) can be readily accounted by the SRD system divergence within D. albomicans.
Numerous so-called “speciation genes” have been identified by mapping and positional cloning in the last three decades but little evidence has been gathered for their roles in establishing the initial RI. Though studies suggest several DMI genes as relics of genomic conflicts [3], or indirectly implicate SRD as the primary evolutionary cause of DMI in fly [17,18] and mouse [33,34], this study is the first one to catch SRD in action and the first account that SRD is driving the evolution of most HMS-causing genetic divergence between two newly formed species. We have also shown the difficulty of simultaneously detecting both SRD and HMS phenotypes, because SRD expression is very sensitive to genetic background (Table 4). This difficulty might explain why so little empirical evidence has been accumulated so far for the “conflict theory” of speciation.
One QTL might contain multiple loci with less effect [32]; the dual functions might be caused by closely linked QTL each with one function only. Even if these possibilities turn out to be true under finer genetic analyses, a weaker version of the “conflict theory” could still be valid because HMS genes can hitchhike with the fixation of SRD genes. This possibility is best demonstrated in a classic case of RI between Mimulus guttatus ecotypes previously thought to be a pleiotropic by-product of adaptive evolution to copper contamination in soil. However, HI and copper tolerance are each controlled by tightly linked but distinct genes [35]. Unlike a scenario that RI is driven by ecological adaptation [36,37], the primary driving force emphasized by the “conflict theory” is intragenomic conflicts.
In a broader sense, our study sheds new light on the relationship between adaptive evolution—conventionally attributed to external biotic or abiotic factors—and speciation, which is generally regarded as a consequence of anagenesis under adaptive evolution [38]. Our study emphasizes that non-adaptive evolution out of intragenomic conflicts might be an important mechanism for evolution [39]. In addition to speciation, evolution of several biological traits might also be driven by intragenomic conflicts, such as mating behavior in some insects and epigenetic regulation of the sex chromosomes [13,40–42]. It would be extremely interesting to see how ubiquitous this mechanism would be in the evolution of many other biological traits.
The simplicity of the genetic architecture of SRD/HMS between D. albomicans and D. nasuta opens the door for future studies to fine map and positionally clone all key genes, and to study their population genetics and genomics as well as biogeography of the speciation process [43,44]. An elucidation of the mystery shrouding the speciation problem appears to be reachable at least for these two species.
Lastly, our study might help to address one long-standing controversy over the role of chromosomal rearrangement in speciation. Chromosomal rearrangements like Robertsonian fusions are often found among closely related species, thus are believed by some evolutionists to have played a major role in RI evolution because the F1 heterozygotes of two karyotypes are often less viable or fertile then the parents [45,46]. The difficulty of this theory is that the less fit heterozygotes would have prevented the new karyotype from spreading in a population, let alone founding new species [47]. D. albomicans with the fused X-3 chromosome has evolved from D. nasuta-like ancestor with separate X and 3rd chromosome. Our study has shown that meiotic drive might indeed have helped the spread of the X-3 fusion and meiotic drive can play an important role in karyotype evolution. However, we have also shown that the current RI between D. albomicans and D. nasuta might actually be caused by genic factors, not necessarily by the chromosomal rearrangements per se. Therefore, we revise the original thesis on the role of chromosomal rearrangement in speciation to emphasize meiotic drive as the means to spread new karyotypes—as M. J. D. White speculated [46]—but karyotypic changes might not be directly causing RI.
Three inbred lines were constructed by sib pair matings for 15 generations from outbred stocks: D. albomicans alb2—from the strain E-10802/MYH01-05, Miyakojima, Okinawa, Japan, 2001; D. albomicans shl2—from the strain E-10815/SHL48, Shillong, India, 1981; and D. nasuta nas3—from the strain G86, Mauritius, 1979. Dr. M. Watada, Ehime University, Japan, kindly provided us these three stocks. For brevity, these three inbred stocks are named alb2, shl2 and nas3. These stocks were crossed to generate the F1, F2 and BC1 hybrids, which were tested for fertility and sex ratio and the result is consistent with previous work [24–27] (S1 Fig).
Based on polytene chromosomes and molecular markers (frequent double peaks in the Sanger sequencing chromatograms), we found all three stocks were still polymorphic for inversions. Multiple single pair matings were set up from alb2, nas3 and shl2. Inversion-free parents were identified based on sequencing select markers. We constructed the stocks alb267 and alb215, free of inversions, from alb2. We produced a standard, more accessible and better quality photograph polytene map from alb267, as compared to the same map published before [48] (S2 Dataset). Similarly, we constructed true-bred stocks nas314 and nas384 from nas3, with different polytene sequences on the third chromosome. We failed to construct true-bred stocks from shl2, presumably due to the recessive sterile mutations located in the inversions. All inversions in the three inbred stocks were identified based on polytenes prepared from various stocks and their hybrids, as summarized in Fig. 1.
Flies were reared on standard Cornmeal-Molasses-Agar food in plastic vials (ϕ2.6 × h9.4 cm). For all crosses, virgin tester females were aged to 5 days before setting up crosses at room temperature (22 ± 1°C).
A pair of salivary glands were dissected out from a wandering third-instar larva—sex determined if necessary by the translucent gonads—in a drop of 45% acetic acid and quickly transferred to a second drop of 45% acetic acid for approximately 3 minutes. Individual glands were transferred to a drop of 2% lactic-acetic-orcein solution and stained for ∼5 minutes, then transferred again to a fresh drop of 2% lactic-acetic-orcein solution on a clean slide. The preparation was covered with a siliconized cover slip. The chromosomes were spread by gentle but firm tapping or pressing. The cover slips were sealed with nail polish. The preparations were stored up to 10 days at room temperature prior to examination under a 100× objective of an Olympus BX51 microscope. All cytological images were documented with an Olympus DP30BW digital camera. Further processing was done with Photoshop CS4 ver11.0.2.
PCR primers of molecular markers were designed based on (1) cDNA sequences prepared from D. albomicans male [21] and (2) their alignments with the annotated homologs from D. pseudoobscura, D. virilis and D. mojavensis (http://flybase.org/). The predicted PCR products fall in the size range of 500–1000 bp and span intron(s) if possible. PCR products amplified from alb2, nas3 and shl2 were Sanger sequenced by Beckman Coulter Genomics (Danvers, MA).
Fixed nucleotide differences among stocks were used to develop allele-specific oligonucleotide (ASO) probes [6]. We have developed a total of 62 ASO markers between alb2 and shl2-hap1, 67 markers between alb267 and nas314, and 54 markers between shl2-hap1 and nas314. Many of these markers were also typed by restriction fragment length polymorphism (RFLP). Technical details of the probes, including PCR primers, ASO probes and wash temperatures, can be found in S1 Dataset.
To prepare DNA from single flies, an individual was quickly ground in a 1.5-ml Eppendorf tube with 200 μl extraction buffer (10 mM Tris pH 8.2, 1 mM EDTA, 25 mM NaCl, 0.4 mg/ml Proteinase K). After a 20-min digestion at 65°C, the tube was incubated at 95°C for 5 min and then chilled on ice. The extracted DNA was spun down briefly before being stored at -20°C.
PCR amplification was performed in a total volume of 10 μl reaction mixture (1× buffer, 0.2 μM forward and reverse primer mix, 0.25 units of Taq polymerase, 150 μM dNTP, and 1 μl DNA template). The amplified PCR products were genotyped by RFLP or ASO probes as previously described [6].
Testes and accessory glands were dissected out from young males (2–3 day 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 hrs at 4°C in 1% paraformaldehyde and 2% glutaraldehyde in 0.067 M phosphate buffer, followed by a post fixation of 1 hr in 2% OsO4 at 4°C. The specimens were treated with 1% uranyl acetate at room temperature and were then dehydrated through ethanol grades (30% to 100%). Only one of each pair of testes was embedded.
Each testis was cut into 4–5 segments with a fine tungsten needle and these segments were then aligned on the bottom of a mold with the apical tip facing out to one end. Sections were cut on a Reichert ultracut-S microtome, followed by staining with uranyl acetate and lead citrate. The grids were observed with HITACHI H-7500 electron microscope at Emory University Apkarian IE Microscopy Core.
Two methods were used to measure fertility and sex ratio:
Standard method. Individual males (females) were crossed to 3 virgin females (males) in a vial for 7 days before the mating parents were discarded or kept for genotyping if necessary. The offspring were sexed and counted 4–5 times until the 19th day after setup. Preliminary tests have shown that F1 hybrids between D. albomicans and D. nasuta produced normal or nearly normal numbers of progeny by this method (S1 Fig), as also suggested by previous work [24,25]. The carrying capacity of the food vial is ∼200 flies so the standard method might not be sensitive enough to measure slightly or even moderately reduced fertility.
Exhaustive mating protocol. We designated a method more sensitive than the standard one to quantify fertility. Throughout the experiments we used 5-day old virgin males or females of the same genotype (alb2) as the tester and controlled the temperature at 22 ± 1°C, because preliminary tests had shown that temperature and tester females had small but significant effects on male fertility of some genotypes (S1 Fig).
For male fertility assay, individual 1-day old males were mated to three tester females for 24 hrs (day 1). The males were subsequently transferred to fresh vials supplied with three virgin females on day 2 and day 3, after which the males were transferred to vials with 12 virgin females to stay in days 4–7 and then to vials with three virgin females for day 8. The 4 + 1 days transfer regime was repeated until the individual males were dead or sterile. To prevent crowding in vials the mated tester females in 1-day vials (days 1, 2, 3, 8, 13, etc.) were transferred to fresh vials every 7 days until they no longer laid fertilized eggs. To reduce the labor cost (by ∼80%) we only sexed and counted offspring from 1-day vials (days 1, 2, 3, 8, 13, etc.) to the 19th day after vial setup, while the offspring from the 4-day vials (days 4–7, 9–12, etc.) were not counted and their numbers were interpolated from the flanking two 1-day vials, assuming the 4-day vials produced twice as many offspring from these two 1-day vials. Towards the end of the protocol, male fertility dropped to only a few offspring per day so all offspring were usually counted from both types of vials.
The above protocol for quantifying male fertility was designed under the assumptions that (1) all functional sperm fertilize eggs, and (2) the interpolation was accurate. In a pilot study we found male mating latency was more than 12 hrs, and the progeny size from the second mating within the same day was much smaller than the first mating. Therefore the first assumption is likely to be valid. The second assumption was shown to be valid also by two pilot experiments in which the alb2 and nas3 males were tested by the above protocol, with additional transfers of 4-day vial females to fresh vials and counting of their offspring. The actual counting and interpolation converges remarkably well (S3 Fig). Therefore the exhaustive mating protocol can be used to “count” the functional sperm produced by a male.
For female fecundity assay, single females were mated to three tester males in a vial and the flies were transferred to a fresh vial every 4 days until the female became sterile or died. Any dead male was replaced with fresh one during the experiment. All offspring were sexed and counted.
Because of the fixed inversions between alb267 (shl2-hap1) and nas314 on the 3rd chromosome, ∼40% of the genome is refractory from meiotic mapping (Fig. 1). On the other hand, the two D. albomicans complements, alb267 and shl2-hap1, are homosequential so that meiotic mapping can cover the whole genome. With these considerations and to maximize the power of QTL mapping from the available lines, three QTL mappings were executed.
In the first QTL experiment (Exp1), we generated a population of 459 males by crossing individual F1 females (alb267/shl2-hap1 from alb267♀ × shl2♂) to nas314 males. After the vials were established, the mated F1 females of the genotype alb267/shl2-hap1 were distinguished from that of the genotype alb267/shl2-hap2 by molecular markers. Each male of the mapping population was phenotyped by crossing to alb2 females per standard method. These 459 males were genotyped for 62 ASO markers that can distinguish the alb267 and the shl2-hap1 alleles (S1 Dataset).
The other two QTL mappings were similarly executed. In Exp2, a population of 442 males was generated by backcrossing the F1 females from alb267♀ × nas314♂ to nas314 males, and was genotyped for 67 ASO markers. In Exp3, a population of 470 males was generated by crossing the F1 females (shl2-hap1/nas314 from shl2♀ × nas314♂) to nas314 males, and was genotyped for 39 ASO markers out of the 54 markers available because only three out of the 18 markers (M8, M24 and M30) on the non-recombining 3rd chromosome were genotyped (S1 Table).
The phenotypes (male fertility and sex ratio) of all three QTL mapping populations are summarized in S4 Fig. Because it is not reliable to calculate sex ratio from small progeny size, we only use the males that sired at least 30 offspring. Thus the sample sizes of sex ratio for these three QTL mappings are reduced to 440, 227 and 340, respectively. Overall, males from Exp2 and Exp3 suffered much greater sterility than from Exp1, while SRD is almost absent from Exp3. This pattern is consistent with earlier observations that the F1 males from alb2♀ × nas3♂ and shl2♀ × nas3♂ were very infertile, while shl2 had only weak SRD alleles genes (Fig. 3).
We first applied the R/qtl package (v1.26) to construct three genetic maps separately from the three QTL mappings [49]. As expected, the 3rd chromosome had normal recombination only in Exp1 but hardly any in the other two mappings (S5 Fig, S1 Table). M68 is not linked to the 2nd linkage group in Exp2, while M41 and M45 are not informative in Exp3. In the end, the genetic maps of Exp2 and Exp3 are far less complete as compared to that of Exp1. Interestingly, there seems to be a cluster of markers around the centromere region on the 2nd chromosome, suggesting the existence of chromosomal rearrangements in that area but we did not detect any polytene evidence for that suggestion.
To map QTL for HMS and SRD, we applied the composite interval mapping (CIM) method implemented in Windows QTL Cartographer (v2.5_008) [50,51]. Male fertility was treated as a continuous variable as either the raw counts (T) or transformed by log10(T+1), or as a binary variable of 1 (fertile) and 0 (sterile), with different biological implications. For example, the difference between sterile (T = 0) and subfertile (say T = 10) is definitely more profound in terms of spermatogenetic defects than difference in fertility, say, of T = 100 and 110; thus the log10 or binary transformation might be closer to biological reality than the raw count T. Sex ratio (k) was also treated as continuous variable. The threshold for significant QTL was determined by 500 times of permutations of the datasets at the level α = 0.05.
The QTL mapping results are plotted in S5 Fig. The presence, location and magnitude of the HMS QTL are often sensitive to data transformation methods. We also applied the multiple interval mapping (MIM) method to evaluate the QTL flagged by CIM and the epistasis, if any, among them [52]. The total genetic components (H2) and their additive parts (h2) of all QTL were also obtained from MIM, as summarized in S2–S4 Tables. A synopsis of QTL mappings by different methods is presented in Fig. 4 and Tables 1 and 2.
For a much improved signal/noise ratio than that from QTL mapping, we thus wished to test the effects of the flagged individual QTL in a uniform and clean background. We used an introgression method to isolate a few chromosomal segments, each containing individual QTL, in a largely nas314 background including the Y chromosome. A typical scheme was shown in S6 Fig. The six QTL, the markers used to monitor their transmission and the approximate sizes of each interval (proportion of the X-3 or 2nd chromosomes based on genetic distance) are: D1 (M105 – M57, ∼3.7%), D2 (M107 – M20, ∼3.8%), D3 (M98 – M157, ∼11.1%), and D4 (M30, ∼5.4%) on the X-3, S3 (M72 – M136, ∼4.0%) and S4 (M66 – M63, ∼25.5%) on the 2nd chromosome. The estimated interval sizes might have large errors because of unequal cross-over frequencies along the chromosomes.
Sex ratio was treated as continuous variable with Gaussian distribution if all the progeny sizes were at least 30; otherwise logistic regression was applied on male and female counts. For summary statistics (mean and s.e.m.) of sex ratio obtained from sub-fertile males that often had progeny < 30, a bootstrapping method was used to avoid spurious results. Other methods were standard as indicated in the text.
|
10.1371/journal.pgen.1002530 | A Quantitative, High-Throughput Reverse Genetic Screen Reveals Novel Connections between Pre–mRNA Splicing and 5′ and 3′ End Transcript Determinants | Here we present the development and implementation of a genome-wide reverse genetic screen in the budding yeast, Saccharomyces cerevisiae, that couples high-throughput strain growth, robotic RNA isolation and cDNA synthesis, and quantitative PCR to allow for a robust determination of the level of nearly any cellular RNA in the background of 5,500 different mutants. As an initial test of this approach, we sought to identify the full complement of factors that impact pre–mRNA splicing. Increasing lines of evidence suggest a relationship between pre–mRNA splicing and other cellular pathways including chromatin remodeling, transcription, and 3′ end processing, yet in many cases the specific proteins responsible for functionally connecting these pathways remain unclear. Moreover, it is unclear whether all pathways that are coupled to splicing have been identified. As expected, our approach sensitively detects pre–mRNA accumulation in the vast majority of strains containing mutations in known splicing factors. Remarkably, however, several additional candidates were found to cause increases in pre–mRNA levels similar to that seen for canonical splicing mutants, none of which had previously been implicated in the splicing pathway. Instead, several of these factors have been previously implicated to play roles in chromatin remodeling, 3′ end processing, and other novel categories. Further analysis of these factors using splicing-sensitive microarrays confirms that deletion of Bdf1, a factor that links transcription initiation and chromatin remodeling, leads to a global splicing defect, providing evidence for a novel connection between pre–mRNA splicing and this component of the SWR1 complex. By contrast, mutations in 3′ end processing factors such as Cft2 and Yth1 also result in pre–mRNA splicing defects, although only for a subset of transcripts, suggesting that spliceosome assembly in S. cerevisiae may more closely resemble mammalian models of exon-definition. More broadly, our work demonstrates the capacity of this approach to identify novel regulators of various cellular RNAs.
| The coding portions of most eukaryotic genes are interrupted by non-coding regions termed introns that must be excised prior to their translation. The excision of introns from precursor messenger RNA (pre–mRNA), is catalyzed by the spliceosome, a large macromolecule composed of both RNA and protein components. Several studies have uncovered connections between pre–mRNA splicing and other RNA processing pathways such as the remodeling of chromatin structure, transcription, and processing events that take place at the 3′ end of the transcript. To date, however, the full complement of factors that function to couple splicing to other processes in the cell remains unknown. Here, we have developed a novel screening methodology in the budding yeast, Saccharomyces cerevisiae, that allowed us to individually examine nearly all of the ∼6,000 genes to determine which factors functionally impact splicing. We identified mutations in components that function at either the 5′ or 3′ end of a gene. Most of these components have previously established roles in other aspects of gene expression, including chromatin remodeling and cleavage and polyadenylation processes, and their identification here provides the first evidence for their roles in coupling these pathways.
| The coding portions of most eukaryotic genes are interrupted by non-coding introns which must be removed prior to the translation of their messenger RNAs (mRNA). Removal of introns from pre–mRNAs is catalyzed by the spliceosome, a large and dynamic ribonucleoprotein complex comprised of five small nuclear RNAs (snRNAs) and at least 100 proteins [1]. Much of our knowledge about the components that comprise the spliceosome as well as their mechanisms of action has been derived from experiments using the powerful genetic tools available in the budding yeast, Saccharomyces cerevisiae. Indeed, although the RNA2 – RNA11 genes originally identified in Hartwell's forward genetic screen preceded the discovery of splicing [2], the mechanistic characterizations of these genes, since renamed PRP2 – PRP11, underlie current models of the splicing pathway. Importantly, because the core components of the spliceosome are highly conserved between budding yeast and humans, the mechanistic details derived from work in yeast have been instrumental in understanding mechanisms of pre–mRNA splicing in higher eukaryotes.
The modern view of pre–mRNA splicing acknowledges the integrated role of the spliceosome with several other aspects of RNA processing. Whereas the historical view of splicing envisioned a cascade of temporal events initiated by transcription, followed by polyadenylation, and finalized with splicing and export of mRNAs from the nucleus, it is now clear that these pathways are not independent from one another but rather are functionally coupled. Strong evidence in both yeast and higher eukaryotes demonstrates that recruitment of the spliceosome to intron-containing transcripts occurs co-transcriptionally [3]–[6], mediated at least in part by physical associations between the C-terminal domain (CTD) of RNA polymerase II and the U1 snRNP [7]. A growing body of evidence also indicates that the landscape of chromatin modifications encountered by transcribing polymerase molecules can dictate the activity of the spliceosome at various splice sites. For example, recent work has identified an enrichment of methylated lysine-36 in the histone H3 protein specifically within exonic sequences, suggesting a possible mechanism for facilitating the identification of intron-exon boundaries [8], [9]. Similarly, the rate of transcription by RNA polymerase II, which can be impacted by chromatin marks, has also been shown to be critical for dictating alternative splicing decisions [10]. Furthermore, it is also clear that splicing is coupled to downstream steps in RNA processing. For example, the yeast Ysh1 protein [11], [12], which is the homolog of CPSF73, the mammalian endonuclease required for 3′ end processing, was originally identified as Brr5 in a cold-sensitive screen for mutants defective in pre–mRNA splicing [13]. Consistent with this observation, recent evidence suggests that transcriptional pausing near the 3′ end of genes is a critical component of pre–mRNA splicing efficiency [14]. Despite the increasing evidence of the interconnectivity of these pathways, in many cases the mechanistic details which underlie these functional relationships remain unclear. Our understanding of these mechanistic connections would benefit from a more complete understanding of the complement of factors through which splicing is connected to these cellular processes.
A variety of recent genome-wide approaches have provided important insights into the connections that exist between the spliceosome and other cellular processes. Two powerful approaches, Synthetic Genetic Array (SGA) Analysis [15] and Epistatic MiniArray Profiling (E-MAP) [16], leverage genetic tools available in yeast to systematically generate millions of double-mutant strains and then carefully quantitate their cellular fitness to determine an interaction score for every pair-wise mutation. On the basis of strong positive or negative genetic interaction scores these approaches have been successfully used to infer functional relationships between many cellular pathways, including several with pre–mRNA processing [17], [18]. Simultaneously, improvements in proteomic methodologies have enabled the direct analysis of protein complexes in organisms as diverse as humans and yeast, allowing for an assessment of all of the stably-bound proteins involved in pre–mRNA splicing in many organisms [19], [20]. While the combination of these and other approaches has provided a global picture of many of the cellular factors that influence the splicing pathway, either directly or indirectly, an important question remains about the functional significance of these factors in the splicing of specific transcripts. Indeed, it has long been known that certain transcripts require the activity of unique accessory factors to facilitate their splicing [21]. Moreover, recent work supports the idea that different transcripts can have a greater or lesser dependence upon the activity of core spliceosomal components for their efficient splicing [22], [23].
Here we present the results of a novel approach that complements the genetic and physical approaches of others by allowing for a direct functional assessment of nearly every gene in the S. cerevisiae genome in the pre–mRNA splicing process. For this work, we developed automated methods that enabled the isolation of total cellular RNA from about 5500 unique strains, each of which contained a mutation in a single gene, and all of which were examined during exponential growth in liquid medium. Using a high-throughput quantitative PCR (QPCR) assay, the relative cellular level of nearly any RNA can be readily determined in the background of each of these strains. By assessing the levels of several different pre–mRNA species, we were able to identify not only those factors which are necessary for the splicing of many transcripts, but also factors that are specifically required for the splicing of a subset of intron-containing genes. Whereas our study specifically examines the levels of several cellular pre–mRNAs, the approach described herein can be easily adapted to study the level of nearly any RNA molecule of interest under a wide variety of cellular growth conditions.
To identify the comprehensive network of cellular factors that lead to a change in splicing efficiency, we developed a high-throughput reverse genetic screen that allowed us to readily assess changes in pre–mRNA levels in the background of ∼5500 Saccharomyces cerevisiae strains, each of which contained a mutation in a single gene. The library of strains contained deletions of non-essential genes [24] as well as conditional mutations in essential genes [25], accounting for mutational access to over 93% of known yeast genes. Using a liquid-handling robot, protocols were developed (see Materials and Methods) that allowed for the simultaneous collection of each of these strains under exponential growth conditions in liquid medium in 384-well plates. Total cellular RNA was isolated robotically from each of these strains using a phenol extraction protocol [23] followed by a glass-fiber purification step [26]. After converting this RNA into cDNA using a random-priming strategy, QPCR was used to directly measure the level of a given RNA species within each strain. Because of the inherent variability between the samples in the cell collection, RNA isolation, and cDNA synthesis steps, the levels of six different RNA species were measured in each of the samples in order to calculate a normalization constant. On the basis of this normalization constant, the relative level of virtually any cellular RNA species can be determined in each of the mutant strains.
As an initial test of our approach we sought to identify the full complement of factors involved in pre–mRNA splicing by determining the relative levels of unspliced U3 small nucleolar RNA (snoRNA) present in each of the mutant strains. The U3 snoRNA is unique in the S. cerevisiae genome in that it is the only known non-coding RNA that is interrupted by a spliceosomal intron [27]. Nevertheless, the U3 transcript has been widely used historically as a splicing reporter, owing to its relatively high basal expression level and the strong accumulation of U3 precursor levels observed in the background of canonical splicing mutants [13], [28], [29]. As shown in Figure 1, the U3 precursor levels are unaffected in the vast majority of the strains examined, with levels varying by less than 1.35-fold from one another for 95% of the strains. Indeed, only ∼200 of the ∼5100 strains that passed our quality filters (see Materials and Methods) showed a change in the relative U3 precursor levels of more than ∼30% from the median value (∼0.35 in log2-transformed space), consistent with our expectation that mutations in most genes will have little or no effect on cellular pre–mRNA splicing efficiency. The tight distribution of relative U3 precursor levels seen within this dataset demonstrates the high precision with which these measurements can be made, and suggests a low false discovery rate for our approach.
To characterize the data generated by this approach we sought to define the biological significance of those strains that showed increased levels of U3 precursor. As an initial analysis, we examined the U3 precursor levels in those strains containing mutations in known splicing genes. Using the GO PROCESS: RNA Splicing as a guide [30], a total of 71 strains in our library were classified as containing mutations in canonical splicing factors (Figure S1), of which 68 passed our quality filters for the U3 precursor dataset (see Materials and Methods). A strong overrepresentation of these splicing factors can be seen within the set of strains showing an enrichment of U3 precursor (Figure 1A). Of the 68 strains containing splicing mutations that passed our quality filters: 53 are found within the top 200 strains (p = 9.28E-64, Fisher's exact test); 38 are found in the top 50 strains (p = 1.33E-66); and the top 14 strains all belong to this list (p = 1.27E-27). Taken together, these data argue strongly that the candidates identified by this approach will be characterized by a high true positive discovery rate.
By contrast, out of the 68 strains containing mutations in known splicing factors for which we obtained high quality data, 15 failed to show an enrichment of U3 precursor levels in this dataset, suggesting either that mutations in these genes don't cause an increase in U3 precursor levels (true negative), or that our approach incorrectly failed to detect the accumulation of unspliced U3 (false negative). To better resolve these possibilities we chose to more completely examine the global splicing fitness of some of these strains using splicing-sensitive microarrays. For every intron-containing gene in the genome, these custom-designed microarrays contain at least three probes (Figure 2A) that allow us to distinguish between spliced and unspliced isoforms [31]. We used these microarrays to assess the global splicing defects of four mutants: two canonical splicing mutants that showed strong U3 precursor accumulation (snt309Δ and lsm6Δ, Figure 2B), and two that showed little or no accumulation (mud2Δ and cus2Δ, Figure 2C). As expected, and consistent with previous work from others [32], the snt309Δ and lsm6Δ strains demonstrate a broad splicing defect, with most intron-containing genes displaying an increase in precursor levels accompanied by a decrease in the amount of spliced mRNA. By contrast, the global splicing profiles of the mud2Δ and cus2Δ strains are markedly different. In the cus2Δ background, few intron-containing genes display a splicing defect: very little precursor accumulation is observed, and there is little if any detectable loss in mature mRNA. The mud2Δ mutation does cause a splicing defect for some intron-containing genes, whereas little change in splicing efficiency is seen for many others. Notably, as seen in Figure 2D, the microarrays of both the snt309Δ and lsm6Δ strains show a strong accumulation of U3 precursor levels, whereas the mud2Δ and cus2Δ strains show almost no accumulation, consistent with our QPCR screen results. It is worth noting that in our experience the behavior of the U3 transcript differs from the other intron-containing genes in that every splicing mutation we have examined that causes an increase in the U3 precursor levels also results in an increase in the total level of U3; the reason for this apparent discrepancy is currently under investigation. Nevertheless, these microarray data demonstrate that our failure to detect an increase in U3 precursor levels in the mud2Δ and cus2Δ strains does not represent a failure of the approach, but rather that these are true negative results.
To better assess the total complement of genes that can impact the splicing of any precursor transcript, we chose to expand our analysis by measuring the precursor levels of several additional intron-containing genes. We chose to examine four ‘canonical’ intron-containing genes (RPL31B, UBC13, TUB3 and TEF5) that vary in terms of intron size, transcriptional frequency, biological function, and the presence or absence of an intron-encoded snoRNA. In spite of these differences, these transcripts are similar to one another in so much as they each contain splice site and branch point sequences that conform to consensus sequences. In addition to these four genes, we chose to examine two intron-containing genes (YRA1 and REC107/MER2) that are known to be poorly spliced under standard growth conditions [21], [33], [34]; as such, we expected the behavior of these two transcripts to be distinct from the efficiently spliced transcripts. For all six of these genes, the precursor levels were measured in all ∼5500 strains. As an initial analysis of this data set, we considered the behavior of the 71 strains containing mutations in spliceosomal components (Figure 3). As expected, precursor accumulation can be detected for each of the canonical intron-containing transcripts in the background of nearly all of the splicing mutations. While all four canonical precursors accumulate in the mud2Δ background, consistent with our microarray data, no precursor accumulation is detected for any of these transcripts in the cus2Δ strain (Figure 3B). In addition, several of the splicing mutants that failed to cause an increase in the U3 precursor levels do cause a splicing defect for these other transcripts. Importantly, the behavior of the Rec107 and Yra1 pre–mRNAs within this subset of strains differs significantly from that seen for the canonically spliced transcripts. Splicing of the Rec107 pre–mRNA shows a strong accumulation in the upf1Δ and upf2Δ strains (Figure 3C), consistent with its known degradation via the nonsense-mediated decay pathway [35]. Because the Rec107 pre–mRNA does not engage the spliceosome during vegetative growth [21], no precursor accumulation is expected in strains containing spliceosomal mutations. Likewise, the Yra1 pre–mRNA shows a strong accumulation in the edc3Δ strain [34], consistent with its previously characterized cytoplasmic degradation pathway. The failure to detect Yra1 pre–mRNA accumulation in strains containing spliceosomal mutations presumably reflects the inherently high levels of unspliced Yra1 transcript present in a wild type cell. Taken together, these data strongly support the capacity of this approach to successfully identify mutations that impact pre–mRNA splicing with low false positive and false negative rates of discovery.
To expand our analysis beyond previously characterized splicing factors, we sought to identify novel mutations that caused an increase in precursor levels in most, if not all, of our canonical intron-containing genes. By determining the rank order of precursor accumulation in each strain for each of the five canonical splicing substrates (U3, Rpl31b, Tef5, Tub3, and Ubc13 precursors), a composite rank order of each strain was calculated as the average of these independent measurements (Figure 4). Remarkably, while the majority of the mutations examined cause little or no change in precursor levels of these four transcripts, the subset of mutations which do cause detectable increases in precursor levels is larger for some of the coding mRNAs than was seen for U3. Interestingly, although there is variation in the number of strains that cause pre–mRNA accumulation of the different transcripts, with Tub3<Tef5<Ubc13∼Rpl31b, strong overlap can nevertheless be identified across the four transcripts. For example, the majority of the strains that cause an increase in the Tub3 pre–mRNA also display an increase in the pre–mRNA levels of the other three transcripts. By contrast, many strains cause a strong accumulation of the Ubc13 and Rpl31b pre–mRNAs without causing a significant change in the Tub3 or Tef5 pre–mRNA levels.
Because the absolute levels of the Rpl31b, Tef5, Tub3 and Ubc13 pre–mRNAs are significantly lower than the U3 precursor levels in most strain backgrounds (Table S1), we considered the possibility that these results reflected a technical artifact associated with measuring the cellular levels of low abundance RNA species in certain strain backgrounds. Importantly, however, the relative levels of the Rec107 pre–mRNA, whose normal cellular level is similar to these other pre–mRNAs, is largely unchanged in the vast majority of the strains examined (Figure 4). Likewise, an analysis of the cellular levels of the Faa1 mRNA, an intronless gene whose transcript abundance is of a similar magnitude as the Rpl31b, Tef5, Tub3 and Ubc13 pre–mRNAs, also shows a nearly constant level in all of the examined strains, further suggesting that there is no inherent bias in detecting low level transcripts. Finally, the Yra1 pre–mRNA, which is inefficiently spliced and has a higher endogenous level than most pre–mRNAs, also shows very little change in the examined strains. Taken together, these results strongly support the conclusion that the levels of the Rpl31b, Tef5, Tub3 and Ubc13 pre–mRNAs are increased in these strains.
Because our approach, as described so far, directly measures the cellular levels of precursor RNA but does not directly determine the efficiency of splicing per se, those mutations which cause an increase in the precursor levels could be doing so simply by increasing the transcriptional frequency of these genes rather than by directly impacting their splicing. To distinguish this possibility from a true splicing defect, we chose to directly calculate the splicing efficiency of the Tef5 transcript by measuring the total cellular level of Tef5 mRNA by QPCR in each strain and using this value to calculate the ratio of unspliced∶spliced RNA in the cell, a classical measure of splicing efficiency. Consistent with a splicing rather than transcriptional cause for precursor accumulation, the measured levels of total Tef5 transcript showed little variation across nearly the entire set of strains (Figure S2). Indeed, nearly every strain that showed an increase in Tef5 pre–mRNA levels also showed a decrease in the splicing efficiency of the Tef5 transcript (Figure 4), suggesting that those mutations affect the splicing of this transcript rather than its transcription. These results strongly suggest that the increased pre–mRNA levels observed in these strains largely reflect changes in pre–mRNA splicing.
To assess the functional significance of the strains displaying increased pre–mRNA levels, we sought to rule out the possibility that mutations which cause a change in overall cellular fitness might indirectly lead to a decrease in overall splicing efficiency. To test this, we compared our precursor accumulation levels with recently described strain fitness data calculated for each of the 5000 non-essential genes [36]. This comparison yielded no correlation between precursor accumulation and cellular fitness (Figure S2), suggesting that cellular growth rate alone is insufficient to explain the observed increase in pre–mRNA levels.
While the precursor accumulation seen for each of the canonical transcripts in the known splicing mutants lends strong empirical support for the overall robustness of our approach, additional analysis was needed to assess the statistical significance of the data we generated. Towards this end, we employed a statistical approach originally developed for analysis of microarray data called Significance Analysis of Microarrays [37], or SAM (see also Materials and Methods). We chose this software because, conceptually, the data generated by our QPCR approach are orthogonal to those from a microarray experiment: whereas a microarray experiment examines the behavior of thousands of mRNAs in a single strain, here we examine the behavior of a single RNA in thousands of different strains. Because similar concerns regarding multiple hypothesis testing apply to both types of data [38], we used this software as a tool for assessing the quality of our data. The results of our SAM analysis were consistent with the qualitative results seen in Figure 4, in so much as the number of strains causing a statistically significant increase in the levels of each precursor species varied depending upon the precursor mRNA in question. A total of 224 strains caused a statistically significant increase in the Rpl31b pre–mRNA levels, 209 strains caused a significant increase in Ubc13 pre–mRNA levels, 146 strains caused a significant increase in U3 precursor levels, 83 strains caused a significant increase in Tef5 pre–mRNA levels, and 78 strains caused a significant increase in Tub3 pre–mRNA levels. The complete list of SAM-identified strains for each RNA species is provided in Table S2. Importantly, many of the SAM-identified strains are found to cause a significant enrichment of the precursor levels of all five of these RNAs, including the majority of strains with mutations in canonical splicing factors.
Interestingly, for some of the species examined, a small number of strains were identified which showed decreased levels of precursor RNA. In certain instances these reflected expected outcomes: a large decrease in the Ubc13 precursor was identified in the ubc13Δ strain, for example. However, in other cases these may indicate important biological phenomena. For example, both the xrn1Δ and the tfg2Δ strains cause a significant decrease in the U3 precursor levels. We have previously shown that deletion of the Xrn1 nuclease paradoxically leads to decreases in many precursor RNAs [39], although the mechanism by which this occurs remains unknown. Likewise, it is unclear whether the decreased precursor resulting from deletion of the TFIIF component Tfg2 reflects an overall decrease in transcription of this gene, or whether this in fact reflects increased splicing efficiency perhaps resulting from a decreased transcription elongation rate [10].
To better characterize the factors that impact pre–mRNA splicing, we examined our lists of SAM-identified candidates for factors that are not canonical components of the spliceosome. As an initial approach, we asked whether any functional categories of proteins were statistically overrepresented within this set of strains. For this analysis, we ordered the strains according to the largest precursor accumulation that they affected for any of the RNA species. We then used the GO::Term Finder program [40] to identify overrepresented classes of genes. As expected, when considering the 50 strains that caused the largest precursor accumulations, a strong enrichment for splicing factors was seen with 30 out of 50 strains containing mutations in genes belonging to the GO PROCESS: RNA Splicing category (p = 1.3E-40 with Bonferroni correction). Interestingly, when the top 100 strains are considered, significant enrichment can also be seen for strains with mutations in factors belonging to the GO PROCESS: Chromatin Remodeling category, with eight different mutants causing precursor accumulation (arp5Δ, arp8Δ, bdf1Δ, npl6Δ, rsc2Δ, rsc9-ts, vps72Δ, and yaf9Δ; p = 1.5E-03). Expanding our analysis to the top 200 candidates increases the enrichment of this category to include twelve factors (adding arp6Δ, swc5Δ, swr1Δ, and taf14Δ; p = 1.8E-04). Interestingly, within the top 200 candidates, significant enrichment is also seen for the GO PROCESS: RNA Catabolic Process category, with 13 different factors being present (ccr4Δ, dis3-ts, dbr1Δ, kem1Δ, lsm2-ts, lsm6Δ, lsm7Δ, prp18Δ, rrp6Δ, rtt101Δ, ski3Δ, ssn2Δ, and upf3Δ; p = 8.5E-03). Whereas some of these factors, such as lsm2-ts, lsm6Δ, lsm7Δ, and prp18Δ are known to directly function in pre–mRNA splicing, the identification of many of these factors presumably reflects their defects in degradation pathways for unspliced pre–mRNAs.
One of the top factors we identified that bridges chromatin remodeling with transcription initiation is the bromodomain factor Bdf1. Bdf1 is a member of the SWR1 complex and, along with its homolog Bdf2, has been shown to interact with the TFIID component of RNA polymerase II [41]. Moreover, BDF1 and BDF2 have been demonstrated to be genetically redundant with one another. Whereas our SAM analysis indicated that the bdf1Δ caused a statistically significant accumulation of most of the canonical precursor species in our experiments, the bdf2Δ strain showed little or no detectable increase in the levels of any of the precursors tested (Figure 5A), and was not considered by SAM analysis to be significant for increases in any of the precursor RNAs. To better characterize the global splicing profile of these two mutants, we again turned to our splicing-sensitive microarrays. Remarkably, a dramatic splicing defect can be seen in the bdf1Δ strain for most intron-containing genes, as evidenced by an increase in the precursor transcript levels with a concomitant decrease in the mature and total transcript levels (Figure 5A). By comparison, the bdf2Δ mutation has almost no effect on cellular splicing, strongly corroborating the specific identification of Bdf1 in our screen. To better assess the mechanism by which Bdf1 impacts pre–mRNA splicing, we monitored U1 snRNP recruitment in the background of wild-type, bdf1Δ, and bdf2Δ strains using chromatin immunoprecipitation coupled to QPCR (ChIP-QPCR). As seen in Figure S3, these experiments show that the deletion of Bdf1 but not of Bdf2 decreases the occupancy of U1snRNP at several intron-containing genes, suggesting impairment of co-transcriptional spliceosomal recruitment in the bdf1Δ strain. A more comprehensive ChIP-Seq experiment will be required to fully characterize the global landscape of genes impacted by the deletion of Bdf1 and further characterize the roles of Bdf1 and Bdf2 in transcription and splicing.
We also chose to further examine several factors that our screen identified that are more classically connected with chromatin remodeling. The lower panels of Figure 5B and 5C show the locations within our U3 precursor dataset of all of the strains containing mutations in components of the SWR1 complex, and the RSC complex, respectively. Notably, mutations in many but not all of the components of these complexes cause a splicing defect of the U3 transcript. Moreover, each of the five precursor species that we examined shows a slightly different susceptibility to the different components of these complexes. We chose to examine the global splicing defects of strains containing mutations in two of these components: Vps72, a member of the SWR1 complex; and Rsc9, a member of the RSC complex. Splicing-sensitive microarrays of the vps72Δ and rsc9-ts strains, respectively, reveal a splicing defect in each strain (Figure 5B and 5C). However, unlike the bdf1Δ strain, the vps72Δ and rsc9-ts strains cause a splicing defect in only distinct subsets of intron-containing genes. Interestingly, the affected subsets of transcripts are neither completely overlapping nor completely independent of one another; rather the microarray data are consistent with our QPCR data in suggesting that mutations in specific chromatin-modifying components can result in aberrant splicing of specific pre–mRNA transcripts.
While an ontology-based approach can successfully identify entire pathways that display enrichment, we were also interested in considering those factors which showed strong pre–mRNA accumulation but whose functional categories were not statistically over-represented at the top of our dataset. Remarkably, while the GO PROCESS: RNA 3′ end Processing wasn't significantly overrepresented as a category within our dataset (9 out of top 200, p = 0.09), several strains with mutations in factors belonging to this category resulted in a strong, statistically-significant accumulation of multiple precursor species. Included among these were: yth1-ts, a zinc-finger containing protein that is the homolog of human CPSF-30; cft2-ts, the homolog of human CPSF-100; and fip1-ts, a component of the polyadenylation factor PF I. To further examine the global splicing defects of each of these mutants, microarrays were performed comparing mutant and wild type behavior after shifting them to both elevated and reduced temperatures. Of the three mutants, the profile seen in the yth1-ts mutation most closely resembles a canonical splicing defect, with more than half of the genes showing an increase of precursor and loss of mature RNA (Figure 6A). Interestingly, the splicing defect is strongest at reduced temperatures even though this strain has only a subtle low-temperature growth defect (not shown). By comparison, neither the cft2-ts nor the fip1-ts strains showed a strong splicing defect at low temperature (not shown), but each mutant was characterized by an unusual phenotype at elevated temperatures. As seen in Figure 6B and 6C, two distinct types of behavior are seen in the cft2-ts and fip1-ts mutants, respectively, that are largely defined by whether or not the affected transcript encodes a ribosomal protein gene (RPG). For a subset of the non-RPG transcripts a canonical splicing defect is apparent, consistent with our QPCR results. Interestingly, the subset of affected non-RPG transcripts is different between the two mutant strains. By comparison, nearly all of the RPG transcripts show a dramatic increase in both the mature and total mRNA levels, with little or no detectable change in precursor levels. The strong increases caused by these mutants suggest that the RPG transcripts may be subject to regulatory control at their 3′ ends. Interestingly, while it has long been known that RPG introns are, in general, longer than non-RPG introns [42], whereas the second exons of RPGs tend to be shorter than non-RPGs [5], we nevertheless find no strong correlation between either intron or second exon length and the strength of the splicing defect seen for these 3′ end mutants (data not shown). The mechanisms by which these 3′ end factors impact pre–mRNA splicing are currently under investigation.
In considering the mechanisms by which candidate factors may be functioning, we sought to determine whether any of the candidates we examined might be indirectly affecting pre–mRNA splicing by changing the cellular levels of known spliceosomal components. Although our splicing-sensitive microarrays were designed primarily to interrogate the splicing status of the ∼300 intron-containing genes in S. cerevisiae, they also contain probes against all ∼6000 protein-coding genes and ∼200 RNA genes, including the spliceosomal snRNAs. Figure S4 shows the relative RNA levels for each of the canonical spliceosomal components, including the snRNAs, in the background of each of the different strains we examined, as determined from our microarray analyses. While these results positively recapitulate the expected changes (for example, the decreases in Snt309 and Lsm6 mRNA levels in the snt309Δ and lsm6Δ strains, respectively), with only a few exceptions, most spliceosomal components appear unchanged in most of the mutants we examined. Importantly, the transcript encoding the Mud1 protein showed dramatic mis-regulation in both the yth1-ts and cft2-ts strains, increasing by more than 10-fold in each background. To test whether Mud1 overexpression might be causing the splicing defects observed in these strains, a strain was constructed where the wild type Mud1 transcript was encoded on a high-copy plasmid. As seen in Figure S5, in spite of the over 30-fold increase in Mud1 levels in this strain, there is no detectable change in pre–mRNA splicing. Therefore, while the mis-regulation of Mud1 levels in these 3′ end mutants suggests that, similar to its human homolog, Mud1 levels in yeast may be subject to negative regulation via its 3′ end processing [43], it nevertheless appears that the splicing defect observed in these strains is not a consequence of Mud1 overexpression.
Interestingly, several of the strains, including bdf1Δ, yth1-ts, cft2-ts, and fip1-ts, showed an ∼2-fold increase in the levels of both the U1 and U2 snRNAs. Although spliceosomes function as an equimolar complex of all five snRNAs, the total cellular levels of the snRNAs vary: in yeast, the U2 snRNA is the most abundant [44], while in mammals the U1 snRNA is most abundant [45]. While recent work demonstrates the cellular defects associated with decreased levels of snRNA [46], it is less clear whether increases in their levels will impart a defect on global splicing. Nevertheless, because each of these strains shows a similar increase in these snRNA levels but distinct splicing defects, it seems unlikely that the changes in snRNA levels alone can explain the observed splicing phenotypes. However, it is not inconceivable that small changes in levels for one or more of these transcripts could lead to the observed splicing defects. As such, additional work will be necessary to determine the functional consequences of these mutations.
Here we present the results of a global survey designed to identify the full subset of cellular factors in the budding yeast, Saccharomyces cerevisiae, that impact the efficiency of pre–mRNA splicing. As a complement to other recently described genetic and physical genome-wide approaches, in this work we have developed an approach that allows for a direct readout of the accumulation of specific RNA species in the background of thousands of different mutant strains. An important strength of a genome-wide screen such as this is its unbiased approach. By directly measuring the splicing efficiency of endogenous transcripts, this method avoids bias generated using reporter constructs. Moreover, the ability to examine numerous different transcripts allowed us to distinguish the natural variation in the spliceosomal factors that are required for efficient splicing of different intron-containing transcripts. Indeed, by systematically examining the precursor levels in the background of each strain, mutations can be identified which result in a change in splicing efficiency regardless of their previously described functions. In the work described here, mutations in scores of genes with no previously known role in splicing were identified, some of which impacted the splicing of all five canonical transcripts examined and some of which impacted only a subset of them. While some of these factors have been further characterized and discussed here, many have not (see Table S2). To be sure, as is the case with all genetic screens, it is impossible on the basis of these screen data alone to ascribe a direct role for any of these candidate factors in the splicing pathway. Rather, the identification of these different factors can be seen as generating a rich dataset from which hypotheses can be generated and tested for their mechanistic underpinnings.
Beyond known splicing factors, the most highly over-represented set of factors identified in this work function in chromatin remodeling. One particularly interesting mutation that was identified was the bdf1Δ mutant. In budding yeast, Bdf1 has been demonstrated to play a role precisely at the interface of transcription initiation and chromatin remodeling. Based in part on its physical interaction with the Taf7 subunit of TFIID, yeast Bdf1 has been proposed to function as the missing C-terminal portion of the higher eukaryotic TAFII250 [41], the largest subunit of the TFIID complex. More recently, it has become clear that Bdf1 interacts with Swr1 and functions in recruiting the entire SWR1 chromatin remodeling complex to nucleosomes. A recent genome-wide study demonstrates that Bdf1 is enriched on the +1 and +2 nucleosomes of actively transcribed genes [47], and that it coincides with the localization of Vps72, another component of the SWR1 complex, and another component which was identified in our screen (Figure 5).
Remarkably, we demonstrate here that the splicing of nearly every intron-containing gene is negatively affected in a bdf1Δ strain, and that the quantitative defect seen in this mutant rivals that seen for canonical splicing mutants. Given its role in global gene expression, one possible explanation for our results in the bdf1Δ strain is that the transcription of some key splicing factor is repressed by this mutation, causing a decrease in splicing efficiency. Indeed, early work on Bdf1 from the Séraphin lab suggested a role in global transcription, including transcription of the spliceosomal snRNA genes [48]. However, our microarray analyses show essentially normal RNA levels of all known splicing components in the bdf1Δ strain (see Figure S4). Moreover, our microarray data assessing the snRNA levels themselves are entirely consistent with Séraphin's original observations and demonstrate that none of the five wild type snRNAs are decreased in cellular level during growth at 30°C in the bdf1Δ mutant; rather, there are subtle increases in the U1 and U2 snRNA levels. Importantly, our ChIP-QPCR experiments in the bdf1Δ strain demonstrate a decreased occupancy of the U1 snRNP on all four intron-containing genes we tested, suggesting the intriguing possibility that Bdf1 plays a direct role in connecting pre–mRNA splicing with chromatin remodeling and transcription initiation.
In considering such a role for Bdf1, it is important to note that the yeast BDF1 gene has a close sequence homolog in the BDF2 gene. These two genes are genetically redundant, in so much as both single gene deletions are viable but the double mutant bdf1Δ/bdf2Δ is lethal. Moreover, it has been shown that these two genes evolved from a single ancestral gene following a whole-genome duplication event [49]. Yet surprisingly, unlike the bdf1Δ strain, the bdf2Δ strain showed no signs of a splicing defect either in our screen or when examined by splicing sensitive microarrays. Moreover, unlike the bdf1Δ strain, there was no apparent decrease in U1 snRNP ChIP-QPCR signal in the bdf2Δ strain. In considering a mechanism whereby Bdf1 connects transcription initiation and chromatin remodeling with pre–mRNA splicing, it is worth noting that, unlike human genes, the majority of yeast genes do not contain an intron. As such, co-transcriptional recruitment of the spliceosome is unnecessary for most yeast genes. We are intrigued by the possibility that, in the time since the duplication event, Bdf2 has evolved to a point where it retains the capacity to recruit RNA polymerase but has lost the ability to efficiently connect splicing with transcription. Such a scenario would explain the differences observed between the bdf1Δ and bdf2Δ microarrays and U1 snRNP ChIP-QPCR data. It would also likely explain the previously published results that Bdf1 shows higher sequence conservation with the C-terminal domain of human TAFII250 than does Bdf2 [41]. Given such a model for the divergence of Bdf1 and Bdf2 functions, the differences in protein sequence between these two proteins may prove informative for deciphering the mechanism of Bdf1 activity.
In addition to the over-representation of factors marking the 5′ end of genes, our screen identified a number of factors involved in the 3′ end processing of mRNAs. Splicing-sensitive microarrays confirm a broad splicing defect in a mutant of Yth1, the homolog of human CPSF30, and transcript-specific splicing defects in mutants of Cft2 and Fip 1, the homolog of human CPSF100 and a component of the polyadenylation factor complex PF I, respectively. In higher eukaryotes, components of the 3′ end processing machinery have been shown to physically associate with components of the U2 snRNP [50] and U2AF65 [51]; moreover, in vitro studies demonstrate a functional link between the pre–mRNA splicing and 3′ end processing pathways [52]. The interactions between these two pathways in mammalian systems have led to the proposal that the 3′ end machinery plays an important role in terminal exon definition. Whereas the exon-definition model for mammalian spliceosome assembly posits that internal exons are defined by interactions between U1 and U2 snRNP components across an exon [53], definition of terminal exons is achieved by interactions between the 3′ end processing machinery and the U2 snRNP (Figure 7A), imposing a functional connection between the pathways. Yet because of the relatively short length of S. cerevisiae introns, and the limited number of genes that are interrupted by multiple introns, splicing in yeast has long been considered to proceed through a model of intron-definition. Nevertheless, the Keller lab recently demonstrated that some conditional alleles of YSH1/BRR5 lead to a decrease in splicing efficiency [54]. Our demonstration here of pre–mRNA splicing defects in the background of additional mutants in 3′ end processing mutants suggests the intriguing possibility that some of the basic interactions that facilitate exon-definition in higher systems may also be present in budding yeast. Indeed, further characterizing the mechanism by which these 3′ end processing factors are affecting splicing in yeast may provide important insights into the mechanisms by which exon-definition is accomplished in higher eukaryotes.
An important strength of an approach such as this is the genome-wide perspective that it provides. Figure 7B shows a model of an idealized transcript along with the functional location of a subset of the factors that have been examined in our screen. It is striking to note that many of the factors identified here function both during transcription initiation (Bdf1 and others) and termination (Yth1, Cft2, Fip1, and others), thereby defining the beginning and ends of the first and last exons, respectively. In this work, we have identified not only those factors whose disruption leads to a functional defect in splicing efficiency, but in many cases the specific transcripts whose splicing is affected.
More broadly, the work presented here demonstrates the feasibility of quantitating cellular RNA levels in the background of large mutant strain collections. While our current approach examined splicing efficiency in the context of optimized growth conditions, a similar approach could be applied to identify factors necessary for efficient splicing under varying cellular or developmental growth states. Likewise, although our work focused on the levels of several pre–mRNA species, this methodology should be directly applicable to assessing the levels of nearly any cellular RNA of interest.
All experiments were performed using haploid strains. To assess the function of non-essential genes, the mat a version of the haploid deletion library from Open Biosystems [55] was used (referred to herein as non-essential strains). Likewise, to assess the function of essential genes, a collection of strains provided by the Hieter lab [25] was used (referred to herein as essential strains). In addition, a collection of strains containing previously characterized mutations in core spliceosomal components was used (from here on considered a part of the essential strains set). A complete list of the strains used in this work is included in Table S3. Unless otherwise indicated, all strains were grown in rich medium supplemented with 2% glucose (YPD) [56]. When appropriate, strains were recovered from frozen glycerol stocks on solid medium supplemented with 200 µg/ml G418 grown at either 30°C (non-essential strains) or 25°C (essential strains). A manual pinning tool (V&P Scientific, cat.#: VP384FP6) was used to transfer cells from solid medium into 384-well microtiter plates (Greiner BioOne, cat.#: 781271) for growth in liquid media. Liquid cultures were grown in an Infors HT Multitron plate shaker at 900 rpm with 80% constant humidity. Breathable adhesive tape (VWR, cat.#: 60941-086) was used to seal the plates and reduce evaporation.
Because the growth rates of the strains being used vary significantly [36], an approach was developed to enable the systematic collection of a similar number of rapidly dividing cells for each strain. An initial liquid culture was grown in 384-well plates for two days, allowing nearly all strains to reach saturation. Because all of the strains being used are derived from a common parental strain, the cell density for each of these strains is nearly identical at saturation, allowing us to effectively ‘normalize’ the cell numbers. Using a liquid handling robot (Biomek NX), 2 µl of saturated culture were used to inoculate a fresh 150 µl of YPD. This culture was allowed to grow for four hours, an amount of time which is sufficient to allow all strains to exit lag-phase and begin exponential growth, but not so long as to result in a large variation in cell densities among the strains (Figure S6). For the non-essential strains, all growth was conducted at 30°C. For the essential strains, the initial growth was done at 25°C (a permissive temperature for all strains), but the saturated cells were back-diluted into plates containing media pre-warmed to 30°C (a non-permissive temperature for many, but not all, of the strains) and allowed to continue growing at 30°C for four hours. For both the non-essential and the essential strains, two independent biological replicates were initiated from each saturated plate. After four hours of outgrowth, cells were harvested by centrifugation at 4000×g for five minutes. The cell pellets were flash frozen in liquid N2 and stored at −80°C until further processing.
Isolation of total cellular RNA was performed using custom protocols written for a Biomek NX liquid handling system. To each frozen cell pellet collected as described above, 50 µl of Acid Phenol: Chloroform (5∶1, pH<5.5) and 25 µl of AES buffer (50 mM sodium acetate (pH 5.3), 10 mM EDTA, 1%SDS) were added. The plates were sealed with plastic CapMats (Greiner BioOne, cat.#: 384070) and vortexed for five minutes at top speed on a plate vortex. The plates were incubated for 30 minutes in a water bath at 65°C with intermittent vortexing. After incubation, the plates were spun for one minute at 1000×g. An additional 35 µl of AES buffer was added to each well, and after mixing the organic and aqueous phases were separated by centrifugation for five minutes at 3000×g. Using a slow aspiration speed, 40 µl of the upper phase containing the RNA were robotically transferred to a new 384-well microtiter plate. The transferred aqueous phase was mixed with 3 volumes of RNA Binding Buffer (2 M Guanidine-HCl, 75% isopropanol) and passed through a 384-well glass fiber column (Whatman, cat.#: 7700-1101) by centrifugation for two minutes at 2000×g. The column was washed twice with two volumes of Wash Buffer (80% ethanol, 10 mM Tris-HCl (pH 8.0)), followed by a final dry spin for two minutes at 2000×g. To remove any contaminating genomic DNA, 5 µl of DNase Mix (1× DNase Buffer, 0.25 units of DNase I (Promega)) was added to each well and incubated at room temperature for 15 minutes. After the incubation, 80 µl of RNA Binding Buffer was added to each well of the 384-well glass fiber plate and spun as before. After washing and drying as above, 15 µl of sterile water was added to each well of the glass-fiber plate to elute the RNA into a clean 384-well microtiter plate (Greiner BioOne, cat. #: 781280). In general, this procedure yielded about 1 µg of total cellular RNA from each cell pellet. The quality of the RNA produced by this protocol is equal to our conventionally purified samples, and the effectiveness of the DNase treatment is demonstrated in Figure S7.
Total cellular RNA was converted into cDNA in 384-well microtiter plates. Of the 15 µl of RNA purified as described above, 10 µl were used in a cDNA synthesis reaction that had a total volume of 20 µl and which contained 50 mM Tris-HCl (pH 8.3), 75 mM KCl, 3 mM MgCl2, 10 mM DTT, 0.5 mM each dNTP, 5 µg dN9 primer, and 60 ng M-MLV RT. Reactions were incubated overnight at 42°C.
The cDNA reactions were diluted 30-fold with water, giving a final concentration of ∼1 ng/µl based on the initial RNA concentration, and used without any further purification as templates in high-throughput quantitative PCR (QPCR) reactions. The QPCR reactions were performed in a reaction volume of 10 µl, containing 5 µl of template (∼5 ng of template), 10 mM Tris-HCl (pH 8.5), 50 mM KCl, 1.5 mM MgCl2, 0.2 mM each dNTP, 0.25× SYBR Green, 5% DMSO, 0.7 ng Taq DNA polymerase, and 250 nM forward and reverse primers. The sequences of the primers used for each targeted RNA are shown in Table S4. Standard curves were generated consisting of 4-fold serial dilutions of genomic DNA and covering a range of 1.6×105 molecules. Each primer pair was well-behaved, showing an amplification efficiency of between 86% and 97% (Figure S8). Two technical replicates were measured for each biologically independent sample, generating four independent measurements for each of the ∼5500 mutant strains.
On the basis of standard curves generated using QPCR, relative nanogram quantities were calculated for every RNA transcript within each of the ∼5500 strains tested. To assess reproducibility, coefficients of variation (CV) were determined for each primer pair and each strain. The vast majority of these were highly reproducible, both overall and on a per plate basis. As an initial quality filter, we chose to exclude any samples for which the CV was greater than 0.25.
Because no simple mechanism exists to normalize for variability in each of our experimental steps, we instead chose to measure the levels of six different RNAs in each of the samples and use these to determine a composite normalization value to account for the overall yield in our procedure. The six RNAs were: U1 snRNA, Scr1 (SRP) RNA, Tef5 mRNA, Tub1 mRNA, Srb2 mRNA and Faa1 mRNA. These RNAs were chosen because their biological functions are diverse and their cellular levels vary over a broad range (∼300-fold, Table S1). For both independent biological replicates of every strain, a composite normalization constant, , was calculated according the following formula:For each primer pair, represents the relative nanogram quantity calculated for an individual sample. Similarly, represents the median value determined for a given primer pair on an individual QPCR plate run. Because of the subtle variations that are apparent from one plate run to the next, we found that this per plate normalization using gave us the most robust data. By determining the ratio of for every primer pair, a relative abundance of total RNA can be calculated for every sample. As seen in Figure S9, a histogram of values follows a normal distribution in log2 space with a variance of 1.5 units. A second filter at the level of values was introduced which allowed for the filtering of samples with very low amounts of cDNA.
For strains that passed this filter, the normalized levels of a given RNA were determined according to the following formula:The relative amount of RNA in a given strain was then determined according to the following formula:For each primer pair, represents the median value of the normalized RNA levels determined within a given QPCR plate. For each biological replicate of every strain, both the and the values are available through Gene Expression Omnibus (GEO) using accession number GSE34330.
To determine the subset of strains that cause a statistically significant increase or decrease in precursor levels, we employed the Significance Analysis of Microarrays, or SAM, program [37]. While this software was originally designed for the analysis of microarray data, a significance analysis of our QPCR data is subject to similar concerns regarding multiple hypothesis testing. For each RNA, SAM analysis was performed on the four values, comprised of both technical and biological replicates that were generated for each of the ∼5500 strains. For each transcript, a one class SAM analysis was performed where the Δ value was adjusted to minimize the false discovery rate (FDR), yielding the following values: for the U3 precursor using Δ = 0.983, FDR = 0.045; for the Tub3 pre–mRNA using Δ = 0.91, FDR = 0; for the Rpl31b pre–mRNA using Δ = 1.061, FDR = 0.003; for the Ubc13 pre–mRNA using Δ = 0.978, FDR = 0.002; and for the Tef5 pre–mRNA using Δ = 0.99, FDR = 0.
The candidate non-essential deletion strains were grown to saturation in YPD at 30°C, then back diluted in 25 ml cultures in flasks at a starting A600∼0.2 and allowed to grow at 30°C until they reached an optical density of between A600 = 0.5 and A600 = 0.7. The candidate essential strains were initially grown at 25°C in YPD, then shifted to the indicated temperatures for 15 minutes after they reached an optical density of between A600 = 0.5 and A600 = 0.7. In parallel with the collection of the mutant strains, wild type isogenic controls were grown and collected under the same conditions as the mutant strains examined. Total cellular RNA samples were isolated, converted into cDNA, and fluorescently labeled as previously described [31]. All microarrays were performed as two-color arrays comparing mutant and wild type strains, each grown under identical conditions. Both raw and processed microarray data are available through GEO using accession number GSE34330.
The U1C-Tap bdf1Δ and U1C-Tap bdf2Δ strains were generated by deleting the appropriate genes in the background of the U1C-Tap strain from Open Biosystems [57] using standard techniques. The strains were grown at 30°C in rich medium supplemented with 2% glucose (YPD) until they reached an optical density of A600∼0.7. The chromatin was cross-linked with 1% formaldehyde for 2 minutes at 30°C. Glycine was added at a final concentration of 125 mM and the cultures were left shaking for another 5 minutes. Cell pellets from 50 ml of culture were collected by centrifugation at 1620×g for 3 minutes, then washed with 25 ml ice-cold 1× PBS and the pellet stored at −80°C. The pellets were resuspended in 1 ml Lysis buffer (50 mM Hepes pH 7.5, 140 mM NaCl, 1 mM EDTA, 1% TritonX-100, 0.1% Na-deoxycholate supplemented with protease inhibitors) and lysed in the presence of 500 µl 0.5 mm glass beads in a beat beater. The lysate was collected by centrifugation at 1000×g for 1 minute, and then pre-cleared by spinning for 15 minutes at 14000 rpm in a tabletop centrifuge at 4°C. The pellet was re-suspended in another 1 ml of Lysis buffer and the chromatin was sheared to an average size of 300 bp (range 100–500 bp) by means of a Bioruptor sonicator. The sample was clarified by 2 cycles of centrifugation at 14000 rpm for 15 minutes in a tabletop centrifuge at 4°C and the resultant chromatin solution frozen and stored at −80 C. From the chromatin samples a 1% Input sample was retained, and then each sample was split equally between a Mock IP and an IP sample. The IP samples were incubated with 5 µl 0.5 mg/ml anti-Tap Antibody (Thermo Scientific, CAB1001). After 2 hours at 4°C on a rotator, 25 µl of protein A/G-agarose resin (#Sc-2003Santa Cruz) was added to all samples and they were further incubated for another 2 h at 4°C. The resin was washed twice with 1 ml Lysis buffer, twice with 1 ml Wash buffer (10 mM Tris-HCl, 25 mM LiCl, 0.5% NP-40, 0.5%Na-deoxycholate, 1 mM EDTA) supplemented with 360 mM NaCl, twice with 1 ml Wash buffer, and finally twice with 1 ml TE. The first wash was a brief one, followed by a 15 minute incubation of the samples on a rotator at 4°C for the second wash. In between washes, the resin was collected by short spins at 2000 rpm in a tabletop centrifuge. The resin was resuspended in 100 ul Elution buffer (50 mM Tris-HCl pH 8.0, 5 mM EDTA, 1% SDS) and the immunoprecipitated material was eluted from the beads by incubating at 65°C for 30 minutes with occasional tapping. To reverse crosslinks, the IPs and the 1% Input samples were incubated overnight in a 65°C water bath. The next day, the samples were treated with 12.5 µl 20 mg/ml Proteinase K solution and incubated at 42° for 2 h. The DNA was then purified by using a Cycle Pure Kit (Omega Bio-Tek, D6492-01) following the manufacturer's instructions and eluted in a final volume of 120 µl.
Quantitative real-time PCR was performed on a Roche Light Cycler 480 machine as described above, using the 1% Input sample to generate a standard curve for each of the primer pairs we used. For the primers used in the screen, the sequences are available in Table S4. The primers for the different regions of actin gene and the PMA1 gene are the same as previously published [5]. For each sample the Mock IP value calculated as percent input was subtracted from the IP value (in percent input). Then, a fold enrichment value was calculated, by dividing these values by the PMA1 value.
An overexpressing plasmid containing a full-length copy of the Mud1 gene including ∼500 bp up- and down-stream of the ORF was transformed into BY4741 (Open Biosystems). This strain and a control strain containing the empty vector were grown in 25 ml minimal media until they reached an optical density of A600∼0.5–0.6. RNA isolation was performed as previously described [31], and cDNA synthesis and Q-PCR were performed as described above. The primer sequences are found in Table S4.
|
10.1371/journal.pgen.1003938 | Co-evolution of Human Leukocyte Antigen (HLA) Class I Ligands with Killer-Cell Immunoglobulin-Like Receptors (KIR) in a Genetically Diverse Population of Sub-Saharan Africans | Interactions between HLA class I molecules and killer-cell immunoglobulin-like receptors (KIR) control natural killer cell (NK) functions in immunity and reproduction. Encoded by genes on different chromosomes, these polymorphic ligands and receptors correlate highly with disease resistance and susceptibility. Although studied at low-resolution in many populations, high-resolution analysis of combinatorial diversity of HLA class I and KIR is limited to Asian and Amerindian populations with low genetic diversity. At the other end of the spectrum is the West African population investigated here: we studied 235 individuals, including 104 mother-child pairs, from the Ga-Adangbe of Ghana. This population has a rich diversity of 175 KIR variants forming 208 KIR haplotypes, and 81 HLA-A, -B and -C variants forming 190 HLA class I haplotypes. Each individual we studied has a unique compound genotype of HLA class I and KIR, forming 1–14 functional ligand-receptor interactions. Maintaining this exceptionally high polymorphism is balancing selection. The centromeric region of the KIR locus, encoding HLA-C receptors, is highly diverse whereas the telomeric region encoding Bw4-specific KIR3DL1, lacks diversity in Africans. Present in the Ga-Adangbe are high frequencies of Bw4-bearing HLA-B*53:01 and Bw4-lacking HLA-B*35:01, which otherwise are identical. Balancing selection at key residues maintains numerous HLA-B allotypes having and lacking Bw4, and also those of stronger and weaker interaction with LILRB1, a KIR-related receptor. Correspondingly, there is a balance at key residues of KIR3DL1 that modulate its level of cell-surface expression. Thus, capacity to interact with NK cells synergizes with peptide binding diversity to drive HLA-B allele frequency distribution. These features of KIR and HLA are consistent with ongoing co-evolution and selection imposed by a pathogen endemic to West Africa. Because of the prevalence of malaria in the Ga-Adangbe and previous associations of cerebral malaria with HLA-B*53:01 and KIR, Plasmodium falciparum is a candidate pathogen.
| Natural killer cells are white blood cells with critical roles in human health that deliver front-line immunity against pathogens and nurture placentation in early pregnancy. Controlling these functions are cell-surface receptors called KIR that interact with HLA class I ligands expressed on most cells of the body. KIR and HLA are both products of complex families of variable genes, but present on separate chromosomes. Many HLA and KIR variants and their combinations associate with resistance to specific infections and pregnancy syndromes. Previously we identified basic components of the system necessary for individual and population survival. Here, we explore the system at its most genetically diverse by studying the Ga-Adangbe population from Ghana in West Africa. Co-evolution of KIR receptors with their HLA targets is ongoing in the Ga-Adangbe, with every one of 235 individuals studied having a unique set of KIR receptors and HLA class I ligands. In addition, one critical combination of receptor and ligand maintains alternative forms that either can or cannot interact with their ‘partner.’ This balance resembles that induced by malfunctioning variants of hemoglobin that confer resistance to malaria, a candidate disease for driving diversity and co-evolution of KIR and HLA class I in the Ga-Adangbe.
| Major Histocompatibility Complex (MHC) class I molecules are present on the surface of most mammalian cells. There they function as ligands for various receptor families on two types of lymphocyte: the cytotoxic T lymphocyte (CTL) of adaptive immunity and the natural killer (NK) cell of innate immunity [1], [2]. NK cells also contribute to reproduction, during formation of the placenta [3]. A key component of all MHC class I molecules is a short peptide, a product of intracellular protein degradation, that is bound during assembly of the MHC class I molecule in the endoplasmic reticulum. After transport to the cell surface, the complexes of peptide and MHC class I molecule are presented for surveillance by NK cell and CTL receptors [4]. In healthy tissue the presented peptides all derive from normal proteins and do not usually stimulate an immune response. In unhealthy tissue, that is infected, cancerous or in other ways damaged, changes occur in the spectrum of peptides presented, which lead to activation of NK cell and CTL mediated immunity [5], [6].
In mammals, the selection pressures imposed by diverse and rapidly evolving pathogens have driven the evolution of gene families encoding a variety of MHC class I molecules [7]–[9]. These include conserved and highly polymorphic MHC class I molecules with species-specific character [10]. The human MHC, the HLA complex on chromosome 6p21, has three highly polymorphic MHC class I genes, (HLA-A, -B and -C) each of which has thousands of alleles [11], [12]. Some of the alleles have a worldwide or continent-wide distribution, others are more localized geographically and the majority constitutes rare variants that have been discovered through sequence-based HLA typing of huge cohorts of potential bone-marrow donors for clinical transplantation [12], [13]. Evolving through mechanisms of point mutation and recombination, pairs of allotypes are distinguished by between one and 51 amino-acid substitutions [14], [15]. Consistent with natural selection having driven this diversification, the common substitutions are predominantly at ‘functional’ positions of the HLA class I molecule that influence the peptide-binding specificity or the site of interaction with one of the lymphocyte receptors that engage HLA class I molecules [16]–[19].
The antigen receptors of CTL bind to the upper face of the HLA class I molecule, which is formed by the α helices of the α1 and α2 domains and the peptide bound between them [20], [21]. The genes encoding these αβ T-cell receptors (TCR) are diversified during T-cell development by mechanisms of somatic recombination and somatic mutation. These processes produce acquired changes that are not passed on from one generation to the next. In addition, the conserved CD8 co-receptor of CTL, binds predominantly to the conserved α3 domain of the HLA class I molecule [22]. Largely conserved is the leukocyte immunoglobulin-like receptor (LILR) B1, which also binds to the α3 domain [23] and is expressed by some NK cells [24].
NK cells and some T cells express killer-cell immunoglobulin-like receptors (KIR) [25]. They bind to the same upward face of the HLA class I molecule as the TCR, with an overlapping but different orientation [18], [26]. KIR recognition of HLA class I is primarily influenced by polymorphisms in the carboxy-terminal half of the α helix of the α1 domain [18], [27]. To a first approximation, KIR recognize four mutually exclusive epitopes of HLA-A, -B and -C molecules [28], [29]: the A3/11 epitope carried by a small subset of HLA-A allotypes, the Bw4 epitope carried by larger subsets of HLA-A and -B allotypes, the C1 epitope carried by many HLA-C and the HLA-B*46 and -B*73 allotypes, and the C2 epitope carried by all the HLA-C allotypes that lack the C1 epitope. Each of these four ligand-receptor interactions is heterogeneous, being further diversified by allelic polymorphism of both the HLA class I and the KIR, as well as by the sequence of the bound peptide [5], [30]–[32]. By providing resistance to specific diseases, this combinatorial diversity is believed to give individuals and populations the means to fight wide ranging pathogen diversity [9], [28], [29], [33].
The KIR locus on chromosome 19q13.4 exhibits an extensive variability in human populations, one comparable to that of the HLA class I genes [28], [34]. KIR haplotypes differ in the content and copy number of KIR genes and are further differentiated by allelic polymorphism of the constituent genes [28], [35]–[37]. On the basis of gene content, human KIR haplotypes, but not their counterparts in other hominoid species [38], divide into two groups [36]. These ‘A’ and ‘B’ haplotype groups are maintained by all human populations and are differentially associated, either alone or in combination with HLA class I, with susceptibility to diverse diseases, reproductive success, and the outcomes of therapeutic transplantation [28], [39]–[46]. The nature of these correlations has suggested a scenario in which the A and B haplotypes are maintained by competing selection on the functions that NK cells serve in resisting infectious disease and in establishing the placenta during the early stages of pregnancy [29].
Although KIR diversity has been studied in numerous (N = 105) human populations at the low-resolution of KIR gene content [47], [48], high-resolution analyses of allelic and haplotypic diversity have been few (N = 4) and involved populations such as the Japanese and Yucpa Amerindians that have restricted genetic diversity as a consequence of historical population bottlenecks [49]–[52]. By contrast, little is known of the KIR system and its interactions with HLA class I in sub-Saharan Africans, the human populations with highest genetic diversity [53], [54]. By using a novel combination of molecular, genetic and computational methods we have defined at high-resolution the rich diversity of KIR and HLA class I in the Ga-Adangbe population of one village in southern Ghana, West Africa.
Variation in the functionally interacting families of killer-cell immunoglobulin-like receptors (KIR) and polymorphic HLA class I molecules was studied in the Ga-Adangbe people of Ghana. To facilitate this genetic analysis, the study population was chosen to comprise 104 mother-child pairs, as well as an additional 27 unrelated individuals.
Initial low-resolution analysis of the Ga-Adangbe KIR locus identified 19 KIR gene-content haplotypes (Figure 1A) and 16 different KIR genotypes (Figure S3). The 53% frequency of the KIR A haplotype (h1) is comparable to the 47% combined frequency of the 18 KIR B haplotypes (h2–h19), consistent with balancing selection having been active on the two haplotype groups [49]. The number of KIR genes per B haplotype varies from four (h9) to twelve (h11), with only two genes, KIR3DL3 and KIR2DL2/3, being detected on every haplotype. By frequency, over 10% of the Ga-Adangbe KIR haplotypes (h5, 7–10, 12, 13) lack one of the three framework genes (KIR3DL3, KIR2DL4 and KIR3DL2) that define the structure of the KIR locus [37] and its organization into centromeric and telomeric regions; haplotypes h5 and h13 lack KIR2DL4, whereas haplotypes h7–10 and h15 lack KIR3DL2. In previous studies of non-African populations such haplotypes were either absent [49], [51] or rare [50], [52]. Haplotype h12 has a duplication of the KIR2DL4 and KIR3DL1/S1 genes, of the sort that has been described previously in Europeans [28], [55], [56] and South and East Asians [57].
Seven centromeric region motifs combine with six telomeric region motifs to form the 19 Ga-Adangbe KIR gene-content haplotypes (Figure 1A). By far the most common motif is tA01, which is fixed on A haplotypes and present at a frequency of 86% in this population. Consequently, the Ga-Adangbe, as well as other sub-Saharan African populations, has significantly reduced gene-content diversity in the telomeric region of the KIR locus compared to non-African populations (p<0.001, Figure 1B). In contrast, centromeric region KIR diversity is much higher and comparable to that of other population groups.
To give a complete comparison of KIR variation in the centromeric and telomeric regions of the Ga-Adangbe KIR locus, we performed high-resolution typing to determine the allelic diversity of the component KIR genes. A total of 175 KIR variants were found, of which 126 involve allotypic differences: 32 of these being previously undiscovered (Figure S4A–C). The individual KIR genes exhibit high heterozygosity (H), particularly the KIR3DL3 framework gene, which is present on every haplotype and has H of 0.93 (Figure S4C). This heterozygosity exceeds that of the highly polymorphic HLA class I genes and is clearly an outlier amongst genome-wide multi-allelic markers from West African populations (Figure S4D–E). With addition of the high-resolution analysis, the 19 gene-content KIR haplotypes become subdivided into 208 allele-level haplotypes (Figure S5); of these a large majority (195/208; 95%) encode unique combinations of KIR proteins and have the potential to be functionally distinct (Figure 2A and S4F). Most diverse is h1, the canonical KIR A gene-content haplotype (Figure 1A), for which there are 108 different allele combinations and 100 allotype combinations (Figure S4F). Individually, none of the 18 KIR B gene-content haplotypes approaches h1 in diversity, but when both gene-content and allotype-content diversity are taken into account the 100 A and 95 B KIR haplotypes have comparable diversity as well as frequency. None of the allele-level KIR haplotypes dominate the Ga-Adangbe population; the frequency of the most common haplotype is only 6% and only 18 of the 195 functionally different haplotypes exceed a frequency of 1% (Figure 2B). Thus the Ga-Adangbe population is seen to have a rich diversity of KIR haplotypes upon which natural selection can operate.
The centromeric region of the Ga-Adangbe KIR locus exhibits a bimodal mismatch distribution, a network indicating successive formation and expansion of haplotypes, and a significantly elevated value for Tajima's D (Figure 3A–C). All these features reflect the presence of a variety of divergent haplotypes that are at comparable frequencies and maintained by balancing selection. In contrast, the telomeric region of the KIR locus displays a unimodal mismatch distribution, a star-like haplotype network pattern (Figure 3C) and a Tajima's D value significantly below that expected for neutrality (Figure 3B), features reflecting the presence of numerous closely-related variants under directional selection. Such difference in the evolution of the centromeric and telomeric KIR regions is not a general feature of human populations, as exemplified by comparison of the Ga-Adangbe with Yucpa Amerindians and US Europeans (Figure S6).
Sliding-window analysis showed that the boundary between the high diversity and low diversity parts of the Ga-Adangbe KIR haplotype does not correspond precisely with the conventional division of the locus into centromeric and telomeric regions (Figure 4). High diversity extends into the KIR3DL1/S1 gene of the telomeric region, but sharply declines at the end of exon 3 that encodes the D0 domain, resulting in low diversity that is maintained throughout the rest of the telomeric KIR region. This result is consistent with our previous analysis of KIR3DL1/S1 polymorphism worldwide, which showed that balancing selection was restricted to the D0 domain in sub-Saharan Africans [35]. The functional consequences are first that mutations in the D0 domain can abrogate cell surface expression [58] or decrease binding to HLA-B [18] and second that reduced diversity in the D1 and D2 domains favors one particular type of ligand specificity [32], [35], [59]. Segments of low diversity that are of comparable length to the one in the telomeric KIR region are infrequent in the genomes of sub-Saharan Africans, as shown from analysis of Yoruba West Africans (p<0.01: Figure 4), a population related closely to the Ga-Adangbe [60], [61]. In summary, intron 3 of the KIR3DL1/S1 gene marks the boundary between a diversified centromeric part and a conserved telomeric part of the KIR locus in sub-Saharan Africans. The centromeric region of the KIR locus encodes inhibitory receptors KIR2DL1 and KIR2DL2/3 that recognize the C1 (KIR2DL2/3) and C2 (KIR2DL1) epitopes of HLA-C, whereas the telomeric region encodes inhibitory KIR that recognize the A3/11 epitope (KIR3DL2) of HLA-A and the Bw4 epitope (KIR3DL1/S1) of HLA-A and -B [37].
In the Ga-Adangbe we identified 26 HLA-A, 32 HLA-B and 23 HLA-C allotypes (Figure 5). The numbers of alleles and their composition are typical of West African populations, which are readily distinguished from other African population groups by clustering analyses based on these genes alone (Figure S7). In contrast to KIR, none of the HLA class I alleles were novel or private to the Ga-Adangbe population. Significantly high values of Tajima's D provide good evidence for balancing selection having acted on all three polymorphic HLA class I genes (Figure 6A), as observed previously for other populations [34]. HLA class I has various roles in immunity and reproduction that involve binding to peptide fragments, and serving as ligands for KIR and other lymphocyte receptors. To assess if any of these functions influenced HLA class I allele-frequency distributions in the Ga-Adangbe we analyzed the amino acid sequence of each of the binding motifs separately. In this analysis, we considered both the binding site for peptide antigens, and the sites of interaction with four types of lymphocyte receptor: the TCR and CD8 of CTL, and the KIR and LILR of NK cells. Of the three extracellular domains of the HLA class I molecule, the α1 and α2 domains mediate interactions with peptide, TCR and KIR, whereas the α3 domain mediates interaction with LILR and CD8 (Figure 6B and Figure S8A–C). Because some of the motifs overlap, we analyzed only those residues located exclusively in each type of binding site. We analyzed the allele-frequency spectrum of each motif using the Ewens-Watterson test [62] and compared the deviation from neutral expectations for each motif using a normalized statistic (Fnd [63]). The results showed strong evidence for balancing selection acting on the peptide-binding residues of all three HLA class I molecules, but no evidence for natural selection acting on the TCR binding motifs (Figure 6C). The CD8 binding site is largely invariant for HLA-B and -C, whereas for HLA-A variation is introduced at residue 245 by A*68:01. Mutation of residue 245 can influence CD8 binding [64] but there was no evidence of this being selected in Ga-Adangbe (Figure 6C). These distinctions among motifs demonstrate that our analysis differentiates the effects of natural selection acting independently on each of the functional motifs of HLA class I molecules. The analysis is also consistent with codon-by-codon tests for selection, which show that HLA class I evolution in hominids has been driven by diversification of the peptide-binding motifs and not TCR or KIR binding motifs (Figure S9). Such independent evolution has likely been facilitated by the extensive intra-locus recombination and gene conversion that shaped HLA class I diversity by shuffling functional motifs among allotypes [14], [15].
Analysis in Ga-Adangbe of the frequency distributions of HLA-A and -B motifs that interact exclusively with KIR3DL1, gave evidence for balancing selection that was statistically significant and of magnitude greater than either the peptide-binding motifs and/or the locus as a whole (Figure 6C). Although the magnitude of the Fnd values approached those of their respective peptide-binding domains, there was less evidence for balancing selection at the KIR-exclusive motif of HLA-B in populations not from West Africa, and the observation only reached statistical significance in the Ga-Adangbe (Figure 6D and Figure S10). Included in the KIR-exclusive motif is arginine at position 83, a component of the Bw4 epitope and the only Bw4 residue necessary for HLA-B binding to KIR3DL1 [65]. In the Ga-Adangbe arginine 83 is present in 16 HLA-B allotypes having a combined frequency of 46% (Figure 5).
HLA-B*35:01/*53:01 and HLA-B*49:01/B*50:01 comprise pairs of HLA-B allotypes that differ only by presence/absence of the Bw4 motif [66]. This difference determines whether these HLA-B alloypes bind to KIR3DL1 (B*53:01 and B*49:01) or do not (B*35:01 and B*50:01) [67]. HLA-B*35:01 and HLA-B*53:01 are both common in the Ga-Adangbe (Figure 5) suggesting that distinction between binding or not binding to KIR3DL1 has been a major influence on the balancing selection acting on HLA-B, and that this variation substantially augments the diversity of peptide-binding function. Further, it implies that the presence/absence polymorphism of Bw4 is driven by the benefits of diversifying the interaction of HLA-B with KIR3DL1, and not its interaction with peptides.
For HLA-A, polymorphic residues within the KIR-exclusive motif include positions 17 and 142 and are provided primarily by the HLA-A*02, -A*30 and -A*68 allotypes. None of these allotypes is known to interact with KIR and all are common in sub-Saharan African populations [13], [68]. In contrast to HLA-A and -B, the HLA-C residues that interact exclusively with KIR are monomorphic (Figure S8) and all expressed HLA-C allotypes are presumed to interact with KIR [31].
To examine the impact of natural selection on the LILRB1-contacting residues of HLA class I we first performed likelihood ratio tests for selection on hominid α3 domains. This analysis revealed evidence for diversifying selection on HLA-C, and codon-by-codon analysis identified the LILRB1-contacting residues for all three HLA class I molecules (Figure S9B). Although statistical confidence from this phylogenetic-based analysis was low (Figure S9B), frequency-based Fnd analysis suggested that balancing selection has acted on the LILRB1-interacting motifs of Ga-Adangbe HLA-A and -B (Figure 6C). Their Fnd values were greater in magnitude than those of the respective peptide-binding motifs and reached statistical significance for HLA-A. In the Ga-Adangbe, HLA-A molecules that bind LILRB1 with low affinity (193A/194V, 47%) are at similar frequency as the high-binding allotypes (193P/194I, 53%) [69].
Together, these results demonstrate that balancing selection has acted on HLA class I in the Ga-Adangbe population, resulting in the evolution of a diversity of ligands for interaction with NK cell receptors. We next measured the scale of KIR and HLA combinatorial diversity and assessed if the interacting receptors and ligands continue to co-evolve.
Each individual in the Ga-Adangbe panel has a unique compound genotype of KIR and HLA-A, -B and -C (Figure S11). Based on the known interactions between KIR and the C1, C2, Bw4 and A3/11 epitopes of HLA class I, we determined the number of functional ligand-receptor pairs for all members of the Ga-Adangbe panel. The frequencies of these values within the panel gave a normal distribution (Figure 7A) with a mean number of ligand-receptor interactions of eight (95% CI of 3–12).
To assess for co-evolution of KIR with HLA in the Ga-Adangbe, we used the Mantel test of congruence between distance matrices to look for population-wide correlation between KIR and HLA class I genotypes [70]. These analyses revealed significant correlations of matrices for KIR3DL1/S1 and HLA-B genotypes (p<0.001), for KIR2DL2/3 with HLA-C (p<0.01), and for KIR2DL1 with HLA-C (p<0.01) (Figure 7B–D). However, no correlations were observed between either KIR3DL1/S1 or KIR3DL2 and HLA-A.
Residues 31, 44 and 86, in the D0 of KIR3DL1, are in complete LD and were correlated in synergistic action with three groups of HLA-B residues (Figure 7B and S12A–B). That the correlation also involves residues of the Bw4 epitope, is consistent with interaction between KIR3DL1 and HLA-B being the underlying mechanism driving their population frequencies. Further contributions from HLA-B are made by residue 114, and three residues in complete LD, 24, 45 and 194; the latter contacting LILRB1 (Figure 7B) and having enhanced diversity in the Ga-Adangbe (Figure 6C). Residues 24 and 114 are located in the peptide binding B and F pockets, respectively, which define the anchor residues of the peptide that is presented by HLA-B (Figure S8 and [71]). This result suggests that sequences of the peptides presented by HLA-B contributed to its co-evolution with KIR3DL1 in the Ga-Adangbe. A previous analysis showed replacement of isoleucine 194 in HLA-B with valine reduced the interaction with KIR3DL1 as measured by NK inhibition [65]. The study also demonstrated that polymorphism at positions in the B and F pockets of the peptide-binding site can impact 3DL1-mediated inhibition, either alone or in concert with residue 194. Moreover, the correlations observed here between HLA-B and KIR3DL1 are all supported by the results of functional studies, which assessed the influence of the sequence of the peptide bound to HLA-B on the binding to KIR3DL1 [18], [27], [32], [72].
Differences between the KIR2DL2 and KIR2DL3 subsets of KIR2DL2/3 allotypes have had major impact in the co-evolution of KIR2DL2/3 with HLA-C. For HLA-C, the major factor in this co-evolution is a group of seven residues in LD (positions 194, 261, 273, 311, 313, 332, 345), which includes residue 194 that contacts LILRB1 (Figure 7C). This group of residues distinguishes HLA-C*07, a common allotype in many populations, from all other C1-bearing HLA-C allotypes (Figure 5). Because of the strong LD between KIR2DL2/3 and KIR2DL1 (D′ = 0.87), this group of residues also correlates with C2-specific KIR2DL1 (not shown) although this receptor does not recognize C1-bearing HLA-C*07 [31]. The analysis revealed an independent influence from residue 49 (Figure 7D) which distinguishes HLA-C*04, the most frequent HLA-C allotype in the Ga-Adangbe (Figure 5), from all other C2-bearing allotypes (Figure 5). Five residues of KIR2DL1 (positions 154, 163, 182, 216 and 245) contribute to its co-evolution with HLA-C. These five residues, which are in complete LD, include residue 182 that contacts HLA, and residue 245 that modulates both ligand-binding and signaling functions [26], [73]. These are all residues that distinguish KIR2DL1*003 from KIR2DL1*004, encoded by the common KIR2DL1 alleles of the centromeric A and B motifs, respectively (Figure 2 and Figure S5). For the cenA-containing KIR haplotypes, which carry KIR2DL3 and KIR2DL1, 80% of the KIR2DL1 allotypes have histidine 182 and arginine 245 and are strong high-expressing C2 receptors, whereas the other 20% of allotypes have cysteine 245 and are weak, low-expressing C2 receptors. In contrast, 80% of the cenB haplotypes carry KIR2DL2 and either lack KIR2DL1 (49%) or encode weak, low-expressing allotypes having arginine 182 and cysteine 245 (31%).
Variable interactions between KIR and HLA class I influence the immunological and reproductive functions of NK cells. Because of the complexity of the KIR gene family, population genetic studies have been limited in large part to low-resolution analyses of KIR gene-content variation [47], [48], [74]. In developing methods for high-resolution KIR genotyping, we previously focused on Asian and Amerindian populations having inherently low genetic diversity because of their demographic histories [49], [51]. At the other end of the human spectrum are sub-Saharan African populations, who have, genome-wide, greatest genetic diversity. Reflecting this general characteristic, are the results presented here from our high-resolution analysis of KIR and HLA-A, -B and -C variation in the Ga-Adangbe population of Prampram, a coastal village in Ghana, West Africa. Segregating in this population are 81 HLA and 175 KIR variants, numbers that are four- to five-fold higher than the 19 HLA and 30 KIR variants we previously described for the Yucpa population of South American Indians [49]. Thus, we find the Ga-Adangbe population to be highly heterozygous, with every individual having a unique compound genotype for KIR and HLA class I. As they have similar levels of KIR gene-content (Figure 1) and HLA class I [13], [75] heterozygosity to other West African populations, the Ga-Adangbe provide an archetypal population for investigating immune diversity. The consequence of genetic individuality is predicted to be functional individuality in the immune responses to viruses and other pathogens against which NK cells and CTL are important elements of the defences of human immune systems.
The unprecedented diversity of HLA and KIR haplotypes and alleles, and their relatively even distributions, argue that strong balancing selection on these loci has been a persistent force in the history of the Ga-Adangbe population. Probable causes of this selection include reproductive success [29] and the fluctuating pressures imposed by the variety of human pathogens in West Africa and their continual evolution to evade the immune systems of their human hosts [33]. Consistent with these roles, we identified strong balancing selection of the centromeric KIR region and co-evolution between KIR2DL1, KIR2DL2/3 and HLA-C. Upon this background of strong balancing selection we have also identified signatures of directional selection on the telomeric region genes of the KIR locus. The telomeric region has a much lower diversity than occurs in non-African populations, due to the low frequency of the telomeric B motifs (14%) and a corresponding increase in the frequency of the telomeric A motif. This bias is consistent with pressure from infectious disease [28] being stronger than that from reproductive disorders [29]. For example, KIR2DS1, a component of telomeric B and thus infrequent in the Ga-Adangbe (Figure 1A) is the major KIR factor that protects against pre-eclampsia in European populations [46].
Although the two gene families are on different chromosomes, low-resolution analysis showed that KIR and their HLA ligands have evolved in concert across populations worldwide [48], [76]. Here, using high-resolution analysis of a well-defined population having substantial genetic diversity, we identified an on-going molecular co-evolution. That the analysis only identified functionally interacting components of known ligand-receptor pairs demonstrates the correlations are due to natural selection and not chance [77]. We also identified the differential action of natural selection on the motifs of HLA class I molecules that interact with lymphocyte receptors. Diversification of peptide binding has been the major outcome of balancing selection on all three HLA class I molecules and has continued throughout hominid evolution to the present day. Through the same time period the TCR-interacting motifs have been evolving under selective neutrality, consistent with T-cell diversity being generated by somatic, not heritable, mutation [20], [21]. Contrasting both of these patterns we detected on-going balancing selection of the KIR-contacting motif of HLA-B, and this selection was strongest in the Ga-Adangbe. Whereas varying selection pressures have resulted in a high number of different peptide binding motifs, selection on the KIR-interacting motif (Figure 4) and its co-evolution with KIR3DL1 (Figure 7) are likely driven by the two extreme phenotypes of receptor ligation or no ligation. This suggests these phenotypes each provide both an advantage and a potential cost to the host. This mode of balancing selection is strikingly similar to the deleterious mutants of haemoglobin that provide resistance to Plasmodium falciparum malaria but also impair erythrocyte function [78].
Illustrating the binary nature of balancing selection at the KIR-interacting motif of HLA-B are two common Ga-Adangbe allotypes that differ only at residues 77–83. HLA-B*53:01 has the Bw4 motif and is therefore a ligand for KIR3DL1 and HLA-B*35:01 does not have the motif. HLA-B*53:01 originated in West Africa as the product of a gene conversion between HLA-B*35 and a second, unknown allele [66]. That it remains localized to West Africa [13] and combines high prevalence with low haplotype diversity is consistent with HLA-B*53:01 having risen rapidly in frequency due to natural selection likely in response to pressure exerted by P. falciparum [79] . Both B*35 and B*53 can elicit CTL responses to this pathogen through distinct but overlapping peptide repertoires [79]. Thus, the capacity of HLA-B*53:01 to also interact with NK cells may contribute to its observed protective effects, whilst parasite strain-specific differences could contribute to its detrimental effects. Supporting this interpretation are the high incidence of malaria caused by P. falciparum in the Ga-Adangbe population [80], its impact on human health and genomes [81]–[83] and associations with combined KIR and HLA genotypes [39]. Moreover, there is no other single pathogen in West Africa that carries such a high pre-reproductive mortality as malaria [33].
In examining the sites on HLA class I that interact with different types of lymphocyte receptor we found that diversity in the LILRB1 binding site on the α3 domain of HLA-A, -B and -C is enhanced through balancing selection. We also identified co-evolution of KIR with HLA class I and also of the LILRB1 interaction with HLA class I. Supporting these results are functional data showing that the LILRB1-contacting residues and the peptide binding motif influence KIR3DL1 binding to HLA-B [18], [27], [32], [65], [72]. Thus, mutations within the LILRB1-binding motif could affect KIR ligation indirectly through their influence on HLA class I structure [65] or aggregation of receptor/ligand complexes [26]. In parallel, diversity in the LILRB1 contact site on HLA class I could serve to thwart viruses, such as cytomegalovirus (CMV), that evolve mimics of HLA class I to protect virus-infected cells from NK cell attack [84]. Any collateral loss of HLA recognition by LILRB1 will be limited through presence of multiple functionally-related receptors, such as other LILR, KIR or CD94/NKG2 molecules [2], [9], [24], [85]. Pointing to the selection pressure exerted by CMV are its impact on individual NK repertoires, prevalence in African populations, and the risk of mortality associated with perinatal transmission of the virus [41], [86], [87].
The research we report here was conducted with approval from the Stanford University School of Medicine Institutional Review Board and the Ghanaian Ministry of Health.
The population we studied were residents of Prampram, a coastal fishing village of 7,000 inhabitants situated 50 km east of Accra and south of the Volta Basin in the Greater Accra region of Ghana. Malaria (98% Plasmodium falciparum) is endemic in Prampram, with a mean of 8.5 infectious bites/person/year [80]. In the course of a study to determine the patterns of malaria infection in children, samples of genomic DNA were obtained from 131 newborn infants and from 104 of their mothers [80]. The subjects are from the Ga-Adangbe ethnic group, which currently comprises 2 million individuals in total. Archaeological data and accompanying historical accounts, combined with linguistic and genetic evidence indicate that Ga-Adangbe ancestors first lived in the region of present-day Nigeria or Burkina Faso before the Bantu expansion (∼3000 years ago) and then migrated to the Volta Basin 750–1000 years ago [88], [89]. The Ga-Adangbe speak a Kwa language of the non-Bantu Niger Kordofanian family. Analysis of autosomal genetic markers indicates that the Ga-Adangbe are closely related to the Akan, also from Ghana [60]. The Akan and other closely-related Ghanaian populations, such as the Ashanti, have similar composition of both mitochondrial and Y-chromosome haplogroups, supporting the demographic model that the Ga-Adangbe derive from a population that lived in West Africa prior to the Bantu migration [90], [91].
Nucleotide sequences were determined for the exons of KIR genes from 16 Ga-Adangbe children who were chosen at random to represent the study population. The sequences of newly discovered alleles were confirmed by re-amplification, cloning and sequencing; or by direct sequencing of the PCR products obtained from homozygous and/or hemizygous individuals. When possible, new alleles were also confirmed by amplification and sequencing of the same gene from the mother. From this dataset of Ga-Adangbe KIR sequences, we developed a pyrosequencing-based method for KIR genotyping that distinguishes all known variants, including those detected in the 16 randomly selected children (Figure S1 and Figure S2). Pyrosequencing provides a semi-quantitative measure of SNP genotypes (the peak-height ratio) that determines both allele identity and copy-number genotype [57]. We further exploited this feature to genotype combinations of KIR genes having exons that are difficult to distinguish using standard genotyping technology. In this manner KIR2DL1 and KIR2DS1, which are different genes with high sequence similarity, were genotyped together, as were KIR2DL2/3 and KIR2DS2. Similar criteria were used to distinguish exons 1 and 2 of KIR2DL5 from those of the related KIR3DP1 pseudogene. KIR2DS3 and KIR2DS5, which are relatively uncommon in the Ga-Adangbe population, were subjected to standard Sanger sequencing in addition to pyrosequencing. The combined method targets 304 coding-region SNPs, of which 190 are non-synonymous, to discriminate 350 KIR alleles (247 KIR allotypes). Following allele-specific genotyping, 20 individuals were chosen either at random, or because of their unusual pyrosequencing patterns, and the nucleotide sequences of their KIR exons determined by standard sequencing. Pyrosequencing reactions were performed using PyroGold reagents and a PSQ HS 96A machine (Qiagen, Valencia, CA).
KIR gene content was confirmed by results from bead-based sequence-specific oligonucleotide probe hybridization (SSOP), which tests for the presence of 13 KIR genes (KIR2DL1-5, KIR2DS1-5 and KIR3DL1-3). The assay was performed using LABType reagents (One Lambda, Canoga Park, CA with KIR lot #4) and detected using a Luminex-100 instrument (Luminex corp. Austin, TX).
The cohort of 235 Ga-Adangbe individuals was genotyped for HLA-A -B and -C at allele-level resolution using bead-based SSOP hybridization that was detected with a Luminex-100 instrument (Luminex corp. Austin, TX). The assays were performed using lots #11 (HLA-A), #14 (HLA-B), and #9 (HLA-C) of LABType SSO reagents (One Lambda, Canoga Park, CA). To identify variants that are common in the Ga-Adangbe but not detected by the probes, we further investigated all individuals who typed homozygous for HLA-A, -B, or -C by sequencing their putative homozygous genes.
PCRs were performed using a Perkin-Elmer 9600 thermal cycler (or a Veriti 96-Well instrument using 9600 emulation mode) with a three minute denaturing step at 94°C, 10 cycles of (94°C 10 s; 65°C 60 s) and 20 cycles of (94°C 10 s, 61°C 50 s, 72°C 30 s). Standard DNA sequencing reactions were performed in forward and reverse directions using BigDye Terminator v3.1 and analyzed using an ABI-3730 sequencer (ABI, Foster City CA). When required, PCR products were cloned using Topo-pcr2.1 vector (Invitrogen, Carlsbad CA) and sequenced using M13 and internal primers. All of the newly-discovered alleles described herein were validated according to the guidelines recommended by the curators of the Immuno Polymorphism Database (IPD) [12]. At least five clones of the desired allele were sequenced from each individual examined. Newly identified allele sequences were submitted to Genbank and the IPD database with accession numbers indicated below and in Figure S2.
KIR genes and alleles were named by the KIR nomenclature committee [92] formed from the WHO Nomenclature Committee for factors of the HLA system, and the HUGO Genome Nomenclature Committee. A curated database is available at http://www.ebi.ac.uk/ipd/kir/ [12]. <D> denotes the number of Ig-like Domains, <L> a Long, inhibitory, cytoplasmic tail <S> a Short, activating, tail and <P> a Pseudogene. A unique DNA sequence that spans a KIR coding region is considered an allele and those that yield unique proteins are considered to define an allotype. The first three digits distinguish the allotypes, the fourth and fifth digits distinguish synonymous variation. To give an example: KIR3DL1*01501 and KIR3DL1*01502 are synonymous variants of the KIR3DL1*015 allele, and encode the KIR3DL1*015 allotype – an inhibitory receptor having three Ig-like domains.
KIR haplotypes are named according to the criteria described by Pyo et al. [93]. KIR haplotypes are divided into centromeric (c) and telomeric (t) regions, or segments, that are of two forms: A and B. The two letters in the haplotype nomenclature define the four types of segment: cA, cB, tA and tB. Following these letters are two digits that uniquely define the different gene-content motifs for each type of segment: for example cA01 and cA02. Following these designations of gene-content motif are two sets of three digits that are separated by colons and distinguish motifs having identical gene content but differing by one or more allelic polymorphisms. The first set of three digits denotes differences that include non-synonymous variation, whereas the second three digits denote differences that are only synonymous or non-coding.
The high heterozygosity observed for each KIR and HLA class I gene in the Ga-Adangbe, coupled with analysis of mother-child pairs, allowed unambiguous deduction of HLA class I and KIR allele-level haplotypes. Core sets of 208 HLA and 208 KIR haplotypes were deduced by segregation analysis in 104 mother-child pairs. These sets of haplotypes were used as priors in PHASE 2.1 [94] analyses which deduced 54 HLA class I and KIR haplotypes from the remaining 27 unrelated individuals. The final data set consisted of 262 independent HLA class I and KIR haplotypes.
Population statistics were calculated from the set of 131 children (2N = 262). For some analyses, in which we estimated the total KIR and HLA diversity in the Ga-Adangbe population, total numbers of 366 independently segregating HLA class I and KIR haplotypes were used (262 haplotypes from the set of unrelated children, plus 104 non-segregating maternal haplotypes (2N = 366)).
The distributions of HLA-A, -B and -C alleles were compared in 108 populations, including the Ga-Adangbe, for which high-resolution genotyping data were available. These comprised 103 of the 497 populations studied by Solberg et al. [13], of which 11 are sub-Saharan Africans, and four additional sub-Saharan populations: Ugandans from Kampala [95], Yorubas from Ibadan in Nigeria [96], KhoeSan from Southern Africa and Hadza from Tanzania [75]. Data from a total of 31,298 individuals were used in the analyses described here.
Statistica 10 (StatSoft Inc. Tulsa OK) was used to perform principal component analysis on the frequencies of every HLA-A, -B and -C allele present in four or more of the 108 populations (242 alleles: 70 A, 129 B, 43 C). Population clustering analysis, performed using STRUCTURE 2.3.3 [97], was restricted to populations where information for each individual was available. The analysis was performed assuming the model of correlated allele frequencies among ancestral clusters, with a 1,000 step burn-in stage, 10,000 step run stage and 5 replicates. The influence of linkage disequilibrium (LD) between markers was reduced by including only HLA-A and -B, which are separated by ∼1.4 Mb.
For comparison of gene-content diversity of centromeric (cen) and telomeric (tel) region KIR haplotypes across worldwide populations, haplotype frequencies were obtained from population studies that discriminated 2DL5cen (KIR2DL5B) from 2DL5tel (KIR2DL5A) and for which the data are available from allelefrequencies.net [74]. There were 72 populations satisfying these criteria with a mean N of 105 individuals per population.
Tajima's D measures the impact on allele-frequency spectra of directional selection favoring a single allele (D<0), or balancing selection favoring multiple alleles (D>0) [98]. Tajima's D was calculated using DnaSP 4.1 [99]. Statistical significance was assessed by comparing the observed values with those expected under neutral-drift equilibrium, in a range of demographic models generated using the program ms [100]. When evidence remains significant under all reasonable demographic models, the allele distributions are unlikely to have arisen through neutral genetic drift. The demographic models were as described previously [35].
Watterson's homozygosity F test provided the first evidence that balancing selection was acting on HLA molecules [101]. The statistic, which is the proportion of homozygotes expected under Hardy-Weinberg equilibrium, was calculated from the frequencies of allotypes for given HLA class I motifs using the exact test described by Slatkin [102] and implemented in the Pypop software package [103]. The reported p-value is the probability of obtaining an F statistic less than the observed value if the motif was evolving under neutrality. It is based on the null distribution of F values simulated under neutrality/equilibrium conditions and on the observed number of alleles (k) of any given motif and sample size (2N). In order to directly compare the magnitude of deviation from neutral expectations for motifs with differing numbers of alleles, we computed the normalized deviate of the homozygosity statistic (Fnd). Fnd is the difference between the observed homozygosity, divided by the square root of the variance of the expected homozygosity. This calculation is implemented in Pypop, with variance values obtained through simulations [63]. Significant negative values of Fnd indicate balancing selection, while significant positive values of Fnd indicate directional selection.
PAML 4.5 [104] was used to identify codons subject to positive diversifying selection. Neighbour-joining (NJ) and Bayesian phylogenetic analyses to provide input for PAML were performed as described previously [35] using Mega 5 [105] and MrBayes 3.2.1 [106]. The MHC-C data set used corresponded to release 2.21 of the IPD database [12] which included 340 alleles unique through exons 2 and 3 (α1 and α2 domains) of HLA-C, plus all unique chimpanzee and orangutan MHC-C alleles having sequences complete through these exons. Similarly for the α3 domains of MHC-A, -B and -C, all unique human, chimpanzee and orangutan exon 4 sequences were used.
Haplotypes of coding sequence were constructed by concatenating the sequences of the KIR alleles identified by pyrosequencing. A gapped alignment was used to account for gene absence and the duplicated copies of 2DL4 and 3DL1 observed in a single individual were not included. Haplotype networks were created with the Hamming distance model using the haploNet function of Pegas 0.4-3 [107]. The node probability was calculated according to Templeton et al. [108] using Pegas 0.4-3. Mismatch distributions were calculated with p-dist and pairwise deletion using Mega 5 [105].
For all the populations described as West African by Tishkoff et al. [83] and having N>20, heterozygosity was calculated for each non-GATA microsatellite. The percentile range was then calculated from these 6659 data points. Heterozygosity was calculated using Nei's unbiased estimator [109].
Distance matrices (p-distance; number of SNPs which differ, divided by number of SNPs) between individuals in the study cohort (N = 131) were calculated from SNP genotypes using the ‘dist.gene’ function in the ‘ape’ (Analyses of Phylogenetics and Evolution: ver. 3.0-6 [110]), package for the R language for statistical computing [111]. Mantel's permutation test for similarity of matrices [70] was implemented for pairwise combinations of distance matrices using the ‘mantel.test’ function of ‘ape’. The function compares the observed value of the z statistic for correlation to a distribution obtained by permuting the rows and columns of data. 10,000 permutations were performed. The SNPs were phased and haplotypes concatenated prior to analysis. In the first round single polymorphic HLA residues were compared with complete KIR genotypes; those showing significant correlation were then tested against single KIR residues. From the LD (r2) values, groups of residues in linkage disequilibrium that contribute to the correlation between genotypes were then identified. Further iterations allowed the identification of single residues and groups of residues having the highest correlation between HLA class I and KIR.
EU272647 (KIR3DL2*029), EU272648 (KIR3DL2*00302), EU272652 (KIR3DL2*049), EU272654 (KIR3DL2*032), EU272657 (KIR3DL2*023), EU272660 (KIR3DL2*024), FJ666320 (KIR3DL2*035), FJ666322 (KIR3DL2*037), FJ666323 (KIR3DL2*038), FJ666325 (KIR3DL2*040), FJ883770 (KIR3DL3*032), FJ883771 (KIR3DL3*033), FJ883772 (KIR3DL3*01406), FJ883773 (KIR3DL3*00903), FJ883774 (KIR3DL3*00208), FJ883775 (KIR3DL3*01502), FJ883776 (KIR3DL3*02502), FJ883777 (KIR3DL3*01602), FJ883778 (KIR3DL3*034), FJ883780 (KIR3DL3*035), GQ478175 (KIR3DL3*02702), GQ906701 (KIR2DL4*013), GU301909 (KIR2DS5*011), GU323350 (KIR2DL1*01201), GU323352 (KIR2DL1*01102), GU323351 (KIR2DL1*01202), GU323353 (KIR2DL1*020), HM211183 (KIR2DL3*018), HM211184 (KIR2DL3*01202), HM211185 (KIR2DL2*011), HM211186 (KIR2DL2*00602), HM235772 (KIR3DL3*056), HM358895 (KIR2DS3*006), JX523641/HM358896 (KIR2DS5*00502), HM602023 (KIR2DL5B*017), HM602024 (KIR2DS3*00106), HQ026776 (KIR2DS5*009), HQ191481 (KIR3DL3*02703), HQ191482 (KIR3DL3*049), HQ609602 (KIR2DP1var1), HQ609603 (KIR2DP1var2), HQ609604 (KIR2DP1var3), HQ609605 (KIR2DP1var4), HQ609606 (KIR2DP1var5), HQ609607 (KIR2DP1var6), JX523632 (KIR2DL4*023), JX523633 (KIR2DL4_19b). Seven KIR3DL1/S1 alleles from this population were reported previously [35].
|
10.1371/journal.pntd.0006705 | Complement receptor 1 (CR1, CD35) association with susceptibility to leprosy | Pathophysiological mechanisms are still incompletely understood for leprosy, an urgent public health issue in Brazil. Complement receptor 1 (CR1) binds complement fragments C3b/C4b deposited on mycobacteria, mediating its entrance in macrophages. We investigated CR1 polymorphisms, gene expression and soluble CR1 levels in a case-control study with Brazilian leprosy patients, aiming to understand the role of this receptor in differential susceptibility to the disease.
Nine polymorphisms were haplotyped by multiplex PCR-SSP in 213 leprosy patients (47% multibacillary) and 297 controls. mRNA levels were measured by qPCR and sCR1 by ELISA, in up to 80 samples.
Individuals with the most common recombinant haplotype harboring rs3849266*T in intron 21 and rs3737002*T in exon 26 (encoding p.1408Met of the York Yka+ antigen), presented twice higher susceptibility to leprosy (OR = 2.43, p = 0.017). Paucibacillary patients with these variants presented lower sCR1 levels, thus reducing the anti-inflammatory response (p = 0.040 and p = 0.046, respectively). Furthermore, the most ancient haplotype increased susceptibility to the multibacillary clinical form (OR = 3.04, p = 0.01) and presented the intronic rs12034383*G allele, which was associated with higher gene expression (p = 0.043), probably increasing internalization of the parasite. Furthermore, there was an inverse correlation between the levels of sCR1 and mannose-binding lectin (initiator molecule of the lectin pathway of complement, recognized by CR1) (R = -0.52, p = 0.007).
The results lead us to suggest a regulatory role for CR1 polymorphisms on mRNA and sCR1 levels, with haplotype-specific effects increasing susceptibility to leprosy, probably by enhancing parasite phagocytosis and inflammation.
| The reasons for which some individuals resist Mycobacteria leprae infection, whereas others contract leprosy and only a subgroup of them become severely affected, are still poorly understood. The complement receptor 1 (CR1) serves as a gate for bacterial entry in macrophages, but its importance in the spread of infection and emergence of symptoms is unknown. Despite having many common structural and regulatory variants, the CR1 gene was investigated only once in a leprosy association study in Malawi. In order to fill in this gap, we investigated if CR1 polymorphisms are co-responsible for differential disease susceptibility in 213 leprosy patients and 297 controls, also measuring mRNA and soluble CR1 levels. Associations were dependent on specific combinations of variants in regulatory and coding regions, which were also associated with gene and protein expression. Thus, this study corroborates the importance of the CR1 receptor in the susceptibility to leprosy and is the first to bring information about CR1 polymorphisms in the Brazilian population, as well as to show the relationship between genotypes and mRNA and sCR1 levels.
| Leprosy has been reported for millennia in many ancient cultures, being currently common and causing great social stigma for affected individuals in India, Brazil, Indonesia, Bangladesh, Democratic Republic of Congo, Ethiopia, Nepal and Nigeria [1–3]. India, Brazil and Indonesia report more than 10 000 new leprosy patients, annually. Together, they figure up 81% of the newly diagnosed and reported patients, globally [4]. In Brazil, the state with the highest prevalence is Mato Grosso (9.03 cases / 10.000 inhabitants) [5].
Leprosy is a chronic infectious disease caused by Mycobacterium leprae and M. lepromatosis [6] that primarily affects the skin and peripheral nervous system, later reaching other organs and systems [7,8]. After being exposed to the pathogen, most of the individuals present resistance to infection. Those who develop the disease can fit within a broad clinical spectrum with two antagonistic poles, from paucibacillary tuberculoid to multibacillary lepromatous disease [9]. This clinical diversity relies on the quality of the immune response, which itself results from genetic variants of the host and environmental factors [10]. Genome-Wide Association Studies (GWAS), all done with the Chinese population, identified polymorphisms in interleukin 23 receptor (IL23R), nucleotide-binding oligomerization domain containing 2 (NOD2) and human leukocyte antigens–antigen D related (HLA-DR) genes as associated with susceptibility to the disease [11–14]. In addition, polymorphisms of genes associated with Parkinson (Parkin—PARK2, parkin coregulated—PACRG and superoxide dismutase 2 -SOD2) and Alzheimer diseases (SOD2) modulate susceptibility to leprosy in independent populations [15,16].
The complement system includes 47 proteins and protein fragments that activate, coordinate and regulate a proinflammatory, proteolytic cascade of the immune response [17]. Altered expression and functional deficiency of complement components modulate susceptibility to diseases. Genetic variants that reduce complement activation or complement-mediated recognition of opsonized elements may enhance resistance against pathogens that depend on phagocytosis to initiate infection, as mycobacteria. They may also decrease inflammation at skin and nerve injuries [18,19]. Indeed, several polymorphisms of genes of the complement system are associated with leprosy: mannose-binding lectin (MBL2), ficolins (FCN1, FCN2 and FCN3), the serine protease associated with them (MASP2—mannose-binding lectin serine protease 2) and complement receptor 1 (CR1, also known as CD35) [10,20–24].
The entrance of mycobacteria into macrophages is mediated by complement receptors such as CR1, which binds C3b/C4b complement fragments deposited on opsonized bacteria [25]. CR1 is mainly expressed by erythrocytes, phagocytes, B and T cells, in membrane-bound or soluble form [26]. It accelerates the decay of C3 and C5 convertases and serves as a cofactor in factor I–mediated cleavage of C3b/C4b into ligands for other complement receptors [27]. In addition, CR1 competes with MASPs for the same binding sites on collectins and ficolins, which initiate the lectin pathway of complement. These molecules possibly act as opsonins, leading to CR1-mediated internalization of pathogens into phagocytes [28]. In malaria, CR1 generates rosettes between Plasmodium falciparum infected and noninfected erythrocytes and mediates sialic acid–independent cell invasion [29,30]. Despite this multitude of functions, genetic association studies of infectious diseases with CR1 single nucleotide polymorphisms (SNPs) are scarce [20,31], and only one investigated this gene in leprosy, reporting a protective association with the McCoyb allele (rs17047660) in a rural region of Malawi [20]. In this work, we aim to fill in this gap by investigating CR1 polymorphisms, mRNA expression levels and sCR1 serum levels in Brazilian leprosy patients.
This transversal case-control study was approved by The Ethics Committee of the Federal University of Paraná (Clinical Hospital) (protocols 218.104 and 279.970). All study participants were informed about the research and signed a term of informed consent.
A total of 213 leprosy patients were recruited from three reference centers: Dermatology Service at the Clinical University Hospital–(UFPR), Regional Center for Metropolitan Specialties (both in Curitiba–PR, South Brazil) and Reference Center for Leprosy and Tuberculosis (Sinop–MT, Central-West Brazil), within a period of 10 years (2005–2015). Data for both patients and controls was collected through interviews and medical records. Cases were ascertained by the responsible dermatologists of each hospital/reference center. Patients were classified as either paucibacillary (PB) or multibacillary (MB), according to clinical presentation and bacilloscopy results, following recommendations of the World Health Organization (WHO). Individuals with confirmed clinical diagnosis of leprosy and older than 18 years were included. Patients with skin conditions other than leprosy (either autoimmune or infectious) were excluded. As controls, we included 297 healthy volunteers and blood donors from the same reference centers or nearby blood banks, that may share patient’s environmental factors, including exposure to the parasite. They were from Paraná Hematology Center–HEMEPAR (South Brazil), Reference Center for Leprosy and Tuberculosis (Sinop–MT) and Adventist Hospital of Pemphigus (Campo Grande–MS) (both in Central-Western Brazil). Individuals older than 18 years and without any pathological skin condition were included. In addition, all participants were classified according to ethnic origin, based on physical characteristics and self-reported ancestry, into Euro and Afro-Brazilian. This strategy has been confirmed by HLA genotyping, where an average sub-Saharan African component of 9% and an average Amerindian component of 5% was identified for the first, and at least 40% of African and 6% of Amerindian ancestry for the last [32,33]. Demographic characteristics of the participants of this study are listed in Table 1.
Blood was collected with and without anticoagulant ethylenediaminetetraacetic acid and DNA extracted from peripheral blood mononuclear cells through commercial kits (Qiagen, Hilden, Germany and GFX Genomic Blood DNA Purification Kit, GE Healthcare, São Paulo, Brazil). For samples from Campo Grande, MS, we used the phenol-chloroform-isoamyl alcohol protocol [34].
We genotyped nine SNPs: rs6656401 (NC_000001.11:g.207518704A>G in intron 4); rs3849266 (NC_000001.11:g.207579645C>T in intron 21); rs2274567 (NC_000001.11:g.207580276A>G in exon 22, exchanging histidine by arginine—NP_000564.2:p.His1208Arg); rs3737002 (NC_000001.11:g.207587428C>T in exon 26, exchanging threonine by methionine—NP_000564.2:p.Thr1408Met); rs11118131 (NC_000001.11:g.207587851C>T in intron 26); rs11118167 (NC_000001.11:g.207608809T>C in intron 28); rs17047660 (NC_000001.11:g.207609511A>G in exon 29, exchanging lysine by glutamic acid—NP_000564.2:p.Lys1590Glu); rs4844610 (NC_000001.11:g.207629207A>C in intron 37); rs12034383 (NC_000001.11:g.207630250G>A in intron 37) (Fig 1). They were selected by: association with any disease, being a tag SNP with r2 ≥ 0.8 in European (Utah–USA), Mexican, Colombian or Yoruba populations of the 1000 Genomes Project [35]; and/or with minor allele frequency (MAF) > 0.05. Noncoding SNPs were also chosen for investigation, due to their possible regulatory effect on gene expression.
We developed two multiplex Polymerase Chain Reaction—Single Specific Primer (PCR-SSP) reactions for simultaneous identification of four SNPs and a simple PCR-SSP for an isolated SNP. Each reaction coamplified a non-specific monomorphic fragment, for quality control. Primer sequences are shown in Table 2. All reactions were carried out in a final volume of 8 μl, containing 20 ng of genomic DNA, 0.2 Mm of each dNTP, 1x Coral Buffer (Qiagen, Hilden, DE). Thermal cycling began with 94°C for 5 min; followed by 33 (simple PCR-SSP and multiplex PCR-SSP 2) or 35 (multiplex PCR-SSP 1) cycles, where each cycle began with 94°C for 20s and ended with 72°C for 40s. We evaluated the amplified fragments after electrophoretic run on 1% agarose gel, stained with Sybrsafe (Invitrogen Life Technologies, Carlsbad, CA) (S1 Fig).
For the simple PCR-SSP, where we discriminated the rs6656401 alleles, annealing temperatures were 58°C for the initial 11 cycles, 55°C for the following 11 cycles and 52°C for the last 11 cycles. We used 0.3 μM of each SSP and 0.1 μM of each control primer, 1.5 mM MgCl2 and 0.04 units of Taq polymerase (Invitrogen Life Technologies, Carlsbad, CA).
In the PCR-SSP multiplex 1, annealing temperatures were 59°C for the initial 8 cycles, 57°C for the following 7 cycles, 55°C for another 10 cycles and 53°C for the last 10 cycles. We used 0.5 μM of SSPs for rs3737002 and rs11118131, 0.6 μM for rs11118167 and rs17047660 and 0.08 μM of each control primer, 1.5 mM MgCl2 and 0.2 units of Platinum Taq DNA polymerase (Invitrogen Life Technologies, Carlsbad, CA).
In the PCR-SSP multiplex 2, annealing temperatures were 63 °C for the initial 6 cycles, 61°C for the following 16 cycles, 59°C for the last 11 cycles. We used 0.18 μM of SSPs for rs3849266 and rs2274567, 0.4 μM for rs4844610 and rs12034383, 0.08 μM of each control primer, 2 mM MgCl2 and 0.3 units of Taq polymerase (Invitrogen Life Technologies, Carlsbad, CA).
All protocols are available in protocols.io in the following dx.doi.org/10.17504/protocols.io.p49dqz6; dx.doi.org/10.17504/protocols.io.p5sdq6e; dx.doi.org/10.17504/protocols.io.p44dqyw.
We collected blood from 33 controls and 46 leprosy patients (87% of which presented with the paucibacillary clinical form), all from Sinop, using tubes of the PAXgene Blood RNA system and extracted total stabilized RNA with the PAXgene Blood RNA Kit (both from PreAnalytiX, QIAGEN / BD). The RNA was reverse transcribed into cDNA with the High Capacity cDNA Reverse Transcription Kit (Applied Biosystems). mRNA levels were measured by quantitative real-time TaqMan PCR in ViiA 7 Real-Time PCR System (Applied Biosystems) and the TaqMan inventoried assay Hs00559348_m1, with a probe specific for the exon-exon junction encoding the CR1 domain most proximal to the cell membrane, common to the most abundant CR1 mRNA transcripts. All assays were conducted in duplicate and the relative mRNA levels were normalized to mRNA expression of the beta-glucuronidase gene (GUSB) using the TaqMan inventoried assay 4333767F. Cq values (threshold cycle) were calculated using the ViiA 7 Software v1.2 (Applied Biosystems, USA), and gene expression with the comparative Cq method 2-ΔΔCq [36].
We measured sCR1 levels in serum of 58 leprosy patients (30 multibacillary and 28 paucibacillary) and 22 healthy controls from both Curitiba and Sinop, with similar ethnic distribution and presenting alleles associated with the disease. sCR1 levels were quantified by ELISA using SEB123Hu kit (USCN Life Science Inc., Wuhan, China), according to the manufacturer’s instructions. The sCR1 levels were correlated with previously measured MBL levels in 30 of the same patients, all from Curitiba [21].
Allele, genotype and haplotype frequencies were obtained by direct counting. We calculated the sample size needed for detecting associations with allele/haplotype frequencies of at least 10% with 95% confidence level and a confidence interval of 5.0, arriving at minimal 384 chromosomes (at least 192 individuals). Thus, our work has enough statistical power for detecting an association with common alleles/haplotypes and susceptibility to leprosy per se, although not necessarily with clinical leprosy forms, since multibacillary and paucibacillary leprosy groups encompassed less individuals (100 and 91, respectively). Distribution of polymorphisms and haplotypes between patients and controls, multibacillary and paucibacillary patients, were compared using exact Fisher test (for haplotype frequencies) and binary multivariate logistic regression (for the dominant and recessive models) with STATA v.9.2 (Statacorps, Lakeway Drive, TX), correcting if necessary, for age, sex, geographic origin and ethnic group distribution. To correct for false discovery rate (q value), we used the Benjamini and Hochberg [37] approach on all significant results. The hypothesis of Hardy-Weinberg equilibrium was evaluated with the exact test of Guo & Thompson, implemented in Arlequin v.5.1. Linkage disequilibrium was evaluated with Haploview 4.2 [38]. Extended haplotypes were manually reconstructed based on linkage disequilibrium and phase information (obtained by the PCR-SSP amplification). Quantitative variables (CR1 mRNA and soluble CR1 levels) were not normally distributed and thus compared using nonparametric methods. They were grouped into patients and controls, multi- and paucibacillary patients, to check if levels alter according to disease status and severity. They were also grouped according to genotypes, to see if polymorphisms influence gene and protein expression. Data were transformed into log10 for better graphical visualization. Normality tests (D’Agostino & Pearson and Shapiro-Wilk test), correlation tests (Spearman) and non-parametric comparisons between medians of CR1 mRNA and sCR1 levels (with Mann-Whitney and Kruskal-Wallis tests) were done using GraphPad Prism v.6 Software (GraphPad Software, La Jolla, CA).
Genotype distribution of controls followed the predictions of Hardy-Weinberg equilibrium. We were able to reconstruct haplotypes based on phase information obtained with SSP amplification and found 18 haplotypes, half of which possibly are recombinants. We confirmed the haplotypes by evaluating linkage disequilibrium (LD) in Euro-and Afro-Brazilians and named them according to the phylogenetic nomenclature suggested by others [39], adapted for recombinant haplotypes [40] (Fig 2, S2 Fig). The most ancestral *1 haplotype is identical to the sequence found in Pan troglodytes (NM_001193675). Haplotype *2 and three recombinant haplotypes (*2.3B2B, *3A2A.2 and *3B2B.2) are unique in that they harbor the G allele in exon 29, encoding p.1590Glu, responsible for the McCoy McCb blood group antigen (rs17047660). On the other hand, all haplotypes of clade *3A and one recombinant (*3A2A.2) encode p.1208Arg, due to the G allele in exon 22 (rs2274567). Yet clade *3B harbors all haplotypes with the A allele in intron 37 (rs12034383). They are probably involved in the generation of most recombinant haplotypes, reaching together a frequency higher than 40% (Table 3). For example, the York Yka+ blood group antigen (corresponding to the T allele in exon 26 or p.1408Met), is solely encoded by one haplotype of this clade (*3B2B), but also by six recombinant haplotypes (*1.3B2B, 1.3B2B.1, *1.3B2B.3A2B.3B1, *3B2B.1, *3B2B.2 and *3B2B.3A2B.3B1). The *4 haplotype seems to be a relatively recent in cis combination, occurring on an isolated branch and harboring Alzheimer’s disease risk alleles rs6656401*A and rs4844610*A [41].
The tree was rooted on the haplotype presenting high sequence identity with Pan troglodytes (NM_001193675). In the phylogenetic nomenclature system, the first clades to diverge are numbered with Arabic numerals. Sublineages of each clade are subsequently designated with capital letters and individual present-day haplotypes are given Arabic numerals, following the schema numerals/letters/numerals, if they diverge further. Recombinants are named according to the most common inferred parental haplotypes, separated by a dot, and citing the most similar parental haplotype, first. Bootstrap values are given on the respective branches. Amino acids are given where nucleotide substitution caused a missense mutation. Dendrogram was constructed following the maximum parsimony method using Mega Software v.6.
Among the recombinant haplotypes, the frequency of those sharing the initial GTHMC combination of the *3B2B haplotype—*3B2B.3A2B.3B1, *3B2B.1 and *3B2B.2 (where the last “HM” mean wild type p.1208Hist and the York Yka+ p.1408Met amino acids), was higher in the patient group, than controls (33/426 or 7.7% vs. 18/594 or 3%, OR = 2.69 [95%CI = 1.49–4.84], Fisher’s exact test p = 0.0011, q = 0.0125). Individuals with the most common of them, the recombinant *3B2B.3A2B.3B1 haplotype, presented twice higher susceptibility to leprosy per se, than controls, independent of age, gender and ethnicity (OR = 2.43, p = 0.017, q = 0.0375). Among leprosy patients, those with the *1 haplotype had an increased susceptibility to the more severe multibacillary clinical forms (OR = 3.04, p = 0.01, q = 0.025) (Table 3).
As expected, we found no correlation between mRNA and sCR1 levels, since sCR1 is the result of enzymatic protein cleavage. Nevertheless, both were associated with intronic SNPs. Patients that carry the rs12034383*G allele presented higher mRNA expression than A/A homozygotes (p = 0.043) (Fig 3A). Furthermore, paucibacillary patients with the rs3849266 and rs3737002 T allele presented a reduction in sCR1 levels, compared with C/C (p = 0.040 and p = 0.046, respectively) (Fig 3B). Finally, there was a negative correlation between sCR1 and MBL serum concentrations in the patient group (Fig 4). Neither CR1 mRNA expression nor sCR1 concentration differed between patients and controls, or between clinical types (S1 and S2 Tables).
More than two decades ago, complement receptor CR1 was identified as a mediator for Mycobacterium leprae’s entry into phagocytes [25]. Despite the importance that CR1 polymorphisms may have in the establishment of infection, only one association study on five missense mutations in this gene, has been conducted [20]. Although the genome-wide association study done with the Chinese population did not reveal any association of leprosy with CR1, it should be noted that this population has a markedly different genetic background from other populations, such as Brazilians [11–14]. Furthermore, as occurs with other genes encoding erythrocyte proteins, the CR1 polymorphism has been modified by natural selection for malaria resistance in populations of endemic regions [29,41]. The present study is actually the first attempt to specifically analyse the influence of CR1 haplotypes, mRNA expression and sCR1 levels in the susceptibility to leprosy.
We found an association of recombinant haplotypes sharing rs3849266*T and rs3737002*p.1408Met with two-fold higher susceptibility to leprosy. The rs3849266 SNP is located in a regulatory region bound by CTCF (CCCTC binding factor), as well as nine other regulatory proteins [42,43]. CTCF separates chromatin domains, regulating transcription [44]. Mutations in the DNA sequence recognized by this protein have been associated with cancer, most probably due to an interference with epigenetic regulation [44]. Yet rs3737002 is a missense variant (p.Thr1408Met) with a high Polyphen score of damage (1.00), configuring the York blood group antigen [45]. Paucibacillary patients with the rs3849266*T and p.1408Met alleles presented a reduction in sCR1 levels. It is possible that the York antigen alters protein conformation and hampers C-terminal cleavage of the CR1 transmembrane region [46] (Fig 5). The reason for which the leprosy susceptibility effect is restricted to recombinant haplotypes, not including the original *3B2B haplotype, leads us to suggest that the association relies on the combined effects of all other alleles configuring this haplotype, including hitch-hiking linked SNPs, not investigated in this study. We did also not find the protective association against leprosy, previously reported in a rural area of northern Malawi with the McCoy p.1590E blood group antigen [20]. Discrepant results are most probably due to ethnic differences, since our population was mainly composed of Euro-Brazilians (more than 70%).
The *1 haplotype, most similar to the ancestral CR1 sequence, increased susceptibility to multibacillary disease. Multibacillary disease is known to be associated with a genetic susceptibility to build a Th2 immune response against mycobacterial spread, whereas susceptibility to leprosy per se is expected to be associated simply with genetic susceptibility to mycobacterial challenge (based especially on the quality of the innate immune response). Thus, it should not be surprising to find different CR1 haplotypes associated with susceptibility to both conditions: CR1*1 with the first and CR1*3B2B.3A2B.3B1, with the second.
The *1 haplotype presents rs12034383*G in intron 37, which was associated with higher mRNA expression and is predicted to reduce the affinity for BAF155, a subunit of the SWI/SNF complex [42] (Fig 5). This complex regulates transcription by altering nucleosomal structure, using ATP hydrolysis [47]. Homozygotes for the rs12034383*A allele may present higher chromatin condensation than carriers of the G allele, reducing CR1 gene expression. The A allele was indeed associated with lower CR1 gene expression in different tissues of healthy individuals (https://www.gtexportal.org/home/snp/rs12034383). Even so, the expression results in this study were obtained with a majority of paucibacillary patients. They shall thus be cautiously interpreted, especially regarding multibacillary leprosy. The A allele occurs in the most frequent haplotypes, belonging to the phylogenetic clade *3B. It is associated with about 20% increased amyloid Aß cerebrospinal fluid levels in homozygote individuals, increasing susceptibility to Alzheimer’s disease [48]. It was also associated with higher erythrocyte sedimentation rate, a measure of ongoing inflammation [49].
Finally, higher sCR1 levels correlated negatively with MBL levels, previously measured [21]. Thus, it may be speculated that not only low MBL levels, but also increased levels of the inhibitory sCR1 protein, may be the products of an immune response regulation, that contribute to reduce inflammation and complement activation of the lectin pathway, at least among leprosy patients. This anti-inflammatory effect can be highly beneficial, protecting against lepromatous disease. Since sample size was very small, this hypothesis shall be further tested in other settings.
In conclusion, we suggest that CR1 polymorphisms and haplotypes enhance susceptibility to leprosy by modulating gene expression and sCR1 abundance, to increase inflammation and parasite phagocytosis. Functional investigations on phagocytic efficiency of the associated haplotypes would help to clarify the role of this receptor in the disease and possibly lead to novel therapeutic strategies.
NCBI: 1378, MIM: *120620, ENSEMBL: ENSG00000203710, hprd: 00398, UniProtKB: P17927.
|
10.1371/journal.pgen.1006543 | Protein Phosphatase 1 Down Regulates ZYG-1 Levels to Limit Centriole Duplication | In humans perturbations of centriole number are associated with tumorigenesis and microcephaly, therefore appropriate regulation of centriole duplication is critical. The C. elegans homolog of Plk4, ZYG-1, is required for centriole duplication, but our understanding of how ZYG-1 levels are regulated remains incomplete. We have identified the two PP1 orthologs, GSP-1 and GSP-2, and their regulators I-2SZY-2 and SDS-22 as key regulators of ZYG-1 protein levels. We find that down-regulation of PP1 activity either directly, or by mutation of szy-2 or sds-22 can rescue the loss of centriole duplication associated with a zyg-1 hypomorphic allele. Suppression is achieved through an increase in ZYG-1 levels, and our data indicate that PP1 normally regulates ZYG-1 through a post-translational mechanism. While moderate inhibition of PP1 activity can restore centriole duplication to a zyg-1 mutant, strong inhibition of PP1 in a wild-type background leads to centriole amplification via the production of more than one daughter centriole. Our results thus define a new pathway that limits the number of daughter centrioles produced each cycle.
| The centrosomes are responsible for organizing the mitotic spindle a microtubule-based structure that centers, then segregates, the chromosomes during cell division. When a cell divides it normally possesses two centrosomes, allowing it to build a bipolar spindle and accurately segregate the chromosomes to two daughter cells. Appropriate control of centrosome number is therefore crucial to maintaining genome stability. Centrosome number is largely controlled by their regulated duplication. In particular, the protein Plk4, which is essential for duplication, must be strictly limited as an overabundance leads to excess centrosome duplication. We have identified protein phosphatase 1 as a critical regulator of the C. elegans Plk4 homolog (known as ZYG-1). When protein phosphatase 1 is down-regulated, ZYG-1 levels increase leading to centrosome amplification. Thus our work identifies a novel mechanism that limits centrosome duplication.
| In mitotic cells the centrosome serves as the primary microtubule-organizing center and consists of two centrioles surrounded by a proteinaceous pericentriolar matrix (PCM). During mitosis the centrosomes organize the poles of the spindle, therefore maintaining appropriate centrosome numbers promotes spindle bipolarity and faithful chromosome segregation. Regulated centrosome duplication is the primary mechanism by which centrosome number is controlled, and involves building a new daughter centriole adjacent to each pre-existing mother centriole. Two features of centriole duplication maintain appropriate centrosome numbers: first, centriole duplication is limited to occurring only once per cell cycle. Second, only a single daughter centriole is assembled in association with each pre-existing mother centriole.
A conserved set of five centriole duplications factors, SPD-2/CEP192, ZYG-1/Plk4, SAS-6, SAS-5/STIL/Ana2, and SAS-4/CPAP are required for daughter centriole assembly and their individual loss results in centriole duplication failure (reviewed in [1]). Conversely, individual over expression of a subset of these duplication factors, Plk4, SAS-6 and STIL/SAS-5, leads to centriole over-duplication (the production of more than a single daughter) leading to a condition known as centriole amplification (the accumulation of an excess number of centrioles) [2–6]. Interestingly, the three factors whose overexpression leads to centriole amplification have been identified as key players in the initial steps of centriole duplication. In human cells Plk4 phosphorylates STIL to trigger centriolar recruitment of SAS-6, which initiates formation of the cartwheel, the central scaffolding structure of the new centriole [7–12]. Similarly in C. elegans the Plk4 homolog ZYG-1 recruits a complex of SAS-5 and SAS-6 through direct physical association with SAS-6 to initiate centriole duplication [13,14].
Because Plk4 overexpression causes the formation of extra daughter centrioles, Plk4 protein levels must be tightly regulated in vivo. One mechanism that regulates Plk4 levels is its SCF-mediated degradation promoted by autophosphorylation, whereby the active kinase induces its own destruction [15–17]. Degradation of Plk4 homologs in Drosophila, (plk4/Sak) and C. elegans (ZYG-1) are similarly regulated by their SCF-mediated targeting for degradation [18,19]. In addition, recent studies have shed light on temporal and spatial regulation of Plk4 levels. Plk4 initially localizes in a broad ring around the mother centriole until, coincident with the initiation of duplication, it becomes restricted to a small focus marking the location of daughter centriole assembly [8,20,21]. Emerging evidence suggests that this transition, which seems to be a key step in ensuring only a single daughter centriole is assembled, relies on spatially regulated Plk4 degradation [8,22]. Because STIL can both activate Plk4 [10,22,23], and also protect it from degradation [22] it is proposed that centriolar recruitment of a small focus of STIL at the G1/S transition triggers broad Plk4 activation and degradation via autophosphorylation, while protecting a local focus. Thus STIL limits the centriolar distribution of Plk4, promoting the assembly of a single daughter centriole. These studies reveal the central role that regulated destruction of ZYG-1/Plk4 plays in controlling centriole number, and highlight the importance of better understanding how the stability of Plk4 is regulated.
Protein phosphatase 1 (PP1) is a major cellular phosphatase that plays well-characterized roles in diverse processes including glycogen metabolism, circadian rhythms and cell division. Humans possess a single PP1α gene, a single PP1β gene (also known as PP1δ) and a single PP1γ gene. The different isoforms are >85% identical and show largely overlapping roles, although some functional specialization has been identified [24]. C. elegans has four PP1 catalytic subunits: a single broadly-expressed PP1β homolog, GSP-1, and three PP1α homologs, GSP-2, which is also widely expressed, and GSP-3 and GSP-4, whose expression is limited to the male germ line [25,26]. The core catalytic subunit of PP1 can directly bind a subset of substrates, but functional specificity is largely conferred by its interaction with regulatory proteins. Over 200 PP1 interactors exist, which regulate PP1 through modulating substrate specificity, enzyme localization or by inhibition or activation of phosphatase activity [27]. Two evolutionarily conserved PP1 regulators that play a role in cell division are inhibitor 2 (I-2) and SDS22. I-2 was originally identified as an inhibitor of PP1 and shows potent inhibitory activity in vitro, although interestingly the yeast homolog of I-2, GLC8, can also stimulate PP1 activity [28–30]. Down regulation of I-2 in Drosophila or human cells leads to chromosome mis-segregation, which is proposed to result from mis-regulation of Aurora B [31,32]. Similarly SDS22 antagonizes Aurora B autophosphorylation, downregulating Aurora B kinase activity [33]. Although it has been noted that PP1α regulates centrosome cohesion, [34,35], no role for PP1 in regulating centrosome duplication has previously been found.
Here we report a novel PP1-dependent pathway that plays a critical role in ensuring that each mother centriole produces one and only one daughter centriole during each cell cycle. We show that PP1 together with two of the most conserved PP1 regulators, I-2 and SDS-22 downregulates ZYG-1 protein abundance. Our data indicate that PP1 regulates ZYG-1 levels post-translationally and we demonstrate that loss of PP1-mediated regulation leads to ZYG-1 overexpression and centriole amplification through the production of multiple daughter centrioles.
ZYG-1 is essential for centriole assembly: when hermaphrodites carrying the temperature sensitive zyg-1(it25) mutation are grown at the non-permissive temperature of 24°C centriole duplication fails, resulting in 100% embryonic lethality. To identify additional regulators of centriole assembly, we screened for suppressors of the zyg-1(it25) phenotype and isolated mutations in 20 szy (suppressor of zyg-1) genes that rescue embryonic survival [36]. We mapped one of these mutations, szy-2(bs4), to the Y32H12A.4 locus on chromosome III, hereafter referred to as szy-2. The szy-2(bs4) mutation is a single base pair change (G to A) in a splice donor site, which is predicted to alter splicing of the szy-2 transcript resulting in a frame-shift (Fig 1A). Analysis of the SZY-2 sequence shows that it is homologous to inhibitor-2 (I-2), a conserved regulator of protein phosphatase 1 [30]. Using antibodies raised against I-2SZY-2 we confirmed that the szy-2(bs4) mutation significantly reduces I-2SZY-2 protein levels, although residual I-2SZY-2 is still detected, indicating that at least some message is properly spliced in the mutant (Fig 1B). At 24°C the embryonic viability of zyg-1(it25); szy-2(bs4) double mutant embryos is 83%, significantly higher than zyg-1(it25) (0% viability). To determine if suppression of the embryonic lethal phenotype was associated with restoration of centriole duplication, we imaged zyg-1(it25) and zyg-1(it25); szy-2(bs4) embryos expressing GFP::tubulin and mCherry::histone (Fig 1C). Normally at fertilization the sperm delivers a single pair of centrioles to the acentrosomal egg, and during the first cell cycle the centrioles separate, duplicate and form the poles of the mitotic spindle (Fig 1C (top panel), S1A Fig and S1 Movie). Subsequent duplication events ensure the formation of bipolar spindles in later divisions. When the first round of centriole duplication fails, as in zyg-1(it25) mutants, the two sperm-derived centrioles still separate and organize the poles of the first mitotic spindle leading to a normal first division. However as each daughter cell inherits a single centriole, monopolar spindles assemble in the 2-cell embryo (Fig 1C (middle panel), S1B Fig and S2 Movie). As this phenotype is indicative of centriole duplication failure, we scored the presence of monopolar spindles in 2-cell stage embryos and found that >80% of centrioles duplicated during the first round of duplication in zyg-1(it25);szy-2(bs4) double mutant embryos (Fig 1C (bottom panel) and 1D & S3 Movie). In contrast, this first duplication event always failed in zyg-1(it25) control embryos (Fig 1C (middle panel) & D). We confirmed that this effect is specific by using RNAi to deplete SZY-2 in zyg-1(it25) worms; RNAi of szy-2 but not the non-essential gene smd-1 (control RNAi), restored centriole duplication to the zyg-1(it25) mutant (Fig 1D). Moreover the deletion allele szy-2(tm3972), in which the C-terminal 190 residues are removed, was also able to suppress the zyg-1(it25) phenotype (Fig 1D). Together these data demonstrate that reducing I-2SZY-2 function suppresses zyg-1(it25) and suggest that modulating PP1 activity may impact centriole duplication.
Because szy-2 encodes the C. elegans homolog of I-2, this suggested that the szy-2(bs4) allele may alter PP1 activity, resulting in the observed suppression of centriole duplication failure. I-2 is conserved from yeast through humans and has been described as both an inhibitor and an activator of PP1 [28–30]. Analyses in vitro have shown that C. elegans I-2SZY-2 can inhibit rabbit PP1 activity, but the in vivo role of szy-2, in particular in relation to endogenous worm PP1 subunits, has not previously been investigated [30]. We sought to determine whether the important function of I-2SZY-2 with respect to zyg-1(it25) suppression was as an inhibitor or an activator of PP1. To determine whether PP1 activity is up- or down-regulated in the szy-2(bs4) mutants we wanted to monitor the phosphorylation levels of a known PP1 substrate. One such substrate is histone H3, which is phosphorylated in the early stages of mitosis; this modification is removed by PP1 [37,38]. To investigate whether I-2SZY-2 is an activator or inhibitor of PP1 we analyzed phospho-histone levels in szy-2(bs4) mutant embryos. We used a phospho-histone H3 antibody to detect phosphorylated histone H3 levels in embryos and found them to be greatly elevated in szy-2(bs4) when compared with the wild type (Fig 1E & 1F), a result reminiscent of PP1 depletion (see below; [25,39]). Phospho-histone levels were elevated in both mixed stage embryos (Fig 1E) as well as during the metaphase stage of mitosis (Fig 1F). This result indicates that the szy-2(bs4) mutation reduces PP1 activity and suggests that I-2SZY-2 normally acts as an activator of PP1. Consistent with I-2SZY-2 being a positive regulator of PP1, szy-2(bs4) mutant embryos exhibit a chromosome mis-segregation defect [36] that is similar to that of embryos depleted of PP1 (S4 Movie). There is precedent for I-2 acting as an activator of PP1: in yeast I-2GLC8 binds and inactivates PP1GLC7, but upon phosphorylation of I-2GLC8, PP1 becomes activated [29]. The phosphorylated residue is conserved in all I-2 homologs including C. elegans, although it is not clear whether I-2SZY-2 is similarly regulated in the worm [30].
Because we had identified a mutation in I-2szy-2 as a suppressor of zyg-1(it25), we speculated that loss of additional PP1 regulatory proteins may have a similar effect. To test this hypothesis we identified seven C. elegans proteins which show a high degree of conservation with human PP1 regulators. We then used RNAi to deplete each factor in zyg-1(it25) worms, and monitored embryonic viability (S2 Fig). RNAi of only the sds-22 gene significantly increased embryonic viability of the zyg-1(it25) embryos. SDS-22 is a leucine-rich repeat (LRR) domain containing protein that localizes PP1 during mitosis, regulating mitotic progression [33,40–42]. Although sds-22(RNAi) had only a modest effect on zyg-1(it25) embryonic viability, direct observation of centriole duplication in zyg-1(it25); sds-22(RNAi) embryos, revealed 70% of centriole duplication events occurred normally (Fig 2B). The disparity between the strength of suppression of embryonic lethality and the strength of suppression of centriole duplication failure is likely explained by a combination of two factors. First, we find that on average, 30% of the duplication events in zyg-1(it25); sds-22(RNAi) embryos fail; such a high degree of cell division failure during early development would result in embryonic lethality. Second, since sds-22 is an essential gene, RNAi knockdown may cause embryonic lethality even though centriole duplication is normal. Since reducing SDS-22 function is able to suppress the zyg-1(it25) phenotype we hypothesized that one of the suppressors of zyg-1(it25) recovered in our genetic screen [36] may have a mutation in sds-22. We therefore sequenced the sds-22 coding region in those szy mutants that mapped close to the sds-22 genetic locus. The szy-6(bs9) strain contained a G-to-A missense mutation in the coding region of sds-22 which results in a single amino-acid substitution (G224S) within the eighth LRR (Fig 2A). To confirm that bs9 was an allele of sds-22 we performed a complementation test with sds-22(tm5187) a deletion allele that exhibits a fully-penetrant larval-lethal phenotype (Fig 2A & 2C). While the szy-6(bs9) homozygotes exhibited minimal embryonic lethality, the trans-heterozygotes (bs9/ tm5187) exhibited 100 percent embryonic lethality confirming that bs9 and tm5187 are allelic (Fig 2C). We have therefore renamed the szy-6(bs9) allele sds-22(bs9). Although by itself, the sds-22(bs9) mutation does not appear to affect centriole duplication, zyg-1(it25); sds-22(bs9) double mutant embryos exhibit an approximately 50% success rate of centriole duplication, comparable to levels seen after SDS-22 RNAi depletion (Fig 2B). Together these data show that reducing the function of either of two PP1 regulatory proteins, I-2SZY-2 or SDS-22, is able to partially suppress the failure of centriole duplication caused by the zyg-1(it25) mutation.
Our data suggest that I-2SZY-2 is an activator of PP1, and work in vertebrate cells has shown that the SDS-22 ortholog positively regulates PP1 activity [33]. Because the szy-2(bs4) and sds-22(bs9) mutations rescue zyg-1(it25) phenotypes it followed that directly lowering PP1 activity should also suppress the zyg-1(it25) phenotype. Embryos express two PP1 catalytic subunits encoded by the genes gsp-1 and gsp-2. Similar to what is seen in the szy-2(bs4) mutant (Fig 1E & 1F), co-depletion of GSP-1 and GSP-2 by RNAi resulted in an increase in phospho-histone H3 staining of mitotic chromosomes (Fig 3A). We next tested whether individually reducing the activity of either PP1βGSP-1 or PP1αGSP-2, could suppress the centriole duplication failure seen in zyg1(it25) embryos. Although the gsp-1(tm4378) deletion allele is sterile, precluding its use in our analysis, we were able to specifically deplete PP1βGSP-1 by RNAi (Fig 3B). When PP1βGSP-1 was depleted in zyg-1(it25) embryos we saw a robust restoration of centriole duplication (90% of centrioles; Fig 3B, 3C, 3D & 3G). In contrast, the putative null allele gsp-2(tm301) [25] did not suppress the zyg-1(it25) phenotype and we never observed centriole duplication in the zyg-1(it25); gsp-2(tm301) double mutant embryos (Fig 3E, 3F & 3G). The C. elegans PP1αGSP-2 and PP1βGSP-1 isoforms show a high degree of identity and are both expressed in the early embryo [25]. Since depletion of only the PP1βGSP-1 isoform can suppress the zyg-1(it25) allele this suggests either 1) a divergence in function of the two enzymes, with only PP1βGSP-1 being involved in the regulation of centriole duplication, or 2) that, despite partial redundancy between the two enzymes, zyg-1(it25) mutants provides a sensitized background where loss of PP1βGSP-1 alone is sufficient to subvert normal controls. Indeed, two observations support the later possibility. First, we have only observed defects in chromosome segregation after co-depletion of both PP1βGSP-1 and PP1αGSP-2 suggestive of some functional redundancy (S4 Movie). Second, we show below that co-depletion of PP1βGSP-1and PP1αGSP-2 disrupts the normal pattern of centriole duplication whereas single depletions of either phosphatase have no effect (see below & S5 Movie).
Our data indicate that individually reducing the function of either I-2SZY-2, SDS-22 or PP1βGSP-1 can partially suppress the zyg-1(it25) centrosome duplication failure. Using GFP fusions we were, however, unable to detect centrosome localization of any of the proteins (S3 Fig). Nevertheless we reasoned that all three proteins work in a common process and sought to determine whether they physically interact. When we immunoprecipitated I-2SZY-2 from C. elegans embryonic extracts, we found by immunoblotting that we could detect co-precipitated PP1βGSP-1 and reciprocally when we immunoprecipitated PP1βGSP-1 we were able to pull down I-2SZY-2 (Fig 3H), confirming that C. elegans I-2SZY-2 is a PP1 binding protein in vivo. Similarly, antibodies against PP1βGSP-1 co-precipitated SDS-22 (Fig 3H). However, we were unable to detect any interaction between I-2SZY-2 and SDS-22, suggesting that a complex simultaneously containing all three proteins does not form. Additional IP-western experiments demonstrated that I-2SZY-2 also interacts with PP1αGSP-2, but did not reveal an interaction between SDS-22 and PP1αGSP-2 (Fig 3I). To confirm and extend these results we also individually immunoprecipitated I-2SZY-2, SDS-22, and PP1βGSP-1 and analyzed the immunoprecipitated material by mass spectrometry (S4 Fig). Using this approach we found that I-2SZY-2 and SDS-22 could independently interact with both PP1βGSP-1 and PP1αGSP-2 but not with each other (S4 Fig). We conclude that components of the PP1-dependent regulatory pathway physically interact but that I-2SZY-2 and SDS-22 do not appear to reside in the same complex.
Finally, we have found that although both I-2SZY-2 and SDS-22 interact with PP1 and positively regulate PP1’s function in controlling centriole duplication, the two regulators do not always function together to control PP1 activity. Specifically we have found that while loss of I-2SZY-2 activity results in an increase in the level of phospho-histone H3 in mitotic chromosomes (Fig 1F), no such increase is observed upon loss of SDS-22 activity (S5B Fig). Thus I-2SZY-2 appears to function independently of SDS-22 to mediate the role of PP1 in regulating histone H3 phosphorylation. Curiously, loss of SDS-22 resulted in an elevated phospho-histone H3 signal as detected by immunoblotting (S5A Fig). Because SDS-22 does not control the chromatin content of phospho-histone H3, we speculate that loss of SDS-22 leads to a mitotic delay; this would explain the elevated signal of phospho-histone H3 in mixed-stage worm extracts.
Given that PP1 is a mitotic phosphatase that antagonizes many known mitotic kinases we wanted to investigate whether these relationships contribute to the regulation of centriole duplication. In Drosophila and human cells, I-2 is implicated in regulation of Aurora B and its depletion results in chromosome mis-segregation [31,32]. Similarly in humans, SDS22 antagonizes Aurora B autophosphorylation, downregulating Aurora B kinase activity [33]. PP1, I-2 and SDS22 therefore seem to cooperatively regulate chromosome segregation by modulating Aurora B activity. In C. elegans PP1 also regulates Aurora B, contributing to its correct localization during meiosis and to chromosome segregation in mitosis [39,43]. Although Aurora B has not been implicated in centriole duplication we wanted to determine whether szy-2(bs4) suppresses centriole duplication failure through upregulation of Aurora B activity. We therefore RNAi-depleted Aurora BAIR-2 in szy-2(bs4); zyg-1(it25) worms and monitored centriole duplication (S6A, S6B & S6C Fig). Although we observed failures in chromosome segregation and cytokinesis consistent with successful Aurora B depletion (S6C Fig;[44]), centriole duplication was unperturbed (S6A Fig), indicating that szy-2(bs4) does not regulate centriole duplication by modulating Aurora B activity.
Since PP1 and I-2 are associated with regulation of Aurora A activity [45] and Aurora A plays a conserved role in centrosome separation and maturation [46,47] we tested whether Aurora AAIR-1 activity is required for suppression of zyg-1 (it25) by szy-2(bs4). After exposing double mutants to air-1(RNAi) we observed reduced SPD-2 localization and incomplete centrosome separation, consistent with loss of Aurora A activity (S7H Fig), however we did not see a perturbation of centriole duplication (S6A, S6Fa & S6Fb Fig), suggesting that Aurora A is not required for suppression of zyg-1(it25) by szy-2(bs4). In vertebrate cells PP1 also has a recognized role in antagonizing Plk1 [48], a kinase with an established role in centriole duplication [49]. However, depletion of plk-1 in zyg-1(it25), szy-2(bs4) embryos did not affect centriole duplication (S6A Fig) even though PLK-1 activity was clearly compromised (S6D Fig). In summary our data suggest that reducing PP1 activity does not restore centriole duplication in the zyg-1(it25) mutant by relieving antagonism, and thus increasing the relative activity, of Aurora B, Plk1 or Aurora A, which are known PP1 antagonists.
During the course of our analyses we noted that the level of the coiled-coil protein SPD-2 was elevated at the centrosome in szy-2(bs4) mutant embryos. SPD-2 is a component of both the PCM and centrioles and is required for centriole duplication and PCM assembly [50,51]. We quantified the effect on SPD-2 levels in szy-2(bs4) embryos expressing SPD-2::GFP and found a 1.5-fold increase in centrosomal SPD-2 levels throughout the cell cycle (S7A Fig). When we compared levels of endogenous SPD-2 at the centrosome, we found a similar increase in centrosome-associated SPD-2 in szy-2(bs4) embryos (S7B–S7D Fig). Since SPD-2 plays a positive role in centriole duplication, we reasoned that its increased levels at the centrosome in the szy-2(bs4) mutant might be responsible for PP1-mediated suppression. To test this possibility, we utilized a codon-optimized spd-2::gfp transgene [52] to overexpress SPD-2 protein in the zyg-1(it25) mutant. Despite a large increase in centrosome-localized SPD-2 (S7E & S7F Fig) we did not see any suppression of zyg-1(it25) embryonic lethality (S7G Fig), indicating that centriole duplication failure persisted. Thus by itself, a general elevation of the level of SPD-2 at the centrosome is not sufficient for suppression of zyg-1(it25). Conversely, we also find that the elevated level of centrosome-associated SPD-2 is not required for szy-2(bs4)-mediated suppression: Aurora AAIR-1 is required for SPD-2 recruitment to the centrosome [50], and depletion of aurora AAIR-1 in the zyg-1(it25); szy-2(bs4) double mutant drastically reduced the amount of SPD-2 at the centrosome, (S7H Fig) nevertheless, centriole duplication was not perturbed (S6A, S6E & S6F Fig). Thus, I-2SZY-2 regulates both centriole duplication and PCM assembly, but the elevated SPD-2 levels at the centrosome observed in szy-2(bs4) mutants do not constitute the primary mechanism by which szy-2(bs4) suppresses the centriole duplication defect of zyg-1(it25) mutants.
Previous work has shown that the failure of centriole duplication in zyg-1(it25) mutants can be rescued by increasing centriole-associated levels of the mutant ZYG-1 protein [19,53]. We therefore sought to determine if decreasing PP1 activity affects ZYG-1 protein levels or localization. To determine whether the szy-2(bs4) mutation affects ZYG-1 abundance, we measured total ZYG-1 levels in early mixed-stage embryos using quantitative immunoblotting. Strikingly, in comparison to the wild type, szy-2(bs4) embryos consistently possessed approximately four-fold more ZYG-1 at the restrictive temperature of 25°C (Fig 4A & 4B). We then used quantitative immunofluorescence to see if the elevated level of total ZYG-1 protein resulted in increased levels of ZYG-1 at centrioles. As expected, the overall increase in ZYG-1 levels in embryos was associated with elevated ZYG-1 at the centrioles (Fig 4C, 4D & 4E). This increase however was cell cycle stage specific with increases observed in prophase and metaphase but not anaphase, of the first cell cycle (Fig 4C). We further sought to determine whether decreasing SDS-22 or PP1 similarly increased ZYG-1 levels. Quantification of centrosome-associated ZYG-1 in prophase revealed elevated levels of ZYG-1 in the sds-22(bs9/tm5187) trans-heterozygous mutant but a minimal increase after RNAi-depletion of PP1βGSP-1 (Fig 4D & 4E). Similarly depletion of PP1αGSP-2 had little effect on ZYG-1 levels at centrioles (Fig 4D & 4E). Since we only observed chromosome segregation defects when we depleted both PP1βGSP-1 and PP1αGSP-2 (S4 Movie), we wondered whether they also acted redundantly to regulate ZYG-1. We therefore co-depleted PP1βGSP-1 and PP1αGSP-2 and found an increase in centrosome-associated ZYG-1, consistent with the existence of redundancy between the two PP1 isoforms (Fig 4D & 4E). Overall our results suggest that reducing PP1 activity leads to an elevation of both total ZYG-1 levels, and of centriole-associated ZYG-1, providing a likely mechanism for the suppression of the zyg-1(it25) centriole duplication defect. Notably, this effect, of reduced PP1 activity leading to increased ZYG-1 levels, seems to be specific as we did not see a similar increase in SAS-6 levels in szy-2(bs4) embryos (S8A, S8B & S8C Fig). Finally, because ZYG-1 has also been implicated in positively regulating centrosome size [53], elevated ZYG-1 levels may also account for the observed increase in centrosomal SPD-2.
We next sought to determine how ZYG-1 levels are regulated by PP1. First, we measured zyg-1 transcript levels by quantitative RT-PCR and found that wild-type and szy-2(bs4) embryos possessed similar levels of zyg-1 mRNA (Fig 4F). The observed increase in ZYG-1 protein levels was not, therefore, due to an effect on transcription or mRNA stability. To test whether decreasing PP1 activity resulted in an effect on ZYG-1 translation efficiency, we crossed the szy-2(tm3972) allele into a strain carrying a GFP::histone reporter expressed under the control of the zyg-1 regulatory sequences (promoter and UTRs) [54]. Since this reporter contains zyg-1 regulatory sequences, but lacks the zyg-1 coding sequence, it allows us to determine whether elevated ZYG-1 protein levels stem from alterations at the level of gene expression (transcription/translation) or from post-translational controls. Quantitative fluorescence intensity measurements of GFP::histone revealed that the szy-2(tm3972) mutation did not increase expression of the reporter relative to that observed in the wild-type strain (Fig 4G & 4H; Student’s t-test p>0.1), indicating that reduced PP1 activity does not increase zyg-1 translation via the 3′-UTR. In order to determine whether PP1 might directly regulate ZYG-1 levels we tested for an interaction between the two proteins. Immunoprecipitation of I-2SZY-2, SDS-22, and PP1βGSP-1 from worm extracts, followed by western blotting or mass spectrometry failed to detect an interaction between any of the three proteins and ZYG-1. Because ZYG-1 is a low abundance protein [14] and interactions between PP1 and its substrates may be transient we cannot however rule out a direct interaction between ZYG-1 and PP1. Cumulatively, therefore, our data point to PP1 regulating ZYG-1 protein levels either directly or indirectly via a post-translational mechanism.
Elevated levels of the human and Drosophila homologs of ZYG-1 are associated with centriole amplification [2,5,6]. Although centriole amplification has not previously been observed in C. elegans embryos, given the increases in ZYG-1 levels we observed in the szy-2(bs4) and sds-22(bs9/tm5187) mutants we were intrigued whether this would be sufficient to subvert the normal regulation of centriole duplication, resulting in supernumerary centrosomes. We therefore performed time-lapse confocal microscopy of embryos expressing GFP::SPD-2 and mCherry::histone. Initial inspection of szy-2 and sds-22 mutants did not reveal any obvious defects during the first two cell cycles; centriole duplication proceeded normally and bipolar spindles assembled in all cells. Unexpectedly however, in embryos strongly impaired for either I-2SZY-2 or SDS-22 function, extra centrosomes were frequently observed at the four-cell stage (Fig 5A and 5B & S6 Movie). Closer inspection of the movies indicated that the extra centrosomes arose approximately synchronously from the spindle poles as the PCM dispersed near the completion of the second mitotic divisions (S6 Movie & S7 Movie). In a single case (out of 79 analyzed) we were not able to trace the origin of one of these centrosomes to a spindle pole. Thus, it is possible that very infrequently in this mutant a centriole arises spontaneously in the cytoplasm via a de novo pathway. The production of extra centrosomes was most prevalent in sds-22(bs9/tm5187) trans-heterozygotes where nearly 40% of the spindle poles gave rise to more than two centrosomes (Fig 5B). Although less frequent, an identical defect was observed in szy-2(tm3972) embryos where supernumerary centrosomes arose with the same timing and spatial pattern as those observed in the sds-22 mutant (Fig 5B). These results suggest that extra centrioles form during the second round of centriole duplication (which occurs in the 2-cell embryo) such that the extra centrioles only become apparent by confocal microscopy when mother and daughter centrioles separate as cells enter the third cell cycle (S1C Fig).
We next followed the fate of 79 centrosomes in sds-22(bs9/tm5187) trans-heterozygotes and found that all eventually accumulated PCM and participated in spindle assembly, often giving rise to multipolar spindles (S7 Movie). A minority (4/79 centrosomes), however, exhibited a one-cell-cycle delay before accumulating normal levels of PCM and participating in spindle assembly (S9 Fig and S7 Movie). This is somewhat reminiscent of the reversible “inactivation” of centrioles observed in Drosophila embryos following over-expression of Plk4 [55]. The production of excess centrosomes continued in the ensuing cell cycles. We quantified the rate of over-duplication following the third centriole cycle of sds-22(bs9/tm5187) embryos and found a similar frequency of over-duplication (35%, n = 96 centrosomes). Thus our results suggest that over-duplication begins during the second centriole duplication event and continues through embryonic development.
To confirm that the excess centrosomes observed in szy-2 and sds-22 mutant embryos arose due to a reduction of PP1 activity, we also followed centrosomes during the first several cell cycles of embryos depleted for one or both PP1 catalytic subunits. Surprisingly, neither depletion of PP1βGSP-1 (n = 10 embryos) nor of PP1αGSP-2 (n = 11 embryos) resulted in the appearance of extra centrosomes during the first three cell cycles (Fig 5E). We therefore co-depleted both catalytic subunits (Fig 5C) and found that excess centrosomes appeared at the four-cell stage in 4/10 embryos (Fig 5D and 5E & S5 Movie). Further, these extra centrosomes could function as microtubule-organizing centers and directed the formation of tripolar spindles (S8 Movie). Interestingly, we also observed occasional anaphase chromatin bridges in these embryos (S4 Movie), whereas in singly depleted embryos chromosomes always segregated normally. We conclude that PP1βGSP-1 and PP1αGSP-2 exhibit some level of functional redundancy in their roles in chromosome segregation and centriole duplication.
To further investigate the origin of the excess centrosomes we analyzed sds-22(bs9/tm5187) embryos by structured illumination microscopy (SIM), which has proven an effective means to observe centriole arrangement and number within the centrosome [56,57]. SIM imaging of SAS-4-stained embryos allowed us to resolve the basic structure of worm centrioles and we could clearly detect the normal arrangement of mother and daughter centrioles (Fig 5Fa). Strikingly, in sds-22(bs9/tm5187) embryos we were also able to find mother centrioles bearing more than one daughter (Fig 5Fb). We first detected extra daughter centrioles at the late 2-cell stage, consistent with the appearance of excess centrosomes at the four-cell stage. Furthermore, the appearance of extra daughter centrioles in association with a single mother is reminiscent of what is seen after Plk4 overexpression [2] and indicates that the excess centrosomes we observe in the 4-cell embryo originate from the formation of extra daughters during centriole duplication. Quantification of our SIM data reveals that while wild-type embryos never exceed two centrioles per centrosome, 18% of sds-22(bs9/tm5187) centrosomes contain more than 2 centrioles (Fig 5G). We did not observe any other unusual centriole configurations such as mother-daughter-granddaughter arrangements indicative of reduplication of centrioles during a single cell cycle. In summary, our data indicate that loss of PP1 activity results in overexpression of ZYG-1 and consequently centriole amplification. Amplification appears to be largely, if not entirely, driven by the production of multiple daughter centrioles. However, we cannot rule out that centriole reduplication and de novo formation also contribute to the excess centrioles observed after PP1 inhibition. In conclusion, we have described a new PP1-dependent mechanism that limits centriole duplication so that each mother produces one and only one daughter per cell cycle.
We have shown that PP1 activity is an important regulator of ZYG-1 levels in the C. elegans embryo and therefore that it is a critical regulator of centriole duplication. Although PP1α has previously been implicated in the regulation of centrosome separation at the beginning of mitosis [34], this is the first indication that PP1 regulates centriole duplication. Our data indicate that PP1 inhibits centriole duplication by restraining the accumulation of ZYG-1 in the embryo either directly or indirectly (Fig 6). Previous work has shown that Plk4 levels can be regulated by SCF-mediated degradation promoted by autophosphorylation [15–17]. Similarly in C. elegans the SCF complex is involved in regulating ZYG-1 levels and depletion of SCF components leads to an increase in centrosome-associated ZYG-1 [19]. Our finding that down-regulation of PP1 activity leads to centrosome amplification implicates PP1 as an additional regulator of ZYG-1 levels.
How does PP1 regulate ZYG-1 levels? Although many PP1 targets have been identified none are known regulators of centrosome duplication. A major substrate of PP1, I-2 and SDS22 is Aurora B kinase and activity of the C. elegans homolog, AIR-2, is inhibited by PP1 [25,39,43]. Down regulation of AIR-2 activity in the zyg-1(it25); szy-2(bs4) double mutant however did not have any measurable effect on the level of centriole duplication suggesting that PP1 is not regulating centriole duplication by antagonizing Aurora B activity. Identical results were obtained from two additional mitotic kinases, Plk1 or Aurora A. Our data indicate that PP1 regulates ZYG-1 levels through a post-translational mechanism. In theory PP1 could regulate the recognition of ZYG-1 by the SCF, however we think that this is unlikely as our results do not match with the canonical mechanism for SCF regulation: substrate recognition by the SCF is regulated by phosphorylation, but the increased phosphorylation associated with PP1 down-regulation would be expected to increase proteosomal degradation, decreasing substrate accumulation. We, however, find an increase in ZYG-1 levels when PP1 activity is reduced. Furthermore, evidence in Drosophila and C. elegans indicates that degradation of Plk4/ZYG-1 is antagonized by PP2A [58,59]. Therefore, our favored hypothesis is that PP1 regulates ZYG-1 protein levels through an SCF-independent mechanism, perhaps by removing a stabilizing phosphorylation, however the identity of the opposing kinase remains unknown. We cannot, however, rule out the possibility that PP1 functions through the SCF pathway in a non-canonical fashion to mediate ZYG-1 degradation. Of note, a KVXF consensus site for PP1 binding [60] does exist in ZYG-1, however it is not conserved even in closely related nematode species. Nevertheless, although this is a common motif by which PP1 interacts with its substrates the presence of this motif is not an absolute requirement for PP1 binding.
Our data suggest that PP1, I-2SZY-2 and SDS-22 cooperate to regulate overall cellular levels of ZYG-1. Down regulation of PP1 activity leads to accumulation of excess ZYG-1 and in embryos expressing a wild-type version of ZYG-1, the elevated activity is sufficient to drive centriole over-duplication. This is similar to the situation in humans and Drosophila where overexpression of homologs of ZYG-1 causes centriole over-duplication [2,5,6]. In agreement with this previous work [2] the appearance of supernumerary centrosomes in C. elegans results from the formation of extra daughter centrioles in association with a single mother. Intriguingly, we find that when PP1 activity is decreased, the initial duplication event (that occurs during the first cell cycle following the female meiotic division) is unaffected, yielding bipolar spindles at the two-cell stage. However, beginning with the second centriole duplication event (that occurs at the two cell stage), centrioles commence over-duplicating resulting in multipolar spindles, abnormal divisions, and ultimately lethality.
Why is the first duplication event unaffected by the loss of PP1-mediated regulation while the ensuing duplication events go awry? One unique feature of the first duplication event is that it involves paternally-inherited (sperm) centrioles while later events involve centrioles assembled in the embryo. The sperm-derived centrioles are not, however, permanently immune to elevated ZYG-1 levels, as we have observed cases where, during the second centriole cycle, all four mother centrioles (the two original sperm-derived centrioles plus the two centrioles produced during the first duplication event) produce multiple daughters. Thus, it seems that the elevated level of ZYG-1 present in the embryo only triggers over-duplication of the sperm-derived centrioles following a one-cell-cycle delay. There are at least three possibilities to explain the pattern of over-duplication in PP1-compromised embryos. First, loss of PP1 activity might only lead to overexpression of ZYG-1 beginning with the second centriole duplication event. This however seems unlikely, as we have shown that loss of PP1 activity results in elevated ZYG-1 levels even in the first cell cycle (Fig 4C & 4D), and that it can suppress the failure of the first duplication event in zyg-1(it25) embryos (Figs 1C, 1D, 2B and 3G). Another possibility is that the critical period for exposure to elevated ZYG-1 is one cell cycle prior to the actual over-duplication event. Thus the sperm centrioles, which presumably first encounter elevated ZYG-1 in the zygote, would only over-duplicate during the second centriole cycle. This model is however inconsistent with previous work showing that when zyg-1 is overactive in the male germ line, centrioles will over-duplicate during spermatogenesis, but when these centrioles are introduced into a wild-type egg they duplicate normally [61]. Thus the prior exposure of sperm centrioles to elevated ZYG-1 activity does not program them to over-duplicate in the zygote. Finally, a third and more likely possibility is that the centrioles need to be exposed to elevated ZYG-1 for two consecutive cell cycles before they will over-duplicate. Thus sperm-derived centrioles are first exposed to elevated ZYG-1 in the zygote, but do not overduplicate until the second cell cycle, after they have been exposed to elevated ZYG-1 activity for two consecutive cell cycles. There is precedent for this type of regulation; during centriole maturation in human cells, Plk1 activity is needed over the course of two successive cell cycles in order for centrioles to fully mature and become capable of organizing PCM and serving as basal bodies [62]. Thus, this mode of regulation might be a common feature of how polo like kinases operate, at least when it comes to the control of centriole function.
Why do increased ZYG-1 levels lead to centriole overduplication in the C. elegans embryo? Normally, levels of centriole-associated Plk4 peak in mitosis before being reduced to a single focus marking the site of procentriole assembly at the G1/S transition [8,20,21]. Centrosome-associated ZYG-1 levels also peak in mitosis [63] and we speculate that elevated ZYG-1 levels prevent it from becoming restricted to a single focus during the ensuing cell cycle, leading to the assembly of multiple daughter centrioles. Ensuring that only a single ZYG-1/Plk4 focus is maintained seems to be a key regulatory step in centriole duplication, however our understanding of how this is achieved remains limited. Clearly excess ZYG-1/Plk4 is sufficient to subvert the normal controls. Since Plk4/ZYG-1 plays an important role in the recruitment of SAS-5/ana2/STIL and SAS-6 at the initiation of procentriole assembly [13,14] we envision elevated ZYG-1 levels may be required at this time. However whether continual exposure to elevated ZYG-1 through two cell cycles is required for overduplication remains an open question.
In summary, we have identified I-2SZY-2, SDS-22, and PP1 as novel regulators of centriole duplication. The involvement of PP1 in regulating centriole duplication has not previously been described, but interestingly centrosome amplification was reported after depletion of I-2 from human cells, suggesting that the function of PP1 in regulating centriole duplication may indeed be conserved [31]. We show that the key function of PP1 is in limiting the availability of ZYG-1. When PP1 activity is decreased, excessive accumulation of ZYG-1 leads to the formation of extra daughter centrioles in association with a single mother, resulting in centriole amplification. Appropriate regulation of PP1 activity is therefore crucial to maintaining correct centrosome numbers. Although the requirement for PP1 in chromosome segregation due to its function at the kinetochores is well documented, our work suggests a novel requirement for PP1 in the maintenance of genome stability by regulating centriole duplication.
All worm strains were maintained at 20°C on MYOB plates seeded with OP50. The strains used in these experiments are listed in supplemental S1 Table. To monitor the effect of the sds-22(bs9) mutation on centriole duplication sds-22(bs9) homozygotes were selected from the OC626 strain using the visible dpy marker which is closely linked to the sds-22 gene. RNAi of SZY-2 was carried out by soaking worms in dsRNA [64], all other RNAi experiments were carried out by feeding worms bacteria expressing dsRNA as previously described [65]. Briefly, bacteria containing the RNAi construct were grown overnight and seeded onto MYOB plates supplemented with 25ug/ml ampicillin and 1mM IPTG. For double RNAi a 50:50 mix of overnight cultures of the two bacterial strains was plated. Worms were placed on RNAi at the L4 stage for 28h before analysis. The sequence contained in the GSP-1 RNAi construct was evaluated using the Clone mapper tool (http://bioinformatics.lif.univ-mrs.fr/RNAiMap/index.html)[66], which confirmed likely specificity for only the gsp-1 gene. For control RNAi we used an smd-1-containing vector.
Embryonic extracts for western blots were prepared and analyzed as detailed previously [59]. Whole worm extracts for the immunoprecipitation (IP) experiments were made according to [67]. Total protein concentration was determined using the Biorad Protein Assay Dye Reagent (Bio-Rad). 1.6 mg of total protein from N2 worms was used for performing each IP. Briefly, for each IP, 30 μl of Dynabeads Protein A (Life Technologies) were incubated with 10 μg of each respective antibody at 4°C for 2 hours. Beads were washed three times in worm lysis buffer (50 mM HEPES (pH 7.4), 1 mM EGTA, 1 mM MgCl2, 100 mM KCl, 10% Glycerol, 0.05% NP-40), re-suspended in 2X Laemmli Buffer (Bio-Rad), boiled at 100°C for 2 minutes and analyzed by western blotting. The following antibodies/reagents were used in this study: polyclonal SZY-2 antibodies were raised and purified against the entire szy-2 ORF (Covance); GSP-1 antibodies were raised and purified against the peptide CQYQGMNSGRPAVGGGRPGTTAGKK (YenZym Antibodies LLC); ZYG-1 [53], phospho-histone H3 (Abcam); GSP-2 [68]; DM1A (Sigma). Mass spectrometry analysis was carried out by the NIDDK mass spec core facility.
RNA was extracted from wild-type and szy-2(bs4) embryos, DNase treated and cDNA made using a superscriptIII first strand synthesis kit (Invitrogen). Forward (ACAGTACGCGGAAGAAATGG) and reverse (CACAGCAACCATCTTTTGGA) primers were used to amplify zyg-1. Primers against ama-1 were used as a control [69]. qRT reactions used iQ SYBR green supermix (Bio-Rad) as directed by the manufacturer.
Fixation and staining of embryos was carried out as described previously [36]. The following antibodies were used at a 1/1000 dilution: DM1A (Sigma), phospho-histone H3 (Abcam), ZYG-1 [19], SPD-2 [50] and anti-SAS-4 [53]. For live and fixed imaging we used a spinning disk confocal microscope which has been described previously [61]. To determine whether PP1 affects translation of the ZYG-1 transcript we shifted worms carrying the reporter construct to 25°C as L4s and imaged embryos the next day. Intensities of chromatin GFP at first metaphase were measured using Metamorph. Levels of ZYG-1 or SPD-2 at the centrosome were determined by quantification of average pixel intensity at the centrosome. Maximal projections of the centrosome were used for quantification of fluorescence in ImageJ 1.40g and background fluorescence was subtracted. Centrosome fluorescence was normalized to controls such that control intensity is 1. For structured illumination microscopy, embryos were immuno-labeled as usual and mounted in Vectashield (Vector Laboratories, Inc.). Samples were imaged with a DeltaVision OMX4 SIM Imaging System (Applied Precision).
|
10.1371/journal.pmed.1002228 | Customised and Noncustomised Birth Weight Centiles and Prediction of Stillbirth and Infant Mortality and Morbidity: A Cohort Study of 979,912 Term Singleton Pregnancies in Scotland | There is limited evidence to support the use of customised centile charts to identify those at risk of stillbirth and infant death at term. We sought to determine birth weight thresholds at which mortality and morbidity increased and the predictive ability of noncustomised (accounting for gestational age and sex) and partially customised centiles (additionally accounting for maternal height and parity) to identify fetuses at risk.
This is a population-based linkage study of 979,912 term singleton pregnancies in Scotland, United Kingdom, between 1992 and 2010. The main exposures were noncustomised and partially customised birth weight centiles. The primary outcomes were infant death, stillbirth, overall mortality (infant and stillbirth), Apgar score <7 at 5 min, and admission to the neonatal unit. Optimal thresholds that predicted outcomes for both non- and partially customised birth weight centiles were calculated. Prediction of mortality between non- and partially customised birth weight centiles was compared using area under the receiver operator characteristic curve (AUROC) and net reclassification index (NRI).
Birth weight ≤25th centile was associated with higher risk for all mortality and morbidity outcomes. For stillbirth, low Apgar score, and neonatal unit admission, risk also increased from the 85th centile. Similar patterns and magnitude of associations were observed for both non- and partially customised birth weight centiles. Partially customised birth weight centiles did not improve the discrimination of mortality (AUROC 0.61 [95%CI 0.60, 0.62]) compared with noncustomised birth weight centiles (AUROC 0.62 [95%CI 0.60, 0.63]) and slightly underperformed in reclassifying pregnancies to different risk categories for both fatal and non-fatal adverse outcomes (NRI -0.027 [95% CI -0.039, -0.016], p < 0.001). We were unable to fully customise centile charts because we lacked data on maternal weight and ethnicity. Additional analyses in an independent UK cohort (n = 10,515) suggested that lack of data on ethnicity in this population (in which national statistics show 98% are white British) and maternal weight would have misclassified ~15% of the large-for-gestation fetuses.
At term, birth weight remains strongly associated with the risk of stillbirth and infant death and neonatal morbidity. Partial customisation does not improve prediction performance. Consideration of early term delivery or closer surveillance for those with a predicted birth weight ≤25th or ≥85th centile may reduce adverse outcomes. Replication of the analysis with fully customised centiles accounting for ethnicity is warranted.
| In developed countries, one-third of stillbirths and infant deaths occur at term.
There are multiple clinical definitions at term of what constitutes a small- or large-for-gestation fetus, with <10th centile and >90th centile commonly used. Whether these statistical thresholds can accurately identify fetuses at risk of mortality or morbidity is unknown.
Customised birth weight centiles (accounting for sex, gestation, and maternal characteristics) are increasingly being adopted by many maternity units. However, whether they can identify term fetuses at risk of death more accurately than noncustomised centiles is unknown.
We examined data on 979,912 term singleton pregnancies over a 19-y period in Scotland. With external validation of our findings on an independent UK cohort (n = 10,515).
We studied the associations of birth weight centiles (noncustomised and partially customised) with stillbirth, infant mortality, admission to the neonatal unit and Apgar score <7 at 5 min. In addition, we assessed whether partially customised centiles perform better in predicting adverse outcomes compared with noncustomised centiles. We were unable to assess fully customised centiles as we did not have data on maternal ethnicity and weight.
We found that birth weight ≤25th or ≥85th centile (both partially and noncustomised) are associated with greater risk of adverse outcomes. Partially customised centiles did not identify more fetuses at risk of death compared with noncustomised centiles.
Adverse outcomes frequently occur in term fetuses. Closer surveillance or earlier delivery of those fetuses with a predicted birth weight ≤25th or ≥85th centile may reduce adverse outcomes.
Replication of the analysis with fully customised birth weight centiles is required.
| Infants who are born at the extremes of birth weight have a higher risk of adverse perinatal outcome [1]. In developed countries, one-third of stillbirths and infant deaths occur at term [2], yet no consensus exists about what defines a small or large fetus or infant at term. A variety of methods have been used, including absolute birth weight (most commonly <2,500 g and >4,000 g or 4,500 g), or statistical thresholds outside the expected birth weight for gestational age (commonly <10th or >90th centile or, for more severe phenotypes, two standard deviations) [3–7]. Whether these thresholds optimally define the risk of perinatal mortality and morbidity at term is unknown. Furthermore, some advocate that birth weight percentiles should account for maternal characteristics known to be associated with fetal growth, such as weight, height, parity, and ethnicity. However, there is conflicting evidence whether customised charts perform better than noncustomised centiles in predicting adverse perinatal outcome [8–11] and the strength of evidence for supporting this approach, particularly for term infants, has been challenged [12,13].
The aims of this study were (1) to determine the shapes and magnitudes of the associations between birth weight centile and infant death, stillbirth, infant mortality and stillbirth combined, Apgar score <7 at 5 min, and admission to the neonatal unit and (2) to compare the accuracy of predicting mortality (infant mortality and stillbirth) using noncustomised (accounting for infant sex and gestational age at birth) birth weight centile charts versus partially customised (additionally accounting for maternal height and parity) birth weight centile charts.
The Privacy Advisory Committee of the Information Services Division (ISD) of the National Services Scotland awarded ethical approval for access to and linkage of the datasets (www.isdscotland.org). All data were nonidentifiable, and individual informed consent from participants was not required.
This study is reported as per Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) guidelines (S1 Checklist).
We linked four Scotland-wide databases: the Scottish Morbidity Record 02 (SMR02), the National Records of Scotland (NRS), the Scottish Stillbirth and Infant Death Survey (SSBIDS), and the General Registrar for Scotland’s death certificate database, to provide comprehensive obstetric, neonatal, infant, and mortality data for all women discharged from Scottish maternity hospitals. Details of the datasets and the quality assurance procedures are provided in Supporting Information (full lists of data collected are available here: http://www.adls.ac.uk/nhs-scotland/maternity-inpatient-and-day-case-smr02/?detail). The data can be accessed from the National Services Scotland (www.isdscotland.org) following approval.
We obtained SMR02 data on all infants delivered in Scotland between 1 January 1992 and 31 December 2010 inclusive, the latter equating to the most recent data available at the time of data extraction. Our analyses were restricted to singleton births, in women aged over 10 y, with a gestational age at delivery between 37 and 43 wk inclusive. We excluded infant deaths due to congenital anomalies or isoimmunisation and stillbirths due to congenital abnormalities. Death ascribed to congenital anomaly was defined as any structural or genetic defect incompatible with life or potentially treatable but causing death.
The primary outcomes were infant mortality (deaths up to 1 y of age), stillbirths (intrapartum or antepartum deaths), combined stillbirth and infant mortality, low Apgar score (<7) at 5 mins, and admission to neonatal unit (special care or neonatal intensive care unit).
The exposures were noncustomised birth weight centiles (sex and gestation specific) and partially customised birth weight centiles (sex and gestation specific and adjusted for maternal height and parity) [14]. Transformation of data to partially customised centiles was performed with the GROW centiles bulk calculator: http://www.gestation.net/GROW_documentation.pdf. Customised centile charts typically also adjust for maternal weight and ethnicity. However, this information was not available in the routine data sources employed. White British ethnicity was assumed for all pregnancies since only 2% of babies were born to women in ethnic minority groups (ISD personal communication) in Scotland over this period. Weight was assumed to be 66 Kg for all women (the default of the GROW bulk calculator and the median of UK women at the start of pregnancy). Given the lack of adjustment for maternal weight and ethnicity, we refer to the adjusted percentiles as "partially customised." Noncustomised percentiles were internally standardized for gestational age (in weeks) and sex. Gestational age has been confirmed by ultrasound in the first half of pregnancy in more than 95% of women in the UK since the early 1990s [15].
We considered that year of delivery, parity, and diabetes (either existing or gestational) might be effect modifiers of the relationships between birth weight and outcomes. Data on year of delivery and parity were obtained from SMR02. The ICD codes recorded on the SMR02 record were used to ascertain gestational diabetes: 6488 (ICD9) O244, O249 (ICD10) and pre-existing diabetes: 250, 6480 (ICD9) E10-14, O240-1, O243 (ICD10). Any associations between birth weight and fatal outcomes might differ in those who die after delivery as a result of delivery complications (anoxia, trauma, or intracranial haemorrhage [ICH]). We, therefore, repeated all main analyses with deaths related to these causes only.
In addition to the procedures and audits used to ensure high-quality data on mortality, we explored the face validity of these data further by assessing associations of known risk factors, such as a previous history of infant/fetal mortality or morbidity and smoking.
On being granted data access, all authors agreed on the statistical analysis plan prior to commencing the main analysis. The analysis of the ALPSAC cohort was undertaken in response to comments from peer reviewers. All analyses were performed using Stata (version 13, StataCorp LP, College Station, Texas) and R version 3.2.3 (R Foundation for Statistical Computing, Vienna, Austria).
Between 1 January 1992 and 31 December 2010, there were 1,062,390 deliveries in Scotland. After pre-specified exclusions, the analysis cohort included 979,912 term deliveries (Fig 1). Among these deliveries 2,765 (0.3%) resulted in death: 1,672 (0.2%) were stillborn and 1,093 (0.1%) suffered infant death. There were 12,044 (1.2%) neonates who had an Apgar score <7 at 5 mins and 62,217 (6.4%) were admitted to the neonatal unit. Table 1 shows the maternal and offspring characteristics used in our main analyses by mortality outcomes (stillbirth and infant deaths compared to infants alive at 1 y) and S1 Table shows additional characteristics used to validate the mortality data by these categories. Women who delivered a stillborn were more likely to be nulliparous, smoke, live in deprived areas, and have pre-existing diabetes. Women who delivered neonates who died within their first year of life were more likely to be younger, live in more deprived areas, smoke, and have a history of previous stillbirth or neonatal death. Stillborn infants and infant deaths had lower birth weight at the time of delivery compared with neonates who survived beyond 1 y of age. S2 Table lists the causes of infant deaths.
For all five outcomes, the patterns of risk were similar but the magnitude of risk was greater for stillbirth and overall mortality. The risk started to increase at approximately the 25th centile of noncustomised birth weight centiles and increased monotonically from that point (Fig 2). For stillbirth, low Apgar score and admission to the neonatal unit there was also some evidence of risk increasing from approximately the 85th centile. Infant mortality risk remained low in all participants who were above the 25th centile. Similar patterns and position of knots were observed for partially customised birth weight centiles for all five outcomes (S1 Fig). When we restricted the analyses to infant mortality attributed to anoxia, trauma, or intracranial haemorrhage the results were similar, though the magnitude of associations were somewhat weaker, for both noncustomised and partially customised birth weight centiles (S2 Fig). S3 and S4 Figs show that knot points and associations were similar in different categories of parity and year of delivery. However, diabetes (existing or gestational) seems to modify the associations between birth weight and fatal events (p for interaction < 0.001) and fatal events increase with increasing birth weight centile (linear association) in infants of mothers with diabetes, whereas in those without diabetes the 25th and 85th centiles knot points were still evident (S5 Fig).
Table 2 shows the risk was similar for all five outcomes for those with a birth weight ≤25th centile or ≥85th centile referent to birth weight between 25th to 85th centiles for both noncustomised and partially customised centiles. In addition, we present the risk of all five outcomes for birth weight based on the traditional thresholds (<10th centile and >90th centile) referent to birth weights between 10th and 90th centiles. The risk of all mortality and morbidity outcomes was only slightly stronger for these more restrictive thresholds than the 25th and 85th centiles inferred by our data.
Discrimination of prediction of mortality was not improved using a partially customised birth weight centile (AUROC 0.61 [0.60 to 0.62]), compared with a noncustomised birth weight centile (AUROC model 2 0.62 [0.60 to 0.63]; p < 0.0001) (S3 Table). The model with partially customised birth weight centiles as the predictor of overall mortality also slightly underperformed in reclassifying pregnancies to different risk categories (for both pregnancies with a fatal and nonfatal event) compared with the model using noncustomised birth weight centiles as the exposure (NRI -0.027 [95% CI -0.039, -0.016], p < 0.001) (Table 3).
Expediting delivery for an anticipated birth weight of ≤25th noncustomised centile rather than <10th (assuming that the intervention would have 69% effectiveness for mortality), would result in pre-emptive delivery of 159,025 additional women to prevent 377 fatal events (stillbirth and infant deaths). Hence, to prevent one death, 422 (95% CI 381 to 468) additional pregnancies below this new threshold would require to be delivered. If the equivalent ≤25th threshold was used for the partially customised centiles, intervention in an additional 463 (95% CI 417 to 516) pregnancies would be required to prevent one fatal event. Adopting the threshold of ≥85th centile for defining large for gestation babies rather than >90th centile and delivering them earlier, assuming 69% effectiveness of intervention, would require an additional 721 (95% CI 598 to 947) or 826 (95% CI 638 to 1,091) pre-emptive deliveries, respectively for noncustomised and partially customised centiles, to prevent one fatal event.
In the ALSPAC cohort partially customised centile charts that did not adjust for maternal weight accurately identified SGA in comparison with fully customised charts, with 97% of those identified as SGA with full-customisation being identified with partial (weight removed) customisation when either the 10th (S4 Table) or 25th (S5 Table) thresholds were used. Accuracy for LGA was somewhat poorer with 84% and 85% being correctly identified, with the 90th and 85th percentiles, respectively (S4 and S5 Tables). With additional change of all ethnicities to white the results did not differ from those shown (for removal of weight only) in S4 and S5 Tables.
Being born too small or too large is associated with an increased risk of mortality and morbidity [1,5,7]. Despite dramatic improvements in maternal and neonatal care, we showed that even at term (37 to 43 gestational weeks) birth weight remains strongly associated with the risk of stillbirth and infant death, low Apgar score, and admission to the neonatal unit. An increased risk of mortality and morbidity was evident at term with birth weights ≤25th and ≥85th centile irrespective of whether noncustomised or partially customised centiles were used, with similar associations observed for potentially preventable infant deaths due to anoxia, trauma, or intracranial haemorrhage. These thresholds may not apply to diabetic pregnancies, in which there is evidence of increasing mortality with greater birth weight. Given that partially customised centiles exhibited weaker associations with mortality than simpler noncustomised centiles, their increasingly wide adoption by health care providers [21] for identifying may not be appropriate for assessing risks of adverse perinatal outcome at term.
Prior to term, fetal biometry is used to detect abnormal growth trajectories, with the overall clinical management of ongoing fetal health and timing and mode of delivery guided by evidence-based guidelines [10]. This is particularly the case for preterm small-for-gestational-age (SGA) fetuses. There is less emphasis and guidance on the management of the term fetus, yet this group contributes one-third of all stillbirths and infant deaths and associated substantial neonatal morbidity. A previous population analysis suggested that there was a progressive increase in perinatal mortality from 37 wk gestation and elective induction of labour at term was associated with decreased odds of perinatal death without an associated increase in emergency caesarean section delivery [22]. There has been no widespread adoption of this policy despite randomised controlled trials also suggesting induction of labour does not increase operative delivery [23,24]. This is largely due to concerns regarding increasing caesarean section rates and that induction of labour is associated with an increased risk of neonatal admission to a special care unit [22]. Our analysis suggests that guidelines on using biometry to predict birth weight in term infants and considering early term delivery for those anticipated to have a birth weight ≤25th centile for gestation week at term may reduce overall mortality while minimising iatrogenic neonatal morbidity [25,26]. Consideration of all available information (e.g., growth trajectory) and additional testing (e.g., umbilical and uterine Doppler) may facilitate differentiation between the pathological and nonpathological cases, with interventions focusing on those identified as pathological [6]. Of importance, delivery at 37–38 wk is associated with the lowest perinatal risk of death [26]. While the DIGITAT study did not show a difference in a composite measure of adverse neonatal outcome between early term induction of labour and expectant management in those with suspected fetal growth restriction (below the 10th centile), the authors suggested intense monitoring of those who are keen on expectant management to minimise the risk of stillbirth and neonatal morbidity [27]. We have previously shown that early term delivery (37–38 wk) is associated with a small increase in special educational needs compared with delivery at 40–41 wk [28]. However, this study was in a general population and not those selected on the basis of evidence that fetal growth may be faltering, hence, it is unclear whether earlier intervention in those whose fetus is predicted to be faltering grow will lead to greater risk of long-term educational outcomes. Interestingly, the same study showed a progressive increase in the risk of special educational needs with decreasing birth weight below the 20th percentile. For infants with a birth weight <3rd percentile, the risk was doubled. Hence, interventions focused on improved care of poorly grown fetuses could potentially have longer term benefits in childhood.
Our results also support considering earlier delivery of those fetuses ≥85th centile at term, with a recent randomised controlled trial suggesting that induction of labour at 37–38+6 wk for those in the >95th centile reduces shoulder dystocia [29]. But there are no clear guidelines for preventing macrosomia and large-for-gestational-age (LGA), which is becoming increasingly common due to the increasing prevalence of obesity among women of reproductive age [30]. Randomised trials of effects on stillbirth or infant mortality are unlikely to be conducted because of the extremely large sample size that would be required and therefore for these rare but extremely important outcomes large population level data like ours are required to guide practice. For anticipated birth weights ≤25th or ≥85th centile, early delivery would be anticipated to reduce mortality and delivery related morbidity, as delaying delivery to more than 39 wk in at-risk infants has previously been shown to increase the number of stillbirths and infants affected by macrosomia [31].
The clinical utility of routine second and third trimester ultrasound for the detection of fetuses with abnormal growth and at increased risk of neonatal complications has recently been reported [6]. An additional benefit of these ultrasounds would be an estimate of birth weight using internationally validated growth charts [3], which could be adopted into a prediction model at 36 wk to guide care at term if comprehensive induction of labour for estimated birth weight beyond ≤25th or ≥85th centiles was to be avoided [3,4].
The process of customisation of birth weight centiles has an additional major difference to the conventional approach, over and above accounting for maternal characteristics. The growth reference used is the distribution of weight at a given gestational age based on ultrasonic estimation of fetal weight, rather than actual birth weight. This is important preterm, as previous studies have shown that preterm birth is associated with fetal growth restriction [32]. A consequence is that use of customised centiles markedly increases the proportion of preterm births which are defined as SGA. As preterm birth is one of the major determinants of perinatal morbidity and mortality, this may explain some of the apparent strengthening of associations following customisation described in previous studies [14]. At term, this effect is less dominant and the relative risk of stillbirth is similar irrespective of whether actual birth weight, customised or an intrauterine reference ranges used [33]. That partially customised centiles exhibit poorer performance in prediction of mortality suggests that the maternal characteristics that they adjust for are not purely physiological and may be associated with pathological impairment of fetal growth resulting in inappropriate misclassification of high-risk term fetuses. Maternal short stature, primiparity, and ethnicity are all independently associated with perinatal mortality, supporting their pathological contribution to fetal growth [34,35]. We were only able to partially customise birth weight charts as we did not have data on maternal ethnicity or weight. However, additional analyses of the independent ALSPAC cohort suggested that in a similar population (i.e., in which the vast majority are white), lack of complete ethnicity data would not have biased our results. Similarly, lack of maternal weight appeared to have very little impact on defining SGA, although the extent of misclassification was somewhat higher for LGA, but the majority of LGA (over 84%) were still correctly classified. Consistent with the lack of major bias due to potentially misclassifying ethnicity in 2% of our study, a recent multicentre international study (n = 20,486) indicated that ethnicity had a minimal effect on fetal growth in healthy women [3], though another smaller study (n = 1,737) suggested that Hispanic, Asian, and black ethnicities are associated with differential growth and up to 245 g lower median estimated fetal weight than white Europeans by 39 wk gestation [36]. However, very few women of reproductive age in Scotland are from these ethnic groups. Maternal height was imputed for 19%, but the distribution of height in the imputed database was the same as women with measured height and when we repeated analyses restricted to those women with complete data on height and all other variables the results were essentially unchanged from those presented here using imputed data (data available from authors on request).
Our study included around one million deliveries. We included all eligible deliveries in Scotland over a 19-y period, thereby avoiding selection bias. We linked four national datasets to maximise data completeness. The routine data sources are subject to quality assurance checks and perform well in terms of completeness and accuracy. The detail provided by these datasets allowed us to examine a range of different outcomes and we used robust methods to calculate the RR of the outcomes of infant mortality, stillbirth, Apgar score <7 at 5 min, and neonatal admission for birth weight at term. Measurement of birth weight was not subject to quality control measures, but represents routine clinical practice, and is not subject to the error that can occur in research assessments of birth weight taking place within the first few days (rather than immediately as in clinical practice) because of the large decrease in weight observed during the first two days of life [37]. We were unable to examine specific nonfatal complications but assumed that substantive neonatal morbidity would be associated with neonatal unit admission. We examined a range of relative risks to identify thresholds, and different populations may wish to accept lower or higher risks of mortality and morbidity to guide thresholds for intervention. In particular, we cannot assume that the thresholds that we have identified here would be appropriate for other (than white European) populations. We acknowledge that using birth weight as a proxy of estimated fetal weight (EFW) may have inflated some of the associations. However, the difference between EFW and actual birth weight has been quoted up to an average of 10% [38] and we have previously shown that universal ultrasonography performs well as a screening test for SGA (under the 10th birth weight centile) with an AUROC of ~0.9 [6]. However, a key area for future research is to identify methods which help discriminate healthy SGA and LGA infants from infants which are SGA or LGA as a consequence of pathological processes. Combinations of tests (e.g., ultrasound and biomarkers) may reduce the false positive rate and allow targeting of interventions to pregnancies with high absolute risks of adverse outcome, and this is an active area of current research.
In conclusion, term fetuses remain at substantive risk of infant death, stillbirth and neonatal morbidity. Birth weight identifies those at greatest risk, and our results support consideration of early delivery, increased surveillance or additional testing for those with an anticipated birth weight ≤25th centile or ≥85th centile (rather than the widely used 10th and 90th centiles, respectively) proposed to reduce adverse outcomes. Replication of our results in other independent large datasets is warranted. A clinical trial looking at morbidity outcomes (since the small number of mortality outcomes renders a clinical trial with primary outcome stillbirth and infant deaths practically unfeasible) will clarify whether early term intervention at the proposed thresholds of birth weight will be beneficial. In Scotland, and similar countries, there is no evidence that partially customised charts perform better at identifying those term infants at risk than noncustomised charts.
|
10.1371/journal.pntd.0001483 | Global Distribution of Outbreaks of Water-Associated Infectious Diseases | Water plays an important role in the transmission of many infectious diseases, which pose a great burden on global public health. However, the global distribution of these water-associated infectious diseases and underlying factors remain largely unexplored.
Based on the Global Infectious Disease and Epidemiology Network (GIDEON), a global database including water-associated pathogens and diseases was developed. In this study, reported outbreak events associated with corresponding water-associated infectious diseases from 1991 to 2008 were extracted from the database. The location of each reported outbreak event was identified and geocoded into a GIS database. Also collected in the GIS database included geo-referenced socio-environmental information including population density (2000), annual accumulated temperature, surface water area, and average annual precipitation. Poisson models with Bayesian inference were developed to explore the association between these socio-environmental factors and distribution of the reported outbreak events. Based on model predictions a global relative risk map was generated. A total of 1,428 reported outbreak events were retrieved from the database. The analysis suggested that outbreaks of water-associated diseases are significantly correlated with socio-environmental factors. Population density is a significant risk factor for all categories of reported outbreaks of water-associated diseases; water-related diseases (e.g., vector-borne diseases) are associated with accumulated temperature; water-washed diseases (e.g., conjunctivitis) are inversely related to surface water area; both water-borne and water-related diseases are inversely related to average annual rainfall. Based on the model predictions, “hotspots” of risks for all categories of water-associated diseases were explored.
At the global scale, water-associated infectious diseases are significantly correlated with socio-environmental factors, impacting all regions which are affected disproportionately by different categories of water-associated infectious diseases.
| Water is essential for maintaining life on Earth but can also serve as a media for many pathogenic organisms, causing a high disease burden globally. However, how the global distribution of water-associated infectious pathogens/diseases looks like and how such distribution is related to possible social and environmental factors remain largely unknown. In this study, we compiled a database on distribution, biology, and epidemiology of water-associated infectious diseases and collected data on population density, annual accumulated temperature, surface water areas, average annual precipitation, and per capita GDP at the global scale. From the database we extracted reported outbreak events from 1991 to 2008 and developed models to explore the association between the distribution of these outbreaks and social and environmental factors. A total of1,428 outbreaks had been reported and this number only reflected ‘the tip of the iceberg’ of the much bigger problem. We found that the outbreaks of water-associated infectious diseases are significantly correlated with social and environmental factors and that all regions are affected disproportionately by different categories of diseases. Relative risk maps are generated to show ‘hotspots’ of risks for different diseases. Despite certain limitations, the findings may be instrumental for future studies and prioritizing health resources.
| Although substantial advances in biomedical sciences and public health measures have facilitated control of many infectious diseases in the past century, the world has witnessed an increasing incidence and geographical expansion of emerging and re-emerging infectious diseases [1], which, together with some other old ones, remain among the leading causes of deaths and disability worldwide [2], [3]. The global environmental, ecological, and socio-economic changes have a significant impact on the distribution, emergence and re-emergence of infectious diseases and are expected to continue to influence such trend [1], [4], [5], [6], [7], [8], [9]. Some recent studies at both global and regional scales have suggested that climatic factors, human movement, and agricultural practices are important factors underlying the distribution, emergence, and re-emergence of infectious diseases [1], [6], [10].
Water is essential for maintaining life on Earth. Meanwhile, water can also serve as a media for hazardous substances and pathogenic organisms, posing substantial health threats to humans through a variety of pathways. During the past few decades, human development, population growth, extreme weather events, natural calamities, and climate change have exerted many diverse pressures on both the quality and quantity of water resources which may in turn impact conditions fostering water-associated diseases. Worldwide, water-associated infectious diseases are a major cause of morbidity and mortality [11], [12], [13]. A conservative estimate indicated that 4.0% of global deaths and 5.7% of the global disease burden (in DALYs) were attributable to a small subset of water, sanitation, and hygiene (WSH) related infectious diseases including diarrheal diseases, schistosomiasis, trachoma, ascariasis, trichuriasis, and hookworm infections [11], [14], [15]. Although unknown, the actual disease burden attributable to water-associated pathogens is expected to be much higher. A total of 1415 species of microorganisms have been reported to be pathogenic, among which approximately 348 are water-associated, causing 115 infectious diseases [5].Yet, their distribution and associated factors at the global scale remain largely unexplored.
Although the linkage between the hydrological cycle and infectious diseases has long been recognized, the underlying mechanisms shaping this relationship at global and regional scales are rarely characterized. Recent developments in hydrology and geo-spatial technology, and increasing availability of spatial socio-environmental information provide an opportunity to explore this issue. Geospatial techniques (e.g. Geographic Information System, or GIS, and spatial analytical techniques) offer a means for developing and organizing spatially explicit information. For example, the availability of information on terrestrial surface water area from the Global Lakes and Wetland Database [16], could allow the exploration of the possible relationship between the availability of terrestrial surface water and distribution of water-associated diseases at the global scale.
In this study, a comprehensive database has been developed for global water-associated infectious pathogens and diseases and socio-environmental information which have been integrated into a GIS database. The overall goal of our study is to explore the possible relationship between global distribution of water-associated infectious diseases and socio-environmental factors. In this study reported outbreaks of water-associated diseases were chosen as the study subject as they were available in the developed database and provided semiquantitative information (e.g. yes or no, and frequency of outbreaks). Our specific aims in this study were to describe the global distribution of reported outbreaks caused by water-associated infectious diseases from 1991 to 2008, to explore potential risk factors associated with spatio-temporal distributions of these outbreaks, and to develop a global risk map for these diseases.
Primary source of information on water-associated pathogens and infectious diseases for the database developed in present study was based on the Global Infectious Disease and Epidemiology Network (GIDEON), a subscription- and web-based comprehensive global infectious diseases database which provides extensive geographical and epidemiological information including outbreaks for 337 recognized infectious diseases in 231 countries and regions. Data in GIDEON are collated through a system of computer macros and dedicated source lists developed over the past 15 years. A monthly search of Medline is conducted against a list of GIDEON key words (similar to Mesh terms in PubMed), and titles/abstracts of interest are reviewed. In addition, all standard publications of WHO and CDC are scanned for relevance before they are collated and entered into GIDEON. The GIDEON infectious diseases database provides a chronological listing of all reported outbreaks of infectious diseases, which are listed by year and country, with specific location information available for the majority of reported outbreaks. For those without specific location information, original publications or reports were searched to extract the information. To assess GIDEON's completeness on the reported outbreaks, a systematic search based on PubMed, ISI Web of Knowledge, WHO and CDC reports was conducted on reported outbreaks (1991–2008) for 10 randomly chosen water-associated diseases. Search terms included names of specific pathogen(s)/disease(s) and country/region, “outbreak”, “epidemic”, and “epidemics”, respectively. Chi-square test was performed to compare results from the independent search vs. that from GIDEON – our results were largely in agreement with that from GIDEON (X2 = 591.2, P<0.001). Based on the database developed, water-associated diseases and their corresponding causal agents were systematically reviewed, together with extensive literature review for relevant environmental, biological, and epidemiologic characteristics. For each disease, the following information was included in the database we developed.
The database included the following information - grid-based global human population density (per km2) based on the 2000 global population dataset, which was developed by Socioeconomic Data and Applications Center (SEDAC) of Columbia University between 2003 and 2005, providing globally consistent and spatially explicit human population information (http://sedac.ciesin.columbia.edu/gpw/); global average accumulated temperature (degree days, with a spatial resolution of 0.5 degree) for the period between 1961 and1990 from United Nation Environmental Protection(http://www.unep.org/), which was based on the degree that the temperature rose above zero degree and the number of days in the period during which this excess was maintained [18]; surface area (km2) of water bodies including large lakes , rivers, and wetland, collected from the global lakes and wetlands database (http://www.worldwildlife.org/); the average rainfall (mm) per year for the period between 1961 and1990 from FAO (http://www.fao.org); and per capita Gross Domestic Product (GDP) which was based on each a country's GDP divided by the total number of people in the country ( http://sedac.ciesin.columbia.edu/ddc/baseline/). The scale of all information collected was converted to one-degree grid in the GIS database.
A total of 1,428 outbreak events had been reported from 1991 to 2008. Outbreaks occurred all over the world and the clusters of reported outbreaks tended to be in west Europe, central Africa, north India and Southeast Asia (Figure 2). Among the reported outbreak events, 70.9% (1,012) were associated with water-borne diseases including 32.9% (471) water-carried, 12.2% (174) water-related, 6.8% (97) water-washed, 2.9% (41) water-based, and 7.3% (104) water-dispersed. 46.7% (667) of the outbreak events were associated with emerging or reemerging pathogens, which appeared in humans for the first time or had occurred previously but were increasing in incidence or expanding into areas where they had not previously been reported [5]. It is found that 49.6% (709) of the outbreak events was caused by bacteria, 39.3% (561) by viruses, and 11.1% (158) by parasites. 6.5% (93) of the outbreak events was caused by agents that could be transmitted by direct contact, 1.1% (16) transmitted through vectors, 63.5% (907) through environmental transmission, and 28.9% (412) by zoonotic routes.
The reported outbreak events had shown a significant increase since 1991, which had been accompanied by a significant increase in the number of published articles (Figure 1, Pearson correlation - 0.935, P<0.001). We used a generalized linear model to test the temporal trend in the outbreak events and found it insignificant (t = 0.046, P = 0.940) after controlling for the publication efforts. The number of published articles was therefore used as a covariate in the subsequent statistical analyses.
Table 1 summarizes analyses of the Poisson models without and with spatially structured random effects using Bayesian inference for the five categories of water-associated diseases. The DIC values of the Poisson model with spatial random effects are smaller than that without spatial structure, suggesting that the spatial models provided a better fit to the data. The Poisson models with spatial structure were therefore used for risk factor analysis and mapping.
The population density was shown to be a significant risk factor for reported outbreaks of all categories of water-associated infectious diseases and the probability of outbreak occurrence increased with the population density. The accumulated temperature was a significant risk factor for water-related diseases only. The analysis suggested that occurrence of water-washed diseases had significantly inverse relationship with surface water areas. Such inverse relationship was also observed between the average annual rainfall and water-borne diseases (including water-carried) and water-related diseases.
Figure 3 (A–F) shows the risk distribution based on the model predictions with the blue indicating lower risk while the red representing higher risk. The model predictions suggested that west Europe, central Africa, north India were at the higher risk for water-borne diseases (e.g. Escherichia coli diarrhea), and notably, that the higher risk for water-borne diseases in west Europe was primarily driven by water-carried diseases (e.g. cryptosporidiosis). West Europe, North Africa, and Latin America tended to be at higher risk to water-washed diseases (e.g. viral conjunctivitis). Risks associated with water-based diseases (e.g. schistosomiasis) were higher in east Brazil, northwest Africa, central Africa, and southeast of China. High risk areas for water-related diseases (e.g. malaria and dengue fever) were clustered in central Africa in particular Ethiopia and Kenya, and north India. For water-dispersed diseases (e.g. Legionellosis), west Europe seemed to be at higher risk.
In the past decade there has been an increasing interest in understanding factors underlying the distribution of infectious pathogens, emerging and re-emerging infectious diseases. Some recent research efforts have been in attempt to determine large-scale ecological factors associated with diversity richness and distribution of infectious and parasitic pathogens [6], socio-environmental determinants of emerging infectious disease [1], and to explore the impact of global environmental change on distribution and spread of infectious diseases [23], [24]. These studies have offered valuable insights into understanding socio-environmental processes and factors underlying the distribution of infectious diseases. In this study, we focused our attention on water-associated infectious diseases and attempted to explore whether these diseases follow similar patterns observed in other studies [1], [6], and whether the distribution and occurrence of these diseases were related to terrestrial water dynamics (e.g. precipitation and land-surface water) together with other socio-environmental factors. The transmission of many infectious diseases is closely linked to water and the water-infectious pathogen interactions exhibit a complicated relationship depending on the transmission characteristics of the pathogens and water's roles in the transmission. The study showed that water-associated infectious diseases and outbreaks were broadly distributed throughout the world but the distribution of specific agents/diseases varied greatly from region to region. The majority of reported outbreaks events were associated with water-borne pathogen including those water-carried. Water-borne diseases have a much broader distribution than other water-associated diseases, suggesting a broader impact of waterborne pathogens in particular those related to fecal-oral route and water, sanitation, and hygiene. In addition to water, other environmental factors have also been recognized to play a significant role in the distribution, transmission, and outbreaks of these water-associated diseases [25], [26], [27].
It should be noted that, though, the outbreaks reported here only reflected “the tip of the iceberg” of the much larger problem. A complete count of outbreaks attributable to water-associated pathogens is impossible as underreporting is a universal problem, and reporting efforts and effectiveness may vary from country to country, and pathogens to pathogens, depending on many factors particularly availability of research and surveillance resources, and epidemiological characteristics of causal agents. In developing countries, outbreaks of many vector-borne infectious diseases such as dengue and malaria [28], [29] and gastrointestinal infections [30] were grossly underreported, partly due to their endemic characteristics. Even in the US, reporting completeness of notifiable infectious diseases varied from 9% to 99%, and was strongly associated with diseases being reported [31]. In general, water-borne pathogens usually exhibit acute manifestations and are more likely to be reported [32]. In contrast, other diseases such as water-based schistosomiasis, a disease of chronic infections and atypical symptoms, are more likely to be underreported. In this study, the primary source of outbreak information was from GIDEON, which is the most comprehensive database on infectious diseases and offers detailed information on epidemiology including distributions and outbreaks of infectious diseases for more than 205 countries and regions, as well as clinical manifestations and treatment associated with each disease [23], [33]. As expected, GIDEON does not include all outbreak information due to underreporting of outbreak events, but we believe that information from GIDEON is representative and provides an overview of available and recognized outbreak data, as argued by some other studies [23], [33].
The distribution of water-associated diseases, like many other infectious diseases, is highly heterogeneous. The spatial structure associated with the distribution of the outbreaks may be important in understanding underlying risk factors. To explore possible associations between socio-environmental factors and the outbreaks at the global scale, two Poisson models (without and with spatial structures) were developed. Among the two models explored, the one incorporating spatial effects provided a better fit to the data. Our findings suggested that the importance of these socio-environmental variables was dependent on the category of water-associated diseases. Human population density was a common significant risk factor for the outbreaks caused by all categories of water-associated diseases, in concurrence with the previous study suggesting that human population was an important predictor of emerging infectious diseases event at the global scale [1]. The accumulated temperature was a significant factor associated with water-related diseases, which was in agreement with many other studies [34], [35], [36], [37]. The transmission of diseases in this category typically involves vectors (e.g. mosquitoes) which require certain energy level (e.g. accumulated temperature) allowing completion of development of vectors and pathogens [10], [38], [39]. In this study, terrestrial surface water area (at each grid-region) was found to be inversely proportional to the outbreak events associated with water-washed diseases such as trachoma. The primary determinant of water-washed diseases is poor personal and/or domestic hygiene typically due to insufficient sanitary water for hygienic purpose, and this has been reported in many site-specific studies [40], [41], [42]. Our result from a large-scale correlation study supported these points of the previous studies, suggesting that regional water availability may be indicative of local water availability which is closely linked to personal and domestic hygiene. Our analysis indicated a negative relationship between average annual rainfall and water-related diseases, in contrast with some previous studies showing that some outbreaks of water-related diseases are positively associated with heavy rainfall events [8], [43], [44], [45]. This can be partly explained by issues related to scale and timing effects – the majority of studies reporting positive relationship between precipitation and waterborne illness was conducted at local scale and typically time-lag effects were considered. Indeed, the rainfall and water-related diseases exhibit complex relationships as shown in previous studies, and many rainfall-driven transmission and outbreaks were dependent on local circumstances. In addition to rainfall, multiple and covarying drivers have also been proposed for seasonal pattern of transmission and outbreaks of many water-associated diseases, including temperature, host demographic and biological characteristics [46], [47], [48]. However, due to lack of global information on seasonal patterns of outbreaks and the driving factors, temporal heterogeneity of outbreaks events, such as seasonality discussed here, was not included in the present study.
Using the best-fitted models we predicted global distributions of relative risks associated with each category of water-related infectious diseases, as shown in Figure 3. Surprisingly, the risk maps show that west Europe and central Africa were all at relatively higher risk for water-borne diseases. A closer look at pathogens associated with the reported outbreaks indicated different dominant species in the two regions – in Africa reports of water-borne outbreaks were primarily associated with Vibrio cholerae, whereas in west Europe giardia, cryptosporidium were common in the water-borne outbreaks, with the latter being particularly related to accidental ingestions of contaminated water (e.g. in recreational settings) and, to some extent, mixed with infections of food-borne sources [49], [50], [51].
Some limitations of the current study are recognized. Although possible reporting bias was adjusted for using publications for each country, the analysis may have missed countries/regions with outbreaks but no publications and/or reports. Second, only a few socio-environmental factors were considered in the present study and it is likely that some other factors might be associated with the outbreaks. In addition, significant prediction uncertainties were noted throughout the outbreak countries and regions, this was partly due to the temporal correlation of the outbreak events which was not considered in the analysis. The addition of such information (e.g. temporal trend of outbreaks in places where repeated outbreaks occurred) to the model may improve model prediction. In spite of these, we think that overall patterns of distribution and associated risk factors presented here are informative and offer insights into global distribution and risk factors associated with water-associated diseases, although further studies on other possible risk factors and modeling approaches to improving prediction are still needed.
In conclusion, our study, to our knowledge, is the first to describe global distribution of outbreaks caused by water-associated infectious diseases and explore possible risk factors underlying the distribution of these outbreaks at the global scale. The risk maps may offer insights for future studies and for prioritizing health resources.
|
10.1371/journal.pcbi.1004582 | Dynamic Redox Regulation of IL-4 Signaling | Quantifying the magnitude and dynamics of protein oxidation during cell signaling is technically challenging. Computational modeling provides tractable, quantitative methods to test hypotheses of redox mechanisms that may be simultaneously operative during signal transduction. The interleukin-4 (IL-4) pathway, which has previously been reported to induce reactive oxygen species and oxidation of PTP1B, may be controlled by several other putative mechanisms of redox regulation; widespread proteomic thiol oxidation observed via 2D redox differential gel electrophoresis upon IL-4 treatment suggests more than one redox-sensitive protein implicated in this pathway. Through computational modeling and a model selection strategy that relied on characteristic STAT6 phosphorylation dynamics of IL-4 signaling, we identified reversible protein tyrosine phosphatase (PTP) oxidation as the primary redox regulatory mechanism in the pathway. A systems-level model of IL-4 signaling was developed that integrates synchronous pan-PTP oxidation with ROS-independent mechanisms. The model quantitatively predicts the dynamics of IL-4 signaling over a broad range of new redox conditions, offers novel hypotheses about regulation of JAK/STAT signaling, and provides a framework for interrogating putative mechanisms involving receptor-initiated oxidation.
| Incomplete reduction of oxygen during respiration results in the formation of highly reactive molecules known as reactive oxygen species (ROS) that react indiscriminately with cellular components and adversely affect cellular function. For a long time ROS were thought solely to be undesirable byproducts of respiration. Indeed, high levels of ROS are associated with a number of diseases. Despite these facts, antioxidants, agents that neutralize ROS, have not shown any clinical benefits when used as oral supplements. This paradox is partially explained by discoveries over the last two decades demonstrating that ROS are not always detrimental and may be essential for controlling physiological processes like cell signaling. However, the mechanisms by which ROS react with biomolecules are not well understood. In this work we have combined biological experiments with novel computational methods to identify the most important mechanisms of ROS-mediated regulation in the IL-4 signaling pathway of the immune system. We have also developed a detailed computer model of the IL-4 pathway and its regulation by ROS dependent and independent methods. Our work enhances the understanding of principles underlying regulation of cell signaling by ROS and has potential implications in advancing therapeutic methods targeting ROS and their adverse effects.
| From initially being perceived as accidental and harmful byproducts of aerobic respiration, reactive oxygen species (ROS) have emerged as important regulators of physiological cell signaling [1]. In particular, due to its relatively long half-life, enzymatic regulation, and specificity for protein thiols, hydrogen peroxide (H2O2) is recognized as an important second messenger in signal transduction [2]. Activation of many classes of cell surface receptors induces transient ROS production by activating NADPH oxidase (Nox) family enzymes; the enzymatically produced ROS play a role in modulating downstream signaling [3–5]. ROS such as H2O2 can either directly react with the thiol functional group of susceptible cysteine residues in redox sensitive proteins or indirectly oxidize protein thiols through an intermediate “relay” protein [6], converting the cysteine to sulfenic acid form [7]. Alternatively, lipid electrophiles may oxidize thiols without ROS directly coming in contact with proteins in the cellular milieu [8]. While further oxidation is irreversible, the sulfenic acid form can be protected by formation of disulfides and sulfenyl amides which can be reduced back by oxidoreductases such as thioredoxin and glutaredoxin [9–11]. Reversible cysteine oxidation can result in transient changes in protein function, such as gain or loss of catalytic activity, at several points in a signaling pathway resulting in systemic changes in cell signaling dynamics [12]. This reversibility has been noted as particularly relevant for the protein tyrosine phosphatases (PTP), due to a conserved low pKa cysteine residue in their active sites [13–15].
Methods for detecting intracellular changes in protein sulfenic acids are developing [16–18] but technical challenges remain to be addressed before quantitative, systems level measurements are possible [19]. Inferring relative contributions of and interactions between various regulatory mechanisms is not straightforward due to inherent complexity of signaling pathways. Computational modeling can be used to simulate multiple oxidative-regulatory events and evaluate their relative importance in a meaningful way. Development of computational models for redox systems has recently gained traction [12,20–23]; however, to date, most applications have focused on modeling systems in which the system structure is well-defined. The difficulty of experimentally monitoring redox events is well-suited for using available signaling data to infer the underlying interactions in signaling networks by computational methods [24].
We selected the IL-4 signaling pathway as a representative redox-regulated network [25] and have complemented easily observable phosphoprotein data with computational methods to analyze the role of putative redox-regulatory mechanisms. The IL-4 pathway signals through Janus kinases (JAK) 1 and 3, which are constitutively bound to the IL-4 receptor chains [26]. Activation of JAK is followed by recruitment and phosphorylation of cytosolic signal transducer and activator of transcription 6 (STAT6) by the activated receptor complex. Phosphorylated STAT6 (pSTAT6) forms a homodimer, the active form of STAT6 that can initiate transcription of target genes after translocating to the nucleus [27,28]. Multiple PTPs, including CD45, SHP-1, PTP1B and TCPTP, are involved in down-regulation of IL-4 signaling. While CD45 inhibits phosphorylation of both JAK1 and JAK3 [29,30], SHP-1 can dephosphorylate either the receptor or JAK molecules [26,31,32]. PTP1B and TCPTP are also reported as regulators of STAT6 in IL-4 signaling [33,34]. Moreover, TCPTP can shuttle between the cytosolic and nuclear compartments making it an important nuclear regulator of STAT6 phosphorylation [33]. Multiple components of the IL-4 pathway described above can act as redox sensors, in that their oxidation confers a functional change in the protein activity [1,35]. Rapid ROS production via PI3K-mediated Nox activation has been observed in IL-4 treated A549 cells, corresponding with concomitant oxidation of PTP1B [25]. ROS such as H2O2 have been reported to reversibly oxidize PTPs including PTP1B, TCPTP, CD45 and SHP-1 following activation of a variety of other cell surface receptors or when H2O2 is added exogenously [5,11,36–38]. Based upon the conservation of the active site cysteine in PTPs, we hypothesized that the other PTPs involved in IL-4 regulation are also redox regulated in a manner similar to PTP1B (Fig 1A). Furthermore, members of the JAK family have been shown to possess a redox sensitive switch and could be involved in redox regulation of IL-4 signaling (Fig 1B). JAK2 and JAK3 have been shown to become increasingly inactive under oxidizing conditions [39,40]. Structural homology between JAK1 and JAK2 [39] as well as indirect experimental evidence [41] suggest that JAK1 could also be inactivated by oxidation. In addition to catalytic activities of the PTP and JAK proteins, the subcellular localization of TCPTP has also been shown to be affected by redox state of the cell with more oxidizing conditions favoring cytosolic accumulation of TCPTP [42] (Fig 1C). All these lines of evidence drawn from various cell types and receptor systems suggest that IL-4 signaling is potentially controlled by multiple mechanisms of redox regulation; however, biochemical and cell-based studies so far have focused on examining individual mechanisms without accounting for competing influences exerted within an intact protein signaling network.
Given that multiple potential points of redox-dependent and independent control points exist in the IL-4 signaling network, how do the various putative mechanisms interact to regulate overall signaling dynamics? In the scope of physiologically relevant levels of oxidants, to what magnitude does this form of post-translational modification (PTM) alter pathway activation, and on what timescales does oxidation dissipate as a signaling mechanism? These questions are particularly challenging because the outcomes of potential redox regulatory mechanisms are qualitatively opposite in nature. For instance, while PTP oxidation could increase signaling activity, JAK oxidation could suppress it. From our novel computational approaches, we conclude that oxidation of multiple PTPs is the dominant mechanism of redox regulation in IL-4 signaling and the coupling of dynamic phosphatase activity with redox-independent mechanisms is critical in explaining IL-4 signaling dynamics both at initial ligand-receptor initiation and down-regulation several hours later. We have developed a systems model of the IL-4 pathway that successfully predicts IL-4 signaling dynamics over a wide range of redox conditions, demonstrating how intracellular modulation of receptor-initiated signaling can occur by altered cellular redox potential.
To examine global oxidative post-translational modifications that may occur across the proteome during IL-4 signaling, we performed redox differential 2D gel electrophoresis (Redox-DIGE) [43,44] comparing unstimulated Jurkat cells to 30 minutes post treatment with 100 ng/mL IL-4. Consistent with the prior report of intracellular oxidation occurring during this time frame [25], we observed a characteristic pattern of Redox-DIGE indicating thiol oxidative modifications of proteins (Fig 1D). We observed three sets of protein spots, green, red, and yellow. As described in the preceding section, upon treatment with IL-4 proteins can be oxidized reversibly (green spots) or irreversibly (red spots). A third set (yellow overlay) representing the majority of proteins visualized in this manner, remained unchanged under IL-4 treatment.
Jurkat cells were stimulated with 100 ng/mL IL-4 and intracellular oxidation was monitored using flow cytometry by staining the cells with CM-H2DCFDA. Fluorescence of the dye increased quickly after addition of IL-4 and the dye approached maximal oxidation 1 hour after IL-4 addition (Fig 2A). To ensure that the saturation in dye oxidation was not due to limitation in loading of the dye, a bolus of excess H2O2 was added to the cells and time course of dye fluorescence was measured. The fluorescence exceeded that observed under IL-4 stimulation showing that IL-4 did not saturate the dye oxidation signal (S1 Fig). Pretreating the cells with 20 μM diphenyleneiodonium chloride (DPI), an inhibitor of phagocytic NOX and other flavoproteins, lowered the baseline oxidation of the dye and significantly suppressed fluorescence/oxidation following stimulation with IL-4 (Fig 2A). Because the oxidation of H2DCFDA is an irreversible process, the fluorescence time courses shown in Fig 2A represent cumulative oxidation of the dye as a function of time. In order to infer instantaneous levels of ROS from the cumulative dye oxidation time courses, Hill curves were fitted to the data points (Fig 2A) and derivatives of these curves were obtained (Fig 2B). The derivatives indicated that intracellular oxidation increased rapidly following IL-4 treatment of Jurkat cells, peaked at approximately 20 min and gradually returned to the baseline level. In DPI pretreated cells, the increase in oxidation was observed to be much lower than that in cells not exposed to DPI.
To study the effects of intracellular oxidation on IL-4 signaling, time-dependent phosphorylation of total intracellular STAT6 (i.e., sum of nuclear and cytosolic pSTAT6; see S2 Fig) was quantified under a variety of oxidative conditions. Treatment of Jurkat cells with IL-4 significantly increased STAT6 phosphorylation within 5 min, and the phosphorylation was sustained for 2 hours (Fig 2C). Cells pretreated with DPI showed significantly lower baseline phosphorylation of STAT6 and responded very weakly to IL-4 stimulation (Fig 2C). Addition of exogenous hydrogen peroxide (10 μM) in combination with IL-4 further increased STAT6 phosphorylation when compared to IL-4 treatment alone (Fig 2D). However, unlike the MAPK cascade [45,12], addition of H2O2 in the absence of IL-4 failed to alter STAT6 phosphorylation from its basal level (Fig 2D).
The pSTAT6 time course in response to IL-4 stimulation of Jurkat cells showed two distinct local maxima over a two-hour period (Fig 2C). Statistical analysis confirmed that the first maximum is likely to occur between 0 and 25 min, and the second between 25 and 120 min (S3 Fig). We hypothesized that the dynamic information contained in the STAT6 phosphorylation time course, especially the characteristic shape of the curve with two peaks, could be used to infer the regulatory mechanisms involved in IL-4 signaling. We sought to use this information to investigate the importance of four distinct regulatory mechanisms described above: i) reversible inactivation of PTPs by oxidation (Fig 1A); ii) reversible inactivation of JAK by oxidation (Fig 1B); iii) ROS mediated cytosolic accumulation of PTPs (Fig 1C); and iv) dependence of nuclear-cytosolic shuttling of STAT6 on its phosphorylation state. While the first three are directly influenced by the redox state of the cell, the fourth is not. The IL-4 signaling network was conceptually divided into 5 regulatory modules and by taking different combinations of these modules, a library of 16 different models was constructed (Fig 3A; S4 Fig). This library covers all possible combinations of the 4 mechanisms listed above. Next, simplified ODE representations of all 16 models were obtained using a rationale-based approach to reduce model complexity. Specifically, linear chains of events such as sequential assembly of active receptor complex, or dimerization of phosphorylated STAT6 were collapsed into a single reaction. JAKs, which are constitutively bound to the receptor, were not modeled explicitly and were assumed to be implicit in the receptor. Different PTPs that can dephosphorylate STAT6 were abstracted as a single generic PTP. Similarly, the “ROS” species in the model is a generic representation of oxidants that can cause direct or indirect thiol-based post-translational modifications that result in modified protein function.
The various network topologies obtained were coded into systems of ordinary differential equations (ODEs) assuming elementary mass action kinetics for all reactions (S2 Text). Next, we assessed these models based on their ability to produce two distinct pSTAT6 peaks with characteristics similar to those seen in the experimental data. Parameters of the models were manually adjusted so that total pSTAT6 dynamics roughly matched the experimentally observed dynamics (rapid increase followed by a slow decrease). For each model, 50,000 sets of parameters were randomly sampled in a fixed space spanning one order of magnitude around the estimated parameter vector, the model was simulated for all sampled parameter vectors and the dynamics of total pSTAT6 were recorded. The predicted pSTAT6 traces from these Monte Carlo (MC) simulations were qualitatively and quantitatively compared with the experimental results as described next to judge the fitness of the models.
The models were first tested qualitatively based on their ability to reproduce the two peaks observed in the experimental data. This was done by counting the number of local maxima in each pSTAT6 trace produced by the model ensemble. All simulations produced two or fewer peaks, and traces that could produce exactly two distinct local maxima were taken to qualitatively match the two peaks observed in the experimental data. Among the 16 models, 7 failed to produce any traces with two distinct peaks for total pSTAT6. Of the remaining 9 models which generated one or more traces with two peaks, 5 models produced fewer than 50 instances (0.1% of number of MC simulations per model) of pSTAT6 time courses with two peaks (Fig 3B). Using this threshold of 0.1%, 12 network configurations were rejected as likely models on a purely qualitative basis due to their inability to reproduce the two peaks observed experimentally. Notably, all 8 models in which STAT6 translocation was independent of its phosphorylation state consistently failed to cross the 0.1% threshold (Fig 3B). This strongly suggests that cycling of STAT6 between nucleus and cytosol is phosphorylation dependent. Only 4 models demonstrated the occurrence of two distinct peaks for more than 0.1% of the 50,000 MC simulations, and all 4 of these models included PTP oxidation as a redox regulatory mechanism (Fig 3B). The model that had PTP oxidation as the only ROS-mediated mechanism generated the most instances of pSTAT6 traces with two distinct peaks (first bar in Fig 3B). The other two redox regulated mechanisms, JAK oxidation and ROS mediated nuclear translocation of PTP, could not cross the 0.1% threshold when acting alone; however, combining one or both of these mechanisms with PTP oxidation allowed two pSTAT6 peaks to occur. Nevertheless, fewer instances of traces exhibiting two peaks were generated when either one of these mechanisms was combined with PTP oxidation, and even fewer when both were added in together.
Quantitative comparison of features of the curves with two distinct peaks generated from the simulations with the experimental data provided further support to PTP oxidation as the prime mechanism of redox regulation in IL-4 signaling. A smoothing spline was fitted to the mean pSTAT6 data and features of the curve including heights of the two peaks, separation between them and the value at the final time point were extracted (Fig 4A). Simulations that produced two distinct local maxima for the total pSTAT6 trace were identified for each model. The features indicated in Fig 4A were extracted from the simulated curves. The ratio of peak heights, separation between the peaks, and the ratio of final value to first peak were computed and compared with the experimental results. Representative results are shown for two models in (Fig 4B and 4C, corresponding signaling networks in 4D and 4E). When only PTP oxidation was included as a mechanism of redox regulation, not only did the model produce the most instances of curves with two peaks, but the features of these curves also conformed well with the measured dynamics (Fig 4B and 4D). However, when additional ROS dependent mechanisms were included in the model, both the number of curves with two peaks and their similarity with experimental data decreased (Fig 4C and 4E). Indeed, the network shown in Fig 4D exhibited better qualitative and quantitative fit to experimental data than all the other networks in the mode library.
Collectively, exploitation of the observed dynamic behavior in our data set allowed us to explore topological features of the IL-4 network that dictate regulation of STAT6 phosphorylation. Two of the four regulatory mechanisms considered emerged as most crucial for recapitulating proper behavior: i) PTP oxidation; and ii) phosphorylation-dependent nuclear-cytosolic translocation of STAT6.
The experimental results presented above taken together with the results from the initial Monte Carlo simulations suggest a complex picture of the IL-4 pathway with many regulatory mechanisms operating in tandem. We constructed a more detailed ordinary differential equation model of the IL-4 signaling pathway using mass action kinetics that incorporates these important regulatory mechanisms and explains the observed dynamics of various molecular species under a variety of experimental conditions (Fig 5). Mechanisms found to be most important from the MC analysis, namely reversible PTP oxidation and phosphorylation dependent nuclear-cytosolic cycling of STAT6, were included into the model. Experimental results on down-regulation mechanisms were used to include SOCS and STAT6 degradation as important control mechanisms.
In the interest of parsimony, several simplifying assumptions were made in the model. The IL-4Rα chain and the γC chain were not modeled separately. Instead, we used an abstraction of the receptor complex in the form of a single transmembrane molecule that binds IL-4 and becomes activated. The JAK1 and JAK3 molecules that are constitutively bound to the receptor chains were also not modeled explicitly and were assumed to be implicit in the receptor molecule. We examined the effects of explicitly modeling JAK and found that the simulation results were not altered significantly by our assumption (S5 Fig). SOCS family proteins have been shown to inhibit JAK/STAT signaling by binding directly to phosphorylated JAK molecules inhibiting their function, or by binding to the receptors and indirectly inhibiting JAK [46]. Since JAK and receptor molecules were abstracted into a single species, SOCS binding to activated receptor was taken to represent both possibilities. Dephosphorylation of the receptor complex was assumed to result in dissociation with SOCS. SOCS1 and SOCS3 are known to affect IL-4 signaling and were modeled together as a generic SOCS molecule. Multiple phosphatases including PTP1B and TCPTP have been shown to act on STAT6 [33,34]. Similarly, multiple phosphatases, such as CD45 and SHP-1, can dephosphorylate the receptor and JAK molecules [26,29]. We assumed that STAT6 and the receptor complex were dephosphorylated by distinct, individual PTPs. All reactions were modeled using the law of mass action except active STAT6 mediated SOCS production; the rate of production of SOCS was assumed to increase monotonically with the concentration of active STAT6 in a saturating fashion to model the eventual saturation of transcription factor binding sites.
Activation of the IL-4 receptor in Jurkat cells induced transient production of ROS as shown in Fig 2B. ROS dynamics estimated from experimental data were used as an input to the model, thus eliminating the need to explicitly model the calcium and PI3K-mediated initiation of Nox as well as numerous mechanisms of ROS clearance [47]. Reduction of oxidized proteins was assumed to follow first order kinetics; in other words, the reducing capacity of the cell was assumed to be constant over time. The more reducing environment of the nucleus [48,49] was modeled by excluding protein oxidation reactions from the nuclear compartment.
An evolutionary strategy algorithm with hyper-mutation algorithm was used to estimate the parameters of the model. The objective function defined in the Methods section was minimized to fit the model to experimentally measured time courses of multiple species across 3 different experimental conditions (S1 Text and S6 Fig). Jurkat cells were stimulated with IL4 with or without pretreatment by either MG132 or cycloheximide (CHX). MG132 is an inhibitor of the proteasome and prevents degradation of phosphorylated STAT6. CHX is an inhibitor of protein synthesis and is expected to inhibit the expression of pSTAT6 inducible suppressors of cytokine signaling (SOCS) thereby elevating the pSTAT6 signal (effects of MG132 and CHX are discussed further in S1 Text and S7 Fig). The fitted model recapitulates the dynamics of i) pSTAT6 in Jurkat cells following IL-4 stimulation with or without pretreatment by either MG132 or CHX (S6 Fig, panels A and B); ii) total STAT6 following IL-4 stimulation with or without MG132 (S6 Fig, panel C); and iii) SOCS3 under the conditions of IL-4 stimulation with or without inhibition by CHX (S6 Fig, panel D).
We tested the detailed ODE model’s ability to predict the dynamics of IL-4 signaling under new experimental conditions of high and low cellular oxidative conditions that were not used to train the model. We simulated two conditions, i) DPI pretreatment followed by IL-4 stimulation, and ii) addition of exogenous H2O2 simultaneously added with IL-4. The experimentally measured ROS profile under DPI pretreatment (Fig 5B) was used as input to simulate the effect of DPI in the model. To mimic the effect of 10 μM exogenous H2O2 added along with IL-4, a bolus of exponentially decaying ROS was added to the experimentally measured ROS profile under IL-4 treatment (Fig 5B). The model predicted severely attenuated STAT6 phosphorylation under DPI pretreatment, whereas addition of exogenous H2O2 with IL-4 was predicted to amplify STAT6 phosphorylation in comparison to IL-4 treatment alone (Fig 5C). Both these predictions matched quantitatively with experimental results, showing that the model is robust enough to predict novel dynamics of IL-4 signaling over a wide range of redox conditions. To further validate the refined IL-4 model, we monitored reversible PTP oxidation using a monoclonal antibody against the oxidized catalytic domain of PTPs. An increase in PTP oxidation at multiple molecular weights was observed after 30 min of IL-4 treatment followed by a slight decrease in PTP oxidation by 60 min (Fig 5D).
While computational models of ROS production and consumption have previously been developed for cells [47,50], competing redox-based mechanisms during cell signaling have not been sufficiently explored by computational systems biology approaches. The possibility of multiple mechanisms of redox regulation specific to the IL-4 pathway poses challenges for deducing these features from observable experimental data. Our Redox-DIGE result of IL-4 induced thiol oxidation on a proteomic scale suggests that a number of reduction and oxidation post-translational modifications occur during active cytokine signaling (Fig 1D). In prior work, theoretical analysis of various mechanisms of redox regulation overlaid on signaling network topologies demonstrated that complex signaling behaviors could arise from such interactions; computational modeling is a powerful tool to unravel this complexity [12]. Here, we have comprehensively examined proposed mechanisms of redox regulation in the IL-4 pathway using computational modeling based on quantitative and easily acquired phosphorylation data to gain insight into the importance of reversible inactivation of PTPs by oxidation in IL-4 signaling.
The model and experimental results collectively suggest ROS may be playing two roles; first, to keep the system ready to respond to an input signal and second, to amplify the response once it is initiated by IL-4 input. Consistent with previous reports in A549 cells [25], IL-4 treated Jurkat cells were observed to quickly increase fluorescence of DCFDA, a molecular probe often used as a surrogate reporter of intracellular oxidation. DPI suppressed this ROS production indicating that Nox, isoforms of which are known to be expressed in Jurkat cells [51], could be involved in the transient ROS production. Pretreatment with DPI also significantly lowered baseline STAT6 phosphorylation and severely attenuated pSTAT6 response to IL-4 treatment; the oxidation of CM-H2DCFDA at the initial time point was also lower in the DPI pretreated cells. We interpret this result as a sustained basal ROS production that maintains a fraction of STAT6 in the phosphorylated state. Failure of the DPI treated cells to respond to IL-4 suggests that baseline ROS also have a role in sustaining the system in a "primed" state so that it is ready to quickly respond to an activating stimulus. Supplementing IL-4 induced ROS with exogenous H2O2 amplified the pSTAT6 response; however, on its own, the same concentration of H2O2 did not elicit STAT6 phosphorylation.
Having confirmed that IL-4 signaling in Jurkat cells is redox regulated, we used computational analysis to understand the control mechanisms by which ROS regulate IL-4 signaling. A key feature of our studies was examining a duration of signaling in which down-regulation occurred (~ 2 hours), which provided information about the summative effects of transient (i.e. reversible) thiol-based oxidation of phosphatases, SOCS feedback, and proteasomal degradation. The experimentally acquired time course of IL-4 induced STAT6 phosphorylation in Jurkat cells had a distinctive shape, presenting two distinct peaks over a two hour time period. This feature combined with combinatorial exploration of simplified models of the pathway proved a remarkably useful tool for selecting models based simply on their ability to reproduce experiment-like pSTAT6 time course with two peaks. When possible combinations of mechanisms were systematically examined, PTP oxidation stood out as the most likely mechanism by which redox-mediated regulation can take place. The other mechanisms considered, namely JAK oxidation and redox-mediated PTP translocation, were not sufficient on their own to explain the observed dynamics. Models with increased complexity that incorporated multiple redox mechanisms did not result in improved fits when tested within the selected parameter bounds. This double failure of the more complex models is compelling evidence in support of the PTP oxidation model, which is, firstly, parsimonious, and secondly, matches the data better, both qualitatively and quantitatively.
An important concern regarding the ability of ROS to play a significant role in modulating cell signaling has to do with the relatively slow rates of H2O2 mediated cysteine oxidation in proteins. Rather than assigning explicitly the mechanism of action or the ROS involved, we used a simple abstraction of intracellular dye data to generically describe phosphatase inactivation regardless of how this occurs at the molecular level. Since intracellular ROS concentration was modeled using scaling of fluorescence-based measurements, it could not be assigned real units in the model. This means that protein oxidation rates are also to be understood in terms of these arbitrary units. However, the scaling factor was chosen such that the absolute value of ROS concentration was in the order of 100 units of ROS in simulations of IL-4 stimulated cell. Since intracellular H2O2 concentration is thought to be in the sub-micromolar range [52], a 1:1 scaling can be assumed between the model's arbitrary ROS units and nanomolarity. Thus, assuming 1 unit of ROS in the model corresponds to 1 nM ROS in the real cell, the simulated ROS levels scale to the order of 100 nM, which is a reasonable estimate of intracellular H2O2 concentration [47,53]. Assuming this scaling, the estimated rate of PTP oxidation turns out to be 2×106 M-1s-1. In a previously published systems model of H2O2 dispersion in Jurkat cells, the second order rate of H2O2 mediated protein oxidation was estimated to range between 10×107 M-1s-1 for the fastest reactions involving catalase and peroxiredoxin to 10×104 M-1s-1 for the average intracellular protein [47]. The estimated rate of PTP oxidation in our model lies within this range. These rates are still much higher than in vitro measurements of PTP oxidation rates, but in vitro estimates themselves have been found to be much slower than observed rates of PTP oxidation (e.g. [14]). Several mechanisms have been discovered which may explain this high apparent rate of oxidation, including localization of ROS to create high concentration [54,55] and transfer of oxidation state through relay proteins [6,53], and this continues to be an area of active investigation.
Based on inferences from analysis of reduced models of redox regulation and knowledge gained from experimental data, we developed a highly detailed model of the IL-4 signaling pathway. This model includes two reversibly oxidized phosphatases, P1 and P2, responsible for receptor complex/JAK and STAT6 dephosphorylation. Because PTPs have been associated with multiple targets within the IL-4 pathway, we did not assign these variables specifically to TCPTP, PTP1B, SHP1, or CD45. The oxPTP immunoblots (Fig 5D) are consistent with model estimates of increased PTP oxidation at 30–60 minutes, with changes observed at molecular weights approximating these PTPs (45, 50, 67.5, and 147 kD, respectively). Interactions between the transient inactivation of PTPs coupled with STAT6 degradation and SOCS feedback allowed the model to successfully predict the pathway dynamics over a wide range of oxidized and reduced experimental conditions. Collectively, the computational analysis indicates that while ROS mediated regulation is a very important arm of the control machinery in IL-4 signaling, systemic behavior of the pathway emerges from interactions of redox and non-redox regulatory mechanisms. For instance, none of the redox regulation mechanisms considered in our analysis were sufficient to explain pSTAT6 dynamics when working alone (Fig 3B), but combining them with other ROS-independent mechanisms changed the behavior of the system qualitatively. In other words, without considering all forms of regulation we would have reached a faulty conclusion.
In developing our models we have made several assumptions that could limit the scope of the results obtained from the modeling analysis. We have assumed mass action kinetics for the biochemical reactions included in the model with lack of compartmentation or spatial definition. While there is ample precedence for modeling interleukin signaling using mass action kinetics [56–59] as well as for modeling redox kinetics [12,47], it is possible that it is not the most appropriate mechanism for some of the reactions. With respect to the library of reduced models, we have tested a fixed set of network topologies within constrained parameter ranges. While the structures are based on current knowledge of the biology of the IL-4 pathway, the parameter space or the set of network structures explored is a subset of all possible topology-parameter combination space. Additionally, in the reduced models as well as the detailed model, we have used several simplifying assumptions (see Results), including ignoring the time-dependent variation of the intracellular reducing pool and abstracting direct and indirect protein oxidation into a single oxidation step. While our model is built with these assumptions and simplifications, we have validated it with independent experimental data, which provides strong support to our modeling strategy.
This quantitative, model-based approach to investigating redox mechanisms responsible for systemic regulation of the IL-4 pathway provides several future venues for further computational and experimental analysis. Hypothetical roles of relay proteins [53] or lipid electrophiles [8,60] as intermediates for electron transfer, for example, or the proximity to Nox for subcellular localization of redox control [55] can be tested further with this modeling framework. Iterative feedback between modeling and experimentation will help elucidate the operating principles in redox regulation of cell signaling, especially as technical advances in measuring redox PTMs continue.
Jurkat T cells were cultured in supplemented RPMI media as previously described [47]. For all experiments Jurkat cells were first serum starved at a concentration of 2×106 cells/ml in media containing 0.5% FBS for 4 hours. Cells were treated with 100 ng/ml human recombinant IL-4 (R&D Systems). The inhibitors cycloheximide (CHX), diphenyleneiodonium chloride (DPI) and MG132 (all EMD Millipore) were administered at 20 μg/ml, 20 μM and 10 μM, respectively. Rabbit anti-STAT6 phosphotyrosine 641 antibody (Cell Signaling Technology) was used at 1:100 dilution. Rabbit anti-SOCS3 antibody (Abcam) was used at 1:500 dilution. R-PE conjugated anti-rabbit IgG (Life Technologies) was used as secondary antibody for flow cytometry at 1:500 dilution. R-PE conjugated anti-STAT6 (BD) was used per manufacturer's recommendation. Jurkat cells were incubated in 5μM CM-H2DCFDA (Life Technologies) for 30 min to measure intracellular oxidation.
Serum starved cells were suspended in PBS at 4×106 cells/ml. For each time point in an experiment, the appropriate inhibitor was added one hour before IL-4 stimulation. Cells were fixed in 1.5% paraformaldehyde for 10 min at room temperature and permeabilized in 90% methanol for 30 min at 4°C. After staining with suitable antibodies, the samples were analyzed on a BD LSR II flow cytometer. Fluorescence data were analyzed using in-house code written in Matlab; mean fluorescence intensity (MFI) was used to summarize the observations for analyzed cell populations. For pSTAT6 and SOCS3 analysis, cells non-specifically labeled only with secondary antibody were used to acquire the background signal. For STAT6, unstained cells (because the primary Ab was fluorophore conjugated) were used as background. To enable comparison between experimental conditions, background corrected MFIs were normalized by dividing by the background corrected MFI of Jurkat cells not treated with inhibitors and IL-4 as follows:
MFInorm=MFIsample−MFIbkMFIuntreat−MFIbk
where, MFInorm is the background corrected, normalized MFI, MFIsample is the experimentally measured MFI of the sample, MFIbk denotes the background MFI, and MFIuntreat is the measured MFI of untreated Jurkat cells.
To detect intracellular redox state, CM-H2DCFDA was added 30 min prior to IL-4 stimulation. After adding IL-4, samples were analyzed on BD LSR II Flow Cytometer to measure MFI of DCFDA. MFI time course of DCFDA stained cells not treated with IL-4 was subtracted as background from all other DCFDA time courses.
For detecting reversibly oxidized PTPs, we used thiol chemistry to irreversibly oxidize cysteines that were oxidized during IL-4 treatment and used immunoblotting with a monoclonal antibody against the irreversibly oxidized catalytic domain of PTPs [61,62] to measure PTP oxidation. Jurkat T cells were serum starved and stimulated with IL-4 as described in the previous section. Cells were lysed with Argon purged lysis buffer (20 mM Tris HCl, 10% glycerol, 1 mM benzamidine hydrochloride, and 10 μg/ml containing 100 mM iodoacetamide to protect reduced thiols and prevent disulfide exchange) for 20 min in the dark in an Argon purged AtmosBag (Sigma Aldrich). Lysates were then sonicated for 10 min at 4°C. Following the removal of insoluble cellular debris by centrifugation at 18000 x g for 20 min at 4°C, excess iodoacetamide was removed from the lysates using Microspin G-25 columns (GE Healthcare). Samples were normalized by protein concentration using measurements obtained by BCA. Samples were then frozen overnight. Reversibly oxidized thiols were reduced with 500 mM DTT for 30 min on ice. Excess DTT was removed with Microspin G-25 columns. Nascent thiols were irreversibly oxidized for 1 h at room temperature using freshly prepared 1 mM pervanadate. Samples were diluted in Laemmli sample buffer for Western blotting according to standard procedures. For measurement of oxidized PTPs, membranes were blocked for 1 h at room temperature with Rockland blocking buffer (Rockland) and incubated overnight with 10 μg/ml Mouse anti-oxidized PTP antibody (R&D Systems) with 0.1% Tween 20. Membranes were washed three times in TBS-T and probed with Donkey anti-Mouse IRDye680 (Licor) with 0.1% Tween 20 and 0.01% SDS for 1 h at room temperature. Membranes were stripped at 50°C for 15 min, blocked overnight at 4°C, and reprobed for 1 h with Rabbit anti-actin (Sigma) in Rockland blocking buffer with 0.1% Tween 20 and 0.1% SDS as a loading control. Membranes were imaged using a Licor Odyssey scanner and ImageStudio software.
After the individual treatments, the harvested cells were resuspended in ice-cold PBS containing 50 mM NEM and transferred to 2.0 mL tubes. The cells were washed twice by 30 s centrifugation in NEM/PBS at 10,000 g. After the second PBS wash, the cells were resuspended in the lysis buffer (50 mM NEM, 40 mM HEPES, 50 mM NaCl, 1 mM EDTA, 1 mM EGTA, protease inhibitors, pH 7.4) to a density of 2 x 107 cells/mL and incubated at 37°C for 5 min. 1% w/v CHAPS was then added to the cell lysate, vortexed, and incubated for a further 5 min at 37°C. The sample was centrifuged at 8000 g for 5 min to remove the insoluble material. SDS was then added to a final concentration of 1% w/v and the supernatant vortexed and incubated for another 5 min at 37°C. The unreacted NEM was removed using Micro Bio-Spin 6 columns (Bio-Rad) equilibrated with the lysis buffer. Protease inhibitors were not present in the lysis buffer from this step. The samples were then reduced with 2.5 mM DTT for 10 min at room temperature. Excess DTT was removed with Micro Bio-Spin 6 columns (Bio-Rad) equilibrated with argon sparged lysis buffer. The samples were immediately labeled with 40 μM CyDye DIGE Fluor Cy3 and Cy5 saturation dyes (GE Healthcare). After 30 min at 37°C, the reaction was quenched with 2.5 mM DTT. Unreacted dyes and salts were removed using Micro Bio-Spin 6 columns (Bio-Rad) equilibrated with the first-dimension rehydration buffer (7M urea, 2M thiourea, 2% CHAPS, 0.28% DTT, and 2% IPG buffer 3–10 NL). Protein concentrations were measured using the 2-D Quant kit (GE Healthcare). Equal amounts of the Cy3 and Cy5 maleimide labeled samples were pooled and resolved by two-dimensional electrophoresis. All experiments were replicated by swapping the dyes between treatments.
40 μg of fluorescently labeled protein was diluted to 200 μL with first-dimension rehydration buffer and absorbed overnight onto 11 cm pH 3–10 IPG strips (Bio-Rad). IEF was performed on Ettan IPGphor 3 (GE Healthcare) instrument for a total of 20,000 V-h at 20°C at 50 μA. Prior to SDS-PAGE, the strips were equilibrated with rocking for 10 min in 75 mM Tris-HCl, pH 6.8, 30% glycerol, 6 M urea, 2% SDS, 0.5% DTT. The strips were loaded onto a 8–16% Precast Criterion Gel and were run for approximately one hour at 200 V. A protein mixture was made in the laboratory by mixing equal amount of Cy3 and Cy5-labeled reduced proteins from a different cell lysate as a control for imaging artifacts. The bromophenol blue dye front from this sample was monitored to determine the completion of the second dimension run. After two-dimensional electrophoresis, gels were transferred to a Typhoon imager (GE Healthcare), and fluorescent spots were viewed using 532 and 633 nm lasers in conjunction with 580 and 670 nm emission filters (band pass 30 nm), respectively.
All modeling, simulation and analyses were performed in Matlab. For Monte Carlo (MC) simulations, all the networks to be analyzed were coded as ordinary differential equation (ODE) systems assuming mass action kinetics and solved using the ode23 numerical ODE solver in Matlab. The ODE system representing the largest model with all regulatory mechanisms and the parameter bounds used for the MC simulations are shown in S2 Text. The systems model of the IL-4 pathway was implemented using the Simbiology toolbox in Matlab. Equations and parameters of the model are presented in S3 Text. Derivatives of Hill curves fitted to experimentally measured DCFDA fluorescence were taken to represent instantaneous intracellular ROS trends. The time derivative of the fitted Hill curve, f(x), was modified to a + bf(x), where a and b are model parameters representing baseline ROS level and a scaling factor, respectively. The modified curve was supplied to the model as an input. Simulated time course x(t) of species x was also similarly scaled to y(t) = αx + βxx(t), where αx and βx are constants defined for species x and are independent of experimental conditions. Different scaling is required for different species because the antibody used to measure each protein has different characteristics and the measured MFI scales differently to actual amount.
Initial estimates of parameters were obtained from [47,56]. A variation of the evolutionary strategy algorithm was developed in house and coded in Matlab to fit the model to experimental data using these initial estimates. The following error function was minimized to obtain the fit:
e=∑i∑j∑t(yij(t)−eij(t)eij(t)σij(t))2
where yij(t) is the scaled and shifted value (as described above) of the jth species under the ith experimental condition at time t; eij(t) represents the experimentally measured value under the same conditions and σij(t) is the standard error associated with the experimental measurement.
|
10.1371/journal.pcbi.1001094 | Informed Switching Strongly Decreases the Prevalence of Antibiotic Resistance in Hospital Wards | Antibiotic resistant nosocomial infections are an important cause of mortality and morbidity in hospitals. Antibiotic cycling has been proposed to contain this spread by a coordinated use of different antibiotics. Theoretical work, however, suggests that often the random deployment of drugs (“mixing”) might be the better strategy. We use an epidemiological model for a single hospital ward in order to assess the performance of cycling strategies which take into account the frequency of antibiotic resistance in the hospital ward. We assume that information on resistance frequencies stems from microbiological tests, which are performed in order to optimize individual therapy. Thus the strategy proposed here represents an optimization at population-level, which comes as a free byproduct of optimizing treatment at the individual level. We find that in most cases such an informed switching strategy outperforms both periodic cycling and mixing, despite the fact that information on the frequency of resistance is derived only from a small sub-population of patients. Furthermore we show that the success of this strategy is essentially a stochastic phenomenon taking advantage of the small population sizes in hospital wards. We find that the performance of an informed switching strategy can be improved substantially if information on resistance tests is integrated over a period of one to two weeks. Finally we argue that our findings are robust against a (moderate) preexistence of doubly resistant strains and against transmission via environmental reservoirs. Overall, our results suggest that switching between different antibiotics might be a valuable strategy in small patient populations, if the switching strategies take the frequencies of resistance alleles into account.
| Infections with bacterial pathogens that are resistant against antibiotics are an important cause of mortality and morbidity in hospitals. One possibility to minimize this burden of antibiotic resistance is to coordinate the use of several drugs at the level of a single hospital ward. Here, we use a computational model of a hospital ward in order to assess the performance of several such strategies that take into account the frequency of antibiotic resistance in the hospital ward. We assume that information on resistance frequencies stems from microbiological tests, which are performed routinely in order to optimize individual therapy. Thus the strategy proposed here represents an optimization at population-level, which comes as a free byproduct of optimizing treatment at the individual level. We find that in most cases our informed strategy can substantially reduce the prevalence of antibiotic resistance. We show that the performance of an informed strategy can be improved substantially if information on resistance tests is integrated over a period of one to two weeks. Overall, our results suggest that switching between different antibiotics might be a valuable strategy in small patient populations, if the switching strategies take the frequencies of resistance alleles into account.
| The increasing prevalence of antibiotic resistance in nosocomial infections is a serious threat for clinical care and an important cause for mortality and morbidlity as well as a substantial driver of health care costs [1]. Several strategies to coordinate the use of different drugs and thereby limiting the spread of antibiotic resistance have been proposed. The most prominent such strategies are Cycling (sequential use of different drugs), and Mixing (simultaneous use of different drugs in different patients). The rationale behind cycling is that strains resistant to the formerly used drug may decrease in frequency or even disappear in the off-period. Mixing, on the other hand, creates a strong environmental heterogeneity that makes it difficult for the pathogen to adapt. Concerning these two strategies, the clinical literature is inconclusive [2], [3], while the consensus in most of the theoretical literature is that mixing almost always outperforms cycling [4], [5], [6] (see however also the discussion in [7], [8], [9]). The intuitive explanation for this pattern is that mixing leads to more heterogeneity and hence hinders the adaptation of the bacterial population against the antibiotic agents [4]. Thus it seems that periodic switching of treatment regimes does not help to alleviate the burden of antibiotic resistance in hospitals. On the other hand, treatment decisions that take institution-antibiograms into account [10] may lead to a cycling-like pattern in which antibiotics are withdrawn when resistance rises and re-instituted when resistance becomes more rare. It is often recommended that resistance surveillance should be used as a guideline for empirical therapy (i.e. therapy that is initiated before microbiological results are available) [11]. However, it is difficult to disentangle the effects of this particular strategy from other simultaneously used approaches such as restriction of antibiotic usage [12].
Here, we use an epidemiological model for a hospital ward to show that contrary to the current views switching between different regimes of empirical therapy (i.e. treatment before the causative pathogen and its resistance profile are known) can reduce antibiotic resistance. The switching regime proposed here differs from the traditional ones in [5] and [4] by taking the frequencies of the resistant strain in the hospital into account. Thus in contrast with “blind” periodic switching strategies, we analyze informed switching strategies (ISS) similar to the ones that arise by antibiogram-guided therapy and show that such strategies can serve as valuable strategy to curb resistance.
In this study we use an epidemiological model (see Figure 1 and Tables 1–2) in order to consider the impact of several different informed switching strategies (ISS), which coordinate the use of two broad-spectrum antibiotics A and B on the level of a single hospital ward (see Table 3 for a mathematical characterization of the considered treatment strategies). The common element of these ISS is that if the resistance mutation against one drug is suspected to have gone extinct only this “resistance-free” drug is deployed. In this way, these strategies exploit the high frequency of stochastic extinctions of resistant strains caused by the small population sizes in hospitals. The crucial question for the practical value of an ISS is whether such a strategy can substantially reduce the burden of antibiotic resistance even if it is based on the imperfect information, which can be obtained realistically. Here we model the following realistic scenario of how such information may be obtained: Commonly, symptomatically infected patients are first treated empirically with a broad spectrum antibiotic, then a resistance profile (microbiological tests) is determined (this usually requires 1–2 days), which guides further therapy (optimally with a narrow spectrum antibiotic). In our model we assume that the ISS are based on the information obtained through these microbiological tests, which are made to guide non-empirical therapy of individual patients. Therefore, the model makes the realistic assumptions that i) the information on the resistance status of a symptomatically infected patient is only available after a delay of 2 days on average and ii) that upon the availability of these test-results, the patient is immediately put on a narrow-spectrum therapy regimen against which the pathogen is susceptible. We consider two main classes of ISS: those, which are based on a snapshot of resistance frequencies (i.e. on the frequency of resistance mutations among the infected patients that are currently in the ward and of which microbiological results have been obtained) or those, which integrate information of resistance over a certain time window. Whereas the first class of strategies is simpler to understand from a population biological point of view, we will argue that strategies of the second class are recommendable for clinical practice. We use two measures to assess the success of the ISS (see method section): the prevalence of resistance mutations in the ward and the number of inappropriately treated patients. In fact, it has been shown that inappropriate initial (empirical) treatment increases mortality since severely infected patients might die before treatment can be adjusted [13], [14]. For both measures the success of an alternative strategy is measured relative to mixing: if mM and mA denote the value of the measure for mixing and the alternative therapy, respectively, then the success of the alternative therapy is quantified by Δm = (mA−mM)/mM. Thus, the more negative Δm, the better the strategy.
In order to study the population biological basis of the ISS, we start by considering the simpler snapshot-based ISS: The negative-frequency-dependent informed switching, ISS-, and the mixing-like informed switching, ISSM. Both strategies deploy only one drug if, among the infected patients in the ward whose resistance status is known, there are both no reports of strains resistant against this drug and at least one report of strains resistant against the other drug (see Table 1). If no resistance mutation is present both drugs are used at equal frequencies. The two strategies differ with respect to their deployment of antibiotics when both resistance mutations are present. In this case, ISS- deploys both antibiotics inversely proportional to the momentary frequency of the corresponding resistance mutations, whereas ISSM deploys both antibiotics at equal frequencies. We find that informed switching clearly outperforms both mixing and periodic cycling (Figure 2). By contrast the difference between the two strategies ISS- and ISSM is marginal (Figure 2). Thus the central aspect of the strategies is the coordinated deployment of antibiotics in those phases when one resistance mutation is extinct in the hospital. This fact indicates that the success of ISS is essentially a stochastic phenomenon, as extinctions are chance effects facilitated by small hospital sizes. In accordance with this interpretation and with earlier work[8], [9], we find that, in the deterministic version of our model, the ISS- strategy leads to no substantial improvement over mixing (results not shown, but see section: effect of population size).
Considering only the current symptomatically infected patients with a microbiological test leads to an imprecise estimate of the resistance frequencies in the ward, as these patients represent only a small fraction of all carriers. However, as the detection of one infection with a resistant strain is indicative of other such infections (which often persist after the detected case has been cleared) the imprecision can, in part, be compensated by integrating the information over several time points. The strategies ISSK (with K = 4,7,14,30,60,90) integrate information over several time-points in the following way: A resistant strain is considered extinct if no symptomatically infected patient with a microbiological test that detected this strain has been in the hospital for the last K days (see Table 3). Thus ISSK integrates the infection status (of patients with a test) over the last K days. Again, if only one of resistance mutations is considered extinct according to the above criterion, then only the corresponding drug is used. If both or no resistance mutation is considered extinct both drugs are used at equal frequencies. The disadvantage of these ISSK is that they require the choice of the length of the integration time window. An alternative way to integrate information on resistance prevalence over several days without this drawback is the following (ISSLast): If at a given time-point the drug resistance-mutations against A and B have been last detected (among the patients with known resistance status) tA and tB days ago, then use that drug for which this time-span is larger. If tA and tB are equal (in particular if both resistance mutations are simultaneously present at the given time point) then both drugs are used at the same frequency. We find that both ways of integrating the information on resistance frequencies can indeed substantially improve the performance of the ISS and that the overall best results can be achieved for ISSLast and for the ISSK with K = 7 or 14 (Figure 3). Accordingly we will focus on these two optimal strategies ISS7 and ISSLast, when assessing in the following the robustness of our results with respect to several important aspects of the model.
One of the major factors determining the success of the ISS is the frequency of resistant strains among incoming patients. The ISS are based on extinctions of the resistance mutations, which is counteracted by the frequent reintroduction of these strains. Accordingly, we find that the success of ISS decreases with increasing influx of resistant strains (Figure 4). However, the advantage of ISS remains substantial unless an unrealistically large fraction of incoming patients carries the resistant strain of the pathogen.
The progression rate is important mainly because it affects the fraction of symptomatically infected patients and therefore the frequency at which the antibiotic is used. With increasing progression rate one would expect an increasing use of the antibiotic and hence an increasing impact of the applied strategy. Indeed, we find that both ISS7 and ISSLast are especially effective for fast progressing diseases (Figure 5). However, even for moderate and low progression rates, the ISS still confer a substantial advantage in terms of reducing inappropriately treated patients.
The preexistence of doubly resistant strains among incoming patients has been argued to render treatment strategies futile, i.e. strategies perform equally bad when doubly resistant strains are present [4], [5]. To some extent, this also applies to the ISS discussed above. Specifically we find that the beneficial impact of the ISS decreases substantially as the fraction of incoming patients with a doubly resistant strain increases (Figure 6). The figure, however, also shows that the ISS can still substantially reduce the prevalence of drug resistance even if the frequency of the doubly resistant strain is as high as 5% among colonized and infected patients.
As the success of the ISS is essentially a stochastic effect, one would expect that it becomes weaker in the presence of an environmental reservoir. This is because an environmental reservoir exhibits a slower turnover of strains (see Table 1) and hence reduces the extinction risk of resistant strains, i.e. the reservoir can act as a “seedbank” for resistant strains. Indeed, we find that the ISS perform slightly worse in the presence of such a reservoir (Figure 7). However this decrease in the strategies' efficiency is very weak and the improvement achieved by applying ISS7 remains substantial even if transmission is uniquely mediated by an environmental reservoir. This indicates that even if the turnover rate of strains in the ward is reduced to that of the reservoir (here: 1/(30 days)) stochastic effects are strong enough to ensure the efficiency of the ISS.
The way in which the benefit conferred by the ISS depends on population size confirms the stochastic nature of this effect: As expected the benefit essentially disappears for very large population sizes when stochastic effects are expected to be small (Figure 8). Regarding the fraction of inappropriately treated patients the magnitude of the benefit decreases monotonically with increasing population size. Regarding the prevalence of resistance, we observe however a slight increase in the magnitude when increasing the population from 20 to 50. Although it is not entirely clear what causes this increase, it might be that for very small population sizes the subpopulation of patients with microbiological tests gives a very inaccurate picture of the resistance prevalence even if integrated over time. An alternative explanation is that in very small populations, extinction events impede resistance emergence, such that resistance is infrequent regardless of the treatment regimen. In this case, the additional reduction of resistance may be small. However, as soon as the population size exceeds 100 the magnitude of the effect decreases strongly (with regard to both measures) and becomes negligible at population sizes of 500 and beyond.
For simplicity, we assumed so far that both resistance genes are symmetrical, i.e. that resistance-costs and prevalence among incoming patients carrying the resistant strain are identical for drug A and B. If this assumption is relaxed, the optimal mixing strategy does not deploy the two drugs at equal frequencies but gives preference to the drug whose resistance mutation is more frequent among incoming patients and less costly. We make the realistic assumption that these two properties coincide: i.e. the less costly mutations are more frequent. We find that in such an asymmetric scenario, the ISS still considerably outperform even the optimal mixing strategy (Figure 9). Interestingly, this scenario provides the only example (apart from a very narrow parameter region in Figure 3) in which the ISS7 strategy can outperform the ISSLast strategy. However, the difference is rather small and depends sensitively on choosing the correct mixing strategy for those phases in which the ISS7 deploys both drugs simultaneously (see definition of ISS in Table 3).
Previous theoretical studies suggest that from the point of view of preventing the spread of resistance mutations, mixing strategies perform at least as good as strategies that switch between different antibiotics [4], [5]. However, selecting treatment based on cumulative ward antibiograms has been shown to increase adequate therapy [15]. Here, we have shown that such a strategy does not only benefit the single patients receiving appropriate therapy, but may also be used to counteract the spread of resistance. In the highly stochastic setting of small hospital wards, mixing (random treatment) can be substantially outperformed by informed switching strategies, which take the frequency of antibiotic resistance mutations into account. Factors that promote the success of such ISS include the absence of multiply resistant strains and a low frequency of singly resistant strains among the incoming patients. However, even if these conditions are not fulfilled, ISS can still substantially alleviate the burden of drug resistance. Moreover, we found that the magnitude of the effect of ISS decreases with increasing fitness cost, remains however substantial even for large fitness costs (results not shown). Given that, especially in the long term, resistance carries very small costs if any [16], the default fitness cost chosen here (s = 0.1) can be considered as an upper bound yielding thus a conservative assessment of the effect of ISS.
In our view, the most promising version of an informed switching strategy is ISSLast. Apart from the fact that this strategy outperformed the other versions in almost every setting, it has the advantage that its implementation would be relatively simple: Essentially, it would only require that the dates at which resistant strains have been detected in a given ward are recorded, and that for every new patient that drug is used for which the last isolation date is most distant. The main case in which we found this strategy not to be the best choice, was if among incoming patients resistance mutations against one drug was much more common than resistance mutations against the alternative drug. In this situation ISSLast clearly outperformed mixing. However, it was slightly worse than the following alternative strategy: Always using the antibiotic less common among incoming patients, except if a resistance mutation against this drug has been detected in the past seven days, in which case only the alternative drug has to be deployed (formally this corresponds to ISS7 with M = 0, see Table 3). However the additional improvement conferred by this strategy was modest and does in our view not outweigh the larger simplicity and robustness (independence of an integration time-window) of ISSLast.
The benefit conferred by the ISS is a result of the underlying stochasticity of resistance prevalence in the hospital. This is demonstrated by our finding that the magnitude of the effect becomes negligibly small as soon as the population size is above 500–1000. This is consistent with the findings of [8], [9] which found in a deterministic model no (or no substantial) improvement is conferred by an “adaptive” strategy similar to the ISS used here. The fact that ISS are very effective for small population sizes but have limited success at large population sizes, suggests that switching strategies should be implemented at the ward level rather than at the hospital level, especially in large hospitals.
The information underlying the switching strategies considered here is a byproduct of microbiological resistance tests, which are usually done in clinical practice in order to optimize individual treatments. While the recent HICPAC guidelines, “Management of Multidrug-Resistant Organisms in Healthcare Settings,” [11] recommend at least annual updates, the strategies proposed here are not based on discrete updates. Instead the information acquired from microbiological tests would have to be integrated into the decisions as it is generated in the course of optimizing individual treatments. Also, in accordance with current recommendations [17], [18] we follow resistance in one single hospital ward, not in the whole institution, such that the cumulative antibiogram of the respective wards should be chosen for informed switching. As the success of an informed switching strategy depends on the quality of the information on the frequencies of resistance genes, the success of the strategy can be further improved by sampling also asymptomatically infected patients (results not shown). However, we think that the strategies we have proposed in this study represent the most realistic option, given that they come at no additional cost other than compiling the available data from microbiological tests. Furthermore, the inclusion of isolates from asymptomatic patients is not recommended by the Clinical and Laboratory Standards Institute [CLSI] [19]. Thus although the extent of information is important for informed switching strategies, a realistic and often available degree of knowledge seems to be sufficient for a successful implementation of the strategy.
In summary, we have shown that coordinated informed switching of the antibiotic deployed in a hospital ward can outperform mixing as a strategy to limit the spread of antibiotic resistance of nosocomial pathogens. This theoretical result is especially interesting, since the impact of surveillance-guided therapy is often difficult to assess [12].
We consider a compartmental epidemiological model that describes a single hospital ward. We further consider two empirical broad spectrum antibiotics, to which we refer as drug A and B. Accordingly, we follow four genotypes: wild type (sensitive to both drugs), resistant against A and sensitive to B, resistant against B and sensitive to A, and resistant against both drugs. Patients are classified as being susceptible (S), colonized (C; i.e. asymptomatic carriers), or infected (I; i.e. symptomatic carriers). Furthermore, the compartments are subdivided according to the treatment status and (for I & C) according to the genotype of the carried strain. In addition, we follow a pathogen reservoir outside the patients (E), which describes environmental contamination but may also describe the dynamics resulting from the transient colonization of health care workers; although these are not modeled explicitly. Finally, we assume that symptomatically infected patients undergo a microbial test (with a rate tR) after which they are switched to an appropriate narrow spectrum antibiotic for which we assume that resistance is negligible. These test-results provide the information on resistance frequencies upon which the ISS are based.
Figure 1 summarizes the population dynamics of the model for a single strain and Table 1 lists the parameters and their default values (which are used if not declared explicitly otherwise). We used parameter values from clinical literature as far as they are available.
We assume a fixed number of 20 beds in the hospital ward. As soon as a bed is free, patients of all classes carrying pathogens of all genotypes may be admitted at frequencies that are assumed to be constant over the observed timeframe (see Table 1). The proportion of incoming patients belonging to the three main compartments S, C and I is determined by the parameters pS, pC and pI (see Table 1). The proportion of patients carrying the genotypes wt, A, B, and AB is given by the parameters pAB, pA0 and p0B. Upon admission patients are not treated unless they are symptomatic carriers (i.e. we focus on non-prophylactic treatment). Upon transition to the “infected” compartment all patients are treated with a broad-spectrum antibiotic according to the current treatment strategy (Mixing, Cycling or informed switching). After clearance of the pathogen, treatment is ceased at a rate of 1/5 d−1.
Here we consider a stochastic version of the model described above. Specifically, the state of the patient-population in the ward is given by the discrete variables referring to the number of susceptible (S) asymptomatically infected (C) and symptomatically infected (I) patients with treatment status y and infection status x. The infection status can be either “infected with the strain susceptible against both drugs” (x = 00), “infected with the strain susceptible against A but resistant against B” (x = 0B), “infected with the strain susceptible against B but resistant against A” (x = A0), or “infected with the strain resistant against both drugs” (x = AB). The treatment status can be either “treatment with no drug” (y = 00), “treatment with drug A” (y = A0), “treatment with drug B” (y = 0B), or “treatment with a narrow spectrum antibiotic” (y = N). We assume that the narrow-spectrum antibiotic is only administered after microbiological tests and that hence the infection status of patients in this treatment class is known. The state of the environmental colonization is given by the density Ex of the strain x in the environment.
The patient population is simulated stochastically according to Gillespie's Direct Algorithm[20]. The full characterization of this model is given by Table 2, which lists the different events (and rates) that constitute the model. The following points should be noted concerning these events:
As the dynamics of the environmental compartment is not directly affected by the fluctuations of the patient population, the variables Ex describing the environmental reservoir are updated according to the ODE system
The success of the treatment strategies (summarized in Table 3) is measured by their impact on the prevalence of resistance given by
(note that double resistant strains are counted twice) and by their impact on the number of inappropriately treated patients given by . (Note that for this measure we take only symptomatically infected patients into account, because it is in that group that inappropriate treatment will have the most severe clinical consequences).
|
10.1371/journal.ppat.1004488 | Acidification Activates Toxoplasma gondii Motility and Egress by Enhancing Protein Secretion and Cytolytic Activity | Pathogenic microbes rely on environmental cues to initiate key events during infection such as differentiation, motility, egress and invasion of cells or tissues. Earlier investigations showed that an acidic environment activates motility of the protozoan parasite T. gondii. Conversely, potassium ions, which are abundant in the intracellular milieu that bathes immotile replicating parasites, suppress motility. Since motility is required for efficient parasite cell invasion and egress we sought to better understand its regulation by environmental cues. We found that low pH stimulates motility by triggering Ca2+-dependent secretion of apical micronemes, and that this cue is sufficient to overcome suppression by potassium ions and drive parasite motility, cell invasion and egress. We also discovered that acidification promotes membrane binding and cytolytic activity of perforin-like protein 1 (PLP1), a pore-forming protein required for efficient egress. Agents that neutralize pH reduce the efficiency of PLP1-dependent perforation of host membranes and compromise egress. Finally, although low pH stimulation of microneme secretion promotes cell invasion, it also causes PLP1-dependent damage to host cells, suggesting a mechanism by which neutral extracellular pH subdues PLP1 activity to allow cell invasion without overt damage to the target cell. These findings implicate acidification as a signal to activate microneme secretion and confine cytolytic activity to egress without compromising the viability of the next cell infected.
| Toxoplasma and related parasites including those that cause malaria are obligate intracellular pathogens that replicate within a specialized compartment termed the parasitophorous vacuole. To infect new host cells these parasites must first escape from the parasitophorous vacuole and other limiting membranes of the currently infected cell. Escape, or egress as it is often called, depends on the timely release of adhesive proteins and lysis factors from secretory organelles called micronemes. Although this secretory event is crucial for egress, the natural environmental cues that trigger microneme secretion remain poorly defined. Here we discover that acidification of the parasitophorous vacuole is sufficient to trigger microneme secretion and promote the activity of a lysis factor called PLP1. We also show that pH-neutralizing drugs inhibit egress and provide evidence of parasitophorous vacuole acidification approximately coinciding with parasite egress from infected host cells. The findings support a working model in which acidification activates microneme dependent motility and lytic activity to execute egress and destruction of infected cells. The results also provide insight into how PLP1 lytic activity is stimulated during egress in an acidic environment and subsequently suppressed by the neutral extracellular environment, thus permitting cell invasion with minimal damage to the next target cell.
| Infectious microorganisms experience diverse microenvironments during infection of a host. Such pathogenic microbes utilize specific cues to assess the environment and trigger an appropriate response that aids in their survival. For example, malaria parasites (Plasmodium spp) sense xanthurenic acid and a drop in temperature upon infecting a mosquito to trigger male gamete exflagellation for sexual reproduction [1]. The related apicomplexan parasite Toxoplasma gondii is thought to utilize the rapid build up of self-made abscisic acid as an intrinsic cue to exit from infected cells [2], an event termed egress. However, other studies have shown that T. gondii can respond to diverse environmental changes including the loss of host cell viability [3], [4] or an increase in reduction potential [5] to trigger egress. Most apicomplexan parasites including, malaria parasites and T. gondii, replicate within a membrane bound compartment termed the parasitophorous vacuole (PV). The PV microenvironment is presumed to be similar to that of the host cell cytoplasm based on the detection of a hypothetical pore that permits fluorescent dyes of <∼1,300 Da freely pass across the PV membrane [6]–[8]. However, the identity of this pore has not been reported for malaria parasites or T. gondii. Also, the malaria PV has much higher Ca2+ concentrations (∼40 µM) than the cytosol of infected erythrocytes (100 nM) [9], indicating a restricted flow of Ca2+. Another report suggested that a membrane potential across the T. gondii PV membrane is maintained by proton and potassium P-type ATPases [10]. The limited and, in some cases, apparently discrepant findings for the PV highlight the need to better understand this microenvironment and how it changes during key events in the life cycle.
T. gondii persists as a chronic infection in an estimated one third of the global human population, causing opportunistic disease in a subset of those infected. It also produces disease in domestic livestock, wild mammals and birds. In humans, the parasite is especially virulent when acquired congenitally or in reactivated disease, which occurs when the host becomes immune-suppressed. Pathogenesis is driven by iterations of the tachyzoite lytic cycle, which includes host cell invasion, replication within the PV, host cell egress and migration to infect a neighboring cell.
Parasite motility and host cell invasion require the coordinated action of parasite proteins secreted from apical secretory organelles called rhoptries and micronemes. Micronemes are Ca+2-regulated secretory organelles that are controlled by phosphorylation-based signaling pathways (reviewed in [11]). Potassium and Ca+2 ion fluxes have been shown to influence parasite motility and egress (reviewed in [12]). High K+ concentrations, mimicking the intracellular state, inhibit microneme secretion and motility, and a drop in external K+ triggers microneme secretion [3]. Although the precise mechanisms of K+ sensing by the parasite are still emerging, it is known that intracellular Ca2+, phospholipase C and at least two Ca2+-dependent protein kinases (CDPKs) are involved [13]. Other studies have shown that Ca2+ release from intracellular stores regulates parasite motility by activating the glideosome and apical secretion of transmembrane micronemal adhesins, which engage the motor to transduce power into motion [14]–[16]. Parasite sensing of environmental K+ is thought to ensure that the motility system is in neutral during intracellular replication, but is available for engagement to rapidly exit from an infected cell. Malaria sporozoites also respond to K+ fluxes [17], implying a conserved mechanism for environmental sensing and regulation of motility.
Earlier investigators of tachyzoite motility additionally found that motility is pH-dependent [18]. Alkaline conditions inhibited motility and acidic buffers induced motility. Here we demonstrate that pH-dependent motility involves the activation of microneme secretion. We also implicate acidification as an enhancer of egress both by promoting microneme secretion and enhancing cytolysis by perforin-like protein 1 (PLP1), a pore forming protein required for efficient egress. Our findings suggest that pH-dependent microneme secretion and activation of PLP1 is another layer of regulating parasite behavior to promote parasite success in changing environments.
Early studies on tachyzoite motility showed that parasite gliding is inhibited by high concentrations of K+ and alkaline pH and promoted by acidic pH [18]. We revisited the effect of pH on gliding by purifying parasites in high K+, alkaline buffer (pH 8.4), switching to the same buffer at neutral or low pH and observing motility over time. Low pH stimulated motility in >90% of observed parasites and motility was sustained in a majority of parasites for at least 15 min (Figure 1A), confirming previous findings [18]. Switching from alkaline to neutral pH led to no significant change in motility.
Since parasite motility is linked to microneme secretion, we tested the effect of pH on microneme secretion by incubating parasites in buffer of varying pH and probing the secreted fraction via immunoblots for the microneme adhesive protein MIC2, a galactose-binding protein MIC4 and PLP1. More microneme secretion was detected at low pH compared to neutral and alkaline pH (Figure 1B–C). In contrast, secretion of dense granule proteins GRA1 and GRA4 was largely independent of pH. Low pH induction of microneme secretion was more effective than stimulation with 1% ethanol, which is commonly used to activate microneme discharge (Figure 1C) [19]. Low-pH induced microneme secretion was sensitive to the Ca2+ chelator, BAPTA-AM, indicating dependence on intracellular Ca2+ (Figure 1D). We also noted pH-dependent differences in the proteolysis of MIC2 and MIC4, which are processed by the micronemal serine protease SUB1. Processing was inhibited by low pH, suggesting the protease SUB1 functions optimally at neutral pH (Figure 1D,E). Acidic pH stimulated microneme secretion despite the presence of high K+, whereas minimal secretion occurred at neutral-alkaline pH in a high K+ environment (Figure 1E). These findings reveal that low pH can overcome the normally microneme suppressive effects of a high K+ environment, which the parasite experiences within an infected cell. The results further suggest that low pH activation of microneme secretion contributes to pH dependent motility.
Having established that low pH overcomes K+ suppression of microneme secretion, we reasoned that exposure of intracellular parasites to low pH should activate microneme secretion and egress despite a high K+ environment. We tested this by permeabilizing infected cells with digitonin in high K+ buffer with varying pH. Minimal egress was observed under alkaline (pH 8.4) or neutral (pH 7.4) conditions (Figure 2A), consistent with a previous report [3]. On the other hand, acidic pH induced parasite egress in a manner related to the degree of acidification. Together with the above data, these findings suggest that low pH is sufficient to stimulate microneme secretion and initiate motility and parasite egress.
Next we tested the extent to which the pH of the PV changes during induced egress and replication by expressing a pH sensitive GFP variant, superecliptic pHluorin, in the parasitophorous vacuolar space of plp1ko parasites (plp1kosepH). plp1ko parasites were used instead of WT parasites to avoid losing the probe from the vacuolar space upon secretion of PLP1 during egress. Superecliptic pHluorin is highly fluorescent at neutral pH but is quenched at low pH [20]. We verified the pH-dependence of fluorescence by measuring the signal in superecliptic pHluorin plp1ko infected cells lysed at low or neutral pH and in live infected cells. Fluorescence was completely quenched in cells lysed at low pH (Figure 2B). A modest but significant drop in fluorescence was detected upon treating infected cells with Ca2+ ionophore (Figure 2C). This decrease in fluorescence was reversed with NH4Cl, a weak base, which accumulates in acidic compartments, raising luminal pH. DCCD, a P-type ATPase inhibitor also partially reversed the drop in fluorescence (Figure 2D). These findings suggest a moderate decrease in vacuolar pH occurs upon egress induction. To observe changes in PV pH during parasite replication and spontaneous egress, we measured the fluorescent signal of live superecliptic pHluorin plp1ko infected cells over the course of intracellular replication,. If the PV pH is neutral, NH4Cl treated cells are expected to have a similar amount of fluorescence as cells in buffer alone. Conversely, if vacuolar pH is acidic, NH4Cl treated cells should show a stronger signal than untreated cells due to the pH neutralizing effect of treatment. Super-ecliptic pHluorin signals increased identically during parasite replication until ∼30 h post-inoculation. At this time point, however, the curves began to diverge, with a substantial suppression of fluorescence that was reversed by NH4Cl treatment (Figure 2D). Microscopic examination of the infected monolayers indicated that most of the parasites remained intracellular or at least within spherical structures representing failed egress events until after 44 h post-infection. These findings imply a population-scale decrease in vacuolar pH occurs late in the replication cycle, prior to or during spontaneous parasite egress.
Next we reasoned if an acidic pH contributes to parasite egress, induced egress should be sensitive to pH neutralization. Consistent with this, we found that parasite egress is suppressed by NH4Cl treatment upon stimulation with Ca2+ ionophore or dithiothreitol (DTT), an egress inducer that activates a PV nucleotide triphosphatase [5] (Figure 3A–B). PV acidification could occur through passive accumulation of metabolic wastes, or active delivery of protons by a proton-pump. We tested the latter by determining the effect of H+-ATPase inhibitors on induced egress. We found that parasite egress was sensitive to the P-type ATPase inhibitor, DCCD (Figure 3C–D), but not the V-type ATPase inhibitor bafilomycin or the H+/K+ exchange inhibitor omeprazole (Figure S1). Egress induced by the phosphodiesterase inhibitor Zaprinast, which triggers microneme secretion and egress via activation of protein kinase G [21], was also sensitive to NH4Cl or DCCD treatment (Figure 3D). The above pH-neutralizing agents did not significantly alter motility (Figure S2A, B) or microneme secretion (Figure S2C) induced by the Ca2+ ionophore in a neutral buffer, rendering it unlikely that that they affected egress by impairing microneme secretion or the parasite motor system. Together these findings suggest a role for acidic pH and a P-type ATPase in egress.
Next we tested the extent that pH-neutralizing agents affect the activity of important egress effectors during induced egress. Previous work has shown the microneme protein PLP1 to be crucial for rapid parasite egress [22]. To determine whether the reduced parasite egress was due to an inhibition of PLP1 activity occurring with pH-neutralization, we tested the effect of these treatments on egress-associated membrane permeabilization. Since PLP1 is necessary for membrane damage during egress [22], we used the membrane impermeable dye propidium iodide (PI) to assess membrane permeabilization upon treating infected cells with Ca2+ ionophore to induce microneme secretion. Parasites were immobilized with the F-actin inhibitor cytochalasin D (CytD) to prevent membrane damage due to actin-myosin dependent parasite motility. Whereas vehicle-treated cells maintained intact membranes, ionophore treatment lead to the permeabilization of the majority of WT-infected cells. This activity was PLP1-dependent since no significant permeabilization was observed in ionophore-treated plp1ko-infected cells (Figure 4B, D). Ionophore-induced membrane permeabilization was also sensitive to both NH4Cl and DCCD, suggesting that pH-neutralization suppresses PLP1 activity (Figure 4A, C). As indicated above, NH4Cl and DCCD did not inhibit parasite gliding motility or microneme secretion, ruling these out processes as possible off-targets of treatment (Figure S2).
Since several members of the protein superfamily to which PLP1 belongs are regulated by pH, we used recombinant PLP1 to determine the extent that its activity is pH-dependence. Using hemolysis of erythrocytes as a measure of PLP1 activity, we observed increased PLP1 cytolytic activity beginning at pH 6.4 and peaking at pH 5.4 (Figure 5A). Approximately 7 times more PLP1 activity was seen at pH 5.4 than pH 7.4. The pH dependent profile of PLP1 lytic activity was similar to that of listeriolysin O (LLO), which Listeria monocytogenes uses to escape from the acidifying primary vacuole after cell entry [23]. The PLP1 cytolytic profile was distinct from streptolysin O (SLO), which displayed a similar amount of lytic activity across a broad range of pH values. LLO and SLO are cholesterol dependent cytolysins (CDC), which are members of the CDC/membrane attack complex-perforin superfamily that includes PLP1 [24].
We next tested the effect of pH on PLP1 membrane binding using erythrocyte ghosts as a model membrane. We observed that PLP1 membrane binding activity mirrors the lytic activity with more binding at acidic pH than neutral pH (Figure 5B). Previous work demonstrated that PLP1 N- and C-terminal domains both contain membrane-binding activity [25]. Subsequently, we tested for pH-dependent membrane binding of full-length, mature PLP1 and the N-and C-terminal domains by sucrose density gradient membrane flotation. At low pH the majority of mature PLP1 was membrane bound (fractions 1–6), whereas at neutral pH most of the PLP1 remained unbound (fractions 7–12). Membrane binding by the N-terminal domain was not significantly affected by pH, whereas the C-terminal domain showed increased binding at low pH (Figure 5C). These findings are consistent with a previous report showing a dominant role for the C-terminal domain in membrane binding [25]. Together the findings suggest that PLP1 cytolytic activity is pH-dependent at the membrane-binding stage.
We next investigated the effect of low pH on parasite invasion. Although our above findings predict that the low pH stimulation of microneme secretion should augment parasite invasion, it should also enhance PLP1 activity if invading parasite secretes it. Parasites were purified in high K+ buffer and allowed to settle on host cells prior to switching the buffer to either DMEM, or DMEM buffered to pH 5.4 or 7.4 and incubating at 37°C for 2 min, then washed and fixed. We found that low pH substantially increases the amount of attached parasites and invaded parasites (Figure 6A), likely due to increased microneme secretion and motility. Interestingly, acidic conditions appeared to reduce the proportion of invaded parasites relative to attached parasites
To test if PLP1- and pH-dependent membrane damage occurs during invasion we pre-loaded host cells with calcein-AM, and WT and plp1ko parasites were allowed to settle on host cells in high K+ Endo buffer to block microneme secretion and invasion. Endo buffer was subsequently removed and replaced with DMEM-like buffer at pH 5.4 or 7.4. After 10 minute's incubation, supernatant was collected and calcein release from host cells was quantified by fluorometry. Results were normalized to detergent lysed cells set at 100%. Host cells incubated with WT parasites at neutral pH did not show a density dependent increase in calcein release, consistent with parasite invasion in the absence of overt damage to host cells (Figure 6B). However, host cells incubated with WT parasites at low pH resulted in a significant parasite density-dependent increase in calcein release, a trend that was not seen for plp1ko parasites at neutral or acidic pH, indicating a requirement for PLP1 in pH-dependent cell damage. Together these findings suggest that low pH stimulates invasion but also promotes PLP1-dependent wounding of target host cells.
Microneme secretion and motility contribute to host cell invasion and egress, but are quiescent during intracellular replication. Earlier work showed that a high K+ environment mimicking the intracellular milieu suppresses parasite motility and egress [3], [18]. A previous report also suggested that T. gondii tachyzoite motility is regulated by pH, with moderately acidic conditions strongly enhancing motility [18]. Here we show that an acidic environment promotes microneme secretion, suggesting a mechanism for pH augmentation of motility. We also found that parasite exposure to an acidic environment overcame high K+ inhibition of microneme secretion, indicating that acidification of the intracellular environment can trigger microneme secretion and egress even if the high K+ intracellular milieu is intact. Consistent with this notion, we found that low pH can drive parasite egress in a high K+ environment. Using the pH-sensitive probe superecliptic pHluorin, we observed that egress induction with A23187 leads to a modest decrease in signal, reflecting a shift towards acidic pH. It remains unclear if A23187-induced PV acidification is due to effects of Ca2+ flux on proton pumps, due to signaling or secretion events downstream of Ca2+ flux, or both. Regardless, using the same probe we also detected a reduction in PV pH late in the replication cycle, possibly near the time of egress. Although these experiments were done with plp1ko parasites to minimize loss of the probe from PLP1 membrane damage, experiments showing inhibitory effects of pH neutralization on egress were done with WT parasites, indicating findings are not strain specific.
Earlier investigations determined that upon host cell entry, the PV avoids fusion with endosomes, preventing acidification and degradation of the invaded parasite [26]. Previous efforts to examine vacuolar homeostasis also found a free-flow of small molecules (<1,300 Da) between the PV and host cytosol [7], implying an equivalent neutral pH in both sites. However, these studies were carried out at 24 h or less post-inoculation. Our results concur with the vacuolar pH being neutral at this time point during replication, but suggest that acidification occurs later during the replicative cycle, perhaps immediately prior to egress. Our attempts to directly measure the intravacuolar pH during spontaneous egress were compromised by low expression, poor signal-to-noise and toxicity of ratiometric pHluorin, and difficulties capturing rare egress events with high spatiotemporal resolution. Breaching these obstacles will require substantial improvements in genetically encoded, ratiometric pH biosensors and imaging technologies.
Further support for a role of pH in egress came from showing that treating infected cells with pH-neutralizing agents reduced parasite egress regardless of the egress inducer. Treatment with the P-type ATPase inhibitor, DCCD, impaired induced egress, implying a role for active proton flux during egress. P-type ATPases, also called E1-E2 ATPases, are an evolutionarily conserved group of cation or lipid transporters. The parasite expresses two P-type ATPases (TgPMA1 and TgPMA2) of the Type IIIA subfamily, which includes H+-ATPases of plants and fungi [27]. TgPMA1 (TGME49_252640) localizes to the parasite plasma membrane and is substantially upregulated in bradyzoites during the chronic stage of infection [27]. Genetic ablation of TgPMA1 led to a reduction in tachyzoite to bradyzoite conversion in vitro [27]. TgPMA1 deficient parasites had no reported growth defect as tachyzoites; however, egress was not assessed. TgPMA2 (TGME49_284598) is expressed in tachyzoites and bradyzoites [27], though its function remains to be characterized. The orientation of TgPMA1 and TgPMA2 is such that they are expected to pump protons from the parasite cytoplasm into the PV, consistent with a possible role in acidification of the PV. It is also possible that a DCCD-sensitive host P-type H+-ATPase contributes to the putative acidification of the PV. At least one plasma membrane multipass protein is known to occupy the nascent PV membrane after parasite invasion [28], raising the possibility that other such host proteins including ion pumps could reside in the PV membrane. The accumulation of excreted metabolic waste products in the PV or infected cell late during intracellular replication could also contribute to acidification. Also, since it remains unclear how abscisic acid triggers egress, the possibility that it contributes to PV acidification warrants consideration.
Other pore forming proteins including LLO are regulated by pH. LLO pH-sensitivity is mediated by three acidic amino acids in the transmembrane helices, which are part of domain 3 [29], [30]. These residues are thought to repel one another at neutral pH leading to protein denaturation and loss of activity. Our findings suggest that PLP1 uses a distinct mechanism, however, since low pH augments PLP1 membrane binding via the C-terminal domain, which is positionally and functionally equivalent to domain 4 of LLO. Although the N-terminal domain also has membrane-binding activity, its membrane binding was not significantly affected by pH. Our findings also indicate that PLP1 membrane binding is pH-dependent, but it remains possible that subsequent steps such as oligomerization and membrane insertion are also pH regulated. The pore forming activity of human perforin is pH-dependent at a step following membrane binding [31]. Thus, pore-formation may be regulated by different environmental conditions at multiple steps of pore-formation. Future structural comparisons of PLP1 with LLO and perforin might illuminate divergent features that modulate the activity of these proteins in the varied environments in which they function.
The increased microneme secretion at low pH also leads to dramatically higher parasite attachment and invasion. However, this augmentation of parasite association with the host cell comes at the expense of increased cell wounding. It is difficult to distinguish the extent to which increased microneme secretion versus increased PLP1 activity contributes to the membrane damage observed. The effect is probably due to a combination of the two factors. Regardless, membrane permeabilization was entirely PLP1-dependent since it was not detected in the PLP1-deficient strain at either pH.
Collectively, our findings suggest a working model in which acidification of the PV during late stage parasite replication or immediately preceding egress augments both microneme secretion and PLP1 activity. That pH 5.9–6.4 is sufficient to promote microneme secretion and PLP1 membrane binding implies that moderate acidification is adequate to enhance microneme- and PLP1-dependent egress. PV acidification is capable of overcoming K+-dependent suppression of microneme secretion and motility, thus it can act as a primary trigger in the high K+ intracellular environment of a viable cell. Initial pH dependent release of PLP1 is expected to cause membrane damage resulting in a loss of K+ from the infected cell, thus further promoting Ca2+signaling, motor activation and microneme secretion to accelerate membrane damage, motility and egress. It should be noted, however, that while motility independent disruption of the PVM during egress is strictly PLP1-dependent [22], the contribution of PLP1 to disruption of the host plasma membrane remains unknown. The model further predicts that the low K+, neutral pH environment experienced by extracellular parasites after egress augments microneme secretion but suppresses PLP1 activity. Thus the extracellular environment is conducive to microneme-based motility, parasite attachment and invasion, but it suppresses PLP1-dependent damage to the membrane of the target cell, thereby permitting correct formation of the PV during invasion. This model does not exclude other levels of regulation e.g., differential accessibility of PLP1 receptors or functionally distinct subpopulations of micronemes [32], which could occur in parallel to ensure maximal PLP1 activity during egress while minimizing membrane damage during invasion. It is also expected that the proposed pH regulatory mechanism functions in parallel with other sensory and signaling pathways to coordinate egress under different circumstances.
Student's t-tests were used to assess differences in quantitative experiments, which were performed at least three times, with technical replicates within each experiment in some cases. Qualitative experiments were performed at least twice and often three to four times.
T. gondii tachyzoites were maintained in human foreskin fibroblasts (HFF) as previously described [33].
Gliding experiments were conducted using a Zeiss temperature/CO2/humidity modulation system on a Zeiss Axio inverted microscope equipped with an Axiocam MRM CCD camera. For pH-dependent gliding, parasites were filter purified in high K+ buffer (145 mM KCl, 5 mM NaCl, 1 mM MgCl2, 15 mM MES, 15 mM HEPES, pH 8.4) and allowed to settle in a glass-bottom petri dish. After parasites had settled, images were collected every 100 ms for 6 min. Then buffer was exchanged for the same buffer adjusted to pH 7.4 or 5.4 and images were collected every 100 ms for 30 min. For inhibitor gliding experiments, parasites were filter-purified in PBS, resuspended and allowed to settle in HBSSC (Hanks buffered salt solution, 10 mM HEPES, 1 mM CaCl2, 1 mM MgCl2). After initial images were collected, buffer was exchanged for 40 mM NH4Cl or 40 µM DCCD in HBSSC and parasite motility was observed as above. Maximum projection images and videos were examined for motile parasites and the percent of motile parasites graphed over time. For inhibitor treatment assays, values were normalized to the percent motile parasites at time zero and the fold change in motility over time was graphed.
Microneme secretion induced with A23187 or ethanol as described previously [14] in the presence of vehicle, 10, 40 mM NH4Cl, or 10, 40 µM DCCD [10]. Low-pH induced secretion was tested by purifying parasites in Endo buffer (44.7 mM K2SO4, 106 mM sucrose, 10 mM MgSO4, 20 mM Tris-H2SO4 (pH 8.2), 5 mM glucose, 3.5 mg/ml BSA) [18], and re-suspending in invasion buffer (110 mM NaCl, 0.9 mM NaH2PO4, 44 mM NaHCO3, 5.4 mM KCl, 0.8 mM MgSO4, 1.8 mM CaCl2) or Endo buffer of the indicated pH at 37°C for 2 min. Calcium-dependence of secretion was tested by pre-incubation with BAPTA-AM as described previously with the following modifications: freshly egressed RH parasites were purified in Endo buffer and pre-treated with BAPTA-AM in Endo buffer at 37°C, pelleted, and then resuspended in 37°C Endo buffer of pH 5.4 or 7.4 with or without BAPTA-AM, incubated for 2 min at 37°C and then placed on ice [14]. To test the effect of potassium inhibition of microneme secretion, freshly egressed RH parasites were purified in Endo buffer at room temperature, pelleted, and resuspended in 37°C Endo buffer with either 45 mM KCl/5 mM NaCl or 5 mM KCl/45 mM NaCl, incubated at 37°C for 2 min, and then placed on ice. For all secretion assays, after incubating on ice, parasites were pelleted at 4°C, the secreted fraction (supernatant) was removed and spun again and the parasite pellet was washed in cold PBS prior to resuspending in boiling SDS-PAGE loading buffer. Secreted fractions and parasite pellets were examined by immunoblotting with the indicated antibodies.
Low-pH induced egress was tested by inoculating HFF in an 8-well chamber slide with 3 µl of freshly egressed RH parasites/well and incubating for 30 h. The slide was then washed twice with warm high K+ buffer (pH 8.4) and the buffer was replaced with buffer of the same composition and varying pH with and without 15 µM digitonin. The slide was incubated at 37°C for 3 min, fixed with 8% formaldehyde and occupied vacuoles were enumerated as previously described [25].
Superecliptic and ratiometric pHluorin vectors were kindly provided by Dr. Gero Miesenbock by material transfer agreement (University of Michigan, SSP no. 13477; Memorial Sloan-Kettering Institute, SK# 19367) [20]. The genes were subcloned into the DsRed vacuolar expression vector [22]. Parasites were transfected with plasmid, transformed parasites were selected for with chloramphenicol and cloned by limiting dilution. Superecliptic pHluorin was highly expressed by parasites in the PV.
pH-sensitivity was tested by inoculating HFF with varying concentrations of super-ecliptic pHluorin expressing plp1ko parasites (plp1kosepH) in a 96-well plate, and incubating for 30 h at 37°C. plp1ko parasites were used to retain the fluorescent signal in the parasitophorous vacuole upon egress induction. Wells were washed twice with warm PBS and PBS without Triton-X-100 was used to measure fluorescence in live, infected cells. PBS (pH 5.4 or 7.4) with 0.1% Triton-X-100 was used to measure fluorescence in lysed cells.
A23187-induced changes in superecliptic pHluorin signal in live, infected cells were tested by inoculating HFF (grown in phenol red-free DMEM) in a 96-well plate with plp1kosepH parasites (2.5×106 parasites/ml, 100 µl/well), incubating for 30 h, washing the wells twice with warm PBS, and adding HBSSC with 2 µM A23187 or DMSO, without or with 20 mM NH4Cl or 40 µM DCCD. The time between the addition of the compounds and fluorescence measurement was approximately 5 min. The majority of plp1ko parasites are unable to egress in this time period.
Fluorescence over time in live, infected cells was observed by inoculating HFF with plp1kosepH (1×105 parasites/ml, 100 µl/well) in a 96-well plate. At the indicated time points, 2 sets of triplicate wells were washed twice with warm PBS. One hundred µl warm HBSSC with or without 20 mM NH4Cl was added and fluorescence was read in a pre-warmed plate reader. Following the fluorescence reading, the plates were reincubated and a new set of wells was used for each time point. At each time point, wells were briefly examined microscopically prior to the fluorescent reading to check if the majority of parasites were intracellular. The experiment was terminated beyond 44 h due to the progression of endogenous egress.
Fluorescence was measured at excitation 485/20 nm and emission 530/25 nm in a BioTek Synergy HT microplate reader at 37°C. Background from uninfected cells was subtracted from the total fluorescence for each condition. Data points represent the average and standard deviation of 2 or 3 independent experiments consisting of triplicate wells in each experiment. All assays were conducted in clear plastic 96-well plates since optimization experiments indicted they performed equally well as black-sided well plates.
We also attempted to measure the pH of the PV using ratiometric pHluorin [20]. Transfection of the original ratiometric pHluorin sequence in the DsRed vacuolar expression vector [22] into T. gondii tachyzoites followed by drug selection and isolation of clones revealed that ratiometric pHluorin was transcribed but not translated as demonstrated by production of mRNA by RT-PCR and lack of detection by fluorescence microscopy, immunoblot or pulse chase 35S-methionine/cysteine metabolic labeling and immunoprecipitation. Ratiometric pHluorin was subsequently codon-optimized, chemically synthesized (GenScript Inc) and subcloned into the DsRed vacuolar expression vector as above. Codon-optimized ratiometric pHluorin was expressed by the parasites, and fluorescent parasites recovered upon drug selection and cloning by limited dilution. Imaging was performed with Zeiss filter sets 21 HE (excitation 340/30 nm +387/15 nm, emission 510/90 nm) and 38 HE (excitation 470/40 nm, emission 525/50 nm). Fluorescent parasites were lost upon prolonged passage despite continuous drug selection, thus limiting the experiments to transiently transfected parasites. However, due to low expression and fluorescence, the exposure times required for signal detection were longer than those needed to measure rapid changes in pH during induced egress. Additionally, PVs of mock transfected parasites were noted to have varying degrees of autofluorescence in the 390 nm channel, which is the pH-sensitive wavelength, giving low confidence in the ability of ratiometric pHluorin to accurately reflect changes in PV pH especially as the total amount of signal was low late in the endogenous replication cycle.
Egress assays with A23187 were conducted as previously described [33] with the following modifications: HFF in an 8-well chamber slide were inoculated with 3 µl of freshly egressed wild type (RH) parasites and incubated for 30 h at 37°C. Wells were washed twice with warm PBS and a 2 minute pretreatment at 37°C of the inhibitor (NH4Cl, bafilomycin, dicyclohexylcarbodiimide (DCCD), omeprazole) in HBSSC was applied prior to addition of 4 µM A23187 (2 µM final concentration) or DMSO with or without the indicated treatment in HBSSC for 2 minutes and fixation in 8% formaldehyde. Inhibitors were purchased from Sigma and tested at the indicated concentrations. Zaprinast-induced egress was tested at 250 µM (final concentration) in the same manner as A23187-induced egress. Immunofluorescence was performed and enumerated for SAG1 and GRA7 and occupied/unoccupied vacuoles as previously described [33].
Egress-associated membrane permeabilization was tested by inoculating HFF cells in an 8-well chamber slide with 2 µl of freshly egressed RH or plp1ko parasites and incubating for 30 h at 37°C. Following 2 washes with warm PBS, wells were treated with 100 µl HBSSC+1 µM CytD with or without NH4Cl or DCCD for 3 min at 37°C. Then 100 µl/well was added of 4 µM A23187/DMSO, 1 µM CytD, 12.5 µg/ml propidium iodide (PI) with the indicated final concentrations of NH4Cl and DCCD and incubated for 3 min. Following the incubation, cells were washed twice with warm PBS, fixed with 4% formaldehyde and stained with DAPI. Membrane permeabilization was quantified by the number of infected cells with PI-positive nuclei.
Recombinant PLP1 and LLO were generated as previously described [33]. Streptolysin O (SLO) was handled according to the manufacturer's instructions (Murex Diagnostics). pH-dependent hemolysis was assessed by washing erythrocytes in PBS (pH 7.4), pelleting the RBC, and re-suspending in PBS of indicated pH (prepared by mixing sodium mono- and diphosphate in different amounts and adjusting pH with HCl or NaOH) with 100 nM recombinant protein. RBC and recombinant protein were incubated at 37°C for 1 h, pelleted, and hemolysis was measured by absorbance at 540 nm of the supernatant.
pH-dependent binding was tested by incubating erythrocyte ghosts, prepared as previously reported, with recombinant protein in PBS of the indicated pH [33]. RBC ghosts and recombinant protein was incubated at 37°C for 30 min; cells were pelleted and washed three times with cold PBS at neutral pH and bound samples were analyzed by SDS-PAGE and immunoblot.
Parasite invasion was tested as previously described [34] with the following modifications: parasites were purified in Endo buffer and allowed to settle on HFF in an 8-well chamber slide at room temperature for 20 min. Then Endo buffer was replaced with either DMEM or DMEM-like buffer (invasion buffer) (110 mM NaCl, 0.9 mM NaH2PO4, 44 mM NaHCO3, 5.4 mM KCl, 0.8 mM MgSO4, 1.8 mM CaCl2) at pH 5.4 or 7.4 and the slide was incubated at 37°C for 2 min. The buffer or media was removed, wells were washed twice with room temperature PBS and the slide was fixed with 0.4% formaldehyde in PBS. Immunofluorescence staining and parasite quantification was conducted as previously described for attached and invaded parasites.
Cell wounding was tested by pre-loading host cells with 1 µM calcein-AM in phenol-red free DMEM and incubating for 1 h at 37°C, followed by two washes with warm PBS. Parasites were filter-purified in Endo buffer and applied to host cells in a 96-well plate by centrifuging at 500 g for 5 min. Supernatant was removed and replaced with 100 µl/well DMEM-like buffer, pH 7.4 or 5.4. Plates were incubated for 10 min at 37°C and centrifuged as above. Fifty µl of supernatant was transferred to another plate and calcein fluorescence was read in a 96-well plate reader.
|
10.1371/journal.ppat.1000022 | SNARE Protein Mimicry by an Intracellular Bacterium | Many intracellular pathogens rely on host cell membrane compartments for their survival. The strategies they have developed to subvert intracellular trafficking are often unknown, and SNARE proteins, which are essential for membrane fusion, are possible targets. The obligate intracellular bacteria Chlamydia replicate within an intracellular vacuole, termed an inclusion. A large family of bacterial proteins is inserted in the inclusion membrane, and the role of these inclusion proteins is mostly unknown. Here we identify SNARE-like motifs in the inclusion protein IncA, which are conserved among most Chlamydia species. We show that IncA can bind directly to several host SNARE proteins. A subset of SNAREs is specifically recruited to the immediate vicinity of the inclusion membrane, and their accumulation is reduced around inclusions that lack IncA, demonstrating that IncA plays a predominant role in SNARE recruitment. However, interaction with the SNARE machinery is probably not restricted to IncA as at least another inclusion protein shows similarities with SNARE motifs and can interact with SNAREs. We modelled IncA's association with host SNAREs. The analysis of intermolecular contacts showed that the IncA SNARE-like motif can make specific interactions with host SNARE motifs similar to those found in a bona fide SNARE complex. Moreover, point mutations in the central layer of IncA SNARE-like motifs resulted in the loss of binding to host SNAREs. Altogether, our data demonstrate for the first time mimicry of the SNARE motif by a bacterium.
| Chlamydiae are obligate intracellular bacteria that have co-evolved with eukaryotic cells and adapted to a wide range of hosts, causing several diseases in humans and animals. For example, one species pathogenic to humans, Chlamydia trachomatis, is the leading cause of preventable blindness and of bacterial sexually transmitted diseases worldwide. Chlamydiae multiply inside a membrane-bound compartment, the inclusion. The exchanges between the membrane of the inclusion and other intracellular membranes are tightly controlled by the bacteria, for example avoiding fusion with some degradation compartments, while acquiring lipids. Inclusion proteins, made by the bacteria and secreted into the inclusion membrane, are thought to play a central role in controlling these interactions, although their exact function is mostly unknown. We have identified, in three inclusion proteins, a motif common to proteins that are essential for the fusion of two compartments in eukaryotic cells, the SNARE proteins. Via this motif, inclusion proteins interact specifically with a subset of SNAREs of the host, which leads to the selective recruitment of intracellular compartments around the inclusion. This study thus provides a striking example of mimicry of the host by an intracellular pathogen.
| Chlamydia are obligate intracellular bacterial pathogens of eukaryotic cells. They infect a variety of animals, including humans, and cause acute and chronic diseases [1]–[3]. Chlamydia replicate primarily within epithelial cells, in a membrane-bound compartment called the inclusion. The membrane of the inclusion is of dual origin, reflecting its position at the interface between host and pathogen. The bacteria use a type III secretion process to translocate into it the Inc proteins, a large family of Chlamydia specific proteins of mostly unknown function [4]–[6]. Host cell proteins might be less abundant in the inclusion membrane, which lacks conventional markers of early and recycling endosomes and avoids fusion with acidic degradative compartments [6],[7]. However, several lines of evidence indicate a contribution of host cell compartments to the inclusion growth [8]–[10]. They suggest that Chlamydiae control their interactions with the host intracellular traffic, allowing some fusion events while avoiding others. In eukaryotic cells, SNARE (soluble NSF (N-ethylmaleimide-sensitive factor) attachment protein receptors) proteins play an essential role in compartment fusion [11]. They share a conserved motif, the SNARE motif, and have been classified as Q-SNAREs (glutamine containing SNAREs) and R-SNAREs (arginine containing SNAREs) based on a highly conserved residue at the centre of this motif [12]. SNARE proteins anchored in two lipid bilayers associate in complexes involving three Q-SNARE and one R-SNARE motifs. Complex formation is needed for the fusion of the two lipid bilayers. More recently, it appeared that SNAREs can also have inhibitory role in membrane fusion, by substituting for or binding to a subunit of a fusogenic SNARE bundle to form a nonfusogenic complex [13]. As central regulators of membrane fusion, SNARE proteins appear as possible targets for intracellular organisms, which often rely on subverting the host intracellular traffic. However, although there have been suggestions for the presence of SNARE-like motifs in Legionnella effector proteins [14], there is no definite example of mimicry of the SNARE motif by an intracellular bacterium. We have previously shown that one Chlamydia inclusion protein might interact with itself to form a complex similar to that of the SNARE complex, and thus facilitate the homotypic fusion of inclusions [15]. This finding led us to hypothesize that one of the functions of some inclusion proteins is to control intracellular trafficking by mimicking SNAREs.
To identify SNARE-like motifs in Chlamydia proteins, we used a bioinformatic approach. SNARE motifs are made of heptad repeat sequences that form coil-coiled structures. Position a and d of the heptads are occupied by hydrophobic residues, except at the centre of the motif, where a very conserved glutamine or arginine residue occupies position d, defining the zero layer. Many coiled-coil regions meet these criteria, hampering the identification of SNARE-like motifs from genomic data. However, in this case, we reasoned that the access to several different Chlamydia genome sequences should allow us to identify SNARE-like motifs in Chlamydia proteins with a high level of confidence. We restricted the search to proteins containing a large bilobed hydrophobic motif, which is the hallmark of Inc proteins [4]. These proteins are good candidates to interact with SNAREs, since they are most probably exposed on the cytosolic face of the inclusion membrane. Each Chlamydia genome encodes more than 40 putative Inc proteins. Among them, we found that the C. trachomatis genome encodes 11 proteins predicted to engage in coiled-coil interactions (see Materials and Methods). We aligned the 11 identified sequences against a SNARE motif profile compiled from 261 referenced SNAREs. The best score was obtained with a carboxy-terminal region of the inclusion protein IncA, which is anchored in the membrane by its N-terminal extremity (Figure 1A). We extended the search for SNARE motifs to all IncA sequences known to date, from a variety of Chlamydia strains. The level of similarity between IncA from different species is low, a characteristic of all Inc proteins [4]. Strikingly, in spite of this low conservation, SNARE-like motifs were identified in all IncA homologues, with the single exception of that of C. pneumoniae (Figure 1B). The conserved polar residue at the position corresponding to the zero layer of the SNARE motif was a glutamine or an arginine residue, which are the canonical residues in SNARE motifs. The presence of SNARE motif characteristics in IncA from six different Chlamydia strains strongly supports the hypothesis that the similarity to eukaryotic SNARE motifs is not fortuitous, and illustrates the usefulness of sequencing many strains of this intracellular bacterium which we can not genetically manipulate.
In addition to this motif, bioinformatics also revealed the presence of a second, membrane proximal, motif, which shared characteristics of SNARE motifs (Figure 1C). We designate it as motif Nter, and the carboxy-terminal motif as motif Cter. For each species, motif Cter always gave a higher score in the alignments with eukaryotic SNAREs than motif Nter. In particular, the polar residue in the zero layer of motif Nter was a threonine rather than a glutamine or an arginine for two Chlamydia species. Using a less systematic approach, we had previously observed that motif Nter showed similarities with SNARE motifs, and shown that it might be involved in the formation of homotetramers of IncA [15].
Finally, among the other C. trachomatis Inc proteins predicted to form coiled-coils, the Inc protein CT813 [16] and the putative Inc protein CT223 also showed similarities with SNARE motifs (amino acids 191 to 264 of CT813, 169 to 236 of CT223), although the alignment with host SNARE motifs was poorer than IncA's, and the zero layer was more difficult to define. Orthologs of CT813 and CT223 are only found in the closely related C. muridiarium species, so in this case sequence comparison could not be used to validate the identification of SNARE-like motifs. We did not identify SNARE-like motifs in the the eight remaining Inc proteins predicted to form coiled-coils.
To determine whether host SNAREs interact with inclusion proteins, we investigated the distribution of several of them in cells infected with C. trachomatis serovar D. Since C. trachomatis IncA SNARE-like motif Cter had a glutamine in its central layer, we hypothesized thas it might interact with R-SNAREs and investigated the localization of several host R-SNAREs, whose distribution and role in intracellular traffic is well documented (see Table 1 for details). In addition to their expected punctate distribution throughout the cell, endogenous Vamp3 and Vamp7 formed a patchy circle around the inclusion. Sec 22, another R-SNARE, did not encircle the inclusion, indicating that the redistribution observed in infected cells did not apply to all SNAREs (Figure 2A and C). Endogenous Vamp4 and Vamp8 expression was too low for detection. However, when cells were transfected with GFP-Vamp8 prior to infection, this protein showed a ring pattern around the inclusion, while GFP and GFP-Vamp4 did not (Figure 2B and C). Transfected GFP-Vamp7 relocalized to the inclusion while GFP-Vamp7 deleted of its SNARE motif (GFP-Longin) did not, indicating that the SNARE motif is necessary for the recruitment to the inclusion membrane.
The circular pattern of specific SNAREs can be due to their accumulation around the inclusion, and/or to their presence in the inclusion membrane itself. To discriminate between these possibilities, we used electron microscopy. In agreement with our observation that SNAREs of different compartments are recruited to the inclusion membrane, we observed an abundance of varied intracellular compartments in the immediate vicinity of the inclusion membrane (Figure 3A). Using immunogold labeling, we confirmed the enrichment of GFP-Vamp8 relative to GFP-Vamp4 around the inclusion, as 35% of the gold particles that labeled GFP-Vamp8 were less than 50 nm distant from its membrane, against 14% for those associated with GFP-Vamp4 (p-value<0.001). Among the gold particles that labeled GFP-Vamp8 within a 50 nm range of the inclusion membrane, 22% were on its membrane, while the rest were associated with compartments around the inclusion (Figure 3B). This result shows that the accumulation of GFP-Vamp8 around the inclusion observed in Figure 2B is mainly due to the recruitment of GFP-Vamp8 positive compartments around the inclusion, rather than to the presence of this host protein in the inclusion membrane. The same probably applies to GFP-Vamp3 and GFP-Vamp7 positive compartments, since endocytic markers of these compartments accumulate around the inclusion and are excluded from it [7]. Our observation that GFP-Vamp8 can reach the inclusion membrane, as well as the recent finding of CD63 on it, support the idea that multivesicular bodies, which contains both markers, might contribute to its growth [10].
The recruitment of Vamp8 positive compartments in the immediate proximity of the inclusion membrane suggests a possible direct interaction between this SNARE and SNARE-like proteins on the inclusion membrane. We next asked whether IncA could mediate the recruitment of Vamp8-positive compartments, and of other intracellular compartments. HeLa cells were simultaneously transfected with plasmids coding for C. trachomatis IncA and with different GFP-tagged SNARE proteins. One day later, SNARE proteins were immunoprecipitated using anti-GFP antibody. IncA was present in the GFP-immunoprecipitate when co-expressed with GFP-Vamp3, GFP-Vamp7 or GFP-Vamp8, but not GFP alone (Figure 4A). Importantly, IncA was not found in the GFP-immunoprecipitate when co-expressed with GFP-Vamp7 deleted of its SNARE motif (GFP-Longin), indicating that the SNARE motif of the SNARE protein is important for the interaction with IncA. It co-immunoprecipitates very poorly with GFP-Vamp4, indicating that IncA interacts preferentially with only a subset of SNAREs. These biochemical observations correlate well with the selective recruitment of SNAREs around the inclusion observed in Figure 2.
In vitro pull-down assays were performed to determine whether the interaction between IncA and host SNAREs was direct. Two populations of liposomes, one containing purified Vamp3, Vamp4 or Vamp8 with an amino-terminal glutathione-S-transferase tag, and one containing purified IncA-His, were mixed and incubated together for 16 hrs at 4°C. Proteins were then solubilized, and Vamps were pulled-down using glutathione-agarose beads (Figure 4B). IncA was pulled-down together with Vamp3 or Vamp8, showing that the interaction between IncA and these SNAREs is direct. IncA was not pulled-down with Vamp4, which correlates with the very weak co-immunoprecipitation of this inclusion protein with GFP-Vamp4 observed in Figure 4A, and confirms that IncA interacts preferentially with only a subset of SNAREs.
IncA is not expressed by C. trachomatis until 10 hrs post-infection [17]. To see whether the timing of the recruitment of SNAREs was consistent with the temporal expression of IncA, we observed the localization of SNAREs early in infection. Eight hours after infection, inclusions were very small, and we could not quantify the recruitment of SNAREs with confidence at this stage (data not shown). Eleven hours after infection, the recruitment of GFP-SNAREs was much less pronounced than 18h after infection (Figure S1), showing that the level of recruitment of SNAREs correlates with the timing of expression of IncA.
To test directly IncA's contribution to the recruitment of SNAREs to the inclusion membrane, we used a strain of C. trachomatis that does not express IncA, Ds5058 [18]. This strain grows significantly slower than the wild type strain [19], indicating that IncA plays an important role in infection. Cells were transfected with GFP-Vamp8, and then infected with the IncA negative strain. The cells were incubated for 48 hrs, in contrast to 24 hrs for the wild type strain, in order to observe inclusions of similar size (Figure 5A). Recruitment of GFP-Vamp8 to the inclusion was observed in about 40% of cells infected with the IncA negative strain, in contrast to 70% of cells infected with the wild type strain. Similarly, the circular pattern of GFP-Vamp3 or GFP-Vamp7 around the inclusion was observed in fewer cells infected with the IncA negative strain than with wild-type bacteria. At this level of analysis, in cases where host SNAREs were recruited to the IncA negative inclusions, no difference could be noted with the pattern of their distribution in cells infected with the wild type strain (Figure 5B and Figure S2).
We used an independent approach to confirm the predominant role of IncA in SNARE recruitment. As IncA associates with itself [15], we reasoned that heterologous overexpression of IncA might titrate the protein on the inclusion membrane and prevent association with SNAREs. Unfortunately, overexpression of C. trachomatis IncA inhibits the development of the bacteria [15], and could not be used for this purpose. However, we had previously shown that IncA from C. trachomatis (CtrIncA) and C. caviae (CcaIncA) shared biochemical properties [15] and we have identified here two SNARE-like motifs in C. caviae IncA (Figure 1B and C). We hypothesized that CcaIncA, if expressed by the infected cell, might associate with endogenous CtrIncA at the inclusion membrane. Indeed, when overexpressed by HeLa cells, CcaIncA was present in the endoplasmic reticulum, as previously observed [15], and around the inclusion membrane (Figure S3). The number of cells in which Vamp3 and Vamp7 were scored as recruited to the inclusion was reduced by about 50% in the population of cells expressing CcaIncA compared to non transfected cells (Figure 5C and Figure S3). Interestingly, overexpression of CtrIncA deleted from its hydrophobic domain (Δ75IncA) was not enriched around the inclusion and had no effect on the recruitment of Vamp3 and Vamp7 to the inclusion (Figure 5C). This result confirms our previous observation that overexpressed IncA needs to be inserted into a membrane compartment to be able to interact with endogenous IncA [15].
These results indicate that IncA plays a predominant, although not exclusive, role in the recruitment of SNAREs around the inclusion. Other inclusion proteins such as CT223 and CT813, which show similarities with SNARE motifs, may contribute to SNARE recruitment around the inclusion. To test this hypothesis, we cloned these two genes in mammalian expression vectors. After transfection, CT223 was not expressed by HeLa cells, as assessed by immunofluorescence, and was not studied further. CT813 was expressed in the endoplasmic reticulum. Co-expression and immunoprecipitation experiments showed that CT813 interacts with GFP-Vamp7 and GFP-Vamp8, and not with GFP alone GFP-Vamp4 or GFP-Vamp7 deleted of its SNARE motif (GFP-Longin). Interaction of CT813 with GFP-Vamp3 could not be assessed because the expression levels of both constructs were low (Figure 6). This result indicates that CT813 can interact with host SNAREs and that, in addition to IncA, several Inc proteins have evolved the ability to interact with SNAREs. Interestingly, more CT813 was pulled-down with Vamp7 than with Vamp8, while more IncA was pulled down with Vamp8 than with Vamp7, suggesting that affinities between inclusion proteins and different SNAREs vary. Other Inc proteins might partly compensate for the absence of IncA and explain why the clinical strain defective in IncA expression was still able to recruit SNAREs around its inclusion, although less efficiently that the wild type strain. Interestingly, CT813 and CT223 are specific to C. trachomatis and C. muridarium, suggesting that different Chlamydia species may have evolved different Inc proteins, targeting the SNARE machinery of their host in subtly different manners. In support of this hypothesis, we observed that, while C. caviae and C. pneumoniae were also able to recruit SNAREs to the inclusion membrane, the level of recruitment varied with strains (Figure S4). C. caviae, whose IncA has SNARE-like motifs and behaves similarily to C. trachomatis IncA [15], recruited the same set of SNAREs as C. trachomatis. In contrast, C. pneumoniae, whose IncA does not possess a clear SNARE-like motif, recruited good level of Vamp8 in the vicinity of its inclusion, but only little Vamp7. This observation suggests that, in C. pneumoniae, other Inc proteins than IncA might specifically interact with a subset of SNAREs.
Importantly, targeting the host SNARE machinery is not the sole method for Inc proteins to interfere with host trafficking: members of the family of rab proteins, which also participate in recognition and fusion of cell compartments [20], were also shown to interact with Inc proteins [21],[22].
Using molecular modelling, we previously showed that motif Nter was fully compatible with the formation of stable homotetramers, associated in a structure similar to that of the SNARE complex [15]. Using the same approach, we now asked whether IncA SNARE-like motifs could fit in the structure of a SNARE complex involving host SNAREs. We chose to use IncA SNARE-like motif Cter because it aligned better with eukaryotic SNARE motifs than motif Nter (Figure 1). We modelled the association of three identical motifs Cter (in place of Q-SNAREs) in association with one SNARE motif from a R-SNARE, for three reasons: (i) IncA has a high propensity to form dimers or other multimeric structures [15], (ii) motif Cter of C. trachomatis IncA classifies as a Q-SNARE, and aligned better with Q-SNAREs than R-SNAREs (Figure 1B) (iii) IncA can interact with R-SNAREs (Figure 4). We modelled the heterotetramers between the SNARE motif of host Vamp8 and a trimer of IncA motif Cter (Figure S5). The model of the complex between three molecules of IncA and Vamp8 is very similar to the structure of the endosomal SNARE complex (Figure 7). In particular, several side-chains of Vamp8 are involved in salt bridges with side-chains in IncA. There is a cluster of salt bridges close to the central layer of the complex, involving residues at positions +5 and +8 from the central Argnine in Vamp8, with resdidues at positions +3 and +6 from the central Glutamine in IncA, and two additional salt bridges N-terminal and C-terminal of the central layer. The patterns of predicted interaction energies, evaluated as the difference in total energy between the complex and separated helices, are similar. The predicted arrangement of glutamine residues around the central arginine of Vamp8 is very similar to that found in the X-ray crystal structure of the endosomal SNARE complex [23] (Figure 7, inset). These data support a model in which IncA makes SNARE complexes with host SNAREs via its SNARE-like motif Cter.
Our models, and numerous reports on eukaryotic SNARE complexes, suggest that, if the IncA SNARE-like motif functions as a bona fide SNARE motif, forming complexes with R-SNAREs, introduction of a large charged amino acid, such as an arginine, in the zero layer of the IncA SNARE-like motif might destabilize the SNARE complex sufficiently to lose the interaction between IncA and host SNAREs. To test this hypothesis, we introduced point mutations in the zero layer of IncA SNARE-like motif Nter (IncAT126R), motif Cter (IncAQ244R) or both (IncAT126RQ244R). All constructs were expressed at a level comparable to wild type IncA when transfected in HeLa cells. However, coexpression of the Q244R mutant together with GFP-Vamps constructs resulted in very weak expression of both transfected genes. The reasons for this phenomenon are unclear, and we could not assay the ability of this mutant to co-immunoprecipitate with GFP-SNAREs. We performed co-immunoprecipitation experiments in cells co-expressing different GFP-SNAREs and either IncA wild type, or mutated in motif Nter, or in both SNARE-like motifs. The single mutant (IncAT126R) co-immunoprecipitated with GFP-Vamp3, GFP-Vamp7 and GFP-Vamp8, to the same extent as IncA wild-type. However, IncA mutated in both SNARE-like motifs (IncAT126RQ244R) did not co-immunoprecipitate with any of the Vamps (Figure 8A). This experiment demonstrates that the interaction between IncA and host SNAREs is mediated by IncA SNARE-like motifs. Further analysis would be needed to determine whether only motif Cter is able to engage in SNARE complexes with host SNAREs, or whether any of the two SNARE-like motifs can do so and mutation of both motifs is needed to lose the interaction. In any case, these experiments demonstrate that IncA SNARE-like motifs function as bona fide SNARE motifs since point mutations in residues known to be critical for the formation of SNARE complexes result in the loss of the interaction between IncA and host SNAREs.
To confirm this important result by a different approach, we investigated the effect of the point mutations in IncA SNARE-like motif on the ability of IncA to engage in homotypic interactions. We have previously shown that heterologous expression of IncA, which localizes at the endoplasmic reticulum, inhibits inclusion development. This effect requires a direct interaction between IncA molecules at the inclusion and on the endoplasmic reticulum, and we hypothesized that it might be due to the formation of homotypic SNARE complexes between IncA molecules present on the two compartments [15]. Cells, transfected with IncA wild- type or with the different mutants, were infected for 20 hrs before fixation and labeling of the bacteria and IncA-His by immunofluorescence. Development of inclusions was largely impaired in cells expressing IncA wild type or IncAT126R (Figure 8B). The Q244R mutation in the IncA SNARE-like motif Cter partially restored the growth of the inclusion in the transfected population, and the double mutation in both SNARE-like motifs restored the growth further, although not totally. Expression of IncA deleted of its hydrophobic domain (Δ75IncA) had no effect on the development of the bacteria, confirming that IncA needs to be inserted in a membrane to disrupt the development of the inclusion [15]. Altogether, this experiment shows that homotypic interaction between IncA molecules is mediated by its SNARE-like motifs. It suggests that both SNARE-like motifs contribute to the interaction, although motif Cter is able to compensate for the mutation in motif Nter, while motif Nter is not fully able to compensate for the mutation in motif Cter. The observation that the double mutant only partially (about 50%) restores the ability of the bacteria to grow in the transfected cells indicates that the point mutations are not sufficient to fully disrupt IncA homotypic interactions.
Altogether, point mutations in IncA SNARE-like motifs impaired its ability to associate with host SNAREs and with itself. These results show that IncA SNARE-like motifs behave as bona fide SNARE motifs, and strongly support our hypothesis that IncA interaction with host SNAREs is mediated by the formation of SNARE complexes.
In agreement with previous reports from several laboratories, we have observed that a variety of cellular compartments accumulate around the inclusion. Here we bring evidence that SNARE mimicry is one mechanism by which the Chlamydia recruit a specific subset of host SNAREs. We have identified SNARE-like motifs in the inclusion protein IncA, and showed that a mutant strain that does not express IncA presents reduced ability to recruit SNAREs around its inclusions. Our data also suggest that other inclusion proteins may use SNARE motif mimicry to interact with SNAREs. These conclusions open two important questions: in which SNARE complexes is IncA engaged, and what are the consequences in term of membrane fusion?
To start answering the first question, we have tested a variety of IncA/SNARE interactions. By co-imunoprecipitation experiments, we found that IncA was able to interact with several different SNAREs, but not all. We observed a good correlation between this result and the specific recruitment of a subset of SNAREs around the inclusion membrane. The basis of the specificity of interaction between IncA and a subset of host SNAREs is not yet known. A model between the SNARE motif of Sec22 and a trimer of IncA motif Cter resulted in an unstable complex (data not shown), a finding that correlates with the absence of recruitment of Sec22 around the inclusion membrane. However, modelling predicted a stable association of Vamp4 with IncA motif Cter, while we observed no or very weak interaction between IncA and Vamp4 (Figure 4). This result suggests that, in addition to motif Cter, other domains of IncA contribute to define its specificity. It may also be that, in vivo, interactions between the inclusion membrane and the host SNAREs involve a combination of SNARE-like motifs from several different inclusion proteins and/or of SNARE motifs from several different host SNAREs present in the same compartment. Dissecting the complexes in which IncA is involved therefore remains a challenging task for the future.
It is also difficult to bring a definite answer to the second question, regarding the mechanistic roles of SNARE-like motifs in inclusion proteins, especially in the absence of tools to manipulate the Chlamydia genome. The three SNAREs recruited to C. trachomatis inclusion membrane are SNAREs involved in endosomal trafficking, while Vamp4 and Sec22, which are more involved in the secretory pathway, were not, suggesting that SNARE mimicry may have preferentially evolved in C. trachomatis to target the endosomal pathway. Depletion of host SNAREs Vamp3, Vamp4, Vamp7 or Vamp8 using siRNA had no impact on the growth of the bacteria (Figure S6). This might be due to an insufficient depletion of the pools, especially in the case of Vamp3, or because other host SNAREs present in the same compartment as the siRNA target can compensate for the depletion. It might also reveal some degree of redundancy in the interactions between the inclusion and different intracellular pathways, as was observed for another intracellular bacterium, Legionella pneumophila [24].
In the absence of functional indications regarding the role of the interaction between inclusion proteins and host SNAREs in infection, we can propose several models. SNAREs are mostly known for their role in membrane fusion. By engaging in fusion competent SNARE complexes, inclusion proteins might permit the fusion of some cellular compartments with the inclusion membrane. If this is the case, the cellular SNARE should be present in the inclusion membrane, at least transiently. By electron microscopy, we observed that Vamp8 is found in the inclusion membrane, suggesting that indeed, IncA-SNARE interaction might promote fusion of specific intracellular compartments, in this case probably multivesicular bodies, with the inclusion. Conversely, SNARE motifs in inclusion proteins might play an inhibitory role on fusion, by engaging in fusion incompetent SNARE complexes [13], or by titrating individual SNAREs. One argument to support this opposite model is the long distance between motif Cter and the transmembrane domain of IncA. SNARE mediated fusion requires proximity between the SNARE complex and the lipid bilayers, which might be difficult to achieve if motif Cter is engaged in the SNARE complex. Another observation that supports this model is the absence of effect of Vamp7 depletion on infection. Although siRNA against Vamp7 reduced its level by more than 90%, it did not affect the growth of the bacteria (Figure S6). Similarily, expression of GFP-Longin, a dominant negative form of Vamp7, did not affect Chlamydia infection. These results support a model in which IncA/Vamp7 interaction prevents Vamp7 mediated fusion of late endocytic compartments with the inclusion. Finally, it is also possible that SNARE-like domains in inclusion proteins serve as mere anchors. By associating with various SNAREs, they contribute to the accumulation of vesicles in the immediate vicinity of the inclusion. This might be beneficial to the development of the bacteria, without requiring fusion to occur.
Chlamydiaceae have very probably been intracellular for several hundred million years [25]. Therefore, a high degree of complexity in the interaction between inclusion proteins and host organelles should be expected. Bacterial proteins involved in these interactions are just beginning to be identified. In light of our results, we propose that co-evolution shaped some of the bacterial inclusion proteins into bacterial SNARE proteins.
C. trachomatis serotype D (27F0734) was from ATCC. C. trachomatis serotype D(s)5058 is a clinical isolate, which does not express IncA, and was kindly given by Drs. D. Rockey and W. Stamm [18]. The GPIC strain of C. caviae and the CWL029 strain from C. pneumoniae were obtained from Drs. R. Rank (University of Arkansas) and G. Christiansen (University of Aarhus, Denmark) respectively. HeLa cells were used in all experiments except for infection with C. pneumoniae, which was performed in Hep2 cells. Infections were performed as described [15]. In some experiments, HeLa cells were transfected 8 to 18 hrs prior to infection and processed for immunofluorescence as described [15]. The inclusion membrane was observed using either anti-IncA (generous gift from Dr Ted Hackstadt, Rockey Mountain Laboratories, NIH, NIAID) or anti-Cap1 polyclonal antibodies which we obtained from New Zealand White rabbits immunized with His-tagged recombinant protein CT529 purified from E. coli. Antibodies against Vamp4 were kindly provided by Dr. Andrew A. Peden (Cambridge Institute for Medical Research, Cambridge), rabbit anti-Vamp3, rabbit anti-Vamp8 and mouse anti-Vamp7 and vectors coding for GFP-Vamp3, GFP-Vamp7, GFP-Longin and GFP-Vamp8 were kindly provided by Dr Thierry Galli (Institut Jacques Monod, Paris), rabbit anti-Sec22 was described previously [26], plasmid coding for GFP-Vamp4 was a gift of Dr Ludger Johannes (Institut Curie, Paris). Endogenous SNAREs were first labeled using the corresponding antibodies, followed with Alexa Fluor-488-conjugated anti-rabbit antibodies (Molecular Probes). In that case, the inclusion membrane could not be visualized with rabbit antibodies, and bacterial DNA was labeled using 0.5 µg/ml Hoechst in the mounting medium. Coverslips were examined under an epifluorescence microscope (Axiophot, Zeiss, Germany) equipped with a 63× Apochromat objective and a cooled CCD-camera (Photometrics, Tucson, AZ), driven by Metaview software (Universal Imaging, Downingtown, PA). For quantification of the recruitment of SNAREs around the inclusion membrane, more than 100 infected cells (and in some case transfected with GFP-SNARE) were counted in each case. Cells were scored as positive when the entire circumference of the inclusion was surrounded by the SNARE protein.
HeLa cells were transfected by electroporation with the indicated constructs, and lysed 24 h later on ice in 50 mM Tris, 150 mM NaCl, 1% Triton X-100, 10 mM EDTA, pH 7.5 and cocktails of inhibitors. Immunoprecipitation of GFP-tagged proteins was performed using anti-GFP monoclonal antibodies (clones 7.1 and 13.1, Roche Applied Science). Immunoprecipitated proteins were incubated in sample buffer, boiled and loaded on acrylamide gel for western blot detection of IncA or of the histidine tag (#14-6757, eBioscience, San Diego CA). Expression level and immunoprecipitation of SNAREs proteins were checked by stripping the membrane and incubating it with anti-GFP antibodies (#sc-8334 Santa Cruz Biotechnology) followed by horseradish peroxidase-linked (HRP) anti-rabbit antibodies (Amersham Biosciences).
GST-Vamp3, GST-Vamp4, and GST-Vamp8 were expressed, purified and reconstituted into a first population of clear liposomes as described [27],[28]. Full-length IncA-His was subcloned in the pet28 vector, expressed in BL21 E. coli for 20 hrs at 16°C, and subsequently purified in buffer A (100 mM KCL, 25 mM HEPES, 10% glycerol, 1% octyl-β-D-glucopyranoside) using the protocol described in [27]. IncA-His was reconstituted into a second population of clear liposome as described [28]. After incubating each GST-SNAREs containing liposome with IncAHis-containing liposomes together for 16 hrs at 4°C, the mixture was dissolved in 2.5% wt/vol n-dodecyl-maltoside (Boehringer). The proteins/lipids mixture was further diluted into buffer A, and GST complexes were pulled-down using glutathione agarose equilibrated in buffer A. The glutathion agarose beads were then pelleted and washed three times with the buffer A. Pull-down complexes were resolved on SDS-PAGE gels and proteins stained with Coomassie blue.
Transfected and infected HeLa cells were fixed with a mixture of 2% (wt/vol) paraformaldehyde and 0.5% (wt/vol) glutaraldehyde in a 0.2 M phosphate buffer (PB) pH 7.4, 4 h at room temperature. After many washing with 0.1 M glycin in PBS, cells were processed for ultracryomicrotomy as described [29]. Incubations were performed in blocking buffer (20 mM glycin, 0.1% of cold water fish skin gelatin, Biovalley, in PBS) and sections were labeled with anti-GFP (#A11122, Invitrogen Life Technologies) and visualized with protein A coupled to 15-nm gold. Sections were observed and acquired under a Philips CM120 electron microscope (FEI, Eindoven, Netherlands). Digital acquisitions were made with a numeric camera Keen View (Soft Imaging System, Munster, Germany).
Construction of CtrIncA and CcaIncA with a carboxyterminal His tag was described elsewhere [15], Δ75IncA-His was constructed in the same way with a deletion of the first 75 amino acids. The same cloning strategy was used to clone CT813 using the primers atgctaccatggctactcttcccaataattgcactt and agtcggtacctatcgaaccacgtcttcctg, but for unknown reasons, the gene corresponding to the full-length protein could not be cloned over several attempts. A clone, designated CT813His*, was obtained, with a deletion of one nucleotide which did not permit expression of the full-length protein, but of a protein missing its 23 first amino acids (initiation of translation from methionine 24). This deletion does not affect the transmembrane domain of CT813, or its SNARE-like motif.
The selection of specific oligonucleotides and the procedures for RNAi experiments to silence the expression of endosomal v-SNAREs (Vamp3, Vamp7, Vamp8) will be published elsewhere (Danglot et al., 2008 in preparation), Vamp4 siRNA (AAGAUUUGGACCUAGAAAUGA) was designed by Dr. Andrew A. Peden (Cambridge Institute for Medical Research, Cambridge). 5×106 cells were electroporated in the presence of 2 pmoles of siRNA at day 1 and again at day 4, plated in 24-well plates and one 10-cm dish at day 5 and infected at day 6 in the 24-well plates. On the day of infection, cell lysates were prepared from the 10 cm dish and normalized to equal protein concentration. The level of expression of each SNARE was assessed by western blot using specific antibodies. Nitrocellulose membrane were incubated with ECL western blotting reagents (Amersham) and processed for detection using a Luminescent Image Analyzer LAS-3000 (Fujifilm). Digital images were acquired using the software Image Reader LAS-3000 v2.2 (Fujifilm) and analyzed by MultiGauge v3.0 (Fujifilm). All the quantification were performed using the expression of β-tubulin to normalize the samples, and for each SNARE, the level of expression is expressed as the percentage of expression in control cells transfected with the si-β-globin. One day after infection, the cells were detached from 24-well plates in 0.5 mM EDTA in PBS, fixed for one hour in 70% ethanol, centrifuged and washed in PBS. Bacteria were then stained using anti-EfTu antibody from Dr Y-X Zhang (Boston) followed by Cy5-coupled anti-mouse antibody (Amersham) and the samples were analyzed by flow cytometry.
Transmembrane domains were identified in Chlamydia proteome using Phobius predictor, which detects in a single search signal peptides and transmembrane helixes [30]. Proteins with two transmembrane domains separated by 2 to 40 amino acids, and with no obvious functional attribution, were considered as putative Inc proteins. Putative Inc proteins predicted to engage in coiled coil using a hidden markov model (Marcoil [31]) were CT119 (IncA), CT222, CT223, CT224, CT225, CT226, CT228, CT229, CT233 (IncC), CT616 and CT813. A database made of SNARE motifs from 261 referenced SNARE proteins was used to perform optimal alignments searches [32].
Models of the complexes were generated essentially as described before [15], see Figure S5. Electrostatic energies were calculated with the “ACE” Generalized Born model [33], implemented in X-plor [34], with an dielectric constant of 1 for the interior of the protein, and a cut-off for electrostatic interactions of 15 Å. To estimate the relative stabilities of the complexes, we calculated the difference between the energies for the tetramer and the four helices separated by 200 Å. Solvatation energies were estimated by the difference in solvent accessible surfaces multiplied by 0.005 kcal/(mol Å2). The profiles shown in Figure 6 represent the sum of electrostatic energies, van der Waals energies, and solvatation energies, averaged over the equivalent residues in the four helices. |
10.1371/journal.ppat.1003997 | Activation of HIV-1 from Latent Infection via Synergy of RUNX1 Inhibitor Ro5-3335 and SAHA | A major barrier to the elimination of HIV-1 infection is the presence of a pool of long-lived, latently infected CD4+ memory T-cells. The search for treatments to re-activate latent HIV to aid in clearance is hindered by the incomplete understanding of the mechanisms that lead to transcriptional silencing of viral gene expression in host cells. Here we identify a previously unknown role for RUNX1 in HIV-1 transcriptional latency. The RUNX proteins, in combination with the co-factor CBF-β, are critical transcriptional regulators in T-cells. RUNX1 strongly modulates CD4 expression and contributes to CD4+ T-cell function. We show that RUNX1 can bind DNA sequences within the HIV-1 LTR and that this binding represses transcription. Using patient samples we show a negative correlation between RUNX1 expression and viral load. Furthermore, we find that pharmacologic inhibition of RUNX1 by a small molecule inhibitor, Ro5-3335, synergizes with the histone deacetylase (HDAC) inhibitor SAHA (Vorinostat) to enhance the activation of latent HIV-1 in both cell lines and PBMCs from patients. Our findings indicate that RUNX1 and CBF-β cooperate in cells to modulate HIV-1 replication, identifying for the first time RUNX1 as a cellular factor involved in HIV-1 latency. This work highlights the therapeutic potential of inhibitors of RUNX1 to re-activate virus and aid in clearance of HIV-1.
| Since it was first discovered in the early 1980s, Human Immunodeficiency Virus 1 (HIV-1), the causative agent of Acquired Immunodeficiency Syndrome (AIDS), has been the focus of intense research. In untreated individuals, the number of CD4+ T-cells in the blood slowly drops over time and the patient eventually succumbs to an opportunistic infection. Although current therapies are capable of managing the virus; they do not represent a true cure. As a retrovirus, HIV-1 incorporates itself into the host genome and survives in the long-lived population of memory T-cells found in the human host. In this study, we examine the roll of a T-cell specific transcription factor (RUNX1) in the control of HIV-1 replication. Through various molecular studies, we show that RUNX1 represses HIV-1 replication in T-cells. By examining samples from patients with HIV-1, we are able to show a negative correlation between viral replication and RUNX1 expression. Finally, we show that an inhibitor of RUNX1 synergizes with Vorinostat, a current lead compound in the quest to re-active HIV-1 and purge the latent pool.
| Human Immunodeficiency Virus type I (HIV-1) is the etiologic agent of Acquired Immunodeficiency Syndrome (AIDS). HIV-1 has a complex life cycle that in part involves a unique transcriptional interaction between the viral Tat protein and its target RNA element (TAR) found in the R sequence of the LTR [1]–[3]. In the absence of treatment, most HIV-1 infected individuals will experience a steady decline in the number of CD4+ T-cells, progress to AIDS and eventually die as the result of acquiring opportunistic infections.
Transcriptional control of HIV-1 occurs in two phases. Basal transcription of the integrated provirus first occurs at a low level in a Tat independent manner [4]. Once the Tat protein is synthesized, viral transcription transits to a Tat-dependent route. Tat binds TAR RNA and recruits a complex of cyclin T1 and CDK9 to the start site of transcription [1] leading to the phosphorylation of the c-terminal domain (CTD) of the RNA Pol II to induce more processive transcription. Tat has also been shown to help initiate transcription through interaction with the TATA Binding Protein as well as various histone modifying enzymes such as CBP/p300 and the PBAF complex [5].
The HIV-1 LTR contains a myriad of transcription factor-binding sites, such as those for SP1 and NF-κB. It is believed that interactions of cellular factors with the HIV-1 LTR determine active transcription versus the establishment of transcriptional latency. HIV-1 latency, a state in which the infected cell produces little to no viral RNA, represents a major barrier to viral eradication in an infected individual [6]–[8].
In mammals, there are three RUNX proteins that can interact with a cofactor, core-binding factor β (CBF-β), to form an active transcription factor complex [9], [10]. RUNX protein binding to CBF-β allows transport of the complex into the nucleus via a localization signal in the RUNX protein [11]. In turn, CBF-β increases the affinity of RUNX proteins for DNA. This complex is essential for proper differentiation of cells of the hematopoietic lineage. Of particular interest is the involvement of RUNX1 in the differentiation and fate selection of CD4+ T-cells [12]–[15]. Specifically, RUNX1 is drastically down regulated when thymocytes progress from double-negative to double-positive during development, and it is also down-regulated when naïve CD4+ T-cells are stimulated through the T-cell receptor (TCR) to become effector cells. In the latter case, RUNX1 down-regulation is associated with increased IL-2 production which is likely critical for CD4+ T-cell function [14].
Mechanistically, RUNX serves a role in the initiation of transcription that is likely achieved through p300 histone acetyltransferase recruitment. Intriguingly, RUNX family members may also recruit repressive factors such as mSin3A, Suv39H1 and histone deacetlyases and serve transcriptional repressor functions. Indeed, RUNX1 is important in repressing CD4 expression in double-negative thymocytes and mature CD8+ T-cells. The choice to act as an activator or repressor is influenced by post-translational modifications of the RUNX protein in the core binding factor [16], [17] and through the recruitment of selective factors to a specific promoter [18], [19].
Three recent publications have highlighted a role for CBF-β in Vif function [20]–[22]. Specifically, CBF-β is capable of binding to Vif and this interaction increases Vif mediated APOBEC3G degradation. A recent study has presented evidence that CBF-β binding by Vif influences RUNX responsive genes [23]. These studies suggest that the Vif protein has evolved to bind CBF-β. The presumption made in the first two studies is that this mechanism is specific to protection of the virus from APOBEC3G. However, the follow-up study suggests the intriguing possibility that Vif may be binding to CBF-β in order to influence RUNX mediated gene expression.
Our current work suggests that RUNX1 may be an important HIV-1 LTR-binding factor that serves a role in latency. RUNX1, as a T-cell specific transcription factor capable of suppression or activation of target promoters, may alter the transcription of the viral LTR during cellular infection. We find that inhibition of RUNX1 by a small molecule inhibitor synergizes with the HDAC inhibitor SAHA to activate HIV-1 latency. Our findings indicate that RUNX1 with its binding partner CBF-β may act to repress LTR transcription and show that this repression might be countered by the viral Vif protein.
To assess the involvement of RUNX1 in HIV-1 transcription, RUNX1 and CBF-β expression plasmids were co-transfected with an HIV-1 molecular clone (pNL4-3) into HeLa cells (Fig 1A). Twenty-four hours post transfection, cells were harvested and protein extracts were prepared and subjected to Western blot analysis of HIV-1 protein expression. Interestingly, the transfection of either RUNX1 or CBF-β alone reduced HIV Gag expression in a dose dependent manner (Fig. 1A, compare lanes 2+3, 4+5 and 7+8 to lane 1). Transfection of RUNX1 and CBF-β together reduced Gag expression to nearly undetectable levels (Fig. 1A, lanes 6, 9, 10 and 11).
We next tested the effect of RUNX1 and CBF-β on spreading HIV-1 infection of Jurkat T-cells. In brief, Jurkat cells were transfected with 0.04, 0.2 or 1 ug of pMaxGFP control, pRUNX1 or pCBF-β expression vector using Amaxa nucleofection. Observation of GFP expression in transfected control cells showed that the transfection efficiency was >75%. Western blot analysis was performed to confirm over-expression of CBF-β and RUNX1 (Supplementary Figure S1). Twenty-four hours after transfection, cells were infected with the NL4-3 virus and cell culture supernatants were sampled at 2, 4, 6, 8 and 10 days post infection. Quantification of reverse transcriptase activity (RT) in the culture supernatants showed typical HIV-1 spreading infection in the cells transfected with the control vector (Figure 1B, C). However, increasing the transfected amounts of either RUNX1 (Fig. 1B) or CBF-β (Fig. 1C) produced reduced HIV-1 replication; with RT levels repressed to 43% and 30% of the control on day 10. Conversely, knock down of RUNX1 or CBF-β expression by ∼60% in a cell line model of latency induced a two-fold increase in RT production (Supplementary Figure S2). Taken together, these data are consistent with an important role served by RUNX1 and CBF- β in HIV-1 replication.
To further characterize the effects of RUNX1 and CBF-β on the transcription of an integrated HIV-1 LTR, we used the TZMbl reporter cell line. TZMbl cells are derived from HeLa cells to express CD4 and HIV-1 entry co-receptors, CXCR4 and CCR5; they contain an integrated HIV-1 LTR driving expression of firefly luciferase and a second integrated HIV-1 LTR driving the beta-galactosidase reporter gene. TZMbl were transfected with a Tat expression vector, together with RUNX1 or CBF-β expression vectors or a combination of the two. Forty-eight hours after transfection, the cells were harvested and equal protein amounts were used to determine β-galactosidase activity (Fig. 2A). We observed that the transfection of one microgram of either RUNX1 or CBF-β expression vector suppressed LTR-driven β-galactosidase expression by 88% and 90% respectively. Transfecting both vectors together repressed LTR-driven β-galactosidase expression by 94%.
We also tested the effect of RUNX1 and CBF-β on the basal activity of the HIV-LTR in 293T cells. 293T cells were transfected with 1 ug pLTR-GL3 (an HIV LTR driven luciferase reporter), 0.5 ug of RUNX1 or CBF-β expressing plasmid and increasing amounts of CBFβ or RUNX1 respectively (Fig. 2B). Transfection of CBF-β alone only modestly repressed LTR-driven transcription (Fig. 2B, 88% of control, 3rd column). Cotransfecting 0.25 ug of RUNX1 plasmid further repressed reporter expression to 62% of baseline (Fig. 2B, 4th column). Transfection with RUNX1 alone reduced reporter activity to 47% of control (Fig. 2B, 5th column). Increasing the levels of transfected CBF-β further reduced activity to a modest degree (Fig. 2B, compare columns 5 to 6 and 7). RUNX1 and CBF-β over-expression had no effect on either CMV or EF1 control promoter constructs and were capable of activating a murine leukemia virus promoter known to be responsive to RUNX1 (BXH2 LTR) (Supplementary Figure S3). Taken together these data demonstrate that RUNX1 and CBF-β are capable of suppressing the promoter activity of integrated or un-integrated HIV-1 LTRs, which is consistent with our earlier findings [24].
The DNA binding site, consensus sequence TGYGGT [25], for the RUNX proteins has been described for several promoters including MHC I, KIR, the BXH2 LTR and RUNX1 itself [25]–[28]. In order to identify potential binding sites in the HIV-1 LTR the TransFac matrix of RUNX1 binding sites was used to search the DNA sequence of the NL4-3 provirus from U3 through Gag using a search algorithm available from the University of Pennsylvania [29]. This analysis revealed 10 potential binding sites in the LTR and none in the 5′ portion of the Gag coding region. Of these 10 sites, six exist in the positive orientation, while four reside on the complementary strand (Fig. 3A and 3B). The conservation of these potential sites was analyzed using fifty-eight clade B sequences that contain a complete 5′LTR from the Los Alamos HIV Sequence Database (Fig. 3B) [30]. This analysis revealed that sites 3, 4 and 5 are not well conserved amongst clade B viruses. However, the remaining seven sites are well conserved. Indeed, these sequences appear to be better conserved than other regions of the LTR, suggesting that they may be selected for function.
Chromatin immuno-precipitations (ChIP) were performed to evaluate binding of RUNX1 to the potential LTR sites. It is known that CBF-β binds to RUNX1 and increases its affinity for DNA; thus, our ChIP analysis was performed in the context of both proteins. Chromatin was isolated from the HIV-1 latently infected ACH2 T-cell line, sheared by sonication, and immunoprecipitated with antibodies for RUNX1, CBF-β or no antibody control. The association of RUNX1 or CBF-β with HIV-1 LTR DNA or Gag DNA as a negative control was measured by qPCR (Fig 3C). ChIP analysis revealed the binding of RUNX1 and CBF-β to the LTR but not Gag DNA.
To specifically identify RUNX1 binding sites within the LTR, mutant reporter constructs were generated. Site-directed mutagenesis was used to change the final two nucleotides (consensus GT will be changed to TG) in potential RUNX binding sites 2, 5, 6 and 7 (Fig. 3A, B). This alters the important fifth residue in the binding site that has previously been shown to abrogate RUNX binding [31]. The four mutant reporter constructs were then used to evaluate RUNX1 responsiveness in 293T cells, which were transfected with 1 μg reporter (pGL3-LTR WT, pGL3-LTR mut2, pGL3-LTR mut5, pGL3-LTR mut6 and pGL3 mut7), 0.5 μg of CBF-β expression plasmid and 0.5 μg of RUNX1 expression vector. Forty-eight hours after transfection, the cells were harvested, lysed and 20 μg of protein extract used for luciferase assay (Figure 3D). RUNX1 and CBF-β expression repressed WT LTR to 25% of control (compare columns 1 and 2). Mutants 5, 6 and 7 were also repressed by RUNX1 and CBF-β. However, mutant 2 was not repressed by RUNX1 and CBF-β expression (compare columns 3 and 4). This suggests that the predicted binding site 2 is a physiological RUNX1 binding site in the LTR.
Recently it was shown that HIV-1 Vif can bind to CBF-β and that this interaction increases APOBEC3G degradation [20]–[22]. The ability of Vif, a cytoplasmic protein, to bind CBF-β may alter the ability of CBF-β to function as a cofactor for the RUNX proteins. CBF-β itself has no nuclear localization signal. Instead, it relies on binding to a RUNX protein partner to be ferried into the nucleus [11]. We wondered if binding of CBF-β by Vif may sequester CBF-β in the cytoplasm. To test this hypothesis, Vif and CBF-β expression vectors were transfected alone or in combination into HeLa cells (Fig 4A). 24 hours after transfection, the cells were stained for Vif, CBF-β and the nuclear marker Lamin B. Vif alone localized to the cytoplasm; and CBF-β alone was seen throughout the cell (Panel A, 2nd column), which is consistent with prior studies [11]. However, in cells that over expressed both Vif and CBF-β (Panel A, 3rd and 4th column), CBF-β was localized entirely in the cytoplasm consistent with the notion that Vif sequesters CBF-β in the cytoplasm.
Based on the published association of Vif with CBF-β and the above data, we explored Vif expression as a model for CBF-β depletion. 293T cells were transfected with 0.1 μg LTR-luciferase reporter and Vif expression vector (Fig 4B). Luciferase activity at forty-eight hours post-transfection was determined. Transfection of increasing amounts of Vif led to increased luciferase expression. Vif binds CBF-β through its n-terminal region, and mutation of residues 21 and 38 in Vif disrupts this binding [20], [22]. To test that Vif binding to CBF-β is required to mediate the observed increase in LTR expression, we utilized a vector that expresses a Vif protein lacking residues 23–43 (Vif ΔD). Transfection with Vif ΔD showed a small, but not dose responsive, increase in LTR driven luciferase expression.
To further confirm the involvement of Vif in rescuing RUNX1 mediated repression of the LTR, we employed the LTR mutant 2 reporter construct that was demonstrated above (Fig 3D) to be un-responsive to RUNX1 or CBF-β overexpression (Fig 4C). Transfection of neither Vif WT nor Vif ΔD was able to substantially increase Luciferase expression from this vector, supporting that the Vif effect on the LTR involves RUNX1 and CBF-β.
The ability of Vif to counteract repression mediated by RUNX1 and CBF-β in a spreading infection was monitored using a ΔVif virus. Jurkat cells were again used as they have been classified as ‘permissive’ due to their lack of APOBEC3G expression and ability to support the replication of viruses lacking Vif [32], [33]. Jurkat cells were transfected with 0.04 μg pRUNX1 or pCBF-β (Fig 4D and E). Twenty-four hours post-infection cells were infected with NL4-3 or NL4-3 ΔVif. Reverse transcriptase activity at day 10 demonstrated no difference in the susceptibility of the viruses to RUNX1 overexpression (Panel D). However, NL4-3 ΔVif was significantly more sensitive to CBF-β overexpression (Panel E). CBF-β repressed WT virus to 58% of the control, while NL4-3 ΔVif was repressed to 23% of the control. This data confirms that Vif is able to counteract CBF-β repression of HIV infection.
In order to determine the relevance of RUNX1 expression in HIV-1 infection in human patients we examined the expression of RUNX1, CBF-β and RUNX3 in primary T-cells. Analysis of CD4+ T-cell populations (Supplementary Figure S4) shows that upon activation of naïve CD4+ T-cells (a population relatively refractory to infection) the expression levels of RUNX1, RUNX3 and CBF-β drop (correlating with greater susceptibility to infection). Interestingly, the expression of these proteins in naïve cells is fairly uniform, but expression in the memory pool (the primary target cells during infection) is much more variable. This person-to-person variability leaves open the possibility that RUNX1 expression levels may correlate with clinical outcomes.
To determine the relevance of RUNX1 expression in HIV-1 infection in human hosts we compared RUNX1 expression to viral load and CD4+ T-cell counts in viremic HIV-1 patients who were not on therapy. Previous studies have shown that two promoters drive RUNX1 expression: a proximal and distal promoter. Each promoter codes a transcript that varies in the 5′ coding region and studies suggest that one isoform may be more important than the other in terms of function [34], [35]. In keeping with this, no significant correlation was seen between total RUNX1 levels and either viral load or CD4+ T-cell counts (Data not shown). However, when examining the specific levels of the promoter proximal RUNX1 transcript in memory CD4+ T-cells we noted a significant negative correlation with viral load and a significant positive correlation with CD4+ T-cell count (Fig 5A and 5B). This was in contrast to promoter distal RUNX1 transcripts that showed no correlation with viral load or CD4+ T-cell count (Fig 5C and 5D). Memory CD4+ T-cells are the primary target of infection and account for the majority of virus seen in the blood of patients. Therefore, these findings suggest the intriguing possibility that RUNX1 expression may have a role to play in the clinical progression of HIV-1 in patients.
The benzodiazepine compound Ro5-3335 was recently identified as an inhibitor of RUNX1:CBF-β interaction and their functions in transcriptional regulation and hematopoiesis [24]. Suppression of RUNX1 or CBF-β by siRNA was capable of reactivating latent cells, suggesting that a pharmacologic inhibitor of RUNX1/CBF-β function may have a similar effect (Supplementary Figure S2). To test this hypothesis the Jlat model of HIV-1 latency [36] was used wherein viral reactivation can be followed by GFP expression. Jlat cells were treated with DMSO, 0.5, 5 or 50 μM Ro5-3335 for 72 hours and re-activation was assayed by flow cytometry (Fig 6A). Treatment with 50 uM Ro5-3335 induced a 2.2-fold increase in the number of GFP positive cells (1.9 to 4.2%). This two-fold activation is similar in magnitude to the effect seen with siRNA (Supplementary Figure S2). We next tested if the effect of Ro5-3335 could be increased by treatment with a second drug. For this purpose, we chose the HDAC inhibitor suberoylanilide hydroxamic acid (SAHA), also known as Vorinostat, which has previously been used to reactivate latently infected T-cells in HIV-1 patients. Treatment of Jlat cells with 1 μM SAHA increased the number of GFP positive cells by 5.1-fold (1.9 to 9.6%) – similar to the induction of RT activity in ACH2 (Supplementary Figure S5). Treatment with increasing concentrations of Ro5-3335 increased the percentage of GFP positive cells in a multiplicative fashion. Treatment with 50 uM Ro5-3335 was capable of activating another 2.9-fold above SAHA alone, for a total activation of 28%. The combination of Ro5-3335 and SAHA yielded similar results in ACH2 (Fig 6B) and TZMbl (Fig 6C) cell lines. Of potential concern to the development of any therapeutic is the issue of toxicity. To address this we treated JLat, ACH2, TZMbl and J-LTR-G (a Jurkat cell line that carries an integrated LTR-GFP reporter) with SAHA and Ro5-3335 (Supplementary Figure S6). Ro5-3335 alone had minimal effects on cell viability. Treatment with SAHA induced significant cell death in all four cell types. Treatment of JLat, ACH2, TZMbl and J-LTR-G with SAHA and increasing concentrations of Ro5-3335 induced further cell death beyond SAHA alone (Fig S5, compare final four data sets for panels A–D). Interestingly, JLat and ACH2 had the greatest cell death (15% and 12% live when treated with SAHA and 50 uM Ro5-3335) as compared to TZMbl and J-LTR-G (50% and 44%). JLat and ACH2 also experienced greater cell death with SAHA treatment alone (55% and 51% live vs 81% and 63% live). ACH2 and JLat both carry complete copies of the virus, whereas TZMbl and J-LTR-G are reporter only cell lines. It is tempting to posit that the increased cell death seen in ACH2 and JLat is due to reactivation of the virus. The multiplicative effect on HIV reactivation seen with Ro5-3335 and SAHA in the three cell lines tested is indicative of possible synergy.
True synergy implies that two compounds are working on the same pathway or mechanism. There are two likely points of interaction that would cause synergy between Ro5-3335 and SAHA. RUNX proteins are capable of recruiting HDACs [19] and this may be happening at the HIV-1 promoter. Alternatively, the inhibition of HDACs by SAHA will induce broad changes in gene expression and this may include changes in RUNX1 and CBF-β expression. To test the later point, we treated cells with SAHA and measured the expression of CBF-β, RUNX1 and RUNX3 by quantitative RT-PCR (Fig 6D–F). Treatment of Jlat with SAHA triggered increased expression of RUNX1 and RUNX3 (Panel D) and a similar pattern was observed for ACH2 (Panel E). Treatment of TZMbl with SAHA also induced RUNX1 expression and induced CBF-β as well (Panel F).
To further verify that the observed synergy of SAHA and Ro5-3335 is due to the involvement of RUNX1 we asked if siRNA knockdown of RUNX1 would synergize with SAHA in the same way as the drug. ACH2 (Fig 6G) and TZMbl (Fig 6H) cells were again transfected with 50 pMol of siRNA against RUNX1 and then treated with either DMSO or 10 μM SAHA at 24 hours post transfection. Twenty-four hours post drug-treatment RNA was extracted from the cells and activation was measured by quantitative RT-PCR against Gag (ACH2) or U5 (TZMbl). Treatment of ACH2 with siRNA alone induced a 4.8-fold increase in expression. siRNA treatment reduced expression of RUNX1 and CBF-β by 52% and 63% respectively. Treatment with SAHA induced a 6.2-fold increase. Treatment with both siRNA against RUNX1 and SAHA led to a greater than 300 fold activation. In TZMbl cells, a similar, but smaller effect was seen. Treatment with siRNA or SAHA alone induced a 2.1 and 2.7-fold increase respectively. Treatment of TZMbl with both siRNA against RUNX1 and SAHA yielded a 6.5-fold increase in transcription. These results are consistent with that seen with Ro5-3335 and SAHA and strongly suggest that RUNX1/CBF-β inhibition may improve the ability of SAHA to activate HIV-1.
The elucidation of RUNX1 involvement in HIV-1 latency described above is heavily dependent upon cell line models of latency. Although these minimalistic systems are useful in determining the possible mechanism of action of a given protein or pathway, they do not represent an ideal model system for evaluating potential therapeutic intervention. The gold standard for evaluating latency is the ability to reactivate latent virus from the PBMCs of HIV-1 patients on suppressive therapy. In order to evaluate the effect of Ro5-3335 in primary cells we obtained PBMCs from five HIV-1 patients on suppressive therapy who had undetectable viral loads for at least 6 months. PBMCs were then treated with SAHA and Ro5-3335 to induce reactivation. In brief, 10×106 PBMCs were divided as needed for multiple conditions and placed in 10 ml RPMI +10%FBS. Cells from one patient were treated with 250 nM SAHA, 250 nM SAHA plus 5 uM Ro5-3335 or DMSO control. A SAHA dose of 250 nM was chosen as it has been shown to be equivalent to the concentration of available drug in the sera of patients [37]. Cells from two patients were treated with the above combinations plus 5 uM Ro5-3335 alone. Finally, cells from the final two patients were incubated with the three conditions plus 1 uM phorbol myristate acetate (PMA). Twenty-four hours after treatment the cells were collected, a small portion (∼300 k cells) was used for flow cytometry to detect activation of T-cells by Ki67 and cell death by vital stain. RNA was extracted from the remaining cells and used for RT-qPCR to detect HIV Gag mRNA (Fig 7A). The background level of Gag mRNA varied from 0.005 to 0.0633 copies/106 GAPDH. In three of the five patient samples treatment with SAHA induced a noticeable increase in Gag mRNA ranging from 1.4 to 6.7-fold. In all five patient samples treatment with SAHA and Ro5-3335 increased the levels of Gag mRNA beyond the levels seen in SAHA treatment (from 3.2 to 75-fold). A Wilcoxon matched pairs test showed a p-value that was approaching significance (p = 0.0625). Control cultures treated with 5 uM Ro5-3335 alone showed no increase in activation as compared to control. Flow cytometric analysis of the cells revealed no increase in T-cell activation (Fig 7B) or cell death (Supplementary Figure S5E).
In this study, we report a role for the RUNX family of transcription factors in repressing HIV-1 transcription driven by the viral LTR. The RUNX1 protein is involved in fate determination of T-cells and control of CD4 expression [12], [14], [38], [39] making its potential involvement in HIV replication of physiological interest. We have identified interaction of RUNX1 and the co-factor CBF-β with the viral LTR through a potential binding site (Fig 3) and that alteration of RUNX1 and CBF-β expression alters viral replication (Fig 1).
Work in in the field of HIV Vif biology has identified a role for CBF-β in Vif mediated degradation of APOBEC3G [21], [22]. We have uncovered new mechanisms for the involvement of CBF-β in HIV infection in the context of Vif expression, demonstrating that another functional consequence of Vif:CBF-β interaction is rescue of the virus from transcriptional repression by RUNX1 (Fig 4). A Vif mutant lacking the region necessary for CBF-β binding [20], [22] is incapable of counteracting RUNX1 repression of transcription, thus codifying the role of Vif in protecting the viral promoter.
Perhaps most interestingly we show that repression of RUNX1 activity by a pharmacologic inhibitor (Ro5-3335) is capable of synergizing with the HDAC inhibitor SAHA (Fig 6 and 7). True synergy implies that two compounds are working on the same pathway or mechanism. There are two likely points of interaction that would cause synergy between Ro5-3335 and SAHA. RUNX proteins are capable of recruiting HDACs [19] and this may be happening at the HIV-1 promoter. Alternatively, the inhibition of HDACs by SAHA will induce broad changes in gene expression and this may include changes in RUNX1 and CBF-β expression. Blocking RUNX1 function during SAHA treatment (via either siRNA or drug) significantly increases the activity of SAHA on the LTR. SAHA has been the recent drug of choice employed in studies attempting to reactivate latently infected memory T-cells – a necessary step in clearing the latent reservoir [37], [40]–[43]. Although the drug potently activates cells in culture and has a measurable effect on viral transcription in the T-cells of patients, it has not been successful in reducing the percentage of infected memory cells [44]. Combination therapy in which SAHA is delivered alongside a RUNX inhibitor may provide greater activation of virus, which in turn could lead to greater cell death or greater response by anti-HIV CD8+ T-cells.
It is of note that Ro5-3335 and a related drug have crossed the HIV literature before. A 1991 screen by Hsu et al. identified Ro5-3335 as a potential inhibitor of HIV-1 transcription [45]. However, the exact mechanism of this inhibition is unknown and several follow up studies paint a very complex picture. Another analysis of potential Tat inhibitors determined that Ro5-3335 did not inhibit Tat-TAR interaction [46]. We found recently that Tat binds RUNX1 with high affinity and inhibits Tat-mediated transcription together with CBF-β [24]. Combined with data presented here, our findings suggest that RUNX1 potentially influences Tat transactivation, and is the true target of Ro5-3335.
An analysis of Ro5-3335 and a related drug (Ro 24-7429) determined that they had no suppressive effect in chronically infected cells [47]. A later analysis by Cupelli and Hsu concluded that the drug might act at the level of initiation [48]. A clinical trial using Ro 24-7429 failed to reduce viral load and the presence of infectious virus in patient plasma, or to increase CD4 T-cell count [49]. Indeed, in our hands the activation of HIV-1 by Ro5-3335 happens in a timeframe beyond what was examined in the initial studies. This may explain why no inhibition was seen in chronically infected cells and suggests that the drug acts in a positive fashion on a fully integrated and chromatinized promoter.
What remains unanswered is how RUNX proteins might shape the clinical outcomes during HIV infection. Analysis of CD4+ T-cell populations (Supplementary Figure S4) shows that upon activation of naïve CD4+ T-cells (a population relatively refractory to infection) the expression levels of RUNX1, RUNX3 and CBF-β drop (correlating with greater susceptibility to infection). These findings are similar to what has been demonstrated in mice [13], [50], [51], although the detection of RUNX3 in CD4+ T-cells suggests a possible difference between humans and rodents. Interestingly, the expression of these proteins in naïve cells is fairly uniform, but expression in the memory pool (the primary target cells during infection) is much more variable. Although microarray databases show general expression of RUNX1 and CBF-β across tissues [52], no definitive studies have yet been performed in human cells which examine expression in specific T-cell subsets. Data presented in figure 5 shows a significant negative correlation between RUNX1 expression and viral load, as well as a significant positive correlation between RUNX1 expression and CD4+ T-cell counts in the absence of therapy. However, further research is needed to identify the causal relationship that drives this correlation.
The human sample collection protocol was approved by the NIH Clinical Center Institutional Review Board as part of a separate ongoing study. Written informed consent was obtained in all cases and all applicable protections of patient rights and privacy applied. For this study specific samples were requested from the sample bank based on given criteria.
Adherent cell lines (293T and TZMbl) were maintained in DMEM supplemented with 10% fetal bovine serum, L-glutamine and penicillin/streptomycin. 293T and TZMbl cells were transfected using Lipofectamine LTX (Invitrogen) according to the manufacturer's instructions.
Suspension cell lines (Jurkat, Jlat, J-LTR-G and ACH2) were maintained in RPMI supplemented with 10% fetal bovine serum, L-glutamine and penicillin/streptomycin. Jurkat and ACH2 cell lines were transfected using the Amaxa Nucleofector (Lonza) with reagent kit V and programs X-005 or T-014 respectively.
Cells for chromatin immunoprecipitation (ChIP) were cross-linked using 1% formaldehyde for 10 minutes at 37°C. Following crosslinking, cells were lysed in 1% SDS, 10 mM EDTA and 50 mM Tris-HCl pH 8.1 and sheared by sonication to less than 1000 bp. Lysate was clarified by centrifugation and diluted 1∶10 with 0.01% SDS, 1.1% Triton X-100, 1.2 mM EDTA, 16.7 mM Tris-HCl and 167 mM NaCl for antibody incubation. Following overnight incubation with antibody, complexes were precipitated using protein A/G agarose beads and washed with low salt (0.1% SDS, 1% Triton X-100, 2 mM EDTA, 20 mM Tris-HCl and 150 mM NaCl), high salt (0.1% SDS, 1% Triton X-100, 2 mM EDTA, 20 mM Tris-HCl and 500 mM NaCl) and lithium chloride (0.25 M LiCl, 1% NP40, 1% deoxycholate, 1 mM EDTA and 10 mM Tris-Hcl) buffers. Complexes were eluted in 1% SDS, 0.1 M NaHCO3 and reverse crosslinked by overnight incubation at 65°C. Proteinase K was used to digest remaining protein before DNA was cleaned using phenol:chloroform and precipitated prior to qPCR with primers for the LTR (174-464) or Gag (bp 1852-2153). Data is represented as signal for each primer pair relative to total input DNA (before IP).
Although positive signal in this LTR primer set might suggest that the relevant RUNX binding site is located between base pairs 174-464, the chromatin used in this analysis was sheared to less than 1000 bases meaning that the binding could be anywhere within the LTR.
RUNX binding site mutants were created in pLTR-Luc using QuikChange XL site directed mutagenesis kit (Stratagene) and primer pairs RUNXMut2, RUNXMut5, RUNXMut6 and RUNXMut7 (see below). Mutants were verified by sequencing.
HeLa CBF-β knockdown (KD) cells (5×106) were transfected with the Vif expression vector pNL-A1 (2.5 μg), the CBF-β vector pCBF-β (1 μg) or were cotransfected with both vectors (2.5∶1 plasmid ratio). Total amounts of transfected DNA were adjusted to 5 μg using empty vector DNA as appropriate. Three hours after transfection, cells were trypsinized and seeded onto cover slips. Cells were fixed 24 hr later in methanol (10 min, −20°C) and then stained with rabbit antibodies to Vif (Vif93; 1∶100) or CBF-β (Thermo Fisher; 1∶100). Nuclear membranes were stained with a mouse monoclonal antibody to lamin B (RDI; 1∶100). Bound antibodies were visualized by Texas-Red or Cy2-conjugated secondary antibodies (Jackson Labs; 1∶100). Images were collected on a Zeiss LSM410 confocal microscope using a Plan-Apochromat 63x/1.4 oil immersion objective (Zeiss).
To analyze the expression levels of RUNX1 and CBF-β in primary human T-cells, I sorted resting and activated naïve and memory CD4+ and CD8+ T-cells from PBMCs. In brief, PBMC were cultured overnight in the presence of absence of SEB to broadly activate T-cells. Flow cytometry was then used to sort CD3+ T-cells into different populations. Resting cells in the unstimulated population were identified as CD69- and activated cells from the SEB treatment as CD69+. Cells were then further classed as naïve (CD27+, CD45RO−) or memory (CD45RO+, CD27−). Finally, T-cell populations were sorted by the presence of CD4 or CD8 T-cell co-receptor. For the reactivation work, cells were analyzed for cell death using Live/Dead fixable Aqua (Life Technologies) and stained for CD3 and Ki67. Populations of T-cells were used to prepare RNA using Trizol reagent and RNA was submitted for RT-qPCR for RUNX1.
Data represented in graphical form is always the average of at least three replicates with standard deviation. Statistical significance was determined using an unpaired Student's t-test (cell culture) Mann-Whitney (Primary cells) or a Spearman's exact test with a cutoff of p<0.05.
Primers. HIV U5 F CTGCATGGGATGGAGGA
HIV U5 R GTTAGCCAGAGAGCTCCCAG
HIV Gag2 F GGTGCGAGAGCGTCAGTATTAAG
HIV Gag2 R AGCTCCCTGCTTGCCCATA
RUNXMut2 F atccttgatctgtggatctcacacacacaaggctacttcc
RUNXMut2 R ggaagtagccttgtgtgtgtgagatccacagatcaaggat
RUNXMut5 F CCAGGGAGGTGTGTGCTGGGCGGGACTG
RUNXMut5 R CAGTCCCGCCCAGCACACACCTCCCTGG
RUNXMut6 F CTGTACTGGGTCTCTCTGTGTAGACCAGATCTGAGCCT
RUNXMut6 R AGGCTCAGATCTGGTCTACACAGAGAGACCCAGTACAG
RUNXMut7 F GGGTCTCTCTGGTTAGCACAGATCTGAGCCTGGG
RUNXMut7 R CCCAGGCTCAGATCTGTGCTAACCAGAGAGACCC
GAPDH F GCTCACTGGCATGGCCTTCCGTGT
GAPDH R TGGAGGAGTGGGTGTCGCTGTTGA
RUNX1 F GATGGCACTCTGGTCACTGTGA
RUNX1 R CTTCATGGCTGCGGTAGCAT
RUNX3 F TTCCTAACTGTTGGCTTTCC
RUNX3 R TAGGTGCTTTCCTGGGTTTA
CBF-β F ACAGCGACAAACACCTAGCC
CBF-β R CAGCCCATACCATCCAGTCT
RUNX1 Proximal F TGCATGATAAAAGTGGCCTTGT
RUNX1 Proximal R CGAAGAGTAAAACGATCAGCAAAC
RUNX1 Distal F TGGTTTTCGCTCCGAAGGT
RUNX1 Distal R CATGAAGCACTGTGGGTACGA
|
10.1371/journal.pcbi.1004530 | Comparative Analysis of Yeast Metabolic Network Models Highlights Progress, Opportunities for Metabolic Reconstruction | We have compared 12 genome-scale models of the Saccharomyces cerevisiae metabolic network published since 2003 to evaluate progress in reconstruction of the yeast metabolic network. We compared the genomic coverage, overlap of annotated metabolites, predictive ability for single gene essentiality with a selection of model parameters, and biomass production predictions in simulated nutrient-limited conditions. We have also compared pairwise gene knockout essentiality predictions for 10 of these models. We found that varying approaches to model scope and annotation reflected the involvement of multiple research groups in model development; that single-gene essentiality predictions were affected by simulated medium, objective function, and the reference list of essential genes; and that predictive ability for single-gene essentiality did not correlate well with predictive ability for our reference list of synthetic lethal gene interactions (R = 0.159). We conclude that the reconstruction of the yeast metabolic network is indeed gradually improving through the iterative process of model development, and there remains great opportunity for advancing our understanding of biology through continued efforts to reconstruct the full biochemical reaction network that constitutes yeast metabolism. Additionally, we suggest that there is opportunity for refining the process of deriving a metabolic model from a metabolic network reconstruction to facilitate mechanistic investigation and discovery. This comparative study lays the groundwork for developing improved tools and formalized methods to quantitatively assess metabolic network reconstructions independently of any particular model application, which will facilitate ongoing efforts to advance our understanding of the relationship between genotype and cellular phenotype.
| Scientists have been mapping the chemical reactions cells use to grow and manage waste since before enzymes were first identified more than 150 years ago. The model yeast Saccharomyces cerevisiae has one of the most extensively studied metabolic networks, including at least 25 metabolic network models published since 2003. If iterative model improvement refines the metabolic network map, we would expect eventual convergence to a full, accurate metabolic network reconstruction. In this study, we looked for evidence of such convergence through comparative analysis of 12 genome-scale yeast models. We conducted simulations and evaluated model features such as predictive accuracy, genomic coverage and the included metabolites and reactions. We found that no single metric for evaluating models can adequately summarize important aspects of model quality. In some cases, we observed tradeoffs between model predictive accuracy and network coverage. We found evidence of incremental changes to the network reconstruction, but not marked shifts in model predictive ability or other metrics clearly arising from changes to the network alone. This work has broader implications to computational reconstruction of metabolic networks for any organism, and suggests that there is opportunity for refocusing the model building process to better support mapping cellular metabolic networks.
| Efforts to map metabolic networks—to describe the full network of anabolic and catabolic biochemical reactions occurring within a cell—have advanced from early biochemical studies of fermentation [1] to contemporary efforts to algorithmically generate pathway diagrams from genomic sequence [2]. Such pathway maps may be augmented with additional metadata to build a digital “reconstruction” of an organism’s metabolic network. In turn, such organism-specific reconstructed metabolic networks may be further supplemented to build mathematical models that are capable of simulating metabolic fluxes [3]. Recently, research efforts have focused on improving the ability to quickly build genome-scale metabolic network models of metabolism and to improve their predictive accuracy [2,4–6]. Comparatively less effort has been spent exploring opportunities for knowledge discovery arising during the process of network reconstruction prior to mathematical simulation [7]. In this work, we emphasize the distinction between metabolic network “reconstruction” and metabolic network “model”. Emphasizing this distinction facilitates an effort to resolve the relative contributions to model predictive accuracy or error arising from the metabolic network structure itself (the “reconstruction”) from those arising from mathematical parameters chosen when building a simulatable metabolic network “model” from the network reconstruction. While a variety of ad hoc quantitative metrics have been applied to evaluate improvements in metabolic network models, quantified assessment of the progress of the underlying reconstructions over time is a nascent effort [8]. This may be, in part, due to the fact that the number of models is so much greater than the number of extensively curated reconstructions. The relative difficulty of curating a comprehensive metabolic network reconstruction compared to generating a draft model is highlighted by the fact that there are currently more than 2,600 functional draft models [6], but only Escherichia coli [9] and Saccharomyces cerevisiae [10] reconstructions have been extensively updated multiple times and revised through curation efforts by multiple research groups over a multi-decade timescale. Since simulation results are more amenable to quantitative analysis than reconstruction quality, reconstructions have generally been assessed indirectly, often in the context of model performance via manuscript discussion of scope (the number of genes, reactions, or metabolites in a model), standards compliance, naming or annotation conventions, reputation of the group that built a particular model, or predictive performance of a derived model for a particular phenotype of interest (commonly used phenotypes include gene essentiality, substrate utilization, growth rate, or product production) [11]. This indirect approach to assessing metabolic network reconstruction quality bears a risk, however, because the model building process itself can obscure important details about the underlying reconstruction (particularly knowledge limitations that may be useful for informing future investigation [12,13]).
The standard reconstruction protocol includes converting a reconstruction to a mathematical model for subsequent debugging [3]. Thus, the ability to create a functional model has come to serve as a minimum threshold for defining the scope of a “draft reconstruction”, and the distinction between “reconstruction” and mathematical “model” has become blurred. Model developers are free to take different approaches when parameterizing model features such as objective function (i.e., biomass composition) [14,15], media definition [16], and reference lists of “essential” genes used for benchmarking model performance [17]. Model developers may use different approaches to gap-filling [4,18,19], trimming dead-end metabolites, establishing an objective function, and adding transport and exchange reactions. In fact, optimization-based approaches have been applied to successfully improve model essentiality predictions by adjusting these parameters [20]. Such algorithmic approaches can improve the predictive performance of a model even in the absence of any changes to the underlying metabolic reconstruction.
Using model performance to drive iterative improvements to metabolic network reconstruction has led to two perverse consequences. First, if two models of the same organism give different predictions, how can a researcher determine whether the differences arise from differences in the reconstructed network or from differences in model parameters? We have previously observed that algorithms such as OptKnock [21] can suggest different targets for metabolic engineering efforts when applied to different models of the same organism (unpublished data). Second, a single metabolic network model can provide only limited information about the quality of the underlying metabolic network reconstruction because there are so many degrees of freedom associated with deriving a model from a reconstruction [7,8]. Comparative analysis of multiple models, now possible at scales not previously feasible [22], provides an opportunity to address these challenges of single-model analysis. Our approach is to conduct comparative analysis of yeast metabolic network models that have been published in the past two decades, while controlling for differing modeling assumptions with a standardized model biomass function, media definition, and common sets of genes considered in the evaluations. An additional benefit of comparative analysis of models spanning a multi-decade timescale is the opportunity for evaluating model predictive performance on data that was not available at the time of reconstruction, which can provide a useful independent validation procedure and provide insights into the degree of overfitting possible in these models through the manual reconstruction process that is very difficult to ascertain otherwise.
At least 25 models of the Saccharomyces cerevisiae metabolic network have been published since 2003 [5,11,17,20,23–40]. Each of these models has been applied successfully to research efforts focusing on advancing biotechnology [41], mapping genotype to phenotype relationships in cellular physiology [42], or developing new methods in computational biology [43]. Previously, researchers have combined comparative analysis of three of these models (iFF708, iLL672, and iND750) with experimental data to refine characterization of cellular phenotypes in 16 environmental conditions [44], and developed tools to facilitate model matching and comparison for synchronous investigation or building composite models [45]. Another two models (Yeast 5 and iMM904) have been evaluated for predicting growth rates of a prototrophic gene deletion library in 20 different conditions [46]. More recent efforts have begun comparative analysis of a broader range of these models [22]; however, we are not aware of previous large-scale comparative analysis efforts that modify model objective functions, reference phenotype lists, and simulated media composition to evaluate the underlying metabolic network reconstructions built for S. cerevisiae.
In this study, we conducted 161 in silico screens of predicted single gene essentiality using 18 different simulated media conditions, 12 different yeast metabolic models, and 13 different biomass definitions. We used this range of simulation parameters to standardize choices made in the development of various models, thus facilitating evaluation of the underlying network reconstructions. Using a binary growth/no growth assessment metric, we evaluated model predictions of the essentiality of three different lists of “essential” genes compiled from literature and database review. Additionally, we conducted simulations of aerobic growth with constraints on glucose, oxygen, and nitrogen exchange reactions singly or in combination to evaluate the correlation between model predictions of maximally achievable biomass flux values and reported experimental growth rates. We also conducted 10 in silico screens of pairwise gene essentiality by different models using their default media and biomass definitions, and compared model predictions to 32,488 gene pairs annotated as synthetic lethal in the Saccharomyces Genome Database. All code for our analysis is available as S1 File. Our key findings include the following. (1) Changes in model scope reflect a history of iterative reconstruction development via collaboration between groups—in other words, each model contains evidence of its history, with stylistic and content evidence of the specific model from which it is derived. Knowledge is propagated between models, but there is also risk of error propagation. Therefore, it is important to revisit assumptions made when earlier models were originally built when evaluating newer models. (2) Model updates tended to fall into two major categories, model scope expansion (i.e. the inclusion of new metabolic processes) or subsequent refinement (i.e. including essentially the same sets of processes but working to improve accuracy). There was a pattern in analyzing the models’ ability to predict KO essentiality that accuracy on average reduced when model scope expansion was done and then improved on subsequent reconstructions aimed at improving the same set of processes. (3) For each model, single-gene essentiality predictions were affected by parameters external to metabolic network structure, such as simulated medium, objective function, and the reference list of essential genes. (4) The correlation between model predictions of maximum biomass flux correlate and reported growth rates are the same for all models when only a single exchange reaction is constrained, but the correlation between model prediction and reported growth rate differ among the models when multiple exchange reactions are simultaneously constrained to experimental values. This difference can be attributed to changes in metabolic network reconstructions independent of model parameters. (5) The predictive ability for single-gene essentiality did not correlate with predictive ability for our reference list of pairwise synthetic lethal genes. Thus, we conclude that the reconstruction of the yeast metabolic network is generally improving, and have demonstrated that comparative model analysis contributes to reconstruction improvement. There remains great opportunity for advancing our understanding of metabolic function through continued efforts to improve the reconstruction of the yeast metabolic network.
We compiled summary statistics for functional yeast metabolic models published since 2003 (Fig 1), including the number of metabolites, reactions, dead-ends, gene-associated reactions, and genomic coverage. When the models are ordered chronologically, none of these statistics demonstrates continuous improvement, perhaps reflecting the differing research objectives that motivated the development of each new model, but also demonstrating the limited ability of any individual statistic for fully describing model quality (e.g. reducing the number of genes in a reconstruction is an improvement if previous iterations included misannotated genes, but such a reduction of genomic coverage could be considered a worse statistic).
We observed a general increase in the number of genes included in models over time, but this increase was not uniform. We also found that increased genomic coverage could be a result of modelers including genomic features that are no longer considered genes. For example, the Biomodels.db model accounts for the greatest coverage of the yeast genome, but includes 28 open reading frames currently annotated as “dubious”, or unlikely to encode a functional protein. Additionally, increases in genomic coverage did not imply improved predictive accuracy: Yeast 6 includes fewer genes than Yeast 5, but improves single-gene essentiality predictions
Similarly, the number of metabolites and reactions and the proportion of dead-end metabolites in yeast metabolic models has generally, but not uniformly, increased over time—and does not coincide with improved predictive ability. For example, the number of metabolites and reactions in Yeast 7 is much larger than that in Yeast 6, though they have the same overall MCC. In contrast, iND750 contains fewer metabolites and reactions than its progenitor model iFF708. The portion of dead-end metabolites—those metabolites that are consumed but not produced in the network or vice versa—has also varied among models, and does not correlate with predictive accuracy.
Next, we evaluated model scope by comparing genomic coverage and the metabolites that could be cross-identified with Chemical Entities of Biological Interest (ChEBI) identifiers with the annotation included with the models. We were unable to directly compare model reactions because of the lack of standardized reaction identification between models, and the lack of an external reaction reference database identifier in any yeast metabolic model, a current limitation for interoperability and comparison in our field. We found that models clustered in groups that reflect their historical development [10], but these clusters differ between gene and metabolite comparisons (Fig 2). Models clustered in 4 groups when comparing genomic coverage: 1) Versions 4–7 of the Consensus Reconstruction; 2) iMM904, iMM904bs, iAZ900, and iTO977; 3) a looser cluster of iFF708, iIN800, and iND750; and 4) the Biomodels.db model. A row-aligned comparative table of genes in each model is included as S1 Table.
Model similarity clusters differed when based upon ChEBI identifier-annotated metabolites, and the clusters were more tightly linked to the research group most closely related to the development of a group of models. When clustered by annotated metabolites, the resulting 5 groupings consisted of 1) Versions 5–7 of the Consensus Reconstruction; 2) iND750, iMM904, iMM904bs, and iAZ900; 3) iFF708, iIN800, and iTO977; 4) Version 4 of the Consensus Reconstruction and 5) the Biomodels.db model.
Improving model ability to predict the essentiality of individual genes for growth has not been the primary motivating factor for developing each new yeast metabolic network model, but essentiality predictions have generally been reported with the publication of each new model to demonstrate their accuracy and utility. However, direct comparison between reported predictive values is complicated by differing simulation conditions. In this study, we did not find a strong trend of continuing improvement in model ability to predict single gene essentiality over time. Instead, we found evidence of an iterative process in which model scope changes (typically leading to a decrease in average predictive accuracy), followed by subsequent curation leading to improved prediction by descendant models (Fig 3). We found that both iIN800, with its expansion of the reconstruction of lipid metabolism, and iND750, with its expansion of compartmentalization, had a lower overall Matthews Correlation Coefficient (MCC) for single-gene essentiality predictions than their progenitor model, iFF708. Subsequently, iMM904 refined iND750, and made more correct predictions of single gene essentiality. Similarly, Yeast 6 refines Yeast 5 and improves predictive ability, but the focus of Yeast 7 on expanded scope does not lead to as large an improvement in single-gene essentiality predictive ability, and iTO977’s focus on expanding model scope to cover some protein modification processes and to provide a scaffold for integrating transcriptomic data does not lead to an improvement in predicting single-gene essentiality compared to its progenitor model, iIN800.
The Biomodels.db model was generated in a methods-development effort to improve automated reconstruction and annotation. The algorithm underlying the Biomodels.db model prioritizes connectivity and defines “functionality” as the ability of the model to predict growth using a generic biomass definition objective function. The Biomodels.db annotates genes with a different nomenclature than the ORF format used by other models, so was incompatible with our gene essentiality screen. We evaluated the Biomodels.db model by other comparative metrics, but did not evaluate its FBA performance here.
We conducted 161 simulated genome-wide single gene deletion screens for gene essentiality by conducting flux balance analysis with an objective of maximizing biomass flux. We used different combinations of simulated media and biomass objective functions and compared model predictions to appropriate reference lists of essential genes, as described in the Materials and Methods section. We found that no single model predicted essential genes best in all simulations (Fig 4). In our simulations, the iAZ900 model had the highest single MCC we calculated (0.83) for a particular condition, and the Yeast 7 model had the highest overall MCC across all the conditions for which we calculated (0.61). As a point of comparison, we calculated a MCC of 0.61 based on the reported results of a gene essentiality screen with a recent model of the E. coli metabolic network [47], which is, along with yeast, widely considered the best studied genome-scale metabolic network model to date.
Although it had the highest observed MCC in one condition, the iAZ900 model did not perform as well in other simulations—it also had the lowest MCC (0.17) for an out-of-sample screen using the iFF708 biomass definition, a very permissive set of exchange reactions, and a reference gene list based upon SGD-reported phenotypes. When the exchange reactions are constrained to reflect a simulated glucose minimal defined media, the iAZ900 MCC for the iFF708 biomass increases to 0.55. Such ranges of model predictive ability were observed for all models across differing simulation conditions, highlighting the importance of controlling for model parameter variation when attempting to compare metabolic network models of a particular organism. In the specific case of iAZ900, the excellent performance of its best condition reflects the authors’ goal in developing iAZ900 –to use an algorithmic approach to improve the iMM904 model by maximizing agreement with a list of genes essential genes reported to be essential by Kuepfer et al. [17]. The reference list of essential genes used in the development of iAZ900 originates from a screen of non-essential genes in the yeast knockout collection in glucose-limited defined medium [48]. This reference list for training the algorithm is one of the reference lists we used for comparative evaluation. iAZ900 did not perform as well at classifying genes as essential when using other reference gene lists. Thus, iAZ900 demonstrates that high model performance can be achieved by one metric, but there is the usual tradeoff between sensitivity and specificity when attempting to generalize a specific metabolic network model to predict phenotypes in new conditions.
Our observation that model performance was influenced by the reference list of genes considered essential when attempting to evaluate model predictive ability demonstrates that the definition for gene essentiality is another parameter that may be tuned as model developers refine their model. In our simulations, model MCC was higher on average when calculated relative to the SGD-based list of essential genes for five of the models (iND750, iIN800, iTO977, Yeast 6, and Yeast 7), and higher relative to the Kuepfer-based list of essential genes for the remaining six models (iFF708, iMM904, Yeast 4, iAZ900, iMM904bs, and Yeast 5). These two groups do not correspond to the clusters identified when comparing model genomic coverage or the clusters identified when comparing annotated metabolites.
All models predicted gene essentiality better when glucose was the simulated primary carbon source than when galactose, glycerol, or ethanol were the primary sources. However, since the reference gene list used for the non-glucose carbon sources was based upon a single screen, we could not determine whether this reflects limitations in the reconstruction of the non-glucose metabolic network, or strain and laboratory-specific effects in the reference data. Historically, the metabolism of non-glucose carbon sources has received less biochemical characterization than glucose metabolism in yeast.
Since the objective function is a tunable parameter that is independent of metabolic network structure, we normalized the objective by selecting a biomass definition that each model could satisfy, as described in the Flux Balance Analysis—Biomass Definition subsection of Methods, below. Thus, we began differentiating between model parameter improvements and network structure improvements to compare the reconstruction underlying different models more directly. We performed Flux Balance Analysis of the metabolic network models using both the biomass definition provided by the model authors, and the biomass function used for the iFF708 model, and found that for all models with different biomass definitions than the iFF708 model, the model predictive power was affected by the objective function used (Fig 5). In every case but the Yeast 4 model, model predictions were better using the model default biomass objective than the iFF708 objective, suggesting that model developers have achieved improved predictive accuracy in part by modifying the objective function, and such improvements have been achieved independently of refinements to the biochemical network reconstruction itself. This approach is not meant to imply that modifications to an objective function would be conducted solely to improve a predictive metric: refinements to the biomass definition also reflect improved measurement of biomass composition and changes to model scope. We selected a common biomass definition for our analysis to evaluate the impact of this particular model parameter.
We conducted FBA-based comparison of media- and objective-normalized model predictions of maximum achievable biomass fluxes with the aerobic growth rates reported by Österlund et al. for “N-limited” and “C-limited” conditions (we did not simulate anaerobic growth since most of the models we are examining do not predict anaerobic growth on a minimal medium). We found that the model predictions of maximally achievable biomass flux correlated with the previously reported “N-limited” growth rates with a correlation of 0.994 when nitrate or nitrite exchange fluxes were constrained to previously reported uptake rates (S3 Table).
The “C-limited” simulations reflected a different behavior. When we constrained the glucose exchange reaction alone, all models had a 0.816 correlation with the reported growth rates (Fig 6A). However, the growth rates labeled “C-limited growth aerobic” by Österlund et al. are not linear over the range of constraints imposed on the glucose exchange reaction, suggesting that carbon (glucose) flux is not the sole growth-limiting factor, particularly at the higher range of glucose flux constraints. The ratio of glucose exchange flux to oxygen exchange flux would be expected to strongly influence maximum achievable biomass flux due to stoichiometric constraints on the oxidation of glucose [49]. We tested model behavior against this expectation by conducting FBA with both glucose and oxygen exchange reactions constrained to values reported by Österlund et al. [37]. When glucose and oxygen exchange reactions were both constrained to experimental values, we observed that the models segregated to 2 groups: biomass flux predictions made by 7 models (iFF708, iIN800, Yeast 5, iTO977, iMM904, and iMM904bs) correlated with observations with a correlation >0.9, and predictions made by the remaining models (Yeast 4, Yeast 6, Yeast 7, iAZ900) had lower correlations (Fig 6B).
We used one-norm minimized FBA [50] to find an explanation for this difference in model predictions and observed unrealistically large fluxes through internal reactions along with unusually large exchange fluxes in models that overpredict biomass flux in high glucose:oxygen growth simulations. Through repeated FBA and manual investigation of high-flux loops, we found that the low-correlation models all had a flux through a mitochondrial aspartate transport reaction. This reaction is not associated with a gene in the iAZ900 model (reaction id “ASSPt2M”), but is annotated with yeast open reading frame YPR021C in the Yeast 4, Yeast 6, and Yeast 7 models (reaction ids “r_1163”, “r_1117”, and “r_1117”, respectively). YPR021C encodes Agc1p, a protein that “fulfills two functions… glutamate transport into mitochondria … and … aspartate-glutamate exchanger within the malate-aspartate NADH shuttle” [51]. Subsequently, we also found this reaction in the iND750 (“ASPt2M”), iIN800 (“AGC1_2”), iMM904 (“ASPt2m”), iMM904bs (“ASPt2m”), Yeast 5 (“r_1117”), and iTO (“AGC1_2”) models. We did not find this reaction in iFF708 or the Biomodels.db models, which do not include mitochondria as a separate compartment. We did not find literature support for including yeast mitochondrial aspartatate transport as reconstructed in these models. Thus, investigating erroneous predictions of maximum biomass flux by four models at simulated high glucose:oxygen flux states allowed us to identify a reconstruction error common to all compartmentalized models, an error that is independent of model parameters.
When we removed this reaction from the models, we found that it did not affect the predictions for the high correlation models, and improved all remaining correlations to >0.9, with the exception of the Yeast 4 model, which still over-predicted the maximum biomass flux at high glucose:oxygen exchange constraint ratios (Fig 6C).
Using the models as distributed (i.e., with tuned biomass definitions and default exchange reaction constraints), we conducted a simulated screen of all pairwise deletions for 10 models (iFF708, iND750, iIN800, iMM904, Yeast 4, iAZ900, iMM904bs, Yeast 5, iTO977, and Yeast 6). Using a strict definition of synthetic lethality in which neither gene is individually essential, but are pairwise essential for growth, we found that the MCC for model prediction of synthetic lethal gene pairs ranged from 0.04 to 0.12, when compared to a list of synthetic lethal gene pairs that we generated using the Saccharomyces Genome Database Yeastmine tool [52] (Fig 1). Additional summary statistics of these screens are included in Fig 7.
Surprisingly, we found that the MCC for synthetic lethal interactions did not correlate with the MCC for single-gene essentiality (R2 = 0.0253). The relatively low predictive ability of these metabolic network models for synthetic lethal gene pairs may be attributed in part to the fact that the reference list of synthetic lethal gene pairs is not well-established due to the challenge of conducting pairwise gene deletion screens in vivo [53], the fact that predictions of multiple perturbations to a genetic network require more complex analysis [54], and synthetic lethality phenotype observations may be greatly influenced by experimental design [16]. We anticipate that evaluating and improving constraint-based phenotypic predictions of multiple-gene deletions will advance hand in hand with efforts to experimentally explore gene interaction networks.
The source code used for conducting these simulations is included as S1 File. Detailed results of all simulations conducted, including lists of true and false predictions for each model in each simulated single-knockout screen, are attached as S2 File.
The main findings of this study are that the iterative publication of many models over the past two decades has generally, but not universally, improved yeast metabolic network reconstructions when assessed by a range of metrics. We also found that current approaches to model development and annotation can hinder direct assessment of the underlying reconstruction. Thus, this study serves to provide an overview of the historical development of yeast metabolic network models over the past two decades, provide methods for evaluating future metabolic models, and highlight opportunities for improving the reconstruction of the yeast metabolic network. It also raises issues that should be considered so that metabolic reconstruction efforts can best contribute to investigations of metabolic processes.
When we directly compare functional models of the yeast metabolic network published over the past two decades, we note generally increasing trends in genomic coverage (using either verified or total number of open reading frames annotating model reactions), number of metabolites, and reactions. We do not discern strong trends in number of dead-end metabolites in models, the percentages of reactions associated with genes, or in predictive accuracy for synthetic lethal genetic interactions. When comparing flux balance predictions with model default or standardized simulated media, objective functions, and reference lists of “essential” genes, we find that predictive power for single-gene essentiality has gradually improved. However, the observed trends are not uniform across all models and simulations we conducted. Each model has its own strengths and weaknesses; as demonstrated in a previous comparative study, “different models may be preferable for use in different applications” [44].
The uneven progress in improving model performance metrics reflects the historical path of iterative model refinement. Different models are developed to address different research questions, and are not necessarily focused on improving gene essentiality predictive accuracy. Thus, each new model may advance (or regress) when compared to previous models depending on the metrics used to assess the model. This is particularly evident when examining the relative performance of single gene essentiality predictions (Fig 3). For example, iND750 greatly expanded compartmentalization in the yeast metabolic reconstruction, but had lower predictive ability for single-gene essentiality than the earlier iFF708 model in our analysis. Development of the iAZ900 model demonstrated the utility of optimization-based procedures for improving model prediction, so it has the highest MCC for the reference conditions used for model development, but not the highest overall MCC across all conditions. The iIN800 model expanded the reconstruction of lipid metabolism, and the iTO977 model expanded the scope of yeast metabolic models to facilitate transcriptomic analysis. As new models integrate and improve upon earlier models, a path dependency on previous modeling or reconstruction efforts emerges. This necessary relationship can lead to iterative improvement, but can also propagate errors and complicate assessment of the reconstruction of the yeast metabolic network. Further, as models have been developed, different research groups have used different tools to validate their models (such as different lists of genes reported to be essential in a particular strain background or experimental condition). Thus, no model should be considered “best” or definitive for all applications. Examining simulations across multiple models may be a prudent approach for building confidence in predictions.
The results of our comparison of predicted maximum achievable biomass flux to measured growth rates emphasize that model users must take great care when imposing multiple constraints prior to conducting FBA, or when interpreting experimental growth rate measurements as being attributable to a single limiting nutrient. If a model user is attempting to compare simulated predictions with observed growth rates that scale linearly with the concentration of a single limiting nutrient, our results suggest that they should test to ensure that the model is operating within a linear range in which the desired nutrient is in fact the sole factor limiting predicted flux to the objective. If operating within such a regime, users could confidently scale FBA results by either varying model parameters (such as ATP maintenance demands) or by simple linear transformation of objective values found via FBA. However, model users must be wary of discontinuities arising from shifts in the limiting nutrient (such as from glucose to oxygen).
The high correlation between predicted biomass flux and observed growth rates in high glucose:oxygen exchange constraint ratio regimes is surprising not only because “normal yeast mitochondrial structures are disrupted when glucose levels are high” [55], but because c. 2,000 genes are regulated by the diauxic shift [56]. These changes are dependent upon concentration, rather than flux [57]. Thus, it is likely that there are sets of constraints that should be applied to a metabolic network for condition-specific modeling. We did not observe that such constraints were necessary for predicting the stoichiometrically-determined shift from glucose-limited to oxygen-limited maximal growth rate over a range of glucose to oxygen exchange flux ratios. Differentiating universal constraints (such as chemical stoichiometry) from condition-specific constraints appears to have great potential as a fruitful avenue for future research efforts.
We found that model gene essentiality predictions are biased by factors that are not reflective of the accuracy or completeness of the metabolic network reconstruction. Such factors include reference gene lists, choice of objective function for flux balance analysis, and simulated media used for in silico screens. However, it is likely that standardizing these factors (as we have done in this study) for comparing models is not sufficient for assessing the quality of the metabolic network reconstruction; model builders must make other choices when developing a model that is amenable to simulation from a network reconstruction. For example, since different models use different approaches to fill gaps in the known metabolic network or to ascribe catalytic function to a poorly characterized yeast genes, different models are likely to include different hypothetical transport or biochemical reactions with different levels of evidence or confidence in the accuracy of the functional role of a protein.
Reconstructing a metabolic network provides an opportunity to highlight areas of uncertainty to productively guide future research efforts. This opportunity is distinct from the utility of mathematical simulation of fluxes using metabolic network models. In fact, deriving a metabolic network model from a reconstruction can obscure the knowledge gaps or uncertainty that can be highlighted during the process of network reconstruction [7]. This risk is particularly acute where poorly understood portions of metabolism are not clearly implicated in the research to which a model is being applied, or when highlighting knowledge gaps or ambiguity may hurt model performance according to metrics used to assess predictive performance or scope. For the yeast models we compared, some of which use current annotation and model exchange protocols and formats, there is no mechanism for a model user to identify knowledge limitations discovered during reconstruction of the underlying network, nor is there sufficient annotation describing the specific techniques used to address such limitations when the model was constructed within the published model itself.
The current state-of-the-art for metabolic network modeling presents a significant barrier to entry for researchers who are not familiar with the idiosyncrasies of each model because these idiosyncrasies are not sufficiently documented within the model structure itself. Thus, though we observed that model predictions of gene essentiality are generally better for models evaluated with a simulated medium containing glucose as the primary carbon source than model predictions when using ethanol, glycerol, or galactose as a carbon source, we cannot conclusively attribute the improvement in glucose-essential prediction to improvements in the reconstruction of the biochemical reaction network because there is no clear mechanism for separating the information contained in the underlying reaction network reconstruction from the modeling assumptions and choices made in deriving a particular metabolic network model. Similarly, we cannot conclusively attribute the relative lack of improvement in predictions with non-glucose carbon sources among models to errors in the reconstruction rather than faulty model assumptions, idiosyncratic objective function definition (i.e., model overfitting), or biological factors such as condition-dependent gene essentiality for genes included in the reference list of “essential” genes.
Selecting appropriate data sets for model validation presents an additional challenge to the reconstruction effort. Specifically defining the media and conditions in which a given gene (or combination of genes) is essential remains an ongoing and important area of research to advance our understanding of metabolism. In the absence of well-defined reference phenotypes, we cannot confidently ascribe the low predictive ability for pairwise essentiality to errors in metabolic network reconstruction, uncertainties in synthetic lethal phenotypes, or physiological processes which are not metabolic, such as gene regulation or cell cycle checkpoint events. Further evaluating and improving model predictive performance for conditional essentiality will be greatly assisted by use of new prototrophic yeast strains and genetic screens in specifically designed media [46].
Despite these methodological challenges, there is benefit to comparing metabolic network models for the same organism for filling gaps and for identifying mistakes and opportunities for further expansion of the metabolic network reconstruction. We note, for example, that iron metabolism is important to mitochondrial function, but is not included in these models. None of the models include folate, chitin, or hypusine in the biomass definition, a model building choice that leads to false negative gene essentiality predictions and dead-end metabolites, but also highlights opportunities for expanding the reconstruction of the yeast metabolic network. Similarly, since most models have been validated with laboratory results from strains originally designed to facilitate genetic investigation (strains which bear auxotrophic markers in their genetic backgrounds) [48], it is likely that the reconstruction of portions of the yeast metabolic network (such as nitrogen and sulphur metabolism) is incomplete. Updating the reconstruction in support of research with a new prototrophic yeast mutant library [46] provides an exciting opportunity for refining our understanding of yeast metabolism.
As different groups refine yeast metabolic network reconstructions and models, there should be a convergence to a full, accurate reconstruction of the complete network. We do not observe evidence that supports marked changes in the reconstruction, such of a marked shift in model predictive ability or genomic coverage in our analysis of models published to date. Further, recent work has observed that many enzymatic functions are not included in existing models and reconstructions [8]. Thus, the effort to reconstruct the yeast metabolic network is incomplete. Increased efforts to expand the scope of reconstruction, such as including signaling and regulatory network processes, may provide a way to advance efforts to reconstruct organism-specific networks.
Our analysis suggests that metabolic network reconstruction efforts could benefit from emphasizing the distinction between reconstruction of known or hypothesized metabolic function, and metabolic models developed for particular applications. A reconstruction may be improved, but model performance may drop by some metrics (for example, adding a parallel metabolic pathway could lead to false negative gene essentiality predictions in the absence of regulatory constraints blocking an available network branch, or conversely, adding condition-specific regulatory constraints could hinder predictive value for conditionally-essential genes in other environments). Similarly, model performance could be enhanced in some cases by removing established biochemistry (and such a choice would be defensible if modeling a particular environment in which a portion of the metabolic network was unavailable due to regulation). Thus, we find that no single metric we used to compare metabolic network models is sufficient to evaluate the progress of the yeast reconstruction efforts. Models should be assessed by gene essentiality predictions, as well as the extent of evidence and annotation for included information, the size of the network, and network connectivity metrics. Unfortunately, current methods for annotating the workflow of model development makes such analysis challenging. In some cases, erroneous model predictions have been computationally corrected through changes that cannot be annotated in exchange formats. Thus, they become obscured, rather than highlighted in a way that would better facilitate further investigation. Similarly, although great efforts have been expended to assess the evidence for information in the published models, none of the SBML files we evaluated included confidence scores or full annotation of literature sources, so these assessments remain internal to a development group and are not effectively propagated to subsequent model users. This is in part a historical artifact—many existing standards such as SBML are intended to distribute models, rather than fully annotated reconstructions. Efforts such as the definition of the Pathway Tools schema [58] lay important ground work towards broader community participation in improving the process of metabolic network reconstruction and metabolic model derivation. Though reconstructing metabolic networks has been the focus of biochemistry for more than a century, computational metabolic network reconstruction is still a young field with great contributions to make. Through this comparative analysis of yeast metabolic network models, we hope to contribute to the ongoing efforts to improve our understanding of metabolism through collaborative network reconstruction, and to highlight opportunities for improving the process of metabolic network reconstruction and model derivation.
All simulations were conducted on a laptop running Windows 7 (Microsoft) using MATLAB 2013a (MathWorks Corporation, Natick, Massachusetts, USA), with SBML Toolbox version 4.1.0 [59], COBRA Toolbox version 2.05 [60], and Gurobi Optimizer version 5.6 (Gurobi Optimization, Inc., Houston, Texas, USA). All code written for this study is included as S1 File.
Models were obtained from public repositories, supplemental information, or research collaborators and modified as follows:
Each of the models is provided in S1 File, formatted as.mat files containing the COBRA Toolbox data structure with any modifications to enable simulation.
Different sets of genes have been observed to be essential for growth in different conditions, and different lists have been used for previous evaluations of model predictive accuracy. Therefore, we generated our own reference lists for the current comparative analysis. We began with the list of 1,120 unique open reading frames annotated as essential by the Yeast Deletion Project (available at http://www-sequence.stanford.edu/group/yeast_deletion_project/downloads.html), then removed YCL004W and YKL192C, based upon literature review [64,65]. Since this list was generated from experiments using a complete medium, we used it as our reference for simulations of growth in a synthetic complete medium (our approach to defining simulated medium is described in the “Flux Balance Analysis—Medium Definition” section below).
For evaluating simulations of growth in a glucose-limited minimal medium, we supplemented the essential gene list with 441 additional open reading frames reported to induce auxotrophy upon deletion. This list was generated by downloading a list of ORFs annotated as auxotroph-inducing in the Saccharomyces Genome Database, removing those already included in the essential list and temperature-sensitive inositol auxotrophs, and modifying the list based on literature-based curation as described in the testYeastModel.m file, which is included in S1 File. Combining the essential ORF list with the auxotroph-inducing list resulted in a list of 1560 open reading frames as our reference list of genes considered essential in a minimal medium.
We also compiled reference lists of essential genes based on the screen of non-essential genes in the knockout collection in defined media with different carbon sources conducted by Kuepfer et al. [17]. This screen evaluated 4869 open reading frames included in the yeast knockout collection. Applying the stringent standard of a score of 0 for ORF essentiality, we classified 59 ORFs as essential in defined medium with glucose as the sole carbon source, 307 with galactose, 291 with glycerol, and 332 with ethanol. When comparing model predictions for these medium, we did not evaluate all ORFs in each model, but instead only characterized the subset of genes in the model that were also evaluated in the Kuepfer et al. screen. These lists of essential genes are included in the testYeastModel_kuepfer.m script, which is included in S1 File.
We built a reference list of synthetic lethal genetic interactions (pairs of ORFs that are not individually essential, but become essential when both are deleted) by querying the Saccharomyces Genome Database with YeastMine [52]. We built an XML query to search for interacting genes where the experiment type was listed as “Synthetic Lethality”. The resulting report was downloaded and imported to MATLAB to generate a list of 32,488 pairs of ORFs that have been annotated as synthetic lethal. The specific XML query is described in the analyze_double_results.m script, which is included in S1 File.
We note that any reference list of gene essentiality is dependent upon experimental conditions, so different researchers may construct such lists in different ways. The predictive accuracy of any model is a function of the standard used, so different reference lists are expected to affect the specific MCC value of its agreement with observation. Thus, it is particularly important that the same list of “essential” genes or gene pairs be used when comparing different models.
We conducted flux balance analysis of each model using both the model-default simulated medium composition, along with media formulations we defined in an effort to standardize model predictions. We defined the following media for our simulations: a minimal medium that enabled predicted biomass production for all the models in which glucose is the sole carbon source; a synthetic complete medium with glucose as the sole carbon source, which was based on previous computational screening efforts [66]; and the synthetic medium defined by Kuepfer et al., using glucose, galactose, glycerol, or ethanol as the sole carbon source.
The minimal medium was simulated by allowing unconstrained exchange of ammonia/um, oxygen, phosphate, sulphate, and setting a constrained uptake of glucose. The iMM904bs and Biomodels.db models did not predict growth using this medium when using their default biomass definitions (biomass definitions are described below in the “Flux Balance Analysis—Biomass Definition” section). To enable FBA, the simulated minimal medium for these models was supplemented: the iMM904bs model required iron exchange, and the Biomodels.db model required the amino acids L-tyrosine, L-lysine, L-isoleucine, L-arginine, L-histidine, L-methionine, and L-tryptophan.
The code used for setting the medium for each model is included in S1 File in the testYeastModel.m and testYeastModel_kupefer.m scripts.
We conducted FBA for all models except the Biomodels.db model with two different objective functions: first, we used the model’s default biomass definition, as included in the published version of the model; second, we used a common biomass definition as similar to the iFF708 biomass definition as each model’s exchange reactions and metabolites allowed. The Biomodels.db model did not predict growth with the iFF708 biomass definition, so we only used its default objective function to verify functionality. We selected the iFF708 biomass definition as a reference standard because it was the objective function that most models could satisfy. For example, iFF708 would not be able to satisfy the Yeast 5 biomass definition due to the expanded sphingolipid requirement in the latter. Our use of the common, older biomass definition was intended in part to separate model improvements that arise from improved reconstruction from those that arise due to a more specific biomass definition.
The code used to set the biomass definitions for each model is included in S1 File in the testYeastModel.m and testYeastModel_kupefer.m scripts.
Accounting for the seven media compositions and two biomass definitions described above, we conducted flux balance analysis to predict single-gene deletion growth phenotypes for each model in fourteen different conditions. We elected to use used a tight threshold of binary growth/no growth prediction when comparing model growth predictions to our reference lists of essential genes because flux balance analysis of metabolic network models may be less predictive for mutant growth rates than for a binary essential/non-essential gene classification [46] and because growth rate predictions may be tuned by adjusting model parameters such as ATP maintenance reaction demands or constraints on carbon source utilization reactions. For this study, a gene was considered to be predicted as essential only if flux balance analysis of a simulated mutant predicted a maximum flux to the biomass objective of less than 1 x 10−6 flux units.
The agreement between model gene essentiality predictions and the reference lists was quantified using the Matthews’ Correlation Coefficient (MCC) (eq 1) [67], a metric that considers true positive, true negative, false positive, and false negative predictions without any assumption of the frequency of observations in the reference dataset. MCC ranges from -1 (when model predictions are the exact opposite of the reference dataset) to +1 (when model predictions match the reference data set).
MCC= TP×TN−FP×FN(TP+FP)(TP+FN)(TN+FP)(TN+FN)
(1)
Where true positives (TP), true negatives (TN), false positives (FP), and false negatives (FN) are defined as in [17]: a true positive prediction is one in which the model predicts that a gene is not essential for biomass production, and the gene has been annotated as not essential. Values for each confusion matrix, along with lists of positive and negative predictions, are included as S2 File.
We also assessed model prediction of synthetic lethality, or double-gene deletion phenotypes, for 10 of the models. When comparing model predictions to the reference gene list, we defined true positive as predictions in which neither gene in a reported synthetic lethal pair is predicted to be essential by itself, and the pair is essential. We defined true negatives as predictions in which neither gene is predicted to be essential by itself, the pair is not predicted to be essential, and the pair is not reported to be synthetic lethal. We defined false positives as predictions in which neither gene is individually predicted to be essential for growth and the pair is predicted to be essential for growth, but the pair has not been reported to have a synthetic lethal interaction. We defined false negative as predictions in which neither gene is individually predicted to be essential and the pair is predicted to be non-essential, though a synthetic lethal interaction has been reported. We categorized incorrect predictions of single-ORF essentiality as “other errors”–such errors were not included in our MCC calculation, since they were accounted for in the in silico single gene knockout screen. We did not modify the models’ biomass definition or simulated medium composition for our double knockout simulation. We also note that our definition of synthetic lethal interactions, which requires a model prediction of greater than 10% of the predicted wild-type biomass flux, is an arbitrary, but strict requirement. It is likely that the MCC for synthetic lethal predictions would be influenced by the choice of minimum biomass flux, and we selected 10% as a representative example for this particular analysis. If slow-growing double mutants are scored as synthetic lethal in an in vivo screen, and included in our reference list of synthetic lethal pairs, a correct model prediction of low biomass flux could be scored as false negative.
Blocked reactions are reactions that cannot carry a flux in a given simulation condition; thus, the number of blocked reactions may change for a given model with different biomass definitions or different allowed exchange reactions. We used the fastFVA module [68] to count the number of blocked reactions for each model when all exchange reactions were allowed to carry flux, and using both the model default and the iFF708 biomass definitions.
Dead-end metabolites are metabolites that either participate in only one reaction, or can only be produced or consumed. Thus, they are a network feature that is not influenced by exchange reaction or biomass definition changes. We counted the number of dead-end metabolites in each model with the COBRA Toolbox detectDeadEnds function.
The code used for blocked reaction and dead-end metabolite analysis is provided in S1 File. Model-specific lists of blocked reactions and dead end metabolites are included as S2 Table.
Similarity of genomic coverage among models was assessed by hierarchical clustering based on pairwise distance of binary vectors of logical values for open reading frames included in a model’s reaction annotation (i.e., 1 if a given ORF is included in a model, or 0 otherwise). The binary vectors are presented as a heat map, and clusters are presented as a clustergram and scatterplot (generated with classical multidementional scaling) in Fig 2.
Different model developers have annotated metabolites in different ways, so we began our comparison of metabolites by expanding the annotation of models by adding identifiers from the Chemical Entities of Biological Interest (ChEBI) database [69] to metabolites where possible. We were able to establish ChEBI annotation for different subsets of metabolites in each model, so this comparison is, by necessity, less comprehensive than comparison of model genomic coverage. The Biomodels.db model annotates metabolites with multiple ChEBI identifiers (reflecting redundancy in the ChEBI database). We chose the first ChEBI identifier when comparing the Biomodels.db model with models derived from manual reconstruction. Other models did not include multiple ChEBI identifiers for annotated metabolites. Like genomic coverage, metabolic coverage was scored with a binary vector of logical values, and the comparison is presented as a heatmap, clustergram, and scatterplot.
A sorted list of genes by models and all code used for scope comparison are included as Supporting Information.
We compared predictions made by media- and objective-normalized models with the aerobic growth rates reported by Österlund et. al. [37] for “N-limited” and “C-limited” conditions (we did not simulate anaerobic growth since most of the models we are examining do not predict anaerobic growth on a minimal medium). We conducted flux balance analysis of each model after standardizing model objective functions to the iFF708 biomass objective, and then applying constraints to the glucose, oxygen, and nitrogen exchange fluxes, first individually and then in combination. We used the measured uptake values reported by Österlund et al. [37] as constraints for each of these exchange reactions.
|
10.1371/journal.pgen.1004380 | Spermatid Cyst Polarization in Drosophila Depends upon apkc and the CPEB Family Translational Regulator orb2 | Mature Drosophila sperm are highly polarized cells—on one side is a nearly 2 mm long flagellar tail that comprises most of the cell, while on the other is the sperm head, which carries the gamete's genetic information. The polarization of the sperm cells commences after meiosis is complete and the 64-cell spermatid cyst begins the process of differentiation. The spermatid nuclei cluster to one side of the cyst, while the flagellar axonemes grows from the other. The elongating spermatid bundles are also polarized with respect to the main axis of the testis; the sperm heads are always oriented basally, while the growing tails extend apically. This orientation within the testes is important for transferring the mature sperm into the seminal vesicles. We show here that orienting cyst polarization with respect to the main axis of the testis depends upon atypical Protein Kinase C (aPKC), a factor implicated in polarity decisions in many different biological contexts. When apkc activity is compromised in the male germline, the direction of cyst polarization within this organ is randomized. Significantly, the mechanisms used to spatially restrict apkc activity to the apical side of the spermatid cyst are different from the canonical cross-regulatory interactions between this kinase and other cell polarity proteins that normally orchestrate polarization. We show that the asymmetric accumulation of aPKC protein in the cyst depends on an mRNA localization pathway that is regulated by the Drosophila CPEB protein Orb2. orb2 is required to properly localize and activate the translation of apkc mRNAs in polarizing spermatid cysts. We also show that orb2 functions not only in orienting cyst polarization with respect to the apical-basal axis of the testis, but also in the process of polarization itself. One of the orb2 targets in this process is its own mRNA. Moreover, the proper execution of this orb2 autoregulatory pathway depends upon apkc.
| After completion of meiosis, the 64 cells in the spermatid cyst begin differentiating into sperm. Sperm are highly polarized cells and a critical step in their differentiation is spermatid cyst polarization. Spermatids are also polarized within the testis, with the heads of the elongating spermatids located basally, while the tails extend apically. We show that aPKC accumulates preferentially on the apical side of the cyst during polarization and is required to correctly orient cyst polarization with respect to the apical-basal axis of the testis. Unexpectedly, aPKC activity is spatially restricted by a mechanism that depends upon the CPEB family translational regulator orb2. orb2 is required to asymmetrically localize and activate the translation of apkc mRNAs during spermatid differentiation. In addition to correctly orienting the direction of cyst polarization, orb2 is required for the process of polarization itself. One of the orb2 regulatory targets in this process is its own mRNA, and this autoregulatory activity depends, in turn, upon apkc.
| Polarity plays a central role in a diverse array of biological contexts in organisms ranging from single cell bacteria to complex multicellular eukaryotes. In multicellular eukaryotes, the steps involved in establishing, maintaining, and transmitting polarity are typically controlled by an interacting set of evolutionarily conserved atypical protein kinase C-partitioning defective proteins (aPKC-PAR proteins). The classic model for polarity determination by the aPKC-PAR machinery is the establishment of the anterior-posterior axis in the C. elegans embryo [1]. Prior to fertilization, anterior determinants, the worm aPKC ortholog PKC-3, PAR-3 and PAR-6, are distributed in a complex around the entire cortex of the egg [2]–[4], while the posterior factors, PAR-1 and PAR-2, are cytoplasmic. PAR-2 is kept off the cortex by PKC-3 dependent phosphorylation, and a similar mechanism may apply to PAR-1 [5], [6]. Sperm entry induces a cytoplasmic flux that relocalizes the PKC-3/PAR-3/PAR-6 complex in the posterior to the anterior cortex. Following the removal of PKC-3 activity from the posterior, PAR-1 and PAR-2 are able to associate with the cortex. Cortical PAR-2 in turn prevents re-association of anterior determinants with the posterior cortex (for review: [7]). This generates a polarized cell in which the PKC-3/PAR-3/PAR-6 complex is distributed along the anterior cortex, while PAR-1/PAR-2 are localized on the posterior cortex. This process also serves to orient the mitotic spindle: the first cell division in the embryo is parallel to the anterior-posterior axes and as a consequence the two daughter cells receive different sets of embryonic determinants [1]. The aPKC-PAR machinery defines polarity in many other contexts besides the establishment of the anterior-posterior axis of the C.elegans embryo. Moreover, as in C.elegans, antagonistic interactions between aPKC-PAR family proteins are critical for establishing and propagate cell asymmetry in many systems (for reviews: [8]–[11].
In most of the well-studied model systems, the aPKC-PAR machinery is deployed to generate undifferentiated cells that are then able to assume different identities. However, polarity can also be a distinguishing feature of terminally differentiated cells. One example of a highly polarized fully differentiated cell is the mature D. melanogaster sperm. At one end of the mature sperm cell is the sperm head, which contains the highly condensed haploid genome encased in a multilayer membrane. The rest of the cell is the nearly 2 mm long flagellar axoneme tail, which is connected to the head by a centrosome-derived structure called the basal body. The formation of this polarized cell commences after meiosis is completed and the 64 interconnected spermatids begin the process of differentiation (Fig. 1A). Each haploid nucleus has a single basal body with a short axoneme surrounded by a membrane cap. In the first steps the basal body inserts into the nuclear envelope, where it functions to nucleate the assembly of the flagellar axoneme. The 64-cell cyst re-organizes so that the spermatid nuclei cluster together on the proximal or basal side of the cyst while the basal bodies and nascent flagellar axonemes localize to the opposite side. The flagellar axonemes then begin elongating towards the mitotic spermatogonia and stem cells at the apical end of the testis (Fig. 1A). At the tip of elongating tails are ring canals (that function to connect the 64 cells in the cyst) and a ciliary sheath that encases the bundled elongating flagellar axonemes. The remainder of the axoneme is ensheathed by mitochondria and a plasma membrane. Once the spermatids are fully elongated, the process of individualization begins. A special actin based structure, called the individualization complex (IC), forms from the mature spermatid nuclei, and then translocates down the axonemes, remodeling the membranes to separate the individual sperm tails and removing excess cytoplasm [12]–[16].
One of the critical steps in sperm differentiation is the polarization of the 64-cell spermatid cyst. There are multiple levels of polarization: At the cellular level, the round, seemingly symmetric spermatids must become polarized so that their nuclei are localized on one side, while the nascent flagellar initiate their growth on the opposite side. At the next level, polarization of individual spermatids in one cyst must be coordinated so that the all 64 of the spermatids in the cyst have the same orientation, nuclei clustered on one side of the cyst and nascent flagellar axonemes on the other. The two somatic cells that surround the cyst are also polarized: the head cyst cell encapsulates the clustered spermatid nuclei while the tail cyst cell covers the growing flagellar axonemes. During the following spermatid elongation and differentiation stages, the two somatic cells expand their volumes without divisions and retain their relative positions. Finally, the elongating spermatids must be oriented correctly with respect to the apical basal axis of the testes so that once differentiation is complete the mature sperm can be readily transferred into the seminal vesicle.
Here we show that aPKC is required to orient the direction of spermatid cyst polarization with respect the apical-basal axis of the testis. When apkc activity is compromised, the direction of polarization appears to be randomized. As in many other biological contexts, aPKC protein is asymmetrically localized to the apical side of the spermatid cyst during polarization. However, the mechanisms used to spatially restrict aPKC accumulation are different from the canonical cross-regulatory interactions between this kinase and other cell polarity proteins that normally orchestrate polarization. Instead, an mRNA localization pathway that is regulated by one of the two Drosophila CPEB family translational regulators, orb2, is responsible for spatially restricting the accumulation of aPKC protein during cyst polarization. We show that Orb2 binds to the 3′ UTR of a special apkc mRNA species, apkc-RA, which is only expressed in post-meiotic cysts, and regulates both its localization and translation. We also find that orb2 functions not only in orienting cyst polarization with respect to the apical-basal axis of the testis, but also in the process of polarization itself. One of the orb2 regulatory targets during polarization is its own mRNA, and this autoregulatory activity depends, in turn, upon apkc.
Although substantial amounts of aPKC can be detected in Western blots of testis extracts (not shown), much of this protein is likely to be somatically derived. In whole mounts of wild type testes probed with aPKC antibody, only a low level of unlocalized protein is detected in spermatocytes before and during the first and second meiotic divisions (Fig. 1B, arrowhead and arrow). However, after meiosis is complete and the 64-cell cysts show the first signs of polarization and axoneme nucleation, prominent aPKC signals can be detected. As shown in a newly polarized cyst in Fig. 1C, aPKC concentrates in “puncta” (arrow) that are clustered in a region of the cyst diametrically opposite to the spermatid nuclei. This region contains the nascent flagellar axonemes and it will form the tip of the elongating spermatid tails. As elongation proceeds and the structure of the elongating flagellar axonemes becomes more tightly organized, there is a band of aPKC protein located near the tip of the elongating sperm tails (Fig. 1D, arrow, insert). During elongation, the protein appears to be arranged into a series of thin stripes that run parallel to the bundled sperm tails and are located near if not at the termini of the 64 growing flagellar axonemes (Fig. 1E, arrow). These stripes remain associated with the tips of the elongated flagellar axonemes until after individualization commences (not shown).
In other biological contexts, proteins that function together with aPKC in cell polarization often exhibit similar or reciprocal localization patterns. To identify other potential players, we examined the expression of other classic polarity regulators Bazooka (Baz: Drosophila Par3), Discs large (Dlg) and Par-1 during spermatogenesis. As shown in Figure S1, high levels of Baz, Dlg and Par-1 were seen at the apical end of the testis, in the region that contains the germline stem cells and the mitotic spermatogonia/spermatocytes and the somatic support cells. These proteins seemed to be mostly expressed in the somatic support cells that surround the developing mitotic and meiotic germline cysts (Fig. S1A, C, E). However, unlike aPKC none showed localized accumulation later on during the differentiation of the spermatid cysts (Fig. S1B, D, F).
In many systems a critical step in the polarization pathway is the asymmetric localization of aPKC. Thus, a reasonable expectation is that the targeting of aPKC to the apical side of differentiating spermatid cysts will orchestrate some aspect(s) of cyst polarization. To explore this possibility we analyzed the effects of apkc mutations and RNAi knockdowns on sperm morphogenesis (Fig. 2). apkck06403 has a P-element insertion between the P2 and P3 promoters that could potentially affect mRNAs produced by the upstream P1 and P2 promoters (see map in Fig.3A). Since this allele is homozygous lethal, we looked for phenotypic effects in heterozygous mutant males. We also examined two viable hypomorphic alleles apkcex55 and apkcex48 that were generated by the excision of the apkck06403 P element [17] [18].
In wild type testes, the orientation of spermtid elongation is invariant: The sperm tails always extend towards the apical end of the testis while the sperm nuclei are pushed towards the base of the testis. As a result, spermatid nuclei clusters are not seen in the spermatogonia/spermatocyte region of the testis (Fig. 2E). However, when apkc activity is compromised, spermatid nuclei clusters appear at the apical side of the testes (Fig. 2B). Fig. 2A shows the frequency of testes that have differentiating sperm oriented in the opposite direction – the spermatid nuclei cluster located apically and the tails elongating basally. Whereas misoriented cysts were rarely if ever seen in wild type, about ¼ of the testes in males homozygous for either of the hypomorphic alleles had misoriented spermatid cysts. Within these testes, 5–20% of the cysts elongated in the incorrect direction. Arguing against simple background effects, we found that apkc was weakly haploinsufficient and incorrectly oriented spermatid cysts were also detected in males heterozygous for the strong loss of function allele, apkck06403. It is worth noting that unlike other genes that have been implicated in cyst polarization (e.g., the exocyst complex: [19]), the defect in these hypomorphic apkc mutants appears to be in choosing the direction of polarization and not in coordinating or executing the polarization process itself. This is suggested by the fact that all 64 of the spermatids in the mutant cysts undergo seemingly normal elongation but in the incorrect direction and we never observed spermatids within the same cyst elongating in opposite directions.
To provide further evidence that apkc is required to properly orient cyst polarization within the testis and also determine whether it functions in the germline, we knocked down apkc by combining UAS-apkc RNAi lines with a germline specific Gal4 driver nanos-Gal4, or a triple Gal4 line (pCOG-Gal4; NGT-40; nanos-Gal4), MTD, that drives higher levels of UAS dependent expression in germline cells in most stages of spermatogenesis [20], [21]. To control for off target effects from RNA interference, we tested two UAS-apkc RNAi lines, UAS-apkc::34332 and apkc:35140, that express double strand RNA, dsRNA-HMS01320 and dsRNA-GL00007, targeting two different exon regions that are common for all apkc mRNA species (Fig. 3A) (see materials and methods) [22].
Orientation defects like those observed in the hypomorphic mutants were also seen when apkc activity was reduced by RNAi knockdown (Fig. 2D). For example, at 22o C 18% of the nanos-Gal4/UAS-apkc:35140 testes had at least one incorrectly oriented spermatid cyst (Fig. 2A). Moreover, amongst the testes that had polarity defects, nearly 20% of the spermatid cysts were elongating in the wrong direction. The frequency of nanos-Gal4/UAS-apkc:35140 testes with incorrectly oriented elongating spermatid cysts increased from 18% to 45% when the temperature was raised from 22o C to 25o C. Even stronger effects were observed when UAS-apkc:35140 was combined with the triple Gal4 line MTD. In this case, every testis had multiple misoriented spermatid cysts (Fig. 2A). In addition to these orientation defects, the spermatids in testes from the MTD/UAS-apkc:35140 knockdown were shorter than wild type (not shown). This defect suggests that apkc activity might be needed to sustain the elongation of the flagellar axoneme.
Since several of the proteins that usually collaborate with aPKC do not seem to be expressed in differentiating spermatids, other mechanisms are likely important for localizing aPKC asymmetrically and facilitating its function in correctly orienting cyst polarization. One such mechanism would be mRNA localization. Support for this possibility comes from the studies of Barreau et al. [23]. They identified a special group of genes that are transcribed only post-meiotically and encode mRNAs that localize in a striking “comet” pattern at the apical end of the elongating spermatid tails. The “head” of the comet, where these mRNAs are most highly concentrated, is in the same region of the elongating tails as the aPKC protein stripes.
To determine if apkc mRNA is similarly localized in differentiating spermatids, we generated a set of oligo probes specific to the large exon that is common to all apkc mRNAs (Fig. 3A). apkc mRNAs can first be detected in the stem cell region of the testis and they persist until late in spermatogenesis (Fig. 4A). Significantly, in differentiating spermatid cysts apkc mRNAs are distributed in a “comet” pattern just like the post-meiotically expressed mRNAs identified by Barreau et al. (Fig. 4A, arrow). The highest concentration of apkc mRNA is in the comet “head” which is close to the leading edge of the elongating sperm tails, while there is a comet “tail” with lower levels that extends backwards towards the spermatid nuclei.
The apkc gene has four different promoters (P1–P4) and, together with an extensive set of alternatively spliced exons and UTRs, it is predicted to generate more than 10 different mRNAs (Fig. 3A). To identify apkc mRNA species that are expressed in the testis we used primers specific for different promoters and the alternatively spliced exons and UTRs (e.g., 1-2/3-4/PF-PR) for RT-PCR (Fig. 3A, 3B). These experiments indicate that testes have a complex mixture of many apkc mRNA species. Interestingly, one of these mRNAs, RA (P1: 5–10; PF-PR), has an unusually long 3′ UTR that is not found in any of the other predicted mRNA species.
As mRNA localization often depends upon special cis-acting elements in the 3′ UTR, we generated in situ probes specific for apkc-RA mRNA species. The apkc-RA expression pattern differs substantially from that of the “bulk” apkc mRNAs that hybridizes to the large common exon probe. Unlike “bulk” apkc mRNA, apkc-RA can not be detected in the pre-meiotic stages, while only low levels are present in the meiotic cysts (Fig. 4B, arrow head & Fig. S2A). Its expression is first detected in the round spermatids after meiosis is complete (Fig. 4B, white arrow), and increases to a maximum during spermatid differentiation (Fig. 4B, yellow arrows & Fig. S2A). Initially apkc-RA is distributed uniformly through the cyst; however, this changes when polarization commences and the spermatid nuclei start to cluster on the basal side of the cyst. At this point apkc-RA mRNAs can be detected near the apical tip of the nascent flagellar axonemes much like aPKC protein (Fig. S2B). As elongation proceeds, apkc-RA mRNAs accumulate in the characteristic comet pattern, and are most highly concentrated in the comet head near the apical tip of the elongating tails (Fig 4B, yellow arrows).
The apkc-RA mRNAs in the comet head are expected to be in close proximity to the aPKC protein stripes that mark the ends of the flagellar axonemes. When we simultaneously probed testes for apkc-RA mRNA and aPKC protein, the aPKC stripes partially overlap with the proximal side of the apkc-RA mRNA comet head (Fig. 4C). This finding would fit with the idea that the apkc-RA mRNAs in the comet head are a source of the aPKC stripes at the leading edge of the growing flagellar axonemes.
Several of the post-meiotic mRNAs identified by Barreau et al. have consensus CPEs (Cytoplasmic Polyadenylation Elements) in their 3′ UTRs. CPE motifs are recognized by a conserved family of RNA binding proteins, CPEBs, which in flies are known to function both in mRNA localization and translational regulation [24]–[26]. Proteins in this family have two RRM-type RNA binding domains in their C-terminal halves, while their N-terminal halves contain polypeptide sequences that have regulatory functions [27], [28]. Inspection of the unique apkc-RA 3′ UTR indicates that it has three canonical CPE motifs (Fig. 3A and Text S2). Several of the other apkc mRNA species (RD, RJ, RK, RL and RM), which share a different UTR sequence, have a single CPE sequence. Thus, an intriguing idea is that a fly CPEB protein binds to the RA mRNA species and perhaps to one or more of these other apkc mRNAs, facilitates their localization to the tip of the elongating sperm tails and controls their translation. Flies have two CPEB genes, orb and orb2, and both are expressed in testes. Of these, orb is not likely to have either of these functions as its message is not translated until after spermatid elongation is complete [29]. orb2, on the other hand, would be a plausible candidate for the regulatory factor. It is required at multiple steps during spermatogenesis including spermatid differentiation. We've found that one of its differentiation functions is to control the translation of post-meiotically expressed mRNAs like orb that have CPE motifs in their 3′ UTRs. Moreover, like these post-meiotic mRNAs, Orb2 protein is distributed in a comet pattern in elongating spermatids [29]. Two other findings lend support to the idea that orb2 might regulate the localization and/or translation of apkc-RA and perhaps other CPE containing apkc mRNAs species in the testis. First, we found that orb2 is required for the apical accumulation of aPKC in embryonic neuroblasts during asymmetric cell division [30]. Second, exogenous biotin tagged apkc-RA RNA can pull down Orb2 protein from adult fly brain extracts [31]. This hypothesis makes several predictions that we have tested for the apkc-RA mRNA species as it has a post-meiotic expression pattern.
The experiments in the previous section demonstrate that Orb2 co-localizes with apkc mRNAs in elongating spermatid cysts and binds directly to the 3′ UTR to an apkc mRNA species that is specifically expressed at this stage of spermatogenesis. If this association is important for apkc activity during spermatid differentiation, orb2 should also be required for correctly polarizing the 64-cell cysts. To test this prediction we examined spermatid cyst polarization in two different orb2 mutants, orb236 and orb2ΔQ. orb236 is a null allele that lacks the entire Orb2 protein coding sequence. Meiosis is blocked in orb236 and 16-cell cysts, which have duplicated their DNA but not undergone the 1st meiotic division, accumulate in the mutant testes. These cysts ultimately exit meiosis and then attempt but fail to differentiate into mature sperm [29]. The other allele, orb2ΔQ, is a hypomorph. It has an in frame deletion that removes a short poly-glutamine domain in the N-terminal half of the protein. While the sequence of the Orb2 protein is altered, its level of expression remains the same [32]. Unlike the null, meiosis is unaffected in orb2ΔQ, and orb2ΔQ males produce some functional sperm and are fertile. However, in a subset of the spermatid cysts there are abnormalities in spermatid differentiation, including polarization defects (Fig. 2A). The polarization phenotype in this hypomorphic orb2 allele closely resembles that seen for apkc, and we will discuss it first.
One of the differentiation defects in orb2ΔQ is in properly orienting the direction of spermatid cyst polarization. This phenotype is observed in slightly over 20% of the orb2ΔQ testes (Fig. 2). Fig 5A and B show spermatid cyst nuclei and individualization complexes (IC) in orb2ΔQ testes visualized by Hoechst staining of DNA (blue, arrowheads) and phalloidin staining of actin (green, arrows). The IC consists of 64 aligned actin cones. After elongation is complete, the IC assembles using the condensed spermatid nuclei as a scaffold and then travels along the flagellar axonemes separating syncytial spermatids from each other. The orb2ΔQ testes in Fig. 5A has two assembled ICs (actin in green) that are still associated with the condensed spermatid nuclei (DNA in blue); however, instead of being located towards the basal end of the testis as in wild type, the two spermatid nuclei clusters and associated ICs are in the spermatocyte region of the testis (“*” marks stem cell position). ICs have a distinct morphology; the flat side faces the direction of IC motion, while the thinner, rounded side faces the cluster of spermatid nuclei. In orb2ΔQ testes, ICs moving in opposite directions were sometimes observed (Fig 5B, yellow arrows indicate moving directions).
While orb2ΔQ has defects in choosing the direction of cyst polarization with respect to the apical-basal axis of the testis, the spermatids within the cyst undergo a coordinated (all in the same direction) and otherwise seemingly normal polarization. In contrast, as illustrated by the disorganized distribution of the Boule translation factor and spermatid nuclei, the process of cyst polarization itself seems to be disrupted in orb236. In wild type, Boule accumulates in a comet pattern near the tip of the elongating flagellar tails just like Orb2 (Fig. 5C, arrows) [29], [33]. In the improperly polarized orb236 cyst in Fig. 5D), there are Boule “pseudo-comets” extending from both ends of the cyst towards the spermatid nuclei, which are scattered near the center of the cyst (arrows). In the partially elongated cyst shown in 5E, Boule is distributed along the apical-basal axis with little evidence of polarization, while the spermatid nuclei are scattered throughout much of the cyst. Although these findings indicate that orb2 is essential for cyst polarization, the possibility remains open that the polarization defects arise at least in part from the block in meiosis.
If the failure to properly orient the direction of spermatid polarization in orb2ΔQ arises because apkc requires orb2 activity in this process, we might expect to observe genetic interactions between these two genes. To explore this possibility we examined spermatid cyst polarization in different orb2/apkc trans-heterozygous mutant combinations. As described above, apkc is weakly haploinsufficient in cyst polarization, and in flies heterozygous for the strong loss of function allele apkck06403 about 15% of the testes had improperly oriented spermatid cysts (Fig. 2A2). By contrast, essentially no polarization defects were evident in testes from flies heterozygous for either orb2ΔQ (not shown) or orb236 (Fig. 2A1). On the other hand, about 40% of the testes from orb236/apkck06403 trans-heterozygous males have misoriented spermatid cysts (Fig. 2A2). While orb2ΔQ has at most only a minimal effect on the number of testes that have incorrectly oriented spermatid cysts when combined with apkck06403, this allele significantly increases the frequency of cyst polarization defects when combined with the weak loss of function apkc mutants apkcex48 and apkcex55(Fig. 2A3). In both trans-heterozygous combinations, the frequency of testes with misoriented spermatid cysts increases more than 10-fold. An even more dramatic genetic interaction is observed when the orb2 null allele, orb236, is combined with either apkcex48 or apkcex55 (Fig. 2A3). In fact, the frequency of testes with misoriented cysts in these two trans-heterozygous combinations is equivalent to that observed in the testes of the corresponding homozygous apkc mutant.
These synergistic genetic interactions support the idea that the functioning of apkc in orienting spermatid cyst polarization within the testis depends upon orb2. An expectation of this hypothesis is that there should be defects in the targeting of apkc mRNA and/or in the localized expression of the aPKC protein when orb2 activity is compromised. We first examined the localization of apkc-RA and “bulk” apkc mRNAs in orb2ΔQ. In orb2ΔQ cysts that were polarized in the incorrect direction, the comet pattern was invariably lost and apkc-RA mRNA was instead distributed almost uniformly along the length of the flagellar axonemes (Fig. 6A, B). Similar results were seen when we used the exon probe to visualize the “bulk” apkc mRNA (Fig 6C) indicating that the localization of other apkc mRNA species in the comet pattern is disrupted in incorrectly polarized cysts. On the other hand, in most, but not all of the orb2ΔQ cysts that polarized correctly and have their axonemes elongating towards the apical end of the testis, apkc-RA and the “bulk” apkc mRNAs were localized in the characteristic comet pattern as in wild type (not shown). The pattern of aPKC protein accumulation in orb2ΔQ parallels that of apkc-RA mRNA: aPKC stripes are absent in flagellar axonemes that are polarized in the wrong orientation (Fig.6E). In contrast, in most of the cysts that are polarized in the correct orientation, aPKC protein stripes are localized at the tips of elongating flagellar axonemes just like wild type (Fig.6D). Interestingly, apkc-RA mRNA is readily detected in Orb2 immunoprecipitates of extracts from orb2ΔQ testes (Fig. 3D). Thus, it seems likely that this deletion partially compromises a step subsequent to binding to the apkc-RA mRNA that is important for localizing the apkc-RA;orb2 mRNP complex and/or activating translation.
We also examined apkc mRNA localization and translation in the null allele orb236. In spite of the fact that meiosis never takes place, the apkc-RA transcript is still expressed in orb236 spermatid cysts. This was also true for orb, which, like apkc-RA, is transcribed post-meiotically in the germline of wild type males [29]. Fig. 6F and G show that the apkc-RA and also “bulk” apkc mRNAs are completely unlocalized in differentiating orb236 spermatid cysts. Likewise, the sharp stripes of aPKC protein seen near the ends of the growing sperm tails in wild type cysts are absent in the null mutant cysts (not shown).
Fully extended flagellar axonemes are up to 1.8 mm in length [12], [13]. In order to mediate the localization and translational regulation of apkc or any other mRNA at the tips of the axonemes as the sperm tails elongate, there would have to be a continuous source of Orb2 in the comet head. One way to maintain high levels of Orb2 at the tip of the elongating axonemes would be a positive autoregulatory loop in which Orb2 helps direct the localization and translation of its own mRNA. An orb2 autoregulatory activity could also help promote the initial polarization of the spermatid cyst. Moreover, there is precedence for CPEB proteins having autoregulatory activity. In the female germline, a two step positive autoregulatory mechanism, in which Orb mediates the localization and translation of orb mRNA, is known to be important in targeting Orb to the oocyte when it is first specified and then ensuring that Orb continues to accumulate within oocyte as it develops [34]. Also consistent with this idea, orb2 mRNAs have consensus CPE motifs in their 3′ UTRs (Text S2).
To explore this possibility further we first asked if orb2 mRNA is associated with Orb2 protein in testes extracts. We found that orb2 mRNA can be immunoprecipitated with Orb2, but not control LacZ antibody (Fig. 3D). We next asked whether orb2 mRNA and protein co-localize in elongating wild type and orb2 mutant spermatids. As shown in Fig. 7A, there is generally a close correspondence between orb2 mRNA and protein in elongating wild type spermatids. Importantly, both are concentrated and overlap extensively in the comet head.
Further evidence that orb2 autoregulates the localization and translation of its own mRNA comes from analysis of orb2ΔQ testes. Just as was seen for apkc mRNAs and aPKC protein, the effects of orb2ΔQ depend upon the direction of polarization. In spermatids that are polarized in the incorrect orientation, orb2 mRNA and Orb2 protein do not accumulate to high levels in the comet head at the end of the elongating flagellar axonemes. Instead, the mRNA and protein are distributed uniformly throughout the elongating sperm tails (Fig. 7B). In contrast, in most of the spermatids that are polarized in the correct direction, orb2 mRNA and protein are localized in the characteristic comet pattern seen in wild type testes (not shown). The apparent failure of orb2 autoregulation in orb2ΔQ cysts that are incorrectly polarized could also explain why apkc mRNAs are not localized in the comet pattern and why there are no aPKC protein stripes.
One explanation for the strong orb2-apkc genetic interactions is that the functioning of the orb2 autoregulatory loop at this stage of spermatogenesis could be intimately linked to apkc activity. To test this idea, we examined Orb2 protein expression in elongating spermatid cysts in testes from the apkcex55 allele and from the MTD/UAS-apkc:35140 RNAi knockdown. There were no obvious effects on Orb2 expression in the hypomorphic allele apkcex55 in either incorrectly or correctly oriented elongating spermatid cysts. On the other hand, we found that the accumulation of Orb2 protein in the characteristic comet pattern at the tips of the elongating flagellar axonemes is disrupted when apkc is knockdown using the MTD/UAS-apkc:35140 transgene combination (Fig. 2G). In this case, the Orb2 comet pattern was absent in cysts that were oriented correctly as well as incorrectly. This finding would be consistent with the idea that spermatid cyst polarity is completely randomized in the strong MTD/UAS-apkc:35140 RNAi knockdown.
In order to produce functional sperm, the spermatid cyst must polarize so that all of the nuclei cluster together on one side of the cyst and the basal bodies and nascent flagellar axonemes on the other side. How the cyst orients with respect to the organ (testes) in which it resides is invariant: the nuclei cluster on the basal side of the cyst, while basal bodies localize on the opposite side of the cyst so that the axonemes can elongate apically towards the stem cell region of the testis. The factors involved in polarizing the cyst and in correctly orienting the polarized cyst within the testis are largely unknown. Here we have identified two genes that function to orient cyst polarization with respect to the apical-basal axis of the testis, apkc and orb2. While the involvement of apkc in determining polarity has been extensively documented in many biological contexts, the spatial restriction of apkc activity by a mechanism that involves localized mRNA translational regulation is by contrast rather unusual. We have found that compromising apkc and orb2 activities can give rise to similar polarization phenotypes. In both instances, the system that orients cyst polarization relative to the apical-basal axis of the testes is disrupted and the sperm tails in a subset of the spermatid cysts elongate in the incorrect direction.
During the first steps of polarization all of the spermatid nuclei in the cyst congregate on the basal side, while aPKC protein accumulates in a series of puncta on the other, apical side, near the ends of the nascent flagellar axonemes. As elongation progresses, the aPKC puncta transform into a series of short stripes at the ends of the elongating axonemes. That this spatially restricted pattern of aPKC accumulation is not only mediated by orb2 but is also likely to be relevant to properly orienting polarization with respect to the main axis of the testis is most clearly illustrated by the phenotypes in the hypomorphic allele, orb2ΔQ. All orb2ΔQ cysts that polarize in the incorrect orientation share a common set of defects. First, instead of being localized in the characteristic comet pattern, apkc-RA and “bulk” apkc mRNAs are spread uniformly though the cyst. Second, the aPKC protein stripes at the ends of the growing flagellar axonemes are absent. Third, while Orb2 protein and orb2 mRNA can be detected in these cysts, neither is localized. This latter finding would provide an explanation for the defects in both apkc mRNA localization and the expression of aPKC protein. By contrast, most but not all of the orb2ΔQ cysts that elongate in the correct direction resemble wild type: apkc and orb2 mRNAs and Orb2 protein are localized in the characteristic comet pattern, with aPKC stripes at the ends of the flagellar axonemes.
One likely orb2 regulatory target is the apkc-RA mRNA. Although apkc mRNAs can be detected with a common exon probe in the adult male germline in all but the last stages of spermatogenesis, the apkc-RA mRNA is unusual in that it is expressed post-meiotically, and like other post-meiotic transcripts, it accumulates in a comet pattern in elongating spermatid cysts. This mRNA is transcribed from one of the four apkc promoters, and differs from the 12 other predicted apkc mRNA species in that it has a unique 3′ UTR. The RA 3′ UTR contain 3 CPE motifs and we've found that Orb2 binds to it both in vitro and in vivo. As is true for other mRNA targets of CPEB proteins, it is likely that orb2 promotes the localized production of aPKC by directly activating the translation of the apkc-RA mRNAs in the comet head. Consistent with this idea, the stripes of aPKC at the ends of the elongating flagellar axonemes are localized on the distal (towards apical tip of the testes) side of the Orb2 protein/apkc-RA mRNA comet head. Several other apkc mRNA species share a different UTR sequence that also contains CPE motifs. While we haven't examined the expression of the mRNAs containing this specific UTR sequence, it is possible that one or more is expressed at this stage of spermatogenesis. Since the apkc mRNAs detected with the common exon probe localize in a comet pattern that is similar to apkc-RA and this localization is disrupted in orb2 mutants, it would appear that if other apkc mRNA species are expressed at this stage of spermatogenesis they could also be orb2 regulatory targets.
Significantly, the regulatory relationship between orb2 and apkc seems to be reciprocal. The first suggestion of cross-regulation came from the synergistic genetic interactions seen in trans-heterozygous combinations between orb2 and apkc mutants. Whereas polarization defects are rarely observed in the testes of males heterozygous for the hypomorphic apkcex48 or apkcex55 alleles, when they are combined in trans with the hypomorphic orb2ΔQ allele the frequency of testes with misoriented spermatid cysts increases by more than 10-fold. Trans-heterozygous genetic interactions between weak hypomorphic alleles are somewhat unusual, and raise the possibility that the two interacting genes are functionally interdependent. Consistent with a cross-regulatory connection, Orb2 is not localized in its characteristic comet pattern in elongating flagellar axonemes when apkc activity is knocked down in the MTD/UAS-apkc:35140 combination. Instead it is uniformly distributed in the elongating axonomes. This finding indicates that the localized accumulation of Orb2 protein in differentiating spermatid cysts depends upon apkc activity. Since Orb2 binds to its own mRNA and appears to be required for both the localization and translation of this mRNA during spermatid cyst differentiation, a plausible hypothesis is that apkc is a component of the orb2 positive autoregulatory loop. While apkc need not function directly, it is interesting to note that Orb2 has three (high stringency) predicted aPKC phosphorylation sites (Ser146, Ser273, and Ser446) [35], and all three are phosphorylated in Orb2 protein isolated from testes (unpublished data). Thus it is possible aPKC phosphorylation facilitates the localized translation of orb2 mRNA, and in turn the localization and translation of apkc mRNAs, by phosphorylating Orb2 protein. In principle, this reciprocal regulatory relationship could help trigger the choice of orientation at the start of polarization and then serve to reinforce this choice.
While our finding would be consistent with a model in which aPKC activity is spatially restricted by a mechanism that depends upon orb2 localizing and regulating the translation of apkc mRNAs, many important questions remain unanswered. One is whether apkc is required only for choosing the direction of cyst polarization within the testis, or if it also has a role in the process of polarization itself. Neither the apkc mutants nor the knockdowns are useful in answering this question as they certainly retain apkc activity. On the other hand, it seems likely that apkc is needed during the formation of the sperm tails since the elongating tails in the strongest RNAi knockdowns are considerably shorter than in wild type. As Orb2 protein accumulation is perturbed in this RNAi knockdown, one of the apkc functions during this phase of spermatogenesis is to maintain high levels of localized Orb2 protein.
A similar question can be asked about orb2. With the caveat that the orb236 cysts might have structural abnormalities arising from the failure to undergo meiosis, the phenotypes of this mutant argue that orb2 is required for polarization per se. If the only apkc function in this process is orienting the direction of polarization, then orb2 would have to have regulatory targets that control the actual process of polarization. In fact, there are several plausible candidates. Studies by Fabian et. al. [19] have shown that phosphatidylinositol 4,5-bisphosphate (PIP2) and components of the exocyst complex are required for cyst polarization. The exocyst complex also mediates plasma membrane addition during spermatid elongation and localizes around the tip of the growing sperm tails. mRNAs encoding 4 of the 8 exocyst complex subunits (sec3, sec8, sec10, and sec15) have CPE sequences in their 3′ UTRs, and their localization and/or translation could be regulated by orb2. Another possible orb2 regulatory target in cyst polarization is cdc42. This small membrane anchored GTPase directs exocyst complex localization during ciliogenesis in kidney epithelial cells, interacting directly with Sec10 [36], [37]. Cdc42 also functions in regulating apical-basal cell polarity by interacting with Par6, and this interaction in turn recruits and localizes aPKC to the membrane [38]. Since mRNAs encoding both Cdc42 and Par6 have CPEs in their 3′ UTRs, their localization and translation could be controlled by orb2. While we have not tested the effects of reducing cdc42 activity, we found that a subset of the spermatid cysts are polarized in the incorrect direction in par6 heterozygous males (data not shown).
The idea that orb2 has other mRNA targets in cyst polarization would also be consistent with the effects of orb2 mutations on asymmetric cell division in the embryo. We previously found that the accumulation of aPKC along the apical cortex of dividing neuroblasts depends upon orb2 [30]. However, one of the other orb2 asymmetric cell division phenotypes is a failure to properly orient the mitotic spindle. Since spindle orientation in dividing neuroblasts is thought to be independent of apkc [17], it would appear that the localized expression of other polarity proteins must also depend upon orb2. In addition to cdc42, another potential orb2 target that has functions in spindle orientation would be the mRNA encoding Inscuteable (insc). In fact, Hughes et. al. [39] have shown that apical localization of the insc mRNA is important for Insc function in neuroblast cell division, and like apkc-RA mRNA it has CPE sequences in its 3′ UTR. Thus, a plausible inference from their studies is that Orb2 may help localize insc mRNA to the apical side of the neuroblast, and then activate its on site translation. (It is interesting to note that there are hints that mRNA localization may also be important when polarization is being initiated in epithelial cells [40]).
While orb2 might regulate many of the same target mRNAs during both spermatid cyst polarization and asymmetric cell division, one apparent difference is in the penetrance of the phenotypes. Cyst polarization seems to be completely disrupted in the absence of orb2 function. In contrast, only a subset of the cells in the embryonic neuronal and muscle cell lineages show obvious defects in asymmetric cell division. One likely reason for this difference is that polarization during cell division can be mediated independently by cross-regulatory interactions between factors specifying the apical and basal domains. Since several of the polarity proteins do not seem to be expressed in the spermatid cysts, it is possible that these cross-regulatory interactions either do not exist, or are not sufficient for polarization. Conversely, the fact that there are even modest cell division phenotypes in neuronal and muscle precursor cell lineages in orb2 mutants also implies that these cross-regulatory interactions are not in themselves sufficiently robust to ensure that these cells always polarize correctly in the absence of orb2 function.
In this context, it is worth noting that the role of orb2 in both asymmetric cell division and spermatid cyst polarization seems to differ from many of the previously documented functions of mRNA localization in processes that depend upon polarization such as cell fate or polarity axis determination. In most of the instances in which mRNA localization is known to have an important role in cell fate or polarity axis determination (e.g., prospero mRNA localization to the basal daughter cell during neuroblast cell division or oskar mRNA localization in the specification of the posterior pole of the oocyte/embryo), the underlying polarity of the cell, cyst, egg, embryo or organ is pre-defined by the activity of an upstream and seemingly distinct polarizing pathway, most typically involving the aPKC/PAR machinery. The mRNAs encoding the cell fate or axis determinants function primarily in the elaboration or execution of this polarity decision, and not in the initial definition of polarity [41]–[43]. In contrast, orb2 seems to be intimately involved in the upstream polarity pathway, helping the aPKC/PAR machinery define and then maintain the underlying polarity. In fact, in the process of correctly orienting spermatid cyst polarization the functional interdependence of orb2 and apkc would seem to be mechanistically equivalent to the cross-regulatory interactions that underpin “classical” polarization by the aPKC/PAR machinery.
Other unresolved questions include the identity of the signals that initiate and orient cyst polarization relative to the testis itself. The latter, cyst orientation, most probably depends upon an external signal and the most likely source would be the two somatic support cells, the head and tail cyst cells, which encase the spermatid cyst (Fig. 1A). These two cells arise from a population of somatic stem cells at the apical tip of the testis and associate with the newly formed germline daughter cell that ultimately gives rise to the 64-cell spermatid cyst. The two somatic cells grow without division, surrounding the germ cells as they undergo mitosis and meiosis. Interestingly, one of the cells ends up on the apical (relative to the testes) side of the newly formed cyst and it expresses the Dlg guanylate kinase. The other is on the basal side and doesn't express Dlg [44]. We confirmed this observation and found that another polarity regulator, Bazooka, seems to exhibit the same expression pattern. Either one of these cells could potentially produced a signal(s) that orients cyst polarization. It is also possible that polarization depends upon an autonomous signal generated in the germline when the cysts commence differentiation. Further studies will be required to answer these and other questions.
apkc mutant alleles apkck06403, apkcex55, apkcex48 and par1-GFP are kind gifts from Dr. Elizabeth Gavis, Dr. Yu-Chiun Wang and Dr. Eric Wieschaus at Princeton University. apkc RNAi lines are obtained from Bloomington Drosophila Stock Center (stock # 34332 and 35140), targeting two different exons common to all apkc mRNAs (sequence: CTGGAGAAGACGATTCGTATA and CAAGCTGTTGGTGCACAAGAA) [45]. MTD multiple driver line is obtained from Dr. Andrew Hudson and Dr. Lynn Cooley from Yale University. The orb2 mutant allele orb236 was generated in the lab [29]. orb2ΔQ is a gift from Dr. Barry Dickson from IMBA, Austria.
Testes were dissected and fixed/stained following standard whole mount staining procedures [29]. DNA was dyed with Hoechst. Testes were further examined under epi-fluorescence microscopes. Percentage of testes with spermatid nuclei near the stem cell/spermatogonial region was recorded. Those were considered defective in orienting the polarization of the spermatid cysts.
Whole mount staining was performed as in [29]. Antibodies used were as follows: mouse anti-Orb2 2D11 and 4G8 IgG (undiluted, developed in the lab) [29], rabbit anti-Bol (1∶1000, a gift from Steven Wasserman) [46], monoclonal anti-β-Tubulin E7 1∶50 (Developmental Studies Hybridoma Bank), rabbit anti-aPKCζ (1∶1000, clone c-20, sc-216, Santa Cruz Biotechnology), rabbit anti-GFP (1∶1000, Cristea Lab, Princeton University). Rabbit polyclonal anti-Dlg-PDZ1 (1∶1000) and guinea pig anti-Bazooka (1∶500) were provided by Yu-Chiun Wang [47]. Actin was stained with Alexa488-phalloidin (Invitrogen, Carlsbad, CA). DNA was stained with Hoechst (1∶1000). Secondary antibodies used were goat anti-mouse IgG Alexa 488, 546 or 647, goat anti-rabbit Alexa 488, 546 or 647 (Molecular Probes, Inc.). Samples were mounted in Aqua-polymount on slides for an inverted Zeiss LSM510 or Leica SP5 confocal microscope.
Fluorescence in situ hybridization was performed as described in [29]. Fluorescent antisense probes for apkc and orb2 were synthesized by Biosearch Technologies (www.biosearchtech.com) or synthesized in Tyagi lab. Forty non-overlapping 17 bp probes targeted at orb2 mRNA sequence from cctggacgatcagatgt to atatgttatttaatctcac were synthesized and labeled with Quasar 670 and used at 1∶100 dilution. For combined in situ hybridization-antibody staining experiment, the in situ hybridization was performed first, followed by a sample re-fix and then standard whole mount antibody staining.
|
10.1371/journal.pntd.0000887 | Complex Interactions between Soil-Transmitted Helminths and Malaria in Pregnant Women on the Thai-Burmese Border | Deworming is recommended by the WHO in girls and pregnant and lactating women to reduce anaemia in areas where hookworm and anaemia are common. There is conflicting evidence on the harm and the benefits of intestinal geohelminth infections on the incidence and severity of malaria, and consequently on the risks and benefits of deworming in malaria affected populations. We examined the association between geohelminths and malaria in pregnancy on the Thai-Burmese border.
Routine antenatal care (ANC) included active detection of malaria (weekly blood smear) and anaemia (second weekly haematocrit) and systematic reporting of birth outcomes. In 1996 stool samples were collected in cross sectional surveys from women attending the ANCs. This was repeated in 2007 when malaria incidence had reduced considerably. The relationship between geohelminth infection and the progress and outcome of pregnancy was assessed.
Stool sample examination (339 in 1996, 490 in 2007) detected a high prevalence of geohelminths 70% (578/829), including hookworm (42.8% (355)), A. lumbricoides (34.4% (285)) and T.trichuria (31.4% (250)) alone or in combination. A lower proportion of women (829) had mild (21.8% (181)) or severe (0.2% (2)) anaemia, or malaria 22.4% (186) (P.vivax monoinfection 53.3% (101/186)). A. lumbricoides infection was associated with a significantly decreased risk of malaria (any species) (AOR: 0.43, 95% CI: 0.23–0.84) and P.vivax malaria (AOR: 0.29, 95% CI: 0.11–0.79) whereas hookworm infection was associated with an increased risk of malaria (any species) (AOR: 1.66, 95% CI: 1.06–2.60) and anaemia (AOR: 2.41, 95% CI: 1.18–4.93). Hookworm was also associated with low birth weight (AOR: 1.81, 95% CI: 1.02–3.23).
A. lumbricoides and hookworm appear to have contrary associations with malaria in pregnancy.
| Intestinal worms, particularly hookworm and whipworm, can cause anaemia, which is harmful for pregnant women. The WHO recommends deworming in pregnancy in areas where hookworm infections are frequent. Some studies indicate that coinfection with worms and malaria adversely affects pregnancy whereas other studies have shown that coinfection with worms might reduce the severity of malaria. On the Thai-Burmese border malaria in pregnancy has been an important cause of maternal death. We examined the relationship between intestinal helminth infections in pregnant women and their malaria risk in our antenatal care units. In total 70% of pregnant women had worm infections, mostly hookworm, but also roundworm and whipworm; hookworm was associated with mild anaemia although ova counts were not high. Women infected with hookworm had more malaria and their babies had a lower birth weight than women without hookworm. In contrast women with roundworm infections had the lowest rates of malaria in pregnancy. Deworming eliminates all worms. In this area it is unclear whether mass deworming would be beneficial.
| In 1994 and 2002 the World Health Organization (WHO) recommended anthelminthics be given to girls, pregnant and lactating women to reduce the burden of anaemia in areas where hookworm and anaemia are common [1]–[3]. Published evidence suggests that mebendazole [4]–[8] or albendazole [9]–[14] administered after the first trimester of pregnancy are safe. However the advantage of routine deworming of pregnant women is debatable, with different studies presenting different results. Several studies reported that systematic anthelminthic administration was associated with less anaemia [2], [4], [6], [10], [11], [13] and with a beneficial effect on birth outcomes, reducing the rates of low birth weight [5], [7]–[9], [12], very low birth weight [15], stillbirth and perinatal death [7]. However a Cochrane review, including three prospective randomised controlled trials studying the effect of deworming in pregnancy, concluded that the evidence to date is insufficient to recommend use of antihelminthics for pregnant women after the first trimester of pregnancy [16]. A recent randomised controlled trial in Uganda showed no benefit of anthelminthic treatment on maternal anaemia, low birthweight and perinatal mortality [17].
There are also conflicting and often confusing results regarding the impact of geohelminth infections on other infectious diseases, and in particular malaria [18]–[24].While some studies have failed to find any relationship between geohelminth infection and malaria [25], others have shown an increased incidence of P. falciparum malaria in presence of geohelminths [26]–[28]. Ascaris (A.) lumbricoides infections were linked to severe P. falciparum malaria in Senegal [29] but they have more often been associated with a beneficial effect on malaria [20], [30]–[34]. Several immunological hypotheses, including modulation of T-helper or dendritic cell responses and cytokine induction, have been proposed to explain these interactions [35]–[40]. There have also been haematological and entomological hypotheses to explain increased incidence [19].
Data from studies specific to pregnancy and helminths are also conflicting. Hookworm, not P. falciparum malaria, was considered the main cause of anaemia in some [41], [42], while others reported an opposite result [43], [44] or did not find any association [15]. Maternal co-infection with P. falciparum and helminths resulted in a significantly lower mean birth weight than with P. falciparum infection alone in Nigeria and Ghana [45], [46]. Two recent studies report an association with lower rates of P.falciparum infection in women co-infected with A. lumbricoides [47], [48]. Yatich and colleagues report a 4.8 (95% 3.4–40) fold increased risk of P.falciparum with any geohelminth and the risk remained significant for hookworm and A. lumbricoides alone [49].
Two cross-sectional helminth surveys (1996 and 2007) conducted among women attending antenatal clinics on the Thai-Burmese border were reviewed to determine whether there was any association between geohelminth infection and malaria in this area, endemic for both P. falciparum and P.vivax malaria species and where there has been no systematic deworming during pregnancy.
The Shoklo Malaria Research Unit (SMRU) has five established clinics on the Thai-Burmese border. One is based in the largest of the refugee camps, Maela (circa 45,000 people); the others are stretched along 100 km of the border and serve a migrant population of circa 50,000 people. Antenatal clinics (ANC) have been operational since 1986 in the camp and 1998 in the migrant population. Malaria transmission is low and seasonal [50]. Treatment is complicated by a high level of multi-drug resistant strains of P. falciparum [51]. There is currently no safe and effective P. falciparum antimalarial drug that can be offered as intermittent presumptive therapy (IPT) or prophylaxis to pregnant women. Active weekly detection and early treatment of malaria has so far been the best method to prevent maternal death from malaria in this area [52].
The ANC performs a weekly malaria smear for all women, 2nd weekly haematocrit, provides routine iron and folate supplementation, and all necessary medical and obstetric care. A mother-to-child HIV transmission prevention program started in 2001 in the refugee camp and was introduced to the migrant population in 2008. HIV prevalence is low (<1.0%) and test uptake high (>90%) [53].
The antimalarial drug regimen for the treatment of P. falciparum in pregnant women was quinine or mefloquine mono-therapy in 1996, and quinine or artesunate with or without clindamycin in 2007. P.vivax episodes are treated with chloroquine alone. In 1996 women with non-severe (mild) anaemia (haematocrit between 20% and 29.9%, HB between 6.7 g/L and 10 g/L) were treated with ferrous sulphate (200 mg three times daily) and folic acid (5 mg daily) until delivery. Women with severe symptomatic anaemia (haematocrit <20%, HB <6.7 g/L) were transfused. In 2007 all women received ferrous sulphate, 200 mg daily and folic acid, 5 mg weekly, from first consultation to delivery and treatment doses as stated previously if they became anaemic. Thailand has no deworming policy for pregnant women nor do the agencies working in the refugee camps.
Women were encouraged to deliver with trained midwives in the SMRU delivery rooms, those requiring Caesarean section were transferred to the nearest Thai Hospital. Gestational age was estimated by Dubowitz score [54] in 1996 and by ultrasound (or Dubowitz for late scans) in 2007 [55]. Birth weight was measured on electronic Seca scales (accuracy 10 grams) or Salter hanging scales (accuracy 50 grams).
Women participating in the surveys were of similarly deprived socioeconomic groups. Refugees in the camps receive food assistance and have access to medical care, but cannot work. Migrants work hard for low wages and lack access to medical care. Both groups are poor and economically weak [56]. Housing of refugee and migrant women are the same. Houses are elevated on poles of wood and walls and floors are made of bamboo with leaf roofing. Most families have their own toilets. Flip flops (flat sandal) are normally worn on the feet in all age groups. Contact with soil that is reportedly highly contaminated with helminth is inevitable [57], more common in the rainy season and in those involved in agricultural work.
Miscarriage was loss of products of conception or foetus before 28 completed weeks of gestation; stillbirth was delivery of a dead foetus aged 28 weeks or more; low birth weight (LBW) was a birth weight of <2500 grams measured in the first 5 days of life, and prematurity a delivery before 37.0 weeks of gestation; congenital abnormality was considered if a major defect was present at birth.
The surveys in 1996 and 2007 were both conducted during the rainy season period (May–Oct) in order to allow comparison without having to take into account seasonality and because malaria peaks at that time of the year. Both surveys were exhaustive. The first survey, when SMRU only worked in the refugee camp, was to determine if worm infection was associated with anaemia. Every woman was asked to participate.
The 2nd survey was done as a response to the preparation of a border wide medical guideline. The refugee camps on the border fall under the care of different NGOs (Non-Government Organisations) and there was debate on deworming in pregnancy. Since the last survey was old it was decided to resurvey pregnant women to determine if there was a need for deworming in pregnancy. At the time of this survey SMRU also provided antenatal care for migrants who have less access to health care than refugees. Hence refugee and migrant women were surveyed if they voluntarily gave a stool sample. Every 5th woman was asked to participate.
Before each survey a general announcement was made to all pregnant women attending the ANC. Participation was voluntary. It was explained that if their stools were found positive for worms they would receive anthelminthic treatment. The importance of providing a fresh stool sample was explained.
Stool samples were examined on site. As Necator americanus and Ankylostoma duodenale ova cannot be differentiated by microscopy, the term hookworm was used. Women with a positive stool test result for hookworm, A. lumbricoides (roundworm) or Trichuris (T.) trichuria (whipworm) were treated with mebendazole 200 mg once daily for 3 days. In 1996 this treatment was given after delivery, in 2007 at the time of diagnosis or after the first trimester.
In this area the natural immunity to malaria is weak because the transmission is very low, so that most patients with malaria parasites become symptomatic. However because of the systematic weekly screening regardless of symptoms, many episodes are detected before symptoms arise. For this reason we cannot strictly speak of incidence or prevalence during the entire follow up period. Furthermore the number of infections of malaria relates not just to transmission but also to the poor response to antimalarial treatment. At the time of the 1996 survey quinine had an estimated failure rate of 23% and mefloquine 28% [58]. This makes it difficult to assign each episode as a new case, as it might be a treatment failure.
In the analysis, women were categorized as “free of malaria” if all the malaria smears done at antenatal visits up to the day of the stool test were negative; women with a positive malaria smear up to the day of the stool test were categorized into one of 3 groups: “P. vivax group” or “ P. falciparum group” or “ mixed infection group”. Women in the P.vivax group only had one or more episodes of P.vivax, women in the P. falciparum group only had one or more episodes of P. falciparum and women in the mixed group may have had a single mixed infection of P.falciparum and P.vivax or on separate occasions a P.falciparum and a P.vivax. Results are given as proportions.
Following the ANC system as well as participation in the stool survey was voluntary. Providing a stool sample did not involve any risk for the pregnant woman. For these reasons no informed consent was obtained. Pregnancy records have been routinely entered to a data recording system since 1987. Ethical approval for analyzing these patient records was given by the Oxford Tropical Research Ethics Committee (reference: OXTREC 28–09).
Stool samples were prepared using the formalin–ethyl acetate sedimentation technique [59] and hookworm ova counts performed. Two wet preparations were done for each sample to increase the sensitivity of detection and verify negative slides. The stool assessment was quantitative: a standard dilution of the stool sample was made and 100 µl (taken with a Gilson pipette) was put on a slide. The entire area under the cover slip (22×22 mm) was examined with the X10 objective. Hookworm ova were counted and the number multiplied by 10 was the estimated number of ova per ml of faeces. Other geohelminths were reported as: 1 ova per slide rare, 2–3 per slide 1+, 4–10 per slide 2+ and >10 per slide 3+. In 1996, all stool samples were quality controlled by laboratory staff (WB) from the Liverpool School of Tropical Medicine and Hygiene with good agreement.
Thick and thin malaria smears were stained with Giemsa and examined under oil immersion; the presence of any asexual blood stage parasite was declared as malaria positive. Smears were declared negative after reading 200 fields.
Blood samples (finger prick) were centrifuged at 12000 rpm for 3 minutes and read using a standard haematocrit reader. The haematocrit value measured the day of the stool test was used to describe anaemia. If not available, the result the closest to the stool test day (but within 8 weeks prior to) was chosen.
Data were entered using Microsoft Access, and analyzed using SPSS version 14 for Windows (SPSS, Benelux inc., Gorinchem, Netherlands) and Epi Info (Centre for Disease Control and Prevention). Student's t-test and Mann-Whitney test were used for comparison of means and ranks respectively. Categorical data were compared using the chi-squared test or the Fisher's exact test, as appropriate. To assess independent predictors of malaria, anaemia and LBW, a multivariate unconditional logistic regression model was fitted using the variables that were significantly associated in univariate analysis.
A total of 829 pregnant women provided a stool sample for examination; this represented 85% (339/401) of the ANC attendees in 1996 and 33% (490/1,485) in 2007. There was no significant difference in the baseline demographics between women who provided a stool sample and those who did not (data not shown).
In 1996 all women were from Maela refugee camp; in 2007, 42% (244/490) were from the camp, while the others attended the migrant antenatal clinics. Participants in the 2007 survey were older (+1.6 years), had their stool test at a later gestational age (+4 wks) and were less anaemic than those in the 1996 survey (Table 1).
Between 1996 and 2007 there was a significant decline in the proportion of women who had: any malaria, 27% (90/339) vs. 20% (96/490), (P = 0.02), P. falciparum malaria, 14% (48/339) vs. 3% (13/490), (P<0.001) and anaemia, 27% (91/339) vs. 18% (90/490), (P = 0.005), but an increase in the proportion of women infected with P.vivax (Table 1).
Overall 70% (578) of the 829 women were infected with at least one geohelminth, including hookworm (43% (355)), A. lumbricoides (34% (285)) or T. trichuria (31% (250)) alone or in combination (Table 2). Prevalence was significantly higher in 1996 than 2007: 81% (95% CI: 76–84) (273/339) vs. 62% (95% CI: 58–66) (305/490), P<0.001. The intensity of worm infections was low, with high hookworm ova counts (≥1000 ova/mL) found in <10% of the positive results, and a maximum count of 2900 ova/mL. Hookworm (Table 2) and T.trichuria intensities of infection decreased between the 2 surveys, while A. lumbricoides decreased in all intensities except the highest group.
Thirty three pregnant women had their first malaria infection after stool testing and were excluded from further analysis related to geohelminths and malaria and anaemia.
For the purpose of this analysis geohelminth infections were assumed to be present until mebendazole treatment was administered, as there was no routine deworming policy.
Overall 153/796 women (19%) had malaria detected at least once prior to, or at the day of the stool test. Most of the women presented with single species infections; 35% (53) had P. falciparum infections only and 54% (83) P. vivax only. P. falciparum and P. vivax simultaneously or on separate occasions occurred in the remaining 11% (17).
The proportion of women with malaria in pregnancy was similar in those with geohelminth co-infection or without: 18% (44/242) vs. 20% (109/554), P = 0.77 (Table 3). There were important differences in the proportions of women with malaria depending on the type of geohelminth co-infection (Figure 1). The highest proportions of both P. falciparum and P.vivax malaria were seen with hookworm (±T. trichuria) co-infections and the lowest with A. lumbricoides (±T. trichuria) co-infections. The protective effect of A. lumbricoides (±T. trichuria) remained significant for P.vivax malaria when stratifying by malaria species (Table 3). The overall proportion of women with malaria in women with A. lumbricoides infections was approximately half that in hookworm infections. Temperature, days of fever, number of episodes of malaria and parasitaemia were not significantly different between the worm groups (data not shown). There were only 3 women with hyper-parasitaemic malaria (more than 4% of the red blood cells infected with P. falciparum), 2 of them without worm infection, and 1 woman with hookworm infection. The relationship between malaria and stool ova counts for the hookworm (±T. trichuria) group and for ova count in the A. lumbricoides (±T. richuria) group was explored (Figure 2). There was no relationship between malaria and the hookworm (P = 0.76) or A. lumbricoides (P = 0.92) stool ova counts. There was not sufficient data to study interaction between geohelminth single infections and their association with malaria, as nearly half of all infections were combinations of worms (Table 2).
The proportions, by age and gravid groups, of women with malaria and those with geohelminths are presented (Figure 3). Age was not significantly associated with malaria, but gravidity was: 25% (49/198) in primigravida vs. 17% (104/595) in multigravida, P = 0.029. The proportion of hookworm infection was higher in teenage women, although this was not significant.
In a multiple regression model, primigravida (AOR: 1.53, 95% CI: 1.01–2.32, P = 0.043) and hookworm co-infection (AOR: 1.66, 95% CI: 1.06–2.60, P = 0.027) remained the two independent factors associated with an increased risk of malaria while the protective effect of A. lumbricoides co-infection remained significant (AOR: 0.44, 95% CI: 0.23–0.86, P = 0.015). Year of survey and ova counts for hookworm and A. lumbricoides were non-significant.
Sixty five women (8%) did not have a haematocrit measurement at the time of or before the stool test and were not included in this part of the analysis. Mean haematocrit was similar whether geohelminth infection was present or not in the remaining 733 pregnant women. The proportion of women with anaemia was higher in women with high intensity hookworm infection compared to those with lower counts, 41% (14/34) vs. 21% (150/699), (P = 0.011); in multigravida compared with primigravida, 24% (134/547) vs. 16% (30/183), (P = 0.024); those who were older than 25 years, 25% (100/394) vs. 19% (64/337), (P = 0.041); and those who had malaria, 30% (45/152) vs. 20% (119/581), (P = 0.021). In a logistic regression model high hookworm load (AOR: 2.05, 95% CI: 1.01–4.20), P = 0.049), malaria (AOR: 1.83, 95% CI: 1.12–2.74, P = 0.004), being multigravid (AOR: 1.79, 95% CI: 1.15–2.78, P = 0.009) and participating in the 1996 survey (AOR: 1.57, 95% CI: 1.03–2.08, P = 0.032) remained independently associated with anaemia.
Pregnancy outcome data were available for 94% (783/829) of women. There were 14 abortions (2%) and 8 stillbirths (1%). Eleven infants were born with congenital abnormalities (8 live-births and 3 stillbirths). Neither stillbirth nor congenital abnormalities were significantly associated with geohelminth infection. The mean gestational age at delivery was 38.9±1.7 [28.4–42.5] weeks. Mean gestational age and the proportion of premature infants were not significantly different in the presence or absence of geohelminth infection.
Birth weight data were available for 87% (648/748) of live-born, normal, singletons. Mean birth weight was 2900±447 [1100–4400] g. The proportion of LBW newborns was significantly higher among primigravida compared to multigravida (25% (41/162) vs. 10% (46/483), P<0.001), in women aged <25 years vs. older (19% (57/301) vs. 9% (30/345), P<0.001), with hookworm infection vs. none (17% (46/278) vs. 11% (41/370), P = 0.048), and in premature vs. term infants (58% (34/59) vs. 9% (53/589), P<0.001). Anaemia and malaria were not significant risk factors for LBW. In a logistic regression model excluding prematurity (the strongest risk factor for LBW (n = 59)), the presence of hookworm infection was independently but weakly associated with LBW (AOR: 1.81, 95% CI: 1.02–3.23, P = 0.041) as was (more significantly so) being primigravid (AOR: 3.27, 95% CI: 1.83–5.84, P<0.001).
The proportion of intestinal geohelminth infections in pregnant women on the Thai-Burmese border is high and comparable to reports from other parts of South East Asia, including Thailand, Burma and Vietnam [60]–[62], while malaria transmission is low [50]. The cross-sectional surveys conducted 11 years apart confirmed a declining prevalence of intestinal parasites but prevalence remains higher than the 20–30% WHO criteria for mass deworming [3], [63]. Anaemia in this population is common but predominantly mild [64] and the association of hookworm infection and anaemia was only significant for the highest intensity of hookworm infection, a finding already reported half a century ago [65].
This site has a unique system of antenatal care established in 1986 in response to a high malaria related-maternal mortality rate(1000/100,000 live births)[52] and lack of any drug to offer as chemoprophylaxis due to multidrug resistant strains of P.falciparum. Women are encouraged to attend ANC on a weekly basis. Attendance is high and most women average more than 10 consultations per pregnancy in both the refugee and migrant settings. Women with detectable parasitaemia on screening are treated regardless of symptoms as they are unlikely to clear parasitaemia without becoming symptomatic [66]. The reduction in malaria (and anaemia) incidence in pregnancy in this population has been described in detail elsewhere [67].
This is the first time, to our knowledge, that A. lumbricoides infection has been associated with a reduced risk of P. vivax. The novelty of this finding might be influenced by the fact that many studies on interactions between worms and malaria were done in Africa, where P. falciparum is the predominant species. Other investigators in African settings have described higher prevalence of falciparum malaria in presence of A. lumbricoides in pregnant women [49] and in children [29]. These differences could be related to acquired immunity, methodological issues or to interactions (other helminths could alter the immune response e.g. schistomiasis or Strongyloides stercoralis). The density of infection of hookworm and A. lumbricoides in these surveys was low, as is usually reported from Asian settings, and no relationship was found between malaria and geohelminth density.
Similar to a study in Kenyan pregnant women A. lumbricoides prevalence increased with gravidity, but whereas they observed the same trend with maternal age this was not observed on the Thai-Burmese border[48]. Hookworm prevalence peaked amongst the lowest age group, and reached a plateau after 25 years of age, which is similar to the pattern reported from Kenyan pregnant women[48].
Co-infection with hookworm was associated in our setting with a significantly higher risk of malaria, but this did not reach statistical significance for the individual plasmodial species. Similar associations have been previously reported for P. falciparum in children [68], [69], adults [19] and pregnant women [48]. In regions of high prevalence it is plausible that helminthes might suppress the ability to clear infections (resulting in a positive association between helminth infection and asymptomatic malaria parasitaemia), or suppress the inflammatory responses that result in clinical disease (resulting in a negative association between helminth infection and clinical malaria disease) [21] but this seems unlikely in this setting where malaria prevalence is low. Subclinical haematological cues may influence the host attractiveness for the vector [19]. In Thailand hookworm was shown to be associated with increased incidence of P. falciparum but not P.vivax malaria [19]. In our setting, the replenishment of pregnant women's iron and folate reserves may have resulted in reticulocytosis and may have increased P.vivax densities [70]. This effect would be expected to be greatest in hookworm infections, and may well explain why the interaction could be observed in our study whereas it could not be in the study of Nacher [19] where patients did not receive haematinics. P. falciparum and to a much greater extent P. vivax prefer to invade young red cells. This would not account for the negative association seen between malaria and A. lumbricoides. However, an alternative hypothesis proposed by some authors is that residential location and spatial aspects of exposure may explain some of the associations between worms and malaria [71]. NGOs in the refugee camps have provided intermittent deworming (6–12 monthly single dose mebendazole) to school children since late 2001. There has been no deworming program for pregnant women or adults. It is likely that sporadic deworming of children and improved footwear or sanitation, has led to the decreased proportion of geohelminths in pregnant women observed between the two survey periods.
Active weekly screening as part of routine antenatal care has made severe malaria a rare event. No association between disease severity and the prevalence of geohelminths could be demonstrated. The decreased proportion of women with mild anaemia between the two surveys could be related to the decrease of geohelminth infection; it also could be due to the reduction in P. falciparum observed on the Thai-Burmese border [72], [73] or to the implementation of anaemia prophylaxis for all pregnant women. In Sierra Leone, the administration of iron and folate supplements had a greater effect on haematocrit than the administration of albendazole [11]. This suggests that deworming to prevent anaemia should not be used as sole strategy against anaemia [74]. If A. lumbricoides coinfection does indeed attenuate malaria, then mass deworming may reduce a potential protective benefit. On the other hand hookworm was associated with a higher proportion of malaria, low birth weight and anaemia suggesting that hookworm should be treated in pregnancy. As there is no specie selective antihelminth at hand, deworming policies should be based on local prevalence and intensity of geohelminths, malaria, and anaemia severity.
The present paper has limitations: the surveys were cross sectional, pooling data from different periods and the sample size may not have been sufficient to detect quantitative effects of the different worm species on different plasmodial species. No plausible explanation has been provided for the observed associations. No socioeconomic, behavioral or environmental factors were available for analysis, however these tend to be uniformly similar across the population of refugees and migrant workers: These people live in poor conditions and all are economically deprived so that it is unlikely to be a confounder in the analysis. The assumption that worms observed in the stool sample were present at the time of malaria is plausible at the population level given the lifespan of worms but it is not possible to ascertain that was always the case in each individual. This may have reduced the precision. Nevertheless, the present paper presents for the first time in the same data set a range of complex interactions between hookworm, A. lumbricoides and both P. falciparum and P. vivax malaria, during pregnancy. Our findings potentially have considerable practical and evolutionary implications. Future trials to confirm or deny the associations observed here require well designed longitudinal studies to account for the observed complex and conflicting interactions.
|
10.1371/journal.pntd.0002373 | Serum Metabolome and Lipidome Changes in Adult Patients with Primary Dengue Infection | Dengue virus (DENV) is the most widespread arbovirus with an estimated 100 million infections occurring every year. Endemic in the tropical and subtropical areas of the world, dengue fever/dengue hemorrhagic fever (DF/DHF) is emerging as a major public health concern. The complex array of concurrent host physiologic changes has hampered a complete understanding of underlying molecular mechanisms of dengue pathogenesis.
Systems level characterization of serum metabolome and lipidome of adult DF patients at early febrile, defervescence, and convalescent stages of DENV infection was performed using liquid chromatography- and gas chromatography-mass spectrometry. The tractability of following metabolite and lipid changes in a relatively large sample size (n = 44) across three prominent infection stages allowed the identification of critical physiologic changes that coincided with the different stages. Sixty differential metabolites were identified in our metabolomics analysis and the main metabolite classes were free fatty acids, acylcarnitines, phospholipids, and amino acids. Major perturbed metabolic pathways included fatty acid biosynthesis and β-oxidation, phospholipid catabolism, steroid hormone pathway, etc., suggesting the multifactorial nature of human host responses. Analysis of phospholipids and sphingolipids verified the temporal trends and revealed association with lymphocytes and platelets numbers. These metabolites were significantly perturbed during the early stages, and normalized to control levels at convalescent stage, suggesting their potential utility as prognostic markers.
DENV infection causes temporally distinct serum metabolome and lipidome changes, and many of the differential metabolites are involved in acute inflammatory responses. Our global analyses revealed early anti-inflammatory responses working in concert to modulate early pro-inflammatory processes, thus preventing the host from development of pathologies by excessive or prolonged inflammation. This study is the first example of how an omic- approach can divulge the extensive, concurrent, and dynamic host responses elicited by DENV and offers plausible physiological insights to why DF is self limiting.
| Dengue virus is the most widespread arbovirus and a major public health threat in the tropical and subtropical areas of the world. As yet, little is known about the molecular mechanisms underlying infection, and there is no specific treatment or vaccine that is currently effective against the disease. Metabolomics and lipidomics provide global views of metabolome and lipidome landscapes and implicate metabolic to disease phenotype. We performed serum metabolic and lipidomic profiling on a cohort of dengue patients with three sampling time points at early febrile, defervescence, and convalescent stages via mass spectrometry-based analytical platforms. Compared with healthy subjects, approximately two hundred metabolites showed significant difference in dengue patients, and 60 were identified. This study revealed that in primary dengue infection, the host metabolome is tightly regulated, with active, early anti-inflammatory processes modulating the pro-inflammatory processes, suggesting the self-limiting phenotype of dengue fever. Major perturbed metabolic pathways included fatty acid biosynthesis, fatty acid β-oxidation, phospholipid catabolism, steroid hormone pathway, etc. This represents a first report on the characterization of the serum metabolome and significantly advances our understanding on host and dengue virus interactions. These differential metabolites have the potential as biomarkers for disease monitoring and evaluation of therapeutic interventions.
| Dengue virus (DENV) is a member of the Flavivirus genus of the Flaviviridae family and is classified into four distinct serotypes. It is estimated that 100 million dengue infections occur worldwide every year, ranging from acute dengue fever (DF) to life-threatening dengue haemorrhagic fever/dengue shock syndrome (DHF/DSS) [1], [2]. Dengue is emerging as a global health concern with the annual average number of DF/DHF cases reported to World Health Organization increasing dramatically in recent years. Furthermore, DENV has spread widely, and all four serotypes are now circulating in Asia, Africa and the Americas [3].
DF is generally self-limiting, but its symptoms can be debilitating and cause considerable incapacitating morbidity, which have a significant health and economic toll in the society. In a small percentage of patients, DF evolves to the more severe forms of DHF and DSS, which are characterized by abnormal hemostasis, vascular leakage and liver damage. There is currently no specific antiviral therapy or vaccine available for DF or DHF/DSS. Furthermore, the underlying molecular mechanisms of DENV pathogenesis are still unclear and why DF resolves in time is largely neglected. Therefore, the pathological differences between the severe DHF/DSS and the mild, self-limiting, DF suggest differential virus-host interactions in the susceptibility to the disease. Both viral and host immune factors seem to be involved, but the role of each is not fully understood [4]. Meanwhile, the lack of an appropriate animal model significantly increases the challenges in the study of DENV pathogenesis.
Metabolomics and lipidomics are rapidly emerging fields of ‘omics’ research that aim to study the global changes of small molecule metabolites and lipids in biological systems in response to biological stimuli or perturbations [5]. Metabolites are the end products of cellular regulatory processes, forming a link between molecular changes and phenotype, and therefore reflect the physiological state of a cell, tissue or organism at a point in time. Lipidomics covers the subset of lipid constituents, which are known to be involved in signalling and infection biology. Cellular homeostasis is interrupted under disease conditions, and the human body would attempt to maintain a basal, internal environment by increasing or decreasing levels of certain endogenous metabolites. Thus, by providing a snapshot of the metabolic status of an organism, metabolomics holds the promise of finding metabolic pathways related to disease processes [6], [7]. Metabolomics has been applied to infectious diseases to study host-pathogen interactions [8], [9], [10]. Similarly, there has been enormous interest in how lipids and their metabolism influence Flaviviridae viral infection, especially from the perspective of microbe lipid usage for virus entry, replication and release [11], [12], [13]. As we redefine the manner in which we understand the complex and dynamic virus-host interactions of dengue infection in vivo, it necessitates capturing the global changes rather than isolated ones. However, few omics studies exist that describe how the human host response biochemically and physiologically during DENV and other flavivirus infection and how it resolves without intervention [14], [15]. Therefore we turned our attention to examine serum metabolites and lipids using a comprehensive approach to globally capture the human responses to DF.
In our previous study, a systematic characterization of serum cytokines, proteome, and markers of macrophage and neutrophil activity was reported from a subset of longitudinally enrolled adult patients with primary dengue infections [16]. In the present study, both untargeted metabolomics and targeted lipidomics on the same serum sample set were conducted with the aim to identify metabolic pathways linked to disease progression and understand the mechanisms of DENV infection. Our results showed that DENV infection caused significant serum metabolome and lipidome-wide changes in DF patients. Sixty differential metabolites were identified in metabolomics analysis and detailed examination of metabolic pathways changes revealed that in primary DENV infections, acute inflammatory responses were quickly met with active pro-resolution counterbalances to restore homeostasis. Our studies also implicated the plausible association of sphingomyelins with CD8+ lymphocyte activation during the early febrile period, and plasmalogen phosphatidylethanolamines and lysophosphatidylethanolamines with platelet numbers during the later phases of defervescence and convalescence. We propose that early widespread host changes in the metabolome and lipidome play an important physiological component in defining the self-limiting outcome of DF.
The early dengue infection and outcome (EDEN) study is a multi-center longitudinal study, which prospectively recruits and follows-up adult dengue patients in Singapore, through early febrile, defervescence as well as convalescence stages of the disease [17]. Enrolment of all eligible individuals was based on written informed consent and the protocols were approved by the National Healthcare group (DSRB B/05/013). All samples were anonymized.
The details of patient recruitment, sample collection and the study protocols of EDED study have been described earlier [18]. In brief, adult patients (>21 years) presenting with acute onset fever (≥38.0°C for less than 72 hours) without rhinitis or other clinical alternatives were included in the study (Visit 1; febrile). Initial dengue diagnosis was made by real time RT-PCR and followed by serology and subsequent serotyping by virus isolation and immunofluorescence using serotype specific monoclonal antibodies (ATCC: HB46-49). Venous blood samples were also collected at fever day 4 to 7 (Visit 2; defervescence) and weeks 3 to 4 (Visit 3; convalescence), aliquoted and frozen at −80°C. ‘Fever day’ here refers to number of days post onset of fever. Classification of DF was made based on the guidelines provided by the World Health Organisation [19]. In brief, acute febrile patients positive for dengue with one or more of the following symptoms: headache, retro orbital pain, myalgia, rash, leucopenia, hemorrhage were classified as DF. Of the 133 dengue patients that were finally enrolled in this study (September 2005–October 2006), 44 DF patients tested negative for dengue IgG antibodies in the acute sera, using a commercially obtained ELISA (PanBio, Brisbane, Australia). These patients were deemed to have primary DENV infection, all of which were included in this study. A detailed hematological and virological analysis was also performed. Additionally we used serum samples from 50 asymptomatic age-matched healthy subjects participated during a hospital staff annual examination as controls (Table 1). This study was approved by the National University of Singapore Institutional Review Board and samples were collected with individual informed written consents.
For metabolomics analysis, a volume of 50 µL from each serum sample was thawed at 4°C. Serum proteins were precipitated with 200 µL ice-cold methanol, which contained 10 µg/mL 9-fluorenylmethoxycarbonyl-glycine as an internal standard. After vortexing, the mixture was centrifuged at 16,000 rpm for 10 minutes at 4°C and the supernatant was divided into two parts, one for liquid chromatography/mass spectrometry (LC-MS) analysis and the other for gas chromatography/mass spectrometry (GC-MS) analysis. All samples were kept at 4°C and analyzed within 48 h. In order to prevent batch effect, the assays were conducted in random manner. Quality control (QC) samples were prepared by mixing equal amounts of serum samples from 10 healthy subjects and 10 DF patients at all the three time points.
For lipidomics analysis, due to limited sample availability, sera from three patients at each visit were randomly pooled (14 pools total). Lipids were extracted from serum using a modified Bligh and Dyer method [20]. Briefly, 900 µL of chloroform-methanol, 1∶2 (v/v) was added to 100 µL serum. After 20 min vortexing and incubation at 4°C, 300 µL of chloroform and 300 µL of ddH2O were added to the mixture and centrifuged at 9000 rpm, 4°C for 2 min. Lipids were then recovered from the lower organic phase after centrifugation. Subsequently, 500 µL of chloroform was added, vortexed at 4°C for 20 min. After centrifugation, lipids were recovered from the organic phase and combined with the previous fraction. The lipid extracts were vacuum-dried, stored at −80°C and analyzed within a week.
The supernatant fraction from sample preparation step was analyzed using Agilent 1290 ultrahigh pressure liquid chromatography system (Waldbronn, Germany) equipped with a 6520 Q-TOF mass detector managed by a MassHunter workstation. The column used for the separation was an Agilent rapid resolution HT Zorbax SB-C18 (2.1×50 mm, 1.8 µm; Agilent Technologies, Santa Clara, CA, USA). The oven temperature was set at 50°C. The gradient elution involved a mobile phase consisting of (A) 0.1% formic acid in water and (B) 0.1% formic acid in methanol. The initial condition was set at 5% B. A 7 min linear gradient to 70% B was applied, followed by a 12 min gradient to 100% B which was held for 3 min, then returned to starting conditions over 0.1 min. Flow rate was set at 0.4 ml/min, and 5 µL of samples was injected. The electrospray ionization mass spectra were acquired in both positive and negative ion mode. Mass data were collected between m/z 100 and 1000 at a rate of two scans per second. The ion spray voltage was set at 4,000 V, and the heated capillary temperature was maintained at 350°C. The drying gas and nebulizer nitrogen gas flow rates were 12.0 L/min and 50 psi, respectively. Two reference masses were continuously infused to the system to allow constant mass correction during the run: m/z 121.0509 (C5H4N4) and m/z 922.0098 (C18H18O6N3P3F24). The stability of the LC-MS method was examined and evaluated by a subset of peaks covering a range of masses, intensities, and retention times across the QC samples (Table S1). Variations of retention time and m/z values of the peaks were less than 0.2 min and 10 mDa, respectively, and the relative standard deviations of peak areas were below 20%, indicating good reproducibility and stability of the chromatographic separation and mass accuracy of mass measurement during the whole sequence.
100 µL supernatant fraction from sample preparation step was dried under nitrogen and derivatised with 150 µL methoxamine (50 µg/mL in pyridine, 37°C×2 h) followed by 150 µL N-methyl-trimethylsilyl-trifluoroacetamide (37°C×16 h). After centrifugation (4°C, 6000 rpm×1 min), the supernatant fraction was injected into GC-MS. The derivatised sample (1.0 µL) was introduced by splitless injection with an Agilent 7683 Series autosampler into an Agilent 6890 GC system equipped with a fused-silica capillary column HP-5MSI (30 m×0.25 mm i.d., 0.25 µm film thickness) as reported previously [21]. The inlet temperature was set at 250°C. Helium was used as the carrier gas at a constant flow rate of 1.0 ml/min. The column effluent was introduced into the ion source of an Agilent Mass selective detector. The transfer line temperature was set at 280°C and the ion source temperature at 230°C. The mass spectrometer was operated in electron impact mode (70 eV). Data acquisition was performed in full scan mode from m/z 50 to 550 with a scan time of 0.5 sec. The compounds were identified by comparison of mass spectra and retention time with those of reference standards, and those available in libraries (NIST 0.5). A summary of the workflow utilized in metabolomics study is shown in Figure S1.
To facilitate interrogation of the serum lipidome in a high-throughput, targeted and accurate fashion, we employed LC-MS/MS in Multiple Reaction Monitoring (MRM) mode to survey more than 200 individual glycerophospholipids and sphingolipids simultaneously at linear dynamic ranges of 0.1 to 100 ng/mL. Prior to ESI-LC-MS/MS, each lipid extract was reconstituted in 400 µL of chloroform-methanol (1∶1 v/v). LC-MS/MS via Triple Quadrupole 6460 with Jet Stream™ (Agilent Technologies) was used for the quantification of glycerophospholipids and sphingolipids (total >200 molecular species). C-18 reversed phase LC (Zorbax Eclipse 2.1×50 mm I.D., 1.8 µm; Agilent Technologies) was used to separate 5 µL reconstituted lipids at 400 µL/min before entering the mass spectrometer. The column thermostat and autosampler temperatures were maintained at 40°C and 6°C, respectively. The optimized mobile phase consisted of 5 mM ammonium acetate in water (solvent A) and 5 mM ammonium acetate in methanol (solvent B). The gradient of the samples was: 0–2 min, 60–100% B; 2–7 min, maintained at 100% B; 7–9 min, 100–60% B and 9–11 min, 60% for column re-equilibration before the next injection. Positive ionization mode [M+H]+ MRM was based on product ion m/z 264.4 [sphingosine-H2O]+ and 266.4 [sphinganine-H2O]+ for ceramide (Cer), monohexosylceramide (MHexCer), lactosylceramide (LacCer) and ceramide-1-phosphate (C1P), and m/z 184.1 [phosphocholine]+ for sphingomyelin (SM). Negative ionization [M-H]− MRM mode based on specific parent and product ion pairs was employed for phosphatidic acid (PA), phosphatidylinositol (PI), lysophosphatidylinositol, phosphatidylethanolamine (PE), lysophosphatidylethanolamine (lysoPE), plasmalogen phosphatidylethanolamine (pPE), phosphatidylglycerol (PG), phosphatidylserine (PS) and lysophosphatidylserine (lysoPS) for identification and quantification of the above glycerophospholipids [22]. The optimized lipid-class dependent mass spectrometry parameters are shown in . Known concentrations (20–100 ng/mL) of non-naturally occurring internal standards, namely Cer d18:1/17:0, GlcCer d18:1/8:0, LacCer d18:1/8:0, SM d18:1/12:0, C1P d18:1/17:0, PA 14:0, PC 14:0, PE 14:0, PG 14:0, PS 14:0 and PI 14:0; Avanti Polar Lipids, Alabaster, Al) were verified of their low abundance in sera, spiked into extracted samples and used in absolute quantification. The linear dynamic ranges, limits of detection, limits of quantification and relative standard deviation of the various lipids are shown in Table S3. The effect of acyl length on signal intensity [23], [24] was compensated by adjusting the collision energy and fragmentor voltage (Table S2). A summary of the workflow utilized in the lipidomics study is shown in Figure S2.
For metabolomics analysis, raw spectrometric data were converted to mzData (LC-MS/MS) and NetCDF (GC-MS) formats via Masshunter (Agilent, US) and input to open-source software MZmine 2.0 for peak finding, peak alignment and peak normalization across all samples [25]. The MZmine 2.0 report table was exported into SIMCA-P software (version 11.0; Umetrics AB, Umea, Sweden) for multivariate data analysis. Data sets were mean-centered and pareto-scaled to adjust the importance of high and low abundance metabolites to an equal level before carrying out statistical analyses. Unsupervised multivariate analysis principle component analysis (PCA) was first performed on LC-MS data as an unbiased statistical method to observe intrinsic trends within healthy controls and DF patients at different visits. To achieve the maximum separation among the groups, supervised multivariate analysis orthogonal partial least-squares discriminant analysis (OPLS-DA) was sequentially applied. Potential differential metabolites were selected according to the Variable Importance in the Projection (VIP) values and the variables with VIP>1 were considered to be influential for the separation of samples in OPLS-DA analysis. In addition, the Kruskal-Wallis test was performed to determine if the differential metabolites obtained from OPLS-DA modelling were statistically significant (p<0.05) among groups at the univariate analysis level.
For lipidomics analysis, undetectable lipids were replaced with 1/10 of the lowest non-zero value measured. In cases where 70% of data are missing from patients, lipids were excluded. Missing data represents 1.8% of the data. In total 207 lipids remained for further analysis. Comparisons between groups were achieved using non-parametric analysis namely two-tailed Mann-Whitney tests and Kruskal-Wallis tests with Dunn's post hoc analysis. Repeated one-way ANOVA was used for paired analysis with the Bonferroni Multiple Comparison post hoc Test.
The structure identification of the differential metabolites was based on a previously described strategy [26]. The identification process is illustrated here with hexanoylcarnitine, one of the ten differential acylcarnitines. First, the element composition C13H25NO4 of the m/z 260.18 ion was calculated based on the exact mass, the nitrogen rule and the isotope pattern by Masshunter software from Agilent (Figure S3A). Then, the elemental composition and exact mass were used for open source database searching, including LIPIDMAPS (http://www.lipidmaps.org/), HMDB (http://www.hmdb.ca/), METLIN (http://metlin.scripps.edu/) and MassBank (http://www.massbank.jp/). Next, MS/MS experiments were performed to obtain structural information via the interpretation of the fragmentation pattern of the metabolite (Figure S3B). In the case of acylcarnitines, m/z 85.0 is the characteristic fragment ion in the MS/MS spectrum [27]. The MS/MS spectra of possible metabolite candidates in the databases were also searched and compared. As a result, the metabolite was identified as hexanoylcarnitine, which was finally confirmed by comparison with the standard (Figure S3C). Both MetaboAnalyst [28] and Ingenuity Pathway Analysis software (IPA, www.ingenuity.com) were used to identify relevant pathways.
In order to obtain a high level snapshot of metabolite flux at the different stages of dengue infection represented in our clinical samples, we examined variances in our data both by principal component analysis (PCA) as well as OPLS-DA. The PCA score plot (Figure 1A) indicates that the metabolome changes in DF patients are reversible. The most predominant metabolome changes in DF patients, as compared with healthy controls, happened at early febrile stage (Visit 1), followed by the defervescence stage (Visit 2) and finally returned to the control levels at the convalescence stage (Visit 3). This indicates that the most intense responses of the DF patients to DENV infection happen within the first 72 hours upon infection, and homeostatic recovery appears to begin around the fourth day from the onset of fever. This, more or less, coincides with the course of viraemia, which has been observed to peak during early febrile phase and are generally undetectable during the defervescence stage. The duration of DF illness in the EDEN study is reported to be 7–14 days [18], and a majority of patients recover within three to four weeks as seen at Visit 3, which could explain the metabolome similarities between patients at Visit 3 and healthy controls in the PCA score plot. In addition, PCA did not reveal distinct metabolome profiles between DENV serotype 1 and DENV serotype 3. This suggests converging host responses between the two serotypes which resulted in DF. While host molecular changes may be present between infections with different DENV serotypes, a larger patient set will be required to tease out the significant subtle differences.
OPLS-DA, which is a more rigorous classification tool, revealed clearer segregation of the healthy controls and DF patients during the three visits (Figure 1B). The quality of the OPLS-DA model was evaluated by R2 (Y) and Q2 parameters, which indicated the fitness and prediction ability, respectively. By using four principle components, the OPLS-DA model showed a R2 value of 0.95 and Q2 value of 0.91, indicative of optimal performance of the model. The OPLS-DA was also applied to GC-MS data, which showed a R2 value of 0.65 and Q2 value of 0.60.
Approximately 200 differential metabolites identified via both LC-MS/MS and GC-MS analyses were selected based on the criteria- VIP>1 in OPLS-DA analysis and p<0.05 in Kruskal-Wallis test (Table S4, S5). Sixty of these metabolites were structurally identified, with 56 through LC-MS/MS analysis, and 8 by GC-MS analysis, 4 of which were common to both these methods. These metabolites belonged to classes such as free fatty acid, acylcarnitine, phosphatidylcholine (PC), lysophosphatidylcholine (LysoPC), LysoPE, amino acid and derivative, sphingolipid, monoacylglyceride (MG), diacylglyceride (DG), purine, bile acid, bile pigment, steroid hormone, glucose and phenol sulfate (Table S4). Consistent with the PCA score plot (Figure 1A), the quantitative heatmap (Figure 2) of these differential metabolites revealed that the metabolome changes were reversible in DF patients. They were significantly perturbed during the early stages, and normalized to control levels at convalescent stage.
These differential metabolites show two contrasting time-course changes, an elevated and a decreased trend, at Visit 1 and Visit 2. Most of the metabolites have either the highest or the lowest levels at Visit 1, which is the acute stage of DENV infection, and gradually return to the control levels at Visit 3, indicating that they might serve as prognostic markers of the disease.
It is interesting that the metabolites within each metabolite class generally showed a similar change trend. Specifically, seven FFAs, including four polyunsaturated fatty acids (PUFAs): arachidonic acid (AA; 20:4n-6; Figure 3A), linoleic acid (18:2n-6), docosahexaenoic acid (DHA; 22:6n-3) and α-linoleic acid (18:3n-3), showed an elevated trend at Visit 1 and Visit 2. The PUFAs are released from cell-membrane phospholipids by various phospholipase enzymes, predominantly phospholipase A2, which is elevated in dengue patients [29]. Similarly, ten acylcarnitines, two sphingolipids (glucosylceramide and SM), two MG, one DG, one steroid hormone (cortisol), two purine nucleosides (hypoxanthine and inosine), one bile pigment (biliverdin), two conjugated bile acids (glycoursodeoxycholic acid and glycocholic acid), and glucose showed an increasing trend at Visit 1 and Visit 2 (Figure 3B–3F). Conversely, a total of twenty PCs, LPCs and lysoPEs showed a decreased trend (Figure 3G). Ten amino acids and derivatives, including tryptophan (Figure 3H), glutamic acid etc, also displayed a downward trend, with the exception of phenylalanine (Figure 3I), pyroglutamic acid, L-leucyl-L-proline, which showed an upward trend. The changes of amino acids observed are consistent with a previous report that described an overall decrease in plasma amino acid levels in dengue patients, with the exception of increased phenylalanine in the acute phase of DF patients [30].
We used two well-known pathway analysis tools- MetaboAnalyst and IPA, to determine if the metabolites identified here revealed underlying biochemical pathways. Both these tools provided similar results, revealing that disturbed metabolic pathways included fatty acid biosynthesis, fatty acid β-oxidation, phospholipid catabolism, arachidonic acid metabolism, sphingolipid metabolism, tryptophan metabolism, steroid hormone biosynthesis, purine metabolism, heme degradation pathway, bile acid biosynthesis, etc (Figures S4). The Networks function in IPA indicated that the most significant changes happened with lipid metabolism and energy production pathways, which is summarized in Figure 4.
OPLS-DA revealed distinct temporal serum lipidome profiles with DENV infection (Figure 5A). To understand the lipids and blood chemistry parameters responsible for the differentiation of the three visits in the OPLS scores plot, the VIP plot was generated (Figure 5B). Lymphocyte numbers, SMs, PCs and lysoPEs influenced the model most significantly. We noted strong negative correlation between total SM with lymphocyte percentage (Pearson r2 = 0.631, 95% C.I. = −0.8848 to −0.6467, p<0.0001; Figures 6A, S5A). The downward trend of SM levels across the visits (repeated measures 1-way ANOVA, p<0.0001) was dominated by SM d18:1/16:0 (m/z 703.7→184.1) which represented ∼1/3 of total SM and exhibited a 5-fold decrease from Visit 1 to 2 (Figure 6B). In cytotoxic T lymphocytes, acid sphingomyelinase, the enzyme that hydrolyzes SM to Cer, is a critical factor in the secretory granule-mediated cell death of virus infected cells [31]. Concordant with SM hydrolysis to ceramides [32], we observed increased ceramide levels paralleling the attenuation of SM (Figure S5B). Unlike the more biologically inert SMs, Cers are potent apoptosis-inducers, and their glycosylated forms, MHexCers or phosphorylated forms, C1Ps are bioactive sphingolipids [33]. The initial low Cer levels were not due to its glycosylation or phosphorylation which remained low (Figures S5C, S5D), but plausibly compartmentized as the less bioactive SM during the early onset of DF. This is evidenced by the relatively high SMs levels in the febrile phase compared to the later defervescence and convalescence phases (Figure 6C). The development of DENV-specific CD4+ and CD8+ T-lymphocyte responses [34], and the activation and contraction of antiviral CD8+ T lymphocytes in the defervescence and convalescence phases of dengue fever has previously been documented [35]. Therefore, the decrease in SM and increase in Cer correlated well with the action of T lymphocyte sphingomyelinase on SM and Cer levels in generating cytotoxic granules [31].
In our study cohort, MHexCers significantly increased from Visit 1 to Visit 3 (repeated measures 1-way ANOVA, p<0.01; Figure S5C), verifying our untargeted metabolomics results. Conversely, the higher glycosphingolipid, LacCer did not show significant differences across the visits (Figure S5E).
In our study we found many lysoPEs and pPEs closely correlated to platelet numbers (Pearson r2 = 0.263–0.4, p<0.0001; Figures 6C, S6A). By contrast, PEs did not correlate with platelet numbers (Pearson r2 = 0.0404, 95% C.I. = −0.4760 to 0.1096, p = 0.217). In addition, their temporal trends differed from pPEs and lysoPEs (Figures S5B–D). These pPEs are polyunsaturated and long-chained lipids (34p:2, 36p:4, 38p:4; 38p:5, 38p:6 and 40p:6. Plasmalogen glycerophospholipids are characterized by vinyl ether bonds at sn-1 position of the glycerol backbone, and two predominant glycerophospholipids are PEs and PCs. We also found that LPEs and pPEs strongly correlated to white blood cell count (Pearson r2 = 0.425, p<0.0001). pPEs with no or single double bonds and of chain lengths more than 40 carbons correlated poorly, suggesting apparent alkyl chain length specificity to white blood cell count. Furthermore, univariate 1-way ANOVA showed that lysoPEs are significantly increased between Visits 2 and 3 (p<0.0001), and to a lesser degree between Visits 1 and 2 (p<0.05; Figure S5B). Taken together, our combined lipidome and blood chemistry analyses revealed that SMs are associated with early host response whereas lysoPEs are more associated with late recovery responses.
DF is a self-limiting febrile illness and in most DF cases, there is complete recovery with little to no manifestations of chronic negative clinical consequences, clearly indicating human-adapted protective mechanisms against DENV in DF patients. The protective mechanisms are complex and unfortunately, poorly studied, in part due to the focus on heterotypic infection and DHF/DSS. The study of the recovery of DF from a physiologic perspective is therefore important in advancing our understanding how humans mount a defense and recover from DENV infection and potentially other flaviviruses-induced illnesses. Our current integrated metabolomics and lipidomics study provides an extensive map of physiologic changes in adult DF patients, through which we have identified altered metabolic pathways linked to DENV infection. The relatively large population size (44 patients and 50 healthy controls) with repeated sampling at multiple time-points of infection, including early febrile, defervescence and convalescence stages, captured dynamic changes in metabolites and the discovery of reliable differential metabolites that were closely associated with dengue pathophysiology. By using both LC-MS/MS and GC-MS, we found that approximately 200 serum metabolites were significantly changed, of which 60 were identified. Lipids, which form a biochemically important subclass of metabolites, were also independently verified and studied. These metabolites constituted a mix of pro- as well as anti-inflammatory mediators, suggesting the rapid host resolution to limit excessive inflammation and curtail further exacerbation of DENV infection. Furthermore, these metabolites showed reversible trend changes, indicating that they might be able to potentially serve as the basis to develop prognostic markers of the disease.
Infection of humans with DENV causes early pronounced immunological reactions that shift towards pro-inflammatory responses [36]. While acute inflammatory response is necessary in initiating pathogen killing, activation of anti-inflammatory and pro-resolution processes is important to prevent excessive pathological inflammatory damage to the host. Omega-6 PUFAs, which are released from cell-membrane phospholipids by phospholipid hydrolysis, including AA and AA-derived mediators such as prostaglandins, thromboxanes, hydroxyeicosatetraenoic acids and leukotrienes (collectively known as eicosanoids) are potent mediators of inflammation and are involved in modulating both the intensity and duration of inflammatory responses [37], [38]. Contrary to the effects of AA, omega-3 PUFAs, e.g. DHA are known for their anti-inflammatory activity [39]. Our results showed an elevated trend of pro- and anti-inflammatory PUFAs during the febrile stage of DENV infection, including both AA and DHA, as are total hydroxyeicosatetraenoic acids [40]. Such effects of active resolution, where AA-associated inflammation incited early during DENV infection is concomitantly resolved by DHA-derived anti-inflammatory mediators, and the suppressed production of pro-inflammatory cytokines [39], is believed to embrace the current model of inflammation whereby the return to homeostasis is an active process, as opposed to a passive one [41]. We also found increased levels of cortisol, a major, cytokine-activated glucocorticoid of potent anti-inflammatory and immunosuppressive properties in the febrile stage of DF. Viral infections are capable of inducing an increase of cortisol in serum during DENV infection, it had been reported that cortisol level at early febrile stage was significantly higher than its level at convalescent stage [42], consistent with our results. The functional role of cortisol in restraining inflammatory and immune responses by inhibiting the production of cytokines and other pro-inflammatory mediators [43] could possibly down-regulate immune responses and protect the DF patients against cytokine-mediated pathologies. Consistent with elevated oxidative stress in adult DF patients [44] and elevated inosine concentrations in the extracellular space at times of inflammation [45], we found an increased serum inosine in the febrile phase. Inosine, potentially formed through nitrosative deamination of adenosine, has been shown to exert wide-ranging anti-inflammatory effects both in vivo and in vitro, such as inhibition of pro-inflammatory cytokine/chemokine production and the enhancement of anti-inflammatory cytokine IL-10 production [46]. Taken together, our metabolomics results reveal a dynamic metabolite flux in dengue where an acute inflammation is limited by swift, dampening responses to restore homeostasis.
In parallel, host immunity has interested dengue investigators in exploring the mechanisms in the immune system underlying DF pathogenesis. In primary infections, neutralizing antibody responses develop after DENV infection and provide persistent protection against the same serotype even in repeated infections; even though this may lead to the enhancement of disease severity upon challenges of a following heterotypic DENV infection. Increasing evidence demonstrate that self-antigenic glycerophospholipids or glycosphingolipids activate the innate immune system, noticeably NKT cells in microbial infections [47]–[50]. Two MHexCers, GlcCer and GalCer, ceramides with β-linked glucose and α-linked galactose, produced by sphingolipid biosynthesis, have been implicated in invariant natural killer T-cell (iNKT) activation during microbial infections [47], [49], [50], and also NKT cells, which have been implicated in DENV clearance in DF patients [51], [52]. Consistent with this, we found increased GlcCer levels with the visits and followed the same trend as lymphocyte numbers. Other immune cells such as dendritic cells, CD4+ and CD8+ lymphocytes play major roles in viral clearance and as targets of DENV infection. On the basis of the aforementioned observations, we postulated that the relationship between SM, Cer concentrations and immune cells describe multiple early, molecular roles of sphingolipids in mediating the host interactions during DF. Central among the sphingolipid enzymes is acid sphingomyelinase, which has been implicated in numerous immune responses including, T lymphocyte activation and positive infection by ebola viruses [53], entry of measles viruses into dendritic cells [54] and exocytosis of cytolytic granules by T cells [31]. The binding of extracellular ceramide to the leukocyte mono-immunoglobulin-like receptor 3 has been recently reported in repressing mast cell activation [55] and serum ceramides could potentially influence mast cell surveillance in DENV infection [56]. Likewise, we also observed the correlation of LPEs and pPEs with white blood cells and may potentially be related to iNKT activation. Plasmalogen lysoPEs (pLPE) are self-antigenic and are required for iNKT stimulation under both physiological conditions and when challenged with bacteria [48]. We expect that, our findings that sphingolipids are associated with lymphocytes during DF may further unravel the significance of T-cell responses and warrants further investigations. The lack of any significant change in phosphatidic acids and its lyso form was intriguing given its commonly assigned roles in platelet activation. However, it could be because unlike DHF, DF does not entail severe thrombocytopenia and our blood chemistry assays suggest the same.
In the wide spectrum of responses occurring simultaneously with pathogenic invasions, some metabolic changes may appear to be general and others more specific. In influenza 2009 pandemic H1N1 infection, extensive changes in serum eicosanoid (e.g. prostaglandin-F1α, prostaglandin-G, LTA4) and linolenic acid levels were observed [57]. While linolenic acid is a FFA common to both H1N1 infection and dengue infection, there was no observable difference in eicosanoids in DF. Another plausible distinctive metabolome change in DF compared to H1N1 is the xanthinine derivatives. In 2009 pandemic H1N1 infection, serum methylxanthine levels were lower during infection relative to after treatment [57]. This result can be interpreted in two ways: theophylline, a methylxanthine has been shown to elicit immunomodulator, anti-inflammatory, and bronchoprotective properties [58] and such properties may help attenuate the bronchi damaging effects of 2009 pandemic H1N1 virus as i) part of the host response [59] or ii) constitute part of the unspecified treatment. Regardless of which is the reason, methylxanthine points towards the lung-injury specific consequence of 2009 pandemic H1N1 infection. On the other hand, in DF, circulating inosine and hypoxanthine levels increased in the febrile phase where systemic inflammation may be concerned [60], [61]. While studies of increased inosine and hypoxanthine levels were demonstrated in models of chronic systemic inflammation, it is conceivable that similar events could be observed in infectious diseases where acute bouts of inflammation occur. During inflammation, increased adenosine-to-inosine editing in cytokine mRNA may interfere with the degradation of cytokine transcripts and support the early pro-inflammatory and resolving responses necessary in the early phase of dengue. This may be reflected in a mixture of pro- and anti-inflammatory cytokines such as significant IL-4, IL-8 and IFNγ increases in the febrile phase of DF [16].
On the other hand, certain metabolite changes may be more general across different pathogenic infections. For example, we found increased phenylalanine levels in the early stages of DF. Tetrahydrobiopterin (BH4) is a co-factor for phenylalanine (4)-hydroxylase (PHA), an enzyme required for converting phenylalanine to tyrosine, and also NO synthase (NOS) [62]. Through immune and inflammatory responses-mediated up-regulation of NOS, NO is elevated during dengue infection [63], [64]. Therefore, NOS competes with PHA for BH4 and this causes an accumulation of unconverted phenylalanine in the blood. This is aggravated in the elevated oxidative stress status in DF patients [44] where superoxide and peroxynitrite increases BH4 oxidation, thereby depleting the available BH4 pool. Yet, an increased level of phenylalanine has been commonly observed in other infectious diseases [65] and given NO's diverse functions in immunity and inflammation responses it is not surprising that such changes could occur in both bacterial and viral infections [66], [67]. Another general effect is energy metabolism. In this study we identified ten acylcarnitines, which are essential intermediates of fatty acid β-oxidation (FAO). FAO in liver is suppressed during infection and inflammation, and increased FFA substrates are then moved away from oxidation and directed toward re-esterification into TG [68]. The elevated levels of acylcarnitines suggest disturbed FAO in DF patients, which is in agreement with the observation that the TCA cycle was affected in an endothelial cell line infected with DENV [14]. Proteomics and lipidomics work by Diamond et al., showed a shift in energy metabolism during HCV infection in Huh7.5 cells, plausibly as a compensatory measure for the energy-demanding biosynthesis [15]. Similarly, a metabolomics study on severe pneumonia identified energy metabolism as a perturbed pathway [10]. Our relatively large number of human samples not only corrobated the in vitro findings but also showed that energy metabolism is likely to be a general perturbed pathway in acute infections. In the absence of large datasets of host-pathogen ‘functional omic’ datasets, we are limited in our ability to make comparisons and evaluate pathways specific to either bacterial or viral infections in humans. However, with increasing need to comprehensively capture the complexity of host-pathogen interactions, and aided by advancements in mass spectrometry and computational biology [69], we expect more datasets made available to the community for meaningful comparisons.
The limitations of our study can be discussed in the context of the host. Firstly, since the current study is restricted to adult primary infections the findings still need to be expanded and evaluated in pediatric patients as well as patients with secondary infections. Secondly, DENV progressively infects a multitude of cells and organs, including dendritic cells, monocytes, macrophages, endothelial cells and liver [70]–[72]. When we approached the host response from a systemic perspective, changes of the chemical composition in the serum could reflect tissue lesions, organ dysfunctions, pathological states and also compensatory responses. Thus pinpointing the source or cause of these changes can be challenging. Nevertheless, our results strongly raise the implication of distinctive metabolome changes with specific physiological responses at different phases of dengue infection. It will require increasing volumes of investigative work to tease apart the sources of these metabolites, including lipids. More importantly, as the body of metabolomic work on in vivo viral infections grows, the potential to derive distinctive and accurate biomarkers alluding to host responses of various viral infections increases.
In summary, our findings provide a first detailed description of the metabolome changes in patients with acute dengue infection and offer a global molecular view that leads to the overall homeostatic physiological outcome of DF. In addition to providing host-pathogen insights into the underlying mechanism in symptom manifestation, metabolites identified in this study might be used both to monitor disease progression as well as evaluate the efficacy of therapeutic interventions. Furthermore, certain pathways might serve as useful therapeutic targets to alleviate severity of dengue. Such work would be our next endeavour.
|
10.1371/journal.pcbi.1003418 | The Role of Thalamic Population Synchrony in the Emergence of Cortical Feature Selectivity | In a wide range of studies, the emergence of orientation selectivity in primary visual cortex has been attributed to a complex interaction between feed-forward thalamic input and inhibitory mechanisms at the level of cortex. Although it is well known that layer 4 cortical neurons are highly sensitive to the timing of thalamic inputs, the role of the stimulus-driven timing of thalamic inputs in cortical orientation selectivity is not well understood. Here we show that the synchronization of thalamic firing contributes directly to the orientation tuned responses of primary visual cortex in a way that optimizes the stimulus information per cortical spike. From the recorded responses of geniculate X-cells in the anesthetized cat, we synthesized thalamic sub-populations that would likely serve as the synaptic input to a common layer 4 cortical neuron based on anatomical constraints. We used this synchronized input as the driving input to an integrate-and-fire model of cortical responses and demonstrated that the tuning properties match closely to those measured in primary visual cortex. By modulating the overall level of synchronization at the preferred orientation, we show that efficiency of information transmission in the cortex is maximized for levels of synchronization which match those reported in thalamic recordings in response to naturalistic stimuli, a property which is relatively invariant to the orientation tuning width. These findings indicate evidence for a more prominent role of the feed-forward thalamic input in cortical feature selectivity based on thalamic synchronization.
| While the visual system is selective for a wide range of different inputs, orientation selectivity has been considered the preeminent property of the mammalian visual cortex. Existing models of this selectivity rely on varying relative importance of feedforward thalamic input and intracortical influence. Recently, we have shown that pairwise timing relationships between single thalamic neurons can be predictive of a high degree of orientation selectivity. Here we have constructed a computational model that predicts cortical orientation tuning from thalamic populations. We show that this arrangement, relying on precise timing differences between thalamic responses, accurately predicts tuning properties as well as demonstrates that certain timing relationships are optimal for transmitting information about the stimulus to cortex.
| Sensory systems serve the purpose of allowing us to extract perceptually relevant features from the environment. Although there are certainly examples of sensory features whose coding originates in the sensory periphery (e.g. auditory frequency, visual color, etc.), the more intriguing and less well understood phenomena involve the emergence of feature selectivity in more central brain structures that do not just inherit the selectivity from the periphery. Perhaps the most well studied of these phenomena is that of orientation selectivity in primary visual cortex (V1), where many if not most neurons in the mammalian primary visual cortex exhibit differential firing activity for visual stimuli at different orientations, despite the fact that the neurons projecting from the lateral geniculate nucleus (LGN) serving as input to V1 exhibit little to no orientation preference on their own [1] (see [2] for a review). This implies that the thalamocortical link is a transformative location for representation of stimuli as collections of particular features rather than samples (i.e. it does far more than simply relay luminance values to the cortex). This transformation can serve as a general model for how sensory systems convey increasing feature selectivity as the information moves to higher-order brain areas. How do these convergent thalamic structures drive cortical feature selectivity, and in what way do populations drive this selectivity?
The mechanistic origin of orientation tuning in V1 has been vigorously explored in the literature [1]–[5]. In their seminal work, Hubel and Wiesel outlined a conceptual model that involved the projection of LGN neurons along a particular axis of orientation to a common cortical target [1], the core connectivity of which was subsequently confirmed in recordings from connected pairs of neurons in LGN and V1 [6]–[8]. Although the relative roles of this feedforward architecture versus cortico-cortico connectivity in sharpening and refining orientation selectivity in such phenomena as contrast-invariance and cross-orientation suppression has been intensely debated [2], [9], the thalamic basis for the origin of the basic selectivity is not in dispute, and by its nature implies a role for the timing of thalamic inputs to the cortical target. That is, the several decade old proposal by Hubel and Wiesel conceptually suggests that an edge activating the subset of thalamic neurons projecting to a common cortical target at the same time would naturally drive the cortical neuron more so than when the thalamic inputs are activated at different times, establishing the orientation tuning for the cortical neuron. However, the precise role of timing of thalamic inputs in the downstream cortical orientation selectivity is not known. In the context of the natural visual environment, it has been shown that LGN neurons (individually and across pairs) are temporally precise to a time scale of 10–20 ms, a level that is matched to what is necessary to capture the timescale of changes exhibited in natural scenes [10]–[12]. Further it has been demonstrated that neurons in the primary visual cortex are extremely sensitive to short intervals between incoming thalamic spikes also on the time scale of approximately 10 ms [13]–[22] and that common cortical convergence is most probable when receptive fields overlap [7], [13]. All of these findings collectively suggest that feature selectivity is likely to arise from the modulation of precise timing among overlapping populations of neurons in LGN and that this modulation drives the coactivation of neurons within the populations. Finally, we have recently shown that considering just the coactivation between pairs of electrophysiologically recorded thalamic neurons reveals in many cases extremely sharp orientation tuning even when the receptive fields are highly overlapped [23].
Here, to explore the role of the precise timing of thalamic spiking in the orientation tuning of the downstream cortical neurons to which the thalamus projects, we utilized experimental population recordings of single units from the LGN region of the visual thalamus in concert with a large-scale thalamocortical model. Specifically, based on anatomical and physiological evidence concerning the convergence of thalamic input to cortical layer 4, we constructed thalamic sub-populations from experimentally recorded thalamic spiking in response to oriented visual stimuli, and systematically controlled the precise timing across the sub-population and its direct impact on the downstream orientation tuning. We found that the conventionally measured tuning sharpness was remarkably invariant over a wide range of peak LGN timing precisions, but the trial-to-trial variability in cortical response was strongly influenced by the timing precision of the LGN inputs. From a decoding perspective of an ideal observer of the cortical response, this complex relationship led to a decreasing error in estimation of orientation with increasing thalamic precision, and a corresponding increase in the information rate, both saturating for peak thalamic precisions of 10–20 ms, a finding which was invariant to the overall width of cortical orientation tuning. Taken together, the results here provide a compelling picture for the role of stimulus-driven thalamic synchrony in the emergence of cortical feature selectivity.
Neurons in layer 4 of primary visual cortex are driven by sub-populations of projecting LGN neurons with receptive fields that are highly overlapped, thus representing a relatively limited area of visual space [24]. Although individual LGN neurons are relatively insensitive to the orientation of drifting sinusoidal gratings, the synchrony across neuron sub-populations is often highly sensitive to the orientation, a product of the relative spatial geometry of the receptive fields and the underlying temporal dynamics of component neurons [23]. LGN populations which share a convergent cortical neuron are both large (approximately 30 neurons [8]) and highly overlapped. Since it is not currently possible to record from such dense and numerous clusters in the LGN, we implemented a population-filling method to quantify the synchronization properties of the sub-population. Specifically, in the population-filling method we utilized simultaneous recordings of spiking activity of small sub-populations of LGN neurons whose receptive fields span a small area of visual space (see Methods).
Single unit activity was collected in response to spatiotemporal white noise, and receptive fields (RFs) were mapped using standard spike-triggered averaging (see Methods). The RFs of a pool of simultaneously recorded LGN neurons are shown in Figure 1A, where the RF for each neuron is represented as the 20% contour. Note that in this recording, 5 neurons were recorded simultaneously, where each of these neurons is represented as a different color in the figure. We have previously provided experimental measures of the distribution of receptive field spacing of pairs of LGN neurons monosynaptically connected to a single cortical cell [8] and populations of LGN neurons to a single cortical orientation column [24], as shown with the dashed gray curve in Figure 1B. Specifically, this measure provides a probability distribution of the distances between receptive fields, as measured by the distance between the RF centers normalized by the diameter of the larger of the two RFs, referred to here in units of receptive field center diameter (RFCD) - see [24].
From experimental data in [24], the distribution of separations was modeled as , where x is the separation in units of RFCD, which is described only for the range of 0.4 to 2.0. Using the neurons in Figure 1A as templates and the relationship in Figure 1B (dashed line) as a rule, we filled out the assumed remainder of the population by translating the receptive fields in visual space, creating a dense and accurate convergent LGN population, as shown in Figure 1C. The receptive field centers were randomly shifted such that the amount of visual space covered did not change relative to the visual space covered by the original simultaneously recorded population. This method resulted in a distribution of RF separations consistent with previous experimental findings (simulated distribution shown with solid black circles, Figure 1B). Note that because the original population was itself elongated in the horizontal axis, the resultant shifts for this population were also mostly horizontal although some receptive field locations also moved vertically. The resultant cluster of receptive fields would be typical for a population that has a major and minor axis as opposed to being more circularly arranged. The resulting aspect ratio of the cluster of RFs in Figure 1A is approximately 2.4∶1, when measured as the ratio of the longer dimension to the shorter dimension of the area covered by the RF contours. It is important to note that this aspect ratio is lower than the majority of existing models [1], [3]–[5], where aspect ratios range from 3 to 4 (but see [5] for a much smaller aspect ratio).
Spiking activity was also collected in response to drifting sinusoidal gratings (0.5 cycles/degree, 5 Hz, 100% contrast - see Methods). The individual LGN neurons had mean firing rates that ranged from 16 to 28 Hz which were relatively insensitive to the stimulus orientation. To generate the population activity in response to the drifting gratings, we utilized the spatially translated RFs as described above, and imposed temporal shifts in the spiking activity based solely on the geometry related to the RF centers, as illustrated in Figure 1D. Specifically, a spatial translation of the RF by x degrees horizontally and y degrees vertically imposes a latency shift of the neural response by an amount proportional to the component of the vector connecting the centers of the two RFs orthogonal to the edge of the drifting grating, scaled by the speed of the drift (see Methods). For the collected datasets, spiking activity was collected at each of eight drifting directions with sinusoidal gratings. For each stimulus condition, each randomly placed neuron was assigned a random trial from the original neuron from which it was derived and the shift latency value was added to all spike times in the chosen trial. In this spirit, we view the trial to trial variability in spiking activity for a single neuron as representative of the across neuron variability on a single trial. The resulting population response at each orientation is shown in Figure 1E. For most orientations, spike times within the population uniformly distributed across the entire trial timespan. However, at 90 and 270 degrees, the spike times line up rather precisely between all neurons in the population, reflecting a high degree of synchrony at these orientations.
The degree of synchrony across this population of neurons is a function of the orientation of the drifting gratings, as well as the variability in spiking timing across neurons within the population. To quantify the synchrony, we used a timing jitter metric, which utilizes the width of the spike-time auto-correlation computed from all spikes in the population (roughly equivalent to the PSTH width). A brief overview of how the auto-correlation is calculated is demonstrated in 2A. The collection of spike times across the input population is collapsed into a single spike train, which represents all the projecting thalamic input on the cortical target neuron. This spike train is then used to calculate all of the pair-wise timing differences between every input spike in the population, the histogram of which forms the auto-correlation estimate. There are two values of interest: the population PSTH (with a width of ) and the “response timescale” of the auto-correlation function (given by ). These related values provide us with an approximation for the synchronization within the neural population. When synchrony is high, the spike time auto-correlation has a narrow width and thus there is little jitter. Alternatively, when synchrony is low, the auto-correlation has an increased width and jitter is very high, a property that is demonstrated in Figure 2B. From top to bottom in the figure, the level of synchrony in the population increases, spike times become more clustered, and the auto-correlation has a correspondingly decreasing width. Note that each auto-correlation covers the lag range from −400 ms to +400 ms. Each auto-correlation function was fit with a Gaussian between −100 and +100 ms to eliminate any effects of periodicity in response to the drifting sinusoidal grating. The corresponding width of this Gaussian fit was then utilized as the measure of timing jitter. As in [10], the timing jitter was defined as the half the latency at which the Gaussian fit is equal to (see Methods and Figure 2A). The timing jitter of the population is shown as a function of the stimulus orientation in Figure 2C, where the random sampling of single trials of the template neuron was repeated 50 times. At the most asynchronous stimulus orientations (in this case perpendicular to the elongated axis of the RFs of the population), the timing jitter was approximately 100 ms. At the preferred orientations, when synchrony was maximized, the timing jitter was approximately 24 ms. The timing jitter as a function of stimulus orientation was fit with a Gaussian function (gray dashed line in Figure 2C) and exhibited a characteristic tuning width of approximately 31 degrees (standard deviation), a finding which was consistent for two of the three animals. In the third animal there was an insufficient number of strongly-driven neurons with identical polarities (ON- versus OFF-center) to allow for a reasonable reconstruction of a population with more than 2 or 3 neurons. With so few neurons, the population displayed more and more properties of the response of a single neuron as opposed to a rough average of multiple neurons and the overall orientation tuning decreased as the population approached the orientation-agnostic response properties of a single input neuron. To determine the generality of our findings here, we utilized other metrics from previously published studies, with a focus on the reliability method used in [25] which is easily adaptable to population data. We found that qualitatively the results were similar to our own findings; just as jitter decreases in our sample population at 90 and 270 degrees (Figure 2C) the reliability across all the neurons in the population is significantly higher at 90 and 270 degrees. We thus expect that the synchronization observed across all neurons in the population is not affected by the metric chosen to measure it.
By construction, the degree of synchrony across the population of neurons in Figure 1D is a function of the orientation of the drifting gratings and across neuron variability in spiking, independent from geometry. The across neuron variability in timing thus set the lower bound of timing jitter in Figure 2C. To more fully explore the role of synchrony in shaping the feature selectivity in the downstream cortical response, we effectively replaced the across-neuron variability in spike timing with variability under our control. Specifically, we utilized a single trial spike-train for a template neuron and introduced the latency associated with the translation of the receptive field as in Figure 1D, but subsequently added variability to each spike time in the form of a Gaussian random variable with zero mean and variance . So long as the population firing rate reaches a particular minimum mean level it does not matter which template neuron is chosen to provide the spike train; we found that nearly all neurons from all three animals provided consistent simulations of cortical activity. Using a single trial has the effect of removing the effects of variable spike count across trials for a particular neuron in addition to providing the exact control over the timing jitter. Of key importance is the value , which is the stimulus-dependent component of timing jitter (see Methods for expanded description). This value is related to but not equal to the timing value measured from the full populations; represents the underlying stimulus-based modulations to synchrony that give rise to the full timing jitter relationship shown in Figure 2C. This timing variability quantity was parameterized as a Gaussian function of and was manually tuned to reproduce the population timing variability curve in Figure 2C. From here on out, when we refer to “minimum timing jitter” we are referring to the minimum value of that occurs at the preferred orientation.
To determine how different levels of input synchrony affect the downstream cortical response and the corresponding feature selectivity, we simulated the cortical layer 4 neuron response to the drifting gratings at different orientations. The previously described populations were used as input to this model, modulating the minimum value of to cover a range of 6 to 40 ms of population timing jitter. To model the cortical neuron, we used a biophysically inspired integrate and fire model — illustrated in 3A — that generates a continuous membrane potential and corresponding firing activity, similar to that in [22] and [23] - see Methods. In brief, the model lumps all input spike times together in a common spike train, laying down a superimposed EPSC for each input spike (all of which thus have equal weighting). This model is represented by the differential equationwith a fixed parameter set to determine the point by point membrane potential and with a fixed time step of 0.05 ms. Membrane potential traces show a clear stimulus-driven modulation [26]–[28] that increases in amplitude towards the population's preferred orientation when averaged over 1000 trials, as shown in Figure 3B. Single trial responses, with the exception of the nonphysiological mechanics of the hard reset, match typical recordings from cortical neurons using examples from Carandini & Ferster [29] as a primary source for comparison. Further, the tuning properties (firing rate and tuning half-width at half-height) match reported values, as will be shown later. The reset mechanics did not adversely affect the accuracy of the results as the spiking statistics and tuning curves were consistent with experimental observations. Cortical spike counts, as shown in Figure 3C rastergrams, increased dramatically as the stimulus approached the preferred orientation, and the underlying stimulus driven events became very clear. Again, these spike count rastergrams are representative of what would be expected from cortical neurons, although this is easier to see in the cortical tuning curves.
By construction of the thalamic input, the model generated cortical responses that exhibited orientation selectivity. Although the original experimental data was collected only for 8 grating orientations, the parameterized construction described in Figure 2 allowed simulation at an arbitrarily fine grain (chosen to be at 1 degree increments here). The resulting mean cortical firing rate across all orientations for a minimum jitter of 6 ms is shown in Figure 4A, which is stereotypical of recorded responses of neurons in the primary visual cortex [29], with higher firing rates possible when using different neurons for thalamic spike times. The cortical firing rate as a function of stimulus orientation was fit with a local Gaussian over a 180 degree span, as shown with the dashed curve. The parametric fits for each of a range of minimum jitter cases are shown in Figure 4B. The colors indicate decreasing levels of synchrony with dark red representing high synchrony (6 ms of jitter) and dark blue representing low synchrony (40 ms of jitter). The overall magnitude of the cortical response decreased with increasing amounts of jitter, as reflected in the overall amplitude of the tuning curves. The sharpness of orientation tuning is quantified though the half-width at half-height (HWHH) of the tuning curve [29], [30]. Consistent with reported values for firing rate, the HWHH tuning width for firing rate was approximately 15 to 16 degrees and was relatively insensitive to the LGN input synchrony (Figure 4C) up until 35 ms of input jitter at which point the tuning width increases by approximately 1.5 degrees. These values are on the lower end of expected tuning widths [9], [29], [30]. Carandini & Ferster [29] noted that due to experimental limitations they cannot discriminate half-widths less than 17 degrees, a value that they find for almost all recorded neurons. On the other hand different studies [31], [32] have reported tuning widths with significant numbers of neurons with small (10–15 degree) tuning widths. Note that the primary results of the analysis were relatively invariant to the actual tuning width, as we will demonstrate later.
The tuning curve is illustrative to see how well a particular stimulus orientation drives a cortical neuron but by itself it does not convey any context as to how well the cortical neuron transmits information about the stimulus. Synchrony clearly modulates the overall amplitude of this tuning but it is unclear how it modulates the transmission of the underlying stimulus information. The ability of an ideal observer of neural activity to extract meaningful information regarding the features of a visual stimulus depends not only on the shape of the tuning curve, but also on the variability of the cortical response and how this variability changes with the stimulus feature. The statistics of the cortical response are summarized in Figure 5. In Figure 5A, the underlying relationship between the mean and variance of the cortical spike count for all stimulus orientations (each individual dot) is illustrated. The relationship clearly demonstrates an increase of spike count variance relative to spike count mean with a slope of approximately 3, which begins to drop when the input is relatively synchronous (6–10 ms of jitter). The variance begins to drop at extreme levels of synchrony as the decreased amount of added timing variance approaches the size of the integration window of the model, and higher synchrony values effectively make the spike count more deterministic. With respect to the relationship between the mean and variance of the cortical response, experimental results have been variable, exhibiting both sub- and supra-linear variability [33]–[42]. So while the orientation tuning width was relatively invariant to the level of synchrony, as shown in Figure 4C, the increased level of synchrony was accompanied by an increased mean firing rate, and thus an increased variance, the effects of which are not immediately obvious from the perspective of an ideal observer. Figure 5B shows the corresponding spike count distributions for the tuning curves in Figure 4B, for the preferred stimulus orientation (90 degrees). The spike count distribution changed dramatically as input synchrony decreased, with asynchronous inputs pinning spike count distributions at the origin and restricting the discriminability at adjacent distributions, a problem not encountered for highly synchronous inputs. From these results we might qualitatively expect that increasing synchrony would lead to increases in information because synchronization appears to give response distributions a greater range over which to vary with stimulus orientation.
Results from both the mean-variance relationship and the per-synchrony peak spike count response distributions thus lead to conflicting expectations on what level of input population synchrony would drive the maximum amount of information about stimulus orientation. In order to solve this inconsistency we must implement a metric that describes concisely how discriminable different stimulus orientations are and determine the effect input synchrony has cortical information transfer. Fisher information quantifies the degree to which response distributions are discriminable, and thus, provide unambiguous information about stimulus features captured in the response distributions. The simplest understanding of Fisher information in the context of the problem here is that it represents the derivative of the tuning curve with respect to the stimulus orientation; regardless of the underlying firing statistics, the peak Fisher information will occur near orientations where the derivative of the tuning curve is highest.
We use the peak amount of information across all stimulus orientations for each level of input synchrony as the metric for the capacity for any particular neuron to inform estimations about the stimulus orientation. By itself the absolute amount of information is an unintuitive quantity. With the goal of determining how synchrony changes the capabilities of cortical neurons to decode specific stimulus features, it is more natural to look at properties of the feature estimator. The inverse of Fisher information is the Cramér-Rao lower bound, a theoretical lower bound on the variance of a maximum-likelihood estimator; decreases in this quantity yield estimates that are more precise and have more confidence. Under the assumption that the stimulus orientation estimator is unbiased, lower estimator variance guarantees lower estimator error. Since we could directly calculate Fisher information in our model, we could also determine what this lower bound was, as shown in Figure 6A. The estimator standard deviation decreased nonlinearly with increasing synchrony, covering a range of relatively precise estimates to very imprecise estimates with a notable saturation at around 20 ms of jitter; synchrony higher than this does not yield rapid gains while decreases in synchrony rapidly decrease the estimator precision. As the Fisher information is directly related to the local slope of the tuning curve this qualitative observation was unaffected, in a relative sense, by the discretization of the tuning curve. The raw information decreased approximately linearly with increasing minimum jitter as shown in Figure 6B (error bars are ±1 S.D.). However, as we will show the degree to which this is not linear has important implications for the efficiency of information transmission by the cortical neuron.
From these results, we naively assumed that a strategy which absolutely increased synchrony would always be best as it would always result in increasing stimulus information. As has been noted in other models which bear some similarities to our own [43], there is a metabolic cost to increasing firing rate which can affect the efficiency of some information representations relative to others. In this case, as shown in Figure 6C, when we normalize the absolute amount of information by the number of cortical spikes, it becomes clear that the peak in transmission efficiency occurred at around 15 ms of thalamic jitter, and a quadratic fit had a peak at 16 ms with a clear decrease in information efficiency away from this peak. In previous studies [10]–[12] we identified that pairwise LGN synchrony in response to natural scenes tends to be from 10 to 20 ms as measured by our scale. As noted, this result was consistent across all simultaneously recorded neurons when these neurons were used as sources for single-trial spike times. A few neurons maintained this quadratic relationship between information transmission efficiency and input synchronization at a peak efficiency closer to 25 ms of timing jitter, slightly lower than expected. These results indicate that populations in the LGN are uniquely arranged to be effectively synchronized by a preferred orientation. This synchronization allows information transmission to be more efficient without sacrificing precision in estimating orientation.
The results presented so far have demonstrated that information efficiency saturates at levels of minimum timing jitter between 10 and 20 ms, without addressing the effect of tuning width. It is clear from existing literature that there is a wide range of tuning widths that are typically measured in neurons in visual cortex [9], [29]–[32] and these changes are reflected in the width of and thus the width of the tuning curve. To investigate the effect of changes in just tuning width we modulated both the minimum timing jitter as well as the tuning width, with the results shown in Figure 7. From 4.1 to 30.8 degrees (HWHH; maroon to light blue dots in Figure 7), which covers the rough range one could expect tuning width to vary, it is clear that the normalized information per spike (IPS) has approximately the same pattern regardless of tuning width. We show normalized information per spike because Fisher information is directly related to the slope of the curve, higher slopes monotonically lead to higher absolute levels of information and as such 4.2 degree and 30.8 degree tuning widths have an order of magnitude difference in their absolute amount of information. The relationship between tuning width and information efficiency is made clearer in the breakouts in Figure 7B for each individual tuning width; with the exception of extremely narrow tuning widths, as the tuning width increases the optimal level of minimum jitter increases but still stays in the range of 10–20 ms. Narrow tuning curves fail to saturate information per spike because very narrow tuning curves effectively contain information about a very small range of orientations and the amount of information is directly related to the diference between baseline and peak firing rates. As an example consider a tuning curve that goes from baseline firing rate to peak firing rate in the span of 2 or 3 degrees (a very narrow tuning curve). In this case higher peak firing rates have a very pronounced affect on the overall amount of information. Since lower jitter always provides higher peak firing rates, narrower tuning curves are always most efficacious at extremely low amounts of jitter. We thus see that the results are valid for a range of primary visual cortex neurons so long as they have tuning widths that are within physiologically measured ranges.
In this work we investigated the role of stimulus-driven synchrony in thalamic populations in the emergence of feature selectivity in primary visual cortex. The complete understanding of this role requires observation of entire thalamic sub-populations which are convergent onto single cortical neurons. Since these populations are too large to record electrophysiologically using current experimental methodologies, we synthesized representative populations from experimental data by randomly choosing recorded trials of neurons from which we could record, when obeying anatomical rules of thalamocortical connectivity [24] (also see below). These populations had an amount of stimulus-driven synchronization that was a direct function of the orientation of a drifting grating stimulus. These synthesized populations allowed us to systematically modulate the underlying spike timing synchrony to investigate the way in which different levels of synchronization affect information transmission. Through a biophysically inspired integrate and fire model that simulates cortical responses, we estimated the resultant cortical orientation selectivity and the corresponding information conveyed about visual stimulus orientation by the cortical response. Ultimately we found that the level of synchronization of the input population had a nonlinear effect on the resulting information contained in the cortical response; higher levels of synchrony led to higher levels of information, but at the expense of a nonlinear increase in firing rate. When taking into account the potential cost of increased firing rate, we found that the most efficient transmission of information was at a level of thalamic synchrony in the range of 10 to 20 ms.
It is important to note that the synchronization of neurons has been widely studied in a number of different contexts. Notably, synchronization of neurons across cortical columns has been previously reported in the visual cortex, proposed as a means to form relationships across regions of the visual field [44]. Additionally, in the context of convergence and divergence of retinal afferents projecting to the LGN, precise correlations have been observed across geniculate neurons that were present in the absence of stimulus driven correlations, and were attributed to the projections of common retinal ganglion cell inputs [13]. In contrast, the current study (and previous studies from our group [11], [23]) specifically examines the role of stimulus driven synchronization/correlation of neuronal firing in the visual thalamus. Our previous investigations have shown that many neurons in the LGN do not exhibit appreciable noise correlations [11]. The focus here is thus on the relationship between the visual input and the resultant synchronization of firing activity across geniculate ensembles, a requisite for robust activation of the downstream cortical neurons to which they project. In the most general case, however, as described in Gray et al. [44], the propagation of neuronal signals would involve a combination or interaction between the synchronization due to ongoing spontaneous activity and the stimulus-driven synchronization due to coordinated activation of neurons sharing the same topology and feature selectivity. Such a “from-any-source” view of synchronization carries with it the possibility that neurons with receptive fields from disparate regions of the visual field could be synchronized by spatially correlated stimuli. For example two very spatially distant LGN neurons could be simultaneously activated by either two unrelated objects or one very long bar of light; synchronization due to these origins are not considered in this model. It is important to note that we explicitly consider only recordings from spatially localized populations, as widely-spaced LGN units do not converge at the same cortical target.
The emergence of orientation selectivity in primary visual cortex is perhaps the most well-studied example of cortical computation to date. As a result, there have been a large number of modeling studies seeking to capture the mechanistic explanation for the primary observation of orientation selectivity, and also to capture a number of related, and more complex functional properties (e.g. contrast invariant orientation tuning, cross-orientation suppression, etc.). Given that there is little if any dispute as to the role of direct feed-forward geniculate input to cortical layer 4 in establishing the basic orientation preference for cortical neurons, models of orientation selectivity have invariably been constructed around a backbone of thalamic input. Although the model from Ringach introduced structured synaptic weightings and connectivity probabilities of thalamic inputs to cortex as a key model element [5], the majority of other models assume relatively simple feedforward excitation structure and differ primarily in the relative strengths of the feedforward or intracortical inhibition [3], [4], [26]–[28], [45], [46]. A specific limitation of most of these previous models is that they explicitly do not directly involve electrophysiological data as thalamic input. For example, one class of models use simulations of thalamic or retinal responses based on the stereotypical difference-of-Gaussians representation of center-surround receptive fields [3]–[5], [45], [46], while others rely on assumed or derived cortical conductances or membrane potential but not on actual thalamic input [26]–[28]. The large majority of previously published models also assume that sinusoidal inputs (i.e. drifting gratings) elicit sinusoidal thalamic responses and that the cortical membrane potential itself is perfectly sinusoidally modulated (as in [9] or [26], [45]). Dating back to the early 1980s there was the observation that drifting sinusoidal gratings produced asymmetric LGN response PSTHs (i.e. a sharp peak at the onset of the stimulus followed by a long tail of decaying response) [47]–[49] and more recently we have directly analyzed the effects of this synchrony in the context of cortical orientation and direction selectivity [23]. We assert that the precise timing and stimulus-driven synchronization of thalamic inputs serves a prominent role in the thalamocortical circuit and in the emergence of cortical feature selectivity.
It is important to note that most, if not all, existing models designed to capture the mechanism behind cortical orientation selectivity rely on spatial arrangements of projecting thalamic inputs that in some cases exceed those observed experimentally [24]. More specifically, the relevant measure for thalamic input is the aspect ratio of the scatter of thalamic receptive fields that form the input to a single cortical layer 4 neuron. Recently, Jin et al. experimentally observed thalamic clusters and showed that the thalamic input to cortical orientation columns has receptive fields that are highly overlapped [24]. Because the scatter of the thalamic receptive fields covers 2.5 receptive field centers in visual space, the average layer 4 cortical neuron should have a maximum aspect ratio of 2.5∶1. The thalamocortical model from Somers et al. was built on an aspect ratio of 3∶1 [3], whereas the model from McLaughlin et al. was built on an aspect ratio of 4∶1 [4]. Similarly large aspect ratios are apparent from the Kayser et. al. model and Finn et. al. models, with ratios approximately 6∶1 and 2.5∶1 respectively [26],[46]. It is clearly the case that inhibitory mechanisms play a significant role in the shaping of the cortical feature selectivity [2], and would only serve to further refine the selectivity established by the direct feedforward thalamic input shown here. Many of the above-mentioned models differ from our presentation here in that they include OFF-center sub-populations in the thalamic population, most commonly offset from the ON-center population as would be implied by the common Gabor-type simple cell receptive field. To keep the model relatively straightforward and simple, we have chosen to focus on just ON-center populations.
The majority of existing models were optimized to explain extra-classical effects of cortical receptive fields with a particular focus on the contrast invariance of cortical tuning width and as such constructed mechanisms specific to this issue. Specifically, it has been widely observed that although peak cortical firing rates are strongly dependent upon stimulus contrast, cortical orientation tuning is largely invariant to stimulus contrast (for review, see [2]). This observation called into question the purely feedforward model of cortical orientation selectivity [2]. Subsequent models augmented the feedforward thalamic input with inhibitory feedforward connections [26] or cortico-cortico inhibition [46] or some combination [2], [3]. We have previously shown that thalamic synchrony is largely unaffected by stimulus contrast [11], and the cortical tuning based on thalamic synchrony is also contrast invariant. The model we have proposed here thus potentially demonstrates a completely feed-forward explanation for contrast invariance. For a fixed minimum jitter amount, as the underlying LGN firing rates across the entire population are modulated by changes in the stimulus contrast, the peak induced firing in the cortical neuron rises and falls. Since the changes in LGN firing are correlated across the LGN population, the synchrony across such a population (with particularly focus on the relationship between stimulus orientation and the synchrony) remains unchanged as a function of stimulus contrast. As demonstrated in Figure 4B for the span of biophysical levels of preferred orientation population synchrony (∼5 to 20 ms), the tuning width of the cortical neuron does not change, indicating that changes in the degree of underlying synchrony do not change the tuning properties. Although the results are not presented here directly, the combination of past and present results suggest that changes in the LGN population response (i.e. the population becomes less active in general) lead to a decreased or increased peak cortical response but the tuning curve widths will be invariant to stimulus contrast.
We used Fisher information as a measure of the efficacy of cortical neurons in representing stimulus features (orientation) in response to changes in the synchrony of an input population. Specifically, we used the peak Fisher information irrespective of the orientation at which the peak occurs. Contrary to previous investigations [50]–[52] in which the absolute value of the Fisher information was used as an important measure of the performance of neural populations, here we sought to capture the relative effects of varying degrees of thalamic synchrony on the information conveyed by a single recipient cortical neuron target. In this case, we assumed that the Cramér-Rao lower bound need not be met and that whatever bias causing deviations from the lower bound is consistent across all simulation conditions. We ensure this by using the same input data and model structure for all conditions so that we can compare relative levels of information across different synchrony conditions for a single neuron. Although this is a simplification of the true amount of information (and indeed no single neuron will saturate this lower bound), in either case the absolute information was consistent with previous studies utilizing experimental cortical data. Yarrow et. al. [52] computed Fisher information for both real and simulated neural populations and found an information level which was approximately consistent with the findings presented here (see their Figure 4 as well as [51] Figure 3, with axes in [52] helping in the conversion from SSI bits to Fisher Information in units of ). This assumption ultimately only affects the reporting of estimator standard deviation (as in Figure 6A) which was not the primary result of the work.
It is also important to note that the application of Fisher Information to cortical tuning curves has deeper roots in estimating cortical population response information transmission. Past work [53]–[57] has in general used constructions where a collection of identical cortical neurons have preferred orientations that uniformly span the orientation spectrum (0 to 360 degrees). In this study we considered only a single neuron in the population. We claim, though, that results which demonstrate information in a single neuron at all stimulus orientations are fundamentally identical to results which demonstrate information in a population at a single orientation. As long as we assume every neuron in the cortical population is conditionally independent, for the questions we ask these two formulations are fundamentally interchangeable. As identified in [54] under the assumption that each cortical neuron in this population is independent, then at every stimulus orientation the overall Fisher information isFurther, in the case that every neuron in the population is also assumed to be identical in response properties, then we can modify the above to read (for any choice of )It is clear though that not all cortical tuning curves are identical and the absolute amount of information is strongly negatively correlated with tuning width. Using this fact as inspiration, we show in Figure 7 that the optimally efficient level of input timing jitter is widely insensitive to the tuning width of the cortical neuron. In this case, even if a cortical population is composed of non-identical independent neurons, each neuron, as well as the population as a whole, will be optimally efficient as long as the thalamic input is synchronous to the 10–20 ms level (thus implying we need no longer assume neurons within the population have identical, but shifted, tuning curves). If we further consider the effects of correlated variability, as in [55], then we can no longer assume the units are independent. Regardless of whether the correlated variability increases or decreases the absolute amount of information (and neither is guaranteed), correlated variability would raise or lower the response rate of the individual neurons in a coordinated manner. Since again our metric is one of relative comparisons, the results presented here are expected to be invariant to correlated variability in the sense that the efficiency of any single neuron may decrease, the peak efficiency will still occur between 10–20 ms (which would still be true for all neurons in the cortical population). Thus our findings directly translate to cortical populations regardless of the independence and homogeneity of tuning properties of the component neurons.
In previous studies of timing precision of individual thalamic neurons [10] and across thalamic pairs [11] in response to natural scenes, we have reported characteristic timescales on the order of 10–20 ms. In these previous studies, measures were taken across long segments of natural scene movies, representing the aggregate of instantaneous firing events whose timing precision clearly varies on an event-by-event basis [12], [58]. The instantaneous synchronization of firing activity across a sub-population of neurons in the context of natural scenes is undoubtedly a complex function of the local properties of the scene, including spatial frequency, temporal frequency, and orientation of the local spatial structure. It is thus the case that the 10–20 ms average timescale reflects a distribution of synchronous events, spanning from synchrony on just a few milliseconds to more asynchronous firing over a timescale of 10€s of milliseconds, unlikely to drive the cortical target. Here, we report that in the context of the modulation of thalamic synchrony through visual stimulus orientation with drifting sinusoidal gratings, the most efficient level of thalamic synchrony in conveying relevant information to cortex is in the 10–20 ms range. This means that, on average, amongst natural scenes and all their various features, the thalamic neural response is tuned to maximize the efficiency of information transfer to the cortex (similar to [22]). As we have investigated only the effects of orientation changes on synchronization and feature selectivity, we expect that this result implies that information efficiency will be similarly optimized for other visual features such as spatial and temporal frequency. Furthermore, it is possible that synchronization optimizes information transmission in entirely different sensory systems, given previous findings in the somatosensory system [59].
Surgical and experimental procedures were performed in accordance with United States Department of Agriculture guidelines and were approved by the Institutional Animal Care and Use Committee at the State University of New York, State College of Optometry.
The experimental data collection has been previously described [23]. Briefly, single-cell activity was recorded extracellularly in the lateral geniculate nucleus (LGN) of anesthetized and paralyzed male cats, with a total of three animals. As described in [60], cats were initially anesthetized with ketamine (10 mg kg−1 intramuscular) and acepromazine (0.2 mg/kg), followed by propofol (3 mg kg1 before recording and 6 mg kg−1 h−1 during recording; supplemented as needed). A craniotomy and duratomy were performed to introduce recording electrodes into the LGN (anterior, 5.5; lateral, 10.5). Animals were paralyzed with vecuronium bromide (0.3 mg kg−1 h−1 intravenous) to minimize eye movements, and were artificially ventilated. Using a seven-electrode matrix, layer A geniculate cells were recorded extracellularly. The multielectrode array was inserted into the brain to record from iso-retinotopic lines across the depth of the LGN, using an angle of 25–30 degrees antero-posterior, 2–5 degrees lateral-central. To a multielectrode array (with inter-electrode separation of 254 µm) we attached a glass guide tube with an inner diameter of 300 µm. As the elevation axis is better represented in LGN than the azimuth axis, some of the populations of LGN receptive fields showed greater lateral than vertical scatter in the visual field [61]. Layer A of LGN was physiologically identified by performing several electrode penetrations to map the retinotopic organization of the LGN and center the multielectrode array at the retinotopic location selected for this study (5–10 degrees eccentricity). While recording, the RASPUTIN software (Plexon, Dallas, TX) was used to capture voltage signals after being amplified and filtered. We isolated single units by independently moving each electrode and the resulting units were spike-sorted online and verified offline using a commercially available algorithm (Plexon, Dallas, TX). Cells were eliminated from this study if they did not have at least 1 Hz mean firing rates in response to all stimulus conditions. Cells were classified as ON or OFF according to the polarity of the receptive field estimate.
For each cell, visual stimulation consisted of multiple repetitions of a drifting sinusoidal grating at 0.5 cycles/degree, at either 100% or 64% contrast. The direction of the drifting grating was varied. The orientation of a particular drifting grating was one of eight possible values: 0, 45, 90, 135, 180, 225, 270, 315 degrees. The convention was that a vertically oriented grating drifting rightward was referred to as 0 degrees, a horizontally oriented grating drifting downward was referred to as 90, and so on. The temporal frequency for all datasets was 5 Hz or 4 Hz. The spatial resolution for the drifting gratings was 0.0281 degrees per pixel. All stimuli were presented at a 120 Hz monitor refresh rate.
Biophysiological levels of LGN population synchrony were measured from multiple sets of simultaneous electrophysiological recordings (between 5 and 7 neurons were recorded simultaneously). A cortical neuron is thought to receive approximately 30 LGN inputs [8] but these neurons are substantially more densely arrayed than we can reasonably hope to record with penetrating electrodes. Population response estimates were achieved by expanding the simultaneous recorded neurons into a population of 30 neurons by replicating the recorded responses and then shifting to a new visual location, restricted within the visual space bounded by the original receptive field locations. This restriction resulted in a population that has a receptive field center diameter distribution that is consistent with [24] (see Figure 1D). To create the population random shifts were allowed in both the vertical and horizontal directions (i.e. the major and minor axes of the population) but the restrictions placed by the original population layout often required greater shifts along one or the other axis. For the example in Figure 1C, the shift restrictions resulted in a visual space coverage of approximately 5 degrees (horizontal) by 2 degrees (vertical). Shifting responses required knowledge of the timing difference in excitation between the old and the new location, defined as the shift latency. The replicated input spike trains occurred in response to sinusoidal gratings and, due to the regularity in the stimulus, the shift latency was relatively easy to calculate. This shift latency was estimated simply by measuring the timing latency between the maximum excitation at the centroids of the receptive fields at both the original location and the shifted locationwhere is the center to center separation of the original and shifted locations, represent the spatial (cycle/deg) and temporal (Hz) frequencies (fixed) of the stimulus itself, and is the angle between the axis connecting the two receptive fields and a line from the shifted location perpendicular to the oriented stimulus bar. A graphical representation of this is in Figure 1D. Each newly created neuron is assigned a random trial from all recorded trials of the original neuron and the shift latency value is added to all spike times within that chosen trial. For the representation of this process in Figure 1E each neuron received a trial from the appropriate stimulus orientation. As the overarching cortical model, though, expands to a much larger set of orientations than originally recorded from, for consistency each newly created neuron was assigned a trial from the recordings performed with a stimulus at a 0 degree orientation. This allows us to preserve the baseline across-neuron timing changes, while capturing the stimulus-driven timing modulations with our parameter, discussed below.
The model was constructed such that all input synapses to the cortical neuron have equal strength and no particular synaptic location (i.e. along the dendrite or at the soma), and accordingly the source of the spikes from within the LGN population has no effect on the actual model output. Since this is the case, we can estimate the input population auto-correlation by collapsing all LGN spike times into a single vector. The auto-correlation is then calculated by subtracting each spike time from all other spike times and calculating the histogram of these pair-wise interspike intervals. Synchronous populations will have a much higher proportion of small intervals (neglecting stimulus periodicity) than asynchronous populations. The auto-correlations are also appropriately normalized to be between 0 and 1. To smooth the auto-correlation and eliminate correlations caused by the periodicity of the input, a Gaussian was fit to the central 200 ms lags in the correlation. We use timing jitter as a metric of synchrony, which is determined by normalizing the Gaussian fit and locating the lag at which this curve is equal to . To relate this number to the PSTH timing jitter (i.e. combined population timing jitter) we must divide by two (see Supplemental in [10] for a complete description). In brief, we define a value which is the “response timescale”. This value is equal to the latency at which the auto-correlation equals . By construction this has the relationship that , where is the timing jitter in the PSTH, our value of interest. This process was performed for all stimulus orientations (in order to maintain phase and timing differences that arise from differences in neuron properties and not just spatial relationships) to describe timing jitter as a function of stimulus orientation. This function was calculated multiple times for different randomly generated populations to estimate the variance that is created by choosing either different visual locations for the component neurons or choosing different recorded trials to represent the neurons in the population.
The observed timing variability in spike times across the population is composed of two aspects; intrinsic neural variability and variability caused by the interaction between the stimulus and the population organization. Our model captured the intrinsic variability by using spike times that were recorded in vivo. On the other hand, while the grating stimulus always evokes firing in the thalamic neurons the timing differences in spike times from neuron to neuron will vary according to orientation of these gratings and the arrangement of the population itself. We capture this stimulus-evoked timing variability in a parameter . This parameter, as a function of stimulus orientation, was manually calibrated such that when used with recorded data we could reconstruct the exact plot shown in Figure 2C. This procedure allows us to capture both the intrinsic and stimulus-evoked sources of spike timing variability even at orientations for which we were not able to collect data.
All simulations and computations were performed in the Matlab programming language (Mathworks, Inc., Natick, MA) using a 64-node grid computer. The integrate and fire model [62], illustrated in Figure 3A, takes spiking activity from the simulated LGN population as input and outputs cortical membrane potential and the associated cortical spike times. It was assumed that each synapse has equal strength. To create the synaptic input current, an exponentially decaying EPSC of defined amplitude () and time constant () was generated for all spike times in the input LGN population. The EPSCs were summed linearly across all LGN inputs to create a single current input at every simulation time point. The cortical membrane potential was modeled with the following first-order differential equation:where is the membrane potential, is the total synaptic current, is the membrane resistance (), is the resting potential (−70 mV), and is the membrane time constant (2 ms). The integration was performed using the forward euler method with a step size of 0.05 ms; since the step size is significantly smaller than any other temporal dynamics or spike timing precision use of a simple euler method is sufficient. When exceeds the threshold membrane potential (), a cortical spike is generated by setting the instantaneous potential to 0 mV followed by a 3 ms refractory period at the reset potential of −65 mV. These values are similar to those we have used previously for similar models [22], [23]. An analysis was performed to determine the approximate sensitivity of the model to each of the above indicated parameters. In general the model is sensitive to parameters which modulate the strength (or efficacy) of input spikes relative to the generated EPSC. Thus the model is sensitive to the EPSC amplitude (; effective values 0.05 to 0.1 nA within acceptable ranges) and the EPSC decay (; effective values 2 to 5 ms) while being robust to changes in threshold and reset potentials (). Sensitivity manifests itself as a change between one of three states; impoverished cortical firing, sufficient cortical firing, and strong cortical firing. Impoverished firing results in a peak information per spike (see Figure 6C) at very low jitter values (as this maximizes the chance to get any spikes) and strong firing demonstrates no discernible peak information per spike for any particular jitter value (as it results in very wide tuning curves).
Taking the perspective of an ideal observer, we approximated the capability of the observer to discriminate between visual stimulus orientations based on cortical activity alone. More specifically, the Fisher information [54]–[57] at each orientation captures the discriminability between where the expectation is taken with respect to . In the case that the probability is zero, we set . We calculated the derivative numerically using increments of 1 degree which was the resolution at which the simulations were performed. To reduce the results of this calculation to a single descriptive value, we report the estimator minimum standard deviation, which is related to the Fisher information through the Cramér-Rao lower bound (assuming the estimator is unbiased):As a metric of efficiency with which the cortical output conveys information about the stimulus, we divide the peak output information by the peak spike count with the goal of identifying how much each individual spike contributes to the overall information; higher values indicate each spike is more efficient at conveying information about stimulus features. This established a penalty for higher firing rates, realizing that there is a metabolic cost to generating action potentials.
Response distributions of the cortical firing rate were estimated based on the simulated data, in order to calculate the Fisher Information. The firing rate varied as a function of and the distributions are given by . The data were explicitly fit to a Poisson distribution, consistent with previous findings [33]–[42] as well as explicitly verified for appropriate fitting against our own data:To generate an accurate estimation of the response distributions at a minimum 250 simulation trials were run, with more trials providing no significant change in the estimated distributions. Note that the distributions change as a function of stimulus orientation, as indicated by . Further, in order to create a smooth description of Fisher information it was necessary that the response distributions be smooth functions of , as even minor fluctuations in the parameter get magnified by differentiation and squaring. To alleviate this, was smoothed with a Gaussian fit which was empirically verified to describe well.
|
10.1371/journal.ppat.1007268 | Inhibition of p38 MAPK in combination with ART reduces SIV-induced immune activation and provides additional protection from immune system deterioration | Differences in immune activation were identified as the most significant difference between AIDS-susceptible and resistant species. p38 MAPK, activated in HIV infection, is key to induction of interferon-stimulated genes and cytokine-mediated inflammation and is associated with some of the pathology produced by HIV or SIV infection in AIDS-susceptible primates. As small molecule p38 MAPK inhibitors are being tested in human trials for inflammatory diseases, we evaluated the effects of treating SIV-infected macaques with the p38 MAPK inhibitor PH-797804 in conjunction with ART. PH-797804 had no side effects, did not impact negatively the antiviral immune response and, used alone, had no significant effect on levels of immune activation and did not reduced the viremia. When administered with ART, it significantly reduced numerous immune activation markers compared to ART alone. CD38+/HLA-DR+ and Ki-67+ T-cell percentages in blood, lymph node and rectal CD4+ and CD8+ T cells, PD-1 expression in CD8+ T cells and plasma levels of IFNα, IFNγ, TNFα, IL-6, IP-10, sCD163 and C-reactive protein were all significantly reduced. Significant preservation of CD4+, CD4+ central memory, CD4+/IL-22+ and CD4+/IL-17+ T-cell percentages and improvement of Th17/Treg ratio in blood and rectal mucosa were also observed. Importantly, the addition of PH-797804 to ART initiated during chronic SIV infection reduced immune activation and restored immune system parameters to the levels observed when ART was initiated on week 1 after infection. After ART interruption, viremia rebounded in a similar fashion in all groups, regardless of when ART was initiated. We concluded that the inhibitor PH-797804 significantly reduced, even if did not normalized, the immune activation parameters evaluated during ART treatment, improved preservation of critical populations of the immune system targeted by SIV, and increased the efficacy of ART treatment initiated in chronic infection to levels similar to those observed when initiated in acute infection but did not affect positively or negatively viral reservoirs.
| The hallmark of Human Immunodeficiency Virus and Simian Immunodeficiency Virus infection in disease-susceptible species is the progressive decline of the CD4+ T cell population and heightened immune activation, which by itself can contribute to CD4+ T-cell death. The cellular pathway regulated by p38 MAPK, which is activated in HIV and SIV infection, can contribute significantly to immune activation. We tested in SIV-infected macaques a p38 MAPK inhibitor in combination with anti-retroviral therapy. This drug is already being evaluated in humans for treatment of immune activation associated with other diseases. We found that, when combined with antiretroviral therapy, the inhibitor PH-797804 significantly reduced a few parameters of SIV-induced immune activation and improved preservation of critical populations of the immune system targeted by SIV, but did not modulate viral reservoirs. Importantly, the addition of the inhibitor to anti-retroviral therapy during the chronic phase of the infection, which is the time when most HIV-infected individuals initiate treatment, permitted a more significant preservation of the immune system compared to antiretroviral therapy alone that was similar to that observed when anti-retroviral therapy was initiated in the acute phase of the infection, which rarely occurs in HIV infection.
| Differences in immune activation have been identified as the single most significant difference between AIDS-susceptible and resistant species [1–9]. Immune activation can be induced by a variety of mechanisms, including stimulation of innate and adaptive immune responses, production of a superantigen, and/or production of activating cytokines and chemokines. It is quite likely that more than one mechanism is occurring simultaneously during HIV infection. As immune activation can trigger T-cell apoptosis, the differential level of immune activation induced by HIV and SIV among the species could explain the more drastic depletion of CD4+ T cells that occurs in AIDS susceptible compared to AIDS-resistant species, as apoptosis occurs significantly less in species that do not develop AIDS [7, 10, 11]. Apoptosis and immune activation are substantially reduced in HIV-infected individuals that are long-term non-progressors [12, 13]. Residual chronic immune activation during ART is considered a contributor to the co-morbidity observed during treatment, mainly accelerated aging-related diseases, including renal dysfunction, atherosclerosis and hypertension, diabetes mellitus, respiratory diseases (e.g. chronic obstructive pulmonary diseases and pneumonia), and HIV-associated neurological disorders [14–17].
Inhibition of immune activation has been explored in a few studies that have targeted different molecules and pathways. A COX-2 inhibitor tested in 13 HIV-infected individuals appeared to reduce immune activation as indicated by reduction of PD-1 on CD8+ T cells, increasing numbers of T regulatory (Treg) cells, and improved recall responses to a T-cell dependent vaccine [18]. p38 MAPK is also involved in the activation of COX-2 and inhibition of p38 MAPK also resulted in the inhibition of COX-2 [19]. However, when tested in a randomized placebo-controlled trial, no significant immunological effects of the COX-2 inhibitor Etoricoxib were observed in ART-treated patients [20]. PD-1 blockade also reduced immune activation [21]. Interestingly, it is p38 MAPK that stimulates the transcription of PD-1 and induces additional molecules inhibitory of T-cell function [22, 23]. Anti-TNFα treatment during SIV infection, whose expression is dependent on p38 MAPK, also reduced immune activation [24]. Instead, blocking IFNα in pDC did not reduce immune activation in SIV infected macaques and the administration of an IFNα agonist in SIV-infected sooty mangabeys did not result in immune activation, making it somehow less likely that the direct activity of IFNα is the cause of immune activation in Rhesus macaques (RM) [25, 26]. Reduced inflammation was observed in SIV-infected RM when ART was combined with IL-21, which also impacted time to rebound, plasma viremia and cell-associated SIV DNA levels after ART interruption but the mechanism of this outcome was not fully understood [27].
Induction of interferon-stimulated genes (ISG) during a viral infection is a consequence of Toll-like receptor (TLR) activation [28]. This leads to increased transcription of IRF7 and triggering of IFNα production, activation of the kinase cascade, with up-regulation, among others, of p38 MAPK. p38 MAPK triggers a signaling pathway that leads to direct activation of transcription factors implicated in many of cellular process including inflammation, cell cycle, apoptosis, and immune response [29, 30]. The p38 MAPK pathway is critical for maintaining a sustained response to type I and type II IFNs, leading to the induction of transcription of ISGs via activation of signal transducer and activator of transcription (STAT) proteins [31–35]. In addition, p38 MAPK plays an important role in regulating IFN-independent transcription of some ISG after TLR7 triggering [36–39] and production of inflammatory cytokines such as IL-1 and tumor necrosis factor alpha (TNFα). Inhibition of the p38 MAPK pathway may form the basis of a new strategy for treatment of inflammatory diseases.
p38 MAPK plays crucial roles in various pathological processes associated with HIV infection, including macrophage activation, neurotoxicity and impairment of neurogenesis, and lymphocyte apoptosis [29, 30]. Increased, active p38 MAPK has been reported in brains of SIV-infected macaques with encephalitis [40]. Interestingly, p38 MAPK has also been implicated in the production and release of IP-10 in astrocytes exposed to HIV-1 and Tat [41–43]. HIV-1 and Tat were reported to activate p38 MAPK in infected or stimulated monocytes and macrophages [44]. We have shown that HIV and SIV Tat modulates primate antigen presenting cells (APC) and that at least a subset of the ISG are not equally affected by SIV infection in APC of AIDS resistant species [45–47]. We found that Tat associated with the MAP2K6, MAP2K3 and IRF7 promoters and that the association resulted increased activation of p38 MAPK and consequent induction of ISG [45–47]. This mechanism of p38 MAPK activation could further and independently chronically contribute to the activation of ISG that results from TLR activation. Collectively, these data indicate that p38 MAPK activation is an important, additional mediator of HIV-associated pathology. Evaluating in vivo the role played by p38 MAPK in HIV replication and immune activation by inhibiting its activity may provide the rationale for the use of p38 MAPK inhibitors in AIDS therapy, in association with ART or when ART is no longer an option.
A series of compounds targeting p38 MAPK were initially discovered and were followed by the development of more potent and specific inhibitors of this protein capable of inhibiting the production of inflammatory cytokines. Different p38 MAPK inhibitors have been shown efficacy in preclinical animal models of a variety of diseases [48–56]. Some of these inhibitors have advanced to clinical studies for rheumatoid arthritis, chronic obstructive pulmonary disease (COPD), post-herpetic neuralgia and neuropathic pain, and osteoarthritis. [54, 57–59]
The diarylpyridinone PH-797804 is a novel, ATP-competitive and reversible potent inhibitor of human p38 MAPK [53]. It specifically inhibits p38α with IC50 value of 26 nM and K(i) value of 5.8 nM and inhibits LPS induced TNFα and IL-1β production in monocytes in a concentration-dependent manner [50]. PH-797804 blocks RANKL and M-CSF induced osteoclast formation in primary rat bone marrow cells. When given orally, PH-797804 reduces TNFα levels in LPS-induced shock of Lewis rats and of cynomolgus monkeys [50]. In randomized, adaptive design, double-blind, placebo-controlled, parallel-group, multicenter trial demonstrated improvements over placebo in lung function parameters and dyspnea in patients with moderate to severe COPD [58, 60].
Here we show that inhibiting p38 MAPK in vivo can significantly impact SIV-mediated immune activation and protect immune cell populations that are negatively affected by the infection. However, the reduction of immune activation is not complete, it is unlikely to be fully controlled by acting on any single activation pathway, and possibly requires exploration of different inhibitor doses and intervention on multiple pathways linked to immune activation, including other kinases like JNK that are also affected by HIV and SIV.
The contribution of p38 MAPK in chronic immune activation during lentiviral infection was investigated in the SIV-infected RM animal model, treated with ART over the course of 60 weeks. RM were infected intravenously (i.v.) with 10 TCID50 SIVmac251 and divided in groups (n = 4, Groups 1 and 2; n = 6, groups 3–6), with some groups receiving ART alone and others ART combined with p38 MAPK inhibitor (Fig 1). ART consisted of two reverse transcriptase (RT) inhibitors, tenofovir (PMPA, 20 mg/kg, and emtricitabine (FTC, 30 mg/kg), and the integrase inhibitor dolutegravir (DTG, 2.5mg/kg) s.i.d, administered i.m. and was initiated in the acute phase of the infection, one week after SIV infection, in groups 5 and 6, or in the chronic phase, once set point was reached, six weeks post-infection, in Groups 3 and 4. PH-797804, 10 mg s.i.d, orally administered, was chosen among other similar compounds for the proposed studies for multiple reasons. It has been tested in humans and cynomolgus monkeys where it could effectively inhibit the acute inflammatory response that follows the administration of lipopolysaccharide (LPS), in particular production of TNFα and IL-6, and it reduced chronic inflammation and bone loss associated with arthritis in mice and rats [50]. In cynomolgus monkeys exposed to a single dose of LPS, PH-797804 dosed intragastrically at 0.001 mg/kg to 1 mg/kg, the dose of 0.1 mg/kg reduced TNFα plasma levels to 20% of the levels observed in animals treated with placebo, while the dose of 1 mg/kg reduced it to less than 10% [50]. This reduction was virtually identical to that observed for TNFα and IL-1β in humans after LPS challenge, where a reduction of 50% of the IL-6 levels was also observed [50]. As primary endpoints, we evaluated differences in expression of surface and intracellular molecules linked to immune activation and plasma levels of inflammatory cytokines. As secondary endpoints, we evaluated the effects that treatments had on viral loads, preservation of central memory (CM) and other CD4+ T cell subpopulations.
After SIVmac251 infection all animals experienced a rapid increase in viremia that peaked before initiation of ART for the groups when ART was initiated on week 6 and was below peak levels in the groups initiating ART one week after infection. ART was effective in suppressing the viral replication in all animals that received it (Fig 2A). The p38 inhibitor PH-797804 was well tolerated and no major side effects were noticed throughout the study. At the dose used in these animals (10 mg s.i.d, orally), plasma viral loads were comparable between groups receiving PH-797804 and those that did not, indicating that the p38 MAPK inhibitor did not affect virus replication, whether administered alone or in combination with ART, and regardless of when ART was initiated. To obtain a preliminary indication of PH-797804 efficacy in vivo, and considering that p38 MAPK does not directly affect ISG expression but does so via directly increasing the activity of the master transcription regulators of ISG expression IRF7 and pSTAT1, we evaluated the percentages of PBMC positive for accumulation of IRF7 and pSTAT1, and the cytokine IP-10, one of the ISG most significantly upregulated in HIV and SIV infections (Fig 2B–2D, S1A Fig). Using intracellular staining and flow cytometry, we found that the accumulation of these proteins was significantly reduced by the inhibitor when percentages from animals receiving ART alone were compared to those receiving ART and inhibitor for the two paired groups, whether initiating ART at week 1 or 6 (area under the curve from week 8 to 60, IRF7: p = 0.01, pSTAT1: p = 0.004, IP-10: p = 0.004 for groups initiating ART at week 1 and IRF7: p = 0.02, pSTAT1: p = 0.02, IP-10: p = 0.04, for groups initiating ART at week 6). This was true whether the analysis was done on total PBMC or CD3+ T-cell subpopulations (S1C Fig). Percentages of the CD3+ subpopulations were comparable at individual time points in paired groups (S1B Fig). Results were very similar when the same analyses were carried out in lymph node mononuclear cells (MNC) (S1D Fig). This result indicates that the selected PH-797804 dose could reduce expression of ISG transcriptional regulators and of the cytokine IP-10.
To exclude that the inhibitor treatment could reduce the effectiveness of the antiviral immune response, we evaluated the levels of antigen-specific cell mediated responses over the course of the treatment. We found that the number of SIV-specific CD4+ and CD8+ T cells were similar in the group pairs and proportionate to the viral loads present in the animals (Fig 3A), indicating that the p38 inhibitor treatment did not grossly altered the magnitude of the anti-viral immune response. We also evaluated whether inhibition of p38 MAPK could impact the expression of PD-1, checkpoint known to increase during infection because of persistent immune activation, resulting in inefficient CD8+ T-cell activity [61]. When investigated on week 60, before removal of ART and inhibitor treatment, we found that the frequency of CD8+ T cells expressing PD-1 was significantly lower in the groups that received the p38 inhibitor combined with ART compared to ART alone, whether ART treatment started on week 1 (p = 0.011) or 6 (p = 0.038) whereas it was comparable in CD4+ T cells (Fig 3B and 3C). We concluded that PH-797804 did not affect negatively the development of anti-SIV immune responses and reduced the expression of the checkpoint inhibitor PD-1 in CD8+ T cells, most likely indirectly via the overall impact on immune activation.
Chronic immune activation has been proposed to be a key determinant of AIDS pathogenesis. A variety of cell surface determinants expressed in the cell membrane are phenotypically associated with T-cell immune activation and provide useful marker to evaluate immune activation levels. In this study, we measured the surface expression of HLA-DR and CD38 and the DNA replication marker Ki-67, expressed intracellularly in PBMC and tissue MNC and these analyses are reported in Fig 4 and Fig 5 as group averages and in S2 Fig and S3 Fig for individual animals. Percentages of HLA-DR+/CD38+ cells were significantly lower in CD4+ and CD8+ T cells of the groups treated with ART plus p38 MAPK inhibitor compared to those treated with ART alone, whether treatment started at week 1 (p = 0.03, p = 0.009, for CD4 and CD8, respectively) or 6 post-infection (p = 0.002, p = 0.003, for CD4 and CD8, respectively), when the areas under the curve of the plotted parameters were compared in pair groups for the entire duration of the treatment (week 8 to week 60) (Fig 4A and 4B). The group receiving ART treatment since week 1 post-infection plus PH-797804 achieved the lowest frequency of immune activation markers in CD4+ and CD8+ T cells, although values did not return to baseline and remain approximately 2-fold higher. A similar trend was observed for the Ki-67 marker, which identifies activated cells, undergoing DNA synthesis and cell duplication (Fig 4C and 4D). When PH-797804 was used alone, levels of immune activation were no different than those observed in the control group, suggesting that the level of immune activation in these animals could not be impacted by the PH-797804 dose used here.
When the same analysis was carried out in in lymph node MNC (Fig 5A–5D), we found that, when combined with ART, the inhibitor impact was more significant in the animals initiating ART at week 6 and not as much in those initiating ART at week 1, where only the difference in CD38+/HLA-DR+ CD4 T-cell percentage was statistically significant, possible because the number of biopsy samples available for analysis (5) was smaller than for PBMC (Fig 4). The same analyses carried out in rectal MNC revealed that, when combined with ART, the inhibitor impact was significant in three of the four measured parameters in the animals initiating ART at week 1 but not in animals treated since week 6 post-infection, when AUCs were compared (Fig 5E–5H). However, differences in rectal CD38+/HLA-DR+ and Ki67+ CD8+ T cells between Group 3 and 4 were significant at 4 of 5 and 3 of 5 time points, respectively, when time point values were individually analyzed. Interestingly, blood and lymph node immune activation parameters of Group 4 (ART+ p38 inhibitor initiated in chronic infection, green lines) were comparable to those observed in Group 5 (ART only, initiated in acute infection, pink lines) (p = ns), supporting a significant benefit of this combination treatment when initiated in chronic infection. This result, although limited to phenotypic markers, is important, considering that ART is rarely initiated in the acute infection phase in HIV+ individuals and more commonly initiated during the chronic phase.
During lentiviral infection, the production of various inflammatory cytokines and biomarkers is known to be significantly higher than in normal subjects [62–65]. The levels of these cytokines, measured in plasma of SIV-infected RM by ELISA, were highest at peak viremia and were reduced during ART or ART plus PH-797804 treatment. The level of IFNα, IFNγ, TNFα, IL-6, IP-10, sCD163, a molecule shed by monocytes as a consequence of immune activation [66, 67], and C-reactive protein (CRP) were lower in the animals treated with ART and PH-797804 and the difference was statistically significant when Group 4 was compared to Group 3 (Fig 6 for group averages and S4 Fig for individual animals). Interestingly, even in this analysis, levels of IFNγ, TNFα, IP-10, sCD163 and CRP in Group 4 became comparable to those of Group 5, which initiated ART one week post-infection (p = ns). Of note is the fact that IFNα, although impacted by both ART and ART+PH-797804, remained with IFNγ the most abundant measured cytokine when compared to pre-infection values, despite undetectable viremia in some of the animals, suggesting an ongoing stimulation of innate responses. As a consequence, it is unlikely that activation of the numerous ISG, not covered here, can be fully controlled by simply inhibiting p38 MAPK, as it is highly likely that levels of ISG expression correlate with the levels of IFNs, which, although reduced, were still abnormal in this setting and not only influenced by p38 MAPK.
Plasma cytokine reduction was mirrored by reduction of percentages of T cells producing some of these cytokines (Fig 7 and S5 Fig). We found that the percentage of CD4+ T cells producing IFNγ and TNFα were significantly reduced in the groups receiving ART plus PH-797804 compared to ART alone, whether the treatment is started on week 1 (p = 0.005 for TNFα and p = 0.02 for IFNγ) or week 6 post-infection (p = 0.04 for TNFα), (Fig 7A and 7B). The percentages of TNFα+/CD8+ T cells in treated animals showed a significant reduction when the PH-797804 was added to ART, regardless of time of ART initiation (p = 0.0001 and p = 0.01 for comparisons between Group 5 and 6 and Group 3 and 4, respectively), while instead differences in the percentages of IFNγ+/CD8+ T cells were significant when Group 3 was compared to Group 4 (p = 0.0008) (Fig 7C and 7D) but not when Group 5 was compared to Group 6. In addition, the percentage of PBMC expressing IFNα was also significantly reduced when values of individual time points in Group 4 were compared to those in Group 3 (Fig 7E, asterisks) but this significance did not extend to the AUC comparative analysis. Taken together these data show that the administration of PH-797804 with ART reduced more significantly than ART alone the production of inflammatory cytokines and that, when added to ART in the chronic phase of the infection, which is the most common occurrence in HIV+ individuals, restored some parameters to the levels observed when ART was initiated in the acute phase, one week after infection.
The hypothesis we tested by adding a p38 MAPK inhibitor to ART was that, if immune activation contributes to the immune system deterioration, reduction of immune activation should result in preservation of immune cells. We evaluated the effect of PH-797804 treatment on percentages of total CD4+ T cells and CM CD4+ T cells, which is considered an earlier prognostic marker in the infection, as CM CD4+ T cells decline earlier than total CD4+ T cells [68]. The percentages of CD4+ T cells and CM CD4+ T cells were significantly higher in Group 4 (ART + PH-797804 since week 6) compared to Group 3 (ART alone since week 6) (p = 0.02 for CD4+ T cells and p = 0.01 for CM CD4+ T cells) (Fig 8A and 8B) and differences were not significant when Group 4 was compared to Group 5 (ART since week 1 post-infection). The same comparisons for Groups 5 and 6, receiving ART since week 1 post-infection, with CD4+ T-cell loss not as pronounced as in the groups initiating ART in the chronic phase, were significant for CD4+ T cells but not for CM CD4+ T cells. This result supports the possibility that the addition of PH-797804 to ART permits a more significant recovery of CD4+ T-cell counts than ART alone when the virus damage of the immune system has been more severe and the combined regimen can compensate for a later initiation of ART.
Multiple studies have suggested that losses of intestinal Th17 and Th22 cells play a critical role in establishing intestinal mucosal immune dysfunction and are associated with the chronic immune activation typical of pathogenic HIV/SIV infections [69–80]. Several studies have reported that reciprocal changes in Th17 cells and Tregs occur during HIV and SIV infections and that the relative balance of Th17 and Treg subsets, expressed as a ratio of Th17 and Treg percentages, provides a prognostic index of disease progression more significant than each percentage considered individually [81–83]. In addition, Th22 cells play an important role in promoting innate immune defenses against bacterial and fungal infections in mucosal tissues, and in maintaining mucosal barrier integrity via mucus production and repair of damaged mucosal tissue [74–78, 80]. Therefore, we measured the impact of PH-797804 treatment on levels of the Th17 and Th22 CD4+ T-cell populations by evaluating their percentages in PBMC with intracellular staining in PBMC and in rectal MNC. We found that SIV infection reduced the Th17/Treg ratio but treatment improved it. The improvement was more significant in Groups 4 and 6, both receiving ART+ PH-797804, compared to Groups 3 and 5 (ART alone) (p = 0.02 for Groups 5 and 6 and p = 0.04 for Groups 3 and 4) (Fig 8C). We also evaluated Th17/ Treg CD4+ T-cell ratio and IL22+CD4+ T-cell percentage in the intestinal compartment, where loss of Th17+ and Th22+ cells during SIV infection is significant, and confirmed in rectal MNC an improved ratio in the groups receiving ART+ PH-797804 (p = 0.01 for Groups 5 and 6 and p = 0.0003 for Groups 3 and 4) (Fig 8E). A more significant recovery of PBMC Th22+ cells was observed when the inhibitor was administered (p = 0.04, Group 6 vs. 5, p = 0.01 Group 4 vs. 3, Fig 8D). Recovery of rectal Th22 CD4+ T cell percentages was significant for Group 6 compared to Group 5 (p = 0.01) while differences were not significant between Group 3 and 4, possibly due to limited samples and higher standard error for the groups, as percentages were higher in the group receiving PH-797804 and similar to those observed in Group 6 (Fig 8F). Taken together, these data indicate that the inhibition of T-cell activation achieved with PH-797804 treatment was significant enough to provide additional benefit to that observe with ART alone and positively impacted preservation or restoration of populations that are affected by chronic SIV infection. Importantly, the addition of the PH-797804 to ART initiated in the chronic phase resulted in immune population recoveries comparable than that observed when initiating ART in the acute phase of infection (compare data reported in green to those reported in pink), an event highly unlikely in most HIV infected individuals.
On week 60 after infection, ART and PH-797804 were interrupted. We investigated the viral burden in lymph node cells at the end of ART by evaluating DNA and RNA gag viral copies in MNC extracted from lymph nodes of animals in Group 3–6. We found that viral loads were significantly lower in Groups 5 and 6 that received ART since week 1 post-infection compared to Groups 3 and 4, where ART was initiated on week 6 post-infection (approximately 3.5 fold lower, p = 0.04), but not significantly different when the groups receiving the inhibitor were compared to the matched group that received ART alone. Similarly, the number of average SIV transcripts/106 cells was lower in Group 5 and 6. When the total number of SIV RNA copies/ SIV DNA copies was calculated to obtain the average SIV genomic RNA transcripts /SIV infected cells, the number is very similar in all groups, ranging between 0.01 and 1 SIV gag RNA copies/ SIV DNA copy, except for one animal in Group 4, where the average is of 8.5 RNA copies/DNA copy (Fig 9A–9C). Averages below one SIV RNA copy / infected cell support the coexistence of SIV DNA+ cells without SIV transcripts, where the infection is latent, and others where number of SIV RNA molecules is higher than the calculated average. Single cell analysis is required to establish the fraction of cells in which transcription is active and quantitative infection assays to evaluate cells producing infectious virus. However, these results indicate that the level of suppression was similar and effective in all groups, compared well to those reported for suppressed HIV+ individuals and suppressed SIV+ macaques [84, 85], that the size of the reservoirs was established early on, and that the more prolonged viremia that occurs with later initiation of ART resulted in larger reservoirs. The addition of PH-797804 at week 6 post-infection did not impact the size of reservoirs, measured 54 weeks after inhibitor initiation. Despite the differences in total viral DNA burden in lymph nodes among groups, viremia measured for the first time after 4 weeks from ART interruption, rebounded to similar values in all animals, regardless of when ART was initiated or PH-797804 treatment was administered (Fig 9D). This rebound was comparable to that observed in HIV infection patients treated for a similar length of time, where 136 of 164 patients had significant and fast viral rebound after ART interruption [86, 87]. Not surprisingly, CD38+/HLA-DR+ percentages, which appears strictly linked to viral replication, rebounded as well and was indistinguishable among groups (Fig 9E). These data support the observation that reservoirs are established early in the infection and that initiating ART in the acute phase reduces but does not eliminate the establishment of reservoirs, possible because it takes time for ART to bring viremia to undetectable levels and reservoirs continue to be seeded for days after ART initiation [88–91]. Decay of reservoirs is time dependent and time to rebound is clearly affected by the duration of ART, which in this trial was restricted to 59 or 54 weeks and it can be significantly longer in ART-treated, HIV infected individuals, evaluated for the same virological parameters.
Immune activation is the differentiating feature between infection is species that progress to AIDS versus species that do not and persistent immune activation remains despite immunosuppressive ART treatment. ART treatment substantially reduces viral loads and, consequently, immune activation. However, CD8+ T-cell activation does not decrease proportionally to the decrease of viral loads and its levels have been inversely linked to the degree of CD4+ T cell reconstitution during ART [92]. HIV-associated pathology and, in particular, HIV associated neurological disorders (HAND), did not decrease after the introduction of ART in the same proportion that one would expect if only dependent on blood viral burden [93, 94]. Immune activation and/or ART toxicity have been postulated as possible explanations for this outcome. The causes of residual immune activation have been attributed to factors such as residual HIV replication, persistent microbial translocation, and viral co-infections such as CMV [references in [92]]. It seems therefore ideal to conceive a treatment that combines ART and inhibitors of immune activation. Only limited data are available for this approach, with focus on combination of ART with Cyclosporin A during acute/early infection for one short cycle [95–98]. A significant effect on the residual CD8+ T-cell immune activation was observed with this approach [97]. We investigated whether suppression of one of the key player in immune activation, p38 MAPK, can impact SIV-mediated immune activation during ART. We found that this treatment, when combined with ART, does positively impact virus-mediated immune activation and permits preservation of subpopulations that are significantly affected during the infection. However, reduction of immune activation did not appear to impact viral reservoirs, known to be established early on in SIV and HIV infections, even when ART was initiated just a few days after infection [91].
The addition of the p38 MAPK inhibitor to ART significantly suppressed multiple parameters of immune activation. The observed suppression was not complete and residual immune activation was approximately 50–65% of that observed in animals receiving ART alone. Only one inhibitor dose was explored and investigation of additional doses would be preferable. As it is for ART, the combination of multiple inhibitors of immune activation pathways, aimed at different targets linked to immune activation, may ultimately provide a more substantial control of immune activation during ART, considering that redundancy exists in the immune system and that p38 MAPK activity can also, in part, be carried out by other kinases like JNK and ERK. Indeed, when we carried out in vitro experiments, we found that features of immune activation that are observed after HIV infection of APC could most significantly be impacted by treatment with a p38 MAPK inhibitor but also by a JNK inhibitor and to a lesser extent by an ERK inhibitor [45–47].
p38 MAPK inhibition provided diverse benefits with improvement of multiple immune parameters and resulted in a significant preservation or better restauration of cell populations that are critical to the immune system (Fig 8). Recovery of percentages of total CD4+, CM and Th22 CD4+ T cells, and of Th17/Treg ratio was more substantial in Group 4 that initiated ART and PH-797804 around the time when viremia reaches the set point (week 6) and when CD4+ T-cell depletion had reached values as low as those in the animals that remained untreated. These results support the possibility that the addition of PH-797804 permits a more significant recovery of CD4+ T-cell populations than ART alone, especially when the damage to the immune system has been more prolonged and substantial. In Group 4, recovery of affected cell populations reached levels comparable to those observed in Group 5 that initiated ART only one week after infection (Fig 8). The fact that the addition of PH-797804 provided a more noticeable benefit when ART was initiated later in the infection, when the viral damage to the immune system has been more severe, suggests that this treatment could be significant in HIV+ individuals, where initiation of ART after reaching viremia set-point is the more common occurrence than ART initiated during the acute stage. Lastly, significant reduction of both PD-1 percentages and MFI in CD8+ T-cells may permit a more effective anti-retroviral immune response. As only one PH-797804 dose was tested, the benefit could further increase if the ideal dose was identified.
IL-1β inhibition with canakinumab has recently been shown to reduce cardiovascular events in patients with coronary arterial disease and has also been shown to decrease immune activation in treated HIV patients, but there are concerns about its safety, particularly infectious complications [99, 100]. We did not detect reduced immune responses to SIV while treating the animals with the p38 MAPK inhibitor. The setting in which the p38 MAPK inhibitor was used (single caged macaques, kept indoor) significantly reduces the exposure to infectious agents and therefore may not provide a sufficient indicator of lack of impact on infection control. However, this inhibitor could be safer than an IL-1β inhibitor, as it is highly selective for p38 MAPK and does not inhibit the other two members of the family of major mitogen-activated protein kinases, JNK and ERK. These kinases are partially, even if not fully, overlapping with p38 MAPK function and, therefore, the inhibition of some pathways affected by p38 MAPK may not be absolute, avoiding their complete shutdown. This could also be a reason why the observed reduction of immune activation was limited, even if significant. Although a full analysis of ISG expression in blood and tissue populations is beyond the scope of the report, this investigation could reveal in more subtle details whether some ISG are more affected than others and provide additional information on the activity of PH-797804 in vivo. Such analysis will be object of future, larger studies. The expectation is that, if carried out in the setting of this study, a reduction of ISG that could mirror the reduced levels of circulating IFNα and IFNγ would be detected in animals treated with PH-797804 and ART, and that a more significant suppression of IFN pathways could only be achieved by interfering with other mitogen activated kinases that can overlap p38MAPK or their upstream regulators.
One expectation of this study was that by reducing immune activation, the availability of activated CD4+ T cells that are a preferred target of infection would also be reduced and therefore reservoir seeding could be impacted. However, we did not find significant differences in reservoir size when Group 3 was compared to Group 4 and Group 5 to Group 6. It is possible that differences in rebound time could have been observed if earlier sampling had been obtained, when Group 3 and 4 were compared to Group 5 and 6. As by week 4 all groups treated with ART, with or without PH-797804, had reached similar levels of viremia, this delay would have been limited and unlikely to impact the course of disease progression. This results also suggests that reducing the immune activation status did not translate in reduced reservoirs and supports the observation that viral reservoirs are set early on in the infection [89, 91], before the beginning of ART, are very long-lived [101], and not necessarily exclusively made of resting, infected cells. However, SIV DNA and RNA viral loads in lymph nodes did not increase because of treatment, suggesting that the antiviral effect of ISG was not impacted.
The fact that rebound occurs without administration of latency reactivating agents seems to support the possibility that reservoirs include cells that are not fully resting. Indeed, rebound is observed in HIV+ individuals without treatment of activating agents, although external immune activating stimuli could contribute to the occurrence. It is also possible that in our trial ART was carried out for a period that is relatively short compared to similar studies done in HIV+ individuals, who have received ART for many years, and that the physiological decay of the reservoirs established before the initiation of treatment was not as advanced, opening the possibility that a difference could be observed only if the animals were kept on ART plus PH-797804 for a longer time.
An alternative possibility is that the immune activation detected during ART suppression is not fully independent of viral gene expression but stems from continuous, low-level virion production in tissues that maintains TLR activation. As HIV transcription and SIV entry inhibitors were not part of the regimen of this trial, the therapy administered could permit partial rounds of the virus life cycle, with release and cell entry of non-infectious virions, even if productive infection cannot be achieved after entry. However, entry or uptake of non-infectious virions could be sufficient to trigger TLR signaling, even in the absence of reverse transcription and integration, and this activation could support persistent immune activation. If complete suppression of SIV antigen production were achieved and viral genomes were truly latent during ART, one does not explain the fact that CD8+ T-cell depletion leads to viremia rebound while ART is still in place in SIV-infected RM, as CD8+ T-cell activity requires antigen production to be effective. Evaluation of TLR-activated pathways in the context of ART alone and with immune activation suppression is an important goal of future studies. The possibility of persistent, low level antigenemia, which would offer a significant additional source for persistent immune activation, is not sufficiently considered and explored and requires further investigation.
The study received institutional review board approval at the Tulane Primate Research Center, where the macaques used in the study were housed. IACUC approval number of the study is: P0236R. Animal care methods are consistent with the recommendations of the panel on euthanasia of the American Veterinary Medical Association. This study was also 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, National Academy Press, 1996) and with the recommendations of the Weatherall report: “The use of non-human primates in research”. The institution also accepts as mandatory the PHS “Policy on Humane care and use of Laboratory Animals by Awardee Institutions” and the NIH “Principles for the Utilization and Care of Vertebrate Animals used in Testing, Research and Training”.
Thirty-two male RM (Macaca mulatta), ranging in age between 2.40 and 4.33 years when the study was initiated, were included in this study and were housed at Tulane Primate Research Center, Tulane University, Covington, Los Angeles. Animals were evaluated for the expression of the following MHC molecules: A*01, A*02, A*08, A*11, B*01, B*03, B*04, B*08, B*17. None of the animals included in the study tested positive for the protective alleles A*01, B*08, and B*17 [102, 103]. Animals were divided in 6 groups and infected intravenously (i.v.) with 10 TCID50 SIVmac251 (day 0). Groups 1 and 2 included four animals, Groups 3 to 6 included six animals. The animals in Group 1 were left untreated as control. Since week 6 after infection, Group 2 received PH-797804 alone, group 3 initiated antiretroviral therapy (ART) and group 4 received ART and PH-797804. Group 5 initiated ART one week post- infection as did Group 6, who also received PH-797804 starting from week 6 post-infection. ART consisted of two reverse transcriptase (RT) inhibitors, tenofovir (PMPA, 20 mg/kg, and emtricitabine (FTC, 30 mg/kg), and the integrase inhibitor dolutegravir (DTG, 2.5mg/kg s.i.d), all administered i.m. [104]. The animals in group 2, 4 and 6 received 3 cycles of PH-797804 (10 mg s.i.d, orally administered), each of 12 weeks (6–18, 28–40 and 48–60 weeks). ART and PH-797804 treatment were interrupted on week 60 and animals were monitored monthly for virus rebound until week 72 post infection.
Blood were collected at various time points, approximately every 4 weeks; rectal and lymph node (LN) tissues were biopsied before infection and at the beginning and end of PH-797804 cycles. Briefly, after Telazol anesthesia, seven to eight biopsies/animal/time points were obtained from the rectum and blood was collected in EDTA. PBMC and plasma were separated using Ficoll-Hypaque gradient centrifugation. Rectal and lymph node biopsy-derived MNC were isolated by digestion with 1 mg/ml collagenase for 1 h at 37°C, passed through a 70-mm cell strainer to remove residual tissue fragments and separated using Ficoll-Hypaque gradient centrifugation [105].
Polychromatic flowcytometric analysis was carried out in PBMC, LN and rectal biopsy MNC according to standard procedures for membrane and intracellular staining, using a panel of mAbs (see below) shown to be cross reactive with RMs [105]. The percentages of CD4+ T cells and CM CD4+ T cells, immune activated CD4+ and CD8+ cells, FOXP3+ T regulatory cells, IFNγ+ and TNFα+ CD4+ and CD8+ T cells, and IP-10+, pSTAT1+ and IRF7+ PBMC were evaluated in unstimulated cells by membrane and intracellular staining (ICS) [107] and reported as the percentage of CD4+, CD8+ T cells, or PBMC that express one or more markers. Accumulation of IL-17 and IL-22 in CD4+ T cells was analyzed after phorbol 12-miristate-13 acetate (PMA,10 mg/ml) and Ionomycin (1mg/ml) stimulation, using the same technique. Briefly, aliquots of PBMC, LN and rectal MNC were re-suspended at 106 cells/ml in medium with or without stimulation and containing Golgi stop (BD Bioscience, San Diego, CA) and incubated at 37°C for ~ 12 hr. The cells were then washed and stained with the appropriate panel of antibodies for 30 minutes in the dark at room temperature, followed by fixation and permeabilization. After permeabilization, the cells were stained intracellularly with monoclonal antibodies against the cytokines of interest for 1 hour in the dark. Data were acquired on BD LSR II flow cytometer using FACSDIVA software. After acquisition, data were analyzed using the FlowJo software.
The following anti-human, macaque cross-reacting, or anti-macaque antibodies were used in this study: anti-CD3-pacific blue/PerCp-Cy5.5 (clone SP34-2), anti-CD4-Amcyan (clone L200, anti-CD8-APC-Cy7 (clone RPA-T8), anti-Ki-67-Alexa Fluor-700 (clone B56), anti-HLADR-PE-CF594 (clone G46-6), anti-TNFα- PE (clone MAb11), anti-IFNγ-Alexa Fluor-700 (clone B27), anti-IRF7-APC (clone K47-671), anti- pSTAT1-Pacific blue (clone 14/ pSTAT1), anti-CD14-Pe-Cy7 (clone M5E2) (all from BD Pharmigen); anti-FOXP3-FITC (clone 206D), anti-IP-10-PE (clone J034D6), anti-CD28-Pe-Cy5 (clone CD28.2), anti-CD95-FITC (clone DX2), anti-PD-1-PerCp-Cy5.5 (clone EH12.2H7), anti-IL-2-APC (clone MQ1-17H12) (all from BioLegend, San Diego, CA); anti-IL-17- PerCp-Cy5.5 (clone- eBio64DEC17), anti-IL-22-APC (clone IL22JOP) (both from eBiosciences, San Diego, CA) and anti-CD38-APC (clone OKT10) from NHP Reagent Resource, Boston, Massachusetts).
The levels of IFNγ, TNFα, IL-6, and IP-10 in plasma were measured using commercially available ELISA kits from U-CyTech, Utrecht, Netherlands, according to manufacturer’s instructions. CRP, sCD163 and IFNα plasma levels were measured using a monkey CRP ELISA kit (Life Diagnostics Inc., West Chester, PA), a sCD163 ELISA kit (My Biosource, CA, USA) and an IFN α ELISA kit (PBL Assay Science, NJ, USA).
Calculations and statistical analyses were performed using the GraphPad Prism version 7 software. Normality distribution values were calculated using the D'Agostino-Pearson omnibus test. When values were normally distributed, p values were calculated using unpaired t-test, when non-normal, Wilcoxon-Mann-Whitney (rank sum) test was applied. Between-group comparisons at individual time points were carried out with Wilcoxon-Mann-Whitney (rank sum) test or unpaired using t-test depending on population normality values. AUC analyses were carried out by calculating an AUC for the time course values of each animal and areas in one group were compared to those is a second group. Results of statistical analyses were considered significant if they produced p values ≤ 0.05.
|
10.1371/journal.pgen.1003814 | dTULP, the Drosophila melanogaster Homolog of Tubby, Regulates Transient Receptor Potential Channel Localization in Cilia | Mechanically gated ion channels convert sound into an electrical signal for the sense of hearing. In Drosophila melanogaster, several transient receptor potential (TRP) channels have been implicated to be involved in this process. TRPN (NompC) and TRPV (Inactive) channels are localized in the distal and proximal ciliary zones of auditory receptor neurons, respectively. This segregated ciliary localization suggests distinct roles in auditory transduction. However, the regulation of this localization is not fully understood. Here we show that the Drosophila Tubby homolog, King tubby (hereafter called dTULP) regulates ciliary localization of TRPs. dTULP-deficient flies show uncoordinated movement and complete loss of sound-evoked action potentials. Inactive and NompC are mislocalized in the cilia of auditory receptor neurons in the dTulp mutants, indicating that dTULP is required for proper cilia membrane protein localization. This is the first demonstration that dTULP regulates TRP channel localization in cilia, and suggests that dTULP is a protein that regulates ciliary neurosensory functions.
| Tubby is a member of the Tubby-like protein (TULP) family. Tubby mutations in mice (tubby mice) cause late-onset obesity and neurosensory deficits such as retinal degeneration and hearing loss. However, the exact molecular mechanism of Tubby has not been determined. Here we show that Drosophila Tubby homolog, King tubby (dTULP), regulates ciliary localization of transient receptor potential protein (TRP). dTULP-deficient flies showed uncoordinated movement and complete loss of sound-evoked action potentials. dTULP was localized in the cilia of chordotonal neurons of Johnston's organ. Two TRP channels essential for auditory transduction, Inactive and NompC, were mislocalized in the cilia of chordotonal neurons in the dTulp mutants, indicating that dTULP is required for proper cilia membrane protein localization. This is the first demonstration that dTULP regulates TRP channel localization in cilia, and thus provides novel insights into the pathogenic mechanism of tubby mice.
| The auditory system allows animals to communicate and obtain information about their environment. The hearing organs transform sound into an electrical signal through a process called mechanotransduction, the conversion of a mechanical force impinging on a cell into an intracellular signal [1]. Although the recent discovery of several molecules involved in mechanotransduction allows interpretation of the biophysical properties of the mechanotransduction process for hearing [2], many additional molecular players in auditory development and function are waiting to be unveiled.
Drosophila melanogaster has been suggested as a model organism to study the fundamental process of hearing [3], [4]. Hearing in the fly is necessary for the detection of courtship songs [5]–[7]. Male-generated courtship song causes females to reduce locomotion and enhances female receptivity, whereas it causes males to chase each other [8]. The ability to hear courtship songs is ascribed to Johnston's organ (JO) in the second antennal segment. Near-field sounds rotate the sound receiver; the third antennal segment and the arista and this rotation of the antennal receiver transmits mechanical forces to the JO in the second antennal segment, which is connected to the third antennal segment by a thin stalk [9]. Each JO sensilla consists of two or three chordotonal neurons and several supporting cells. The outer dendritic segments of the JO neurons are compartmentalized cilia which are directly connected to the antennal sound receiver via extracellular caps. The distortion of the junction between the second and third segment stretches the cilia and stimulates the JO neurons.
Several transient receptor potential (TRP) channels have been shown to be required for Drosophila hearing transduction and amplification [4], [10]–[14]. Mutation in nompC, the Drosophila TRPN channel, resulted in substantial reduction of sound-evoked potentials [4]. Reports showing that NompC and TRP-4 (the C. elegans ortholog of NompC) are bona fide mechanotransduction channels support the idea that NompC is the Drosophila hearing transducer [15], [16]. Two Drosophila TRPV channel, inactive (iav) and nanchung (nan), mutants showed complete loss of sound-evoked action potentials [11]. However, they have not been considered to be the hearing transduction complex per se; rather they are thought to be required to amplify the electric signal generated by the hearing transduction complex, since Iav and Nan reside in the proximal cilia which are distant from the distal cilia where NompC is localized and mechanical force is directly transmitted [17], [18]. A recent study which employed a new method to measure subthreshold signals from the JO neurons suggested the opposite possibility that the TRPV (Iav and Nan) complex is the hearing transduction complex modulated by TRPN (NompC) [14]. Although the exact roles of each TRP in Drosophila hearing are still controversial, it is clear that TRPN and TRPV have essential and distinct roles in Drosophila hearing.
Several attempts have been made to identify molecular players regulating the function of the ciliated mechanoreceptor neurons. Gene expression profiling identified chordotonal organ-enriched genes from campaniform mechanoreceptors, developing embryo chordotonal neurons, and the second antennal segment [19]–[21]. Alternatively, chordotonal neuron-specific genes were identified by searching for regulatory factor X (RFX)-binding sites, because ciliogenesis of the chordotonal neurons mainly depends on the RFX transcription factor [22]. However, so far only a limited number of genes involved in TRP channel localization in the JO neuron cilia have been identified and characterized, including axonemal components and intraflagellar transports (IFTs) [17], [23]. IFTs are indispensable for the formation and maintenance of cilia as well as for the transport of proteins along the microtubules in and out of the cilia [24]–[26]. Therefore, mutation of many of the characterized genes results in not only delocalization of the TRPs but also profound structural abnormality in cilia, rendering it difficult to delineate the gene functions specific to TRP localization.
Tubby is the founding member of Tubby-like proteins (TULPs) [27]. Loss-of-function of the Tubby gene exhibits adult-onset obesity, retinal degeneration, and hearing loss in mice. The Drosophila genome encodes one Tubby homolog called King tubby (hereafter designated dTULP), which shares approximately 43% amino acid identity with mouse Tubby (Figure S1A) [28]. At the embryonic stage, dTULP is expressed in various types of neuronal cells including the chordotonal neurons. Although previous expression analyses and bioinformatic approaches detected dTulp in the chordotonal organs, its presence did not attract much interest because of its distribution in various neuronal cell types [22].
In this study, we aimed to investigate the novel molecular function of dTULP in Drosophila hearing. dTULP is localized to the well-defined ciliary structure of Drosophila auditory organs. Loss of dTULP has no effect on the ciliary structure of the JO neurons, but NompC and Iav localization in cilia was severely altered. These data demonstrate a new role of dTULP as a regulator of TRP localization in the hearing organs.
To test whether dTULP plays a role in Drosophila hearing, we generated two dTulp mutant alleles by ends-out homologous recombination [29]. The first allele was dTulp1, which harbours a deleted C-terminal containing the conserved “tubby domain” (residues 220 to 460; Figure 1A). The second allele, dTulpG, was generated by replacing an N-terminal portion of the dTULP coding region (residues 18 to 261; Figure S1B) with GAL4 coding sequences at the site corresponding to the initiation codon of the short splicing variant of dTulp. Genomic PCR analyses showed that the dTulp genomic locus was deleted in dTulp1 and dTulpG flies (Figure 1B and Figure S1B). We raised antibodies to dTULP, which recognized a 51 kDa protein as predicted in wild-type fly extracts on a Western blot, and confirmed that dTULP was not detected in dTulp1 and dTulpG fly extracts (Figure 1C). Both alleles are homozygous viable and fertile.
Since both dTulp1and dTulpG mutant alleles showed postural problems and uncoordinated movement, we performed a climbing assay. Flies were banged down to the bottom of a vertical tube and the percentage of the flies climbing above half of the height of the vertical tube within 10 seconds was recorded as the climbing index. dTulp1, dTulpG, and transheterozygote flies exhibited a decreased climbing index compared to control flies (Figure 1D). Introduction of a P[acman] clone containing the dTULP coding region (CH321-59C17) in the dTulp1 mutant background rescued this phenotype [30]. These data suggested that dTulp mutants may have functional defects in the JO neurons [13].
To check for hearing defects in dTulp mutant flies, we recorded extracellular sound-evoked potentials in wild-type and dTULP-deficient flies. Sound-evoked potentials were completely abolished in dTulp1, dTulpG, and dTulp1 in trans with a deletion that completely removed dTulp, Df(2R)BSC462. Genomic rescue using the P[acman] clone produced sound-evoked potentials similar to those in the wild-type, suggesting that the hearing defect was specifically due to dTulp ablation (Figure 1E and 1F).
To test whether dTULP is expressed in the JO neurons, we first attempted to take advantage of the GAL4/UAS system using the dTulpG allele. However, the GAL4 reporter inserted in dTulpG was not expressed. This may be caused by inserting GAL4 at the site corresponding to the initiation codon of the short splicing variant of dTulp rather than the long splicing variant.
Therefore, we performed immunohistochemistry with dTULP antibodies. We found that dTULP was expressed in the cilia as well as the cell body of the chordotonal neurons (Figure 2A, left). We did not detect dTULP immunoreactivity in the JO neurons in dTulp1flies, indicating that the immunosignal is specific for dTULP (Figure 2A, right). To further characterize the ciliary localization of dTULP, we compared the localization of dTULP with that of Iav and NompC. The subcellular localization of Iav and NompC are in the proximal and distal cilia, respectively, in a mutually exclusive manner (Figure 2B) [11], [17], [18], [31]. dTULP staining extended from the proximal to distal cilia with a much weaker signal observed in the distal portion (Figure 2C and 2D). The mouse Tubby protein has been reported to shuttle from the plasma membrane to the nucleus upon Gq-coupled G protein-coupled receptor (GPCR) activation [32]. dTULP was also detected in the cell body as well as the nucleus in the JO neurons (Figure 2A and Figure S7B). We also found that dTULP was expressed in other types of sensory neurons with cilia (Figure S2).
To examine whether the dTulp mutants have developmental defects in the JO neuron structure, we observed the expression of a membrane-targeted GFP (UAS-mCD8:GFP) driven by the pan-neuronal promoter (elav-GAL4) in the JO neurons. We found no gross structural abnormalities in dTulp1flies (Figure 3A). Electron microscopy of the JO revealed that most dTulp mutants had normal ciliary ultrastructure (Figure 3B). Approximately 9.3% of chordotonal scolopidia appeared abnormal in terms of cilia number or cap-cilia connections (Figure S3). In addition, we did not observe any discernible changes in the expression of the dendritic cap protein NompA, which transmits mechanical stimuli to the distal segment of chordotonal neurons in dTulp mutants (Figure 3C) [33]. These observations suggest that structural changes in the JO cannot account for severe hearing impairment in dTulp mutants.
Mutations of trps, including iav and nompC, cause hearing defects in Drosophila [4], [11]. To investigate the possibility that dTULP controls the expression of TRPs and other genes which are indispensable for Drosophila hearing, we performed quantitative PCR analysis of such genes and no significant differences in expression levels were present between wild-type and dTulp1 antennae (Figure S4). This suggested that dTULP plays other roles in Drosophila chordotonal neurons rather than as a transcription factor that controls transcription of known hearing related genes, although we cannot exclude the possibility that dTULP regulates the expression of hearing related genes we did not survey.
Next we examined the ciliary localization of Iav and NompC in the dTulp mutants. Surprisingly, Iav was not localized to the proximal cilia in dTulp1 flies (Figure 4A). Furthermore, NompC, characteristically localized to the distal cilia (Figure 2B), was redistributed toward the proximal cilia (Figure 4B). Spacemaker (Spam) is an extracellular protein which protects cells from massive osmotic stress [34]. Localization of Spam was also altered in dTulp mutants from its two typical locations: the luminal space adjacent to the cilia dilation and the scolopidium base (Figure 4C) [35]. Introduction of the dTulp+ transgene rescued the localization of Iav, NompC, and Spam (Figure 4).
IFTs are involved in the localization of Iav, NompC, and Spam [17], [23]. Because IFT mutants show similar phenotypes to the dTulp mutant, we investigated the localization of IFT proteins in dTULP-deficient flies. Ciliary localization of the two IFTs, NompB (the ortholog of human IFT-B, IFT88) and RempA (the ortholog of human IFT-A, IFT140), was unaffected in dTulp1 mutants (Figure S5).
To further address the functional relationship between dTULP and IFTs, we examined distribution of dTULP in three IFT (nompB, rempA, and oseg1) mutants and a retrograde motor dynein heavy chain (beethoven) mutant. Although the rempA, oseg1, and beethoven mutants show different degrees of defective cilia structure, dTULP is localized to the deteriorated cilia of each mutant, suggesting that rempA, Oseg1, and beethoven are not required for the transport of dTULP into the cilia (Figure S6A–S6C). Since the nompB mutant does not develop cilia structure, dTULP was present in the inner segment at a high level (Figure S6D) [36]. However, it is possible that other IFTs may play a role for dTULP ciliary localization even though the IFTs we examined are not involved in ciliary localization of dTULP.
Mammalian Tubby have two distinct domains: nuclear localization signal (NLS) and phosphoinositide (PIP)-binding domain. An NLS, which allows Tubby to translocate into the nucleus, resides in the N-terminal region of Tubby [32]. Recently, a short stretch of amino acids including the NLS in TULP3, a mammalian member of the Tubby-like protein family, has been reported as an IFT-A binding domain [37]. A PIP-binding domain in the C-terminal tubby domain allows Tubby to be localized under the inner leaflet of the plasma membrane through binding to specific phosphoinositides. These domains are also conserved in dTULP (Figure 5A).
In order to investigate the mechanism by which dTULP regulates the ciliary localization of Iav and NompC, we introduced mutations into the putative NLS/IFT-binding (dTULPmutA), PIP-binding domain (dTULPmutB), or both domains (dTULPmutAB) of dTulp cDNA and generated UAS-wild-type dTulp (UAS-dTulpwt), UAS-dTulpmutA, UAS-dTulpmutB, and UAS-dTulpmutAB transgenic flies, respectively. To eliminate positional effects, all transgenes were integrated into the same loci using site-specific recombination with an attP landing site on the third chromosome [38].
To test the effect of each mutation on the subcellular localization of dTULP, we examined the subcellular localization of dTULPwt, dTULPmutA, and dTULPmutB in Drosophila salivary glands. dTULPwt was detected mainly in the plasma membrane and nucleus (Figure S7A). Mutations in the NLS/IFT-binding domain or PIP-binding domain of dTULP resulted in significant exclusion from the nucleus or accumulation in the nucleus, respectively, which suggested that the NLS/IFT-binding and PIP-binding properties of mouse Tubby are conserved in dTULP in Drosophila salivary glands (Figure S7A). However, the localization of dTULPwt, dTULPmutA, and dTULPmutB in the JO neurons in terms of the cell body and nuclear distribution was virtually the same (Figure S7B). These data suggested that dTULP is not shuttled between the plasma membrane and the nucleus in the JO neurons and these domains may have other functions in the JO neurons rather than controlling the translocation of dTULP from the plasma membrane to the nucleus.
To evaluate the functional consequences of each mutation, we expressed dTULPwt, dTULPmutA, dTULPmutB, or dTULPmutAB in the JO neurons of dTulp1 flies. The expression of dTULPwt in the dTulp mutant background restored the distribution and the expression level of Iav and NompC similar to those of wild type (Figure 5B and 5C). The expression of dTULPmutA or dTULPmutB rescued the Iav trafficking defect of the dTulp mutant, but the expression levels of Iav in the proximal cilia in dTULPmutA- or dTULPmutB-expressing flies were reduced compared to those of dTULPwt-expressing flies (Figure 5B and 5E). NompC localization to the distal cilia in dTULPmutA- or dTULPmutB-expressing flies was similar to that in dTULPwt-expressing flies (Figure 5C). dTULPmutAB could not rescue the Iav or NompC localization defects of the dTulp mutant. This difference was not due to the expression levels of the mutant dTulp transgene since the expression levels of mutant forms of dTULP were similar to those of wild-type dTULP (Figure S8).
Next, we examined whether the different degrees of rescue of Iav and NompC localization was due to differential ciliary trafficking of variant forms of dTULP. The ciliary expression level of dTULPmutB was similar to that of dTULPwt, whereas the ciliary expression levels of dTULPmutA and dTULPmutAB were reduced compared with those of dTULPwt (Figure 5D and 5F). These data suggested that the putative NLS/IFT-binding domain of dTULP has a regulatory function to control the trafficking of dTULP into the cilia.
Consistent with immunohistochemical analyses, dTULPwt fully rescued the hearing defect of the dTulp mutant. dTULPmutA and dTULPmutB restored a partial function and dTULPmutAB had no such activity (Figure 5G and 5H).
In the current study, we demonstrate that dTULP is a cilia trafficking regulator in the Drosophila hearing system. Mutation of dTulp results in hearing loss due to the mislocalization of two TRP channels, Iav and NompC, which are ciliary membrane proteins. In addition, Spam, whose localization is dependent on the IFT machinery, is also mislocalized in dTulp mutants.
How does dTULP regulate the ciliary distribution of TRPs in the JO neurons? Several studies have shown that mutations in IFT machinery or cilia components result in mislocalization of Iav, NompC, and Spam, along with abnormal axonemal structure [17], [23]. It is notable that, in contrast to IFT or cilia component mutants, ciliogenesis and maintenance appear normal in dTULP-missing flies. Furthermore, the altered distribution of Iav, NompC, and Spam in dTulp mutants was not due to the mislocalization of IFT proteins, since the localization of two IFTs (NompB and RempA) was normal in dTulp mutants (Figure S6). These data suggest that dTULP acts downstream of the IFTs to regulate TRP localization.
Even though the mutation of dTulp affected the trafficking of both Iav and NompC, the compartmentalized ciliary localization of Iav and NompC is differentially regulated by dTULP. An individual mutation in either the putative IFT- or PIP-binding domain reduced Iav expression levels in cilia, whereas NompC localization was not altered until both domains were mutated. Even after the double mutations in both domains of dTULP, NompC is still situated inside the cilia, but in abnormal locations. These findings demonstrate that ciliary entry of NompC is not dependent on dTULP while the distal ciliary localization of NompC is dependent on dTULP. One possibility is that dTULP allows NompC to disengage from the IFT complex at the distal cilia so that NompC is enriched in the distal cilia through the mechanism that required both IFT- and PIP-binding domains. It is also possible that the distal ciliary localization of NompC is regulated by an unidentified factor(s) whose ciliary localization is dTULP-dependent as is Iav.
Both the putative IFT- and PIP-binding domains play important roles in the proper Iav distribution in cilia, but they appear to have different roles. Even though the IFT- or PIP-binding mutant forms of dTULP could only partially rescue the ciliary levels of Iav to the similar extent, the mutation of the IFT-binding domain reduced the ciliary levels of dTULP while disruption of the PIP-binding domain had no effect on the ciliary levels of dTULP. These findings suggest that two domains play distinct roles in the regulation of the ciliary localization of Iav. The IFT-binding domain is the motif required for the ciliary entry for dTULP, and the PIP-binding domain is not related to dTULP ciliary entry itself, rather it affects recruitment of Iav-containing preciliary vesicles to dTULP. By these two linked steps, Iav localization to cilia would be facilitated by dTULP.
In mammals, IFT-A directs the ciliary localization of TULP3 through physical interaction between TULP3 and the IFT-A core complex (WDR19, IFT122, and IFT140), and in turn, promotes trafficking of GPCR to the cilia. Indeed, the depletion of individual IFT-A core complex components affects the ciliary localization of TULP3, which results in the inhibition of GPCR trafficking to the cilia [37]. It appears that dTULP and TULP3 have the similar molecular mechanisms to regulate ciliary membrane proteins. However, unlike TULP3, dTULP ciliary access is not dependent on IFT-A. dTULP ciliary trafficking was not affected by the mutation of Oseg1 (an ortholog of human IFT-A, IFT122) or rempA (an ortholog of human IFT-A, IFT140). Furthermore, the presence of dTULP in cilia did not determine the normal localization of Iav. For example, in the rempA mutant, even when dTULP was localized to the cilia (Figure S6B), Iav was not found in cilia [23]. Taken together, dTULP facilitates the relay of preciliary vesicles to the IFT complex at the base of cilia rather than moving together with ciliary membrane proteins into the cilia as an adaptor between IFT and cargo. dTULP may have other additional roles in cilia, which needs to be explored in the future. Based on our finding that dTULP but not Iav could be found in cilia of IFT mutants, it is also possible that recruitment of Iav-containing preciliary vesicles requires dTULP and additional unknown factors, whose function is altered in IFT mutants. Thus, Iav-containing preciliary vesicles may not be able to form stable interactions with dTULP and IFTs.
After the cloning of the Tubby gene two decades ago, one promising hypothesis has been that Tubby is a transcription factor, since Tubby translocates to the nucleus upon GPCR activation and the N-terminal region of Tubby has transactivation potentials [32], [39]. However, candidate target genes for Tubby have not been identified. Tubby is thought to have additional functions including vesicular trafficking, insulin signaling, endocytosis, or phagocytosis [40]–[43]. It is still not clear how these molecular functions lead to the in vivo phenotypes observed in the tubby mouse. Meanwhile, several studies have hinted at possible connections between the phenotypes of tubby mutant mice and ciliary dysfunction. Tubby mice phenotypes comprise syndromic manifestations that are commonly observed in ciliopathies such as Bardet-Biedle syndrome [44] and Usher syndrome [45], [46]. Recently, GPCR trafficking into neuronal cilia was reported to be misregulated in tubby mice [47]. Mutation of Tulp1, a member of the TULPs, in human and mice, exhibits retinal degeneration due to the mislocalization of rhodopsin [48]. TULP3 represses Hedgehog signalling, which is a crucial signalling cascade in cilia, via the regulation of the ciliary localization of GPCRs [49]. Our current study provides additional supports for the idea that TULPs play an important role in ciliary signalling and that the tubby mouse syndrome might be due to the ciliary defects.
In contrast to mammalian cells, only specialized cell types have the ciliary structure in Drosophila, and the expression of dTULP is not restricted to organs with the ciliary structure, which suggested that dTULP may have other roles not related to the ciliary function [28]. For example, dTULP mediates rhodopsin endocytosis in Drosophila photoreceptor cells which do not have cilium in contrast to its mammalian counterpart [50].
In summary, we demonstrate an intriguing role of dTULP in governing the ciliary localization of TRP proteins. This is the first in vivo evidence showing that dTULP may have important roles in the maintenance of ciliary functions by regulating the localization of ciliary proteins, thereby maintaining sensory functions.
All fly stocks were maintained in regular laboratory conditions (conventional cornmeal agar molasses medium, 12-h light/12-h dark cycle at 25°C, 60% humidity). Iav-GFP and NompA-GFP were reported previously [13], [33]. RempA-YFP and NompB-GFP were from M. Kernan. Y. Jan and M. Noll provided UAS-NompC:GFP and Poxn-GAL4, respectively. Df(2R)BSC462, elav-GAL4, UAS-mCD8:GFP, AB1-GAL4, F-GAL4, and Orco-GAL4 were from the Bloomington Stock Center (Bloomington, IN).
We employed ends-out homologous recombination to generate dTulp mutant alleles. To make the dTulp1 allele, 3 kb genomic DNA at the 5′ and 3′ ends of the tubby domain (220 to 460 residues) coding sequence was PCR amplified from w1118 and subcloned into the pw35 vector. The primer sequences for the 5′ homologous arm of the pw35 vector are 5′-AAAGCGGCCGCCACCGGTGACATCCTCATGTTC-3′ and 5′-AAAGCGGCCGCGTTGCATCACGAACTGGTCGATATTG-3′. The primer sequences for the 3′ homologous arm of the pw35 vector are 5′-TGAGCTGGCTGGGATCCTCGGGTTGG-3′ and 5′-GTGGATCCTTCCTGGTTGGCATCACGTTGAC-3′. To generate the dTulpG allele, we used the pw35GAL4loxP vector in which GAL4 and white are flanked by loxP sequences so the cassette can be removed by introducing Cre recombinase. We subcloned the 3 kb of genomic DNA from each of the 5′ and 3′ ends of the dTULP coding region (18 to 261 residues) into the pw35GAL4loxP vector. The primer sequences for the 5′ homologous arm of the pw35GAL4loxP vector are5′-ACAGATCTCACCGTCGCCTGGCTCAGTGCCC-3′ and 5′-GTGGTACCCAGCTGGCGCTGCAAAGCAGTTAAATC-3′. The primer sequences for the 3′ homologous arm of the pw35GAL4loxP vector are5′-AAAGCGGCCGCGTGGGTTATTGATAGTGATCCTCTA-3′ and 5′-AACCGCGGCGTACAGAATACTCCCTGTTCATGTCT-3′. We generated transgenic flies by germ line transformation (BestGene Inc., Chino Hills, CA) and screened for the targeted alleles as described previously [51]. Targeted alleles were subjected to outcross for five generations into a w1118 genetic background.
We amplified dTulp cDNAs from cDNA clones (RE38560) with PCR and subcloned the fragments into the pUASTattB vector. These constructs were subjected to further modification. We generated the dTulpmutA and dTulpmutB mutant constructs using site-directed mutagenesis to change the sequence encoding R23QKR to L23AAA, and K292LR to A292LA, respectively. The dTulpmutAB construct was generated by introducing the mutation corresponding to dTulpmutB into the dTulpmutA construct. To generate genomic rescue transgenic flies, we obtained the BAC clones CH321-59C17 from the BACPAC Resource Center (Oakland, CA) and used these as genomic rescue constructs. Transgenic flies were generated using PhiC31 integrase-mediated transgenesis on the third chromosome to minimize position effect (Bloomington stock number 24749).
Sound-evoked potentials were recorded as described by Eberl et al [4]. Briefly, the fly's antennal sound receivers were stimulated by computer-generated pulse songs. Neuronal responses were detected using a recording electrode inserted in the junction between the first and second antennal segment and a reference electrode was inserted in the dorsal head cuticle. The signals were subtracted with a DAM50 differential amplifier (World Precision Instruments, Sarasota, FL) and digitized using Superscope 3.0 software (GW Instruments, Somerville, MA). Each trace represents the average responses to 10 stimuli.
For whole-mount staining, antennae were dissected at the pupa stage and the labellum and legs were prepared at the adult stage. Salivary glands were dissected from third instar larvae. For antenna sections, fly heads were embedded in OCT medium and 14 µm frozen cryostat sections were collected. Dissected tissues and sections were fixed for 15 min with 4% paraformaldehyde in 1× PBS containing 0.2% TritonX-100 (PBS-T) and washed three times with PBS-T. The fixed samples were blocked for 30 min with 5% heat-inactivated goat serum in PBS-T and incubated overnight at 4°C in primary antibodies diluted in the same blocking solution. The tissues were washed three times for 10 min with PBS-T and incubated for 1 h at room temperature in secondary antibodies diluted 1∶500 in blocking solution. Following three washes with PBS-T, the samples were mounted with Vectashield (Vector Laboratories, Burlingame, CA) and examined using a Zeiss LSM710 confocal microscope (Jena, Germany).
To quantify Iav-GFP and dTULP expression levels in cilia, all samples were prepared at the same time and all confocal images were obtained under the same conditions. The pixel intensity of each protein was measured using Zen Software (Jena, Germany). Iav-GFP intensity was measured without immunostaining.
Rabbit dTULP antibodies were raised by injecting animals with a purified His-tagged dTULP fusion protein (residue 95–339), followed by affinity purification. The primary antibodies were used in immunohistochemistry at the following dilutions: rabbit anti-dTULP, 1∶400; 22C10, 1∶200 (Hybridoma Bank, University of Iowa); 21A6, 1∶200 (Hybridoma Bank); rabbit anti-Orco, 1∶1,000 (gift from L. Vosshall); rabbit anti-NompC, 1∶20; rabbit anti-GFP, 1∶1,000 (Molecular Probes, Eugene, OR); mouse anti-GFP, 1∶500 (Molecular Probes). The secondary antibodies used were Alexa 488-, Alexa 568-, and Alexa 633-conjugated anti-mouse or anti-rabbit IgG (Molecular Probes; 1∶500). DNA and actin were visualized by DAPI and Alexa Fluor 633 Phalloidin (Molecular Probes) staining, respectively.
Fly head or antennae lysates from each genotype were subjected to electrophoresis on SDS-polyacrylamide gels and transferred onto polyvinylidene fluoride membranes. The membranes were blocked for 1 h with 5% nonfat milk plus 0.1% Tween-20. Membrane-bound proteins were analyzed by immunoblotting with primary antibodies against dTULP (1∶1,000) and tubulin (Hybridoma Bank, 1∶2,000).
Fly heads were dissected and fixed in 2% paraformaldehyde, 2.5% glutaraldehyde, 0.1 M cacodylate, and 2 mM CaCl2, pH 7.4. The tissue was embedded in LR white resin. Thin sections were cut, mounted on formvar-coated single slot nickel grids, counterstained with uranyl acetate and lead citrate, and examined on a Hitachi H-7500 electron microscope (Hitachi, Tokyo, Japan).
Total RNA was extracted from adult antennae using Trizol reagent (Invitrogen, Carlsbad, CA). cDNA was generated from 0.5 µg of RNA from each genotype using the SuperScript III First Strand Synthesis System (Invitrogen). Quantitative PCR was performed using an ABI7500 real-time PCR machine (Applied Biosystems, Foster City, CA) and the ABI SYBR green system. Transcript levels were normalized to rp49 as an internal control and the ΔCT (CT = threshold cycle) method was used to calculate the relative amount of mRNAs.The primers used for qRT-PCR are listed in Table S1.
Fifteen 3- to 6-day-old flies were placed in an empty fly food vial. The climbing index is the fraction of flies that climb halfway up the vials in 10 s after being tapped down to the bottom of the tube. We performed each experiment twice and used the average of the two trials to calculate the climbing index.
Data shown are the mean ± SEM. To compare two sets of data, unpaired Student's t-tests were used. ANOVA with the Tukey post-hoc test was used to compare multiple sets of data. Asterisks indicate statistical significance.
|
10.1371/journal.pntd.0002303 | Assessment of Transmission in Trachoma Programs over Time Suggests No Short-Term Loss of Immunity | Trachoma programs have dramatically reduced the prevalence of the ocular chlamydia that cause the disease. Some have hypothesized that immunity to the infection may be reduced because of program success in reducing the incidence of infection, and transmission may then increase. Longitudinal studies of multiple communities would be necessary to test this hypothesis. Here, we quantify transmission using an estimated basic reproduction number based on 32 communities during the first, second, and third years of an antibiotic treatment program. We found that there is little to no increase in the basic reproduction number over time. The estimated linear trend in the basic reproduction number, , was found to be −0.025 per year, 95% CI −0.167 to 0.117 per year. We are unable to find evidence supporting any loss of immunity over the course of a 3-year program. This is encouraging, as it allows the possibility that repeated mass antibiotic distributions may eliminate infection from even the most severely affected areas.
| Trachoma, caused by repeated infections by the ocular strains of Chlamydia trachomatis, is the most common infectious cause of blindness in the world. Treatment for trachoma includes mass azithromycin treatments to the entire community. To reduce the prevalence of infection, the World Health Organization (WHO) advocates at least three annual community-wide distributions of oral antibiotics in affected areas, with further mass treatments based on the prevalence of trachoma. Trachoma programs have dramatically reduced the community prevalence of infection, and some have argued that lowered prevalence of infection may lead to reductions in immunity, and that less immunity may in turn lead to increased transmission from what infection remains. Here, we used a stochastic transmission model to analyze data collected from a 3-year antibiotic treatment program (a 32-community, cluster-randomized clinical trial in Tanzania) to assess whether or not transmission actually increases during elimination campaigns. We found no evidence supporting any increase in transmission over the course of the program. The absence of a short term increase in transmission as the prevalence decreases is good news for trachoma programs.
| The World Health Organization has targeted trachoma for elimination by the year 2020 [1]. Repeated mass oral azithromycin distributions have been a cornerstone of the treatment strategy. Theoretically, repeated treatments may eventually eliminate infection from even the most severely affected areas [2], [3]. In practice, distributions have dramatically reduced the prevalence of infection in a number of locations [3], [4], [5], [6], [7], [8], [9]. However, there remains concern that resistance may develop or that loss of immunity may prevent complete elimination [10], [11], [12], [13], [14]. While no stable chlamydial drug resistance has yet been observed, a loss of immunity is possible. Individuals who had not been exposed to infection recently might have less protection than they had had when the infection was more prevalent. An increase in transmission during the course of a program could indicate loss of immunity. Multiple communities would need to be monitored, to assess whether random fluctuations may explain observed differences over time. Here, we analyzed multiple communities from the Tanzanian portion of the Program for the Rapid Elimination of Trachoma trial (PRET [15]) using a stochastic model of transmission, to assess the initial reproductive number for transmission over time.
Communities were monitored as part of a cluster-randomized, trachoma treatment trial in Tanzania [15], [16]. In brief, 32 communities in Tanzania were randomized in a two by two factorial design. The first factor was the use of standard versus enhanced coverage with annual mass antibiotic treatment; the second factor was the use or disuse of a rule whereby mass antibiotic administration would be discontinued based on ongoing monitoring. In fact, the use of this rule never led to discontinuation of mass antibiotic administration during the first three years. Thus, all 32 communities received treatment at baseline, 12, and 24 months. A baseline census was conducted in all 32 communities, and again at 12, 24, and 36 months. One hundred randomly selected children aged 0–5 years were examined at baseline, and at 6, 12, 18, 24, 30, and 36 months post baseline. A dacron swab was passed 3 times over their inverted right upper conjunctiva, and processed for the presence of chlamydial DNA as previously described [15]. Our stochastic transmission model was fit to the estimated prevalence of infection at 6, 12, 18, 24, 30, and 36 months.
The study received ethical approval from institutional review board (IRB) of the Johns Hopkins University School of Medicine, the University of California San Francisco, and the Tanzanian National Institute for Medical Research, and was carried out in accordance with the Declaration of Helsinki. All subjects provided informed consent. The informed consent given was oral, because 1) verbal consent is the most ethical way to obtain consent, due to the high illiteracy rates in the study area, 2) IRB approved the use of the oral consent procedure for this study, 3) this oral consent is documented on the registration form for each study participant prior to examination in the field.
We constructed a stochastic transmission model of transmission of Chlamydia trachomatis infection over time [17], [18], [19], [20]. For community j (j = 1,…,32), we assumed a population of size Nj, taken from the number of pre-school children found in the census at the time of treatment (baseline, 12 months, or 24 months). We assumed a classical SIS (susceptible-infective-susceptible) model structure, assuming that the force of infection is proportional to the prevalence of infection in the population with proportionality constant β, and a constant per-capita recovery rate γ (month−1). Between periods of treatment, we assumed that the probability pi(t) that there are i infectives in the population obeyed the following equations:(1)andThese equations were applied to each village j, though we have suppressed a subscript j for clarity in Equation (1).
To estimate the transmission coefficient, we used data collected six months and twelve months after each treatment. The model was fit to each of three years. For comparing transmission rates, we initialized the model with observations taken six months after treatment, and we estimated the transmission parameter based on values observed six months after that. Thus, we modeled the time periods from 6 to 12 months, 18 to 24 months, and 30 to 36 months.
For each year, initial values for pi(t) were determined from as follows. From a population of size Nj of which the number Y of infectives equals i, the probability that s positives are observed from a sample of size Mj are sampled is given by the hypergeometric distribution: . We assumed a beta-binomial prior (where the shape parameters and were computed from the observed distribution of infection of 32 villages at 6 months, 18 months and 30 months, and B(x,y) is the beta function [21])) for all values of y. Application of Bayes' theorem yields(2)For each community j, we used the most recent census data to determine the community size Nj. The initial condition was determined from Equation (2), and the system numerically integrated for six months. Let Sj be the number of positive individuals detected in the sample at the end of the period (for community j). Given the number i of infected individuals, we computed the probability of the observed data according to (where Mj here denotes the sample size at the end of the period). We assumed independent communities, and thus we might maximize the sum of the logarithm of the above expressions (summing over all communities).
We assumed specific values of γ (see Table 1) and estimated the value of β that maximizes the total loglikelihood given . Standard errors were obtained from the observed Fisher information. We estimated the change, m (year−1), in the transmission coefficient per year by finding maximum likelihood estimates of β1 and m (with β2 = β1+m, β3 = β1+2m). For statistical comparison, we also used all time periods to estimate a single (constant) transmission coefficient for all three periods. The basic reproduction number is given by = β/γ; thus, estimated values of may be computed by dividing the estimated transmission coefficient by γ. The estimated annual change in the basic reproduction number, , may be computed by .
We also estimated an alternative model in which instead of varying the transmission coefficient over time, we instead assumed a constant transmission coefficient, and instead modeled the recovery rate in year i (i = 1,2,3) according to , where (month−1 year−1) is the annual change in recovery rate.
Previous models have estimated the duration of infection 1/γ to be from 3 months to 12 months [17], [19], [20]. As a base case, we assumed a mean duration 1/γ of 6 months; as a sensitivity analysis, we varied the mean duration from 3 to 18 months.
All calculations were performed using R (version 2.14.1, R Foundation for Statistical Computing, Vienna, Austria).
The numbers of 0–5 year-old children tested for the presence of ocular chlamydia were 3199 (baseline), 3198 (month 6), 3191 (month 12), 3200 (month 18), 3199 (month 24), 3194 (month 30), and 3153 (month 36). The estimated prevalence of ocular chlamydial infection by PCR at baseline was 22.0% (standard deviation 10.1%), at 6 months 10.5% (SD 4.7%), at 12 months 13.0% (SD 6.4%), at 18 months 7.1% (SD 4.4%), at 24 months 8.6% (SD 7.3%), at 30 months 3.5% (SD 2.5%), and at 36 months 4.7% (SD 3.3%).
Assuming a mean duration of infection of six months with beta-binomial prior (choosing the shape parameters to match the observed mean and variance of all villages cross sectionally at each initialization time), we found the basic reproduction number to be = 1.39 (95% CI: 1.28 to 1.49). For the first year, = 1.40 (95% CI: 1.26 to 1.55); for the second year, = 1.38 (95% CI: 1.09 to 1.67), and for the third year, = 1.35 (95% CI: 0.92 to 1.78).
The estimated change per year in the reproductive number, , was found to be −0.025, 95% CI −0.167 to 0.117, see Table 1. Similar findings were obtained when we assumed other values for the mean duration of infection: three months, = −0.041 (95% CI: −0.122 to 0.039) year−1; twelve months, = 0.012 (95% CI: −0.248 to 0.273) year−1; eighteen months, = 0.054 (95% CI: −0.326 to 0.434) year−1. Regardless of the assumed duration of infection, we find point estimates for the annual change in the basic reproduction number which are near zero. The confidence intervals are wider when a longer duration of infection is assumed, and these intervals include zero (i.e., no change) in every scenario.
Similar findings were obtained when a different choice of prior was used. Specifically, we assumed a uniform distribution as the prior distribution for the number of infected individuals; this yielded an overall = 1.30 (95% CI: 1.19 to 1.41) based on a pooled estimate assuming a constant β for all three years (i.e. assuming m = 0) and a mean duration of infection of six months. The corresponding estimate for the estimated change per year in the reproductive number is = −0.084 (95% CI: −0.227 to 0.058) year−1. Choosing other values of the mean duration together with the uniform prior similarly yielded the following results: three months, = −0.072 (95% CI: −0.152 to 0.009) year−1; twelve months, = −0.103 (95% CI: −0.366 to 0.161) year−1; eighteen months, = −0.118 (95% CI: −0.502 to 0.266) year−1.
We estimated the change in the recovery rate γ, assuming a constant transmission coefficient (the optimal value when infection duration was assumed to be 6 months). The estimated linear trend in recovery rate was found to be 0.013 (95% CI: −0.009 to 0.035) month−1 year−1, with the estimated recovery rate in the first year given by 0.177 (95% CI: 0.154 to 0.20) month−1 year−1. This model yields a substantially similar interpretation as the previous model.
Using a transmission model and data collected from a 32-community, cluster-randomized clinical trial in Tanzania, we found no evidence of increased transmission from the 1st through the 3rd year of treatment. In fact, our estimates of the reproduction number of the infection were very similar for each year, suggesting no loss of immunity.
Others have proposed an arrested immunity hypothesis, in which the development of protective immune responses is decreased as the duration of chlamydial infection is decreased. It has been suggested that an increased incidence of infection in the presence of a decreased seroprevalence in Finland and British Columbia is due to this phenomenon [10], [22]. Trachoma programs have offered an ideal setting to test this hypothesis. A study in Vietnam suggested that a single community with more antibiotic treatment had more rapid return of infection than another less treated community, and that this may be due to a loss of immunity [11]. Without a larger number of treated communities, it was not possible to assess whether the magnitude of this paradoxical result would have been expected by chance alone [23]. In this present study, the large number of longitudinally monitored communities offered more power, yet we were unable to document any evidence of increased transmission with more treatment.
There are several reasons this analysis of these data might fail to detect an increase in transmission, even if such an increase in fact exists. Two years may not be long enough for immunity to wane, although the period in which an earlier study suggested waning was only one year [11]. Uncertainty in the average duration of infection is a potential source of model misspecification, although a sensitivity analysis suggests that the reproductive number and estimated change are not sensitive to the value of infection duration. Even though the trial from which these data came is one of the larger trachoma studies performed, 32 communities may not be a large enough sample size to detect a modest increase in transmission. It is possible that a loss of immunity did occur, but that any effect on transmission was balanced by a decrease in transmission due to other factors; other studies have reported that per-infectious case transmission may decrease with decreasing prevalence, perhaps due to a decrease in the diversity of strains at lower prevalence [19], [24]. Finally, we have assumed that transmission is proportional to the number of infectious cases and number of susceptible cases (mass action); if this is not the case, then this may have masked increased transmission at the later, lower prevalence [19].
Models have predicted that if transmission per infectious case remains constant, repeated distributions can eliminate infection from even the most severely affected communities [2], [3]. Longitudinal studies have confirmed that local elimination is possible [5], [6], [25], [26], [27]. However, these successes might not hold in the future, if antibiotic resistance were to develop, or if a loss of immunity resulted in increased transmission. The absence of a short term increase in transmission as the prevalence decreases is good news for trachoma programs.
|
10.1371/journal.pntd.0004938 | West Nile Virus Temperature Sensitivity and Avian Virulence Are Modulated by NS1-2B Polymorphisms | West Nile virus (WNV) replicates in a wide variety of avian species, which serve as reservoir and amplification hosts. WNV strains isolated in North America, such as the prototype strain NY99, elicit a highly pathogenic response in certain avian species, notably American crows (AMCRs; Corvus brachyrhynchos). In contrast, a closely related strain, KN3829, isolated in Kenya, exhibits a low viremic response with limited mortality in AMCRs. Previous work has associated the difference in pathogenicity primarily with a single amino acid mutation at position 249 in the helicase domain of the NS3 protein. The NY99 strain encodes a proline residue at this position, while KN3829 encodes a threonine. Introduction of an NS3-T249P mutation in the KN3829 genetic background significantly increased virulence and mortality; however, peak viremia and mortality were lower than those of NY99. In order to elucidate the viral genetic basis for phenotype variations exclusive of the NS3-249 polymorphism, chimeric NY99/KN3829 viruses were created. We show herein that differences in the NS1-2B region contribute to avian pathogenicity in a manner that is independent of and additive with the NS3-249 mutation. Additionally, NS1-2B residues were found to alter temperature sensitivity when grown in avian cells.
| West Nile virus (WNV) is a mosquito-borne virus that has caused outbreaks in humans in many regions of the world. Birds are the natural hosts for WNV. However, different strains of WNV cause different disease outcomes in birds. Here, we compared two WNV strains, one of which causes higher mortality and generates more virus in American crows than the other. Previous research has shown that this difference is due in large part to a difference between the two strains at a single amino acid in the NS3 gene; however, this difference does not completely explain the observed effect. Here we show that another region of the viral genome also affects disease outcomes in American crows, and changes the sensitivity of the virus to temperature when grown in bird cells. These findings help us to understand the genetic features that affect WNV infection and disease outcomes in its natural host. Detection of such features in new strains of WNV and related viruses could help to understand and predict future outbreaks.
| West Nile virus (WNV) is the most widely distributed flavivirus in the world, occurring on all continents except Antarctica [1,2]. Recent human disease outbreaks in Europe and North America have brought increased scientific and public health attention to WNV; however, WNV may also cause significant underreported disease in developing countries [1,3–8]. Despite some advances, significant gaps remain in our knowledge of the ecological and genetic determinants of WNV transmission and disease.
WNV is maintained in avian reservoir hosts and is transmitted by Culex spp. mosquitoes [9,10]. Infection rates of mosquito vectors with WNV are proportionate to the virus titer in the infectious blood meal, with host sources generating titers below approximately 105 plaque-forming units (pfu)/ml sera considered to be poorly infectious to mosquitoes [11–14]. In contrast, birds of the family Passeridae can develop very high viremia titers, up to approximately 1010 pfu/ml in some corvids, and are considered to be the most relevant reservoir hosts that drive the force of epizootic/epidemic transmission [15–17].
For maximum transmissibility, WNV strains must be able to replicate at a variety of temperatures, from approximately 14°C external temperatures experienced by mosquitoes to 45°C body temperatures of febrile avian hosts [18–22]. Strains that cannot withstand the high temperatures experienced by febrile birds are expected to be at a competitive disadvantage for viremogenesis and subsequent transmission [18,23]. Indeed, flavivirus strains and mutants that are temperature sensitive (ts) in vitro are frequently also attenuated in vivo [23–27].
WNV, like other members of the Flavivirus genus, encodes a polyprotein that is post-translationally processed into three structural proteins (the capsid protein C and envelope proteins prM and E) and seven nonstructural proteins (NS1, NS2A, NS2B, NS3, NS4A, NS4B, and NS5). WNV is phylogenetically divided into at least five lineages, with the majority of circulating and epidemic strains belonging to lineage 1 [2]. WNV was isolated in the Americas for the first time in New York in 1999, and rapidly spread across the continent [28]. The NY99 strain, representative of the East Coast genotype of lineage 1a WNVs, has been extensively studied and is widely used as a model strain for WNV studies. The current strains circulating in North America represent a different genotype that is derived from the NY99 ancestor [29]. An alternative lineage 1a WNV strain that was isolated in Kenya, KN3829, shares a high genetic identity with NY99 with a total of 11 amino acid differences between the two strains (Table 1) (Genbank: AF196835 [NY99] and AY262283 [KN3829]).
American crows (AMCRs; Corvus brachyrhynchos) infected with North American strains of WNV exhibit high levels of mortality and high viremia titers [15,30–32]. In laboratory infection studies, WNV NY99 typically elicits a viremia of over 108 plaque forming units (pfu)/ml sera, and 100% mortality within approximately 6–7 days [15,18,26,30,33]. Due to their high susceptibility and visibility, AMCRs have been used as a sentinel species for WNV circulation in North America [34]. Despite the high genetic relatedness with NY99, KN3829 exhibits a strikingly different avian virulence phenotype, eliciting very low viremia and limited mortality in AMCRs [18,30,33]. Previous research has demonstrated that a single, positively-selected amino acid substitution at residue 249 in the NS3 helicase gene of WNV is strongly associated with virulence in AMCRs [30]. The NY99 strain encodes a proline residue at this position, while KN3829 encodes a threonine residue (Table 1). Introduction of an NS3-P249T mutation into the NY99 backbone reduced AMCR viremia by almost 106-fold, while the reciprocal mutation in the KN3829 virus (KN3829-NS3-T249P) increased viremia to a similar degree [26,30].
However, a residual difference in virulence between the two strains was not attributable to the NS3-249 amino acid difference. KN3829 and KN3829-NS3-T249P elicited approximately 10-fold lower viremia than NY99-NS3-P249T and NY99, respectively [30]. Mortality was also reproducibly lower with the NS3-249T mutant virus created in the NY99 backbone. Therefore, we hypothesized that other amino acid polymorphisms, or differences in the 3′ untranslated region (UTR), could account for this difference in pathogenesis and/or be associated with stabilization of the KN3829 virus. To test this hypothesis, we generated chimeric virus constructs between the infectious clones of NY99 and KN3829, and used the resulting viruses for evaluation of pathogenic potential in AMCRs and growth at standard and elevated temperatures in avian cell culture.
Infectious clones of NY99 and KN3829 were described previously [18]. To create chimeric constructs, we divided the viral genome into segments based on conveniently located restriction sites: NgoMIV at nucleotide (nt) 2495 (in NS1; used for ligation of the two-plasmid system during virus rescue); KpnI at nt 5341 (in NS3); KpnI at nt 7762 (beginning of NS5); and AatII at nt 10203 (end of NS5). Segments from the wild-type KN3829 and NY99 infectious clones, as well as the KN3829-NS3-T249P mutant virus, were interchanged using these restriction sites. Chimeric virus strains were named based on the KN3829-specific genome segments they contained (Fig 1). NS1-2B point mutations were created in the KN-IC (CG plasmid) infectious clone by site-directed mutagenesis as previously described [23].
Rescue of infectious clone-derived virus was described previously [18]. Briefly, the 5′ and 3′ plasmids of NY99, KN3829, and mutant and chimeric viruses were digested with NgoMIV, ligated, and linearized with XbaI (New England Biolabs) before in vitro transcription with the Ampliscribe High-Yield T7 Transcription kit (Epicentre Biotechnologies). Viral RNA was transfected into BHK-21 cells by electroporation. When >50% of cells displayed cytopathic effect, supernatant was harvested, centrifuged to remove cellular debris, and stored at -70°C until titration by plaque assay. RNA was extracted from stocks from individual clone-derived viruses and viral genomes were sequenced as described previously [26].
Vero, BHK-21, and duck embryonic fibroblast (DEF) cells were maintained in DMEM containing 10% FBS, 100 U/ml penicillin, and 50 μg/ml streptomycin. For determination of growth kinetics, DEF cells were inoculated with virus at an MOI of 0.1. After a one hour adsorption at 37°C, cells were washed three times with Dulbecco’s PBS (Life Technologies), growth medium was replaced, and cells were placed in incubators at either 37° or 44°C. Supernatant was sampled daily for five days. 30 μl of each sample was added to 270 μl of fresh medium containing 20% FBS, frozen, and stored for titration as above.
AMCR peripheral blood mononuclear cells (PBMCs) were isolated using Histopaque-1077 (Sigma-Aldrich) and maintained in RPMI containing 10% FBS, penicillin/streptomycin as above, and 1 μg/ml Fungizone (Life Technologies), as described previously [35]. PBMCs were inoculated with virus at an MOI of 10, incubated for one hour at 37°C or 42°C, and then centrifuged at 1500×g, resuspended in fresh growth medium, and incubated at the same temperature. One third of the well volume was sampled and replaced daily for six days, and samples were stored as described above.
After-hatch year AMCRs were trapped using cannon nets in Bellvue, Colorado between 2004–2007. Crows were banded, bled, and tested for pre-existing immunity to WNV and St. Louis encephalitis virus using plaque reduction neutralization tests as previously described [33]. AMCRs were housed at Colorado State University in groups of 2–3 in 1-m3 cages and fed an ad libitum mixture of dry dog and cat food. Groups of 16 AMCRs were inoculated subcutaneously with 1500 pfu of parental, chimeric, or point mutant WNV in a 100 μl volume. Inoculated AMCRs were bled by jugular venipuncture daily for seven days. Whole blood was diluted 1:10 in DMEM containing 10% FBS and penicillin/streptomycin. Blood samples were allowed to coagulate at room temperature before centrifugation for 10 min at 4000 × g, and were stored at -70°C until titration by plaque assay. AMCRs were monitored daily for 14 days and any birds displaying signs of WNV disease, such as ataxia, incoordination, or difficulty feeding, were euthanized by intravenous phenobarbital overdose. All surviving birds were euthanized at day 14 in the same manner.
RNA was extracted from selected samples of PBMC culture supernatant or AMCR blood using a Viral RNA mini kit (Qiagen) as described previously [26]. RT-PCR was performed using a SuperScript III One-Step RT-PCR kit (Life Technologies) and primers WNV5032F (5′-GGAACATCAGGCTCACCAATAGTGG-3′) and WNV5497R (5′-CTTTGTGGAAATGTAACCTCTTGCTGC-3′). The resulting RT-PCR product was sequenced with the same primers.
All statistical calculations were performed using GraphPad Prism v. 6.04 or R v3.2.2. Statistical analysis of in vivo data was performed by synonymizing groups based on NS1-2B genotype (NY or KN) and NS3-249 genotype (Pro or Thr). Survival curves were compared using a log-rank test. Viremia was regressed on dpi assuming polynomial trend and normal errors. The model includes a fixed effect for each modified region of the two viruses, and a random effect for replicates. Times at which peak viremia occurred were estimated from the fit. Standard errors for differences in peak viremia were computed using the delta method and incorporate uncertainty from estimating both time of peak viremia and value of peak viremia. Results were adjusted to account for multiple comparisons, achieving an overall Type I error rate of 0.05.
For temperature-sensitivity data in DEF cells, a semiparametric, mixed model was fit to the titer data. The model includes a fixed effect for each modified region of the two viruses, a random effect for replicates, and temperature-specific mean titer curves. The temperature-specific components were characterized by second degree penalized splines with truncated power basis. The solution to the fit and estimated variances were obtained by computing the best, linear, unbiased predictors of the penalized spline’s representation as a linear, mixed model [36]. Times at which peak titer occurred were estimated from the fit. Standard errors for differences in peak titer were computed using the delta method and incorporate uncertainty from estimating both time of peak titer and value of peak titer. Results were adjusted to account for multiple comparisons, achieving an overall Type I error rate of 0.05.
Trapping of AMCRs was performed under US Fish and Wildlife Scientific Collecting Permit MB-032526 and MB-082812. Birds were collected under US Fish and Wildlife Services and Colorado Parks and Wildlife permits with permission of private land owners as well as the managers of the Colorado State Fisheries Unit in Bellvue, CO. Field studies did not involve endangered or protected species. All animal studies presented herein were approved by Institutional Animal Care and Use Committees at the University of California, Davis (approval number 12874) and Colorado State University (approval number 10-2078A). All protocols and practices for the handling and manipulation of crows were in accordance with the guidelines of the American Veterinary Medical Association (AVMA) for humane treatment of laboratory animals as well as the ‘‘Guidelines to the Use of Wild Birds in Research” published by the ornithological council 3rd edition (2010).
We constructed chimeric virus constructs between NY99 and KN3829 to determine which differences between the two strains, other than the previously described NS3-249 site [26,30], contribute to avian virulence and pathogenesis (Table 1; Fig 1). All chimeric viruses could be grown in vitro in rodent (BHK-21) and primate (Vero) cell lines to titers comparable to those attained by the wild-type parental infectious clone viruses (at least 7 log10 pfu/ml).
AMCRs were inoculated with virus derived from infectious clones of WNV NY99, KN3829, or chimeric plasmids with proline or threonine residues present at the NS3-249 locus. As described previously [18,26,30,33], viremia in NY99-inoculated AMCRs peaked at approximately 109 pfu/ml serum, while KN3829 elicited only approximately 105 pfu/ml serum (Fig 2A). NY99 infection induced 100% mortality within six days post-infection (dpi), whereas 13/16 (81%) of AMCRs infected with clone-derived KN3829 virus survived to 14 dpi (Fig 2E).
As expected, viruses encoding a proline residue at the NS3-249 locus elicited higher viremia (95% CI for difference in peak viremia 0.3 logs to 2.0 logs) (Fig 2A and 2B) and mortality (Fig 2E and 2F) rates than those containing a threonine residue. However, within both Pro and Thr-containing groups, constructs containing the NS1-2B region of NY99 induced statistically higher peak viremia and mortality than those containing the NS1-2B region from KN3829 (Fig 2B and 2F). When groups were synonymized based on genotype at NS1-2B (NY or KN) and NS3-249 (P or T), survival distributions were significantly different (p < 0.001) among all groups (Fig 2F). The peak viremia was significantly different between all pairs of groups except KN/Pro and NY/Thr (Fig 2C and 2D). The structural genes of WNV did not have an apparent effect on pathogenesis in AMCRs. Mortality did not differ between strains that were identical with the exception of their structural genes (p > 0.05) (e.g. compare KN-str/KN-NS3-4B and NY-str/KN-NS3-4B). The difference in mean peak viremia titers between these strains was 0.31 logs (95% CI of -0.1 logs to 0.7 logs). Similarly, the 3′ UTR did not have a detectible effect on viremia or mortality (p > 0.05) (i.e. KN-str/KN-NS1-2B and KN-str/KN-NS1-2B 3’).
We hypothesized that, given the importance of the NS3-249 position for viral replication in AMCRs, infection with Thr-containing viruses may have imposed selective pressure, leading to potential mutations at this site. Therefore, viral RNA extracted from sera collected at 4 dpi from AMCRs infected with KN-str/KN-NS3-4B and NY-str/KN-NS3-4B (8 AMCRs each) was spot sequenced. These viruses were chosen because they grew relatively well in AMCRs compared to other Thr-containing constructs. Of the 16 samples sequenced, only three maintained a Thr residue at NS3-249 with no detectable mutations in the viral population. Eight had mutated to contain an alanine residue at this position. One had mutated to an asparagine residue. The other four contained mixed sequences at the locus. Two had a mixture of alanine and threonine, one had a mixture of alanine and proline, and the last sample contained a mixture of alanine, aspartic acid, and threonine. No other mutations were detected in the surrounding NS3 region.
To further analyze the effects of the differences between NY99 and KN3829, AMCR PBMCs were inoculated with the chimeric virus constructs. As described previously [35], replication in AMCR PBMCs at 37°C correlated with NS3-249 genotype (Fig 3A). However, when the temperature was increased to 42°C, the body temperature of AMCRs, only viruses containing both a proline residue and the NS1-2B region from NY99 were able to replicate (Fig 3B).
High variability was observed among the three replicate infections with the KN-str/KN-NS3-4B and NY-str/KN-NS1-4B viruses grown at 37°C. One replicate of the KN-str/KN-NS3-4B culture attained a final titer that was over 100-fold higher than the titers attained by the other two replicates (Fig 3C). Similarly, one replicate of the NY-str/KN-NS1-4B culture attained a titer at 5 dpi that was over 10-fold higher than the other two cultures. Viral RNA was isolated from these two culture supernatants and the region surrounding the NS3-249 region was sequenced [35]. No mutations were found in the NS3-249 residue. However, mutations were found in nearby residues in single high-titered replicates. Specifically, a NY-str/KN-NS1-4B sample contained a mixed population of wild-type and NS3-E251K mutant virus, while a KN-str/KN-NS3-4B sample contained a mixed population of wild-type and NS3-T246I mutant virus. The proximity of these mutations to the NS3-249 site, which modulates PBMC replication, suggests that they may be the cause of the improved growth in these replicates.
There are three amino acid differences between NY99 and KN3829 in the NS1-2B region (Fig 1). In order to determine the relative contribution of these three differences to the observed changes in AMCR virulence, we created single NS1-S70A, NS2A-A52T, and NS2B-A103V point mutants in the KN3829 infectious clone backbone. Inoculation of AMCRs with these point mutants led to viremia and mortality that were not distinguishable from wild-type KN3829 (Fig 5).
We show here that replication and virulence of WNV in American crows are modulated by the NS1-2B region of the genome in addition to the previously described effect of the NS3-249 residue. The effects of these two genomic regions are independent and additive in vivo. Although the effect of the NS1-2B region is relatively subtle (approximately 10-fold) compared with the effect of NS3-249, it is reproducible and statistically significant. This finding underscores the importance of the flaviviral nonstructural proteins for virulence and viral replication in the natural reservoir host.
As previously described, the NS3-249 site evidently modulates replication in leukocytes [35]. Viral constructs containing a threonine at this position consistently failed to replicate in PBMC culture, while those containing a proline replicated well. No other determinants detectably affected growth in PBMCs at 37°C. However, constructs containing the NS1-2B region from KN3829 were unable to replicate at 42°C. Thus, we conclude that the NS3-249 residue is a determinant of replication in AMCR PBMCs, while the NS1-2B region is a determinant of temperature sensitivity. As febrile AMCRs typically attain body temperatures above 42°C [18], this suggests a possible explanation for the decreased in vivo viremia observed in AMCRs infected with these constructs. The combination of leukocyte replication and temperature sensitivity effects may explain the relative in vivo virulence of the various constructs.
These results suggest that temperature sensitivity may play an important role in WNV pathogenesis in birds. Interestingly, in North American WNV strains, an NS1-K110N mutation, in combination with a mutation in NS4A, has been associated with in vitro temperature sensitivity in DEF cells [23]. Although these findings are not directly comparable to those shown here, they also point to a potential role for NS1 in mediating temperature sensitivity in WNV.
Interestingly, temperature sensitivity of WNV in DEF cells was not directly correlated with temperature sensitivity in AMCR PBMCs. Although previous work found a slight effect of the residue at position NS3-249 on temperature sensitivity in DEF cells in the NY99 genetic background, this effect was not evident in the KN3829 and chimeric genetic backbones assessed here [26]. Instead, the structural genes, NS1-2B region, and NS3-4B region exclusive of NS3-249 all appeared to modulate the differences in DEF cell temperature sensitivity between NY99 and KN3829. This is consistent with chemical mutagenesis studies in dengue virus, in which temperature sensitivity was conferred by mutations in a variety of positions throughout both the structural and nonstructural regions [37]. Temperature sensitivity of flaviviruses is evidently a complex phenotype that can be conferred by a variety of mutations, likely with different underlying mechanisms. These mechanisms may include protein stability and protein-protein interactions, among others. The difference between AMCR PBMCs and DEF cells also indicates that temperature sensitivity results may be dependent on the system used for testing. Although the use of DEF cells for temperature sensitivity testing is convenient, AMCR PBMCs are likely more phenotypically relevant.
None of the three individual amino acid differences in the NS1-2B region between NY99 and KN3829 could individually explain the effect of the overall region on temperature sensitivity or in vivo virulence. Thus, the overall effect of this region evidently requires two or more of the amino acid differences. This is consistent with previous results in a dengue-2 vaccine virus study, which showed that a combination of mutations at NS1-53 and NS3-250 was required to make the vaccine virus fully temperature sensitive [38]. The single NS3-250 substitution did not increase temperature sensitivity of the dengue vaccine virus, while the NS1-53 substitution alone only caused subtle temperature sensitivity. Future studies will be required to fully understand the effects of these NS proteins on pathogenesis and temperature sensitivity.
Alternatively, the synonymous nucleotide changes in the NS1-2B region could have an effect at the RNA level, which would not be captured by amino acid point mutations. An RNA secondary structure motif is required for the production of the frameshifted NS1′ protein in WNV and closely related flaviviruses [39]. Mutations that alter this secondary structure can change the ratio of full-length to frameshifted polyprotein, affecting WNV pathogenesis in mice and house sparrows [40,41]. Both NY99 and KN3829 encode the frameshift motif and would be expected to produce NS1′. There are two amino acid differences between the NS1′ coding sequences of NY99 and KN3829, one of which is encoded by the same nucleotide polymorphism that encodes the tested NS2A-A52T mutation. The role of the amino acid sequence of NS1′ is not well understood, and it is possible that the polymorphism not tested here could play a role in pathogenesis and temperature sensitivity. Other cryptic RNA motifs that have not yet been described could also play a role.
The functions of the flaviviral NS1, NS2A, and NS2B proteins are not fully understood, making it difficult to determine why these proteins apparently affect temperature sensitivity in PBMC culture and replication and virulence in vivo. The NS1 and NS2A proteins, in particular, have apparent roles in immunomodulation and immunopathogenesis, in addition to their roles in viral replication [42–47]. A silent mutation in WNV-Kunjin virus NS2A that affects the NS1′ frameshift motif also has been shown to alter interferon induction, and an amino acid change at the same position affects apoptosis in vitro and virulence in mice [41,48]. Given that the differences in avian pathogenesis observed here appear to be modulated at least in part by replication in immune cells, these immunomodulatory functions may be relevant. Further research on these nonstructural proteins will aid in understanding their role in temperature sensitivity and avian virulence.
Subtle effects of differences among viral strains could have an amplified effect on a larger scale. Although the addition of the NY99 NS1-2B region to virus backbones containing the NS3-249-Thr residue only increased peak viremia titers by approximately 100-fold, AMCRs infected with these viruses experienced viremia titers above 105 pfu/ml for 3–4 days. In contrast, AMCRs infected with the corresponding strains containing the KN3829 NS1-2B region experienced 0–1 days of viremia above 105 pfu/ml. As 105 pfu/ml is the approximate titer required for infection of mosquitoes, this relatively subtle difference could lead to an increased chance of transmission to a mosquito [11–14]. Furthermore, if these determinants in NS1-2B are present in non-North American or alternative lineage WNV strains, increased viremia titers could weaken the potential selective pressure for development of NS3-249P mutations.
These observations highlight the importance of understanding of the determinants of WNV replication and pathogenesis in relevant avian reservoir hosts, including the AMCR. Unraveling the viral genetic factors influencing the infection of different avian species will provide insight into emergence mechanisms of WNV and related flaviviruses. This behavior cannot be predicted based on studies of mammals such as mice, which exhibit physiological, immunological and cytological differences from birds that preclude use as a relevant model system for the selective pressure these viruses undergo during enzootic/epizootic transmission cycles.
|
10.1371/journal.pgen.1002062 | COL4A1 Mutations Cause Ocular Dysgenesis, Neuronal Localization Defects, and Myopathy in Mice and Walker-Warburg Syndrome in Humans | Muscle-eye-brain disease (MEB) and Walker Warburg Syndrome (WWS) belong to a spectrum of autosomal recessive diseases characterized by ocular dysgenesis, neuronal migration defects, and congenital muscular dystrophy. Until now, the pathophysiology of MEB/WWS has been attributed to alteration in dystroglycan post-translational modification. Here, we provide evidence that mutations in a gene coding for a major basement membrane protein, collagen IV alpha 1 (COL4A1), are a novel cause of MEB/WWS. Using a combination of histological, molecular, and biochemical approaches, we show that heterozygous Col4a1 mutant mice have ocular dysgenesis, neuronal localization defects, and myopathy characteristic of MEB/WWS. Importantly, we identified putative heterozygous mutations in COL4A1 in two MEB/WWS patients. Both mutations occur within conserved amino acids of the triple-helix-forming domain of the protein, and at least one mutation interferes with secretion of the mutant proteins, resulting instead in intracellular accumulation. Expression and posttranslational modification of dystroglycan is unaltered in Col4a1 mutant mice indicating that COL4A1 mutations represent a distinct pathogenic mechanism underlying MEB/WWS. These findings implicate a novel gene and a novel mechanism in the etiology of MEB/WWS and expand the clinical spectrum of COL4A1-associated disorders.
| Muscle-eye-brain disease (MEB) and Walker-Warburg Syndrome (WWS) are devastating childhood diseases that belong to a subgroup of congenital muscular dystrophies (CMDs) characterized by ocular dysgenesis, neuronal migration defects, and congenital myopathy. Genetic studies have revealed a number of genes involved in the etiology of CMDs, and subsequent studies show that alterations in dystroglycan glycosylation underlie MEB/WWS. However, over half of MEB/WWS patients do not have mutations in known genes encoding glycosyltransferases, suggesting that other genes are involved. Here, we describe a novel and genetically complex mouse model for MEB/WWS and identify putative heterozygous mutations in COL4A1 in two MEB/WWS patients. We identify a novel gene implicated in the etiology of MEB/WWS, provide evidence of mechanistic heterogeneity for this subgroup of congenital muscular dystrophies, and develop an assay to test the functional significance of putative COL4A1 mutations. Our findings represent the first evidence for a dominant mutation leading to MEB/WWS–like diseases and expand the spectrum of clinical disorders resulting from Col4a1/COL4A1 mutations.
| Congenital muscular dystrophies (CMDs) involving ocular and cerebral malformations are devastating childhood diseases. Fukuyama congenital muscular dystrophy, Muscle-Eye-Brain disease (MEB), and Walker-Warburg Syndrome (WWS) are clinically and mechanistically related forms of CMD [1]–[4]. Patients present at birth or as infants with muscle weakness, hypotonia or even severe myopathy leading to fatal respiratory insufficiency. Clinical presentation varies between patients but often includes myofiber necrosis and fibrosis, replacement by adipose tissue, split muscle fibers and the presence of non-peripheral nuclei. In addition, while most patients exhibit marked elevation of serum creatine kinase (CK), others have CK levels within the normal range [5].
Ocular and cerebral phenotypes demonstrate variable penetrance and expressivity between individual patients. For instance, ocular dysgenesis can occur in either, or both, the anterior (Peters' anomaly, Rieger syndrome, cataracts, buphthalmos and developmental glaucoma) and posterior portions of the eye (retinal dysplasia, retinal detachments and optic nerve hypoplasia) [6]–[8]. Variable neurological manifestations, including mental retardation and epilepsy, may be at least partially explained by cerebral cortical malformations including cobblestone lissencephaly, cerebellar hypoplasia, hydrocephalus and encephalocele.
Genetic and biochemical studies led to the identification of a number of genes involved in the etiology of CMDs and revealed that alterations in post-translational processing of dystroglycan underlie MEB/WWS [9]–[15]. In these ‘dystroglycanopathies’, hypoglycosylation of dystroglycan disrupts ligand binding and impairs muscle fiber attachment to the extracellular matrix. Central nervous system pathology is proposed to be secondary to defective interactions between radial glial cells and the pial basement membrane [16]–[18]. Despite these major advances, over half of MEB/WWS patients do not have mutations in known genes encoding glycosyltransferases [19], [20], suggesting that other genes in this pathway contribute to disease or that independent mechanisms are responsible. Identifying new pathways involved in MEB/WWS will be an important breakthrough that could open new avenues for understanding and ultimately treating CMDs.
We have recently discovered the first mutations in the gene coding for the ubiquitous basement membrane protein type IV collagen alpha 1 in mice (Col4a1) and humans (COL4A1) [21]. COL4A1 is the most abundant basement membrane protein and is ubiquitously present in basement membranes with few exceptions. The collagenous domain (a long stretch of Gly-Xaa-Yaa repeats that forms a triple helix) accounts for over 90% of the protein. Mutations in this triple helix-forming domain are well documented to be pathogenic in several types of collagens, including type IV collagens. The mutation we identified in mice (referred to as Col4a1Δex40) disrupts a splice acceptor site, causing exon 40 to be skipped and concomitant deletion of 17 amino acids from the collagen triple helical domain which interferes with proper folding, assembly and secretion of heterotrimeric COL4A1 and COL4A2 [21], [22]. To date, eleven out of twelve other Col4a1 mutations reported in mice [23], [24] and seventeen out of twenty one COL4A1 mutations identified in humans [21], [22], [25]–[34] are missense mutations within the triple helical domain, demonstrating that alterations of this domain are highly pathogenic.
In mice and humans, semi-dominant Col4a1/COL4A1 mutations are highly pleiotropic with variable expressivity. The tissue distribution and severity of pathology depends on genetic and environmental factors but commonly include cerebrovascular diseases, ocular and renal defects [21]–[24], [35]. Here, we broaden this spectrum to include MEB/WWS. We show that, depending on the genetic context; Col4a1+/Δex40 mice recapitulate the pathophysiological hallmarks of MEB/WWS, including ocular anterior segment dysgenesis, optic nerve hypoplasia, cortical lamination defects and myopathy. In addition, we identify heterozygous mutations in the triple–helix-forming domain of COL4A1 in two MEB/WWS patients. Together, these findings support COL4A1 mutations as a novel genetic cause of MEB/WWS. Importantly, COL4A1 is not directly related to post-translational modification of dystroglycan. We show that the mechanism is independent of dystroglycan glycosylation and instead is probably due to decreased COL4A1 levels in basement membranes. COL4A1 mutations are pleiotropic and these data describe another clinically distinct group of diseases that can result from alterations in a single gene.
Ocular hallmarks of MEB/WWS include anterior segment dysgenesis and optic nerve hypoplasia. Depending on the genetic context, mutations in Col4a1 can cause both anterior segment dysgenesis and optic nerve hypoplasia although the underlying pathogenic mechanism(s) remain unexplored [23], [24], [35]. During development, COL4A1 is present in the inner limiting membrane of the retina [36] and inner limiting membrane disruptions can perturb retinal ganglion cell (RGC) localization and cause apoptotic death [37]. To determine if excess RGC death during development caused optic nerve hypoplasia, developing retinas were immunolabeled for Islet-1 (ISL1) and Laminin, which mark newly specified RGCs [38] and basement membranes, respectively. At embryonic days (E) 14, E16 and E18, RGCs were located in the innermost part of the retina in Col4a1+/+ mice, with only occasional ISL1 immunoreactivity detected in the outer retina (Figure 1D). In contrast, in Col4a1+/Δex40 mice, the thickness of the ISL1 positive layer was highly variable and there were more displaced ISL1 positive cells detected in the outer retina (arrows in Figure 1H and 1L). Laminin immunoreactivity revealed focal disruptions in the inner limiting membranes of Col4a1+/Δex40 animals (asterisks in Figure 1F) that were not observed in Col4a1+/+ animals. Moreover, in contrast to Col4a1+/+ eyes where the vasculature is closely associated with the inner limiting membrane, the hyaloid vasculature in Col4a1+/Δex40 eyes is most often found in the vitreous (Figure 1B and 1F) and, in one extreme case, the posterior chamber was devoid of detectable vasculature (Figure 1J).
During normal retinal development, approximately 50% of RGCs undergo programmed cell death as the visual system matures [39], [40]. Induced or genetic disruption of the inner limiting membrane perturbs RGC localization and leads to RGC apoptosis during embryogenesis [37]. Therefore, we hypothesized that Col4a1+/Δex40 mice might exhibit increased RCG apoptosis. To test this, we co-labeled retinal sections with antibodies against ISL1 and activated Caspase-3 (Figure 2A–2F) and calculated the apoptotic index by counting the number of ISL1/Caspase 3 double-labeled cells (Figure 2G). As retinas from Col4a1+/Δex40 mice were smaller than retinas from Col4a1+/+ mice (Figure 2H), we normalized the number of double-labeled cells to the retinal cross-sectional area. Col4a1+/+ mice had low levels of ganglion cell apoptosis at E14 and E16 that increased approximately 2-fold by E18. Although apoptotic rates in Col4a1+/Δex40 mice at E14 and E16 were not significantly different from those observed in Col4a1+/+ mice, there was a significant increase in apoptosis of ISL1 positive cells in Col4a1+/Δex40 eyes compared to Col4a1+/+ eyes at E18 (p<0.05) – especially among mislocalized ISL1 labeled cells (Figure 2F). Of note, one E18 Col4a1+/Δex40 eye had an extremely high apoptotic index (12.9) that was considered an outlier and was removed from subsequent statistical analyses (see Materials and Methods). Interestingly, we did not observe vasculature in the posterior chamber of this eye suggesting that retinal vasculature might directly contribute to the inner limiting membrane or affect ganglion cell viability in other ways. Together, our data support that Col4a1 mutation leads to focal disruptions of the inner limiting membrane and that optic nerve hypoplasia results both from reduced production of retinal neurons and from mislocalization and subsequent apoptosis of ganglion cells during development.
Based on our observations in the retina, we predicted that Col4a1+/Δex40 mice might also show pial basement membrane disruptions and cerebral cortical lamination defects that model cobblestone lissencephaly seen in MEB/WWS. To test our hypothesis, we stained coronal brain sections from adult Col4a1+/+ and Col4a1+/Δex40 mice with cresyl violet and detected abnormalities in all of the Col4a1+/Δex40 but none of the Col4a1+/+ mice examined (Figure 3). All Col4a1+/Δex40 mice had focal and variable cerebral cortex lamination defects ranging from mild distortions and ectopias to severe heterotopic regions devoid of obvious lamination (Figure 3B–3F). Occasionally, Col4a1+/Δex40 mice displayed enlarged ventricles or major structural abnormalities (Figure 3G and 3H). Immunolabeling with the pan–neuronal marker NeuN confirmed the neuronal identity of the ectopic cells (Figure 4). Col4a1+/Δex40 mice also had subtle but consistent defects within the hippocampus (Figure S1). The CA1, CA3 and dentate gyrus layers of Col4a1+/Δex40 mice were less tightly organized and generally more dispersed compared to Col4a1+/+ mice and local perturbations were common. As it is the case in other animal models of MEB/WWS [41], [42], enhanced glial fibrillary acid protein (GFAP) immunoreactivity, which is reflective of astrocytic gliosis, was observed in the hippocampus and cerebral cortex of Col4a1+/Δex40 mice (Figure S2).
To determine if cortical malformations were congenital or acquired, we used bromodeoxyuridine (BrdU) pulse labeling in utero to evaluate the localization of neurons that underwent terminal cell division during defined stages of embryogenesis. The locations of cells labeled at E14 or E16 were determined at birth (P0). In all Col4a1+/+ mice, BrdU-labeled neurons were uniformly distributed primarily in the superficial cortex (Figure 5A and 5F). In contrast, the distribution of BrdU-labeled cells in mutant animals demonstrated that cortical lamination was disorganized. Focal and variable lamination defects were completely penetrant in Col4a1+/Δex40 mice (Figure 5B–5E and 5G–5J). Laminin labeling of basement membranes in P0 mice revealed discontinuous pial basement membranes in all mutant mice, notably in areas adjacent to ectopias (Figure 6). Together, these findings demonstrate that Col4a1+/Δex40 mice have abnormal neuronal localization typical of cobblestone lissencephaly observed in MEB/WWS patients and suggest that these congenital defects are secondary to breaches in the pial basement membrane.
Because Col4a1+/ex40 mice display ocular and cerebral abnormalities characteristic of MEB/WWS, we hypothesized that they would also have myopathy. To test this, we first confirmed that COL4A1 was present in skeletal muscle basement membrane by immunolabeling (Figure S3) and performed functional, biochemical and histological analyses of muscles from young and aged Col4a1+/+ and Col4a1+/Δex40 mice. At 3 months of age, Col4a1+/Δex40 mice performed significantly worse than controls in a test of peak grip force (Figure 7A). Next, we analyzed serum CK activity before and after exercise. We found no significant difference in CK levels between control and mutant mice at baseline, however, Col4a1+/Δex40 mice had a significant elevation in CK activity following exercise compared to pre-exercise Col4a1+/Δex40 and post-exercise Col4a1+/+ mice (Figure 7B). Consistent with these functional and biochemical data, we also detected histological differences between Col4a1+/+ and Col4a1+/Δex40 muscles. Compared to Col4a1+/+ littermates, Col4a1+/Δex40 mice had occasional split muscle fibers and a significant increase in the number of non-peripheral nuclei – a measure of myopathy (Figure 7C–7E). Importantly, the severity of myopathy was not markedly affected by age.
We have shown previously that ocular dysgenesis is genetic context–dependent and that mutant F1 progeny of C57BL/6J and CAST/EiJ crosses (CASTB6F1) are morphologically rescued [35]. As shown in Figure 7C, in contrast to what is observed on the C57BL/6J background, the number of non-peripheral nuclei was not increased in CASTB6F1 Col4a1+/Δex40 mice compared to their Col4a1+/+ littermates. Moreover, there were significantly fewer muscle fibers with non-peripheral nuclei in CASTB6F1 Col4a1+/Δex40 mice compared to C57BL/6J Col4a1+/Δex40 mice (p<0.05), indicating that the CAST/EiJ strain has one or more loci that can also ameliorate Col4a1-induced myopathy.
Col4a1+/Δex40 mice display multiple hallmarks of MEB/WWS raising the possibility that COL4A1 mutations might cause CMD-like diseases in human patients. To test this, we performed direct sequence analysis of genomic DNA from a cohort of 27 patients with CMD (see Table S1 for primers). Fifteen patients had diagnoses of WWS, two had diagnoses of MEB, and nine were not specifically classified as either but had CMD with variable ocular and cerebral involvement. To enrich for patients that may have dominant or semi-dominant mutations, rather than recessive mutations, most of the patients (23 out of 27) were chosen from non-consanguineous families. Finally, all patients were negative for mutations in genes currently known to underlie MEB/WWS-like diseases including LARGE, POMT1, POMT2, POMGNT1, FKTN and FKRP.
We identified several coding and non-coding sequence variants in COL4A1 (Tables S2, S3, S4). Twelve coding variants were silent and were either observed in controls or were not predicted to be splice–site-altering variants [43]. We identified four non-synonymous coding variants. Two of the four non-synonymous variants were previously identified SNPs that were highly polymorphic in patients and in controls and therefore deemed unlikely to be pathogenic. The two remaining SNPs were rare, missense variants that have not been previously reported in dbSNP and were heterozygous in independent patients. The first mutation was identified in a patient diagnosed with WWS that had lissencephaly, hydrocephalus, Dandy-Walker malformation, optic nerve hypoplasia and was hypotonic with CMD (see Figure S4 and Text S1 for clinical details). An adenine to guanine transition (A3046G) in exon 36 substituted a methionine residue for a valine residue within the triple helical domain at amino acid number 1016 (p.M1016V) and was not detected in 282 control chromosomes (Figure 8A).
The second mutation was identified in a patient with unspecified CMD with ocular and cerebral involvement. The patient had mild gyral abnormalities, hydrocephalus, retinal dysplasia, seizures and elevated CK (see Figure S4 and Text S1 for clinical details). A cytosine to guanine transversion (C3946G) in exon 44 substituted a glutamine residue for a glutamate residue within the triple helical domain of the protein at amino acid number 1316 (p.Q1316E) (Figure 8A). Although this cytosine to guanine transversion was not detected in 286 control chromosomes, it was present in the paternal DNA.
The first variant, COL4A1M1016V, is a mutation of a methionine residue in the Y position of the Gly-Xaa-Yaa repeat in the triple helix-forming domain. This amino acid is highly conserved and the analogous residue is a methionine in all vertebrate orthologues analyzed (Figure 8B) and in the COL4A1 paralogues COL4A3 and COL4A5 (data not shown). Moreover, there is precedence for a Y position methionine to valine mutation in COL4A5 causing Alport syndrome [44]. The second variant, COL4A1Q1316E, is a mutation of a glutamine in the Y position residue within the triple helix-forming domain. This glutamine residue is conserved in most vertebrates (Figure 8B). There is a strong preference for basic residues in the Y position of the Gly-Xaa-Yaa repeat and the mutation represents a substitution from a neutral amino acid to an acidic amino acid. Pathogenic mutations within the collagenous domain often perturb triple helix assembly and mutations in regions of low thermal stability are predicted to be more disruptive [45]. According to an algorithm that predicts collagen stability [46], both mutations modestly reduce the thermal stability of their respective region (Figure 8C) although the biological relevance of this is equivocal.
Pathogenic mutations in the triple helix forming domain of several types of collagens impair secretion of the collagen heterotrimers and concomitantly, misfolded proteins accumulate within cells. To assess the functional significance of the COL4A1M1016V and COL4A1Q1316E mutations, we developed an assay to test the impact of the mutant proteins in a human cell line. We stably transfected HT1080 cells with wild–type or mutant COL4A1 cDNAs and determined the relative levels of intracellular and secreted COL4A1. To validate the assay, we tested the effect of the COL4A1G562E mutant allele that is established to cause familial small vessel disease in human patients [22], [26]. As we predicted, when compared to HT1080 cells transfected with wild–type COL4A1, significant intracellular accumulation of COL4A1 and concurrent decrease in secreted COL4A1 was observed in cells transfected with the COL4A1G562E mutant cDNA (Figure 8D). The two putative pathogenic mutations were functionally tested using the same assay. Overall, the ratio of secreted/intracellular COL4A1 for the COL4A1M1016V mutation was reduced, however the results were variable (n = 9) and did not reach statistical significance. In contrast, and similar to the established COL4A1G562E mutation, the COL4A1Q1316E mutation clearly impaired COL4A1 secretion leading to intracellular accumulation (p<0.001), supporting the hypothesis that the COL4A1Q1316E mutation is pathogenic. Importantly, the accumulation of both the monomeric and heterotrimeric forms of COL4A1 suggests that the COL4A1G562E and COL4A1Q1316E mutations impaired triple helix assembly and/or stability.
Biochemical analyses revealed that known MEB/WWS-causing mutations act via hypoglycosylation of dystroglycan [11]. Although COL4A1 is not directly involved in post-translational dystroglycan modification, misfolded COL4A1Δex40 proteins in the endoplasmic reticulum (ER) might indirectly impair dystroglycan post-translational modification. Similarly, ER stress can produce reactive oxygen species [47]–[50] and exposure to reactive oxygen species can result in dystroglycan de–glycosylation [51]. Thus, Col4a1-induced pathogenesis might still act indirectly via dystroglycan hypoglycosylation or de–glycosylation. To test for alterations in dystroglycan expression and/or post-translational processing, we performed Western blot analysis on wheat germ agglutinin (WGA)-enriched skeletal muscle extracts and compared the amount and mobility of α– and β– dystroglycan between Col4a1+/+ and Col4a1+/Δex40 mice using a polyclonal β–dystroglycan antibody and the IIH6 α–dystroglycan antibody that recognizes the fully glycosylated, functional form of α–dystroglycan [11]. Immunoreactivity to, and mobility of, the precursor dystroglycan protein (WGA-unbound fraction) and α– and β– dystroglycan (WGA-bound, glycoprotein enriched fraction) were indistinguishable between Col4a1+/+ and Col4a1+/Δex40 mice (Figure 9). Consistent with this finding, immunolabeling of muscle sections for α– and β– dystroglycan showed similar expression patterns in the sarcolemma membrane of Col4a1+/+ and Col4a1+/Δex40 mice. These data indicate that dystroglycan expression, localization and post-translational modification are not altered in Col4a1+/Δex40 mice and that myopathy arises via disruption of the basal lamina.
In this study, we show that COL4A1 mutations cause multiple pathophysiological hallmarks of MEB/WWS in mice and possibly MEB/WWS in human patients. Mice harboring a semi-dominant Col4a1 mutation have ocular dysgenesis, cerebral cortical lamination defects and myopathy. Optic nerve hypoplasia results, at least in part, from mislocalization and increased apoptosis of RGCs during development. Cerebral cortical malformations range from subtle lamination defects to large structural abnormalities. Myopathy is mild, but consistent and significant, and is exacerbated by exercise but not by age. Importantly, myopathy was genetic background-dependent, which implies that in resistant genetic contexts Col4a1- induced myopathy might not be detected but that on permissive genetic backgrounds, myopathy could be severe. Two out of 27 MEB/WWS patients tested had COL4A1 mutations and in at least one case the mutation interfered with protein secretion. Moreover, unlike all other known causes of MEB/WWS, the cellular mechanism underlying Col4a1-induced pathology is independent of dystroglycan post-translational modification.
Until now, the precise pathogenic mechanism resulting from COL4A1 mutations in human patients was poorly characterized and was only presumed to impair protein secretion and cause intracellular accumulation. Here, we have established and validated a cellular assay that demonstrates this effect. However, the relative contribution of COL4A1 deficiency in the basement membrane versus toxic intracellular accumulation to pathogenesis is difficult to dissociate and other possible mechanisms exist. Although the ratio of secreted to intracellular protein for the COL4A1M1016V mutation was not statistically different from wild–type there was an overall reduction in the ratio. It is important to bear in mind that absence of statistical significance does not necessarily imply absence of biological relevance. For instance, more efficient intracellular degradation of some misfolded proteins could prevent detection of an altered ratio of secreted to intracellular COL4A1 for some mutations. Alternatively, these data could indicate a third pathogenic mechanism whereby mutant proteins are secreted and exert a detrimental extracellular effect. In support of such a mechanism, the COL4A1M1016V mutation occurs within a putative binding site for the matricellular glycoprotein SPARC (secreted protein acid and rich in cysteine) [52] and could therefore act by disrupting protein-protein interaction in the extracellular matrix.
These variants represent the first heterozygous mutations in MEB/WWS patients. Although generally considered to be recessive, the inheritance pattern in many MEB/WWS patients is unknown. Phenotypic variability, reduced penetrance, and de novo mutations can all explain how dominant mutations might appear recessive in small families. Here, we observed non–penetrance of the COL4A1Q1316E mutation in the father of the affected child. Families with COL4A1 mutations are only now starting to be identified but apparent non-penetrance, and asymptomatic carriers have already been reported [29], [53]. Importantly, Col4a1-associated phenotypes in mice are not only influenced by environmental factors but are also genetic context–dependent, and reduced penetrance could reflect modifier loci. Alternatively, genetic mosaicism of dominant collagen mutations can explain asymptomatic carriers [54].
We identified mutations in two out of 27 CMD patients tested. DNA was limiting and prevented us from also sequencing COL4A2; however, evidence from mice and C. elegans support that mutations in COL4A2 may cause phenotypes similar to those resulting from mutations in COL4A1 [24], [45], [55]. It will be very important in the future to determine whether the COL4A1 findings extend to COL4A2. This could have an even broader relevance as COL4A1 mutations are pleiotropic with variable penetrance and expressivity from organ to organ. It is possible that COL4A1 mutations underlie CMD in patients with different, non-MEB/WWS-like, subclasses of the disease. Based on our current data, the most likely patients are those with mild myopathy and those with clinical manifestations that overlap with some of the other COL4A1-related phenotypes described in non-MEB/WWS syndromes including porencephaly, renal disease and cerebrovascular disease. Collectively, these findings provide the impetus for COL4A1 and COL4A2 mutation analysis in further cohorts of patients with CMD and/or congenital cerebral malformations.
COL4A1 mutations have been identified in families with a spectrum of diseases affecting the cerebral vasculature, although pathologies of the eyes, kidneys and muscles have also been reported [21], [22], [25]–[29], [31], [33], [56], [57]. Notably, COL4A1 appears to be an important genetic cause of porencephaly – a condition usually diagnosed in infants and characterized by large cerebral cystic cavities that communicate with the ventricles. Importantly, the cortical malformations present in MEB/WWS and described in the current manuscript are clinically and mechanistically distinct from porencephalic cavities, which are predicted to result from pre– or peri–natal hemorrhages in the germinal matrix. Col4a1 mutations are pleiotropic and our findings expand the phenotypic spectrum resulting from Col4a1 mutations in mice. We propose that COL4A1 mutations in human patients will reflect the pleiotropy observed in mice and will be involved in the pathogenesis of diseases clinically distinct from those reported previously. Optic nerve hypoplasia and cortical neuronal migration defects have not been described in human patients with COL4A1 mutations and a role for COL4A1 or COL4A2 in MEB/WWS-like diseases was not previously suspected. Interestingly, six families with COL4A1 mutations and cerebrovascular disease also had muscle cramps and CK elevation [28], [33]. Although the degree of CK elevation in Col4a1 mutant mice is less than that observed in dystrophin-glycoprotein complex related disorders, it is comparable to that seen in other types of dystrophies involving extracellular matrix proteins, including collagen VI -associated Bethlem myopathy and Ullrich CMD. While the exact mechanism underlying myopathy in collagen VI-related disorders is unclear, several lines of evidence suggest that mitochondrial dysfunction may be an important player [58].
We show that genetic context is an important factor contributing to the variable penetrance and severity of Col4a1-related diseases and that the CAST/EiJ strain can modify myopathy. These findings also imply that certain genetic contexts might exacerbate myopathy and that COL4A1 mutations could also cause severe CMD. Importantly, genetic modifiers can be general or tissue specific and tissue-specific modification could help explain how mutations in a single gene can contribute to such diverse phenotypes. We also demonstrate that gene–environment interactions contribute to phenotypic variability. We have previously reported that cesarean delivery can reduce the risk of perinatal intracerebral hemorrhages and here we show that elevation in CK activity is exercise–dependent. Allelic differences could also help explain variable expressivity and severity of COL4A1-associated diseases between patients. For instance, a mutation that affects heterotrimer assembly might have broader effects than a mutation that specifically affects interactions with cell surface receptors, growth factors or other extracellular matrix molecules. Interestingly allelic differences are already emerging pointing to genotype/phenotype correlations in some families [28], [33].
Pial basement membrane integrity is critical for normal cortical development and insufficient COL4A1 in basement membranes renders them prone to disruption. Breaches in the pial membrane cause alterations in cortical neuronal distribution [17], [59], [60]. These are not intrinsic neuronal migration defects but are secondary to disorganization of the cortical marginal zone, and to defects in anchorage of glial endfeet and formation of the Cajal-Retzius cell layer [16]. Reactive gliosis and increased GFAP labeling may reflect broader breaches of the glia limitans including the blood brain barrier [11], [42]. Thus, while some structural defects might be due to disruptions of the pial basement membrane, other pathology might be attributed to disruptions of vascular basement membranes and/or defects in blood brain barrier function.
Understanding the precise mechanism underlying myopathy is complicated by the presence of COL4A1 in basement membranes of the sarcolemma, myotendonous junctions and neuromuscular junctions. Notably, Col4a1 mutant mice have transient neuromuscular junction abnormalities that reportedly resolve by 3 weeks of age [61]. Determination of the primary site of pathogenesis for myopathy will likely require conditional expression of the Col4a1 mutation. However, evidence from other model systems suggests that the primary mechanism is load-induced muscle fiber detachment. For instance, the most common form of CMD (MDC1A) is caused by mutations in the basement membrane component laminin alpha 2 (LAMA2) [62] and in zebrafish with mutations in the LAMA2 ortholog, muscle contractions lead to muscle fiber detachment and loss [63]. This mechanism is consistent with exercise-induced CK elevation in Col4a1 mutant mice. This mechanism is also consistent with the observation that C. elegans Col4a1 mutants die with ruptured basement membranes shortly after muscle contractions begin [45], [55]. A similar mechanism in alveolar basement membranes could also explain why newborn Col4a1 mutant mice have respiratory distress immediately after starting to breathe [22].
Thus, myopathy resulting from mutations in basement membrane components LAMA2 and COL4A1 might share a common pathogenic mechanism whereby contraction-induced load leads to muscle fiber detachment. Given this potential shared mechanism and the phenotypic variability of COL4A1 mutations, we propose that MDC1A patients that are negative for LAMA2 mutations are suitable candidates for screening for mutations in COL4A1 and COL4A2. Importantly, if a component of the pathology is secondary to deleterious consequences of compromised blood brain barrier and load-induced myopathy, there is the potential for therapeutic interventions to blunt the severity of this devastating disease. For example, conditions that promoted protein folding and increased COL4A1 secretion in C. elegans mutants were able to rescue muscle contraction-induced basement membrane disruptions and promoted viability and survival [55]. Thus, chemical chaperones, or other methods to promote protein folding, might have therapeutic potential in human patients and improve the prognosis for MEB/WWS patients.
Heads were equilibrated in 20% sucrose in phosphate buffered saline (PBS) overnight at 4°C, embedded in OCT compound (Sakura Finetek), and flash frozen using dry ice/ethanol. For each genotype we analyzed six eyes at E14 and E16 and eighteen eyes at E18. For each eye, at least 3 position-matched, transverse sections (20 µm) were co-labelled with antibodies against Islet-1 (1∶10, Developmental Studies Hybridoma Bank 39.4D5sup and laminin (1∶1000, Abcam), or Islet-1 and activated Caspase-3 (1∶1000, R&D Systems). Sections were blocked (Tris-buffered saline with 0.1% Triton-X (TBST) containing 5% normal goat serum (Invitrogen)) for 1 hour at room temperature (RT) and washed twice for 2 min with TBS. Labeling was performed in TBS containing 3% BSA (Sigma) overnight at 4°C. Sections were incubated with Alexa Fluor 488 goat anti-mouse and Alexa Fluor 594 goat anti-rabbit secondary antibodies (1∶1000, Invitrogen) for 2 hours at RT and cover-slipped with Vectashield Hard Set with DAPI (Vector Labs).
Adult brains were fixed by trans-cardiac perfusion with 4% paraformaldehyde (PFA) in 0.1 M sodium phosphate, pH 7.4 then post-fixed in PFA at 4°C overnight and prepared for cryosectioning as described above. For Nissl staining, coronal sections (25 µm) were hydrated for 5 min in buffer (0.2% sodium acetate, 0.3% glacial acetic acid) then incubated for 10 min in Nissl stain (0.02% cresyl violet in buffer). Sections were washed in water (2×2 min) and de-stained for 15 sec (0.3% glacial acetic acid in 70% ethanol) before being dehydrated, cleared in xylene and cover-slipped with Permount (Fisher).
For GFAP and NeuN immunostaining, sections were incubated in 1% H202 in PBS for 15 min, washed in PBS, blocked in PBS containing 5% NGS, 0.1% Triton-X, and 1% BSA in PBS for 1 hour at RT. Sections were incubated with anti-GFAP antibody (1∶500, Chemicon) or anti-NeuN antibody (Chemicon 1∶500) overnight at 4°C, washed in PBS containing 0.1%Triton-X, and incubated for 2 hr with biotinylated goat anti-rabbit or goat anti-mouse antibody (1∶500, Vector Labs). Sections were then incubated in avidin-biotin solution (Vector Labs) for 90 min and immunoreactivity was visualized by treating sections with diaminobenzidine (DAB, Vector Labs). Sections were dehydrated, cleared in xylene and cover-slipped with Permount (Fisher).
For BrdU labeling, pregnant mice were injected with BrdU (50 mg/kg) at gestational days 14, or 16. At P0, heads were fixed in 4% PFA for 4 hr, embedded in OCT, snap frozen in dry ice/ethanol bath. Cryosections (25 µm) were incubated in sodium citrate buffer (10 mM sodium citrate, pH 6.0, 0.05% Tween-20) for 30 min starting at 90°C and allowed to cool, rinsed in water, incubated in 1 M HCl containing 0.2 mg/ml pepsin for 10 min at RT, in 2 N HCl for 20 min at 37°C. Sections were then washed and labeled with anti-BrdU (1∶30, Accurate Chemical) and biotinylated goat anti-rat secondary (1∶500, Vector Labs). For laminin immunolabeling of P0 brains, heads were processed as described in the retinal analyses section and 20 µm coronal sections were immunolabeled with anti- laminin antibody (1∶1000, Abcam).
Peak grip force was determined using a Grip Strength Meter (Columbus Instruments) and using the average from 3 consecutive trials on each animal. For CK measurements, blood was drawn before and after exercise from the tail vein in a hematocrit tube, centrifuged for 5 min, and the serum was collected. CK activity was measured using CK-NADP assay (Raichem) and a microplate reader (Biorad). Exercise was 30 minutes with a 15° downhill grade on a treadmill equipped with a shock plate (Columbus Instruments). Animals were started at 7 m/min and increased 3 m/min every 2 min until maximum of 16 m/min speed was reached. Non-peripheral nuclei were evaluated in quadriceps and tibialis anterior muscles dissected and frozen in liquid nitrogen-cooled isopentane. Cryosections (8 µm) were stained with hematoxylin and eosin for histopathology or labeled with anti-pan laminin antibody (Sigma), followed by an anti-rabbit AlexaFluor-488 secondary antibody, and the nuclei were labeled with DAPI. The observers were masked to genotypes and counted between 2000–5000 fibers per animal.
For dystroglycan enrichment and immunoblotting, quadriceps muscle biopsy (100 mg) was homogenized in TBS (pH 7.5) containing 1% Triton-X and protease inhibitors (Thermo Scientific Pierce). The soluble fraction was incubated overnight at 4°C with 200 ul of wheat germ agglutinin (WGA)-agarose beads (Vector Laboratories) to enrich for glycosylated proteins. Beads were washed three times with 1 ml TBS containing 0.1%Triton-X and protease inhibitors (Thermo Fisher) and were boiled in SDS-PAGE loading buffer for 5 minutes. Proteins (50 µl) were separated on 10% SDS-PAGE and transferred to polyvinylidene fluoride (PVDF) membranes (BioRad). Membranes were blocked for 2 hours at room temperature in 5% non-fat milk in TBS containing 0.1% Tween-20, incubated overnight with a mouse monoclonal anti- α-dystroglycan antibody recognizing the glycosylated form of α-dystroglycan (IIH6; 1∶1000, Millipore), or rabbit polyclonal anti- β-dystroglycan antibody (raised against the C-terminus of dystroglycan precursor; 1∶400, Santa Cruz) diluted in blocking buffer. Membranes were washed in TBS containing 0.1% Tween-20, incubated 2 hours at room temperature with horseradish peroxidase-conjugated secondary antibody raised in donkey (anti-mouse IgM, and anti-rabbit IgG, respectively; 1∶10 000, Jackson Immunoresearch) diluted in blocking buffer. Immunoreactivity was visualized using chemiluminescence (SuperSignal West Pico Chemiluminescent Substrate, Thermo Scientific).
For immunostaining, muscle cryosections (25 µm) were incubated with either β-dystroglycan (1∶150) or with α-dystroglycan (1∶400) antibody diluted in blocking buffer (10% normal donkey serum, 0.2% BSA in PBS containing 0.1% Triton-X) overnight at 4°C and incubated with Alexa Fluor 488 conjugated donkey anti-rabbit (Invitrogen) or cy5- conjugated donkey anti-mouse IgM (Jackson Immunoresearch).
For COL4A1 immunolabeling, muscle cryosections (25 µm) were fixed in acetone for 10 min, rinsed in TT buffer (50 mM Tric-HCl, pH 7.4, 0.1% Tween-20), incubated in acid solution (0.1 M KCl/HCl, pH 1.5), washed three times in TT buffer and incubated for an hour in blocking buffer (10% normal donkey serum and 2 mg/ml BSA in Tris buffer). Sections were then incubated with rat anti- COL4A1 (H11) monoclonal antibody (1∶200, Shigei Medical Research Institute, Japan) diluted in TT buffer overnight at 4°C and for 1 hour with Alexa Fluor 488 conjugated donkey anti-rat secondary antibody (Invitrogen) and cover-slipped using mowiol mounting medium (Calbiochem).
Images were captured using an AxioImager M1 microscope equipped with an AxioCam MRm digital camera for fluorescence or AxioCam ICc3 for brightfield and AxioVision software (Zeiss).
Genomic DNA (10 ng/µL) sequencing was performed using ABI BigDye v3.1 and analyzed using Sequencher software (Gene Codes Corporation). Control samples are from ethnically matched adults with no history of neurological disease.
HT1080 Human fibrosarcoma cells were transfected using Superfect reagent (Qiagen) with the expression vector pReceiver–M02 vector (GeneCopoeia) containing a CMV promoter upstream of GFP (to evaluate transfection efficiency), wild-type Col4a1 or mutant Col4a1 cDNA cloned and a neomycin resistance cassette, allowing stably transfected cells to be selected using G418 (Invitrogen). After 12 days of G418 selection (600 mg/ml), individual surviving clones were isolated and expanded in presence of 600 mg/ml of G148.
Stably transfected HT1080 cells were cultured in Dulbecco's modified Eagle's medium (DMEM) supplemented with penicillin, streptomycin, nonessential amino acids, glutamine, sodium pyruvate, G418 (250 mg/ml for maintenance), 10% Fetal Bovine Serum (FBS) and ascorbic acid (50 mg/ml) at 37°C in 5% CO2 in a humid atmosphere until they reach 80–90% confluence. Then, cells were serum-deprived for 24 hours under the same culture conditions and harvested and protein were extracted from HT1080 cells using cell extraction buffer containing 0.05 M Tris-HCl pH 8.0, 0.15 M NaCl, 5.0 mM EDTA, 1% NP-40, and protease inhibitors (Pierce) at 4°C. After centrifugation (13000 rpm, 20 min at 4°C), the soluble fraction of the whole cell lysate was collected for subsequent Western blot analysis. The conditioned medium was collected at the same time and supplemented with protease inhibitors.
Determination of protein concentration in the soluble fraction was performed using a DC protein assay (BioRad) and 45 ug of total protein were separated on a 4–15% gradient SDS-PAGE under non-reducing conditions and transferred to polyvinylidene fluoride (PVDF) membranes (BioRad). For the conditioned medium, the volume loaded on the gel was adjusted based on the protein concentration of the soluble fraction, which is assumed to reflect the number of cells for a given sample. Membranes were blocked for 2 hours at room temperature in 5% non-fat milk in TBS containing 0.1% Tween-20, and overnight at 4°C in 3% BSA in TBS. Membranes were then incubated with rat anti- COL4A1 (H11) monoclonal antibody (1∶100, Shigei Medical Research Institute, Japan) in 1% BSA in TBS for 3 hours at room temperature and were washed in TBS containing 0.1% Tween-20, incubated 2 hours at room temperature with horseradish peroxidase-conjugated secondary antibody raised in donkey (anti-Rat IgG 1∶10 000, Jackson Immunoresearch) diluted in 5% non-fat milk in TBS containing 0.1% Tween-20. Immunoreactivity was visualized using chemiluminescence (SuperSignal West Pico Chemiluminescent Substrate, Thermo Scientific). Densitometric analysis was performed on low exposure images using the NIH Image J software (National Institutes of Health). For quantitative analysis of the ratio of secreted to intracellular mutant COL4A1 protein, the amount of COL4A1 detected in the conditioned medium was divided by the amount of intracellular COL4A1 normalized with actin.
In experiments where two groups were compared, samples were compared by a Student's T-test with p<0.05 considered significant. In experiments where more than two groups were compared, samples were analyzed by one-way ANOVA followed by a Tukey's post hoc test with p<0.05 considered significant. For calculation of apoptotic index, one E18 Col4a1+/Δex40 eye had an extremely high apoptotic index (12.9) that was considered an outlier by Grubbs' test (p<0.01 vs. all mutants and Z = 2.93248; vs. all samples Z = 3.26997) and was removed from subsequent statistical analyses.
|
10.1371/journal.ppat.1006102 | Cell-Based Screen Identifies Human Interferon-Stimulated Regulators of Listeria monocytogenes Infection | The type I interferon (IFN) activated transcriptional response is a critical antiviral defense mechanism, yet its role in bacterial pathogenesis remains less well characterized. Using an intracellular pathogen Listeria monocytogenes (Lm) as a model bacterial pathogen, we sought to identify the roles of individual interferon-stimulated genes (ISGs) in context of bacterial infection. Previously, IFN has been implicated in both restricting and promoting Lm growth and immune stimulatory functions in vivo. Here we adapted a gain-of-function flow cytometry based approach to screen a library of more than 350 human ISGs for inhibitors and enhancers of Lm infection. We identify 6 genes, including UNC93B1, MYD88, AQP9, and TRIM14 that potently inhibit Lm infection. These inhibitors act through both transcription-mediated (MYD88) and non-transcriptional mechanisms (TRIM14). Further, we identify and characterize the human high affinity immunoglobulin receptor FcγRIa as an enhancer of Lm internalization. Our results reveal that FcγRIa promotes Lm uptake in the absence of known host Lm internalization receptors (E-cadherin and c-Met) as well as bacterial surface internalins (InlA and InlB). Additionally, FcγRIa-mediated uptake occurs independently of Lm opsonization or canonical FcγRIa signaling. Finally, we established the contribution of FcγRIa to Lm infection in phagocytic cells, thus potentially linking the IFN response to a novel bacterial uptake pathway. Together, these studies provide an experimental and conceptual basis for deciphering the role of IFN in bacterial defense and virulence at single-gene resolution.
| While the type I interferon response is known to be activated by both viruses and bacteria, it has mostly been characterized in terms of its antiviral properties. Listeria monocytogenes, an opportunistic Gram-positive bacterial pathogen with up to 50% mortality rate and a variety of clinical manifestations, is a potent activator of interferon secretion. In mouse models, interferon has been previously implicated in both restricting and promoting L. monocytogenes infection. Here, we utilized a high-throughput flow-cytometry based approach to screen a library of human interferon I stimulated genes (ISGs) and identified regulators of L. monocytogenes infection. These include inhibitors that act through both transcriptional (MYD88) and transcription-independent (TRIM14) mechanisms. Strikingly, expression of the human high affinity immunoglobulin receptor FcγRIa (CD64) was found to potently enhance L. monocytogenes infection. Both biochemical and cellular studies indicate that FcγRIa increases primary invasion of L. monocytogenes through a previously uncharacterized IgG-independent internalization mechanism. Together, these studies provide an important insight into the complex role of interferon response in bacterial virulence and host defense.
| Mammalian cells encode numerous pattern recognition receptors (PRRs) that sense invading pathogens and initiate innate immune responses through cytokine and chemokine production [1]. With viral pathogens, the type I interferon (IFN) family of cytokines serves as a first line of defense and is essential for controlling virus replication and pathogenesis. The IFN-induced antiviral response results from the transcription of hundreds of interferon-stimulated genes (ISGs), many of which inhibit different steps of the viral life cycle [2, 3]. Although less studied, the type I IFN response is also induced by many bacterial pathogens including Legionella pneumophila, Helicobacter pylori, Francisella tularensis, Yersinia pseudotuberculosis, Mycobacterium tuberculosis, and Listeria monocytogenes [4]. However, the role of type I IFN in bacterial infection remains unclear and systematic studies to uncover the breadth of ISGs targeting a bacterial pathogen have not been carried out.
We chose to clarify these aspects of IFN biology by using Listeria monocytogenes (herein referred to as Lm) as a model pathogen as its cellular life cycle has been described in detail and it exhibits a complex relationship with the mammalian IFN response system [5]. Lm is a Gram-positive food-borne pathogen that causes severe and life threatening disease in immunocompromised individuals, pregnant women, elderly and children [6]. Upon invasion of enterocytes, hepatocytes, or phagocytes, Lm gains access to the cytoplasm by lysing the primary phagosome. Lm rapidly replicates in the cytoplasm and spreads to adjacent cells via actin-based protrusion machinery [7]. Recent studies show that Lm stimulates the type I IFN response by secreting cyclic diadenosine monophosphate (c-di-AMP) that activates the Stimulator of Interferon Genes (STING). Activation of STING results in IRF3 phosphorylation and transcription of IFN genes [8, 9]. Notably, STING-deficient mice fail to produce IFNβ in response to Lm infection [10].
While the relationship between IFN and in vivo Lm infection has been firmly established, some discrepancies do exist between these studies. Early work showed that IFNβ increases the tolerance of mice to intravenous systemic Lm infection [11]. Similarly, Ifnar1 is required for resistance of mice to Lm invasion through the intestinal tract, further demonstrating a protective effect of IFN for a natural route of infection [12]. However, more recent studies indicate that mice lacking a functional type I IFN receptor (Ifnar1-/-) display greater resistance to intravenous Lm infection, suggesting that IFN exacerbates systemic Lm infection [13–15]. The type I IFN response has also been found to suppress adaptive immunity against Lm, since Sting-deficient mice exhibit greater numbers of cytotoxic lymphocytes and show protection from Lm reinfection after immunization [16]. These various effects of type I IFN on Lm infection likely reflect the different routes of Lm infection and the pleiotropic roles of IFN in distinct tissue environments or cellular populations encountered by the pathogen. Nevertheless, it is clear that type I IFN plays a significant role in shaping the host-pathogen interaction in vivo.
Because IFN induces a robust transcriptional response, the regulatory role of IFN in bacterial infection likely depends on the cellular expression of ISGs. However, the functions of most ISGs in immunity have not yet been elucidated due to the technical challenges of studying complex transcriptional responses at single-gene resolution. Recently, overexpression screens have been designed to study individual ISG functions [17–20]. While these approaches have proven to be highly successful for identifying genes that potently suppress invasion, replication, or egress of a wide variety of viruses, similar screening methodologies have not yet been adapted for bacterial pathogens. Here, we performed a gain-of-function screen of over 350 human type I IFN ISGs to identify genes that regulate Lm infection. This screen revealed potent bacterial restriction factors including MYD88, UNC93B1, TRIM14, AQP9, and MAP3K14. We demonstrated that the signaling adaptor MYD88 restricts Lm infection through the stimulation of a robust host gene expression program. In contrast, TRIM14 inhibited Lm infection through a non-transcriptional mechanism, thus suggesting that IFN stimulates diverse antibacterial properties. Importantly, we identified the human high affinity immunoglobulin receptor FcγRIa (CD64) as an IgG-independent enhancer of Lm internalization and established its role in the entry of Lm into phagocytic cells. Taken together, these findings reveal effector molecules involved in the complex relationship between Lm and the IFN response system, and open new avenues for exploring the cell-autonomous immune regulation of other bacterial pathogens.
We sought to employ a gain-of-function screening approach to identify ISGs that regulate Lm infection of host cells. We first optimized the screening conditions by determining the suitability of the host cell type previously used for ISG screens [17, 18] and by defining the optimal conditions of Lm infection. Since Lm is known to potently induce IFN expression [5], we chose human STAT1-deficient fibroblasts as the primary host cell type for infection [21]. These cells lack functional STAT1, have defective IFN responses, and therefore limit the spurious activation of ISGs during bacterial infection. To screen hundreds of ISGs in a single experiment, we optimized Lm infection of STAT1-deficient fibroblasts for compatibility with multicolor flow cytometry with auto-sampling functionality [17]. GFP-expressing Listeria monocytogenes 10403s (GFP-Lm) was added to fibroblasts at a multiplicity of infection (MOI) 5 in the absence of antibiotics. GFP-Lm was incubated for 90 minutes with host cells prior to adding gentamicin-containing media to eliminate non-internalized extracellular bacteria. Approximately 10% of host cells were infected by GFP-Lm at this time point. Cells were then incubated for 1, 2, 4 and 6 hours, providing a temporal evaluation of infection. As shown in Fig 1A, we observed an increase in the percentage of GFP-positive host cells over time indicating that GFP-Lm readily infects STAT1-deficient fibroblasts.
Lm infection progresses through a series of well-defined stages including entry, vacuole escape, cytoplasmic replication, and cell-to-cell spread (Fig 1A). Since ISGs could potentially affect any stage of the Lm lifecycle, we assessed the ability of flow cytometry to identify blocks in each of these distinct stages. STAT1-deficient fibroblasts were infected with mutant Lm strains that lack key virulence factors that are critical for cellular entry (Lm ΔinlAΔinlB), phagosomal escape (Lm ΔhlyΔplcAΔplcB), and cell-to-cell spread (Lm ΔactA). As expected, Lm ΔinlAΔinlB, lacking the major Lm invasion proteins InlA and InlB, exhibited a severe infection defect observed as early as 1 hour following initial infection (Fig 1B). In contrast, Lm ΔhlyΔplcAΔplcB mutant lacking listeriolysin O (LLO) and phospholipases required for phagosomal rupture and escape, invaded cells similar to wild type Lm, yet the percentage of infected cells and their bacterial burden (intensity of the GFP signal) did not increase over the time course of infection (Fig 1C). This observation is consistent with the Lm ΔhlyΔplcAΔplcB phenotype in which the pathogen trapped in the phagosome survives but is unable to replicate [22–25]. Finally, Lm ΔactA lacking the ActA protein required for intracellular actin-based motility invaded cells initially but failed to spread from cell to cell. Importantly, the GFP fluorescence intensity of infected STAT1-deficient fibroblasts increased over the course of infection due to bacterial replication and accumulation in initially invaded cells (Fig 1D). Thus, flow cytometry is a well-suited method to measure Lm infection of STAT1-deficient fibroblasts as it is capable of detecting specific infection defects that may arise due to ISG expression.
We next asked whether ISGs that control viral infection also regulate bacteria. Briefly, STAT1-deficient fibroblasts were transduced with bicistronic lentiviral vectors driving constitutive expression of an ISG and a red fluorescent protein TagRFP (Fig 2A). Cells expressing ISGs in a one-gene one-well format were then challenged with a GFP-Lm and resulting infection was analyzed by flow cytometry (Fig 2B). Infection rates were quantified as a percentage of GFP-positive host cells (a measure of infection) among the RFP-positive cell population (a measure of ISG expression). Firefly luciferase (Fluc), which did not affect Lm infection (S1 Fig), was used as a negative control. A panel of ISGs that enhance (MCOLN2, LY6E) or inhibit (IFI6, RTP4, TREX1, IRF2, IRF7, P2RY6, and IFITM3) yellow fever virus (YFV) had no effect on Lm infection (Fig 2C). In addition, in the absence of IFN signaling the cytosolic DNA and RNA sensors MB21D1 (cGAS) and DDX58 (RIG-I) [26, 27] as well as OASL, an ISG that inhibits hepatitis C virus [17, 28] did not inhibit Lm (Fig 2C). Thus, effects of ISGs can be differentiated between model bacterial and viral pathogens.
Next, we expressed a library of over 350 ISGs in a one-gene to one-well format [17] and performed Lm infection as described above. Infectivity of Lm obtained from the average of two screen replicates is shown as a dot plot in Fig 2D (S1 Table). The majority of ISGs had little effect on Lm infection with the cellular bacterial burdens falling within two standard deviations of the population mean (Z-score less than 2, Fig 2D). We considered inhibitors of Lm infection as those ISGs that restricted infection with Z-score greater than 2. Six ISGs fulfilled these criteria including PKRD2, UNC93B1, MYD88, AQP9, MAP3K14, and TRIM14 (Fig 2D). In addition, two genes FCGR1A and SCO2 enhanced Lm infection with Z-score greater than 2. Repeat trials with independent lentiviral preparations confirmed statistically significant inhibitory effects for all six ISGs and an enhancing effect for FCGR1A (Fig 2E).
The identification of UNC93B1 and MYD88 as cell-intrinsic inhibitors of Lm provided positive validation of the ISG screen. These genes are key components of the immune response to pathogens and function to properly target Toll-like Receptors (TLR) to subcellular compartments and to propagate NF-κB signal transduction, respectively [29, 30]. Consistent with the initial ISG screen, a dose-response to infection revealed that MYD88 and UNC93B1 reduced Lm infectivity (defined as the MOI of Lm required to achieve 50% cellular infection during the course of 8 hours) by 9.7-fold and 5.7-fold compared to firefly luciferase control (Fig 3A).
To determine if the anti-Lm activity of MYD88 results from increased NF-κB transcriptional response as predicted, we isolated mRNA from STAT1-deficient fibroblasts ectopically expressing MYD88. RNA-seq revealed 107 genes that were upregulated over 2-fold (S2 Table) and included NF-κB signature genes involved in inflammation (e.g. IL6, IL8, IL1B, CXCL1-3, CXCL5, CCL2), signal transduction (e.g. NFKB1/2, RELB, IRAK2, C1QTNF1), cell adhesion (e.g. ICAM-1, LAMB3, MMPs), and complement activation (C3) (S2 Table). To then establish the role of NF-κB activation in the observed antibacterial activity of MYD88, we tested the function of a naturally occurring single nucleotide polymorphism of MYD88 (rs1319438) that confers a S34Y substitution. While this mutation does not affect the interaction of MYD88 with IRAK1, IRAK4, and Mal, it disrupts MYD88 signaling and NF-κB activation by preventing oligomeric Myddosome complex formation required for downstream signaling (Fig 3B) [31, 32]. The cellular expression level of MYD88 S34Y was comparable to wild type MYD88 (Fig 3C), however the mutated protein failed to activate NF-κB (Fig 3D). More importantly, MYD88 S34Y did not inhibit Lm infection (Fig 3E). These findings indicate that MYD88-dependent suppression of Lm infection results from strong NF-κB transcriptional activation and currently unknown effector mechanisms.
The ISG screen also revealed AQP9, PKRD2, MAP3K14, and TRIM14 as potent inhibitors of Lm infection, suggesting that these proteins may harbor novel antibacterial activities. Aquaporin 9, encoded by AQP9, is a transmembrane channel involved in water and small solute transport [33], whereas PRKD2 and MAP3K14 are kinases implicated in membrane trafficking and immune signaling, respectively [34, 35]. It is currently unclear how expression of these genes blocks Lm infection. Among the newly identified anti-listerial ISGs, TRIM14 exhibited the greatest inhibitory activity (Fig 2E). Interestingly, this protein has recently been linked to antiviral defense through several independent mechanisms ([36–38]) but has not been previously implicated in antibacterial immunity.
TRIM14 is a member of the tripartite motif-containing (TRIM) gene superfamily that includes proteins involved in innate immunity, transcriptional regulation, cell proliferation, and apoptosis [39]. While several family members exhibit anti-viral functions [40], their role in bacterial pathogenesis remains poorly understood. We found that expression of IFN-inducible TRIM5, TRIM21, TRIM25, TRIM34, and TRIM38 had no effect on Lm infection (Fig 4A), suggesting that TRIM14 is a unique anti-bacterial effector among the IFN-stimulated TRIMs. As shown in Fig 4B, domain architecture of TRIM14 is distinct from other family members as it is does not encode the RING E3-ligase domain typically found within the N-terminal tripartite motif, and is therefore likely to function through an alternative mechanism [39]. We next asked if TRIM14-mediated antibacterial activity could be attributed to one of its structural domains. Based on the available crystal structures of truncated TRIM proteins [41, 42], we generated TRIM14 constructs consisting of either the B-box with coiled-coil (residues 1–255) or the PRY/SPRY domain (residues 158–442). Notably, separate domains of TRIM14 had no effect on Lm infection (Fig 4C), indicating that the full-length TRIM14 is required for the anti-bacterial activity.
TRIM14 was recently reported to play an important role in IFN and NF-κB activation during viral infection [36, 37]. It associates with the mitochondria wherefrom it links MAVS and NEMO to NF-κB and IRF3-activated transcription [37]. It has also been shown to positively regulate type I IFN signaling by inhibiting cGAS degradation [36]. However, our studies were performed in STAT1-deficient fibroblasts that cannot be activated by IFN, suggesting that the anti-Lm activity of TRIM14 is not associated with this ascribed function. Furthermore, a critical lysine in TRIM14, K365, was shown to be required for IFN activation by TRIM14 in mitochondria [37]. However, we found that K365 was not necessary for the antibacterial function of TRIM14 (Fig 4C). Finally, ectopic expression of TRIM14 in STAT1-deficient fibroblasts induced the expression of 36 genes by over 2-fold and only 5 genes had greater than 5-fold increase compared to over 100-fold increase in TRIM14. Ingenuity Pathway Analysis failed to identify possible upstream transcriptional regulators (Fig 4D and S3 Table), and the observed transcriptional response did not exhibit an NF-κB or IRF3 signature as would be predicted if TRIM14 regulated MAVS and NEMO as previously reported.
To then determine if TRIM14 functioned through a transcription-independent mechanism during infection, we compared host mRNA produced during Lm infection (6 hours) in cells expressing TRIM14 or luciferase as a control. Lm infection altered expression of hundreds of genes in luciferase-expressing cells (S4 Table). As expected, TNFα, NF-κB and IL1A were among the strongest predicted upstream regulators (Fig 4E and S4 Table). Interestingly, expression of TRIM14 did not alter the host transcriptional response to Lm infection (S4 Table), further suggesting that TRIM14 has a direct anti-bacterial function in host cells. Taken together, these results indicate that the gain-of-function ISG screening technique can resolve direct mechanisms of inhibition of bacteria, similar to what has been demonstrated for viruses [17].
In addition to anti-bacterial ISGs, we identified a high affinity immunoglobulin receptor FcγRIa (CD64) as an enhancer of Lm infection. A dose-response experiment indicated that FcγRIa potentiated Lm infectivity by over 100-fold (Fig 5A). Consistent with the flow cytometry measurements, a greater number of individual bacteria were found in the cytoplasm of FcγRIa-expressing cells (Fig 5B) and FcγRIa increased the total number of cell surface protrusions emanating from infected cells (Fig 5C).
Because FcγRIa is a cell surface expressed protein, we hypothesized that it may enhance Lm infection by promoting primary internalization into host cells or secondary cell-to-cell spread. To distinguish between these possibilities, we visualized Lm infection foci, which are formed from Lm invasion of a single host cell followed by rapid cell-to-cell transmission. Cell monolayers were infected with very low doses of Lm (at MOI 0.015, 0.05, 0.1) and the formation of foci was evaluated 30 hours post infection (see Materials and Methods). Cellular expression of FcγRIa increased the total number of infection foci compared to control (Fig 6A). However, the diameter and surface area of individual foci were not altered in presence of FcγRIa (Fig 6B). Thus, FcγRIa enhances efficiency of primary Lm invasion, yet has little effect on secondary cell-to-cell spread.
We next asked if FcγRIa potentiated Lm entry by coordinating interactions with the host Lm internalization receptors E-cadherin or c-Met [43]. We introduced frameshift mutations into CDH1 (encoding E-cadherin) and MET (encoding c-Met) by CRISPR/Cas9 resulting in non-coding genetic disruption of these loci (S2A–S2F Fig). As expected, the invasive capacity of Lm was significantly attenuated in CDH1/MET-deficient cells (S3G Fig), which could be restored by ectopic expression of either CDH1 or MET (Fig 6C). Remarkably, ectopic expression of FcγRIa in CDH1/MET-deficient cells increased Lm infection to levels comparable with MET complementation (Fig 6C and 6E). In addition, mutant Lm ΔinlAΔinlB lacking the invasins InlA and InlB that directly bind host surface proteins E-cadherin and c-Met, respectively, readily infected CDH1/MET-deficient cells expressing FcγRIa (Fig 6D, 6F and 6G). Therefore, FcγRIa supports bacterial uptake independently of “classic” host Lm internalization receptors E-cadherin and c-Met as well as bacterial invasins InlA and InlB.
Fcγ receptors bind the Fc (antigen non-specific) region of IgG antibodies produced as a part of adaptive response to infection in mammals. The human Fcγ receptor family includes activating receptors FcγRIa, FcγRIIa, FcγRIIc, FcγRIIIa and FcγRIIIb, as well as an inhibitory receptor FcγRIIb. Crosslinking of activating Fcγ receptors by IgG typically results in the phagocytosis of opsonized particles and cellular activation, facilitating destruction of the pathogens and induction of inflammation, respectively [44]. In humans, FcγRIa is constitutively expressed on monocytes and macrophages, and its expression is upregulated by type I and II interferons and other signaling molecules, such as IL-10 [45]. It consists of three extracellular immunoglobulin (Ig)-like domains, a single transmembrane domain, and a short cytoplasmic tail that does not contain any known signaling motifs. During receptor engagement with IgGs, FcγRIa recruits the accessory immunoreceptor tyrosine-based activation motif (ITAM)-containing γ-chain (FcεRIg). Clustering of the FcγRIa with γ-chain triggers intracellular signaling cascades involving Syk and Src family kinases necessary for FcγRIa-mediated particle phagocytosis [46, 47]. Additionally, FcγRIa has been shown to interact with FcγRIIa, using its ITAM-motif to signal in the absence of the γ-chain [48].
To compare the mechanism of Lm invasion to the classic IgG-coated particle uptake though FcγRIa alone in the absence of possible crosstalk with other Fcγ receptors [44, 48], we developed a model of Fc-receptor functions in a non-phagocytic cell type [46, 49–51]. We first reconstituted IgG-coated particle internalization via FcγRIa. U-2 OS cells were transduced with a lentivirus expressing FcγRIa, or Fluc as a negative control. Latex beads were coated with human IgG and labeled with anti-human secondary antibody conjugated to Alexa Fluor 488 (green). The IgG opsonized particles were incubated with U-2 OS cells for 1.5 h at 37°C and then shifted to 4°C to inhibit further uptake. Cell-surface bound beads were differentiated from internalized beads by incubating samples with anti-human secondary antibody conjugated to DyLight 405 (blue) without cell permeabilization (Fig 7A). Under these conditions, internalized beads are protected from the secondary antibody and are visualized as green beads by fluorescence microscopy. In contrast, surface-bound beads are labeled with both green and blue secondary antibodies.
As expected, luciferase-expressing U-2 OS cells showed no interaction with IgG-coated beads. In contrast, FcγRIa recruited IgG-beads to the cell surface, but revealed low levels of bead internalization (Fig 7B and 7C). This may be anticipated since U-2 OS cells do not express endogenous γ-chain (FcεR1g). Indeed, co-expression of FcγRIa with the γ-chain (FcεR1g) fully reconstituted FcγRIa-mediated internalization of IgG-coated beads (Fig 7C). We also cloned and tested another Fcγ receptor–FcγRIIa–a low-affinity immunoglobulin receptor that possesses its own internal ITAM motif and therefore, does not require interaction with the γ-chain for particle internalization. FcγRIIa mediated similar high levels of IgG-coated bead phagocytosis (Fig 7B and 7C). Thus, we have established a robust and simplified cellular system to study the function of individual human Fcγ receptors in context of both particle opsonization and pathogenic Lm infection.
As shown in Fig 5A, Lm readily invaded U-2 OS cells expressing FcγRIa. Surprisingly, this phenotype did not require co-expression of the ITAM-containing γ-chain, suggesting that Lm internalization by FcγRIa occurs through a distinct mechanism compared to IgG-coated particle uptake. It has been previously reported that FcγRIa interacts with the γ-chain exclusively through the transmembrane domain [52]. We therefore asked if this region of FcγRIa was necessary for Lm internalization. We targeted the extracellular Ig-like domains of FcγRIa to the cell surface via a GPI-anchor signal of LFA-3 (FcγRIa-GPI) [53]. This chimeric protein was expressed on the cell surface similar to the wild type FcγRIa (Fig 8A). Notably, as shown in Fig 8B, FcγRIa-GPI induced the same level of Lm infection as the wild type protein (3.03 fold), further confirming that FcγRIa does not interact with the γ-chain during Lm internalization. Additionally, FcγRIa does not require interaction with any other signaling protein through the transmembrane domain.
The ability of Lm to be internalized by FcγRIa in the absence of the signaling γ-chain suggested that the recognition of Lm might also occur independently of IgG opsonization. Several lines of evidence support this conclusion. First, Lm infection was not enhanced by expression of other members of the Fc receptor family (Fig 8C), including FcγRIIa that, as shown in Fig 7C, was able to internalize IgG-coated beads. Second, FcγRIa potentiated Lm invasion in serum-free (and therefore, IgG-free) conditions (Fig 8D). Third, reducing the FcγRIa affinity for all types of IgG up to 100-fold by introducing an H174E mutation in the D2 Ig-like domain [54] did not affect its ability to enhance Lm infection (Fig 8E). Finally, FcγRIa had no effect on the infection rate of other intracellular bacteria Shigella flexneri or Salmonella Typhimurium (Fig 8F and S3A and S3B Fig). Therefore, internalization through FcγRIa is independent of non-specific pathogen opsonization with serum IgG. Together, these data indicate that Lm invades cells independently of the well-established route of phagocytosis, which involves IgG opsonization and ITAM-mediated intracellular signaling through the FcγRIa-γ-chain complex.
Since our data revealed a novel mechanism of FcγRIa function that included Lm internalization independent of IgG opsonization, we hypothesized that Lm might express a cell surface factor required for Lm-FcγRIa interaction. To determine if this molecule was a general feature of Listeria genus or specific to pathogenic Lm, we assessed the ability of FcγRIa to confer invasiveness to a closely related but non-pathogenic species L. innocua. While L. innocua strains are genetically heterogeneous and may encode various combinations of genes shared with Lm [55, 56], L. innocua CLIP 11262 has been fully sequenced and annotated. It has been shown to lack all major genes required for Lm pathogenesis (inlA, inlB, hly, actA, etc.) [57]. As show in Fig 8G, in the absence of FcγRIa, invasion rates of non-pathogenic L. innocua CLIP 11262 were 78.43-fold lower than those of invasion-deficient Lm ΔinlAΔinlB. Importantly, while FcγRIa expression increased internalization of Lm ΔinlAΔinlB, it did not allow invasion of the CLIP 11262 strain (Fig 8G). Similarly, FcγRIa did not increase internalization of an unsequenced L. innocua strain DUP-104 (Fig 8G), suggesting that the bacterial ligand is specific to Lm and is not shared with non-pathogenic Listeria strains.
To determine the contribution of FcγRIa to Lm infection in a naturally phagocytic human cell type that expresses endogenous FcγRIa, we disrupted cell surface expression of FCGR1A in THP-1 human monocytes using a lentiviral CRISPR/Cas9 system [58] (Fig 9A and 9B). Lm infected 54.45 ± 2.19% wild-type THP-1 cells compared to 44.48 ± 2.98% of FCGR1A-deficient cells (Fig 9C) representing a statistically significant decrease in infection (p = 0.0095, n = 3). These data suggest that Lm is internalized through multiple pathways with 18.13 ± 8.1% (n = 3) of the total host cell infection mediated by FcγRIa (Fig 9D). To then determine if the observed decrease in Lm infection was due to the newly defined mechanism of FcγRIa-Lm interaction described above rather than a general defect in IgG-coated pathogen internalization, we performed experiments in serum-free conditions. A significant reduction in Lm infection was observed in FCGR1A-deficient THP-1 (44.52 ± 1% infected wild type cells compared to 38.44 ± 0.69% of FCGR1A-deficient cells; p = 0.0010, n = 3) (Fig 9C) with a 13.63 ± 2.52% (n = 3) relative contribution of FcγRIa under these conditions (Fig 9D). Thus, endogenous FcγRIa contributes to Lm invasion of phagocytic monocytes independently of IgG opsonization similar to what was observed in the reconstituted cellular system.
Having established that human FcγRIa enhances Lm infection, we next asked whether mammalian FcγRIa orthologs exhibit similar functions. Species-specific FcγRIa coding sequences were commercially synthesized, and included (1) mouse (naturally resistant to oral Lm infection), (2) sheep and rabbit (known to be susceptible to Lm), and (3) panda (uncharacterized susceptibility to Lm infection). All FcγRIa orthologs were expressed on the cell surface of U-2 OS cells as determine by IgG-coated latex bead binding assays (Fig 10A). We then co-expressed these receptors with the γ-chain and tested whether they were fully functional in human cells by measuring the rates of IgG-opsonized particle internalization. All tested FcγRIa induced similar levels of IgG-bead phagocytosis (Fig 10B), suggesting that they were indeed functioning as internalization receptors for opsonized particles. Next, we assessed the ability of non-primate FcγRIa to potentiate internalization of Lm. FcγRIa of mouse, sheep and panda failed to enhance Lm infection (Fig 10C). Moreover, murine FcγRIa did not affect Lm infection even when co-expressed with the γ-chain in murine cells (Fig 10D). Unexpectedly, rabbit FcγRIa was found to potentiate Lm internalization in the absence of the γ-chain (Fig 10D). Rabbit is a natural host for Lm and exhibits severe listeriosis upon infection [59]. Analysis of the multiple sequence alignment of FcγRIa from these species did not pinpoint a single residue or a motif that was common between human and rabbit yet divergent from other FcγRIa proteins tested, suggesting a more complex interaction between host and pathogen molecules (S4 Fig). Nevertheless, these data indicate that FcγRIa-Lm interaction is not only pathogen-specific (Fig 8F), but also demonstrates host protein tropism.
The host type I interferon response is stimulated by numerous bacterial pathogens. However, the roles of individual ISGs in restricting bacterial infection are not well characterized. To address this gap in the knowledge of IFN biology, we adapted a gain-of-function screening approach to identify cellular regulators of Lm infection among approximately 350 type I ISGs. The screen revealed strong cell-autonomous inhibitors of Lm infection, such as TRIM14, AQP9, MYD88, UNC93B1 and MAP3K14. Interestingly, we also identified the human immunoglobulin receptor FcγRIa as an enhancer of Lm internalization, suggesting an intriguing possibility that bacterial pathogens have evolved virulence factors to directly hijack the IFN response system.
We identified type I IFN-stimulated inhibitors of Lm infection that function through the upregulation of complex gene expression profiles (e.g. MYD88) and/or through direct anti-microbial mechanisms (e.g. TRIM14). These ISGs may contribute to the regulation of Lm in a wide variety of tissue environments. For example, upregulation and activation of MYD88 in TLR-expressing lymphocytes would result in the expression of NF-κB-regulated genes with broad antibacterial activity. Our data indeed suggest that a MYD88-induced transcription program suppresses Lm infection through NF-κB activation (Fig 3C–3E). Notably, Lm has been previously reported to counteract host defense systems, including interfering with NF-κB activation, thus dampening the overall inflammatory response to infection [60]. Our findings now indicate that inhibition of NF-κB by Lm may protect the pathogen from previously unknown cell-autonomous immune mechanisms. Further studies are needed to confirm this speculation. Another strong inhibitory ISG, TRIM14 is widely expressed throughout the body, including organs targeted by Lm, such as intestine and liver [37, 61]. However, this is not the first study to implicate TRIM14 in anti-microbial defense. Recent studies characterized TRIM14 as an antiviral protein that activates both NF-κB and type I IFN through bridging MAVS and NEMO proteins as well as inhibiting cGAS degradation [36, 37]. Interestingly our data support an alternative mechanism for the function of TRIM14. We found that TRIM14 inhibited Lm infection in cells with defective IFN responses and that ectopic expression of TRIM14 did not alter the host transcriptional profile induced by Lm (Fig 4). Further studies are needed to reveal the precise inhibitory mechanisms of TRIM14 as well as other antilisterial ISGs including PRKD2, AQP9, and MAP3K14 identified here.
Perhaps the most surprising discovery of this work is that the immunoglobulin receptor FcγRIa mediates Lm uptake, contributing to Lm invasion of phagocytic monocytes and macrophages. This finding is particularly insightful since these cells are not only an important target of Lm infection, but also aid the transmission of Lm to peripheral tissues during infection [62]. Currently, the precise molecular mechanisms of Lm internalization in phagocytic cells have not been characterized in detail and are believed to be mediated by C3bi and C1q complement receptors and phagocyte scavenger receptors [63, 64]. However, our studies now suggest that Lm hijacks an alternative pathway to invade phagocytic cells through an immunoglobulin-independent interaction with FcγRIa. While studies presented here have elucidated many key aspects of the internalization process (see below), several questions remain unanswered: (1) what is the nature of the IgG-independent interaction between Lm and FcγRIa resulting in Lm uptake by the host cells, (2) what is the cellular mechanism of FcγRIa-mediated Lm internalization; and, finally, (3) what are the consequences of this interaction for both pathogen and host in terms of pathogen proliferation and disease outcomes.
In this study, we provide compelling evidence that Lm is internalized by FcγRIa independently of IgG opsonization (Fig 8). The most direct explanation for these findings is that Lm directly engages FcγRIa at the surface of immune cells. However, the identity of the bacterial surface protein involved in the interaction remains unclear. FcγRIa was unable to induce invasion of a non-pathogenic L. innocua (Fig 8G), thus narrowing down the search for the ligand to a small subset of Lm-specific surface proteins [57]. Importantly, we have ruled out the involvement of the most well-characterized Lm invasion factors—internalins InlA and InlB (Fig 6), known to act individually or in concert to trigger Lm entry into a wide variety of non-phagocytic cells [43, 65]. Other Lm-specific proteins previously implicated in the entry of Lm into target cells, virulence factor ActA, as well as less-characterized Vip, LapB, and Auto may be involved in FcγRIa interaction [43, 66–68]. Both biochemical studies on candidate bacterial surface proteins as well as unbiased genetic screens will help determine if the FcγRIa ligand is a known invasion protein or a novel factor previously not implicated in Lm infection. Notably, a similar IgG-independent interaction has been demonstrated between FcγRIa and Escherichia coli K1 [69]. These bacteria have been shown to invade macrophages as a result of the interaction of the bacterial Outer membrane protein A (OmpA) with FcγRIa. It therefore appears that targeting IgG-independent functions of FcγRIa may be a general pathogenic strategy to evade immune clearance during systemic infection.
While work presented in this study clearly indicates that FcγRIa facilitates entry of Lm into host cells, the cellular signaling mechanisms required for this process remain unknown. Since FcγRIa itself does not contain any known signaling motifs, the FcγRIa-mediated phagocytosis of IgG-coated particles requires receptor interaction with the ITAM-domain containing γ-chain, which in turn mediates downstream signaling, triggering cytoskeleton rearrangement and particle internalization [46, 70]. Interaction of FcγRIa with the γ-chain occurs exclusively through the transmembrane domain of the receptor [52]. However, we found that GPI-anchored FcγRIa preserved its ability to internalize Lm in the absence of the transmembrane domain (Fig 8B), indicating that both transmembrane and intracellular domains of FcγRIa were dispensable for this process. Thus, our data reveal the existence of an alternative non-canonical mechanism of FcγRIa internalization. It is currently unclear if FcγRIa-mediated uptake of Lm resembles the extensively characterized mechanism of Lm uptake by non-phagocytic cells through E-cadherin and c-Met receptors. Lm-induced clustering of these receptors leads to the recruitment of clathrin-mediated endocytosis machinery, actin cytoskeleton organization, and modulation of the phosphoinositide metabolism at the site of bacterial adhesion, resulting in the engulfment of the pathogen by zipper-like mechanism [71]. It will be of interest to define the involvement of actin, clathrin, and intracellular signaling pathways in the FcγRIa-mediated Lm entry.
It is intriguing to speculate on the potential role of FcγRIa in Lm pathogenesis. We found that a small but reproducible percentage of THP-1 infection (~18%) was dependent on cell surface expression of endogenous FcγRIa (Fig 9). Therefore, our data reveal the existence of at least two distinct pathways for Lm invasion including a canonical phagocytic pathway and a novel FcγRIa-mediated pathway described here. We hypothesize that Lm may have evolved surface molecules to engage the FcγRIa internalization pathway and bypass cell-mediated killing induced by other phagocytic routes of internalization. Consistent with this idea, Lm did not specifically engage the major phagocytic Fcγ receptor FcγRIIa involved in pathogen clearance in neutrophils and monocytes (Fig 8C). In addition, previous studies have demonstrated fundamental differences in intracellular signaling pathways, receptor trafficking, antigen presentation, and kinetics of oxidative burst triggered by high-affinity IgG receptors (FcγRIa) compared to low affinity receptors (FcγRIIa) [72]. Thus, the ability of Lm to exploit the high affinity IgG receptor rather than being phagocytosed through the canonical opsonization pathway by FcγRIIa, may provide an opportunity for invaded Lm to produce phagosome rupture factors and escape into the cytoplasm. While this scenario has not yet been substantiated in vivo, the challenge for future studies will be to examine Lm internalization by FcγRIa in primary human cells revealing the role of FcγRIa in Lm pathogenesis.
In conclusion, our flow cytometry based screening approach not only uncovered type I IFN stimulated suppressors of Lm infection but also revealed a novel Lm uptake pathway, which may play an important role in human Lm infection and disease pathogenesis. This work also opens up new experimental avenues to examine the role of IFNs, and potentially other immune modulatory transcriptional programs, in the pathogenesis of a wide range of bacterial species, including both intracellular bacteria that replicate in either vacuoles or cytoplasmic environment, and extracellular bacteria that may be affected by secreted ISGs.
Listeria monocytogenes 10403s constitutively expressing Green Fluorescent Protein (GFP), L. monocytogenes DP-L2319 (10403s Δhly ΔplcA ΔplcB), and DP-L3078 (10403s ΔactA) strains were a gift from Dan Portnoy (UC Berkeley). L. monocytogenes 10403s ΔinlA ΔinlB, expressing GFP (LM 131) was kindly provided by Manuel Amieva (Stanford). To generate Lm 10403s Δhly ΔplcA ΔplcB and ΔactA pactA::GFP strains pPL2-GFP construct was chemically transformed into E.coli SM10, followed by conjugation with the DP-2319 strain. GFPmut2 was PCR amplified from the genomic DNA of Listeria strain LM124 and then cloned downstream of the actA proximal promoter (200bp upstream) in the pPL2 vector. pPL2 was used to integrate genes at the tRNAArg locus of the Listeria chromosome [73].
L. innocua strains DUP-104 [LCDC 81–861] and BAA-680 (CLIP 11262) (genome sequencing strain) were obtained from ATCC. Additionally, Shigella flexneri strain M90T (serotype 5) with pBBRMCS1-GFP plasmid and GFP-expressing Salmonella Typhimurium str. SL1344 expressing pBBR1MCS 6Y GFP were used.
STAT1-deficient fibroblasts (an SV40 large T antigen immortalized skin fibroblast line, kindly provided by Jean-Laurent Casanova, Rockefeller University) were grown in RPMI Medium 1640 (Gibco, Thermo Fisher Scientific), supplemented with 10% Fetal Bovine Serum (FBS) (Gibco, Thermo Fisher Scientific) and non-essential amino acids (NEAA) (Gibco, Thermo Fisher Scientific). HEK293A (Jack Dixon, UC San Diego), HEK293T (Paul Bieniasz, Aaron Diamond AIDS Research Center), U-2 OS (ATCC), and MEF (Charles Rice, Rockefeller University) cells were maintained in Dulbecco's Modified Eagle Medium (DMEM) (Gibco, Thermo Fisher Scientific), supplemented with 10% FBS and NEAA. THP-1 cells (ATCC) were cultured in RPMI Medium 1640, ATCC modification (Gibco, Thermo Fisher Scientific), supplemented with 10% FBS and NEAA.
cDNA for human FCGR1B, FCGR2B, FCGR3A, FCGR3B, FCER1A, FCER2A, FCER1G, FCAR1 were obtained from the Ultimate ORF Clones (96-well plate) collection (Life Technologies) as Gateway-compatible pENTR clones. cDNA for human FCGR2A was a gift from Dr. Eric Hansen (UTSW). These genes were amplified by PCR with primers encoding attB sites. Polymerase chain reaction (PCR) products were purified with the QIAquick PCR Purification Kit (Qiagen) and then recombined into a pDONR221 vector using BP Clonase II Enzyme mix (Life Technologies). BP reactions were transformed into chemically competent DH5a Escherichia coli, and colonies verified by sequencing. Resulting pENTR clones were further recombined into a pTRIP.CMV.IVSb.ires.TagRFP Destination vector [17] using LR Clonase II Enzyme mix (Life Technologies). LR reactions were transformed into DH5α cells and verified by sequencing.
pLenti CMV Puro DEST (w118-1) for generation of stable cell lines was a gift from Eric Campeau (Addgene plasmid #17452) [74]. FLUC, FCGR1A, and FCER1G (referred to as γ-chain) were introduced using LR Clonase II Enzyme mix (Life Technologies) as described above.
Point mutations and truncations were generated by PCR of the corresponding pENTR clones using a QuikChange II XL Site-Directed Mutagenesis Kit (Agilent) and primers designed according to manufacturer’s instructions. Glycosylphosphatidylinositol (GPI) anchored FcγRIa (previously described in [53] was generated by overlap extension PCR, using FCGR1A and LFA3, obtained from the Ultimate ORF Clones (96 well plate) collection (Life Technologies), as templates.
Sheep (NM_001139452.1), rabbit (XM_008264510.1), and panda (XM_011217915.1) FCGR1A cDNA were codon optimized for expression in human cells using Codon Optimization Tool (Integrated DNA Technologies) and synthesized as gBlocks Gene Fragments (Integrated DNA Technologies) with addition of attB sites. Mouse (NM_010186) FCGR1A cDNA was synthesized as a pENTR clone (GeneCopoeia, Inc). Genes were recombined into pDONR221 and subsequently into expression vector pTRIP.CMV.IVSb.ires.TagRFP Destination vectors as described above.
pX335-U6-Chimeric_BB-CBh-hSpCas9n(D10A) was a gift from Feng Zhang (Addgene plasmid # 42335) [75]. LentiCRISPR v2 was a gift from Feng Zhang (Addgene plasmid # 52961) [58].
Lentiviral pseudoparticles were generated as previously described [17].
Lentiviral transduction was performed as previously described [17]. Briefly, cells were seeded in 24-well tissue culture plates at a density of 7x104 cells per well and transduced the following day with lentiviral pseudoparticles via spinoculation at 1,000 x g for 45 min in medium containing 3% FBS, 20mM HEPES and 4 μg/ml polybrene. 6 h after spinoculation, pseudoparticle-containing media was removed and replaced with full cell culture medium, containing 10% FBS and NEAA. For subsequent bacterial infection, cells were split 1:2 48h after transduction. For generation of stable expressing cell lines using pLenti CMV Puro DEST (w118-1), cells were transduced with the lentivirus and selected for puromycin resistance for 7 days 48h after transduction.
Listeria monocytogenes was inoculated from a frozen stock and grown for 13 h at 30°C in brain–heart infusion media (BHI) (Difco, BD Biosciences) without shaking. 1 ml of bacteria was then washed in phosphate buffer saline (PBS) and resuspended in 1ml of PBS. A 1:10 dilution of the bacterial suspension was used to read the optical density at 600 nm (OD600). Bacteria were then added to each well of cells to achieve multiplicity of infection (MOI) of 10, unless otherwise stated, and incubated for 90 min at 37°C, 5% CO2 (unless otherwise noted). Culture media was then removed and replaced with media supplemented with 25 μg/ml gentamicin (Quality Biological) and cells incubated at 37°C, 5% CO2 for the indicated period of time. STAT1-deficient fibroblasts were infected with Lm for 6 h, HEK293A –for 4 h, MEF– 3.5 h, THP-1–6 h, unless otherwise stated in figure legend, U-2 OS–see specific figure legends.
L. innocua infection was performed following a similar protocol with MOI of 10, invasion was measured as described in “Measuring intracellular bacterial burden” (see below) 1 h following 1.5 h initial infection.
For Lm infection of THP-1 cells, 8x104 cells were seeded per well in 96-well tissue culture plates in 10% FBS/RPMI or serum-free RPMI. 24 h later later Lm infection was performed as described above (MOI = 5). Following 1.5 h initial invasion time, gentamicin-containing media was added to the wells (final concentration 30 μg/ml) and infection was allowed to proceed for 6 h. Contribution of FcγRIa to Lm infection in each independent experiment was calculated using the following equation: [(percent infected wild type cells)–(percent infected FCGR1A-deficient cells) / [(percent infected wild type cells)] x 100%.
To visualize bacterial infection by epifluorescence microscopy, cells were washed once in PBS, fixed in 3.7% formaldehyde in PBS for 10 min at room temperature. Cells were then washed three times in PBS and incubated for 2 min in 4',6-diamidino-2-phenylindole (DAPI) solution.
Shigella flexneri strain M90T was inoculated from a frozen stock and grown overnight at 30°C in BHI medium (Difco, BD Biosciences). Bacteria were then back-diluted 1:50 and incubated at 37°C until reaching OD600 ≈ 0.5–0.6. Bacteria were then washed in 1×PBS and incubated at 37°C for 15 min in 0.003% Congo red. Bacteria were added to each well to achieve MOI = 10 and centrifuged at 1000 x g for 10 min at room temperature to facilitate bacterial adherence. The plates were then incubated for 90 min at 37°C, 5% CO2. The media was removed and replaced with media supplemented with 50 μg/ml gentamicin (Quality Biological) and cells incubated at 37°C, 5% CO2 for 4.5 h. Cells were washed once with PBS before collecting for flow cytometry analysis.
Salmonella Typhimurium strain SL1344 was inoculated from a frozen stock and grown at 37°C in BHI (Difco, BD Biosciences) in a glass flask with high aeration overnight, then subcultured (1:30) and grown for 3 h at 37°C. 1 ml of bacterial suspension was then washed in PBS and resuspended in 1ml of PBS. 1:10 dilution of the bacterial suspension was used to read the optical density at 600 nm (OD600). Bacteria were added to each well to achieve MOI = 100 and incubated for 1 h at 37°C, 5% CO2, washed three times with PBS and incubated at 37°C, 5% CO2 in medium supplemented with 100 μg/ml gentamicin (Quality Biological) and cells incubated at 37°C, 5% CO2 for 8 h. Cells were washed again with PBS before collecting for flow cytometry analysis.
YFV-17D-Venus infection was performed as previously described [17].
For flow cytometry analysis, cells were detached from the tissue culture plate by incubating in 150μl of Accumax Cell Dissociation Solution (Innovative Cell Technologies, Inc.) for 5 min at 37°C, transferred to V-bottom 96-well plates, pelleted by centrifugation at 800 x g for 5 min, resuspended in 1% paraformaldehyde (PFA) and incubated at 4°C for at least 30 min. Fixed cells were then pelleted at 800 x g for 5 min and resuspended in 150μl of 1×PBS containing 3% FBS. Plates were stored at 4°C if flow cytometry was not carried out immediately. Samples were analyzed using a Stratedigm S1000 flow cytometer equipped with 405nm, 488nm and 561nm lasers. Data was analyzed using FlowJo Software (Treestar).
Following Lm or L. innocua infection, mammalian cells were washed three times with 1×PBS and then lysed by incubating in 0.5% Triton X-100 for 5 min at room temperature, followed by vigorous pipetting to complete the lysis. Intracellular bacterial burden was determined by plating serial dilutions of suspension on BHI-agar plates, incubating at 37°C, and counting bacterial colony forming units (CFU) the next day. Additionally, serial dilutions of bacterial culture used for infection were plated to obtain the inoculated CFU. Invasion was quantified by using the following equation: [CFU recovered per well/CFU inoculated per well] x 100% = invasion and normalized to control values, if needed.
To detect surface expression of FcγRIa V450 Mouse anti-Human CD64 (BD 561202) and V450 Mouse IgG1, κ Isotype control (BD 560373) antibodies were used according to the manufacturer’s protocol. Briefly, adherent cell (4x105 cells per well) were washed once with PBS, detached from the surface by incubating in 150μl of Accumax Cell Dissociation Solution (Innovative Cell Technologies, Inc.) for 5 min at 37°C, transferred to V-bottom 96-well plates, pelleted by centrifugation at 300 x g for 5 min, washed once PBS and staining buffer (2% FBS in 1×PBS). Cells were then resuspended in 50μl of staining buffer and 2.5μl of fluorescently tagged antibody was added. Cells were incubated for 30 min at room temperature, in the dark. After incubation, cells were washed twice in staining buffer, resuspended in 150μl of staining buffer and analyzed immediately by flow cytometry.
Cells were washed once with PBS and lysed using RIPA Lysis and Extraction Buffer (Pierce, Thermo Fisher Scientific) supplemented with Protease Inhibitor Cocktail (Sigma). Total protein concentration was determined using the BCA Protein Assay Kit (Pierce, Thermo Fisher Scientific). Proteins were separated on SDS-PAGE and transferred to 0.45 μm nitrocellulose membranes (Biorad). Membranes were then blocked with 5% (w/v) skim milk (Difco, BD) in Tris-buffered saline with 0.1% Tween 20 (TBST) for 1 h at room temperature and immunoblotted with primary antibodies in TBST containing 5% nonfat milk at 4°C overnight, followed by incubation with appropriate secondary antibodies coupled to horseradish peroxidase (HRP) for 1 h at room temperature. Proteins were detected using ECL Western Blotting Substrate (Pierce, Thermo Fisher Scientific). The following antibodies were used in this study: anti- MYD88 (AF2928, R&D Systems), anti-E-cadherin (BD 610181, BD Biosciences), anti-c-Met (CST 4560, Cell Signaling Technology), anti-actin (a-2066, Sigma Aldrich), goat anti-rabbit (31460, Thermo Fisher Scientific), donkey anti-goat (sc-2020, Santa Cruz Biotech), goat anti-mouse (115-035-146, Jackson ImmunoResearch).
RNA was isolated from STAT1-deficient fibroblasts, ectopically expressing the gene of interest, using an RNeasy Mini Kit (Qiagen) according to the manufacturer’s instructions. For each condition, two independent replicates were prepared. Further procedures were performed at the UTSW Next Generation Sequencing Core (McDermott Center). The quality of the total RNA samples was first confirmed on a 2100 Bioanalyzer (Agilent) using the total RNA 600 Nano Kit (Agilent) and amount of RNA quantified using the Qubit RNA Assay kit (Life Technologies). 4μg of total RNA with an RNA Integrity Number (RIN score) above 8, were further processed as described in TruSeq Stranded mRNA Sample Preparation Guide (Illumina). Samples were fragmented at a lower temperature than recommended (80°C for 4 min instead of 94°C for 8 min) to obtain 400-800bp libraries. Additionally, 12 PCR cycles were performed, instead of 15 cycles recommended by the protocol. Resulting libraries were analyzed on 2100 Bioanalyzer (Agilent) using DNA High Sensitivity Kit (Agilent) and quantified using Qubit. Sequencing was performed on Illumina Hiseq2500 with 100 bp paired end reads. Further procedures were performed at the UTSW Bioinformatics Core (McDermott Center). Sequencing reads were trimmed to remove adaptor sequences and low quality bases using fastq-mcf (v1.1.2–806, https://expressionanalysis.github.io/ea-utils/). Filtered reads were then mapped to human genome (hg19) using Tophat (v2.0.10) [76], guided by igenome annotations (https://ccb.jhu.edu/software/tophat/igenomes.shtml). Duplicate reads were marked but not removed. Expression abundance estimate and differential expression test were performed using Cufflinks/Cuffdiff (v2.1.1) software [76]. Differential expression was considered as statistically significant when q-value was lower than 0.05, fold change was greater than 2, and FPKM value of at least one sample was greater than 0.01. The upstream regulator analyses were generated through the use of QIAGEN’s Ingenuity Pathway Analysis (IPA, QIAGEN Redwood City, www.qiagen.com/ingenuity).
Guides targeting exon 3 of CDH1 and exon 3 of MET were designed using the Optimized CRISPR Design Tool (http://crispr.mit.edu/), and cloned into the pX335-U6-Chimeric_BB-CBh-hSpCas9n(D10A) vector as previously described [77]. For each guide pair, 4x105 HEK293A cells were seeded in a 6-well plate, and the following day were transfected with 1 μg of GFP-N3, and 1 μg of the positive and negative guides, according to the FuGENE 6 (Promega) protocol. Approximately 48 h post transfection, fluorescence-activated cell sorting (FACS) was used to deposit single GFP-positive cells into 96-well plates. Approximately 2 weeks after sorting, colonies were transferred to 24-well plates in duplicate, and screened for reduced GFP-Lm infection. Whole cell lysates of putatively edited clones were prepared in RIPA buffer, and western blot against either E-cadherin (BD 610181) or c-Met (CST 4560) was carried out according to a standard protocol. Further, DNA from the samples with substantially lower infection than the wild type control and no detectable E-cadherin or c-Met as determined by western blot was extracted using the Quick Extract kit. PCR using Phusion High-Fidelity DNA Polymerase (NEB) was carried out genomic primers to genotype the indels for CDH1 and MET by cloning into the Zero Blunt cloning vector (Life Technologies) with subsequent Sanger sequencing at the UTSW Sequencing Core.
CDH1 guides: 1: ATAGGCTGTCCTTTGTCGAC; 2: CTCGACACCCGATTCAAAGT
MET guides: 1: GTATGCTCCACAATCACTTC; 2: GGCTACACACTGGTTATCAC
Two guides targeting exon 3 of FCGR1A were designed using the Optimized CRISPR Design Tool (http://crispr.mit.edu/), and cloned into lentiCRISPR v2 vector as previously described [58]. Lentiviruses were generated as described above and used to transduce THP-1 cells. Lentivirally transduced cells were selected in 2 μg/ml puromycin for 7 days 48h after transduction. The absence of FcγRIa on the cell surface in the generated cell line was confirmed by antibody staining as described above.
FCGR1A guides: 1: CTGGGAGCAGCTCTACACAG; 2: CACTGTGTAGAGCTGCTCCC.
In vitro phagocytosis assay was performed as described previously [78]. U-2 OS cells, stably expressing Fluc or FcγRIa were first transduced with lentivirus coexpressing TagRFP and Fluc, FcγRIIa or FcεRIg (γ-chain), to generate desired gene combinations. 48 h after transduction, cells were plated at 7x104 cells/ml in 4-well chamber slides (Falcon). The day before the assay latex beads (3.87μm in diameter) (Bangs Laboratories, PS05N/6749) were opsonized with human IgG by washing a 10% slurry of beads in 1×PBS and mixing overnight with 1.5 mg/ml human IgG (Jackson ImmunoResearch). The day of the experiment, beads were washed in 1×PBS and labeled with an Alexa Fluor 488 AffiniPure Donkey Anti-Human IgG (H+L) (green) antibody (Jackson ImmunoResearch), while rotating at room temperature for 1h. Following secondary labeling, beads were washed, resuspended in DMEM and added to cells in chamber slides. Slides were centrifuged at 300 x g for 1 min, and then placed 37°C for 90 min. After the incubation, slides were placed on ice and washed with ice-cold medium to inhibit further phagocytosis. Extracellular beads were then labeled with DyLight 405 AffiniPure Donkey Anti-Human IgG (H+L) (blue) antibody (Jackson ImmunoResearch) for 10 min on ice. Cells were washed 5 times with ice-cold 1×PBS and fixed with 3.7% PFA for 20 min at room temperature. Next, cells were washed with 1×PBS and incubated with 100 mM glycine for 10 min at room temperature to quench PFA. All samples were washed twice with 1×PBS and chamber removed from the slide. When the excess liquid dried, the coverslips were mounted on the samples with ProLong Gold reagent (Molecular Probes, Life Technologies). Samples were observed with a fluorescent microscope Zeiss Observer Z1. Numbers of green and blue beads were counted for 80 Red Fluorescent Protein (TagRFP)-positive cells per sample, two technical replicates per gene. Phagocytosis efficiency was measured as a percentage of internalized beads, determined by subtracting the number of extracellular (blue) beads from the total (green) beads, divided by the number of total (green) beads.
Cells were plated at 1.4x105 cells/well in a 12-well plate and transduced the next day with FcγRIa or Fluc-encoding lentiviruses as described above. Two days post-transduction, cells were infected with wild type Lm for 1 hr (MOI = 0.015, 0.05, 0.1), washed with medium, supplemented with 50 μg/ml gentamicin (Quality Biological), and then gently overlaid with 1.5ml/well of DMEM, containing with 10% FBS, 0.4% agarose, and 20 μg/ml gentamicin (Quality Biological). The overlay was allowed to stabilize for 15 min at room temperature, when plates were moved back to an incubator at 37°C. Foci of Lm infection were visualized 30 h after initial infection by adding 200μl of 5mg/ml (3-(4, 5-dimethylthiazolyl-2)-2,5-diphenyltetrazolium bromide (tetrazolium MTT) (Sigma) solution to each well and incubating at 37°C for 3 h. Plates were scanned and foci of infection quantified using ImageJ software.
Cells were plated at 1.4x105 cells/well in a 12-well plate and transduced the next day with lentiviruses as described above. Two days after transduction cells were split 1:2 on circular glass coverslips in 12-well plates, and the next day infected with Lm, according to the standard protocol. After infection samples were fixed in 2.5% glutaraldehyde in 0.1 M cacodylate buffer for a minimum of 2 h. Further procedures were performed at the UTSW Electron Microscopy Facility. Fixed cells were rinsed in the fixation buffer and fixed with Osmium tetroxide as secondary fixative. After several water rinses they were dehydrated in serial concentrations (50%, 70%, 85%, 95%, 100%), and critical point dried. The samples were coated for 30s with gold palladium and viewed in the Zeiss Sigma VP FE scanning electron microscope. Images were acquired using the Secondary Electron 2 (SE2) detector.
STAT1-deficient fibroblasts were seeded into 48-well plates at a density of 2.5x104 per well and transduced the following day with lentivirus expressing the gene of interest. 24 h later cells were transfected with 200 ng of the reporter plasmid pNF-kB–luciferase and 150 ng normalization vector pLacZ (to correct for transfection efficiency using a beta-galactosidase assay). 24 h after transfection, cells were lysed and luciferase was measured according to manufacturer protocol (Luciferase Assay System, Promega). LacZ expression was measured in a β-Galactosidase Activity Assay with ortho-Nitrophenyl-β-galactoside (ONPG), and used to normalize luciferase values for each sample.
All experiments were performed in as three independent replicates, unless otherwise stated. For experiments where only two groups of samples were compared, unpaired t-test was used to determine if difference between groups was statistically significant. To determine statistical significance in experiments with three or more groups of samples, one-way analysis of variance (ANOVA) with Dunnett’s procedure for multiple comparisons was used. Data analysis was performed in GraphPad Prism software.
|
10.1371/journal.ppat.1004307 | A Locus Encompassing the Epstein-Barr Virus bglf4 Kinase Regulates Expression of Genes Encoding Viral Structural Proteins | The mechanism regulating expression of late genes, encoding viral structural components, is an unresolved problem in the biology of DNA tumor viruses. Here we show that BGLF4, the only protein kinase encoded by Epstein-Barr virus (EBV), controls expression of late genes independent of its effect on viral DNA replication. Ectopic expression of BGLF4 in cells lacking the kinase gene stimulated the transcript levels of six late genes by 8- to 10-fold. Introduction of a BGLF4 mutant that eliminated its kinase activity did not stimulate late gene expression. In cells infected with wild-type EBV, siRNA to BGLF4 (siG4) markedly reduced late gene expression without compromising viral DNA replication. Synthesis of late products was restored upon expression of a form of BGLF4 resistant to the siRNA. Studying the EBV transcriptome using mRNA-seq during the late phase of the lytic cycle in the absence and presence of siG4 showed that BGLF4 controlled expression of 31 late genes. Analysis of the EBV transcriptome identified BGLF3 as a gene whose expression was reduced as a result of silencing BGLF4. Knockdown of BGLF3 markedly reduced late gene expression but had no effect on viral DNA replication or expression of BGLF4. Our findings reveal the presence of a late control locus encompassing BGLF3 and BGLF4 in the EBV genome, and provide evidence for the importance of both proteins in post-replication events that are necessary for expression of late genes.
| Epstein-Barr virus (EBV) is linked to the development of several types of cancer. Synthesis of structural proteins, a group of proteins that forms the protein shell around the viral genome, is essential for EBV infection and pathogenesis. Genes encoding structural proteins are expressed late in the viral life cycle after amplification of the viral genome. The mechanism controlling expression of this group of proteins represents a longstanding conundrum in EBV and other DNA viruses. In this report, we demonstrate that two EBV regulatory proteins control synthesis of mRNAs encoding viral structural proteins. These two proteins are: BGLF4, a protein kinase conserved in all herpesviruses, and BGLF3, a protein of unknown function with no cellular counterparts. We present evidence that the enzymatic activity of BGLF4 is required after replication of viral DNA to stimulate expression of structural proteins. BGLF3 and BGLF4 are expressed from the same locus in the genome; the two proteins work in concert and independently to promote expression of viral genes encoding structural proteins. Our findings provide novel insights into control of expression of genes encoding viral structural proteins. The enzymatic activity of BGLF4 is a potential target for development of new antiviral drugs.
| Late genes encode structural proteins necessary for virion assembly. A common theme among DNA viruses is the strict dependence of late gene expression on the onset of viral DNA replication. Disruption of replication, using inhibitors or mutating a replication-essential gene, blocks synthesis of late products. The link between these two processes led to models that focus on genome amplification as the principal regulator of late gene expression. These models propose that changes in DNA modifications such as a decrease in methylation as a result of de novo DNA synthesis, or displacement of viral or cellular repressors bound to elements in late promoters by the replication machinery trigger late gene expression. However, mechanisms that link late gene expression to replication have not been elucidated. Reports in herpesviruses suggest that replication per se is not sufficient to activate late gene expression. Pioneering work by Ren Sun and colleagues in murine herpesvirus-68 (MHV-68) identified four early viral proteins, ORF18, 24, 30 and 34, to be required for expression of late genes but dispensable for viral DNA replication [1], [2], [3], [4]. Homologs of ORFs 18, 24 and 34 in human cytomegalovirus (hCMV) map to three unique long (UL) sequences 79, 87 and 95, respectively [5], [6]. The function of the MHV68 and hCMV proteins in activating late gene expression has not been elucidated. In Epstein-Barr virus (EBV), the only protein so far characterized as essential for activation of late genes and not DNA replication is BcRF1, a homolog of ORF24 in MHV-68 and UL87 in CMV [7]. BcRF1 is a TATA box binding-like protein that specifically binds to a non-canonical TATA element (TATT) present in most late promoters [8], [9].
Viral late promoters differ from viral promoters of other kinetic classes and cellular promoters that rely on transcription factor-binding sites located upstream of the TATA box. Activation of late promoters is primarily dependent on a distinct TATA box and a downstream initiator sequence that spans the transcription start site (TSS) [9], [10], [11]. Involvement of upstream elements in regulation of late promoters is postulated to modulate transcription efficiency [12], [13], [14], [15].
BGLF4 is the only Ser/Thr protein kinase encoded by EBV [16]. Homologs of BGLF4 are conserved among the herpesvirus family [17]. These conserved human herpesvirus protein kinases (CHPKs) share sequence and positional similarities, but exhibit both unique and overlapping substrate specificities [18], [19]. BGLF4, like other CHPKs, functionally mimics cellular cyclin-dependent kinases but displays broader substrate specificity [17], [19], [20], [21], [22]. For instance, protein array phosphorylation experiments identified 21 EBV proteins as putative substrates of BGLF4; half of these proteins are shared targets with CDK1/cyclinB [23]. A number of cellular cyclin-dependent kinase substrates are also modified by BGLF4 such as pRB, p27, condensin, MCM4, stathmin, elongation factor 1 delta, and nuclear lamin A/C [19], [20], [21], [22], [24], [25], [26], [27], [28], [29]. Phosphorylation of lamin A/C, another CDK substrate [30], by BGLF4 causes dissolution of the nuclear lamina, a step considered essential for egress of viral capsids from the nucleus [27], [31], [32].
BGLF4 has been extensively investigated to determine its potential role in viral DNA replication. Expression of BGLF4 occurs during the early phase of the lytic cycle and reaches maximal levels following viral DNA replication [33]. BGLF4 localizes to replication compartments, a site of viral DNA replication and late gene expression [34], [35]. BGLF4 modulates gene expression by its capacity to phosphorylate transcription factors, histones and chromatin modifying enzymes, such as BMRF1, ZEBRA, EBNA2, EBNA-LP, IRF3, UXT, HDAC1, H1, and TIP-60 [16], [36], [37], [38], [39], [40], [41], [42], [43], [44], [45]. However, the evidence in favor of a role of BGLF4 in regulating early gene expression and viral DNA replication during the EBV lytic cycle is conflicting. Abolishing expression of BGLF4 had no significant effect on early gene expression [46] and had little (<2-fold) to no impact on viral DNA replication [28], [31], [32], [44], [45], [47]. Perhaps the most established phenotype for disrupting the expression or the activity of BGLF4 is the significant reduction in the amount of virus released by lytic infected cells [27], [31], [48]. This phenotype is consistent with a crucial role for BGLF4 in events occurring after viral DNA replication.
In this study, we employed knockout and knockdown approaches to thoroughly investigate the role of BGLF4 in regulating lytic viral gene expression and its correlation with DNA replication. We found that the kinase activity of BGLF4 is necessary for expression of late genes but not viral DNA replication. Using RNA-seq analysis we demonstrated that expression of most late genes is reduced in the absence of BGLF4. We identified a late control locus encompassing BGLF4 and BGLF3. Knockdown of BGLF3 abolished late gene expression independent of any effect on the level of BGLF4 or viral DNA replication. Our findings identify two EBV genes as regulators of late gene expression and establish the presence of additional checkpoints beyond viral DNA replication that are necessary for expression of EBV structural proteins.
BGLF4 is transcribed as a multicistronic message that also encodes BGLF5, a DNA alkaline exonuclease involved in host cell shutoff [47], [49]. To study the role of BGLF4 in regulating the lytic cascade we used delta G4/G5 cells, which are 293 cells harboring BGLF4/BGLF5-null EBV bacmid virus [47]. Cells were harvested 48 h after transfection to allow for analysis of viral DNA replication and late gene expression. Typically, transient expression of ZEBRA in EBV positive cells induces expression of three major kinetic classes of lytic genes, these are: very early, early and late genes [50], [51]. In delta G4/G5 cells, we found that ZEBRA activated the EBV lytic cycle as demonstrated by expression of the early bmrf1 gene encoding the DNA polymerase processivity factor (Fig, 1A lane 2). However, expression of the late BFRF3 protein (FR3), the minor viral capsid protein, was noticeably low, only 2.5-fold higher in cells transfected with ZEBRA relative to empty vector (CMV) (Fig. 1A, lanes 1 and 2). Complementation with BGLF4 significantly enhanced the level of FR3 to 20.5-fold relative to empty vector and 8.2-fold relative to ZEBRA-transfected cells (Fig. 1A, lane 3). Complementation with BGLF5 did not enhance expression of FR3, a result suggesting that the effect on late gene expression is specific to BGLF4. BMRF1 is a bona fide substrate of the BGLF4 kinase [16], [36], [52]. Therefore, hyper-phosphorylation of BMRF1 is a marker for the intracellular kinase activity of BGLF4. Hyper-phosphorylated BMRF1 was only detected in lytic cells transfected with the BGLF4 expression vector (Fig. 1A, compare lanes 2 and 5 with 3 and 4). The experiment was repeated to confirm that BGLF4, and not BGLF5, reproducibly up-regulated expression of FR3 when expressed alone or together with BGLF5 (Fig. S1).
Since FR3 is a late protein, we examined the possibility that lack of BGLF4 might disrupt viral DNA replication and thus compromise expression of FR3. To assess the effect of BGLF4 on the extent of viral genome amplification we prepared genomic DNA from aliquots of the same cells that were examined for protein expression. The relative concentration of viral DNA was assessed using primers specific to the upstream region of oriLyt and quantitative polymerase chain reaction (qPCR). We found that expression of BGLF4 slightly increased viral DNA replication by 1.8-fold relative to cells transfected with ZEBRA alone or with ZEBRA plus BGLF4 and BGLF5 (Fig. 1B, lanes compare lane 3 with 2 and 4). Over-expression of ZEBRA and BGLF5 reduced the extent of EBV genome amplification by 50% relative to ZEBRA alone (compare lanes 2 and 5). Thus lack of BGLF4 lead to a modest reduction (1.8-fold) in viral DNA replication but caused a more pronounced defect (8.2-fold) on synthesis of FR3 late protein.
Lysine 102 in BGLF4 corresponds to a conserved lysine present in the catalytic domain of protein kinases [53]. Substitution of lysine 102 to isoleucine abolishes the kinase activity of BGLF4 [20]. To determine whether the kinase activity of BGLF4 is required for its effects on viral DNA replication and synthesis of FR3, we expressed BGLF4 or BGLF4(K102I) together with ZEBRA in delta G4/G5 cells and harvested the cells after 48 h. We did not detect any difference in viral DNA replication as a result of expressing the kinase-active or -inactive forms of BGLF4. Both wild-type and mutant BGLF4 equally enhanced viral DNA replication by 2.6-fold relative to ZEBRA alone (Fig. S2B, lanes 2, 4, and 6). However a clear difference was observed between BGLF4 and BGLF4(K102I) in promoting expression of the FR3 late gene (Fig. S2A). Wild-type BGLF4 increased expression of FR3 by 9-fold whereas the kinase dead mutant of BGLF4 enhanced the level of FR3 by 1.66-fold. Wild-type and mutant forms of BGLF4 protein were equally expressed. The level of BMRF1 protein did not change when the kinase activity of BGLF4 was disrupted (Fig. S2A), but the hyper-phosphorylated form of BMRF1 was only detected in cells expressing active BGLF4 and not the kinase dead mutant. These findings show that the kinase activity of BGLF4 is necessary for enhanced expression of FR3 but is not required for the low level of stimulation of viral DNA replication by BGLF4.
The previous experiment suggested that the kinase activity of BGLF4 is necessary for up-regulating expression of the late FR3 protein. To further examine this point, we expressed increasing concentrations of wild-type or mutant BGLF4 in the presence of a constant amount of ZEBRA. After 48 h, transfected delta G4/G5 cells were harvested and the levels of FR3, BMRF1, BGLF4 and ZEBRA proteins were assessed using Western blot analysis. By comparing the effect of the two forms of BGLF4, we found that as we increased the level of wild-type BGLF4 there was a progressive increase in the amount of FR3 protein expressed. The increase in the amount of BGLF4 protein expressed was associated with an increase in the level of hyper-phosphorylated form of BMRF1 (Fig. 2A). On the contrary, in cells expressing increasing levels of the kinase inactive form of BGLF4, expression of FR3 remained constant at a low level and BMRF1 was not hyperphosphorylated.
Using qPCR we studied the effect of expressing various concentrations of BGLF4 and BGLF4(K102I) on viral DNA replication; we prepared DNA from the same cells used in figure 2A. We only detected a change in the amount of viral DNA when we transfected 6 µg of wild-type BGLF4. At this input, replication increased by 1.6-fold relative to expression of ZEBRA alone (compare lanes 2 and 5). Findings from Figures S2 and 2 strongly indicate that the effect of BGLF4 on late gene expression is kinase dependent and is separate from its effect on viral DNA replication.
In the next experiment we asked whether the effect of BGLF4 on expression of FR3 could be detected at the transcript level, and whether BGLF4 regulated mRNA expression of other late genes. We prepared total RNA from fractions of the same samples that were previously used to generate figure S2 and employed quantitative reverse transcriptase polymerase chain reaction (RT-qPCR) to assess the level of five lytic transcripts. Four of these EBV transcripts, encoding the major and minor capsid proteins (BcLF1 and BFRF3), the major glycoprotein gp350/220 (BLLF1), and a tegument protein (BLRF2), have late kinetics and one, encoding BMRF1, has early kinetics. When co-expressed with ZEBRA, wild type BGLF4 or BGLF4(K102I) enhanced the level of BMRF1 by 2.17 and 1.8-fold, respectively (Fig. 3). This result suggested that BGLF4 had a limited stimulatory effect on expression of the early BMRF1 transcript in a manner independent of its kinase activity. However, the kinase activity of BGLF4 was necessary for efficient expression of the four late transcripts examined. The level of all four late transcripts was less in delta G4/G5 cells expressing ZEBRA alone or ZEBRA plus G4(K102I) compared to co-expression of ZEBRA and wild-type BGLF4 (Fig. 3). Expression of late transcripts was reduced by an average of 7.9-fold (range 9.7 to 5.4) in cells expressing ZEBRA alone and a mean of 8.4-fold (range 11.6 to 5.8) in cells expressing ZEBRA plus the kinase dead mutant. These results indicate that the kinase activity of BGLF4 markedly up-regulates expression of late but not early transcripts.
To eliminate the possibility that the markedly enhanced synthesis of EBV late products in delta G4/G5 cells following ectopic expression of BGLF4 was due to its overexpression, we used siRNA to knockdown endogenous BGLF4. For these experiments we used BZKO cells, which are 293 cells containing an EBV bacmid that lacks a functional gene for ZEBRA, to study the effect of silencing BGLF4 on viral DNA replication and late gene expression. Transfection of ZEBRA (2 µg) induced the lytic cycle and activated expression of endogenous BGLF4 (Fig. 4A, lane 2). To knock down expression of BGLF4 we co-transfected four concentrations, 2, 5, 10 and 15 nM, of siRNA specific to the viral protein kinase (siG4) (Fig. 4A, lanes 3 to 6). Cells were harvested after 48 h and expression of the FR3 protein was assessed using Western blot analysis. We found that all four concentrations of siG4 abolished expression of BGLF4. The reduction in the level of the late FR3 protein correlated with the amount of siG4 used to knockdown expression of BGLF4. This reduction in the level of the FR3 protein ranged between 9.8- to 12.3-fold relative to lytic cells without siG4 (Fig. 4A, lanes 3 to 6). We used qPCR to assess the effect of abolishing expression of BGLF4 on replication of the EBV genome. Knocking down of endogenous BGLF4 reduced EBV DNA levels by 2.1- to 2.4-fold (Fig. 4B, compare lane 2 with lanes 3 to 6). The effect of siRNA to BGLF4 on replication was similar in magnitude to that observed in previous experiments using delta G4/G5 cells in which expression of either wild-type or the kinase dead mutant of BGLF4 enhanced expression of BMRF1 and viral DNA replication by ∼2-fold (Figs. 1, S2 and 2).
To attempt to segregate the effect of BGLF4 on viral DNA replication from that on late gene expression we gradually increased the level of BGLF4 expressed by titrating the amount of siG4 (15, 10, 5 and 2 nM) co-transfected together with fixed concentrations of plasmids encoding ZEBRA (2 µg) and BGLF4 (3 µg). At the highest concentration of siG4 (15 nM), the level of the FR3 protein was markedly reduced by 7.5-fold compared to cells without siG4 (Fig. 4A, compare lanes 7 and 8). However, the level of viral DNA in the same cells was not compromised as a result of knocking down BGLF4; on the contrary the amount of viral DNA present was slightly higher relative to cells expressing ZEBRA alone or ZEBRA plus BGLF4 (Fig. 4B, lanes 8–11). As the level of BGLF4 expressed increased due to decreasing the amount of siG4 transfected, there was a proportional increase in the amount of the FR3 late protein, but the level of viral DNA replication was unaffected. These results establish a direct relationship between the level of BGLF4 and expression of the late FR3 protein independent of any effect of BGLF4 on viral DNA replication.
Specificity of siRNA represents a major concern in this widely used approach to down-regulate gene expression. To determine whether the effects of siG4 on late gene expression were a result of down-regulating BGLF4 or an off-target effect, we generated silent mutations in the bglf4 gene that abolished complementarity between the BGLF4 transcript and siG4 without altering the amino acid sequence. Transfection of siG4 is expected to eliminate expression of endogenous BGLF4 and late genes. However, providing the siG4-resistant BGLF4 ectopically should restore expression of late genes, only if the effect of siG4 is target specific. Two forms of the siG4-resistant BGLF4 were generated, kinase active and kinase dead, referred to as rG4 and rG4(K102I), respectively. To study the effect of siG4 on expression of rG4 and rG4(K102I), we transfected 2089 cells, 293 cells containing wild-type EBV bacmid, with two concentrations of siG4 (10 and 20 nM) together with expression vectors of ZEBRA and rG4 or rG4(K102I). Expression of ZEBRA by itself in 2089 cells triggered the lytic program including expression of the endogenous BGLF4 (Fig. 5A, lane 2). Both concentrations of siG4 were sufficient to abolish expression of ectopic and endogenous BGLF4 in cells co-transfected with ZEBRA and BGLF4 (Fig. 5A, lanes 4 and 5). Expression of rG4 and rG4(K102I) was not affected by the same concentrations of siG4 (Fig. 5A, lanes 7, 8, 10 and 11). This outcome allowed us to assess the specificity of siG4 by restoring expression of BGLF4 and studying its effect on late gene expression. If siG4 non-specifically abolished expression of other proteins that are crucial for late gene expression, then restoring expression of BGLF4 by rG4 would not be sufficient to rescue expression of FR3. We found that expression of rG4 and ZEBRA in the presence of siG4 restored expression of FR3 to normal levels (lanes 7 and 8). However, providing rG4(K102I) instead of rG4, failed to restore the level of FR3 protein (lanes 10 and 11). These findings showed that down-regulation of FR3 in cells transfected with siG4 is due to abolishing expression of BGLF4 with intact kinase activity. Lane 9 shows that cells expressing the transfected kinase dead protein and endogenous BGLF4 are permissive for expression of FR3. While rG4 restored late gene expression, neither rG4 nor rG4(K102I) had any significant effect on viral DNA replication (Fig. 5B).
We studied the effect of silencing BGLF4 on the level of eight lytic transcripts in 2089 cells expressing ZEBRA either alone or together with rG4 or rG4(K102I) (Fig. 6). Two of these transcripts, BRLF1 and BMRF1, are synthesized prior to viral DNA replication. The brlf1 gene encodes the Rta protein, a transcription activator and a replication protein, and the bmrf1 gene codes for the polymerase processivity factor. Both proteins play vital roles in the process of viral DNA replication [54], [55]. The other six transcripts, with late kinetics, encode the following proteins: major glycoprotein gp350/220 (BLLF1), major capsid protein (BcLF1), tegument protein (BLRF2), triplex capsid protein (BDLF1), scaffold protein (BdRF1), and minor capsid protein (BFRF3). Knockdown of endogenous and transfected BGLF4 by siG4 reduced the level of late transcripts by a median fold-change of 9.1. Conversely, expression of rG4 restored the amount of late transcripts to a level equivalent to that observed in cells transfected with ZEBRA alone. However, expression of rG4(K102I) failed to up-regulate the level of all six late transcripts and resulted in a median reduction of 8.6-fold relative to ZEBRA alone. These results demonstrate a specific role for the kinase activity of BGLF4 in regulating the level of several late viral transcripts.
To study the effect of BGLF4 on the EBV transcriptome during lytic induction, we performed deep sequencing analysis on RNA purified from 2089 cells transfected with empty vector, ZEBRA or ZEBRA plus siG4. To demonstrate that siG4 had silenced BGLF4 in this particular experiment we assessed the levels of BGLF4 and FR3 proteins by Western blot. As expected, ZEBRA induced expression of BGLF4 and FR3; addition of siG4 abolished expression of BGLF4 and the late FR3 protein (data not shown). We used the Expectation-Maximization (EM) algorithm in RSEM to estimate gene expression levels and EBSeq to identify differentially expressed genes [56], [57]. Figure 7 compares the number of normalized reads corresponding to each viral transcript expressed in the absence and presence of BGLF4. The change in expression is represented as a log base 2-fold. Red bars designate late genes; blue bars designate latent and early kinetic classes of lytic genes. Determination of the kinetic classes of EBV genes was based on viral gene classification reported by Yuan et al [58]. Viral transcripts with a negative fold-change represent EBV genes whose expression is BGLF4-dependent, and vice versa. The most striking result was that silencing of BGLF4 down-regulated the level of thirty-one late transcripts. The level of only three late transcripts, encoding BVRF1, BILF1 and BMRF2, increased by 1.2-, 1.8-, and 3.2-fold, respectively (Fig. 7). BVRF1 is a tegument protein and is a homolog of HSV UL25 (cork), BILF1 is a virally encoded G protein coupled receptor, and BMRF2 is a transmembrane envelope protein [59], [60], [61]. siRNA to BGLF4 did not reduce the level of transcripts encoding components of the replication machinery. Silencing of BGLF4 impaired expression of two lytic transcripts of unknown kinetic class, namely, BGLF3.5 and BGLF3 (Fig. 7) [58]. The reduction in the level of these two transcripts in lytic cells expressing BGLF4 compared to cells without BGLF4 was 11.6-fold for BGLF3.5 and 7.3-fold for BGLF3. The functions of BGLF3.5 and BGLF3 in the EBV lytic cycle remain to be characterized. However, orf34, a homologue of BGLF3 encoded by murine gamma herpesvirus-68 (MHV-68), is an early gene that was found to be essential for late gene expression and not viral DNA replication [4]. These findings demonstrated that BGLF4 globally enhanced expression of EBV late genes.
RNA-seq analysis showed that silencing of BGLF4 reduced expression of BGLF3 and BGLF3.5. We sought to validate the effect of siG4 on the level of transcripts encoding BGLF3 and BGLF3.5 using RT-qPCR. The data presented in figure S3 represent the mean of two biological replicates. Similar to our observations with the RNA-seq analysis, we found that expression of the bglf3 and bglf3.5 genes was reduced in cells lacking BGLF4 by 6-fold and 22-fold, respectively.
The effect of siG4 on BGLF3 and BGLF3.5 could be attributed to two different mechanisms: siRNA to BGLF4 simultaneously reduces the BGLF3 and BGLF3.5 transcripts; BGLF4 regulates synthesis or stability of BGLF3 and BGLF3.5 mRNAs. The bglf4, bglf3.5 and bglf3 genes are located in a transcription unit with a single poly(A) signal (Fig. 8D). Previous reports suggested that transcripts synthesized from this region are co-terminal [33], [62]. To assess the number and size of transcripts expressed from this locus we purified RNA from 2089 cells transfected with ZEBRA. Using Northern blot analysis and probes directed to the unique sequences in BGLF4 and BGLF3.5 we detected five RNAs that were 4.3, 3.5, 3.3, 2.2 and 1.8 kb long (Fig. 8A and C). Based on the lengths of the genes and the distribution of TATA boxes present in this locus, the 4.3 kb RNA correlates with transcript A (BGLF3-BGLF3.5-BGLF4-BGLF5-BBLF1). siRNA to BGLF3 abolished the 4.3 kb transcript without affecting the level of the other transcripts (Fig. 8C). This result demonstrates that the 4.3 kb RNA is the only transcript that contains the BGLF3 sequence, which is consistent with the composition of transcript A. The 3.5 kb RNA corresponds with transcript B (BGLF3.5-BGLF4-BGLF5-BBLF1); the 3.2 kb RNA with transcript C (BGLF4-BGLF5-BBLF1), and the 1.8 kb RNA with transcript D (BGLF5-BBLF1) (Fig. 8D). The 1.8 kb RNA was detected by the BGLF4 probe presumably as a result of containing BGLF4 sequence in the 5′ UTR of RNA species D (Fig. 8A). A 2.2 kb RNA was only detected using the BGLF4 probe (Fig. 8A); its size does not correlate with any of the predicted RNA transcripts.
siRNA to BGLF4 abolished four RNA species: 4.3, 3.5, 3.3 and 2.2 kb. This outcome conforms to the predicted composition of transcripts A, B, and C, in which all three transcripts contained the BGLF4 sequence (Fig. 8D). Thus, targeting BGLF4 with siG4 silences expression of BGLF3.5 and BGLF3 as well. Transfection of siG4 resulted in two new RNA species that were 1.6 and 3 kb long and were only detected by the BGLF4 probe (Fig. 8 A). The appearance of these two RNAs is consistent with the capacity of siG4 to trigger cleavage of the BGLF4 containing transcripts.
To determine whether BGLF4 has the capacity to increase the level of BGLF3 mRNA either by up-regulating its expression or enhancing the stability of its mRNA, we compared the level of BGLF3 mRNA in 2089 cells transfected with ZEBRA, ZEBRA plus siG4 or ZEBRA plus siG4 and rG4 (siG4 resistant BGLF4). Addition of siG4 reduced the level of BGLF4 and BGLF3 transcripts expressed from the endogenous viral genome. Ectopic expression of rG4 increased the level of BGLF3 transcript by 4-fold relative to cells transfected with ZEBRA plus siG4 (Fig. S4).
These results suggest that siG4 reduces the level of BGLF3 mRNA by two different mechanisms; it directly silences expression of BGLF3, and it knocks down expression of BGLF4, a protein that augments expression or stability of the BGLF3 transcript.
To investigate the possibility that BGLF4 might exert its effects on late gene expression by controlling expression of a second tier of late gene regulators we studied the effect of siRNA to BGLF3 and BGLF3.5 on synthesis of late products. 2089 cells transfected with two different concentrations of ZEBRA expression vector (1 and 2 µg DNA) in the presence of siRNA were harvested after 48 h. Cell lysates were analyzed using Western blot assays. The experiment illustrated in figure 9A is a representative of two biological replicates. Knockdown of BGLF3 significantly reduced expression of the late FR3 protein by an average of 9.5-fold, but siRNA to BGLF3 (siG3) had no effect on expression of ZEBRA, BGLF4, or EA-D (Fig. 8B and 9A). Silencing of BGLF3.5 reduced FR3 expression by an average of 2.7-fold. However, siRNA to BGLF3.5 (siG3.5) slightly reduced expression of BGLF4 but not ZEBRA or EA-D. To assess the effect of both siRNAs on viral DNA replication we purified DNA from the same cells and examined the level of viral DNA using qPCR. Neither of the two siRNAs reduced the level of viral DNA in cells expressing ZEBRA (Fig. 9B). We used RT-qPCR to demonstrate that the siRNA to BGLF3 reduced the level of the endogenous BGLF3 transcript by 2.7-fold, but had no effect on the level of BMRF1 mRNA (Fig. S5). Expression of four other late transcripts, BDLF1, BdRF1, BLLF1, and BcLF1, was reduced as a result of silencing BGLF3 (Fig. S6). These results demonstrate that BGLF3 is necessary for progress of the EBV lytic cycle into the late phase.
To demonstrate that the effect of siG3 on late gene expression was specific to knockdown of the bglf3 gene, we inserted silent mutations in the region of the BGLF3 mRNA that is recognized by siG3. These mutations disrupt the complementarity between siG3 and the BGLF3 transcript without altering the amino acid sequence of the protein. Insertion of these mutations resulted in a form of BGLF3, referred to as rG3, which is resistant to siG3. Co-transfection of ZEBRA plus siG3 in 2089 cells reduced the level of FR3 by 7-fold relative to cells solely transfected with ZEBRA (Fig. 10A, lanes 2 and 3). Ectopic expression of rG3 in cells transfected with ZEBRA and siG3 restored expression of the late FR3 protein to a level equivalent to that observed in cells expressing ZEBRA alone (Fig. 10A, compare lanes 2 and 4). The experiment was repeated twice and similar results were obtained. To assess the effect of rG3 on expression of other late genes, we purified total RNA from the same set of samples and examined expression of BMRF1, an early gene, and two late genes: BcLF1, and BdRF1. Using RT-qPCR, we found that siG3 reduced the level of the two late transcripts but had no effect on the level of the early BMRF1 transcript; co-transfection of rG3 rescued expression of BcLF1 and BdRF1 (Fig. 10B). These results demonstrate that BGLF3 is indispensable for expression of late genes.
To determine whether BGLF3 or BGLF3.5 can substitute for the role of BGLF4 in activation of late gene expression, we knocked down BGLF4 by transfecting siG4 and expressed BGLF3 and BGLF3.5, separately and together, using expression vectors. We found that provision of BGLF3 and BGLF3.5 was not sufficient to trigger expression of the late FR3 protein in absence of BGLF4 (Fig. S7). Similarly, ectopic expression of BGLF3 and BGLF3.5 did not restore late gene expression in cells treated with phosphonoacetic acid (PAA), an inhibitor of viral DNA replication (Fig. S7, lanes 8 and 9). These results demonstrate that BGLF4, BGLF3 and viral DNA replication are three necessary components for progress of the EBV lytic cycle into the late phase.
Classification of lytic viral gene expression into pre- and post-replication temporal phases is applicable to the life cycles of all herpesviruses. While kinetics of expression of pre-replication genes has been extensively studied, little is known about the order of events that take place prior to expression of late genes. Our findings reveal a previously unknown transitional step between EB viral DNA replication and late gene expression. We demonstrate that the kinase function of BGLF4 is necessary for optimal expression of late genes. Inactivating the kinase activity of BGLF4 markedly impairs late gene expression without impeding viral DNA replication; the kinase dead mutant supported viral DNA replication to the same extent as wild-type BGLF4 (Fig. S2, 2, and 5). Disrupting the kinase activity or the expression of BGLF4 selectively reduced the level of late transcripts (Fig. 3, 6 and 7). Analysis of the EBV transcriptome revealed that expression of 31 late genes was reduced in the absence of BGLF4 (Fig. 7). Silencing of BGLF4 did not reduce the level of most early transcripts including those encoding replication proteins. Using siRNA to compare the expression pattern of EBV genes with and without BGLF4 we identified two non-late genes, bglf3 and bglf3.5, whose expression was significantly reduced as a result of silencing BGLF4 (Fig. 7, 8 and S3). By knockdown experiments, we found that BGLF3 is an independent regulator of late gene expression. siRNA to BGLF3 markedly reduced the level of FR3, a canonical late protein, and the transcripts of several other late genes without affecting endogenous expression of BGLF4 or viral DNA replication (Fig. 8B, 9A, 10 and S6). Expression of siRNA resistant forms of BGLF3 and BGLF4 provided evidence that each protein plays an independent and indispensable role in regulation of late gene expression (Fig. 5, 6, and 10). In summary, our results identify a control locus composed of BGLF4 and BGLF3 that regulates synthesis of EBV encoded structural proteins (Fig. 11).
To establish a distinct regulatory role of BGLF4 in activation of late gene expression it is imperative to exclude that this role results from an effect on viral DNA replication. Several groups have assessed the effect of knocking down or knocking out BGLF4 on the process of EBV genome amplification. Two separate research groups reported that knocking out the bglf4 gene did not affect the level of viral DNA synthesized during the lytic cycle [28], [32]; three other groups found that lack of BGLF4 reduced viral DNA replication by 1.4- to 2-fold [31], [44], [47]. We found that ectopic expression of BGLF4 in BGLF4 null cells enhanced viral DNA replication by 1.6 to 2.6-fold (Fig. 1, S2, and 2) in agreement with the second group of studies. These slight differences in the effect of BGLF4 on viral DNA replication could be attributed to the cell background or the approach used to knockout the bglf4 gene. In our knockdown experiments we found that silencing of BGLF4 in BZKO cells reduced viral DNA replication by 2-fold and in 2089 cells by 1.1-fold (Fig. 4 and 5). The effect of BGLF4 on viral DNA replication was independent of its kinase activity. In delta G4/G5 cells, expression of either wild-type BGLF4 or the kinase inactive mutant enhanced replication to same extent (Fig. 2 and S2). Thus, we conclude that the kinase activity of BGLF4 is required for its effect on late gene expression but not for its modest enhancement of viral DNA replication.
Since expression of late genes is stipulated by the onset of viral genome amplification we sought to segregate the effect of BGLF4 on replication from its effect on expression of late genes. We achieved this goal using two different approaches: first, we varied the concentration of siRNA to BGLF4. We demonstrated that at a very low concentration of BGLF4 protein there was significant reduction in late gene expression without any significant compromise in viral DNA replication (Fig. 4, compare lane 7 with lane 8). As we increased the concentration of BGLF4, viral DNA replication remained unchanged (Fig. 4B) while late gene expression increased proportionally (Fig. 4A). Second, we supplied a kinase dead form of BGLF4 to cells with a knockout of the bglf4 gene. Viral DNA replication was rescued to normal levels but there remained a marked defect in late gene expression (Fig. S2 and 2).
The possibility that BGLF4 possesses additional functions that are independent of its kinase activity is supported by the capacity of ORF36, a BGLF4 homolog encoded by MHV68, to repress the function of class 1 and 2 histone deacetylases (HDACs) [43]. Repression of HDAC1 and 2 by ORF36 is mediated by protein-protein interactions, rather than by phosphorylation. Expression of either wild-type or kinase inactive ORF36 enhances MHV68 early gene expression and viral DNA replication [43]. The effect of BGLF4 or BGLF4(K102I) on EB viral DNA replication may also be due to the capacity to interact with and disrupt the function of HDAC1 and 2 during the early phase of the lytic cycle. BGLF4 interacts with HDAC1 and 2 [43]; the biological significance of such interaction for temporal control of the EBV life cycle has not been studied.
Expression profiling of EBV genes during lytic infection affirms that most late genes are subject to regulation by BGLF4 (Fig. 7). By comparing the EBV transcriptome during lytic infection in the absence and presence of BGLF4 we found that BGLF3 and BGLF3.5 represent the two most down-regulated non-late genes in cells transfected with siG4 (Fig. 7). The effect of siG4 on the level of BGLF3 and BGLF3.5 mRNAs could be attributed to two different mechanisms. First, in addition to targeting the BGLF4 mRNA, siG4 might concurrently knock down the transcripts encoding BGLF3 and BGLF3.5. bglf3, bglf3.5 and bglf4 are nested within a transcription unit containing five overlapping open reading frames that also includes bglf5 and bblf1, respectively (Fig. 11). Due to the presence of a single canonical poly(A) signal at the 3′ end of bblf1, transcription from this locus is likely to result in co-terminal transcripts (Fig. 8D) [33], [62], unless cryptic variants of the poly(A) signal exist at the 3′end of some genes [63]. In figure 8A, knockdown of BGLF4 abolished all the RNA species that were predicted to contain BGLF4 sequence. This result is in favor of co-terminal expression of transcripts encoded by the late control locus. Second, BGLF4 might regulate the level of the BGLF3 and BGLF3.5 mRNAs. In Fig. S4, siG4 down-regulated the level of BGLF3 mRNA; provision of rG4 overcame the effect of siG4 and increased the level of the endogenous BGLF3 transcript by 4-fold. BGLF4 might either enhance the activity of the BGLF3 promoter or enhance the stability of the BGLF3 mRNA.
Silencing of BGLF3 had no discernable effect on the protein level of BGLF4 (Fig. 8B and 9A). A possible explanation of this result is that BGLF3 is expressed at a low level during the late phase of the lytic cycle. Northern blot analysis revealed that the 4.3 kb transcript, the only RNA containing the BGLF3 sequence, was considerably less abundant relative to the other RNA species synthesized from this locus (Fig. 8A and C). Therefore, knockdown of the 4.3 kb RNA (Fig. 8D, transcript A) presumably had more impact on the level of BGLF3 relative to the overall level of BGLF4 containing mRNAs (Fig. 8C and D). In two independent experiments, abolishing the BGLF3 containing transcript by siG3 reduced the abundance of the FR3 protein without affecting the level of the BGLF4 protein (Fig. 8B and 9A).
Contrary to siG3, siRNA to BGLF3.5 (siG3.5) reduced the level of the BGLF4 protein by 3-fold (Fig. 9A, lanes 3 and 7). BGLF4 has three transcriptional start sites; two sites mapped to regions within the bglf3.5 coding sequence and a third site mapped upstream of the translation start codon of bglf3.5. Reduction in the level of BGLF4 protein by siG3.5 is likely due to the ability of siG3.5 to target transcripts that encode both BGLF3.5 and BGLF4 (Fig. 11). In KSHV, ORF36, the homolog of BGLF4, is robustly translated as a downstream cistron from a polycistronic transcript that initiates with ORF35, the KSHV homolog of BGLF3.5 [64], [65].
The kinase activity of BGLF4 and not the mere synthesis of the BGLF4 transcript or protein is necessary for stimulation of late gene expression (Fig. 2, S2, 3, 5, and 6). In knockout and knockdown experiments of BGLF4, expression of the kinase active, and not the kinase dead, form of BGLF4 activated synthesis of late transcripts. Attempts to complement the lack of BGLF4 with expression of BGLF3, BGLF3.5 or both did not restore expression of late genes (Fig. S7). Therefore, the role of BGLF4 in late gene expression is not limited to regulating the level of BGLF3; BGLF4 has an unknown but indispensable function in the mechanism regulating synthesis of late products. BGLF3 is an ortholog of the MHV68-encoded ORF34 [4]. Mutant MHV68 virus that lacks the orf34 gene underwent viral genome amplification but failed to express late products. Here, we found that the function of MHV-68 ORF34 in inducing late gene expression is conserved in EBV BGLF3. Silencing expression of BGLF3 using siRNAs abolished synthesis of the FR3 late protein and the transcripts of several late genes without affecting the level of BGLF4 or viral DNA replication (Fig. 8, 9, and S6). Providing a form of BGLF3 that is resistant to siG3 annulled the effect of the siRNA on expression of late genes. The exact role of BGLF3 in regulation of EBV late gene expression has not been established. Knockout of ORF34, the murine homolog of BGLF3, abolished recruitment of RNAPII to late promoters. In a related observation, we found that BGLF3 interacts with the C-terminal domain of RPB1, the large catalytic subunit of RNAPII (data not shown). These results represent the first demonstration of the indispensable roles of BGLF4 and BGLF3 in regulation of EBV late gene expression. However, ectopic expression of BGLF4 or BGLF3 does not complement the lack of viral DNA replication (Fig. S7, lanes 8 and 9, and data not shown). Therefore, viral DNA replication as well as the function of BGLF4 and BGLF3 represent three essential components for stimulation of late gene expression. Our findings have scientific and translational implications. The findings provide new insight on regulation of late gene expression, one of the main puzzles in virology, and emphasize the importance of targeting the kinase activity of BGLF4 for development of new antiviral drugs.
The vector encoding the ZEBRA protein was prepared as previously described [66]. Constructs expressing wild-type BGLF4 and the kinase dead BGLF4(K102I) were a kind gift of Dr. Mei-Ru Chen [16]. BGLF4 forms that are resistant to the corresponding siRNA (siG4) were generated by introducing silent point mutations using the following mutagenic primer: 5′-GTGACCAACATTGATGACATGACGGAGACATTATACGTCAAATTACCTGAAAACATGACGCGCTGTGATCACCTCCCCATTACC-3′ and its complementary strand: 5′-GGTAATGGGGAGGTGATCACAGCGCGTCATGTTTTCAGGTAATTTGACGTATAATGTCTCCGTCATGTCATCAATGTTGGTCAC-3′.
2089, Bam Z knockout (BZKO), and delta BGLF4/BGLF5 (delta G4/G5) are 293 human embryonic kidney (HEK) cells stably transfected with bacmids containing wild-type, BZLF1 null, and BGLF4/BGLF5 null EBV B95.8 genomes, respectively [47], [67], [68]. Mutations in wild type EBV bacmid (2089) were generated by homologous recombination in which the bzlf1 gene and the bglf4 gene were replaced with the kanamycin resistance gene [47], [67]. Disruption of the bglf4 gene abolished expression of BGLF4 and BGLF5. The cells were cultured in Dulbecco's modified Eagle medium (DMEM) supplemented with 10% fetal bovine serum (FBS), and antibiotics (penicillin-streptomycin at 50 units/ml and amphotericin B at 1 µg/ml). Hygromycin B (Calbiochem) 100 µg/ml was added to the medium to select for 293 cells containing wild type or mutant EBV bacmids. Transfection of eukaryotic plasmids was performed in 25 cm2 flasks using 36 µl DMRIE-C transfection reagent (Invitrogen) mixed with 2 µg of ZEBRA expression vector to induce the lytic cycle, and 3 µg of plasmids encoding wild-type or kinase dead BGLF4.
Silencing expression of BGLF4, BGLF3 and BGLF3.5 was attained using RNA interference technique. Transfections were performed using Lipofectamine 2000 (Invitrogen) following the manufacturer's protocol. Several siRNAs were tested to knockdown BGLF4, BGLF3, and BGLF3.5. The most specific siRNA was then used in knockdown experiments to study the role of each of these proteins during the lytic cycle. All transfections were carried out in OPTI-MEM medium. Cells were incubated at 37°C in 5% CO2 incubator and harvested after 48 h of transfection.
Harvested cells were re-suspended in sodium dodecyl sulfate (SDS) sample buffer at 106 cells/10 µl. Proteins were separated in 10% SDS-polyacrylamide gels and transferred to nitrocellulose membranes (Bio-Rad). The membrane was blotted with specific antibodies to cellular and viral proteins. The BGLF4 antibody was raised against amino acids 1 to 220 in rabbit. BGLF4(1–220) was expressed in Escherichia coli from a pET-22b vector and purified using nickel affinity chromatography. The EA-D (BMRF1) monoclonal antibody (R3.1) was kindly provided by G. Pearson [69]. Anti-ZEBRA and anti-BFRF3 are polyclonal rabbit antibodies and were described previously [54]. FLAG-tagged proteins were detected with anti-FLAG mouse monoclonal antibody (Sigma). β-actin was detected using a mouse monoclonal antibody (Sigma). The antigen-antibody complex was detected by autoradiography using 125I-protein A.
RNA was prepared from cells harvested 48 h after transfection using Qia-shredder and RNeasy Plus products from Qiagen. The concentration of RNA in each sample was determined by measuring the optical density at 260 nm. The level of viral transcripts was assessed in 100 ng of total RNA using iScript One-Step RT-PCR with SYBR green (Bio-Rad) in a total volume of 25 µl. The level of 18S rRNA was measured to normalize for the total amount of RNA. Each sample was analyzed in triplicates and the fold change in expression was calculated using the ΔΔCT formula. The sequences of the primers used are available upon request.
Three strand-specific sequencing libraries were produced from total RNA of 2089 cells transfected with empty vector, ZEBRA or ZEBRA plus siG4 as previously described [70]. The libraries were run on HiSeq 2000, generating approximately 86, 47 and 49 million reads, respectively. The generated reads were single-end and each read was 76 bp long. Adapter sequences, empty reads, and low-quality sequences were removed. Also, the first 16 and last 4 nucleotides in each read were trimmed using the FASTX toolkit (http://hannonlab.cshl.edu/fastx_toolkit/index.html) to remove low quality bases. Trimmed reads were mapped to the human reference genome (hg19) with a known transcriptome index (UCSC Known Gene annotation) using Tophat v2.0.8 [71]. Those reads that did not map to the human genome were later mapped to the EBV genome (GenBank accession number NC_007605.1) with a known transcriptome annotation [62]. We used the Expectation-Maximization (EM) algorithm in RSEM [57] with Bowtie 2 [72] to estimate gene expression levels. EBSeq within RSEM pipeline was used to identify differentially expressed genes [56].
Cells were harvested and re-suspended in 400 µl of lysis buffer containing 50 mM Tris-HCl [pH 8.1], 1% SDS, and 10 mM EDTA. The cell lysate was sonicated twice, 10 pulses each, and centrifuged for 10 min in a microfuge at full speed. The supernatant was diluted 10-fold in dilution buffer containing 16.7 mM Tris-HCl [pH 8.1], 0.01% SDS, 1.1% Triton X-100, 167 mM NaCl, and 1.2 mM EDTA. The samples were subjected to Proteinase-K digestion (Roche) followed by phenol-chloroform extraction to remove proteins. The DNA was purified using Qiagen PCR-purification kit. The total DNA content was determined by measuring the absorbance at 260 nm. The extent of viral genome amplification was quantitated using the IQ Sybr Green SuperMix kit (Bio-Rad). The sequences of the forward and reverse oriLyt primers are: 5′-TCCTCTTTTTGGGGTCTCTG-3′ and 5′-CCCTCCTCCTCTCGTTATCC-3′. The relative concentration of DNA was calculated based on a standard curve constructed from different concentrations of oriLyt. The level of viral DNA was normalized to control sample transfected with empty vector (CMV).
|
10.1371/journal.pgen.0030118 | Semi-Automatic Classification of Skeletal Morphology in Genetically Altered Mice Using Flat-Panel Volume Computed Tomography | Rapid progress in exploring the human and mouse genome has resulted in the generation of a multitude of mouse models to study gene functions in their biological context. However, effective screening methods that allow rapid noninvasive phenotyping of transgenic and knockout mice are still lacking. To identify murine models with bone alterations in vivo, we used flat-panel volume computed tomography (fpVCT) for high-resolution 3-D imaging and developed an algorithm with a computational intelligence system. First, we tested the accuracy and reliability of this approach by imaging discoidin domain receptor 2- (DDR2-) deficient mice, which display distinct skull abnormalities as shown by comparative landmark-based analysis. High-contrast fpVCT data of the skull with 200 μm isotropic resolution and 8-s scan time allowed segmentation and computation of significant shape features as well as visualization of morphological differences. The application of a trained artificial neuronal network to these datasets permitted a semi-automatic and highly accurate phenotype classification of DDR2-deficient compared to C57BL/6 wild-type mice. Even heterozygous DDR2 mice with only subtle phenotypic alterations were correctly determined by fpVCT imaging and identified as a new class. In addition, we successfully applied the algorithm to classify knockout mice lacking the DDR1 gene with no apparent skull deformities. Thus, this new method seems to be a potential tool to identify novel mouse phenotypes with skull changes from transgenic and knockout mice on the basis of random mutagenesis as well as from genetic models. However for this purpose, new neuronal networks have to be created and trained. In summary, the combination of fpVCT images with artificial neuronal networks provides a reliable, novel method for rapid, cost-effective, and noninvasive primary screening tool to detect skeletal phenotypes in mice.
| Transgenic mice are key models to shed new light on gene function during development and disease. Reliable high-throughput screening tools will facilitate the identification of transgenic mice with distinct phenotypes. In particular, alterations of the skull are difficult to detect by visual inspection due to its very complex morphological structure. Here, we used high-resolution flat-panel volume computed tomography (fpVCT), a novel semi-automatic screening tool to image skull-shape features of mice. The resulting 3-D datasets were combined with artificial neuronal networks and complex nonlinear computational models to permit rapid and automatic interpretation of the images. Compared to the extremely laborious landmark-based analysis, the manual work in our approach was reduced to the control of skull segmentation of images obtained by fpVCT. We applied our approach to genetically altered mice and various mouse strains and showed that it is an accurate and reliable method to successfully identify mice with skeletal phenotypes. We suggest the new methodology will also be a valuable tool for an in vivo, rapid, cost-effective, and reliable primary screen to identify skull abnormalities generated by random mouse mutagenesis experiments.
| Following the sequencing of the mouse and human genomes, attention has now focused on assessing gene function by gain-of-function mutations or targeted deletion of genes to address their function in vivo. However, many transgenic or knockout mice display a mild pathology without overt phenotypic alterations, which is clearly of utmost importance in understanding human diseases. This, in turn, has created an enormous demand for effective tools to assess the phenotype of mouse models so that gene expressions can be understood in a biological context [1]. However, the development of high-throughput mouse mutagenesis protocols requires a time- and cost-effective mode for primary testing of phenotypes.
In previous work, noninvasive imaging techniques such as computed tomography (CT) and magnetic resonance imaging have been applied to the anatomical phenotyping of transgenic mouse embryos [2–4] as well as in the brain and skulls of mouse models [5–7]. The measurement of 3-D coordinates as biological landmarks on the skull was used to analyze craniofacial phenotypes in mouse models for Down syndrome [8]. Similarly, metabolic profiling of cardiac tissue through high-resolution nuclear magnetic resonance spectroscopy in conjunction with multivariate statistics was used to classify mouse models of cardiac disease [9].
These imaging technologies for rapid visualization of large regions of anatomical structures have several important advantages over classical histology. The differential comparison of a large dataset of images using traditional radiological observation and a well-trained eye, especially between complex skeletal structures, is often inadequate. Therefore, automated analysis of images to detect skeletal phenotypes in mouse models will be highly advantageous.
Here, we have performed flat panel-based volume computed tomography (fpVCT) for rapid high-resolution imaging of bone structures in combination with artificial neuronal networks (ANNs) that are complex nonlinear computational models, designed much like the neuronal organization of a brain [10–15]. These networks are composed of a large number of highly interconnected processing elements, termed neurons, working in parallel order to model complicated biological relationships without making assumptions based on conventional statistical distributions. Neuronal networks learn by example so the details of how to recognize the phenotype of the skull are not needed. What is needed is a set of examples that are representative of all the variations of the phenotype [12,13]. Such neuronal networks have already been applied to characterize the variability of anthropological features of the human nasal skeleton [14] and to analyze and classify human craniofacial growth [15]. Here, fpVCT imaging enables the 3-D visualization of small anatomic details of bone structures. By selecting subvisual information from these fpVCT datasets of the skull, we applied ANNs to predict skeletal phenotypes in mouse models.
For visualization of the feature space structure, here, we analyzed the automatically generated skull-shape features with principle component analysis (PCA) and cluster analysis. PCA simplifies multidimensional datasets to lower dimensions and consequently transforms them into orthogonal linear to a new coordinate system such that the greatest variance comes to lie on the first coordinate, the second greatest variance on the next coordinate, and so on [16]. Cluster analysis is the partitioning of data into subsets, so that the data in each subset share some common traits with the use of some defined distance measurements [16,17].
The method was applied to knockout mice of a subfamily of tyrosine kinase receptors, discoidin domain receptors (DDRs), which are selectively expressed in a number of different cell types and organs; upon collagen activation DDRs regulate cellular adhesion and proliferation as well as extracellular matrix remodeling [18,19]. Lack of DDR2 resulted in reduced chondrocyte proliferation and shortening of long bones and the snout [20]. In contrast little is known about skeletal abnormalities of DDR1-deficient mice [21].
The purpose of this study is to present a rapid method for primary screening of skeletal phenotypes using fpVCT, allowing detailed nondestructive imaging of the skull in vivo. Using skull-shape features semi-automatically calculated from fpVCT datasets in combination with ANNs, we were able to successfully classify adult knockout mice with various bone malformations as well as identify mouse populations with subtle skeletal abnormalities with high accuracy.
All animals were maintained under pathogen-free conditions and housed in accordance with German animal welfare regulations. All animal protocols were approved by the administration of Lower Saxony, Germany. For this study homozygous and heterozygous DDR1- and DDR2-deficient mice on inbred C57BL/6 backgrounds, five DDR1/2 double knockout (DDR1−/−//DDR2−/−) mice, as well as C57BL/6 wild-type and severe combined immunodeficient (SCID) mice strain CB-17/ Ztm-scid, of different ages and sexes were used. To allow nearly complete ossification of the skull, all mice with the exception of the 14-d-old double knockout mice were older than 50 d. For this study we used 85 mice in total: 29 DDR1/2+/+, 25 DDR2−/−, ten DDR2+/−, nine DDR1−/−, five DDR1−/−//DDR2−/−, and seven SCID mice. For descriptive statistics, see Table 1.
The genotype of mice was verified by standard techniques. DNA was isolated from tail biopsies, and polymerase chain reaction (PCR) was performed as described previously [20,21]. These results were compared to the outcome obtained by fpVCT datasets in combination with an ANN.
Mice were anesthetized with vaporized isoflurane at 0.8 − 1% concentration throughout the imaging session and centered on the fpVCT gantry axis of rotation. The fpVCT prototype used in this study was developed and constructed by General Electric Global Research (http://www.ge.com/research). It consists of a modified circular CT gantry and two amorphous silicon flat-panel X-ray detectors, each of 20.5 × 20.5 cm2 with a matrix of 1,024 × 1,024 detector elements and a resolution of 200 μm. The fpVCT uses a step-and-shoot acquisition mode. Standard z coverage of one step is 4.21 cm. The mice were placed perpendicular to the z-axis of the system in order to scan the whole mouse with one rotation. All datasets were acquired with the same protocol: 1,000 views per rotation, 8-s rotation time, 360 used detector rows, 80 kVp, and 100 mA. A modified Feldkamp algorithm was used for image reconstruction resulting in isotropic high-resolution volume datasets (256 × 256 matrix, resolution about 200 μm).
To characterize the skull differences between the various mouse lines, we used landmark-based geometric analysis. Distances on the side-view-projection of the 3-D rendered mouse skull (D1, D2, and D3) as well as the curvature of the occipital region (C1) were manually measured (Figure 1). To calculate the curvature, the contour part between the inflection point on the superior region of the head as well as the inflection point after the sharp bend on the occipital region were manually isolated (Figure 1). For comparability of curvature values, all contours were stretched to the same length. Contours were represented in Freeman code (chain code). Therefore, each of the eight possible directions from one to the next contour point was encoded by a number from 0 to 7. We started with 0 on the right and used increasing numbers in a clockwise orientation. Since the determination of the local curvature from the changes of these direction codes results in rather noisy curves [22], we calculated the local curvature for a surrounding of six contour points, three forward and three backward of the actual point [23]. C1 represents the sum of these curvature values.
Segmentations of skulls were done using a simple threshold algorithm in the head region in order to segment all voxel with values above the threshold. For this purpose all parts of the segmentation that do not represent the skull were removed. All segmented voxel were referenced by the letter S.
Skull orientation and size were standardized by computing the centroid (Equation 1) and mass centroid axis, which are the eigenvectors of the inertial tensor (Equation 2), and rotating the skull in a such way that the x-axis points along the main mass centroid axis. To calculate the inertial tensor, the mass of a voxel is required. Bone consists of structures expressing different CT-numbers, and skull is surrounded by soft tissue that is characterized by low CT-numbers. Therefore, a distribution of these CT-numbers with a left skew gauss-like shape due to the partial volume effect was observed. The real surface of the bone is within a surface voxel. Because surface voxel must not contribute to features with the same magnitude as voxel representing solid bone, we determined the mass of a voxel by the weight function g(v) (Equation 3). Applying this equation voxel with a density equal to the mean density of bone had more influence in further calculations than others.
Mass centroid:
Inertial tensor:
Weight function:
Moments:
vr = radius component, vυ = polar angle, and vφ = azimuth angle of position vector v
Maximal dimensions along every axis were determined after the segmentation procedure and skulls were isotropically rescaled between −1 and 1 in all dimensions. These transformed skulls provide a basis to calculate size- and orientation-independent form features.
In a global depiction the components of the inertial tensor (Equation 3) are second-order moments. Here we used as features the moments (Equation 4) up to an order of four [24]. Since mouse skulls can be described as ellipsoid, these moments are not generated in Cartesian but in spherical coordinates. With the index of formula (Equation 4) we obtained the following 34 moments: m001, m010, m100, m002, m020, m200, m011, m101, m110, m003, m030, m300, m012, m102, m021, m120, m201, m210, m111, m004, m040, m400, m013, m103, m031, m130, m301, m310, m022, m202, m220, m112, m121, and m211.
If we consider Cartesian coordinates, it is possible to compare the second-order moments with a mass distribution away from a rotation axis. This means that a high second-order moment indicates that the mass of parts of the skull is distributed farther from the dedicated axis. Assuming that the mass density is nearly the same between the different skulls, a higher value second-order moment connotes the skull being more expanded perpendicular to that axis. Third-order moments impact the aberration from rotation symmetry perpendicular to the dedicated axis, and fourth-order moments impact the outliers. Since these predictions can be applied accordantly for spherical coordinates, form features were now encoded in these statistical parameters, in total 34 moments, independent of skull size and orientation. Sex of the animal (0 = female and 1 = male) and the normalized age, af (Equation 5) were added as features 35 and 36.
The feature af was calculated under the assumption that ossification of the skull has an exponential behavior and is up to 90% completed after 60 days [25]. So the codomain of (Equation 5) is the interval (0, 0.9] for mice that are between 1 and 60 d old.
a = age in days.
Finally, all features were transformed in an interval [−1, 1] to raise the stability of the training process of the artificial networks. For segmentation and feature calculation purposes, the algorithms were implemented using MSVC++ 6.0 and the additional libraries QT4.2 (Trolltech, http://www.trolltech.com) and VGL2.4 (Volume Graphics GmbH, http://www.volumegraphics.com). A C++ implementation of transformation and feature generation part is presented in Protocol S1.
To assess mouse phenotypes by skull images, several different ANNs were initially tested. Here, skull-form features were used to train multilayer perceptron network models developed with the Stuttgart Neuronal Network Simulator (SNNS) in context with a back-propagation-learning rule [26]. A multilayer perceptron consists of one input layer, some hidden layers, and an output layer. Such a network performs a transformation from an input pattern, which is the summary of all input values applied to the input neurons, into an output pattern and can be used for classification purposes by assigning classes to the output pattern. Hidden layers do not interact with the outside and are only necessary to perform the classification process.
For training and validation of a neuronal network, at least two different classified datasets have to be applied: one for the training process in order to impact the needed transformation between input pattern and the known output pattern in the network via the learning rule and one to test the network. The latter set consists of data not presented during the learning process. To test for accuracy, all network responses for the test dataset were compared manually with the known output pattern validated by genotyping of the mice. Here, the training dataset contains images from a minimum of three times more mice than the test dataset and the distribution of the pattern between these two datasets were chosen randomly to avoid bias.
For statistical analysis we performed PCA, cluster analysis, and Student's t-test. PCA and clustering analysis were done with PAST, free statistics software [27]. We used an agglomerative hierarchical cluster method with a Euclidean metric that groups step-by-step samples together with the lowest distance. The result of the algorithm is displayed as a dendrogram.
Furthermore, to determine whether the means of measured distances and curvatures between two mice strains differ significantly, we used the unpaired Student's t-test performed online with http://www.physics.csbsju.edu/stats/t-test.html. All data were expressed as mean ± SD, and statistical significance was established at a p-value less than 0.05.
Noninvasive imaging by fpVCT with 200 μm isotropy and 8-s scan times enabled the selective 3-D visualization of the skull of individual adult DDR1-and DDR2-deficient mice as well as wild-type littermates (Figure 1). Image comparison by visual inspection of C57BL/6 wild-type mice of different sex and age demonstrated minor interindividual differences within the complex skull morphology formed by the cranium and mandible. These minor variations in skull shapes are due to sexual dimorphisms and age dependent ossification. Individual skulls of mice from the control strain differ mainly in relative size and are characterized by the forming of the tympanic bulba, the infraorbital hiatus, and the grade of ossification of the frontal and occipital bone (Figure 1, left panel).
Representative examples of skull images from the various knockout mice in comparison to controls are shown in Figure 1, right panel. For comparative analyses of the various skulls we introduced landmarks located on the cranium that show marked differences mainly between DDR2-deficient mice and wild-type controls. Curvature C1, located at the posterior neurocranium, depicts the bend of parietal, interparietal, and occipital bone. As plotted in Figure 2A and 2B the local curvature is a nonconstant function, which differs between the knockout mice on different contour positions. We introduced C1 as the sum of the local curvature values representing the value of the cumulative curves on the maximal contour position. D1 represents the maximal distance between the external occipital protuberance and the incisor teeth. The distance between the center of the inner ear and the incisor teeth is labeled with D2, whereas D3 is the distance between the center of the inner ear and the external occipital protuberance (Figure 1, left panel). To take into account different skull sizes, we introduced new parameters such as D*2 and D*3, which are calculated from D2 and D3 by dividing the data with the total skull distance of D1. The nasal bone is represented by F1.
The neurocranium of DDR2-deficient mice displays a more round shape in comparison to controls illustrated by an increase in the magnitude of curvature C1 (Figure 1, right panel). This is clearly depicted by different cumulative curvatures of C1 especially in the late part of the curve of DDR2-deficient mice in comparison to their wild-type controls (Figure 2B and 2C). Skulls of DDR2−/− mice are jolted as shown by a reduced length of D1 (20.31 ± 0.58 mm versus 22.68 ± 0.45 mm for control). Furthermore a spherical skull shape is characteristic for DDR2-deficient mice as demonstrated by a reduced mean value for D*2 (0.77 ± 0.016 versus 0.784 ± 0.009 for control; p = 0.068; Figure 2D) and a significant increased value for D*3 (0.34 ± 0.009 versus 0.30 ± 0.005 for control; p < 0.0001; Figure 2E). Characteristic of DDR2-deficient mice is the nasal bone demonstrated by the landmark F1 that appears to be altered and of different shape in comparison to controls by visual inspection (Figure 1, right panel).
Alterations between wild-type and heterozygous DDR2-deficient mice are very subtle and can hardly be depicted by comparative morphological analysis using traditional radiological observation (Figure 1). This is shown by no significant differences of D*2 (0.79 ± 0.018 versus 0.78 ± 0.009 for control; p = 0.76; Figure 2D) and the curvature C1 (−94.98 ± 14.30 versus −84.96 ± 8.56 for control; p = 0.81, Figure 2C). Only the value for D*3 was significantly altered in DDR2 +/− mice (0.29 ± 0.004 versus 0.30 ± 0.005 for control; p = 0.023; Figure 2D).
DDR1-deficient mice are known to be smaller. Here, one characteristic feature of the skull of DDR1-deficient mice is an altered curvature progression compared to their wild-type controls (Figure 2A). However, the value for C1 (−90.24 ± 4.29 versus −84.96 ± 8.56 for control; p = 0.21) was not significantly different (Figure 2C). Distance measurements also revealed no significant differences to the values obtained in wild-type controls. D*2 was reduced in wild-type mice (0.80 ± 0.005 versus 0.78 ± 0.009 for control; p = 0.097) whereas D*3 was within the same range (0.30 ± 0.009 versus 0.30 ± 0.005 for control; p = 0.73; Figure 2D and 2E)
Comparative landmark based analysis of skulls of DDR2-deficient mice confirmed statistical significant bone deformations compared to controls, whereas analysis in DDR1-deficient mice determined no skull alterations.
PCA and cluster analysis were performed in the complete 34-D feature space to visualize and evaluate its structure and the existence of clusters corresponding to distinct mice types. Figure 3A displays the two-dimensional subspace for the two main components of the PCA for a subset of mice. For better visualization sample points of one class were interconnected with lines. No cluster could be detected, indicating a poor conditioned feature space. Cluster analysis applied to skull features of DDR2-deficient mice, and their wild-type controls do not discriminate between the two mouse populations (Figure 3B).
To evaluate the influence of sex- and age-dependent skull-shape differences, we performed PCA for an age- and a sex-matched subset, respectively, using datasets from DDR2-deficient mice and their wild-type controls (Figure 3C and 3D). Skull shapes of both male and female controls and female DDR2−/− mice displayed strong sex related differences shown by higher interclass distances in comparison to intraclass distances. Especially, control mice were grouped into two subsets of male and female mice. Only the cluster of the DDR2-deficient mice was very widespread, thus overlapping the clusters of female DDR2−/− and male control mice. Excluding male DDR2−/−, the feature space can be divided into two half planes for male and female mice. Interestingly, the male DDR2-deficient mice can be subdivided into two subgroups consisting of male and female “looking” mice (Figure 3C). In the PCA of the age dependency of the feature vectors including young mice, we depicted age related varieties, but were not able to determine any functional relationship between increase in age and the position of the feature vector (Figure 3D).
Since the skull-shape features used in this study are dependent on sex, we applied in further experiments both female and male mice to train the ANNs. In order to suppress the age dependent effects we only used mice older than 50 days with the exception of DDR1−/−//DDR2−/− mice, which were scanned ex-vivo when 14 d old.
ANN 1 shown in Figure 4A was developed to identify DDR2-deficient mice displaying a known skeletal phenotype between control littermates. Neuronal network 1 received one import layer with 36 neurons to import the skull features calculated from fpVCT datasets. It consists of two hidden layers with five and three neurons and of one output layer with two neurons, N1 and N2. All neurons were connected with short cuts and trained with the back-propagation momentum-learning rule.
A total of three different phenotypic assessments corresponding to DDR2−/−, DDR2+/−, and control mice were encoded in the output pattern as demonstrated in Figure 4A. We considered “high activation” as activation of the output neuron over 50% and “low activation” as activation under 50%. The following interconnections between output pattern and the corresponding genotype were defined: high activation of output neuron N1 with simultaneous low activation of N2 represents the DDR2−/− genotype; high activation of N2 associated with low activation of N1 represents the DDR2+/+ genotype; while low activation of both N1 and N2 illustrates the heterozygous DDR2 genotype. To train neuronal network 1, datasets of skull images from 39 adult mice, 18 DDR2−/−, 8 DDR2+/− mice, and 13 wild-type control mice were included. The accuracy of the trained neuronal network 1 to predict DDR2 genotypes in mice was tested with an additional dataset of skull images of eight mice. As demonstrated in Table 2, all tested mice were successfully classified by this method. A 30% activation of neuron N2 indicates that a heterozygous DDR2 mouse displays only a mild skeletal malformation. All results were confirmed by standard genotyping techniques applying PCR with genomic DNA.
Network 1 was also shown to successfully discriminate between various two mice types. In particular this simple network could also distinguish between DDR1−/− mice and their wild-type controls as well as between SCID mice and C57BL/6, the wild-type littermates of DDR-deficient mice (unpublished data).
We applied our methodology to a larger cohort of mice derived from more than one strain of knockout mice. We used DDR1-deficient as well as SCID mice in combination with homogenous and heterozygous DDR2-knockout mice and their controls to train ANN 2. This multilayer perceptron network consists of one input layer with 34 neurons, two hidden layers with five and three neurons each, and one output layer with five neurons (Figure 4B). We included 49 mice in the training process including 18 DDR2−/−, six DDR2+/−, five DDR1−/−, and 16 C57BL/6 wild-type mice as well as four SCID mice. The experimental dataset consisted of skull images of 16 mice randomly chosen from the mouse cohort. As shown in Figure 4B, each mouse population was identified by one output neuron, N1 − N5. The output neuron, which shows maximal activation, determines the phenotype that the network estimates for a given input pattern independent of the values of the other output neurons. DDR2-deficient mice were classified with high accuracy as demonstrated by the most highly activated output neuron, N2, being consistent in all three cases (Table 3). Since skull features of DDR1−/−, DDR2+/−, and DDR1/2+/+ as well as SCID mice display only minor phenotypic differences among each other, these genotypes were more difficult to classify in a network with trained datasets from skulls of all five populations. This is shown by low activation of the corresponding output neurons, N3 for DDR1−/− and N4 for SCID mice, in comparison to the high activation observed in N2 for DDR2−/− mice. However, the transgenic status of DDR-deficient mice could be reliably predicted with the exception of DDR2 heterozygous mice and one SCID mouse, which were incorrectly classified as control mice by showing maximal activation of N5 (Table 3). It appears that the major differences in skull formation between DDR2−/− and control mice determine the network, thus making it more difficult to classify mice with phenotypes possessing only minor bone abnormalities.
ANN 3 was applied to distinguish between C57BL/6 wild-type control, DDR2+/−, and SCID mice that were identified in network 2 as control mice (Figure 4C). Neuronal network 3 received 34 input neurons and one output layer with two neurons, N1 and N2. This network also consisted of two hidden layer with five and three neurons. Low activation of both N1 and N2 represented SCID mice, high activation of output neuron N1 with simultaneous low activation of N2 represented C57BL/6 wild-type mice, whereas high activation of both N1 and N2 represented DDR2+/− mice. The network was trained with 21 datasets, seven DDR2+/− mice, nine C57BL/6 wild-type mice, and five SCID mice. As shown in Table 4, all mice were successfully classified by skull-based features using ANN 3.
In an additional experiment we presented to the network 2 skull features of five DDR1/2 double knockout mice (DDR1−/−//DDR2−/−) without a new training process. Even when skull-shape features of these mice were not encoded in the network, the network response (Table 5) clearly depicted the existence of two skull shapes in all five mice tested, both related to DDR1- and DDR2-deficient mice. Therefore, the phenotype of the DDR double knockout mouse appears to be a superposition of the skull shapes from DDR1−/− and DDR2−/− mice.
This study presents a rapid cost-effective primary screening method for comparing and identifying mutant mice with abnormalities in skeletal development out of the increasing number of mouse models that are now being generated where genes have been “knocked out,” “knocked in,” or mutated.
We have developed an ANN-based intelligent system for image interpretation of large 3-D fpVCT datasets. ANNs are interconnected groups of artificial neurons that use a mathematical or computational model for information processing based on a connectionist approach to computation [12]. High-resolution 3-D imaging fpVCT allows a detailed visualization of the mouse skeleton with clear contours. The delineation of the anatomical details of bone structures in mice by fpVCT imaging has been previously described [9].
Comparative morphological analysis is typically difficult in the skull and extremely laborious due to its very complex skeletal structure [6]. In our study, large 3-D fpVCT datasets were simplified to skull-shape features on the basis of high-order moments of the whole skull, making phenotyping of mouse models a much simpler, cost-efficient, and semi-automatic process [23]. One of the benefits of this semi-automatic classification method is that the manual work is reduced to the control of the segmentation process of skulls for which anatomical knowledge is not necessary, resulting in a rapid and therefore high-throughput application.
Since each of the 34 skull-shape features are based on all skull voxels, and therefore the features include all possible skull alterations, the method was found to be very reproducible and reliable for mouse classification. Therefore, in order to analyze novel phenotypes with skull abnormalities the same feature space can consistently be applied. Although, other statistical methods might be appropriate to analyze these complex biological relationships, we have chosen neuronal networks for the classification process because of several advantages of this method. Neuronal networks are able to elucidate nonlinear problems and learn by example, so the details of the complex morphological skull structure on the basis of which the mice classification is made, are not needed [12,13]. Therefore, the use of neuronal networks is highly favorable in our study, characterized by poorly conditioned features space and the comparison of mice types, in which the distinct alterations between the different skull phenotypes are not known yet.
Our results indicate that this computational-intelligence scheme based on 34 skull features is capable of identifying genetically modified mice with skeletal abnormalities observed on the five different trained mouse populations.
PCA and cluster analysis of these skull features were not able to discriminate between the different knockout mice demonstrating a poorly conditioned feature space and thereby showing that a nonlinear classification is a prerequisite to a correct determination of genetically altered mice models [15,16]. Here, features implemented in the ANNs were independent of skull sizes and orientations, thus making this method suitable to exclude interindividual variations of skull growth within mice of one group. Although features are related to age, this method is able to successfully discriminate between mice types when using mice older than 50 d. This goes in line with the observation that skull shapes of mice change with growth, but remain nearly constant 15 d postnatal [28]. The high significance of these features is also shown by the fact that even the smallest differences can be automatically detected, such as alterations in skull shapes related to sex. The algorithm was first tested on a cohort of DDR2-deficient mice with a known skeletal phenotype displaying shortage of long bones and a shorter snout [20]. In this study, this phenotyping method enables us to reliably detect DDR2-deficient mice within a cohort consisting of homozygous and heterozygous DDR2 mutants. Even heterozygous DDR2 mice with a subtle phenotype were correctly determined and identified as being different from their wild-type control by this method based on fpVCT imaging.
So far, no obvious skull abnormalities have been observed in DDR1-deficient mice. A reduced bone calcification has only been described in the fibula [21]. However, the imaging technique in combination with a neuronal artificial network trained only with skull-shape features generated from DDR1-deficient mice, and control mice were successful in discriminating clearly between DDR1-genotypes. With this method DDR1-deficient mice were identified as mice that show, in contrast to C57BL/6 wild-type animals, differences in skull formation. Landmark-based analysis of three distances and the occipital curvatures confirmed the presence of skull abnormalities of DDR2 knockout mice, thereby clearly defining a skeletal phenotype for these mutants. In contrast, DDR1-deficient mice with a subtle phenotype were not significantly altered in these features.
The screening tool based on skull-shape features is successful in discriminating between mouse strains displaying no overt differences in skull formation, for example distinguishing SCID mice from C57BL/6 wild-type mice. Even five double knockout mice for DDR1 and DDR2 not used in the training process, were identified by our semi-automatic classification method as a superposition of the classes related to the single features of DDR1−/− and DDR2−/− mice. This suggests that the algorithm discriminates not only between trained mice, but identifies different skull-shape traits.
The imaging technique, in combination with a more complex neuronal network, was also valuable to reliably discriminate DDR2-deficient mice between five different mouse populations, including SCID mice and DDR1-deficient mice. The challenge of creating a network for all presented mouse genotype-related phenotypes together is to balance the combined feature space of all classes. Therefore, the application of this ANN trained with features out of datasets from DDR2-deficient mice with a marked skeletal defect did not allow discrimination between the heterozygous DDR2-deficient mice, SCID mice, or their controls, which all show similarity in skull bones. However, thereafter they were successfully classified by applying the more specialized neuronal network for the three subtle mouse phenotypes.
In conclusion, we have introduced a novel semi-automatic screening method for skeletal phenotyping by applying neuronal networks in combination with fpVCT, so far limited to five different mice models. This methodology seems to be a powerful tool for the rapid detection of living mice with skull abnormalities. In the future, this technique is expected to be a standardized, cost-effective, primary screen to identify mice with skeletal differences out of a wide spectrum of genetically altered mice on the basis of random mutagenesis as well as transgenic and knockout mice. For successful identification of novel mutant mice with bone abnormalities, skull-shape features have to be calculated to create and train a novel neuronal network. However, because the introduced features are calculated automatically and include information of every skull voxel, they should easily be implemented to new skull shapes. Even a minor training error will indicate the existence of alteration in skull shapes. The degree of reliance to predict a skull phenotype is directly related to the correspondence of the neuronal network response for a second set of mice to their genotype. Therefore, skull alterations of genetic modified mice in comparison to their control littermates are depicted by correct classification of the mutant mice in a separate class.
Though successful in predicting the corresponding phenotype of various mouse populations in the primary screen, this methodology is not suitable for defining the exact bone deformation underlying the gene defect. The challenge of the next step will be to characterize the anatomical phenotypes in mice in more detail with stereological parameters by nondestructive visualization of complex skull structures by fpVCT imaging. Comparative analysis of multiple datasets of bone images will then allow to us to identify differences in corresponding anatomical sites between control and mutant groups, which has been demonstrated recently for MR imaging [6].
Since the overall shape of the skull depends upon coordinated development of separate bony, dental, and cartilaginous elements and functioning of soft tissue components, the identification of knockout mice with skull abnormalities and their characterization will help to further understand the role of major genes that are involved in the cascade of developmental processes necessary for the proper development of a functioning skull. |
10.1371/journal.ppat.1002435 | Targeting of Mycobacterium tuberculosis Heparin-Binding Hemagglutinin to Mitochondria in Macrophages | Mycobacterium tuberculosis heparin-binding hemagglutinin (HBHA), a virulence factor involved in extrapulmonary dissemination and a strong diagnostic antigen against tuberculosis, is both surface-associated and secreted. The role of HBHA in macrophages during M. tuberculosis infection, however, is less well known. Here, we show that recombinant HBHA produced by Mycobacterium smegmatis effectively induces apoptosis in murine macrophages. DNA fragmentation, nuclear condensation, caspase activation, and poly (ADP-ribose) polymerase cleavage were observed in apoptotic macrophages treated with HBHA. Enhanced reactive oxygen species (ROS) production and Bax activation were essential for HBHA-induced apoptosis, as evidenced by a restoration of the viability of macrophages pretreated with N-acetylcysteine, a potent ROS scavenger, or transfected with Bax siRNA. HBHA is targeted to the mitochondrial compartment of HBHA-treated and M. tuberculosis-infected macrophages. Dissipation of the mitochondrial transmembrane potential (ΔΨm) and depletion of cytochrome c also occurred in both macrophages and isolated mitochondria treated with HBHA. Disruption of HBHA gene led to the restoration of ΔΨm impairment in infected macrophages, resulting in reduced apoptosis. Taken together, our data suggest that HBHA may act as a strong pathogenic factor to cause apoptosis of professional phagocytes infected with M. tuberculosis.
| Cell death is a common outcome during infection with a number of pathogenic microorganisms. Therefore, defining the factors responsible for killing of host cells is important to uncovering mechanisms of pathogenesis. World-wide, two billon people are latently infected with Mycobacterium tuberculosis, which is still killing 2–3 million people each year. Heparin-binding hemagglutinin (HBHA) protein of M. tuberculosis is known to interact specifically with non-phagocytic cells and to be involved in dissemination from lungs to other tissues. Nevertheless, the role of HBHA in phagocytic cells such as macrophages, which are the first cells of the immune system to encounter inhaled pathogens, has been unknown. In the present study, we suggest HBHA as a critical bacterial protein for macrophage cell death. After M. tuberculosis infection or HBHA treatment of macrophages, HBHA targeted to mitochondria and then caused mitochondrial damage and oxidative stress, which eventually lead to apoptosis. A mutant of M. tuberculosis lacking HBHA induced less apoptosis with moderated mitochondrial damage. These experiments provide a candidate virulence factor which may be a novel target for tuberculosis treatment.
| Tuberculosis remains a serious global problem, although many researchers have made a persistent effort for several decades. Mycobacterium tuberculosis, a major causative agent of pulmonary tuberculosis, is responsible for 1.8 million deaths per year worldwide [1]. Innate immune system plays a critical role in antimicrobial host response during the early stage of M. tuberculosis infection. Alveolar macrophages mediate innate immunity by phagocytosing pathogens and are the chief defense against M. tuberculosis, which can survive and replicate within phagocytes [2]. The course of tuberculosis rests on the outcome of the interaction between the bacterium and host macrophage. Therefore, a better understanding of these complex interactions is critical to controlling mycobacterial infection.
Many bacterial and viral pathogens utilize various strategies to manipulate host machinery to serve their own needs. Apoptotic cell death has been regarded as an innate cellular response to limit the multiplication of intracellular pathogens [3], although the precise mechanism of the direct antimicrobial action in infected macrophages undergoing apoptosis is unclear. Generally, infectious intracellular pathogens tend to prevent host cell apoptosis during an early stage of infection. However, they may also induce host cell apoptosis with a specific aim to subvert the host attack, such as immune and inflammatory response, at later stages [4], [5].
A number of reports have indicated that M. tuberculosis does indeed inhibit host cell apoptosis, while at the same time it induces pro-apoptotic signals. Recent studies showed that only virulent mycobacterial species can inhibit apoptosis induction in primary human alveolar macrophages [6], THP-1 [7], [8], and J774 macrophage cell lines [9]. Virulent M. tuberculosis reportedly induced the apoptotic death of host cells. For example, enhanced apoptotic response was detected in alveolar macrophages recovered from patients with pulmonary tuberculosis [10], [11]. Extensive apoptosis was also observed in caseating granulomas from lung tissue samples obtained from patients with tuberculosis [12], [13]. Several apoptosis-inducing factors of M. tuberculosis, such as 19-kDa glycolipoprotein (Rv3763) [14], PE_PGRS33 (Rv1818c) [15], ESAT6 (Rv3875) [16], and 38-kDa lipoprotein (Rv0934) [17] are reported.
Heparin-binding hemagglutinin adhesin (HBHA) is a 28-kDa multifunctional protein found on the surface and culture filtrates of mycobacteria. It has hemagglutination activity and binds to sulfated glycoconjugates such as heparin and dextran sulfate [18]. HBHA interacts specifically with non-phagocytic cells and is essential for the infection of lung epithelial cells and extrapulmonary dissemination of M. tuberculosis [18], [19]. Protective immunity induced by HBHA is observed in M. tuberculosis-infected mouse models, indicating that HBHA is a protective antigen [20]. Recent studies suggest that HBHA is a useful diagnostic marker for tuberculosis [21]. We also identified and characterized HBHA as a serologically active mycobacterial antigen in a previous study, whereby HBHA binds strongly to the immunoglobulin M of patients with tuberculosis [22]. Although HBHA function in mycobacterial pathogenesis has been extensively studied, the role of HBHA on professional phagocytes, such as macrophages, is still poorly understood.
The aim of the present study was to characterize the biological effects of M. tuberculosis HBHA on macrophages. We found that HBHA induced apoptosis in murine macrophages and investigated its underlying mechanism. Here, we show that HBHA treatment caused a loss of mitochondrial transmembrane potential (ΔΨm) and the release of cytochrome c from purified mitochondria in vitro, as well as mitochondria of intact cells, and HBHA was efficiently targeted to mitochondria of macrophages.
We first sought to determine whether HBHA could induce macrophage apoptosis. Apoptosis was assessed by quantifying DNA fragmentation, which is considered a hallmark of apoptosis, in the cytoplasmic fractions of dying cells using a commercially available ELISA kit. The incubation of RAW 264.7 cells with HBHA resulted in a significant increase in the release of oligonucleosomal fragments into the cytoplasm in both dose- and time-dependent manners as compared to compared to control cells (Figure 1A and 1B). Cell death was significantly greater in cells treated with HBHA as compared to buffer-treated control cells. As lactate dehydrogenase was not detected in the cell culture supernatant during HBHA treatment, the possibility that HBHA-induced death is necrosis was excluded (Figure S1). We used native antigen 85 complex (Ag85) as an unrelated control protein. The Ag85 of M. tuberculosis is the major secreted protein and fibronectin-binding protein, and shows strong immunoreactivity [23], [24]. Similar results were observed in bone marrow-derived macrophages (BMDMs); like PBS-treated BMDMs DNA fragmentation was not detected in Ag85-treated cells, whereas dramatic DNA fragmentation was observed in HBHA-treated cells (Figure 1C). HBHA-induced apoptosis was further confirmed by examining the nuclear morphology of dying cells using a fluorescent DNA-binding agent, 4′-6-diamidino-2-phenylindole (DAPI). As shown in Figure 1D, control cells treated with buffer had intact nuclei. In contrast, within 48 h of HBHA treatment, RAW 264.7 cells clearly exhibited condensed or fragmented nuclei indicative of apoptotic cell death. We further analyzed the caspase dependency of HBHA-induced apoptosis. Western blot analysis showed that the cleavage of caspase-3, caspase-9, and poly(ADP-ribose) polymerase (PARP) was evident in cells incubated with HBHA for 48 h (Figure 1E). Inhibition of caspases by a pan-caspase inhibitor, zVAD-fmk, attenuated the HBHA-induced DNA fragmentation, indicating that HBHA induces caspase-dependent apoptosis (Figure 1F). These results suggest that macrophages treated with HBHA undergo caspase-dependent apoptosis.
The mitochondrion acts as a central executioner in response to apoptotic stimuli, allowing signals from various inputs to converge [25]. We investigated whether HBHA treatment affected the structural and biochemical integrity of mitochondria. Mitochondrial damage was assessed by examining mitochondrial ΔΨm, which was determined by staining cells with 3,3′-Dihexyloxacarbocyanine (DiOC6), a dye that incorporates into mitochondria with intact membrane potential [26], for flow cytometric analysis. As shown in Figure 2A, a significant loss of ΔΨm was observed in RAW 264.7 cells incubated with HBHA as indicated by a decrease in DiOC6 intensity. Analysis of the time course for examination of ΔΨm onset showed a noticeable dissipation of ΔΨm after 18 h of HBHA treatment, which further decreased with time. A similar result was obtained in BMDMs incubated with HBHA (Figure 2B). These results suggest that mitochondrial damage appears as a subsequent event in the intracellular action of HBHA.
Apoptosis at the mitochondrial level involves the oligomerization of the pro-apoptotic protein Bax [27], leading to permeabilization of the outer mitochondrial membrane (MOMP) and release of cytochrome c [28]. We performed immunocytochemistry to detect Bax translocation and cytochrome c release. An antibody recognizing the Bax N-terminus, which is exposed by the activation of Bax and its insertion into the mitochondrial membrane, was used. Figure 3A shows the translocation of Bax distributed evenly in the cytoplasm to the mitochondria in macrophages as evident by the colocalization of Bax with Mitotracker, a potential-sensitive dye specific for mitochondria. In PBS-treated cells, cytochrome c showed a punctate pattern that colocalizes with Mitotracker, whereas the faint signal for cytochrome c and the decreased colocalization with Mitotracker were detected in HBHA-treated cells, indicating cytochrome c release. These results were confirmed by performing subcellular fractionation and Western blot analysis (Figure 3B). HBHA caused a decrease in cytochrome c immunoreactivity in the mitochondrial fraction with a concomitant increase in the cytosolic fraction and vice versa for Bax immunoreactivity. Collectively, these findings suggest that the apoptotic effect of HBHA on macrophages is associated with cytochrome c release and Bax translocation. To determine whether Bax activation is necessary for HBHA-induced apoptosis, we knocked down the level of Bax by transfecting RAW 264.7 cells with Bax siRNA. The Bax protein level was significantly reduced in cells transfected with Bax siRNA; Bax protein in control siRNA-transfected cells was unchanged (Figure 3C). We then determined the effect of knockdown Bax on HBHA-induced apoptosis in RAW 264.7 cells. As shown in Figure 3D and 3E, HBHA-induced increase in DNA fragmentation was blocked and ΔΨm loss was restored by Bax knockdown, suggesting that Bax activation is required for HBHA-induced macrophage apoptosis.
Enhanced reactive oxygen species (ROS) production, characteristic of early apoptotic events, can be both a cause and a consequence of changes in ΔΨm [26], [29]. To examine the involvement of ROS generation on HBHA effects in macrophages, the oxidation of DCF was monitored by flow cytometry and fluorescent microscopy (Figure 4A). Compared to PBS or Ag85, HBHA significantly induced the increase of intracellular hydroperoxide in macrophages. To determine the requirement of ROS increase in HBHA-induced apoptosis, the effect of HBHA alone or in combination with N-acetylcysteine (NAC), a general ROS scavenger, on DNA fragmentation was assessed. NAC pretreatment effectively inhibited HBHA-induced DNA fragmentation (Figure 4B) as well as ROS production, suggesting that ROS increase is essential for the apoptotic response caused by HBHA.
Studies have suggested that some infectious intracellular pathogens regulate apoptosis of their host cells by targeting proteins to mitochondrial membranes that either induce or inhibit MMP [30]. We addressed the question of where HBHA is localized in mitochondria of HBHA-treated cells. Therefore, the possibility that HBHA interacts with the mitochondrial compartment was examined. Confocal microscopic analysis revealed the presence of HBHA in the mitochondria of HBHA-treated cells, as evidenced by a significant overlap between HBHA and Mitotracker (Figure 5A). Subcellular fractionation and Western blot analysis consistently showed that large amounts of HBHA were detected in the mitochondrial fraction, but not in the cytosolic fraction, where little HBHA was observed (Figure 5B). In contrast, the minimum of Ag85 were detected in cytoplasmic fraction of macrophage treated with Ag85, suggesting that it is not able to pass through plasma membrane. Furthermore, to determine whether HBHA was imported into mitochondria, we isolated mitochondria from cells treated with HBHA. The purified mitochondria were subsequently digested with proteinase K. As shown in Figure 5C, HBHA disappeared in mitochondria digested with proteinase K, indicating that HBHA adheres to the outer membrane of the mitochondria.
We next determined whether HBHA induced cytochrome c release from isolated mitochondria. As shown in Figure 6A, isolated mitochondria from RAW 264.7 cells released cytochrome c after HBHA treatment, whereas the buffer control or Ag85 did not stimulate this release in a cell-free assay. We also examined the effect of HBHA on the collapse of membrane potential in purified mitochondria. For this, mitochondria incubated with HBHA were stained with DiOC6, and the fluorescence intensity was monitored by flow cytometry (Figure 6B). A significant shift to a lower intensity was observed in mitochondria treated with HBHA as compared to buffer control or Ag85, indicating the decrease in ΔΨm. These data provide evidence that similar to the event that occurs in macrophages, HBHA can solely induce mitochondrial damage in a cell-free system, indicating that Bax translocation to mitochondria is not essential for of ΔΨm loss and cytochrome c release.
HBHA is a secreted protein in M. tuberculosis as well as a surface-associated protein [18]. To examine whether HBHA is also transported to mitochondria during M. tuberculosis infection, BMDMs were infected with H37Rv wild type and mutant disrupted in hbhA. Immunofluorescence microscopy of infected cells revealed that a part of HBHA colocalized with mitochondria (Figure 7A). Purified mitochondrial fraction of these cells contained a considerable amount of HBHA protein, although a large portion of HBHA were observed in cytosolic fraction (Figure 7B). These findings demonstrate that HBHA is efficiently transported to mitochondria of infected macrophages. To analyze the effects of HBHA on macrophages in the context of the bacterium as a whole, we compared the relative ability of M. tuberculosis H37Rv wild type and mutant disrupted in hbhA to induce apoptosis and ΔΨm collapse in macrophages. A reduced DNA fragmentation and an increased intact mitochondria were observed in BMDMs infected with mutant strain compared to its parent (Figure 7C and 7D), which was noticeable when macrophages were infected at MOIs of 5 and 10 but not at an MOI of 25 (Figure S2A). On the other hand, there was no significant difference in LDH release between cells infected with two strains (Figure S2B). Similarly, more significant DNA fragmentation and ΔΨm loss were detected in cells infected with M. smegmatis ectopically expressing HBHA compared to cells infected with the M. smegmatis control.
HBHA is involved in the interaction of mycobacteria with alveolar epithelial cells [19]. To determine whether these cells exposed to HBHA undergo apoptosis, human type II A549 pneumocytes were treated with purified HBHA for 48 h. As shown in Figure 8A and 8B, neither DNA fragmentation nor ΔΨm collapse was observed in HBHA-treated A549 cells. Immunofluorescent microscopy showed that a very faint green signal was detected in A549 cells incubated with HBHA, indicating that HBHA enters A549 cells much less efficiently (Figure 8C, upper panels). To confirm this issue, A549 cells were infected with M. tuberculosis wild type and hbhA deficient strains. Like experiments conducted in macrophages, M. tuberculosis infection led to severe ΔΨm dissipation, accompanied by the partial presence of HBHA in mitochondrial compartments (Figure 8B and 8C, lower panels). In contrast, a decrease in the percentage of cells displaying loss of ΔΨm was observed in A549 cells infected with the hbhA deficient strain (Figure 8B). These data suggest that cell entry and targeting to mitochondria of HBHA are essential for ΔΨm loss and apoptotic response.
Programmed cell death is emerging as a major effect of bacterial pathogenesis. Numerous studies have shown that M. tuberculosis infection can increase the rate of macrophage apoptosis [31], [32]. Pro-apoptotic activities of a growing number of mycobacterial components have recently been described [14]–[17]. Nevertheless, data regarding the identities of the mycobacterial molecules involved and the underlying apoptotic mechanism are still scarce. We showed here that intracellular HBHA is targeted to mitochondria in murine macrophages, which leads to ΔΨm dissipation and eventual apoptosis. Although the possibility that HBHA may interact with cytosolic molecules or other cell compartments cannot be ruled out completely, these connections clearly appear to be insignificant. To our knowledge, the present study is the first description of a mycobacteria-encoded protein stimulating apoptotic cell death via a mitochondria-dependent pathway in macrophages.
M. tuberculosis HBHA is a protein that is both surface-associated and secreted. HBHA is involved in the binding of M. tuberculosis to type II pneumocytes, but not to professional phagocytes such as macrophages, and is required for the dissemination of tubercle bacilli from the lungs to other tissues [19]. In this respect, its impact on macrophages has received relatively little attention. However, HBHA was recently demonstrated to have the capacity to bind to complement component C3, and recombinant HBHA was found to mediate the attachment of latex beads to murine macrophage-like cells in both C3-dependent and -independent manners [33]. M. tuberculosis can bind to the complement receptors and is subsequently introduced into the phagocytic cell [34]. These results raise the possibility of the interaction between HBHA and macrophages during mycobacterial infection.
Mitochondria are central organelles in which a variety of key events in apoptosis occur, including the release of cytochrome c, changes in electron transport, ΔΨm collapse, altered cellular oxidation–reduction, and participation of pro- and anti-apoptotic Bcl-2 family proteins [26]. Presently, mitochondria are regarded as the targets for the manipulation of many bacterial and viral pathogens determining the fate of infected host cells [35]. Moreover, mitochondrial damage has been suggested to play a critical role in the outcome of macrophage infection with M. tuberculosis [36]. These findings offer the potential of mycobacterial components for the regulation of programmed cell death at the mitochondrial level.
MMP is regulated by endogenous molecules, including Bcl-2 family members such as Bax [37]. The Bax present in the cytosol under normal conditions fosters the loss of ΔΨm and releases cytochrome c and apoptosis-inducing factor (AIF) from mitochondria after its introduction into the mitochondrial compartment [26]. Indeed, mitochondrial translocation of Bax was observed in macrophages treated with HBHA, and the interaction of HBHA with mitochondria resulted in cytochrome c release in murine macrophages. However, Bax translocation may not be essential for mitochondrial dysfunction by HBHA, as evidenced by a mitochondrial cell-free assay in which HBHA caused ΔΨm loss and cytochrome c release in vitro. Not surprisingly, we observed the activation of caspases 3 and 9 and subsequent cleavage of PARP after incubation of macrophages with HBHA. In contrast, we found no evidence of cytosolic or nuclear translocation of AIF induced by HBHA (data not shown), indicating that it is not involved in HBHA-induced cell death. ROS generation with ΔΨm modulation and caspase-9 activation is known to be a major component of the mitochondrial pathway of apoptosis [38]. ROS are predominantly produced in the mitochondria and lead to the modulation of ΔΨm, which finally results in apoptosis [39]. Our results indicate that HBHA induces macrophage apoptosis through ROS generation and ΔΨm collapse, suggesting that these play an essential role in HBHA-induced apoptosis.
Our results indicate that cellular entry is essential for mitochondria-mediated apoptotic effect of HBHA, although the mechanism by which HBHA internalized by host cells remains unresolved. In A549 cells infected with M. tuberculosis but not cells incubated with purified HBHA, the severe ΔΨm collapse and the presence of intracellular HBHA in mitochondrial compartment were observed. There was a significant increase in the percentage of cells with intact ΔΨm, when A549 cells were infected with the mutant strain lacking HBHA gene. We cannot rule out that these results might come from decreased number of mycobacteria in cells, because invasion of A549 cells, but not macrophages, by HBHA-deficient strain compared with parental strain was reduced [19]. Moreover, HBHA induced ΔΨm loss and cytochrome c release in purified mitochondria from not only RAW 264.7 cells but also mouse liver (data not shown). Thus, no impact on viability of epithelial cells treated with HBHA might be due to the absence of intracellular this protein.
What host molecules physically and functionally interact with intracellular HBHA and how do they then induce mitochondrial dysfunction? In the present study, proteinase K digestion in vitro showed that intracellularly inserted HBHA is attached to the mitochondrial surface but is not imported into mitochondria, indicating that HBHA probably interacts with integral outer membrane molecules. Several mitochondria-targeted proteins encoded by pathogens interact with voltage-dependent anion channel (VDAC). The porin B from N. meningitidis is a VDAC-targeted protein [40]. Hepatitis B virus X protein also co-localizes to mitochondria where it interacts with a particular VDAC isoform, HVDAC3 [41]. Anti-apoptotic members of the Bcl-2 family, such as Bcl-2 and Bcl-xL, are located in mitochondrial membranes where they inhibit cytochrome c release from mitochondria and thereby prevent downstream caspase activation. Pro-apoptotic members of the Bcl-2 family, such as Bax, can translocate into mitochondria and induce MMP [42]. These Bcl-2-like proteins can be prominent targets of bacterial proteins [30], [43]. Recombinant HBHA used in the present study was a His-tagged fusion protein. To determine the interaction between HBHA and VDAC or the Bcl-2 family proteins, HBHA and interacting molecules were purified by Ni-NTA affinity chromatography, followed by immunoblotting against them. However, HBHA showed no direct interaction with VDAC or Bcl-2 family members (data not shown). In addition, the possibility of HBHA nonspecific binding to mitochondria cannot be excluded. The C-terminal region of HBHA contains several cationic lysine-rich repeats where methylation can occur [44]. This region may work like natural antibiotic peptides which form cationic residues on one end and interact with anionic molecules such as phospholipids to disrupt negatively charged membranes and result in apoptosis [45].
Virulent M. tuberculosis induces necrosis of the infected macrophages by inhibiting the repair process of plasma membrane; this leads to cellular lysis and reinforces the spreading to the adjacent infection sites [46]–[48]. Recent reports suggested that high intracellular burden of virulent M. tuberculosis induces host cell death via a new caspase-independent apoptotic pathway involved in the bacterial escape and extracellular replication [49], [50]. Because gain of function mutation in HBHA enhanced the apoptogenic potency of M. smegmatis (Figure 7C), it is plausible that HBHA may be the factor that allows M. tuberculosis to escape from the infected macrophages at high intracellular burden. However, similar levels of apoptosis were observed between macrophages infected with M. tuberculosis H37Rv wild type and mutant disrupted in hbhA at an MOI of 25 but not low MOIs (Figure S2A). Further, HBHA deficiency had no influence on the macrophage necrosis caused by M. tuberculosis at both low and high MOIs (Figure S2B), indicating no involvement of HBHA in bacterial escape from the macrophages at an early stage of infection.
Studies on the comparison of virulent and attenuated mycobacterial strains have demonstrated that the latter has much stronger apoptotic activity in macrophages. This concept is supported by the identification of genes that inhibit apoptosis of host cells [51]–[53]. In this sense, our claims that HBHA targets to the mitochondria of host cells in the induction of apoptosis may be confusing. However, there is cumulative evidence suggesting that virulent M. tuberculosis induces host cell apoptosis. Furthermore, the transcriptional profiling of cells infected with virulent M. tuberculosis showed increases in the expression of both pro- and anti-apoptotic genes [54], [55]. Collectively, it is highly likely that M. tuberculosis infection results in pro- and anti-apoptotic response of host cells. The final outcome may depend on the nature and activation status of the host cell. Although the pro-apoptotic response is inarguably beneficial to the host, it may provide a favorable circumstance for the induction of necrotic cell death and subsequent bacterial escape to the adjacent cells, which may provide a clue for HBHA function during M. tuberculosis infection [46], [49], [50].
Taken together, the present study suggests the possibility that the M. tuberculosis HBHA may be an apoptosis-inducing factor of mycobacteria, although the molecular mechanism by which HBHA causes loss of ΔΨm remains unknown. Future work should focus on the exploration of host targets of HBHA and the mechanism by which HBHA modulates ΔΨm and cytochrome c release in detail, as well as identification of the HBHA domain essential for its activity in mitochondrial dysfunction.
All animal procedures were approved by the Institutional Animal Care and Use Committees of Chungnam National University (Permit Number: 2010-3-9). All animal experiments were performed in accordance with Korean Food and Drug Administration (KFDA) guidelines.
Antibodies against caspase-3, caspase-9, and VDAC were purchased from Cell Signaling Technology Inc (Beverly, MA). The anti-PARP and anti-β-actin, anti-Bax, and anti-Tom40 antibodies were obtained from Santa Cruz Biotechnology (Santa Cruz, CA). Antibodies against cytochrome c (for immunofluorescence, clone 6H2.B4; for Western blot analysis, clone 7H8.2C12) were acquired from BD Pharmingen (San Diego, CA), and the anti-cytochrome oxidase subunit IV (COX IV) antibody was purchased from Abcam (Cambridge, UK). Dichlorodihydrofluorescein diacetate (H2DCFDA), DAPI, and DiOC6 were obtained from Molecular Probes (Eugene, OR) and zVAD-fmk and NAC were purchased from Calbiochem (San Diego, CA).
Mycobacterium smegmatis strains, recombinant HBHA protein from M. smegmatis, and antiserum to HBHA were produced and prepared as described previously [22]. Ag85 was purified from the culture filtrate protein of M. tuberculosis H37Rv (ATCC 27294), as previously described by Lim et al [24]. Parental and mutant (hbhA deletion) Mycobacterium tuberculosis 103 were kindly provided by Dr. Camille Locht (Institut Pasteur de Lille, Lille, France) [19]. HBHA proteins were used in experiments after lipopolysaccharide (LPS) inactivation with polymyxin B (Invivogen, San Diego, CA), a known pharmacological antagonist of LPS.
RAW 264.7 murine macrophage cell line and A549 human alveolar epithelial cell line were cultured in Dulbecco's modified Eagle's medium (DMEM; Lonza, Walkersville, MD) supplemented with 10% fetal bovine serum (FBS; Hyclone, Logan, UT), 1% HEPES, and 1% l-glutamine at 37°C with 5% CO2. BMDMs were obtained from 6–8-week-old female C57BL/6 mice. Briefly, bone marrow cells from the femur and tibia were cultured in DMEM that contained 2 mM l-glutamine, 100 U/mL penicillin, 100 µg/mL streptomycin, 10% FBS, and 25 ng/mL recombinant mouse M-CSF (R&D system, Minneapolis, MN) at 37°C with 5% CO2. After 4 days, non-adherent cells were removed and differentiated macrophages were incubated in antibiotic-free DMEM until use.
One day before transfection, RAW 264.7 cells were plated and grown at 37°C to 70% confluency in complete medium without antibiotics in 6 well plates. One micrograms of a Bax siRNA (Bioneer, Deajeon, Korea, sense: CCGGCGAAUUGGAGAUGAA; anti-sense: UUCAUCUCCAAUUGGCCGG) or a noncomplementary siRNA were transiently transfected into RAW 264.7 using Lipofectamine 2000 transfection reagent (Invitrogen, Carlsbad, CA, USA), according to the manufacturer's instructions.
Cells were seeded in 96-well flat-bottom culture plates. After incubation with recombinant HBHA proteins, cells were collected, washed with PBS, and processed for quantification of cytoplasmic histone-associated DNA fragments formed during apoptosis using an enzyme-linked immunosorbent assay (Cell Death Detection ELISA PLUS; Roche Diagnostic) according to the manufacturer's instructions.
The release of LDH from RAW 264.7 cells incubated with recombinant HBHA or from BMDMs infected with M. tuberculosis was measured using a Cytotoxicity Detection Kit plus (Roche, Indianapolis, IN) according to the manufacturer's protocol. Relative cytotoxicity was calculated using the following equation: Cytotoxicity (%) = % of LDH released from the infected cells/maximum LDH released.
ΔΨm was assessed by measuring retention of the lipophilic cationic dye DiOC6 in mitochondria. Cells were harvested and incubated in a DiOC6 solution (10 nM in fresh medium) for 20 min at 37°C in the dark. The cells were then washed and resuspended in PBS. Immediately after PBS washing, ΔΨm was measured by sorting the cells using FACSCanto (BD Biosciences). Dead cells were excluded by forward and side-scatter gating. Data were acquired by analyzing an average population of 10 000 cells using CELLQuest software (BD Biosciences).
Cells were seeded onto glass coverslips in 12-well plates. Nuclear changes were analyzed by DAPI staining. After cells were incubated with HBHA for the indicated times, they were fixed with 4% paraformaldehyde and incubated with DAPI (10 µg/mL) for 10 min in the dark. The nuclei of stained cells were visualized using an Olympus BX50 fluorescence microscope (Olympus Optical Co., Hamburg, Germany). To determine the localization of cytochrome c or Bax, cells treated with HBHA were incubated in pre-warmed medium containing 100 nM of Mitotracker Red (Molecular Probes), fixed in 4% paraformaldehyde, permeabilized with 0.1% Triton X-100, and then stained with anti-cytochrome c or anti-Bax and Alexa-488-conjugated secondary antibody (Jackson Immuno Research Laboratories) before confocal microscopy. The subcellular localization of HBHA was analyzed using a confocal microscope (LSM510 META; Carl Zeiss). The cells incubated with Mitotracker Red were fixed, permeabilized, and stained with an anti-HBHA antibody followed by a fluorophore-conjugated antibody (anti-mouse IgG Alexa-488). After DAPI staining, cells were imaged with a confocal microscope.
Subcellular fractionation was performed as previously described [56]. Briefly, cells were incubated on ice for 5 min in 100 µL of ice cold CLAMI buffer (200 mM sucrose, 70 mM KCl, 200 µg/mL digitonin in PBS) and centrifuged at 1,000 × g for 5 min at 4°C. The supernatants (cytosolic fractions) were stored at −80°C and the pellets were resuspended in 50 µL of IP buffer (50 mM Tris-Cl, pH 7.4, 150 mM NaCl, 2 mM EDTA, 2 mM EGTA, 0.2% Triton X-100, 0.3% NP-40) containing protease inhibitor cocktail (Roche Diagnostics Corporation, Indianapolis, IN) and incubated on ice for 10 min. The samples were centrifuged at 10 000 × g for 5 min at 4°C and the supernatants (mitochondrial fractions) were stored at −80°C until use in further experiments.
Cells were detached, centrifuged, and lysed in lysis buffer (10 mM Tris, pH 7.4, 5 mM EDTA, 150 mM NaCl, 1% Triton X-100, 1 mM PMSF, protease inhibitor cocktail). Protein concentrations were determined with the Bradford assay and 30 µg of protein was separated with SDS-PAGE, followed by electrotransfer to a nitrocellulose membrane (Hybond-ECL; Amersham Pharmacia Biotech). The blots were probed with primary antibodies at optimized concentrations followed by horseradish peroxidase-conjugated secondary antibodies. The enhanced chemiluminescence system (ECL; Amersham/GE Healthcare) followed by exposure to chemiluminescence film was used to visualize proteins.
Intracellular ROS were evaluated through staining cells with H2DCFDA. Cells were incubated in 10 µM H2DCFDA for 30 min at 37°C, washed, and detached. Resuspended cells were washed and immediately analyzed by flow cytometry using FACSCanto. At least 10,000 cells per sample were analyzed using CellQuest Pro acquisition and analysis software.
Mitochondria were isolated from 1 × 108 RAW 264.7 cells as described previously [57]. Briefly, cells were harvested by centrifugation at 600 × g and resuspended in ice-cold IB buffer (10 mM Tris-MOPS, 200 mM sucrose, 1 mM EGTA/Tris, pH 7.4). All subsequent centrifugations were performed at 4°C. The cells were then homogenized with 35 strokes in a glass potter after incubation for 10 min on ice. Cell debris was removed by centrifugation at 600 × g for 10 min, and then the supernatant was centrifuged for 10 min at 7 000 × g to precipitate mitochondria. The pellet was then resuspended in EB buffer (10 mM Tris-MOPS, 125 mM KCl, 100 µM EGTA/Tris, 1 mM KH2PO4, pH 7.4). An aliquot of the preparation was incubated with HBHA for 1 h at 37°C and centrifuged for 10 min at 7 000 × g. The pellet containing mitochondria was resuspended in the same buffer and stained with DiOC6. An average population of 50 000 mitochondria was analyzed by flow cytometry. Alternatively, the proteins contained in the supernatant were concentrated with by ultrafiltration using a 3-kDa cutoff Centricon device (Amicon, Millipore, Bellerica, MA). Immunoblot analysis for cytochrome c was performed as described above.
The data represent the mean ± standard deviation (SD) from at least three independent experiments. Statistical analyses were performed using unpaired Student's t tests with Bonferroni adjustment. A P-value of <0.05 was considered significant.
|
10.1371/journal.pbio.0060006 | Conformational Equilibria in Monomeric α-Synuclein at the Single-Molecule Level | Human α-Synuclein (αSyn) is a natively unfolded protein whose aggregation into amyloid fibrils is involved in the pathology of Parkinson disease. A full comprehension of the structure and dynamics of early intermediates leading to the aggregated states is an unsolved problem of essential importance to researchers attempting to decipher the molecular mechanisms of αSyn aggregation and formation of fibrils. Traditional bulk techniques used so far to solve this problem point to a direct correlation between αSyn's unique conformational properties and its propensity to aggregate, but these techniques can only provide ensemble-averaged information for monomers and oligomers alike. They therefore cannot characterize the full complexity of the conformational equilibria that trigger the aggregation process. We applied atomic force microscopy–based single-molecule mechanical unfolding methodology to study the conformational equilibrium of human wild-type and mutant αSyn. The conformational heterogeneity of monomeric αSyn was characterized at the single-molecule level. Three main classes of conformations, including disordered and “β-like” structures, were directly observed and quantified without any interference from oligomeric soluble forms. The relative abundance of the “β-like” structures significantly increased in different conditions promoting the aggregation of αSyn: the presence of Cu2+, the pathogenic A30P mutation, and high ionic strength. This methodology can explore the full conformational space of a protein at the single-molecule level, detecting even poorly populated conformers and measuring their distribution in a variety of biologically important conditions. To the best of our knowledge, we present for the first time evidence of a conformational equilibrium that controls the population of a specific class of monomeric αSyn conformers, positively correlated with conditions known to promote the formation of aggregates. A new tool is thus made available to test directly the influence of mutations and pharmacological strategies on the conformational equilibrium of monomeric αSyn.
| Natively unstructured proteins defy the classical “one sequence–one structure” paradigm of protein science. In pathological conditions, monomers of these proteins can aggregate in the cell, a process that underlies neurodegenerative diseases such as Alzheimer and Parkinson. A key step in the aggregation process—the formation of misfolded intermediates—remains obscure. To shed light on this process, we characterized the folding and conformational diversity of αSyn, a natively unstructured protein involved in Parkinson disease, by mechanically stretching single molecules of this protein and recording their mechanical properties. These experiments permitted us to observe directly and quantify three main classes of conformations that, under in vitro physiological conditions, exist simultaneously in the αSyn sample. We found that one class of conformations, “β-like” structures, is directly related to αSyn aggregation. In fact, their relative abundance increases drastically in three different conditions known to promote the formation of αSyn fibrils. We expect that a critical concentration of αSyn with a “β-like” structure must be reached to trigger fibril formation. This critical concentration is therefore controlled by a chemical equilibrium. Novel pharmacological strategies can now be tailored to act upstream, before the aggregation process ensues, by targeting this equilibrium. To this end, single-molecule force spectroscopy can be an effective tool to tailor and test new pharmacological agents.
| A significant fraction (possibly as much as 30%) of proteins and segments of proteins in eukaryotic proteomes has been found to lack, at least partially, a well-defined three-dimensional structure. Proteins belonging to this class are usually called natively unfolded proteins (NUPs) [1]. NUPs have been found to play key roles in a wide range of biological processes like transcriptional and translational regulation, signal transduction, protein phosphorylation, and the folding of RNA and other proteins [2]. The conformational heterogeneity of NUPs allows them to adopt conformations that trigger pathogenic aggregation processes. In fact, NUPs are involved in the pathogenesis of some of the most widespread and socially relevant neurodegenerative diseases, such as Alzheimer and Parkinson [3–5]. Despite intensive research, the folding and the aggregation mechanisms of NUPs remain a major unsolved problem.
Theoretical studies depict the apparent structural disorder of NUPs as the result of the coexistence of a complex ensemble of conformers ensuing from a rugged energy landscape [6]. Five clusters of conformations, each with its own characteristic tertiary structure, were identified by molecular dynamics studies on the Alzheimer β peptide [7]. Traditional bulk experiments and spectroscopies have recently been providing experimental evidence of the conformational diversity of these proteins [3,5]. Because of their inherent ensemble averaging, however, these methodologies cannot reveal the full complexity of the conformational equilibria of NUPs. Single-molecule methodologies can single out the structures adopted by individual molecules within a complex conformational equilibrium [8–14].
We decided to approach the problem of the characterization of the conformers of α-synuclein (αSyn), which is a prototype of this class of proteins. αSyn is a 140–amino acid (aa) protein expressed primarily at the presynaptic terminals in the central nervous system, and it is thought to be physiologically involved in endoplasmic reticulum–Golgi vesicle trafficking [15]. αSyn is involved in the pathogenesis of several neurodegenerative diseases, called synucleopathies. Intracellular proteinaceous aggregates (Lewy bodies and Lewy neurites) of αSyn are hallmarks of Parkinson disease [16] and multiple system atrophy [17]. Three naturally occurring mutations in the αSyn protein sequence—A30P, A53T, and E46K—have been identified so far in human families affected by familial Parkinsonism [18–20]. These mutant proteins display an increased tendency to form nonfibrillar aggregates [21] and Lewy bodies–like fibrils in vitro [22].
The fibrils spontaneously formed by αSyn by a nucleation-dependent mechanism are rich in β structure [23,24]. The transition from the natively unfolded monomeric state to fibril is therefore a process of acquiring structure. This process is still under strong debate. Evidence is accumulating that the monomeric αSyn, under in vitro physiological conditions, populates an ensemble of conformations including extended conformers and structures that are more compact than expected for a completely unfolded chain [25–32]. The marked differences between the scenarios depicted in those studies are mostly determined by the different time scales of the ensemble averaging of the different methods used. Moreover, it is difficult for bulk methodologies to single out the monomeric state in the presence of soluble oligomers when they form quickly in solution [33]. On the contrary, the single-molecule force spectroscopy (SMFS) approach reported here describes, by design, the conformational equilibrium of the monomeric form.
The different structures assumed by αSyn have been commonly investigated by adding to its buffer solution different chemicals, such as methanol or trifluoroethanol [34], metal cations like Cu2+ and Al3+ [35,36], or sodium dodecyl-sulfate (SDS) micelles [37–39] in order to shift the conformational equilibrium toward the form under investigation. A previous force spectroscopy experiment showed that a relevant 12-aa segment of αSyn is conformationally heterogeneous [40]. The approach we report can span the full conformational space of the whole protein and also identify poorly populated conformers of the monomeric αSyn in in vitro physiological conditions. Three distinct classes of structures in equilibrium were identified: random coil, a mechanically weak fold, and “β-like.” Their populations were also monitored under conditions known to influence aggregation, such as the presence of Cu2+, high buffer concentration and, most importantly, the pathogenic mutation A30P.
To stretch an individual αSyn molecule by AFM, we need handles to connect one end of the protein to the tip and the other to the substrate. To this aim, we followed the design proposed by J. Fernandez for the study of the random coiled titin N2B segment [41]. A chimeric polyprotein composed of a single αSyn module flanked on either side by three tandem I27 domains (Figure 1A, 3S3) was expressed [42–44]. These domains act as molecular handles to mechanically stretch a single αSyn molecule. They also introduce well-characterized fingerprint signals into the recorded force curves that make it possible to identify the different αSyn conformations. The design is such that if the number of unfolding signals coming from I27 modules is larger than four, we are sure to have also mechanically stretched the αSyn module in the middle (Figure 1A). Among the curves showing mechanical unfolding events, however, only those featuring at least six unfolding peaks were selected and analyzed. This choice reduced the statistical sample even more, but it allowed us to recognize, in a very stringent way, the signatures of the different conformations of the αSyn moiety on each construct molecule that had been stretched.
To probe the native-like conformer population of αSyn, we performed experiments in a 10 mM Tris buffer solution. We found that the profiles of the selected force curves can be classified into three main classes. Two were unambiguously assigned to well-defined classes of conformers: one with the typical mechanical behavior of random-coil chains and the other of β-like structures. We propose that the profiles of the third class correspond to fairly compact architectures, likely to be sustained also by interactions among different modules of the construct.
In the class of traces depicted in Figure 1B, the force curve exhibits (from left to right) a long initial region, without any significant deviation from the worm-like chain (WLC) behavior [45], followed by a saw-tooth pattern with six consecutive unfolding events, in addition to the last one that corresponds to the final detachment of the molecule from the tip. The initial region corresponds to the extension of a chain that occurs at low force and without significant energy barriers limiting its extensibility.
The six unfolding peaks are spaced by ∼28 nm. This spacing between the peaks corresponds to an 89-aa chain (0.36 nm per amino acid [46]), i.e., to the increase in length of the protein after the unfolding of one I27 domain. These six unfolding peaks correspond to the characteristic fingerprint of the mechanical unfolding of the I27 modules [41]. We can therefore infer that, in this case, the AFM tip picked up the 3S3 construct molecules at the His-tag terminus, while the other end was tethered to the gold surface by the C-terminal cysteines.
The location of the first unfolding peak of I27, corresponding to the contour length of the construct molecules prior to any unfolding event, proves that the preceding featureless part of the trace can be unambiguously assigned to the αSyn chain. In fact, the measured contour length that fits this peak is 77 ± 4 nm. Subtracting the length of the six, still folded, I27 domains from this value (4.5 nm each [47]), a value of 48 ± 4 nm is obtained. This length corresponds to the chain of 140 aa of the αSyn. Therefore, this featureless initial part is the signature of αSyn conformers with the mechanical properties of a random coil. Their average persistence length was estimated by fitting the WLC model at 0.36 ± 0.05 nm. About 38% of the molecules showed this mechanical behavior in Tris/HCl buffer 10 mM (Figure 2).
A significant proportion of force curves with seven regularly spaced unfolding peaks in the 200-pN range (in addition to the last one corresponding to the final detachment) (Figure 1C) was also recorded.
The presence of a number of unfolding events greater than that of the I27 modules in the construct cannot be ascribed to a possible simultaneous pulling of more than one 3S3 molecule, because pulling two multidomain constructs at the same time would not likely lead to a uniform separation between the I27 unfolding events. Moreover, we never obtained a significant and uniform set of reproducible curves with eight or more unfolding peaks with 28-nm separation. Curves with seven unfolding events were well reproducible, and their statistics were unambiguously modulated by conditions able to trigger aggregation: e.g., ionic strength, the presence of Cu2+ ions and, most importantly, pathogenic mutations (see below). The appearance of seven unfolding events cannot come from a construct accidentally expressed with seven, instead of six, I27 domains because of the cloning strategy (see Materials and Methods). The occurrence of a seventh peak due to the stretching of 3S3 dimers can be also ruled out. Dimers could form in solution via disulfide bonds between the terminal cysteines, but those bonds tend to dissociate into thiols in the presence of gold, because the gold–sulfur bond is more stable than the sulfur–sulfur bond [48]. Each monomer contains two terminal cysteines: one of them could be involved in the dimerization and the other could bind to the gold surface. Even in this unlikely event, the length of the tethered chains extending from the surface is the same as that of a nondimerized construct. Therefore, also in the case of a dimer tethered to the surface, more than six I27 unfolding peaks with the same separation cannot be recorded. We nevertheless tested the sample using dithiothreitol (DTT) to avoid any disulfide-bonded dimer formation. Under these conditions, the statistics of different populations was comparable to those in the standard buffer, and we still recorded a significant proportion (∼10%) of seven-peaked curves.
Because of the previous considerations, we therefore assign one of the seven peaks to the unfolding of the αSyn moiety. The length (95 aa) of this αSyn β-like folded section accidentally coincides with that of the I27 domain. This coincidence hinders the possibility to discriminate the peak of the αSyn from the six of the I27 domains. Nevertheless, the assignment of these curves to the unfolding of the αSyn moiety is confirmed by the position of the first unfolding peak, i.e., by the contour length of the construct molecules prior to any unfolding event. As shown in Figure 3, the position values correspond to a chain composed of the six I27 folded modules, plus the αSyn moiety with its C-terminal segment of 50 aa fully unfolded, and the remaining 95 amino acids folded into a structure with the same contour length as a folded I27 domain (solid line). The low propensity to fold of the 50 aa of the very acidic αSyn C-terminal tail has been extensively documented [38,39,49]. Segments of the remaining 95 amino acids are instead known to fold under different conditions into an α helix [38,39] or, in the amyloid, into a β sheet structure [31]. It must be noted that in about 40% of the molecules, the contour length of the same folded section is larger than that corresponding to 95 aa. The αSyn structural diversity therefore includes also β-like chain portions with different lengths. This interpretation is confirmed by comparing the variance of the folded section of seven-peaked curves with that of I27 modules (Figure 3).
The 200-pN unfolding force of all the seven peaks indicates that the folded section of αSyn has the same mechanical properties of the I27 β-sandwich structure. At the moment, without any independent structural characterization, we consider and label this folded structure of the αSyn moiety just as β-like, in accordance with its mechanical behavior. Nevertheless, its mechanical behavior is in agreement with a β sheet content in the β-like class of conformers. It is unlikely that the α-helical content we observed by means of circular dichroism (CD) (see below) correlates with the “β-like” conformers. In fact, whereas β-structures, like those of titin modules, such as I27 [41,50], or tenascin [51], unfold at forces in the range of 100–300 pN (at loading rates of the order of 10−5 N/s), the α helix domains, in the same conditions, are always observed to unfold at forces almost one order of magnitude smaller [52–55].
In conclusion, these curve profiles provide clear evidence that in 10 mM Tris/HCl buffer, about 7% of the molecules (Figure 2) contain a segment of the αSyn chain of about 95 aa folded into a structure with the mechanical property of the I27 β-sandwich structure. This percentage of the β-like structures, as we will see below, can be related with conditions leading to pathogenic aggregation.
The remaining force spectroscopy curves (Figure 1D) show single or multiple small peaks (sometimes with a plateau- or dome-like appearance) superimposed on the purely entropic WLC behavior of the trace preceding the six saw-tooth–like peaks.
The geometry of our construct made it possible to exclude that those small peaks might correspond to the rupture of aspecific αSyn-gold interactions. In fact, if the unstructured αSyn was adsorbed on the surface, upon pulling the construct, we would have recorded the first event at a distance from the tip contact point corresponding to the length of the three I27 modules (∼13.5 nm). The mechanically weak events we observed instead took place at an average distance from the contact point of 60 ± 26 nm with no events below 20 nm. They are therefore not compatible with αSyn-gold interactions. We assign these signals to the rupture of mechanically weak interactions placed at short and long distances along the chain.
The average forces of those single or multiple small peaks of the profiles are in the 64 ± 30–pN range (well above the noise level), without a defined hierarchy; often stronger peaks precede weaker ones, hinting topologically “nested” interactions. From the difference between the contour length estimated at those small peaks and that at the first I27 unfolding peak, one can measure the size of the topological loop enclosed by the interactions whose rupture is monitored by the different peaks. The resulting broad distribution of these distances monitors the ample multiplicity of these interactions as discussed in Protocol S1.
More than 50% of the molecules showed short- and long-distance mechanically weak interactions in 10 mM Tris/HCl buffer (Figure 2). These interactions were also monitored by ensemble-averaged fluorescence spectroscopy. The fluorescence comes from the tryptophan residues of the I27 domains which are absent in αSyn. The fluorescence spectra reported in Figure 4A prove that interactions between the I27 handles and tracts of the αSyn moiety do take place, as shown by the broadening of the spectrum of 1S1 with respect to that of the 3T construct (See Figure 1A and Materials and Methods section for constructs description) and by the 5-nm shift of the λmax. The possibility of partial I27 unfolding leading to Trp exposure and broadening of the spectrum is ruled out by the CD data and by our force curves, which show that I27 domains are as tightly folded in the 3S3 construct as in an I27 homopolymer. A broadening due to subtle conformational effects on the I27 domain that expose the I27 Trp residue is possible, but even in this case, the fact that this broadening happens only when the αSyn moiety is inserted in the construct proves that direct interaction is taking place.
CD spectra of 1S1 and 3T were recorded in which 1S1 shows some α-helical content in the αSyn moiety (Figure 4B). . Subtraction of the contribution of the I27 linkers (2/3 of the CD of 3T recorded in the same 10 mM Tris/HCl buffer) from the CD spectrum of 1S1 reveals a profile that is different from that of αSyn in the same buffer condition (Figure 4C) but similar to that of the same protein in the α helix structure induced by the addition of SDS [33]. This α-helical content might be induced by the interactions between the αSyn moiety and the I27 domains as discussed below and in Protocol S1.
It is well known that multivalent metal cations like Cu2+ can accelerate αSyn aggregation [35,36]. To validate our approach and to investigate how metal cations influence the conformer equilibrium of αSyn, we performed SMFS experiments on the 3S3 construct in 10 mM Tris/HCl buffer in the presence of 1 μM CuCl2. The low concentration of copper was chosen to target the His 50 specific copper binding site of αSyn (dissociation constant Kd = 0.1 μM) [36].
The presence of 1-μM Cu2+ moderately, but significantly (χ2 statistical significance p < 0.01), alters the relative distribution of the αSyn conformers with respect to plain 10 mM Tris/HCl (see Figure 2). In particular, the relative population of the β-like conformers more than doubles (from 7.2% to almost 18%), with a parallel decrease of the signals coming from mechanically weak interactions. An increase (from 38% to 47%) of random coil-like curves is also observed.
The A30P mutation is a pathogenic, naturally occurring human αSyn variant, that correlates with familial Parkinsonism [19]. The mutant protein displays an increased rate of oligomerization [56] and impaired degradation by chaperone-mediated autophagy [57]. We tested the 3S3 αSyn-A30P construct to evaluate the capability of our methodology to probe different conformational propensities in mutants of the same protein. We found that the A30P mutation induces a striking shift in the conformational equilibrium of αSyn with β-like curves being around 37% of the sample and again, a corresponding decrease of signals coming from mechanically weak interactions (Figure 2). In contrast with wild-type αSyn incubated with Cu2+, the A30P mutant does not induce an increase of random coil curves that are exactly in the same proportion observed in wild-type αSyn.
Another condition known to speed up αSyn aggregation is high ionic strength [26,28]. SMFS experiments on the 3S3 wild type construct were performed also in 500 mM Tris/HCl buffer. As reported in Figure 2, the frequency of the three types of profiles radically changed in different ionic strength conditions. The most remarkable result is, again, the significant increase in the population of the β-like structures with buffer concentration (up to about 28%) and the parallel decrease of the percentage of the mechanically weak structures. An increase of random coil curves is also observed, as occurs in the presence of Cu2+, but unlike the case of the A30P mutant.
We have identified the signatures of three classes of conformers in monomeric αSyn at the single-molecule level. One of these classes includes structures that are mechanically indistinguishable from a random coil; the other two classes include β-like structures and structures kept together by short- and long-distance mechanically weak interactions (Figure 1). We have also observed that their equilibrium shifts significantly depending on solution conditions or sequence variants related to pathological aggregation. The important result that emerges from these data is the direct correlation between conditions known to increase the αSyn aggregation propensity and the relative size of the β-like population (Figure 2).
We observed a marked increase of the population of “β-like” conformers under three very different conditions known to accelerate αSyn aggregation. This result links the population of those αSyn monomeric conformers to the process of αSyn aggregation. The first condition is the presence of a μM concentration of Cu2+. Our results in this condition agree with the observation of a metal-induced partially folded intermediate by Uversky, Li, and Fink [35]. Also Rasia et al. suggested a compact set of metal-induced conformations, noticing that the specific binding of Cu2+ to the αSyn N terminus requires the formation of a metal-binding interface (pivoted on His 50), which possibly involves residues that are widely separated in the primary amino acid sequence [36].
The second condition is the A30P mutation. Nuclear magnetic resonance (NMR) experiments have observed a much more flexible average conformation of the αSyn mutants A30P and A53T. The increased average flexibility of αSyn allows the protein to sample a larger conformational space. [58]. Interestingly, the mean hydrodynamic radius of αSyn is not affected by the A30P and A53T mutations [21,59], thus showing that the increased flexibility is compatible with the population of compact folded structures like those singled out by our experiments.
The third condition is a radical increase of the ionic strength. Our results in 500 mM Tris/HCl can be reconciled with the model proposed by Hoyer et al. [26] and by Bernado et al. [28] to explain the well-documented phenomenon of the increased αSyn fibril formation with increasing ionic strength. According to that model, the increased fibril formation is explained just on the basis on an increased freedom of the fibrillogenic NAC region caused by the release of its interaction with the negatively charged C-terminal tail. The increased ionic strength of the buffer leads to a more efficient charge shielding of the strongly acidic C-terminal tail, thus relieving its electrostatic self-repulsion. This in turn leads to the lowering of the protein-excluded volume and increases its flexibility. According to our data in Figure 2, we should add to this model a shift of the conformational equilibrium toward the β-like structures that takes place on increasing the charge shielding.
Any assignment of force spectroscopy signals to a definite secondary canonical structure must be supported by independent structural data. We have labeled as β-like those conformers with a mechanical behavior closely matching those of structures rich in β sheets. The correlation of the population of these structures with aggregation conditions, which enrich β sheet content in αSyn, supports this labeling. Evidence of some β sheet content in the monomeric state of αSyn was previously reported in the literature. Most recently by means of NMR spectroscopy in supercooled water at minus 15 °C, it was found that the αSyn chain, cold-denatured to an hydrodynamic radius equivalent to that displayed by the same protein in 8-M urea, retains a surprising amount of unpacked β strand content that correlates with the amyloid fibril β structure [32]. The packing of these β strands into compact structures like those observed by us is thus likely to occur in nondenaturing conditions and at physiological temperatures. This NMR result supports our observation of β-like conformers in the monomeric state of αSyn and links them to the amyloid β structure. The presence of β sheet structures was indicated also by Raman spectra of this protein in aqueous solution[33]. In the same investigation, CD spectroscopy proved unable to detect any β content. Correspondingly, the CD spectra of αSyn recorded by us in 10 mM and 500 mM Tris were practically superimposable. We conclude that CD is not a technique sensitive enough to detect partial β-sheet content in the αSyn sample. A fraction of β-sheet/extended structure of about 19% was also detected, again not by CD, but by Fourier transform infrared (FTIR) spectroscopy in dried films of αSyn [60]. This fraction is much larger than that estimated by our experiments in 10 mM Tris/HCl buffer (see Figure 2). However, the conditions of the SMFS and FTIR experiments were markedly different, and in the latter case, some template-mediated formation of β structures due to the packing of the αSyn molecules in the dried films required by the FTIR measurements cannot be ruled out.
In conclusion, despite the fact that force spectroscopy data cannot directly assign a specific secondary structure to the conformers we have labeled as β-like, it is most likely that they have significant β sheet content.
By now, any structural characterization of the mechanically weak interactions events monitored by the small peaks in force curves as in Figure 1D (right panels) is at best tentative and falls outside the focus of the present work. A more detailed characterization of these events is, however, within the range of capabilities of the techniques proposed here and is being currently addressed in our laboratory (see Protocol S1 for preliminary measurements). A plausible explanation of the short- and long-distance mechanically weak interactions we observed cannot exclude the interaction between positively charged residues on the αSyn N terminal and the negatively charged surface of I27 modules [61]. It has been documented that αSyn in contact with negatively charged surfaces assumes an α helix structure [37–39,62,63]. We might expect a similar structural transition in the αSyn moiety also from the contact with the I27 modules within the 3S3 or 1S1 constructs (see Protocol S1). This transition is indicated by the CD spectra of the 1S1 construct in 10 mM Tris/HCl (see Figure 4B). We propose that the small peaks like those shown in Figure 1D and assigned to the mechanically weak interactions can be the signature of the interaction between αSyn, possibly in α helical form, and the flanking I27 modules. It is not surprising that more than one of those signals are present in the same force curves, because multiple interactions of this type can occur at the same time in the same molecule. It should be noted that the same transition does not take place when free αSyn is mixed in solution with I27 modules of the 3T construct (see Protocol S1). An electrostatic model, based on the interaction lengths calculated from the positions of the small peaks in the force curves like those displayed in Figure 1D (right panels), is proposed in Protocol S1.
Notably, these short- and long-distance mechanically weak interactions are observed to be in equilibrium with the β-like structures. The population of the former always decreases while that of the latter increases. This result is in accord with the observation by Zhu et al. that a driving force to α helical structures inhibits αSyn fibril formation [60] and also rule out any template-mediated β sheet imprinting by the I27 linkers. This conclusion is confirmed by the data on 500 mM Tris/HCl buffered solutions showing that when electrostatic interactions between the αSyn moiety and the flanking I27 linkers are decreased, the population of β-like conformers increase. We can also expect entropic effects due to the presence of the flanking I27 domains to drive the protein toward more extended conformations rather than compact conformations [64,65].
These considerations indicate that the design and use of alternative linkers or experimental strategies may prove useful in the future to further discriminate the effective conformational distribution of αSyn from alterations due to the interaction with the linkers.
For the first time, to our knowledge, we applied the AFM-based single-molecule mechanical unfolding methodology to a multimodular protein containing the αSyn moiety. This approach brings into play three main methodological capabilities inaccessible to the bulk ensemble–averaged spectroscopies previously applied to study the structure of αSyn and other natively unstructured proteins.
The first is the possibility to work strictly at the single-molecule level, thus ensuring that the conformer distribution of the monomeric αSyn is detected and quantified without interference from oligomeric soluble forms of the protein and therefore of any possible intermolecular imprinting toward the amyloidogenic β structures.
The second capability is that of spanning the conformational space of the protein under investigation and of directly catching and quantifying all of its conformers with a lifetime longer than 10−3 s. These conformers, because of their longer life time, might be the most biologically relevant. Three classes of the monomeric αSyn conformations, including random coil, mechanically weakly folded and β-like, were characterized by our experiments. They could be detected even in low concentration without the necessity of selectively enhancing one of them by adding specific agents to unbalance the conformational equilibrium, as most commonly done so far with bulk ensemble–averaged experiments.
The third capability is that of following shifts in the population of these classes of conformers in response to changing the solution conditions or the protein sequence and to detect them even if scarcely populated. In the case of αSyn, conditions known to promote oligomerization and aggregation—like the presence of Cu2+, the A30P mutation, or a radical increase of ionic strength—markedly shift its conformational equilibrium toward the β-like form at the expense of other structures. These results indicate that the β-like curves contain the signature of the structural precursor to αSyn oligomerization. We suggest that the different aggregation propensities and, ultimately, the pathogenicity displayed by αSyn under different environmental conditions or point mutations can be triggered by unbalancing the delicate equilibria among αSyn conformers.
These capabilities suggest that in the near future, single-molecule methodologies will play a crucial role in studies of the folding equilibria of the NUP monomers and, in particular, in the detection and quantification of the conformers that can lead to aggregation of those proteins. Our results suggest the feasibility of single-molecule approaches to the testing of novel pharmacological or biophysical therapies for pathologies involving the conformational equilibria of NUPs.
We followed the protein construct design proposed by J. Fernandez for the study of the random coiled titin N2B segment [41]. Chimeric polyproteins were obtained starting from pAFM1–4, pAFM5–8, and pAFM(I27)3mer vectors, kindly provided by Professor Jane Clarke (Cambridge University, United Kingdom) and constructed according to [43]. αSyn or its A30P mutated sequences were amplified by PCR using two different pairs of primers, each containing unique restriction sites. A first pair contained KpnI and XbaI sites, and a second one contained SacI and BssHII sites. The original eight I27 module plasmid was reconstituted from pAFM1–4 and pAFM5–8, obtaining the pAFM8m vector. pAFM8m was then digested with KpnI and XbaI and ligated to the amplified αSyn sequence, then cleaved by the same enzymes in substitution of the two central titin modules to give the pAFM3s3 vector (see Protocol S1). By a similar strategy, the pAFM(I27)3mer vector was digested with SacI and BssHII, and the central titin module replaced by αSyn sequence, obtaining the pAFM1s1 vector. The obtained expression plasmids, pAFM3s3 and pAFM1s1, code for two chimeric polyproteins composed of a single αSyn module flanked on either side by three tandem I27 domains or by just one, named 3S3 and 1S1, respectively. The two pAFM8m and pAFM(I27)3mer vectors (coding for two recombinant poly(I27) proteins named 8T and 3T) were transformed into Escherichia coli C41(DE3) cells [66] (obtained from Professor John E. Walker [Medical Research Council–Dunn Human Nutrition Unit, Cambridge, United Kingdom] with the agreement of the Medical Research Council center of Cambridge). The cells were grown and the expression of proteins was induced as described in [43]. Recombinant proteins were purified by Ni2+-affinity chromatography in 20 mM sodium phosphate buffer pH 8, 500 mM NaCl; the elution from the resin was obtained with 20 mM imidazole. After dialysis, proteins were kept at −80 °C in phosphate buffered saline (PBS) with 15% glycerol. The purification gel is shown in Protocol S1.
CD measurements were carried out on a JASCO J-715 spectropolarimeter interfaced with a personal computer. The CD spectra were acquired and processed using the J-700 program for Windows. All experiments were carried out at room temperature using HELLMA quartz cells with Suprasil windows and an optical path length of 0.1 cm. Spectra were recorded in the 190–260 nm wavelength range using a bandwidth of 2 nm and a time constant of 2 s at a scan speed of 50 nm/min. The signal-to-noise ratio was improved by accumulating at least four scans. All spectra are reported in terms of mean residue molar ellipticity [Θ]R (deg cm2 dmol−1).
Fluorescence emission spectra were recorded on a Perkin-Elmer LS 50 spectrofluorimeter equipped with a thermostated cell compartment and interfaced with a personal computer using the FL-WinLab program for Windows. Sample measurements were carried out using a HELLMA ultra-micro cell with Suprasil windows and an optical path length of 10 × 2 mm. Fluorescence spectra were obtained at 25 °C using an excitation wavelength of 288 nm, with an excitation bandwidth of 4 nm and emission bandwidth of 4 nm. Emission spectra were recorded between 290–380 nm at a scan rate of 60 nm/min.
Due to the well-known structuring effects of divalent metal ions on αSyn [35], an accurate elemental analysis of the buffer was performed to exclude artifacts in our results due to metal contamination. The high concentration Tris-buffer solution (500 mM) was analyzed for metal contents by atomic absorption spectroscopies. The measured concentrations were Cu = 0.2 ± 0.1 nM, Zn = 3.5 ± 0.1 nM, Fe = 0.9 ± 0.1 nM, and Ca = 22.5 ± 0.1 nM. These values are two orders of magnitude lower than the concentration required to induce structural effects on αSyn [67].
Gold (Alfa Aesar, 99.99%) was deposited onto freshly cleaved mica substrates (Mica New York Corp., clear ruby muscovite) in a high-vacuum evaporator (Denton Vacuum, model DV502-A) at 10−5 Torr. Before deposition, the mica was preheated to 350 °C by a heating stage mounted behind the mica to enhance the formation of terraced Au(111) domains. The typical evaporation rate was 3 Å/s, and the thickness of the gold films ranged around 300 nm. The mica temperature was maintained at 350 °C for 2 h after deposition for annealing. This method produced samples with flat Au(111) terraces. These films were fixed to a glass substrate with an epoxy (EPO-TEK 377, Epoxy Tech.). They were then separated at the gold–mica interface by peeling immediately before functionalization with the desired molecules. This procedure produced gold substrates with a flat surface morphology due to the templating effect of the atomically flat mica surface [68,69].
For each experiment, a 20 μl drop of 3S3 construct solution (160 μg/ml) was deposited on the freshly peeled gold surface for about 20 min. SMFS experiments were performed using a commercially available AFM system: Picoforce AFM with Nanoscope IIIa controller (Digital Instruments) using V-shaped silicon nitride cantilevers (NP; Digital Instruments) with a spring constant calibrated by the thermal noise method [70]. The pulling speed was 2.18 μm/s for all experiments. The buffer used was Tris/HCl (10 mM or 500 mM, pH 7.5; the 10 mM buffer was obtained by diluting the 500 mM buffer with milliQ ultrapure water). For CuCl2 experiments, the protein was deposited in a drop with the addition of a final concentration of 1 μM CuCl2 and left on the surface for about 20 min, and the experiments were carried out in 10 mM Tris/HCl with 1 μM CuCl2. Control experiments in DTT were made in 50 mM DTT Tris/HCl buffer.
The force curves were analyzed using the commercially available software from Digital Instrument (Nanoscope v6.12r2), custom Origin scripts and Hooke, a Python-based home coded force spectroscopy data analysis program (M. Sandal, unpublished work). Force curves were analyzed fitting each peak with a simple WLC force versus extension model [45] with two free parameters: the contour length L and the persistence length p (Equation 1). The I27 modules were characterized in terms of the length of the polypeptide chain extended after each unfolding event.
To assess the statistical validity of the comparison between data obtained in 10 mM Tris/HCl buffer and those obtained in other conditions, standard chi square tests were performed. The differences between the 10 mM Tris data set and the other data sets are significant, with p < 0.01.
The UniProt KB (http://www.ebi.ac.uk/trembl/index.html) accession number for α-synuclein (αSyn) is P37840 (SYUA_HUMAN). The Protein Data Bank (http://www.pdb.org) entry for the I27 domains included in the polyprotein is 1tit. The Online Mendelian Inheritance in Man (OMIM) reference number for familial Parkinsonism is 163890.0002.
|
10.1371/journal.pgen.1002960 | Rif2 Promotes a Telomere Fold-Back Structure through Rpd3L Recruitment in Budding Yeast | Using a genome-wide screening approach, we have established the genetic requirements for proper telomere structure in Saccharomyces cerevisiae. We uncovered 112 genes, many of which have not previously been implicated in telomere function, that are required to form a fold-back structure at chromosome ends. Among other biological processes, lysine deacetylation, through the Rpd3L, Rpd3S, and Hda1 complexes, emerged as being a critical regulator of telomere structure. The telomeric-bound protein, Rif2, was also found to promote a telomere fold-back through the recruitment of Rpd3L to telomeres. In the absence of Rpd3 function, telomeres have an increased susceptibility to nucleolytic degradation, telomere loss, and the initiation of premature senescence, suggesting that an Rpd3-mediated structure may have protective functions. Together these data reveal that multiple genetic pathways may directly or indirectly impinge on telomere structure, thus broadening the potential targets available to manipulate telomere function.
| Impaired telomere elongation eventually results in telomere dysfunction and can lead to diseases such as dyskeratosis congenita, which is associated with bone-marrow failure and pulmonary fibrosis. Cancer cells require continuous telomere maintenance to ensure continued cellular proliferation. Therefore the regulation of telomere function, both positively (in the case of dyskeratosis congenita) and negatively (for cancer), may be of therapeutic benefit. In this study we have used yeast to determine which genetic factors are important for a certain telomeric structure (the loop structure), which may help to maintain chromosome ends in a protected state. We found that multiple genetic factors and pathways affect telomere structure, ranging from metabolic signaling to specific telomere-binding proteins. We found that proper chromatin structure at the telomere is essential to maintain a telomere fold-back structure. Importantly, there was a strong correlation between telomere structure and function, as the mutants found in our screen (looping defective) were often associated with rapid senescence and telomere dysfunction phenotypes. We believe that, through the regulation of the various genetic pathways uncovered in our screen, one may be able to both positively and negatively influence telomere function.
| The physical ends of linear chromosomes resemble double-strand breaks (DSBs) in many respects with the exception that DSBs result in the activation of the DNA damage response and are eventually subject to repair; activities to which telomeres are refractory [1]. This essential quality of telomeres exists as a result of their repetitive sequence that is bound by specific proteins (shelterin and CST complexes), which in turn inhibit DNA damage checkpoints, DNA repair activities and exonuclease-mediated degradation [2]–[3]. In yeast, the CST (Cdc13-Stn1-Ten1) complex is essential for viability and prevents the accessibility of 5′ exonucleases (primarily Exo1) to the telomere [4]–[6]. Upon inactivation of CST with temperature-sensitive alleles of CDC13 and STN1, cells undergo a DNA damage-mediated checkpoint arrest due to the accumulation of single-stranded (ss) telomeric DNA [6]–[8]. In parallel, the Rap1, Rif1 and Rif2 complex, also contribute to telomere end protection by limiting telomeric ssDNA accumulation and subsequent checkpoint activation [9]–[10].
In most human somatic cells, telomeres shorten during each cell division due, in part, to the end-replication problem [11]–[12]. Eventually, the loss of telomeric DNA leads to telomere dysfunction, checkpoint activation and cellular senescence. Some cell types as well as most cancer cells avoid telomere attrition-induced senescence by expressing the specialized reverse transcriptase, telomerase. Telomerase elongates telomeres through the iterative addition of short sequence repeats to the 3′ ends of telomeres, compensating for the end-replication problem [12]. Wild type S. cerevisiae constitutively express telomerase, however the cellular senescence phenotype can be induced following its inactivation/deletion [13].
In yeast, reporter genes become silenced when placed in the vicinity of telomeres [14]. This telomere-induced silencing is dependent on the Sir2/3/4 lysine deacetylation (KDAC) complex, which is recruited to chromosome ends via the telomere binding protein, Rap1 [15]. Apart from the Sir2/3/4 complex, other KDACs also contribute to the heterochromatic constitution of telomeres and sub-telomeres. The class I KDAC, Rpd3 (the yeast ortholog of human KDAC1), consisting of two sub-complexes, Rpd3L and Rpd3S [16], also localizes to telomeres and is important to establish the euchromatin/heterochromatin boundary in the sub-telomeric regions [17], as well as to prevent hyper-silencing [18]. The class II KDAC, Hda1, also contributes to chromatin regulation at yeast telomeres [19]. The relationships between heterochromatin and telomere structure/function remain unclear.
It has been postulated that telomere protection may stem, in part, from a higher-order chromatin structure. Analysis of telomeric DNA from human and mouse cells has revealed that the telomere terminus can be hidden in a lariat-like structure termed a t-loop [2], [20]–[21]. T-loops, thought to form through the strand invasion of the 3′ telomeric overhang into the double-stranded region of the telomere, have also been found in chickens, worms, plants, and protozoa [22]–[25]. Via electron microscopy, telomeric loops have also been observed in yeast (K. lactis) with over-elongated telomeres [26], and the telomere associated S. pombe protein, Taz1, has been shown to re-model model DNA substrates into t-loops [27]. However, due to the small size of yeast telomeres it has been difficult to both prepare and analyze wild type length yeast telomeres via electron microscopy [26]. In the budding yeast, S. cerevisiae, both genetic and chromatin immunoprecipitation-based experiments have revealed that wild type telomeres do fold-back onto themselves and into the subtelomeric region [26], [28]–[30], suggesting that loops or fold-back structures are indeed important for telomere function in yeast. Apart from the Sir2/3/4 deacetylase complex in budding yeast being important for this fold-back [29], and the shelterin component, TRF2 in human cells being required for t-loop formation [20], [25], the regulation of such telomeric structures remains poorly understood. In this study we have taken an unbiased genome-wide screening approach in yeast to better understand how telomere structure/fold-back is regulated in vivo. We demonstrate that multiple biological processes influence telomere structure, including the state of the subtelomeric heterochromatin as dictated by multiple lysine deacetylases. Furthermore, we find that there are direct correlations between the inability of a telomere to fold-back and telomere dysfunction, implying that the loop structure may make important contributions to telomere protection.
By placing a TATA-less galactose-inducible UAS (upstream activating sequence) downstream of the URA3 gene (from hereon referred to as construct 2), URA3 transcription is only achieved when the UAS loops back and comes into proximity with the URA3 promoter [28]–[29] (Figure 1A). Fold-back-induced transcription only takes place when this construct is integrated at the telomere and does not occur when it is integrated at an internal chromosomal locus [29]. Transcription of URA3 results in lethality on media containing the drug 5-fluoroortic acid (5-FOA), providing a robust readout (cell death on 5-FOA) for successful telomere looping.
To better understand how the telomere fold-back structure in yeast is regulated, we introduced construct 2 into the yeast haploid deletion collection using the synthetic genetic array (SGA) procedure [31], resulting in the construction of ∼4800 haploid deletion mutants harboring construct 2 (Figure 1B). Robotic pinning of these strains in quadruplicate onto galactose media in the presence and absence of 5-FOA revealed potential looping defective mutants that grew on 5-FOA (Figure 1B, bottom panel example of looping defective mutant). All positively scoring mutants were independently re-constructed and spotted as serial dilutions onto galactose +/− 5-FOA media in duplicate. We confirmed 112 yeast mutants that were defective for telomere looping and subsequently ranked them qualitatively for growth on 5-FOA (Table 1, Figure S1A). Using the Cytoscape BinGO plugin [32], the statistically over-represented GO (gene ontology) categories were determined for our positive scoring candidates (Figure 1C). The confirmed mutants formed the “positive hit set” whereas the “reference set” consisted of all 4800 genes screened. The analysis used a hypergeometric test and significance was tested at 5% (p<0.05) after applying Benjamini & Hochberg False Discovery Rate (FDR) correction for multiple testing. This protocol revealed histone deacetylation as a significantly enriched GO term in the GO-Cellular Component and the GO-Biological Process ontologies. Moreover, the Rpd3L, Rpd3S and Hda1 KDAC complexes were specifically over-represented (Figure 1C, Table 1). We introduced the looping construct into deletion mutants of all members of the Rpd3L/S and Hda1 complexes (including those that did not score positive in the screen) and determined that all complex members tested were important for wild type-like telomere structure (Figure 1D). Importantly, we replicated the 5-FOA plates onto media lacking uracil to ensure that the FOA resistance observed was not due to inactivation of the URA3 gene (Figure 1D). FOA resistant strains maintain the ability to grow on media lacking uracil due to low basal levels of the URA3 transcript (see Figure S1D). To confirm that there was not an inherent problem of inducing transcription within the subtelomere of these mutants, we generated and introduced construct 4 (Figure 1E) into Rpd3L, Rpd3S and Hda1 mutants where the TATA-less UAS was placed upstream of URA3 and found that (unlike with construct 2) upon galactose induction all mutants were dead on 5-FOA containing media (Figure 1E). Figure 1E demonstrates that in the mutants of the Rpd3L/S and Hda1 complexes, if the UAS in construct 2 were able to loop back to the URA3 promoter it would be able to induce transcription to an extent that would result in cell death on FOA, as is the case with wild type cells. We excluded that telomere length variation may affect the looping read-out in the rpd3Δ and hda1Δ mutants, as no significant changes in telomere length were detectable when comparing the KDAC mutants to isogenic wild type cells (Figure S1B, S1C). Finally, we demonstrated that the plate read-out effects that we have observed with construct 2 on 5-FOA are due to changes in levels of the URA3 transcript (Figure S1D) as has previously been reported [29] and not an unrelated artifact of 5-FOA.
In conclusion, an unbiased genome-wide screen has implicated lysine deacetylation through the Rpd3L, Rpd3S and Hda1 complexes in promoting a structural change (likely a fold-back) at budding yeast telomeres.
To demonstrate that the looping defect we observe is not specific to modified telomere 7L (construct 2), we employed a previously established chromatin immunoprecipitation (ChIP) technique where it has been shown at a natural telomere (telomere 6R) that α-Rap1 antibodies are able to precipitate subtelomeric DNA greater than 2 kb away from the start of the telomeric tract, despite the fact that chromatin is sheared into fragments of 0.5 kb [30]. From this study, it was concluded that the subtelomeric ChIP signal from cross-linked Rap1 extracts was a result of the telomere looping back into the subtelomeric region (Figure 2A, top). We predicted that the subtelomeric signal would be lost in mutants identified in our above-described screen (Figure 2A, bottom). In agreement with previous reports, we could detect cross-linked Rap1 at a position 0.5 kb and to a lesser extent 1 kb away from the subtelomere/telomere transition point in wild type cells, indicative of a telomeric loop-back structure (Figure 2B). Unlike the previous report [30], we did not detect reproducible differences at positions farther than 1.5 kb from the telomere (not shown), which is likely due to the smaller chromatin fragment size used in our ChIP protocol. Strikingly, the Rap1 signal was diminished in the subtelomere in hda1Δ mutants and lost to a greater extent in sin3Δ cells (Rpd3L/S common subunit), consistent with the 5-FOA assay using construct 2 (Figure 2B, Figure 1D). sir4Δ cells were used as a looping defective positive control for the Rap1 ChIP assay (Figure 2B) [29] and indeed displayed the greatest loss of Rap1 signal in the subtelomeric region. Importantly, the bulk of our chromatin was sheared to 0.3 kbp fragments or less, excluding the possibility that our subtelomeric signals come from inefficient sonication (Figure S2A). Furthermore, Rap1 protein levels were not affected in any of the above-mentioned mutant backgrounds (Figure S2B).
In order to rule out the unlikely possibility that Rap1 spreading into the subtelomere may account for a portion of the ChIP signal in the assay described above (Figure 2A), we repeated the ChIP experiments using an epitope-tagged Cdc13-TAP (Tandem Affinity Purification) allele. Cdc13 associates with the 3′ ssDNA telomeric overhang, and therefore is not prone to spread into the subtelomere. Furthermore we reasoned that by using Cdc13-TAP we would be able to more easily reconcile differences at the -1 kb position due to its distal positioning at the 3′end. Consistently we were also able to detect Cdc13-TAP associated with subtelomeric DNA 1000 bp upstream of the telomeric tract (indicative of a fold-back), and the signal was reduced to background levels in the sin3Δ mutant (Figure 2C). The sin3Δ mutation did not result in decreased Cdc13-TAP protein levels, which could have potentially accounted for the reduced ChIP signal (Figure S2C). It is important to note that we did not enrich significant amounts of subtelomeric DNA at the -6 and -500 bp positions following the Cdc13-TAP ChIP (Figure 2C) although Cdc13 has been previously shown by similar methods to localize to subtelomeres [33]. We interpret this to indicate that our sonication was extremely efficient and fragments above 400 bp (the approximate length of the wild type telomere) were extremely rare and did not give a signal significantly above background (untagged control). In order verify this notion we deleted telomerase (EST2) in Cdc13-TAP cells and let telomeres shorten over 25 and 50 generations (Figure S2D and S2E). We predicted that upon telomere shortening we would be able to increasingly detect a signal at the -6 position as telomeres would be shorter than our sheared chromatin fragments (see Figure 2C for visualization). Indeed, we found that as telomere length decreased to under 300 bp (our average chromatin fragment size) we were able to detect a robust Cdc13-TAP ChIP signal at the -6 position at natural telomere 6R (Figure S2F).
In summary we have confirmed that the looping defect observed in hda1Δ and sin3Δ (Rpd3L/S subunit) mutants using construct 2 and 5-FOA as a read-out (Figure 1D) can be recapitulated using an independent method (ChIP) at natural telomere 6R (Figure 2B, 2C).
Among the list of looping defective mutants we were intrigued that RIF1, a regulator of telomere length, was also implicated in promoting a fold-back structure (Table 1). Since Rif2 works in parallel with Rif1 to regulate telomere length we introduced construct 2 into rif2Δ cells and found that like rif1Δ cells, rif2Δ mutants also displayed a looping defect (Figure 3A). Using the Rap1 ChIP assay (Figure 2A) it was also evident that both rif1Δ and rif2Δ mutants had structural defects in terms of folding back into the subtelomere (Figure 3B). As with the above described Rap1 ChIP, we confirmed that Rap1 protein levels were not altered in rif1Δ and rif2Δ cells which may have accounted for observed differences (Figure S3A). From here on we have performed further analysis only with the rif2Δ mutant rather that rif1Δ cells due to the fact that apart from telomere length regulation, Rif1 also plays an important role in telomere capping [10], checkpoint regulation [34]–[35] as well as telomere localization [36], which greatly complicated the interpretation of rif1Δ cells and their genetic interactions. Ongoing studies are directed at better understand the contributions of Rif1 in promoting the telomere fold-back structure.
We constructed double mutants between rif2Δ and mutants of the Rpd3L, Rpd3S and Hda1 complexes harboring construct 2 in order to assess potential genetic interactions between different pathways involved in telomere looping. Whereas the Rpd3S specific mutants (eaf3Δ and rco1Δ) displayed a slight additive growth advantage on 5-FOA when combined with rif2Δ mutants compared to the respective single mutants (Figure 3C, bottom panels) there were no additive effects with rif2Δ mutants and the Rpd3L complex (sap30Δ and rxt2Δ) (Figure 3C, top panel). sin3Δ rif2Δ double mutants were not additive in comparison to the respective single mutants (Figure 3C, middle panel), as Sin3 belongs to both L and S complexes. As expected, rxt2Δ (Rpd3L) and eaf3Δ (Rpd3S) double mutants had a slight additive looping defect in comparison to the single mutants and further deletion of RIF2 did not exacerbate this defect (Figure 3D). The looping defect of hda1Δ and hda2Δ mutants was also additive with rif2Δ mutants (Figure 3E). The results of this genetic epistasis analysis suggest that Rif2 and Rpd3L may function together in a common pathway to promote a telomere fold-back.
To mechanistically understand the genetic relationships between the Rpd3L complex and Rif2, we performed ChIP experiments to determine if the rif2Δ mutation had an effect on Rpd3L (Rxt2-TAP) localization at telomeres. Subtelomeric DNA was enriched above non-tagged wild type control cells (background) with an epitope-tagged Rxt2 allele both close to (-6 bp) (Figure 3F) and up to 2000 base pairs away from the telomeric tracts (Figure 3F) as previously described [18]. This enrichment was decreased to near background levels in a rif2Δ mutant (Figure 3F). The loss of ChIP signal is not due to altered expression levels of Rxt2-TAP in rif2Δ mutants as confirmed by western blot analysis (Figure S3B). Unlike Rpd3L, the ability to cross-link the Rpd3S complex (Rco1-TAP) to subtelomeric regions was not altered in rif2Δ cells (Figure 3G and Figure S3C). Together these data suggest that the Rif2 promotes a structural alteration at telomeres through the recruitment of the Rpd3L KDAC complex. Moreover, the Hda1 KDAC as well as the Rpd3S complex promote the same fold-back structure, but independent of the Rif2/Rpd3L pathway.
To better understand the function of the loop structure and whether or not it may have a protective role at the telomere we impaired looping (deletion of SIN3) in various genetic backgrounds where telomere function was compromised. In cdc13-1 sin3Δ double mutants we observed a temperature-dependent synthetic lethality in the double mutant, compared to the respective single mutants (Figure 4A), indicating that partially uncapped telomeres (cdc13-1) may require a Rpd3-mediated structure for viability. The negative genetic interaction was suppressed by the further deletion of EXO1, the nuclease responsible for the majority of telomere resection at dysfunctional telomeres (Figure 4A). To ensure that this interaction was a direct consequence of telomere dysfunction, we assayed the accumulation of telomeric ssDNA following the shift of nocodazole-arrested cdc13-1 and cdc13-1 sin3Δ cells from 23°C to the semi-permissive temperature of 26°C (Figure 4B). In agreement with the negative genetic interaction (Figure 4A), we observed an Exo1 dependent increase in telomeric ssDNA in the double mutant above that seen in the cdc13-1 single mutant (Figure 4B). Importantly, we did not detect an increase in ssDNA at telomeres in sin3Δ single mutants when compared to isogenic wild type control cells (Figure S4A).
To determine if Rpd3 dependent telomere structure may have an influence on the rate of cellular senescence, we compared senescence onset in both est2Δ rad52Δ (where HR and telomerase-mediated telomere elongation are impaired) and est2Δ rad52Δ sin3Δ mutants. The deletion of SIN3 resulted in a dramatic increase in the rate of cellular senescence in the absence of telomerase and homologous recombination (Figure 4C). The accelerated loss of viability associated with the sin3Δ mutation is specifically related to the absence of telomerase as rad52Δ cells maintain viability to a similar extent as rad52Δ sin3Δ cells (Figure 4C). To better understand the cause of premature senescence in sin3Δ mutants, genomic DNA was prepared from the senescence curves (Figure 4C) and both single stranded telomeric DNA accumulation and telomere length were analyzed. As is the case when combined with the cdc13-1 mutant, we found that sin3Δ est2Δ rad52Δ mutants had increased levels of ssDNA at telomeres compared to isogenic est2Δ rad52Δ cells (Figure 4D). The telomere shortening rate is unaffected between sin3Δ est2Δ rad52Δ and est2Δ rad52Δ strains (Figure S4B, S4C), however there is evidence of early rapid telomere loss events at some telomeres in the triple mutant compared to the isogenic double (Figure S4B).
Taken together these results imply that the Rpd3 lysine deacetylase is essential to prevent excessive nuclease-mediated resection specifically at uncapped telomeres. In the absence of a telomere lengthening mechanism this resection may lead to excessive telomere shortening. We propose that this protective function may involve the formation of a fold-back structure at the chromosome ends.
We have demonstrated that multiple biological processes influence the ability of telomeres to form a higher-order fold-back structure. High-throughput screening coupled to stringent bioinformatic analysis, has revealed that class I (Rpd3) and class II (Hda1) KDAC activities were among the most significantly enriched biological processes required to promote the formation/maintenance of a telomere loop. In addition, the Rap1 binding proteins, Rif1 and Rif2, which localize directly to telomeres, were implicated in telomere fold-back establishment. Through genetic epistasis analysis, we found that rif2Δ and mutants of the Rpd3L complex did not have additive structural defects at telomeres, suggesting that they may function together in a single pathway. This genetic interaction was confirmed by demonstrating that Rpd3L was no longer able to localize to telomeres in rif2Δ cells. Furthermore, we have shown that the telomeres in mutants with looping defects are more susceptible to uncapping, nucleolytic degradation, telomere loss and promote accelerated rates of cellular senescence in the absence of telomere maintenance. A recent study has reported that the Rpd3 complex is required to prevent chromosome end fusions in Drosophila melanogaster [37]. Together, these data suggest that the regulation of such telomere fold-back structures may be conserved, and furthermore indicate that multiple cellular processes, apart from those that directly impinge on DNA metabolism, may have effects on telomere structure/function.
Although we have focused on the telomere dysfunction phenotypes associated rpd3 mutants, it is of interest that many of the other mutants recovered in our telomere structure screen have been previously identified to have negative synthetic interactions with cdc13-1 (rrd1Δ, swr1Δ, arp6Δ, spe1Δ, spe3Δ, oca1Δ, oca2Δ, oca5Δ, elp4Δ, pep8Δ, rif1Δ, yme1Δ, htd2Δ, ski3Δ, rim101Δ, ede1Δ) [10], [38]–[39] as well as an increased rate of replicative senescence (hda1Δ, sin3Δ, sec28Δ, med1Δ, asf1Δ, rif1Δ, rif2Δ, arp6Δ, elp4Δ) [40]–[41] (see Table 1 for a complete overview of overlaps between our screen and other selected telomere function screens). Moreover, the Sir2/3/4 complex, which is required to form a telomere fold-back structure [29] also prevents premature senescence [42]. This overlap between our screen and previously published data suggests that telomere structure may make significant contributions towards preserving telomere integrity when telomere function is compromised (Figure 5). Consistently, sin3Δ single mutants do not exhibit increased ssDNA accumulation at telomeres nor do they have any changes in telomere length in comparison to isogenic wild type strains. This would suggest that in the absence of a telomere fold-back, the CST complex and other capping factors are sufficient to maintain a protected state (Figure 5). However, when CST function is compromised (e.g. cdc13-1) in combination with an inability to fold-back, resection becomes accelerated (Figure 4B, 5). In terms of telomere-induced cellular senescence in the absence of telomere maintenance, the increased resection in non-looped mutants would also lead to increased telomere shortening (Figure 5). It will be of interest to perform an extensive genetic epistasis of all mutants isolated in the loop screen in order to determine if the negative interactions with cdc13-1 are epistatic (i.e. due to a fold-back defect).
There was also a large overlap between mutants found in our screen and mutants that have been implicated in both the positive and negative regulation of telomere length (vps28Δ, trk1Δ, ctk1Δ, mrt4Δ, rif1Δ, sur4Δ, nut1Δ, siw14Δ, leo1Δ, hit1Δ, pcp1Δ, pdx3Δ) [43]–[44]. This would suggest that either telomere looping has a direct effect on telomere length regulation, or conversely, may indicate that telomere length changes impinge on the ability to form a fold-back. Our results suggest that telomere looping does not affect telomere length homeostasis directly, as many of the mutants recovered in our screen, even those with “strong” looping defects, have wild type telomere length. On the other hand, telomere length changes could indeed have a drastic effect on telomere structure in terms of the chromatin alterations that occur with respect to length changes. Telomere shortening, for example, results in de-silencing in the subtelomeric region [45] due to the decreased capacity of shortened telomeres to recruit the Sir2/3/4 histone deacetylase complex, which is required for loop formation in yeast [29]. Long telomeres, in contrast, promote a hyper-silenced state in the subtelomere [45], much like what occurs in rif1Δ and rif2Δ mutants, where, in the case of the latter mutant, Rpd3L fails to localize to telomeres. Indeed, Rpd3 mutants (L and S) are hypersilenced in the subtelomeric zone. One possibility would be that long telomeres (as seen in rif2Δ) mutants fail to properly localize Rpd3L, which leads to a subsequent fold-back defect. Although our screen implicates proper length regulation as a key regulator of telomere looping, they remain correlative and require further investigation in order to draw concrete conclusions.
Whereas both the Rpd3S and Hda1 complexes were additive with rif2Δ mutants in terms of a looping defect, the Rpd3L complex was epistatic. This relationship was confirmed mechanistically as we noticed that Rpd3L is not able to properly localize to telomeres in rif2Δ cells. These epistasis analyses revealed that multiple KDACs contribute to telomere looping (Figure 5). Rpd3L, Rpd3S and Hda1 all promote telomere looping in parallel pathways whereas the relationship between the Sir2/3/4 complex and the other KDACs in terms of telomere structure remains enigmatic. The KDACs are best known for their deacetylation of histones and they are known to contribute significantly to silencing at subtelomeric loci [18]. Consistent with a connection between chromatin modification and the telomere fold-back structure, we also found that many members of the Swr1 chromatin remodeling complex as well as the histone chaperone Asf1 (Table 1) were important for telomere folding. A future challenge will be to determine the targets of the KDACs. Although subtelomeric histones are prime candidates, telomeric proteins themselves may be targets. Furthermore, it will be important to characterize how the other mutants revealed in the screen contribute to telomere looping and to understand if these mutants are epistatic with the KDACs. Indeed, it has been difficult to understand why mutations that affect such diverse biological pathways may have synthetic growth defects with cdc13-1 or in some cases, senesce rapidly [38]–[39], [41]. Since many signaling pathways activate effectors via activation/repression of target genes through chromatin remodeling and/or histone acetylation/deacetylation, we propose that activation or repression of these pathways may influence the ability of the KDACS (Rpd3/Hda1 and Sir2) to act at telomeres and in turn directly or indirectly influence telomere structure.
This work has uncovered multiple regulators of the telomere fold-back structure, including lysine deactylation and chromatin remodeling. The results of our screen correlate well with screens that have been performed to elucidate genes implicated in telomere function and cellular senescence suggesting that the fold-back structure may be important for chromosome end protection. Previous models have speculated that the fold-back in yeast may be important to establish silent chromatin within the telomeric/subtelomeric loci [46]–[47]. As an alternative, it was also suggested that silent telomeric chromatin may be required to establish a particular architecture that contributes to chromosome end protection [47]. Our results indicate that silent chromatin can be established in the absence of a telomere fold-back since many of the mutants recovered in our screen are not compromised for silencing (Table 1) or even have slightly enhanced silencing (e.g. rpd3 mutants). Sir2/3/4-mediated silencing as well as other chromatin modifications are however important to establish a telomere loop, which likely promotes end protection. Interestingly, Rpd3 dependent histone deacetylation has been shown to prevent Sir2/3/4 protein spreading towards the centromere [17], which may potentially deplete SIR protein levels immediately adjacent to the telomere, raising the possibility that SIR2/3/4 disruption and RPD3L/S disruption may be one in the same in terms of a fold-back defect. Further characterization of the yeast fold-back structure and the relationship to silent chromatin will be essential in order to clarify these issues. In summary, by understanding how different biological processes impinge on chromosome end structure, we increase the possibilities to manipulate telomere function, both positively and negatively, which may have important implications for diseases that stem from telomere dysfunction.
Standard yeast media and growth conditions were used [48]. Yeast strains used in this study are listed in the Table S1.
For spotting assays, yeast cells were incubated overnight at appropriate temperature in YPD. Cells were diluted to OD600 0.5 and spotted in ten-fold dilutions onto 2% raffinose, 1% galactose plates either with (+FOA) or without (−FOA) 5-FOA. Cells were incubated for 3 days at proper temperature, imaged and then replica plated on SD-URA plates for 2 days at the same temperature before imaging.
Telomere PCR was performed using 100 ng genomic DNA diluted in 1× NEB4 buffer and water. Samples were denatured for 10 min at 96°C and cooled to 4°C. Tailing mix (4 U/µl terminal transferase (NEB), 1× NEB4 buffer, 1 mM CTPs) was added to a final concentration of 10%. Tailing reaction was performed as the follows: 37°C 30 min, 65°C 10 min, 96°C 5 min, arrest at 65°C. 3× volume of preheated PCR-MIX (1 µM oligo dG reverse primer, 1 µM telomere specific forward primer either 1L, 6R, 7L or Y′, 0.267 mM dNTPs, 0.083 U/µl Phusion polymerase (NEB), PCR buffer (89.11 mM Tris-HCl pH 8.8, 21.28 mM (NH4)2SO4, 6.65% glycerol, 0.0133% Tween-20) was added and PCR reaction was performed using: 95°C 3 min, 45 cycles: (95°C 30 s, 68°C 15 s, 72°C 20 s), 68°C 5 min, hold on 12°C.
Samples were mixed with DNA loading buffer and separated on a 1.8% agarose gel for 30 min at 100 V. Bands were detected using LAS-4000 (Fujifilm) and quantified using Multi Gauge Software (Fujifilm). A complete lis of oligonucleotides used in this study can be found in Table S2.
Spore-colonies of dissected heterozygous diploids were suspended in water and diluted in 5 ml YPD medium to a final concentration of OD600 0.01. Cells were incubated for 24 h at 30°C and absorption at 600 nm was measured. Cultures were re-diluted to OD600 0.01 in 5 ml YPD and inoculated for further 24 h at 30°C. Each day cell samples were harvested and genomic DNA was prepared for telomere length analysis (Quiagen genomic DNA prep. Kit). Population doublings (PD) were calculated as log2(OD60024 h/0.01). All PD values refer to PD after the spore colony had been harvested from the dissection plate (about 25 generations). Graphs were made in Prism5 (GraphPad).
Synthetic Genetic Array (SGA) methodology was used (strains R1459 and R1460) to obtain haploid gene deletion mutants containing either construct 1A or construct 2 (Tong and Boone, 2006). Cells were then replica-pinned onto media containing galactose, with and without 5-FOA. Growth on 5-FOA was compared between construct 2-containing deletion mutants and construct 1-containing mutants (comparison 1). Growth of construct 2-containing mutants was also compared with and without the presence of 5-FOA (comparison 2). Construct 2-containing mutants that grew better by either comparison 1 or 2 were selected for validation. Validation was carried out by manually crossing and dissecting tetrads from independent starter strains followed by duplicate spot assays onto media with and without 5-FOA.
Yeast cells were grown over night at 30°C and diluted to OD600 0.2. They were grown until exp. Phase (OD600 0.6–1.0), crosslinked for 8 min (20 min for Rxt2-TAP-ChIP, 10 min for Cdc13-TAP-Chip) with formaldehyde (final conc. 1.2%) and quenched with glycine (360 mM final). After adjusting the volume to the same OD all samples were washed two times with 1× PBS, resuspended in FA Lysis Buffer (50 mM HEPES-KOH pH 7.5, 140 mM NaCl, 1 mM EDTA pH 8, 1% Triton X-100, 0.1% Sodium deoxycholate) and lysed with Matrix C tubes via FastPrep (6.5 M/s, 2°—30 sec with 1 min break). Cell extracts were recovered, centrifuged and the soluble potion of the lysate was discarded. Pellets were resuspended in FA buffer +SDS (2% final) and split up for sonication. Chromatin was sheared 30 sec on/off for 15 min. Supernantant (ChIP extract) was diluted to 1 mg/ml protein concentration in FA buffer and used for immunoprecipitation (IP).
Pre incubated protein G sepharose beads (washed with 1×PBS, FA Buffer and pre-incubated with 5% BSA for 1 h at 4°C) were added to the 1 mg/ml solution to perform an addition precleaning step before the IP (1 h at 4°C). After precleaning anti-Rap1 antibody (Santa Cruz) was added to the solution (1∶100) and incubated with fresh beads over night at 4°C, rotating.
For Tap-ChIPs (Rxt2 and Cdc13) IgG-Sepharose Beads (washed with 1× PBS and FA-buffer) were added to the 1 mg/ml solution and IP was incubated over night at 4°C.
Sonication efficiency was tested via cleaning 100 µl of the ChIP extract and performing agarose gelelectrophoresis. IP was washed with FA-Lysis buffer, FALysis buffer 500 (FA buffer with 500 mM NaCl), Buffer3 (10 mM Tris-HCl pH 8 1 mM EDTA pH 8, 250 mM LiCl, 1%NP-40, 1% Sodium deoxycholate), and TE (pH 8). For elution buffer B (50 mM Tris-HCl ph 7.5, 1% SDS, 10 mM EDTA pH 8) was added and IP was incubated at 65°C for 8 mins. For reverse-crosslinking proteinase K was added to the IP and INPUT control (ChIP extract, 1 mg/ml solution without IP) and incubated at 65°C, rotating overnight. Samples were cleaned with Quiagen “QIAquick PCR Purification Kit” and qPCR analysis was performed using Roche standard PCR protocol for Sybr-Green detection with 55°C annealing temperature, all oligonucleotides used are listen in Table S2. Measured ct values were corrected to INPUT and normalized to the actin signal using the following formulas.
Rap1 ChIP:Rxt2-TAP ChIP:Cdc13-TAP ChIP:
Cells were grown overnight at 23°C in 10 ml YPD. Saturated cultures were diluted to OD600 0.2 in 150 ml YPD and incubated at 23°C until they reach log phase (0.6–0.8). Nocodazol (20 µg/ml final) was added and cells were incubated for a further 3 h at 23°C, shaking. Cells were checked under the microscope until >90% were largebudded. “Pre” samples were harvested and cells were subsequently shifted to 26°C.
Additional samples were collected for all time points (30 min, 60 min, 90 min and 120 min) after the shift. For ssDNA analysis, dot blotting was performed. DNA was extracted using genomic DNA Kit (Quiagen). Isolated DNA was either denatured using 0.2 M NaOH and 65°C for 15 min or kept on ice for native conditions.
For blotting, 4 µg DNA (native) or 0.5 µg (denatured) were suspended in 200 µl 2×SSC and loaded to the dot blot apparatus using nylon membrane (GE Healthcare Amersham H-bond). After crosslinking (UV Stratalinker 2400, Stratagene) DIG labeling (DIG labeled probe oBL207) and detection was performed as described by the product guidelines (Roche DIG oligonucleotide 3′labeling KIT).
Cells were grown overnight at 30°C in 5 ml SD medium containing 2% Raffinose (S-Raf). Saturated cultures were diluted to OD600 0.2 in 8 ml S-Raf and incubated at 30°C until they reach log phase (0.6–0.8). Cells were split and 2% galactose or 2% glucose (final) were added. Cells were incubated for 2 1/2 h at 30°C, shaking. Cells were centrifuged down and RNA was extracted and Northern Blotting was performed as described previously [42]. URA3 and actin were detected using DiG labeled PCR products (Roche, “DiG High Prime” labeling) gained from a PCR reaction with oBL17, oBL18 for URA3 and oBL292, oBL293 for actin (1. 98°C 30 sec, 2. 98°C 10 sec, 3. 60°C 30 sec, 4. 72°C 1 min, 5. 72°C 5 min, 12°C forever, repeating steps 2–4 for 33 cycles). Quantification was performed using Multi Gauge software (Fujiifilm) and signal was displayed as URA3 over actin.
3 ml of culture (log phase) was centrifuged at 13000 rpm for 2 min. Pellets were resuspended with 150 µl solution 1 (0,97 M β-mercaptoethanol, 1,8 M NaOH) and incubated on ice for 10 min. 150 µl, 50% TCA was added and cells were incubated 10 min on ice, centrifuged at 13000 rpm for 2 min at 4°C and the pellet was resuspended with 1 ml acetone. Solution was centrifuged at 13000 rpm for 2 min at 4°C and the pellet was resuspended in 140 µl UREA buffer (120 mM Tris-HCL pH 6,8, 5% Glycerol final, 8 M Urea final, 143 M 2-mercaptoethanol final, 8% SDS final, a little bit of bromphenol blue indicator). Protein extract was incubated 5 min at 95°C, centrifuged and loaded on a pre-cast gradient Gel (BioRad).
|
10.1371/journal.pntd.0002781 | Characterization of the Early Inflammatory Infiltrate at the Feeding Site of Infected Sand Flies in Mice Protected from Vector-Transmitted Leishmania major by Exposure to Uninfected Bites | Mice exposed to sand fly saliva are protected against vector-transmitted Leishmania major. Although protection has been related to IFN- γ producing T cells, the early inflammatory response orchestrating this outcome has not been defined.
Mice exposed to uninfected P. duboscqi bites and naïve mice were challenged with L. major-infected flies to characterize their early immune response at the bite site. Mostly, chemokine and cytokine transcript expression post-infected bites was amplified in exposed compared to naïve mice. In exposed mice, induced chemokines were mostly involved in leukocyte recruitment and T cell and NK cell activation; IL-4 was expressed at 6 h followed by IFN-γ and iNOS2 as well as IL-5 and IL-10 expression. In naïve animals, the transcript expression following Leishmania-infected sand fly bites was suppressed. Expression profiles translated to an earlier and significantly larger recruitment of leukocytes including neutrophils, macrophages, Gr+ monocytes, NK cells and CD4+ T cells to the bite site of exposed compared to naïve mice post-infected bites. Additionally, up to 48 hours post-infected bites the number of IFN-γ-producing CD4+T cells and NK cells arriving at the bite site was significantly higher in exposed compared to naïve mice. Thereafter, NK cells become cytolytic and persist at the bite site up to a week post-bite.
The quiet environment induced by a Leishmania-infected sand fly bite in naïve mice was significantly altered in animals previously exposed to saliva of uninfected flies. We propose that the enhanced recruitment of Gr+ monocytes, NK cells and CD4 Th1 cells observed at the bite site of exposed mice creates an inhospitable environment that counters the establishment of L. major infection.
| Sand flies transmit Leishmania parasites during bloodfeeding. Salivary molecules are deposited alongside parasites and can reshape the host's immune response to infection. Exposure to uninfected sand fly bites or immunization with salivary molecules protects the host against Leishmania infection. Here we show that mice exposed to bites of uninfected Phlebotomus duboscqi sand flies are protected against P. duboscqi-transmitted L. major and characterize the formerly unknown early cellular infiltrate at the bite site following L.major vector-transmission. The kinetics and nature of the inflammatory response at the bite site of exposed mice were notably different from those of naïve mice showing an amplified expression of cytokines and chemokines after parasite transmission. The transcripts reflected a faster and more robust infiltrate of immune cells to the bite site of exposed mice composed of neutrophils, macrophages, monocytes, NK cells and CD4+ T cells. In addition, there was an increased influx of activated IFN-γ producing CD4+ T cells and Granzyme B-producing mature NK cells in exposed animals. These findings suggest that the observed robust and persistent proinflammatory response in exposed animals restrict parasite multiplication.
| Early studies demonstrated that mice exposed to saliva of the vector sand fly Phlebotomus papatasi develop immunity to its salivary proteins that confers powerful protection against Leishmania major [1]–[3]. These initial observations were validated by the identification of distinct salivary proteins from various vector species that conferred protection from Leishmania infection in immunized animals [4]–[9]. Further characterization of the adaptive immune response to saliva demonstrated that it is cell-mediated and dependent on IFN-γ-producing Th1 CD4 cells [5], [6], [10], [11].
In naïve animals, Leishmania parasites as well as sand fly saliva have each been associated with suppression of the initial proinflammatory immune response thereby promoting parasite survival [7], [11]–[18]. Co-injection of saliva or its components together with Leishmania parasites was shown to exacerbate cutaneous leishmaniasis (CL) infections producing larger lesions and a higher parasite burden. Enhancement of Leishmania infections by saliva was attributed to the immunomodulatory properties of salivary proteins that act early during infection to promote downregulation of dendritic cells and macrophage function and the production of anti-inflammatory cytokines that favor parasite establishment [19]–[23]. Leishmania parasites also orchestrate their own suppression and deactivation of the function of immune cells including macrophages [24]–[30] and NK cells [31], [32] recruited to the site of infection promoting their survival.
Additionally, in the context of vector-transmission, the parasite-derived promastigote secretory gel (PSG) was also shown to modulate the bite site to the advantage of the parasites. PSG shed by the Leishmania in the sand fly midgut is regurgitated by the fly during an infective bite and was shown to exacerbate Leishmania infection by acting on host macrophages to promote their alternate activation and parasite survival [33].
In any natural transmission event, an infected sand fly bites the host co-injecting saliva, PSG and parasites into the wound. We postulate that during vector-transmission of Leishmania to animals previously exposed to saliva or immunized with a protective salivary molecule, a rapid influx of immune cells is recruited to the bite site by the deposited saliva creating a proinflammatory environment that overcomes the suppressive nature of saliva, PSG and Leishmania. Such an inflamed bite site in exposed animals is likely to adversely affect the co-deposited parasites controlling their growth at an early stage and permitting a slower, damage-limited, development of Leishmania-specific immunity and conferring long-term protection from leishmaniasis [5], [6], [10], [11]. Here, we investigate the kinetics and phenotype of the cellular infiltrate arriving at the site of an infective bite in saliva-exposed mice compared to naïve animals to elucidate the significance of the early inflammatory response in saliva-mediated protection from leishmaniasis.
The Animal Care and Use Committee at the National Institute of Allergy and Infectious Diseases adheres to the U. S. Government Principles for the Utilization and Care of Vertebrate Animals Used in Testing, Research, and Training and maintains animals in accordance with the PHS Policy on Humane Care and Use of Laboratory Animals, the
Guide for the Care and Use of Laboratory Animals, and the Animal Welfare Act and Animal Welfare Regulations user guidelines and has reviewed and approved all experimental procedures involving animal use for this study under animal protocol LMVR 4E.
C57BL/6 female mice, 4–6 weeks old (Charles River Laboratories) were maintained under pathogen free conditions.
Leishmania major clone V1 (MHOM/IL/80/Friedlin) amastigotes were used for sand fly infection. Amastigotes were washed with PBS, counted and mixed with mouse blood. Before mixing, mice serum was heated for one hour at 56°C. Blood containing 3×106 L. major/ml were used to artificially feed sand flies using a glass chamber covered with a chicken skin membrane as previously described [2]. Blood-fed sand flies were separated 24 hours after infection and kept contained in secure paper cups in an incubator at 25°C and 75% relative humidity. Flies were offered 30% sucrose in soaked cotton balls. Before transmission sand flies were dissected to examine the quality of the infection and quantity of metacyclic promastigotes as previously described [34]. Sand flies were used for transmission experiments on days 13 to 14 after infection.
Phlebotomus duboscqi (Mali strain) sand flies were reared at the Laboratory of Malaria and Vector Research, NIAID. Exposure of mice to uninfected flies was carried out on a weekly basis for a total of three exposures. P. duboscqi sand flies used for these experiments were emergent females (5–7 day old) left without sugar overnight. Mice were anesthetized intraperitoneally with ketamine (Phoenix Pharmaceuticals, St. Joseph, MO) according to the weight of the animal. Ten sand flies were placed in plastic vials covered at one end with a 0.25 mm nylon mesh and left to feed on the right ear in the dark. The ear of anesthetized mice was pressed flat against the mesh surface of the vial containing the flies using custom-designed clamps. Sand flies were allowed to feed for up to 30 minutes and were then examined for blood to assess exposure success. For challenge, 10 L. major-infected sand flies were placed in individual vials and applied to the left ear of mice two weeks after the last exposure to uninfected flies as described above. Infected flies were applied to the contralateral ear of exposed mice to demonstrate that exposure to uninfected flies generates systemic immunity. Infected sand flies were allowed to feed for two hours in the dark after which they were examined for blood to assess the success of transmission
For measurement of Leishmania lesions the largest diameter was recorded on a weekly basis using a Digimatic caliper (Mitutoyo). Parasite quantification was performed using a limiting dilution assay as previously described [1], [35]. Briefly, total ear tissue homogenates were serially diluted (1∶5) in 96-well flat bottom microtiter plates containing 50 µl biphasic medium prepared using NNN medium with 10% of defibrinated rabbit blood overlaid with 200 µl Schneider's (Gibco, NY) supplemented with 20% heat inactivated fetal bovine serum, 2 mM L-glutamine, 100 U/ml penicillin and 100 µl/ml streptomycin. The number of viable Leishmania in each ear tissue was determined from the highest dilution at which Leishmania promastigotes could be grown after 7 days of incubation at 26°C.
The expression profile of cytokines, chemokines, and related inflammatory genes was generated using the mouse inflammatory cytokines and receptor Oligo GEArray (OMM-011; Superarray). This array contains 112 genes representing cytokines, receptors and housekeeping genes. Six and 12 hours after exposure to uninfected sand flies or after challenge with L. major-infected sand flies, total RNA was isolated from the left ear of each mouse using the RNeasy Mini Kit (Qiagen) according to the manufacturer's instructions. RNA (6 µg) from a pool of five ears was amplified and labeled with biotin 16-UTP (Roche Diagnostics) using the SuperArray TrueLabeling-RT Enzyme kit (Superarray). The resulting biotinylated cRNA was hybridized overnight to the Oligo GEArray membrane. After washing and blocking the array membranes, alkaline phosphatase-conjugated streptavidin was added to the membrane followed by the CDP-Star substrate. A chemiluminescent signal was acquired using the Image Station 2000 MM (Kodak). The data was analyzed using the GEArray Expression Analysis Suite (Superarray). Analysis parameters were set to local background correction and normalized to a set of housekeeping genes included in each membrane. Results were expressed as fold increase in the intensity of the captured signal over the control group (ear tissue not exposed to bites). Only genes showing at least a four-fold or higher change in expression in at least three out of four independent experiments were considered.
The genes that showed a four-fold or higher change in expression over the control group (ear tissue not exposed to bites) using the GEArray were validated by Real time PCR. Total RNA from individual mice ears was used for the synthesis of cDNA (Superscript III, Invitrogen) following the manufacturer's instructions. The cDNA was amplified with the 480 Master SYBR Green I mix (Roche Diagnostics) and gene specific primer sets for IFN-γ, IL-4, CXCL13 and CCL25 using the LightCycler 480 (Roche Diagnostics). A standard curve for each set of primers was generated as recommended by the manufacturer. The expression levels of the genes of interest were normalized to endogenous 18S RNA levels. The results are expressed in fold change over gene expression in the control group.
Bone Marrow Dendritic Cells (BMDCs) were obtained from the femurs of mice and cultured in RPMI medium enriched with 10% heat-inactivated fetal bovine serum (HyClone), 100 U/ml penicillin, 100 µg/ml streptomycin, 2 mM L-glutamine, 40 mM Hepes, 5×10−5 M 2-mercaptoethanol, plus 20 ng/ml GM-CSF (Prepotech). The medium was replaced by fresh complete RPMI containing 10 ng/ml of GM-CSF on day 3.
Cells were recovered from the ear dermis as described previously (Belkaid et al, 1998; Peters et al, 2008). For characterization of leukocyte populations the cells were washed, incubated for 30 min at 4°C with anti-CD16/CD32 (BD Fc block, 2.4G2; BD Pharmingen), and stained ex vivo with a combination of surface markers including PerCP-labeled anti-CD4 (RM4-5) or PE -labeled anti-CD4 (GK1.5), FITC-labeled anti-Ly6G (RB6-8C5), APC-labeled anti-TCR-β (H57-597), PerCP-Cy5.5 labeled anti-NK1.1 (PK136), PE- or PerCP-labeled anti-CD11b (M1/70), APC-labeled anti-F4/80 (BM8), for 30 minutes at 4°C. For intracellular stains, cells (2×106) were cultured with 2.5×105 Bone Marrow Dendritic Cells (BMDCs) in one mL of RPMI 1640 containing 10% FBS, L-glutamine and penincilin/streptomycin in flat-bottom 48-well plates at 37°C and 5% CO2 for 18 h with the addition of 20 ng of PMA and 500 ng of ionomycin and 1 ul of Brefeldin A (BD Golgi Plug; BD Pharmingen) during the last 4 h of culture. The cells were then washed with PBS and blocked with anti-CD16/CD32 (BD Fc block, 2.4G2; BD Pharmingen) for 30 minutes at 4°C. Cells were stained with a combination of PerCP-labeled anti-CD4 (RM4-5) or PE-labeled anti-CD4 (GK1.5), APC-labeled anti-TCR-β (H57-597), PerCP-Cy5.5 labeled anti-NK1.1 (PK136), FITC or PerCP-Cy5.5-labeled anti-CD11b (M1/70) and PE-labeled anti-CD27 (LG.3A10) or 30 minutes at 4°C, washed twice, fixed and permeabilized with Cytofix/Cytoperm Plus (BD Pharmingen) and stained with FITC-labeled anti-IFN-γ (XMG 1.2) and/or APC-labeled anti-Granzyme B (GRB05, Invitrogen). A minimum of 100,000 cells was acquired using a FACSCalibur flow cytometer (BD Biosciences). The data were analyzed using the Flow Jo software.
Graphs and statistical significance were prepared and analyzed using GraphPad Prism Software 5.0 (GraphPad, San Diego, CA). Data from parasite numbers were log transformed before conducting statistical tests. The unpaired t test with Welch's correction or one-way analysis of variance followed by the Tukey-Kramer post-test was used to evaluate statistical significance among groups. A p value<0.05 was considered statistically significant.
Mice were exposed to bites of uninfected P. duboscqi sand flies three times at two-week intervals and were challenged alongside naïve mice with ten L. major-infected sand flies two weeks after the final exposure. A mean parasite load of 3.8×104 and a mean percent metacyclics of 89% are representative of the infection status of sand flies used in transmission (Fig. 1A). Saliva-exposed mice controlled the infection while naïve mice displayed an increasing lesion size that peaked four weeks post-infection (Fig. 1B). Additionally, there was a three log reduction (p = 0.016) in the number of parasites recovered from the ear of exposed compared to naïve mice (Fig. 1C).
The expression of inflammatory cytokines in the ear tissue of saliva-exposed and naïve mice was determined at 6 h and 12 h after exposure to uninfected sand flies or after challenge with L. major-infected sand flies using the Oligo GEArray Mouse Inflammatory Cytokines and Receptors Microarray that targets key genes involved in inflammation (Fig. 2A). The genes with a reproducible four-fold or higher change in signal intensity compared to unbitten ear tissue are represented in a heat map (Fig. 2B, 2C). Transcript expression was distinct in naïve compared to exposed mice following uninfected bites. Following infected vector bites, the expression profile of inflammatory genes in sand fly-exposed mice was amplified and displayed a dramatic difference compared to naïve mice (Fig. 2B, 2C). Interestingly, the presence of Leishmania parasites augmented gene expression in exposed mice but had a suppressive effect on naive animals (Fig. 2B, 2C). In exposed mice, the induced chemokine receptors and ligands were mostly involved in leukocyte recruitment and T and NK cell activation. Additionally, some genes related to allergic responses were also induced. Certain chemokines were induced by both uninfected and infected sand fly bites including CCL17, CXCL11, CXCL13, CXCR3, CCR7 and CCR8 at 6 h, and CXCL11, CCL2 and CXCL14 at 12 h post-transmission. Of those, CXCR3 and CXCL13 expression were the most prominent. CXCR3 is preferentially expressed on activated Th1 and NK cells and its ligand interactions promote polarization towards a Th1 effector immune response while CXCL13 selectively attracts B cells. Nevertheless, the presence of Leishmania parasites modulated the inflammatory response in exposed mice, particularly at 12 h after transmission. The inflammatory response of sand fly-exposed mice to infected bites intensified with a strong expression of CCL9, CCL19 and CCL25, chemokines known to recruit lymphocytes, monocytes and dendritic cells (Fig. 2B). CCL5, a chemokine that recruits leukocytes and can activate NK cells in the presence of IFN-γ or IL-2 and CCR3, a receptor highly expressed in eosinophils and basophils but is also detected in T cells, were also highly induced (Fig. 2B). A similar pattern was observed for cytokine expression. Compared to uninfected bites, there was a generalized suppressive effect on cytokine expression in naïve animals following infected bites while exposed mice showed an augmented response (Fig. 2C). Sand fly-exposed animals expressed IL-4 at 6 h after transmission and a more pronounced IFN-γ and iNOS2 as well as IL-5 and IL-10 at 12 h (Fig. 2C). In contrast, the naïve group again displayed an overall quiet response with only a moderate expression of TGF-β at 6 h (Fig. 2C). The array expression profiles were validated by real-time PCR for several genes (Fig. 2D, 2E).
Focusing on infected bites, we investigated the kinetics and phenotype of the cells recruited to the bite site at distinct time points up to one week after transmission. The augmented gene expression observed at the bite site of saliva-exposed mice was reflected in the local inflammatory response following transmission. Generally, the number of leukocytes increased with time in both naïve and exposed mice. However, exposed animals sustained a higher number of leukocytes compared to naïve mice at 6 h, 24 h and 48 h post-transmission that was significant (p<0.05) at 6 h and 48 h (Fig. 3A). The peak of cell recruitment in exposed animals was observed 48 h post-transmission (Fig. 3A). By one week post-transmission, the number of leukocytes present at the bite site was similar in both groups. Of note, cellular infiltration at the bite site was considerably faster in the exposed group with a mean of 3.2×106 recruited cells per ear at 6 h post-transmission compared to only 1.5×106 cells per ear in naïve mice (Fig. 3A).
We further characterized the major cell types comprising the leukocyte infiltrate after infected sand fly bites (Fig. 3B). Compared to naïve animals, exposed mice recruited a significantly higher number of neutrophils to the skin 6 h post-transmission (Fig. 3C). By 48 h all cell populations analyzed apart from neutrophils (NK cells, CD4+ T cells, inflammatory Gr1+ monocytes and macrophages) were present in significantly higher numbers (p<0.05) at the bite site of exposed compared to naïve mice (Fig. 3C). At this timepoint, the number of recruited neutrophils was significantly higher in naïve compared to exposed animals. One week post-infected bites, the number of neutrophils had subsided in both naïve and exposed mice (Fig. 3C). In contrast, NK cells and Gr1+ monocytes persisted in the bite site of exposed mice in a pronounced manner that was significantly higher (p<0.05) compared to naive mice a week post-infected bites (Fig. 3C).
We were interested to further explore the activation state of CD4+ T cells and NK cells recruited to the bite site. The percent of IFN-γ-producing CD4+ T cells increased with time for both exposed and naïve mice (Fig. 4A). At 24 h and 48 h post-transmission, the absolute number of IFN-γ-producing CD4+ T cells was significantly higher in exposed (70×103 and 92×103, respectively) compared to naïve animals (51×103 and 16.7×103 respectively) reflecting a more rapid and intense recruitment of these cells in exposed mice (Fig. 4A). One week post-transmission, the naïve mice recruited a comparable number of IFN-γ-producing CD4+ T cells to the bite site (Fig. 4A).
Similarly, 24 h and 48 h post-transmission, the absolute number of IFN-γ-producing NK cells was significantly higher in exposed (34.8×103 and 6.5×103, respectively) compared to naïve mice (zero and 0.7×103, respectively) (Fig. 4B). Surprisingly, despite a significant increase in the number of NK cells recruited to the bite site of exposed mice at the one-week timepoint (Fig. 3B), very few were IFN-γ-producing cells (Fig. 4B). Gating on mature CD11b+CD27−Granzyme+ NK cells (Fig. 4C), 8% (16×103) of NK cells present at the bite site of exposed mice displayed a cytolytic phenotype 48 h post-transmission (Fig. 5). This number increased to 18% (67×103) a week post-transmission (Fig. 5). Of note, compared to naïve mice, the number of cytolytic NK cells observed at the bite site of exposed mice was 50- and 6-fold higher 48 h and one week post-transmission respectively (Fig. 5).
Transmission of Leishmania parasites by an infected sand fly bite is a complex event. During blood feeding, a cocktail of parasites, saliva and PSG, each with distinct immunomodulatory properties, is deposited into the host's skin. These components together with tissue damage caused by probing trigger a potent inflammatory response to an infected bite. The development of adaptive immunity to Leishmania parasites is dependent on the nature of the inflammatory response triggered during the first hours after infection where the local environment dictates what happens downstream of the immune response [36]. We have previously demonstrated that adaptive immunity to sand fly saliva or certain of its immunogenic proteins confers protection against vector-transmitted L. major [2], [10]. Similarly, the initial inflammatory response to L. major following needle injection of the parasites has been well described [37]–[41]. However, the immune response to needle-injected Leishmania parasites differs significantly from that observed following vector-transmission [2], [10], [14], [42]. Additionally, only one study compared the acute immune response to infected vector bites in naïve mice compared to animals protected by exposure to P. papatasi saliva, a sand fly vector of CL [2]. The authors reported a 9- and 15-fold higher IFN-γ- and IL-12-producing cells, respectively, at the bite site in exposed compared to naïve mice six hours post-transmission [2]. In exposed mice, this initial response was followed by a strong delayed–type hypersensitivity response that was associated to the observed protection from CL [2].
Here, we first established that exposure to bites of uninfected P. duboscqi sand flies, another important vector of CL, confers protection against vector-transmitted L. major. Next, we determined the nature of the inflammatory response at the bite site of exposed mice compared to naïve animals following uninfected and Leishmania-infected sand fly bites. Interestingly, the inflammatory response to uninfected and infected bites in naive compared to exposed mice was different. In both cases, the inflammatory response was weaker in naïve mice and was further suppressed by Leishmania parasites. In contrast, the number and level of chemokines and cytokines expressed in exposed mice was augmented as early as 6 h and 12 h following vector-transmission of Leishmania parasites suggesting that immunity to saliva overcomes immune suppression by the parasites that are known to down-modulate the immune response [18], [43]. The hyper-induction of cytokines and chemokines in the exposed mice translates to a more intense leukocyte infiltration compared to the one observed for naïve mice 24–48 h post infected bites. Our findings are distinct from those of Carregaro et al. [11] that noted a decrease in the majority of cell types including neutrophils, macrophages and CD4 T cells recruited to ears of BALB/c mice exposed to Lutzomyia longipalpis saliva compared to naive mice 24 h post-needle challenge with saliva. The observed differences may be attributable to several parameters that varied between the two studies including the mice strains, sand fly vector species and mode of exposures/challenge used. Of note, the magnitude of the cellular response observed in both naïve and exposed mice in the present study is stronger than reported by studies using needle-injected parasites [11], [15] or a lower number of sand flies [14] and emphasizes the potency of vector-transmission.
In the present study, the infiltration of leukocytes to the skin at the bite site was amplified in exposed compared to naïve mice, beginning with neutrophils at 6 h post-transmission and expanding to include macrophages, monocytes, NK cells and CD4 lymphocytes at 24–48 h. Interestingly, the pattern of cell recruitment for neutrophils and NK cells may be indicative of the occurrence of two waves of cell recruitment that possibly contribute to the persistence of cells at the bite site in exposed mice. Recruitment of neutrophils to the bite site was significantly larger in exposed mice at 6 h and was maintained up to 48 h. In contrast, neutrophils were only higher in naïve mice at 48 h after infected bites. The response subsided in both the naïve and exposed groups by one week post-transmission. This differs from recent findings that describe the persistence of neutrophils up to 1–4 weeks in large numbers at the bite site as a feature of vector-transmission to naïve mice where they provide safe passage of parasites into macrophages [14], [42]. Peters et al. [14] demonstrated that neutrophils rescue L. major parasites, promoting a silent entry of parasites into macrophages thereby enhancing their survival. Similarly, neutrophils were quickly recruited to the injection site of L. infantum and contained the majority of intracellular parasites up to 24 h after infection [36]. Taking into consideration that these studies were conducted only in naïve animals, it is clear from our transcript expression data that the suppression exerted by Leishmania parasites in naïve animals is overcome in exposed mice. We hypothesize that due to an intensified proinflammatory environment at the bite site of exposed animals, the activation state and nature of the interaction between macrophages and neutrophils may be altered to the detriment of the parasites contributing instead to disease control. Certainly, the number of recruited macrophages as well as Gr1+ inflammatory monocytes was significantly higher in exposed compared to naïve mice at 48 h post-transmission. A recent study showed that inflammatory monocytes are rapidly recruited to L. major lesions and contribute to killing of the parasites [44]. These inflammatory monocytes express the chemokine receptor CCR2 and are recruited by locally induced CCL2 from activated platelets [44]. Here, compared to naïve mice, exposed animals expressed elevated levels of both CCL2 and CCR2 12 hours post-transmission that correlate to the increased migration of Gr1+ monocytes to the site of bite at later timepoints where may contribute to early parasite control.
The proinflammatory nature of the bite site in exposed mice corresponds to a rapid and sustained recruitment of NK and CD4 T cells to the ears of exposed mice. The phenotype of the CD4+ T cells and NK cells was dramatically different in exposed and naïve groups; both cell types produced significantly higher levels of IFN-γ gamma in exposed compared to naïve mice as early as 24 h (NK cells) and 48 h (CD4+ T cells) post-infected bite. This upholds previous findings where protection from Leishmania observed in animals exposed to uninfected sand fly bites or following immunization with a single salivary molecule consistently correlated with the development of a DTH response and IFN-γ production around 48 h post-infection [10]. NK cells in exposed mice further develop into a CD27−/granzyme+ cytotoxic phenotype at 48 h and persist at the bite site up to a week post-infected bites. Many studies have implicated activated NK cells in protection from Leishmania [32], [45]–[48], however, their significance following vector-transmission of L. major needs to be further elucidated.
Taken together, the results obtained here demonstrate that exposure to P. duboscqi saliva generates a strong inflammatory response following an infected bite that is markedly different from that observed in naïve animals. In contrast to naïve mice, the parasites co-deposited with saliva into the bite site of exposed animals rapidly encounter an influx of proinflammatory leukocytes that persist for a prolonged period after transmission. Of significance, the proinflammatory response of exposed mice parallels findings in a study recently conducted in humans naturally bitten by P. duboscqi [49]. Biopsies taken at the site of a DTH response 48 hours after experimental bites were dominated by lymphocytes, macrophages and high levels IFN-γ indicative of a Th1 response. Importantly, a DTH response to bites was predominant in naturally exposed humans being observed in 75% of individuals aged 1–15 years [49]. The profound modulation of the feeding site in exposed individuals underscores the importance of vector-derived factors in understanding the pathogenesis and control of vector-borne diseases.
This report represents the first description of the kinetics and nature of the local cellular infiltrate in saliva-exposed mice that are protected from disease following vector-transmitted L. major.
|
10.1371/journal.pbio.1002233 | The Actin Nucleator Cobl Is Controlled by Calcium and Calmodulin | Actin nucleation triggers the formation of new actin filaments and has the power to shape cells but requires tight control in order to bring about proper morphologies. The regulation of the members of the novel class of WASP Homology 2 (WH2) domain-based actin nucleators, however, thus far has largely remained elusive. Our study reveals signal cascades and mechanisms regulating Cordon-Bleu (Cobl). Cobl plays some, albeit not fully understood, role in early arborization of neurons and nucleates actin by a mechanism that requires a combination of all three of its actin monomer–binding WH2 domains. Our experiments reveal that Cobl is regulated by Ca2+ and multiple, direct associations of the Ca2+ sensor Calmodulin (CaM). Overexpression analyses and rescue experiments of Cobl loss-of-function phenotypes with Cobl mutants in primary neurons and in tissue slices demonstrated the importance of CaM binding for Cobl’s functions. Cobl-induced dendritic branch initiation was preceded by Ca2+ signals and coincided with local F-actin and CaM accumulations. CaM inhibitor studies showed that Cobl-mediated branching is strictly dependent on CaM activity. Mechanistic studies revealed that Ca2+/CaM modulates Cobl’s actin binding properties and furthermore promotes Cobl’s previously identified interactions with the membrane-shaping F-BAR protein syndapin I, which accumulated with Cobl at nascent dendritic protrusion sites. The findings of our study demonstrate a direct regulation of an actin nucleator by Ca2+/CaM and reveal that the Ca2+/CaM-controlled molecular mechanisms we discovered are crucial for Cobl’s cellular functions. By unveiling the means of Cobl regulation and the mechanisms, by which Ca2+/CaM signals directly converge on a cellular effector promoting actin filament formation, our work furthermore sheds light on how local Ca2+ signals steer and power branch initiation during early arborization of nerve cells—a key process in neuronal network formation.
| The organization and the formation of new actin filaments by polymerization of actin monomers has the power to shape cells. The rate-limiting step in actin polymerization is “nucleation”—a process during which the first actin monomers are assembled with the help of actin nucleators. This nucleation step requires tight temporal and spatial control in order to achieve proper cell morphologies. Here, we analyse signaling cascades and mechanisms regulating the actin nucleator Cobl, which is crucial for the formation of dendritic arbors of nerve cells—a key process in neuronal network formation. We show that the calcium (Ca2+)-binding signaling component calmodulin (CaM) binds to Cobl and regulates its functions. Using 3-D time-lapse analyses of developing neurons, we visualized how Cobl works. We observed local accumulation of CaM, Cobl, actin, and syndapin I—a membrane-shaping protein—at dendritic branch initiation sites. We find that Ca2+/CaM modulates Cobl’s actin-binding properties and promotes its interactions with syndapin I, which then serves as a membrane anchor for Cobl. In summary, we i) show a direct regulation of the actin nucleator Cobl by Ca2+/CaM, ii) demonstrate that the molecular mechanisms we discovered are crucial for shaping nerve cells, and iii) underscore how local Ca2+ signals steer and power branch initiation during early arborization of neurons.
| Metazoan life critically relies on the formation, organization, and dynamics of actin filaments, which are, for example, crucial for shaping and movement of membranes and entire cells. The polar and extremely arborized morphologies that neurons develop during pre- and postnatal brain development are a prerequisite for signal processing in neuronal networks. Their development seems to be promoted by cytoskeletal structures and local calcium signals. These Ca2+ signals are mediated by N-methyl-D-aspartic acid (NMDA)-type glutamate receptors, voltage-gated calcium channels, and ryanodine receptors [1–3] and seem to be sensed by the Ca2+-binding protein calmodulin (CaM; M19312.1; GI:203255), because CaM kinases (CaMKs) downstream of CaM were observed to be involved in dendritogenesis [4,5].
Prime effector machinery that may power early neuromorphogenesis would be proteins with the ability to trigger the formation of new actin filaments in a spatially and locally well-controlled manner. The well-established actin filament-promoting components, i.e., the Arp2/3 complex and Formins, are controlled by Rho-type GTPases [6–9]. Actin nucleators that respond to Ca2+/CaM signals directly are not known. With Cobl (NM_172496.3; GI:162135965) and JMY (NM_021310.3; GI:326633181), two members of the novel class of Wiskott-Aldrich syndrome protein (WASP) Homology 2 (WH2) domain-based actin nucleators [10,11] have been implicated in the development of early neuronal morphology in different ways [12,13]. Cobl nucleates actin filaments by a mechanism that requires a combination of all three of its C-terminal actin monomer-binding WH2 domains [12]. In vitro, a WH2 domain–containing C-terminal fragment of human Cobl additionally increased actin dynamics by severing filaments [14]. Functional studies in neurons showed that Cobl plays some, albeit not fully understood, role in early arborization of hippocampal neurons and of cerebellar Purkinje cells [12,15]. These functions rely on associations of Cobl with the F-BAR protein syndapin I (AF104402.1; GI:4324451) and the F-actin-binding protein Abp1 (NM_001146308.1; GI:226423870) as well as on all three C-terminal WH2 domains of Cobl [12,15,16].
Here, we reveal a signaling pathway that controls the molecular functions of Cobl and identify a direct link between Ca2+/CaM signaling and the early morphogenesis of neurons. We demonstrate that Cobl is a direct target of Ca2+-activated CaM and that Ca2+/CaM activity and CaM interactions are crucial for the Cobl-mediated development of dendritic arbors. Ca2+/CaM modulates Cobl’s actin binding properties. Furthermore, we show that CaM association promotes Cobl’s interactions with syndapin I and that Cobl, CaM, F-actin, and syndapin I accumulate at branch initiation sites.
Taken together, our study addresses the means of Cobl regulation and reveals several molecular mechanisms that enable the Ca2+/CaM signaling pathway to steer Cobl. With Cobl, we identified an actin filament–promoting effector by which local Ca2+/CaM-signaling directly powers cellular morphogenesis during early neuronal network formation.
The extremely arborized morphologies that neurons develop during their early morphogenesis are a prerequisite for the formation of all neuronal networks. By some, yet unknown means, the underlying reorganization of cell shape involves Cobl, a protein that has the ability to promote the formation of new actin filaments [12,14]. Dendritic arborization has furthermore been suggested to be controlled by local Ca2+ influx. Indeed, 3-D-time-lapse calcium imaging in developing hippocampal neurons showed that protrusions, which formed in areas marked by transient Ca2+ influx, were originating from sites enriched for F-actin. These F-actin-rich sites either already existed prior to the calcium signal (Fig 1A) or F-actin accumulated after a calcium pulse (Fig 1B).
Analyzing the distribution and dynamics of the actin nucleator Cobl in developing neurons, we observed that the formation of protrusions from dendritic structures coincided with Cobl enrichment. Branching was a dynamic process with branches being initiated, shrinking back to the mother dendrite and being reinitiated until they were firmly established and grew out. Cobl accumulated in spatially restricted dendritic sites prior to the induction of almost all branching events (Fig 1C, arrows; S1 Fig; S1 Movie). Quantitative analyses showed that Cobl was highly enriched at branch initiation sites 30 s before initiation of protrusion. The intensity of GFP-Cobl at such sites was more than twice as high as at adjacent control region of interest (ROI), whereas GFP showed no intensity differences when control ROIs were compared to branch initiation sites (Fig 1D).
Dual imaging of GFP-Cobl and LifeAct-RFP revealed that Cobl accumulations and dendritic branch inductions were accompanied by local F-actin formation. Interestingly, the signals of both Cobl and F-actin were particularly high at the base of initiated branches. Furthermore, we observed that the maximum of Cobl accumulation hereby usually preceded that of F-actin accumulation (Fig 1E; S2 Movie).
In line with the live imaging data, immunostainings of dendrites of developing neurons showed that sites marked by increased F-actin were often marked by protrusive morphology and displayed accumulations of endogenous Cobl (Fig 1F).
In contrast to other cytoskeletal effectors that give rise to new actin filaments, Cobl has no obvious Rho-type GTPase-binding domains or other regulatory modules that may help to control this powerful cellular machine. We therefore conducted a yeast-2-hybrid screen with BD-Cobl1001–1337 as bait to identify cellular mechanisms that may be responsible for the observed local control of Cobl during branch initiation. Hit#3737 encoded for AD-CaM∆1–32. Retransformed AD-3737-plasmid led to robust reporter gene activity when combined with BD-Cobl1001–1337 (Fig 2A).
Coprecipitations confirmed that GFP-Cobl1001–1337 and GFP-Cobl specifically bound to immobilized CaM (Fig 2B).
Successful and specific reconstitutions of GFP-Cobl/CaM and GFP-Cobl1001–1337/CaM complexes at defined sites in intact COS-7 cells using mitochondrially targeted CaM (Mito-mCherry-CaM) demonstrated the in vivo relevance of the Cobl/CaM interactions (Fig 2C and 2D; S2 Fig).
Immunohistological examinations showed that Cobl and CaM display overlapping expression. Cells with pronounced Cobl expression in the hippocampus were also marked by significant CaM expression (Fig 2E, arrows). The same was true for the cerebellum. In particular, Purkinje cell dendrites relying on Cobl for branching [15] showed marked anti-CaM immunosignals (Fig 2F, arrows).
These findings suggested Cobl to be a Ca2+/CaM-regulated cytoskeletal effector. Indeed, the Cobl/CaM interaction was also observable in coprecipitation analyses with endogenous Cobl from rat brain lysates. The Cobl/CaM interaction was strongly Ca2+dependent and occurred in a wide range of Ca2+ concentrations (tested were physiological concentrations down to 2 μM) (Fig 2G).
We next aimed at further corroborating the Cobl/CaM interaction by coimmunoprecipitations of the endogenous proteins with suitable anti-CaM antibodies. A subpool of endogenous Cobl indeed was specifically coimmunoprecipitated with CaM from rat brain lysate (Fig 2H).
We next examined whether Cobl-enriched sites in dendrites may represent sites of Ca2+/CaM signaling. 3-D-time-lapse studies showed a dynamic coenrichment of Cobl and CaM at sites of protrusion initiation (Fig 3A; S4 Fig; S3 Movie).
Calcium imaging using GCaMP5G showed that also Ca2+ signals correlated with both Cobl accumulation and protrusion initiation (Fig 3B, arrow).
To experimentally test the effects of Ca2+/CaM signals on Cobl’s functions in dendritogenesis, we next employed two different CaM inhibitors: W7 (N-(6-Aminohexyl)-5-chlor-1-naphthalinsulfonamide) [17] and CGS9343B (1,3-dihydro-l-[1-[4-methy1-4H,6H-pyrrolo[1,2-a][4,l]-benzoxazepin-4-y1-methy1]-4-piperidinyl]-2H-benzimidazol-2-one(1:1) maleate) [18] (Fig 4). W7 and CGS9343B both completely suppressed Cobl-induced dendritic arborization in developing neurons (Fig 4A–4H). Also in control cells, W7 and CGS9343B incubation caused some reduction of dendritic branch points. These effects were smaller than the strong suppression of the Cobl overexpression phenotype indicating that the CaM inhibitors indeed suppress Cobl functions and do not act in a parallel pathway. The effect of W7 and CGS9343B in control cells may include an inhibition of the functions of endogenous Cobl that has been demonstrated to be a crucial factor in dendritic arborization [12] and is expressed throughout the development of the brain (S5 Fig). Similar to dissociated neurons, significant reductions of dendritic branch points by W7 and CGS9343B were also observed in Purkinje cells in developing cerebellar slice cultures (S6 Fig). Interestingly, the observed impairments mirror quite well the Coblloss-of-function phenotype in cerebellar slices [15].
Highlighting the neuronal morphology of primary rat hippocampal neurons at days in vitro (DIV)6 with PM-mCherry in 3-D-time-lapse studies showed many dynamic protrusions originating from Cobl-enriched sites in untreated neurons. After addition of CGS9343B, this dynamic behavior rapidly came to a standstill. Neuronal structures became static, and Cobl-enriched sites appeared less frequently (Fig 4I; S4 Movie). Quantitative analyses showed that the frequencies of protrusion initiation decreased to about 10% of the values of the respective cells before addition of CaM inhibitor (Fig 4J).
Taken together, Cobl-mediated dendrite branching requires Ca2+/CaM signaling events.
To unravel the mechanisms of CaM’s crucial role in Cobl-mediated dendritogenesis, we next mapped the CaM binding sites. Surprisingly, we identified multiple CaM interactions with Cobl. The N-terminus contains at least three independent CaM-binding areas (aa48-112, 147–176 and 175–229) (Fig 5A and 5B).
Coprecipitation and corecruitment studies furthermore showed that CaM binding additionally involved regions in the middle of Cobl (aa750-1005) (Fig 5C) as well as in front of the C-terminal WH2 domains (aa1001-1176) (Fig 5D; S7 Fig). Half-maximal binding was reached at 0.68 and 0.95 μM Ca2+ for N- and C-terminal parts of Cobl, respectively (S8 Fig). These values are in line with the cooperative Ca2+ binding of CaM, the kD of which decreases to ~0.5 μM upon target binding in the case of CaMKII (NM_009792.3; GI:161086915) [19–21].
Our observations raised the exciting possibility that Ca2+/CaM modulates Cobl’s cytoskeletal functions. As neither full-length Cobl nor extended C-terminal parts of Cobl can be purified [12,14], and in vitro reconstitutions of actin polymerization are hampered by Ca2+-containing physiological buffers, we focused on actin coimmunoprecipitations and in vitro reconstitutions of actin binding to explore putative Ca2+/CaM-induced modulations of Cobl functions.
Coimmunoprecipitation of endogenous actin with GFP-Cobl1001–1337, GFP-Cobl1176–1337, and GFP-Cobl1206–1337 unveiled that Ca2+ promotes actin binding. This positive effect was independent of CaM binding and also independent of the first WH2 domain neighboring the CaM binding interface but prominently involved the second and third WH2 domain of Cobl (Fig 5E–5G). Interestingly, the Ca2+-mediated increase of actin binding was not reversed upon subsequent lowering of Ca2+ levels (Fig 5F).
Since the increased actin association was more striking for GFP-Cobl1176–1337 lacking the CaM binding site than for GFP-Cobl1001–1337 containing an interface for direct CaM binding (S9 Fig), we next addressed the effects of Ca2+/CaM specifically on the first WH2 domain. Interestingly, the first WH2 domain (Cobl1176–1224) required the CaM-binding region (Cobl1001–1176) for coimmunoprecipitation of actin (S10 Fig). This specific coimmunoprecipitation of actin by Cobl1001–1224 was strongly suppressed by 500 μM Ca2+, which may only be reached in direct vicinity of Ca2+ channels, as well as by lower Ca2+ levels (2 μM), which are more commonly and more widely reached in neurons (Fig 5H).
Importantly, in vitro reconstitutions with purified proteins demonstrated that this suppression of actin binding of Cobl’s first WH2 domain solely involved actin, Cobl and Ca2+/CaM. In the presence of CaM, we observed a Ca2+-specific CaM/Cobl complex formation and a statistically significant reduction of actin binding. In contrast, in the absence of CaM, no such difference between Ca2+ and Ca2+-free conditions was observed (Fig 5I; S11 Fig). Thus, the suppression of the actin binding of the first WH2 domain involves direct Ca2+/CaM association and relies on the Ca2+ sensor CaM.
The first WH2 domain is crucial for Cobl’s actin filament promoting function [12,14], and actin filament formation as well as Cobl accumulation was observed at the initiation point of Ca2+-triggered, newly forming dendritic protrusions (Fig 1). Both observations strongly suggested that actin filament formation at branch initiation sites plays a supporting role during dendritic arborization. It was therefore interesting that Ca2+ was overall promoting actin association of the Cobl C-terminus, despite a simultaneously occurring Ca2+/CaM binding-mediated suppression of the actin association of the first WH2 domain. This suggested that actin binding may be transiently increased even further once Ca2+ levels drop again, as under such conditions the dissociation of CaM may release the suppression. In line with a transient effect, the Ca2+/CaM-mediated block of actin binding to the first WH2 domain indeed was fully reversible upon reduction of Ca2+ levels (Fig 5J).
The observation that CaM inhibition impaired Cobl accumulation at branch initiation sites (Fig 4I and 4J) prompted us to study the influence of Ca2+/CaM on Cobl’s association with the cell cortex. This process involves Cobl Homology domain interactions with syndapin I [16]. Specific coimmunoprecipitation as well as specific reconstitutions of Cobl–CaM interactions in intact cells confirmed that also the CaM interactions with the Cobl Homology domain are of relevance in vivo (Fig 6A–6C).
We thus addressed the exciting hypothesis that Ca2+/CaM signaling may not only control Cobl’s actin cytoskeletal functions but may also orchestrate Cobl’s membrane association. Whereas we were unable to purify the full Cobl Homology domain, we succeeded in purifying an alternative protein (TrxHis-Cobl54–450). In vitro reconstitutions with liposomes revealed that the N-terminus of Cobl itself has membrane-binding activity and therefore floated with liposomes irrespective of calcium presence (Fig 6D and 6E).
Interestingly, Ca2+/CaM addition effectively suppressed the direct membrane-binding ability of the Cobl Homology domain (Fig 6F; upper middle panel).
Cortical targeting of cytoskeletal effectors is a key aspect in shaping cells. The Ca2+/CaM-mediated suppression of Cobl’s direct lipid association thus was puzzling. We therefore analyzed a putative regulation of Cobl/syndapin I interactions. In vitro reconstitutions demonstrated that direct and simultaneously occurring interactions of Cobl with CaM and syndapin I give rise to complexes containing all three components (Fig 7A and 7B).
Strikingly, the anchoring of Cobl to membranes via the F-BAR domain protein syndapin I was not suppressed by Ca2+/CaM addition. Complexes containing all three components, i.e., Cobl, CaM, and syndapin I, floated with liposomes (Fig 7C–7E).
Together, the formation of complexes composed of all three components and their ability to associate with membranes suggested that Cobl’s intrinsic lipid association constitutively ensures some Cobl presence at the cell cortex. Upon association of Ca2+/CaM, this ability of Cobl is switched off. As a consequence, F-BAR domain–mediated membrane associations by syndapin I [22,23] start to dominate the spatial control of Cobl at the cell cortex. Consistent with such a scenario, we observed that addition of the Cobl-binding SH3 domain of syndapin I did not suffice for restoring Cobl’s membrane association in the presence of Ca2+/CaM (Fig 7F).
Thus, upon Ca2+/CaM association, Cobl localization becomes fully dependent on SH3 domain interactions and on F-BAR-mediated membrane association of syndapin I.
This regulatory mechanism would be even more effective if syndapin I associations were promoted upon Ca2+/CaM. In order to address this directly in vivo, we conducted quantitative coimmunoprecipitation studies. We observed a Ca2+-mediated increase of Cobl1-408/syndapin I complex formation at both 500 μM (not shown) and 2 μM calcium (+55.3 ± 17.0%; p < 0.05) when compared to conditions without calcium (Fig 8A and 8B; S12 Fig).
This increase was, to a large extent, reversible. At least upon prolonged incubation with ethylenglycol-bis(aminoethylether)-N,N,N′,N′-tetraacetic acid (EGTA) after Ca2+ stimulation, syndapin I coimmunoprecipitation intensities only remained moderately increased and were not significantly different from control anymore (+21.0 ± 11.4%; Fig 8A and 8B).
The striking promotion of Cobl’s association with syndapin I was dependent on Cobl’s ability to associate with the calcium sensor CaM because a Cobl mutant incapable of binding to CaM (Cobl1-408∆CaM) did not respond to changes of Ca2+ levels. Instead, Cobl1-408∆CaM coimmunoprecipitated constant amounts of syndapin I (Fig 8A and 8C).
Endogenous coimmunoprecipitations from rat brain lysates showed a corresponding calcium-mediated increase of Cobl/syndapin I complex formation that was consistent with the ~60% increase of syndapin I association in the heterologous coimmunoprecipitations using the Cobl Homology domain described before (Fig 8D and 8E).
In primary neurons undergoing dendritogenesis (DIV7), GFP-Cobl and Flag-mCherry-syndapin I colocalized at discrete sites within the dendritic arbor (Fig 9A). 3-D-time-lapse recordings showed that protrusions emanated from sites that were enriched for both syndapin I and Cobl. Both proteins showed very good spatial overlap at nascent dendritic branch points. Similar to the dynamic behavior of Cobl during dendritic branch induction, also syndapin I was found to accumulate at sites of branch initiation shortly before branch induction started. After protrusions had been established and grew, syndapin I and Cobl both redistributed to a more disperse localization in the mother dendrite and in the formed branch (Fig 9B; S13 Fig).
Immunostainings of endogenous Cobl and syndapin I in developing neurons confirmed the presence of syndapin I-enriched sites in dendrites that also showed accumulations of anti-Cobl signals. Often such sites were not symmetric but protruded from one side of the dendrite and may thus represent initiation sites for dendritic branching (Fig 9C).
To address whether CaM association is crucial for orchestrating Cobl functions during dendritogenesis, we next decided to employ GFP-Cobl mutants lacking N- and C-terminal CaM-binding interfaces or combinations thereof (Fig 10A). The importance of the most C-terminal CaM-binding area identified (Fig 5) hereby was addressed in form of two separate mutants, as further biochemical experiments revealed that the interface Cobl1001-1176 contained at least two areas that independently interact with CaM (Cobl1001-1101 and Cobl1100-1176). This increased the number of independent CaM interface on Cobl to at least six (Fig 10A; S14 Fig).
Cobl∆CaM-N,C lacked all CaM-binding areas that we had identified and consistently did not associate with CaM anymore (Fig 10B). The Cobl N-terminus lacking the CaM interfaces (Cobl∆CaM-N) still associated with the two other components that are known to bind to the Cobl Homology domain and are critical for the functions of Cobl in vivo, Abp1 and syndapin I [15,16] (S15 Fig). Thus, this mutant should allow for dissecting the known requirements of syndapin I and Abp1 association from a putative importance of CaM associations in functional studies. Importantly, despite preserved Abp1 and syndapin I interactions, Cobl∆CaM-N,C overexpression did not result in the extensive dendritic arborization observed upon overexpression of wild-type Cobl. Thus, the N-terminal and the more C-terminal CaM-binding interfaces are crucial for inducing Cobl overexpression phenotypes (Fig 10C–10L).
We next tested the functional importance of the identified CaM interfaces individually. Similar to Cobl∆CaM-N,C, also the individual deletions Cobl∆CaM-N and Cobl∆CaM-C were unable to trigger dendritic arborization (Fig 10F, 10G, 10K and 10L). Thus, Cobl functions seem to require both CaM binding sites of the N-terminal part of Cobl, which we showed to be involved in modulating the syndapin I interactions (Fig 8), as well as C-terminal CaM binding sites, which we unraveled to modulate actin associations (Fig 5).
Cobl with even smaller deletions (Cobl∆CaM-C’, Cobl∆CaM-C”, and Cobl∆CaM-C”‘) also failed to give rise to the Cobl gain-of-function phenotype. Instead, the morphology of transfected neurons remained indistinguishable from control cells (Fig 10H–10L). Therefore, the CaM binding area Cobl1001-1176 addressed in our mechanistic studies of Ca2+/CaM-mediated actin association was critically required for Cobl-mediated dendritic branching.
To address whether these data reflect a requirement of both N- and C-terminal CaM binding areas for physiological Cobl functions, we next conducted Cobl loss-of-function experiments and corresponding rescue experiments (Fig 10M–10V). We subjected cerebellar slices to gene gun transfections with GFP-reported plasmids. As described previously [15], Cobl RNAi-impaired dendritic branching of Purkinje cells. This Cobl loss-of-function phenotype was rescued by resupplying the cells with RNAi-insensitive, wild-type Cobl (GFP-Cobl*) (Fig 10M–10O and 10V).
In contrast, substitution of Cobl with any of the four CaM-binding–deficient mutants Cobl∆CaM-N,C Cobl∆CaM-N, Cobl∆CaM-C, and Cobl∆CaM-C’ did not only fail to rescue the Cobl loss-of-function phenotype, but impaired dendritogenesis further (Fig 10P–10S and 10V). Cobl∆CaM-C” and Cobl∆CaM-C”‘ also were unable to rescue the Cobl loss-of-function defects in dendritic branching in the developing slice cultures (Fig 10T–10V).
Thus, Cobl’s crucial role in dendritic branch induction critically relies on both the N- and the C-terminal CaM association sites, for which we have revealed the molecular mechanisms of Cobl regulation.
Shaping neurons demands that Ca2+ signals are converted into locally restricted and transient activity of force-generating effectors [24]. The activity of such effectors must be targeted to the dendritic plasma membrane and must cease to exist once a branch is induced successfully. Whereas other actin cytoskeletal effectors are controlled by Rho-type GTPases, we here describe a direct interaction of the calcium sensor protein CaM with a powerful cytoskeletal effector remodeling neuronal morphology, the actin nucleator Cobl (Fig 11).
To our knowledge, only few direct links of Ca2+/CaM to the actin cytoskeleton have been discovered in neurons thus far. However, none of them explains how Ca2+ signals may trigger dendritogenesis. Neither the described CaM associations with brain-enriched spectrin isoforms [25] nor competitive binding of the F-actin bundling protein α-actinin and CaM to NMDA receptors and L-type Ca2+ channels [26,27] offer obvious mechanisms bringing about dendritic arborization.
With Cobl, we have identified a CaM-associating component that effectively promotes the local formation of actin filaments. Actin cytoskeletal forces have the power to shape cells and cellular compartments. Whether other WH2 domain-containing actin nucleators [10,11] or actin filament formation via the Arp2/3 complex and/or Formins also are directly controlled by Ca2+/CaM remains to be addressed.
We found by yeast-2-hybrid, coprecipitations, coimmunoprecipitations, and corecruitment studies in intact cells that CaM associates with Cobl. Reconstitutions with purified proteins demonstrated that Cobl’s interactions with CaM are direct. The Ca2+ dependency of the interactions strongly suggested that Cobl/CaM complexes are involved in translating Ca2+ signals sensed by CaM into cellular answers. Colocalizations in different parts of the brain and a dynamic coappearance of Cobl and CaM at induction sites of dendritic protrusion support this hypothesis.
Several lines of evidence from functional studies underscore the importance of Cobl/CaM interactions during dendritogenesis: CaM inhibitors fully suppressed Cobl-mediated dendrite formation and branching in quantitative end-point analyses of fixed neurons and in live imaging studies. Our data strongly argue that a direct association and not merely CaM signaling is required, as overexpression of Cobl mutants incapable of associating with CaM failed to mimic Cobl-mediated effects on dendrite formation. Furthermore, re-expression of such Cobl mutants failed to rescue the Cobl loss-of-function phenotype in dendritic branching of cerebellar Purkinje cells. These findings clearly point to a critical function of CaM associations with Cobl.
Our analyses revealed multiple independent CaM binding areas. At least three of them reside in the Cobl Homology domain. In addition, at least three further binding motifs are located in Cobl’s C-terminal part. Usually, the CaM lobes wrap around target segments, burying them in a hydrophobic channel, and thereby enforce alter target conformations [28]. Consistent with such putative conformational changes, we observed dramatically altered properties of both the N- and C-terminus of Cobl upon Ca2+/CaM signaling. Overall, calcium strongly promoted the actin association with the C-terminus of Cobl in a manner that was independent of CaM association and independent of the first WH2 domain of Cobl. In principle, such Ca2+-mediated effects could be due to different mechanisms, i) Ca2+ modulates the properties of actin (G-actin or G-actin/F-actin balance), and this is reflected by changed association with Cobl WH2 domains, ii) Ca2+ modulates the properties of the Cobl WH2 domains number 2 and 3, and this leads to changed actin affinities or iii) Ca2+ signals to further cellular components, and these either modulate actin properties, Cobl properties, or both.
The observed Ca2+-mediated increase in Cobl’s overall actin binding occurred despite a simultaneous suppression of the actin binding of the first WH2 domain upon CaM association with neighboring sites, as demonstrated by quantitative coimmunoprecipitation studies and by in vitro reconstitutions with purified components. In line with CaM acting as calcium sensor protein, this suppression was released with decreasing Ca2+ levels, whereas the Ca2+-promoted actin binding of the Cobl C-terminus persisted. Since all three WH2 domains need to work together and the first WH2 domain is crucial for actin filament formation [12,14], the release of the suppression of the first WH2 domain after a transient Ca2+ signal would elegantly ensure that Cobl responds to calcium transients. In contrast to simple on/off mechanisms, transient activations would also exclude Cobl activity during longer lasting NMDA and Ca2+/CaM signaling, such as during excitotoxicity conditions. Indeed, such conditions do not promote filament formation but are marked by filament loss [29]. Related to high Ca2+ levels during excitotoxicity-reducing actin dynamics and filament formation, high calcium was reported to inhibit the F-actin-driven formation of filopodia-like dendritic spines and stabilized them, whereas lower calcium signals were reported to promote the formation of these F-actin-rich structures. It therefore is conceivable that, besides the crucial role in dendritogenesis, Cobl-mediated functions and Ca2+/CaM-mediated control of Cobl may in addition play some role in synapse formation and/or plasticity processes, as these involve both Ca2+ signaling and actin cytoskeletal reorganizations [30].
We observed local F-actin accumulations at the base of dendritic protrusions. We have demonstrated earlier that Cobl overexpression promotes dendritic arborization, and Cobl loss-of-function results in a reduction of branches [12,15]. Correlative 3-D-time-lapse studies, inhibitor studies, and mutational analyses demonstrated that dendritic branching is associated with Cobl, F-actin and syndapin I at branch initiation sites and is controlled by changing calcium levels, by Ca2+/CaM signaling and by direct Cobl association of CaM.
We have identified a direct plasma membrane association of Cobl that also was CaM-regulated. This ability of the Cobl Homology domain was suppressed upon CaM association. At the same time, Ca2+/CaM promoted Cobl’s association with syndapin I [16,31] and thereby promoted indirect membrane associations of Cobl. Direct membrane association of Cobl may ensure its general availability at the plasma membrane and may also explain why syndapin I loss-of-function only partially suppressed Cobl’s cortical localization [16]. Upon CaM binding, syndapin I interactions will increasingly influence Cobl’s localization. Indeed, our studies revealed that dendritic branching events correlated with accumulations of both syndapin I and Cobl.
In line with this conclusion, syndapin I has been demonstrated to be involved in dendritogenesis [16]. Interestingly, we observed accumulations of syndapin I specifically at nascent dendritic branch sites. It thus seems that F-BAR domain-mediated membrane curvature sensing and/or induction by syndapin I [22,32] spatially steers the actin nucleator Cobl at the cell cortex. Consistently, mutational overexpression analyses in neuronal cultures as well as rescue experiments of cerebellar Cobl loss-of-function revealed that CaM binding to the actin-binding C-terminal part and to the syndapin I-binding N-terminal part of Cobl were crucial for Cobl functions in dendritogenesis.
Together, our work unveils that Cobl is regulated by the calcium sensor protein CaM and reveals the Ca2+/CaM-controlled molecular mechanisms that are crucial for Cobl’s cellular functions. The regulation by Ca2+/CaM seems to distinguish Cobl from established actin nucleators, such as the Arp2/3 complex and Formins, which are directly and indirectly regulated by Rho-type GTPases.
Our examinations of the Ca2+/CaM-mediated mechanisms that control Cobl’s activity in neurons furthermore provided deep insights into how local Ca2+ signals steer and power branch initiation during early arborization of nerve cells.
Plasmids encoding for GFP-Cobl and deletion mutants thereof were described previously [12,15,16] and generated by PCR and subcloning into pEGFP (Clontech), respectively.
GST fusion proteins of Cobl were generated by subcloning into pGEX-5X-1.
A plasmid encoding for TrxHis-Cobl54–450 was generated by PCR and subcloning into a pET32 vector (Novagen).
The yeast-two-hybrid bait used was generated by subcloning Cobl 1001–1337 into pGBTK7 (Clontech).
Plasmids encoding for GFP- and GST-tagged syndapin I full-length and SH3 domain, respectively, were described previously [33,34]. Flag-mCherry-syndapin I was generated using a modified, pCMV-based vector originally expressing Flag-GFP [12]. GST-Abp1 SH3 was described [35].
GST-CaM, TrxHis-CaM, Flag-mCherry-CaM, and Mito-mCherry-CaM were generated by subcloning rat CaM from pCMV-myc-CaM (kindly provided by L. C. Russo [University of São Paulo, São Paulo, Brazil]) into pGEX, pET32, Flag-mCherry, and Mito-targeting vectors [36], respectively.
RNAi constructs directed against mouse Cobl coexpressing GFP and RNAi-insensitive GFP-Cobl full-length (GFP-Cobl*), respectively, as well as the scrambled RNAi vector were described previously [12,15]. RNAi vectors coexpressing GFP-Cobl mutants defective for CaM-binding were generated by combining the generated Cobl mutants with RNAi-resistant parts of Cobl in pRNAT. As a result, using flanking NheI and SmaI sites, GFP-Cobl*Δ48–229 (GFP-Cobl*ΔCaM-N), GFP-Cobl*Δ714–1175 (GFP-Cobl*ΔCaM-C), GFP-Cobl*Δ48–229,714–1175 (GFP-Cobl*ΔCaM-N,C), GFP-Cobl*Δ1002–1175 (GFP-Cobl*ΔCaM-C’), GFP-Cobl*Δ1002–1100 (GFP-Cobl*ΔCaM-C”) and GFP-Cobl*Δ1101–1175 (GFP-Cobl*ΔCaM-C”‘) replaced the GFP reporter.
LifeAct-RFP and-GFP were gifts from K. Murk (Charite Universitätsmedizin Berlin, Germany) described previously [37]. Vectors expressing PM-targeted (farnesylated) mCherry and GFP-GCaMP5G were kindly provided by M. Korte and R. Köster, respectively (both TU Braunschweig, Germany).
Correct cloning by PCR was verified by sequencing in all cases.
Guinea pig and rabbit anti-syndapin I, anti-TrxHis, as well as anti-GST antibodies, were raised and affinity-purified as described previously [34,38]. Guinea pig anti-Cobl antibodies were raised and affinity-purified as described [15]. Polyclonal rabbit anti-GFP antibodies were from Abcam and monoclonal mouse anti-GFP antibodies (B34, JL-8) were from Covance and Clontech, respectively. Monoclonal mouse (M2) and polyclonal rabbit anti-Flag as well as anti-MAP2 (HM-2) antibodies were from Sigma. Polyclonal rabbit anti-CaM antibodies (#4830) were from Cell Signaling Technology. Mouse anti-CaM antibodies (23-132-27) were from DSHB. Monoclonal mouse anti-CaM antibodies (G-3) were from Santa Cruz Biotechnology. Monoclonal mouse and polyclonal rabbit anti-actin antibodies were from Sigma. Polyclonal rabbit anti-MAP2 antibodies were from Abcam. Phalloidin AlexaFluor®488 and MitoTracker were from Molecular Probes.
Secondary antibodies used included, Alexa Fluor488- and 568-labeled goat anti-guinea pig antibodies, Alexa Fluor488- and 568-labeled donkey anti-mouse as well as Alexa Fluor647-labeled goat anti-mouse antibodies, Alexa Fluor488-labeled donkey anti-rabbit, Alexa Fluor568- and 647-labeled goat anti-rabbit antibodies and AlexaFluor680-labeled goat anti-rabbit and anti-mouse antibodies (Molecular Probes); goat anti-rabbit, anti-guinea pig, and anti-mouse-peroxidase antibodies (Dianova); DyLight800-conjugated goat anti-rabbit and anti-mouse antibodies (Pierce) and donkey anti-guinea pig antibodies coupled to IRDye680 and IRDye800, respectively, (LI-COR Bioscience).
Rabbit skeletal muscle actin was from Cytoskeleton. GST- and TrxHis-tagged fusion proteins were purified from E. coli as described previously [16,39].
Direct protein–protein interactions were demonstrated by coprecipitations with combinations of recombinant TrxHis- and GST-tagged fusion proteins purified from E. coli and CaM-sepharose 4B (GE Healthcare), respectively. TrxHis-CaM regulation of GST-Cobl1001–1224/actin complex formation was demonstrated in 10 mM HEPES pH 7.4, 0.1 mM MgCl2, 1% (v/v) Triton X-100 (EGTA-free lysis buffer) with 250 mM NaCl, 500 μM Ca2+, 0.2 mM ATP, and 0.5 mM DTT supplemented with EDTA-free protease inhibitor cocktail and 200 μM calpain inhibitor I (Sigma).
Studies of TrxHis-Cobl54–450/GST-syndapin I/GST-CaM complex formation and controls were done in EGTA-free lysis buffer with EDTA-free protease inhibitor cocktail, 200 mM NaCl, and 500 μM Ca2+.
Liposome-binding assays were conducted with lipids from Folch-fraction type I (Sigma) essentially as described previously [22,40].
Analyses addressing direct interactions of GST-Cobl1001-1176 with CaM were done in lysis buffer with 150 mM NaCl containing 0 and 1 μM Ca2+, respectively (set according to [41]). Eluted proteins were analyzed by SDS-PAGE and subsequent anti-TrxHis and anti-GST immunoblotting.
24–48 h after transfection, HEK293 cells were washed with PBS, harvested and subjected to sonification for 10 s and/or lysed by incubation in lysis buffer containing EDTA-free protease inhibitor Complete (Roche) and 120–150 mM NaCl for 20 to 30 min at 4°C. Cell lysates were obtained as supernatants from centrifugations at 16,000 xg (20 min at 4°C).
For coprecipitation experiments, extracts from HEK293 cells expressing different GFP fusion proteins were incubated for 3 h at 4°C with purified GST-fusion proteins immobilized on glutathione sepharose beads (GenScript) as described [33]. Bound protein complexes were eluted with 20 mM-reduced glutathione, 120 mM NaCl, 50 mM Tris/HCl pH 8.0.
For coprecipitations with CaM, HEK293 cell lysates were prepared in EGTA-free lysis buffer containing 150 mM NaCl, and EDTA-free protease inhibitor cocktail and 200 μM calpain inhibitor I. Cell lysates were supplemented with either 1 mM EGTA or to be tested Ca2+ concentrations.
For binding curves, Ca2+ concentrations ranging from 0 to 500 μM were set according to [41]. After incubation with 25 μl CaM-sepharose 4B for 3 h at 4°C and washing, bound proteins were isolated by boiling in SDS sample buffer.
Lysates, supernatants, and eluates were analyzed by immunoblotting using anti-GST and anti-GFP antibodies, respectively.
For binding curves, quantitative immunoblotting experiments (n ≥ 3) were conducted and data expressed as percent binding, %(Elution/∑Elution)/(Elution/∑Elution+Supernatant/∑Supernatant) with pEGFP control values subtracted. Sigmoidal dose-response curves were fit from Graphpad Prism and modified using Adobe Illustrator.
Lysates from HEK293 cells transfected with GFP-Cobl106–324 or GFP together with Flag-mCherry (FlagC)-tagged CaM or Flag-mCherry were incubated with 2 μg of rabbit anti-Flag antibody and nonimmune rabbit IgG (Santa Cruz Biotechnology) prebound (3 h, 4°C) to protein A/G-agarose (Santa Cruz Biotechnology) in lysis buffer containing 50 mM NaCl.
Coimmunoprecipitations of endogenous actin together with GFP and GFP-tagged Cobl deletion mutants were done according to [12] with slight modifications. Lysates of HEK293 cell were incubated in EGTA-free lysis buffer containing 100 mM NaCl, 5 mM DTT, 200 μM calpain inhibitor I and varying amounts (2 μM, 500 μM) of Ca2+ or no free Ca2+ (EGTA addition) with 5 μg rabbit anti-GFP antibody/well (6-well plate) for 3 h at 4°C.
Reversibility of Ca2+-induced suppression of actin binding of Cobl1001–1224 by CaM association was tested by incubating 2 μM Ca2+-treated samples subsequently with 1 mM EGTA for 2 h. The CaM-independent Ca2+-induced (500 μM Ca2+) increase of actin binding to Cobl’s C-terminal part (Cobl1176–1337) was tested for reversibility by subsequent EGTA addition (5 mM EGTA final). Antibody-associated protein complexes were isolated with protein A/G-agarose (2 h, 4°C), washed with a buffer with the respective Ca2+ concentrations and eluted by boiling in a mix of 8 M urea and 4x-SDS-sample buffer. The eluates were immunoblotted with anti-GFP and anti-actin antibodies and analyzed quantitatively using fluorescently labeled secondary antibodies and a LI-COR Odyssey System.
Quantitative coimmunoprecipitation analyses of Flag-syndapin I with GFP-Cobl proteins were done similarly except that the lysis buffer lacked DTT and contained 75 mM NaCl and protein A-agarose (Santa Cruz Biotechnology) was used. Confirmatory experiments were additionally done at 100 mM NaCl for consistency with previously published conditions for heterologous syndapin I/Cobl coimmunoprecipitations [16].
Coimmunoprecipitations of endogenous Cobl and CaM were performed using rat brain lysates in lysis buffer with 30 mM NaCl, 500 μM Ca2+ and 200 μM calpain inhibitor I using mouse anti-CaM (G-3) antibodies according to coimmunoprecipitation procedures described [16]. Coimmunoprecipitations of endogenous Cobl and syndapin I were performed using rat brain lysates in lysis buffer with 30 mM NaCl and 400 μM calpain inhibitor I with and without 2 μM Ca2+ using guinea pig anti-CoblDBY antibodies as described [16], except that antibodies were added to brain lysates and then isolated with protein A-agarose.
The amounts of coimmunoprecipitated proteins under different conditions were normalized to the amount of immunoprecipitated GFP-Cobl and Cobl, respectively, and expressed as percent difference from Ca2+-free conditions. Statistical analyses were performed using one-way ANOVA with Tukey’s post-test. *p < 0.05, **p < 0.01, ***p < 0.001.
Lysates from HEK293 cells transfected with GFP-CaM were incubated with 5 μg rabbit anti-GFP antibodies, 10 μl rabbit anti-CaM antibodies (4830, Cell Signaling), 2 μg mouse anti-CaM (G-3, Santa Cruz) antibodies and 50 μl mouse anti-CaM (23-132-27 DSHR) antibodies, respectively, for 3 h, 4°C in EGTA-free lysis buffer containing 30 mM NaCl, 0.5 mM Ca2+ EDTA-free protease inhibitor cocktail and calpain inhibitor I and thereafter for 2 h with protein A/G-agarose (4°C). 5 μg rabbit and mouse anti-IgG was used as specificity controls. Immunoprecipitated material was analyzed by immunoblotting with anti-GFP antibodies and the anti-CaM antibodies, respectively.
Culturing of HEK293 and COS-7 cells and immunolabeling were essentially as described [42]. HEK293 and COS-7 cells were transfected using TurboFect (Thermo Scientific). Mitochondria of COS-7 cells were stained with 0.2 μM MitoTracker Deep Red FM (Molecular Probes) in medium at 37°C for 1 h and cells were subsequently fixed with 4% PFA for 7 min.
Primary rat hippocampal neuron cultures for immunofluorescence analyses were prepared, maintained, and transfected (at DIV4) as described previously [16,43,44]. Fixation was done at DIV6 in 4% (w/v) PFA in PBS pH 7.4 at RT for 4–6 min. Permeabilization and blocking were done with 10% (v/v) horse serum, 5% (w/v) BSA in PBS with 0.2% (v/v) Triton X-100. Phalloidin stainings and antibody incubations were done in the same buffer without Triton X-100 according to [42,44].
Brain sections of adult (7 weeks) male mice were prepared and immunolabeled as described [15].
Brain preparations at different developmental stages, mRNA and cDNA preparation, as well as RT-PCR, were done according to [15] using the primers GCTCCGGAAGACTGCAGAACA (forward-WH2; positioned at exon border 12/13) and CGAGCAAGGGAACCTTTCTTAGTC (reverse-WH2; positioned at exon border 14/15) for Cobl detection and the primers ATTGACCTCAACTACATGGTCTACA (forward) and CCAGTAGACTCCACGACATACTC (reverse) for GAPDH as control.
Confocal images were recorded using a Leica TCS SP5 microscope (equipped with 40x/0.75dry and 63x/1.4 oil objectives, LAS AF software), a Zeiss LSM Meta 510 (using the 20x/0.5dry objective and ZEN software) or a Zeiss AxioObserver.Z1 microscope equipped with an ApoTome. Both Zeiss microscopes were equipped with Plan-Apochromat 100x/1.4, 63x/1.4, 40x/1.3 and 20x/0.5 objectives and an AxioCam MRm CCD camera (Zeiss). Digital images from Zeiss microscopes were recorded by ZEN2012 or AxioVision Software (Vs40 4.8.2.0). Image processing was done by Adobe Photoshop.
Primary hippocampal neurons undergoing dendritic arbor formation were transiently transfected with Lipofectamine 2000 at DIV6. For imaging, the culture medium was replaced by 20 mM HEPES pH 7.4, 0.8 mM MgCl2, 1.8 mM CaCl2, 5 mM KCl, 140 mM NaCl, 5 mM D-glucose (live imaging buffer) adjusted to isoosmolarity using a freezing point osmometer (Osmomat 3000; Gonotec). For inhibitor studies, 10 μM (final) CaM inhibitor CGS9343B (Tocris) was used in DMSO and accompanied with the respective solvent control (0.1% DMSO final). Live imaging was conducted 16–28 h after transfection in an open coverslip holder placed into a temperature- and CO2-controlled incubator built around a spinning disc microscope. The microscope was a motorized Axio Observer combined with a spinning disc unit CSU-X1A 5000 and equipped with a 488 nm/100 mW OPSL laser, a 561 nm/40 mW diode laser and a QuantEM 512SC EMCCD camera (Zeiss). Images were taken as Z-stacks (stacks of 7–17 images at Z-intervals of 0.31 μm depending on cellular morphology) at time intervals of 10 s and 3 s (Ca2+ imaging) with exposure times of 50–200 ms and 3%–12% laser power using a C-Apochromat objective (63x/1.20W Korr M27; Zeiss).
Image processing was done using ZEN, Imaris software, and Adobe Photoshop.
Quantitative comparisons of signal intensities at dendritic branch initiation sites with signal intensities at dendritic sites that were not branching were done by placing a circular ROI at the branch initiation site (covering the dendrite diameter) and on an adjacent dendrite section (distance 2 ROI diameters). The fluorescence signal intensities of GFP and GFP-Cobl were measured at both sites 30 s prior to initiation of protrusion and expressed as relative enrichments relative to control ROI.
Frequencies of protrusion initiation from dendrite sections of neurons incubated with DMSO control and CaM inhibitor CGS9343B in DMSO were compared to those prior to treatment and expressed as protrusions per 10 μm dendrite section and min.
Dendrite analyses of transiently transfected (DIV4) hippocampal neurons immunostained for MAP2 (NM_013066.1; GI:6981181) were performed with ≥2 independent neuronal preparations on 2–6 independent coverslips per condition in each assay at DIV6. Neurons were sampled systematically on each coverslip. Morphometric measurements were based on anti-MAP2 immunolabeling and performed with ImageJ according to [16,44]. The number of dendrites and the number of dendritic branching points were determined from 53 to 197 neurons (DIV6) for each condition. The CaM inhibitors W7 (N-(6-Aminohexyl)-5-chlor-1-naphthalinsulfonamide; Tocris) and CGS9343B (1,3-dihydro-l-[1-[4-methy1-4H,6H-pyrrolo[1,2-a][4,l]-benzoxazepin-4-y1-methy1]-4-piperidinyl]-2H-benzimidazol-2-one(1:1) maleate; Tocris) were applied 30 h after transfection at final concentrations of 10 μM and 0.1% DMSO for 18 h. All data were normalized to internal GFP controls run in parallel in each individual experiment and neuronal preparation.
Preparation of cerebellar slices, gene gun transfection and morphometric analyses of Purkinje cell dendrites in the Molecular Layer of the cerebellum were done as described [15].
Statistical analyses were performed using GraphPad Prism 5 and 6, respectively, and one-way ANOVA with Tukey’s post-test. *p < 0.05, **p < 0.01, ***p < 0.001.
|
10.1371/journal.pbio.1000201 | In Silico Reconstitution of Actin-Based Symmetry Breaking and Motility | Eukaryotic cells assemble viscoelastic networks of crosslinked actin filaments to control their shape, mechanical properties, and motility. One important class of actin network is nucleated by the Arp2/3 complex and drives both membrane protrusion at the leading edge of motile cells and intracellular motility of pathogens such as Listeria monocytogenes. These networks can be reconstituted in vitro from purified components to drive the motility of spherical micron-sized beads. An Elastic Gel model has been successful in explaining how these networks break symmetry, but how they produce directed motile force has been less clear. We have combined numerical simulations with in vitro experiments to reconstitute the behavior of these motile actin networks in silico using an Accumulative Particle-Spring (APS) model that builds on the Elastic Gel model, and demonstrates simple intuitive mechanisms for both symmetry breaking and sustained motility. The APS model explains observed transitions between smooth and pulsatile motion as well as subtle variations in network architecture caused by differences in geometry and conditions. Our findings also explain sideways symmetry breaking and motility of elongated beads, and show that elastic recoil, though important for symmetry breaking and pulsatile motion, is not necessary for smooth directional motility. The APS model demonstrates how a small number of viscoelastic network parameters and construction rules suffice to recapture the complex behavior of motile actin networks. The fact that the model not only mirrors our in vitro observations, but also makes novel predictions that we confirm by experiment, suggests that the model captures much of the essence of actin-based motility in this system.
| Networks of actin filaments provide the force that drives eukaryotic cell movement. In a model system for this kind of force generation, a spherical bead coated with an actin nucleating protein builds and rockets around on an actin “comet tail,” much like the tails observed in some cellular systems. How does a spherically symmetric bead break the symmetry of the actin coat and begin to polymerize actin in a directional manner? A previous theoretical model successfully explained how symmetry breaks, but suggested that the subsequent motion was driven by actin squeezing the bead forwards—a prediction refuted by experiment. To understand how motility occurs, we created a parsimonious computer model that predicted novel experimental behaviors, then performed new experiments inspired by the model and confirmed these predictions. Our model demonstrates how the elastic properties of the actin network explain not only symmetry breaking, but also the details of subsequent motion and how the bead maintains direction.
| The directed assembly of actin networks drives the motility of most eukaryotic cells [1]. Specialized cellular factors assemble actin into different network types, each with a unique architecture and cellular function [2]. One of the most well-studied actin assembly factors is the Arp2/3 complex, a seven-subunit protein complex that nucleates new filaments from the sides of pre-existing filaments to create entangled, dendritic filament arrays [3],[4]. These arrays behave like viscoelastic gels with an elasticity that depends on the degree of branching, and which break or rip under relatively low stress [5].
In vivo, dendritic networks built by Arp2/3 complex form the lamellipod that drives the movement of eukaryotic cells [3],[6] as well as the “comet tails” whose assembly drives the intracellular movement of endosomes [7],[8] and intracellular pathogens [9] such as Vaccinia virus [10] and Listeria [11]. Construction of these motile networks in vivo requires a set of highly conserved accessory proteins, including capping protein, cofilin, and profilin, that function together with the Arp2/3 complex in a simple biochemical cycle converting monomeric actin into crosslinked polymer and back again [6],[12]. Motile, dendritic actin networks can also be constructed in vitro by recombining purified components of the actin assembly cycle [13]–[16]. These reconstituted actin networks have become a powerful tool for studying how individual protein–protein interactions control the large-scale behaviors of cytoskeletal systems.
The simplest way to initiate assembly of such motile, dendritic actin networks in vitro is the “bead motility” system, in which micron-sized beads are uniformly coated with factors that activate the Arp2/3 complex to nucleate actin networks at their surfaces [16],[17]. These networks form spherically symmetric shells that eventually “break symmetry” and produce stable, asymmetric comet tails that propel the bead along, maintaining direction [14],[16],[18], moving smoothly or pulsing depending on conditions [19],[20]. In this work, we concentrate on how a geometrically and biochemically symmetric bead can first break symmetry then maintain asymmetry to produce directed smooth or pulsatile motion.
Spatially localized nucleation of actin filaments combined with global inhibition of filament elongation by capping protein restricts filament growth to a well-defined zone, e.g., the Listerium surface [21], lamellipodial plasma membrane [22], etc. On the spatial scale of filaments, a Brownian ratchet mechanism has been proposed [23],[24] to explain how actin polymerization uses the energy of ATP hydrolysis to rectify Brownian fluctuations, exerting force at the surface, as new actin monomers, as new actin monomers add onto existing filaments and extend the network. Although the specific details may vary [25]–[27], spatially localized network extension fueled by ATP hydrolysis is the basis of all polymerization-driven motility models.
Several theoretical frameworks have been proposed to explain actin-based symmetry breaking and bead motility (reviewed in [28]). Some are based on filament-scale descriptions of actin assembly and crosslinking [29],[30], while others take a more coarse-grained approach based on the bulk mechanical properties of crosslinked polymer networks [17],[19],[20],[31]–[34]. One such coarse-grained model is the Elastic Gel model [19],[31], which provides an intuitive explanation for symmetry breaking. In this model, symmetry breaking occurs when new actin network, continuously deposited at the surface of the bead, displaces older portions of the network radially outward. Expansion of the older network stretches it like the surface of an inflating balloon until, at a critical threshold, circumferential stress causes a rupture in the network (either by melting [33] or cracking [35] the shell) and breaks the symmetry of the system. This mechanism fits the experimental observations of symmetry breaking [16],[19] better than mechanisms inferred from filament-based descriptions of the network [30]. Pulsatile motion has been suggested to result from an unstable balance between the pushing forces and the drag from attached filaments [20].
Explaining the smooth directional motility of symmetrically coated beads has proved more challenging. One attempt, the Soap-Squeezing model [31], is an extension of the Elastic Gel model that offers an explanation of propulsive force. In this model, surface-associated polymerization stretches older network outwards orthogonal to the direction of motion, storing energy, which it releases by contracting orthogonally, squeezing the bead forward like a hand squeezing a wet bar of soap. However, photobleaching data showing the movement of the network as it leaves the bead demonstrate that orthogonal squeezing does not occur [17], and whereas treating the network as an incompressible fluid flowing from the bead surface can explain the observed motion [17], this violates the elastic nature of the gel required to explain the initial symmetry breaking. How, then, does sustained motility occur?
In this paper, we examine the essence of actin-based bead motility by reconstituting it in silico from the network's fundamental viscoelastic properties. Just as reconstituting actin-based motility in vitro from a minimal set of purified protein components demonstrates their necessity and can show how they contribute to the large-scale behavior, reconstituting actin-based motility in silico allows us to demonstrate the necessity and specific contributions of a minimal set of higher-level network properties (e.g., elasticity, crosslinking, etc.), and demonstrate the mechanisms of motility on a mesoscopic scale. To do this, we use a framework we call the Accumulative Particle-Spring model (APS model) in which the viscoelastic actin network is represented simply as a set of particles, subject to viscous drag and coupled by springs that break when strained beyond a certain limit. New Particle-Spring network is created at the bead surface, just as the in vitro actin network polymerizes at the bead surface [16], and we find that this simple system is sufficient to reproduce a range of the behaviors of actin networks, including symmetry breaking and motility.
Our simulations enable us to explore the feasibility of hypothesized mechanisms of force and movement generation, using Ockham's razor to determine the essence of the behavior by exploring the minimal requirements to produce the observed results. We validate the model by checking the results and predictions of the simulations with in vitro experiments in which we reconstitute symmetry breaking and motility from purified proteins. To the extent that the model is valid, we are able to make explanatory claims for the mechanisms involved in symmetry breaking and motility, determining 1) the stress and strain distributions in a growing symmetric actin shell and in a comet-like tail, 2) where the symmetry break is initiated (outer or inner surface of the actin shell), 3) the 3-D structure and dynamics of the break, 4) what determines the transition from smooth to pulsatile motility, and 5) how symmetry breaking occurs for nonspherical objects.
To perform our in vitro bead motility experiments, we evenly coated 5-µm diameter beads with ActA and added them to motility mix (see Materials and Methods). ActA activates Arp2/3 to nucleate an actin network that grows in a tightly localized zone at the bead surface, breaks symmetry, and propels the bead on an actin comet tail (Figure 1A–1D and Video S1).
To find out how well bead motility can be explained simply by the viscoelastic properties of the network, we created a computational model that simulates the behavior of a generic viscoelastic network deposited stochastically at the surface of a bead. The model starts at t = 0 with no network, then nucleates nodes at a constant rate and with an even distribution across the bead surface, crosslinking new nodes to their neighbors with links that behave as simple Hookean springs that break if extended too far (Figures S1 and S2). See Materials and Methods and Section S1 of the supporting text (Protocol S1) for full details of the model, and Tables S1 and S2 for the experimental bases for the model assumptions. We tuned the model parameters (spring constant, crosslinking probability, etc.) to produce qualitatively similar observations to the in vitro system (see Model Robustness, Section S3 of the supporting text (Protocol S1) for the effects of varying each parameter. Table S3 lists the corresponding names in the code for simulation parameters mentioned in the main text). This simple model exhibits both symmetry breaking and motility behavior that reproduces the sequence of events seen in vitro (Figure 1E–1H, Video S2).
Our experimental observations and our simulations share several features. As the shell grows, it becomes denser near the surface of the bead. When the thickness of the shell reaches approximately the radius of the bead, a clear crack develops, and the bead exits the shell, then the shell opens, crescent-like, and motility proceeds, leaving a low-density and somewhat irregular comet-like tail behind the bead. Figure 1I–1L show the underlying 3-D nature of the simulated network, with the network links colored by tensile stress (Videos S3 and S4).
Although the simulations share many of the features of the experiments, we noticed that the shell shows a close to perfect arc for the experimental conditions in Figure 1, but the simulations robustly show a more V-like shape with a dent in the center of the inner high-density region of the shell (compare Figure 1C and 1D with 1G and 1H). This implies either a failure of the simulation to capture an essential behavior of the network, or a condition of the in vitro system that we did not include in the simulations.
To determine the cause of the dent, we examined the 3-D mechanics of symmetry breaking in our simulations. Figure 2A and 2B show 3-D top and side views of a representative simulated shell after the bead has moved away from the shell, demonstrating that even though the bead is unconstrained in three dimensions, the symmetry break and shell opening occur along only one axis. A rip in the outer shell often accompanies the dent, as seen in Figure 2A (arrow) and the corresponding 2-D projection view shown in Figure 2C. To understand why symmetry breaking occurs within one plane, we looked at how the shell cracks. Figure 2-D shows an earlier 3-D view of the same simulation, just as the crack completely fractures the shell; isosurfaces show the densest region of the network in green to highlight the shape of the shell, and the extent of the lower-density actin network (semitransparent). The symmetry-breaking crack is a straight line, as opposed to either lightning-like fracture(s) along the weakest regions of the network, or a circular hole opening to allow the bead to escape. The consequence of this straight-line break is that the 3-D stresses in the network are relieved in a 2-D manner—essentially splitting the 3-D spherical shell into two hemispheres that open apart from one another like a clamshell, causing large stresses at the hinge. When this 3-D geometry is viewed from above, the hinge appears as a dent, seen in Figure 2A and 2C. The crack that opens the two hemispheres often continues all the way around the outer network, resulting in the rip in the outer shell that accompanies the dent. For only one rip to occur, as soon as a crack begins, circumferential tension must relax quickly around the bead before a second crack begins. We can reduce this relaxation around the bead by increasing the strength of attachments with the nucleator (Figure S16), which prevents the network moving relative to the bead and makes the second crack progressively more prominent.
For the experimental conditions in Figure 1, we had intentionally confined the bead closely between a slide and coverslip to prevent it moving out of focus while we took data. Having seen how the crack propagates around the bead in the simulations, we hypothesized that the lack of a dent seen in the experiments might be a result of this constraint on the network preventing the crack propagating to the rear of the bead. To test this, we ran the same simulation while constraining the network between two planes (we also excluded nucleation from the very top and bottom 10% of the bead to prevent artifacts caused by this material having nowhere to go). Figure 2E and 2F correspond to 2C and 2D, but for this constrained shell (interactive 3-D representations are included in Figure S4). The constraint creates a toroidal shell that also breaks in a straight-line crack, but unlike the breaking of the spherical shell, the broken toroidal shell relaxes into a much more perfect arc, with the dent much reduced and the shell more closely resembling those seen in the experiments.
If our simulations are a valid model for the behavior of the actin network, they predict that if we were to perform the symmetry-breaking experiment in an unconstrained 3-D volume in vitro, it would produce a clamshell break with a dent in the shell opposite the break site as we see in the simulations. To test this, we performed the in vitro experiment using 5-µm diameter ActA-coated beads while controlling the headspace of the reaction with glass spacer beads of either 5.1-µm diameter for the constrained condition or 15.5 µm for the unconstrained condition. Because the 3-D shell structure is hard to interpret from a single 2-D microscope image, we reconstructed the 3-D shells from confocal z-stacks. We fixed the reaction after symmetry breaking (see Materials and Methods) to prevent movement while the z-stack was acquired; so for experiments, we are only able to capture the 3-D geometry at one time point after symmetry breaking has occurred, in contrast to having every time point in the simulations. Figure 2G and 2H show an example of a 2-D projection and 3-D reconstruction of a confocal stack of an unconstrained bead, confirming the distinctive bilobed structure, and V-shaped shell with central dent. Figure 2I and 2J similarly show the constrained condition with the near-perfect arc. (Beads tend to settle by gravity so that the tail and wide axis of shell are parallel to the coverslip, with shell cracks in the z-direction.) Figures S5 and S6 contain further examples of 2-D projections and 3-D reconstructions of symmetry breaking. Shell geometry for constrained beads was extremely consistent, always showing the near-perfect arc. Unconstrained beads showed less regularity, but always showed shells with shapes consistent with linear cracks; on one occasion, we observed a shell with a three-way opening (Figure S6B).
To confirm that the mechanics of symmetry breaking in our simulations reflect those seen in vitro, we tracked the deformations of the shell during in vitro symmetry breaking using fluorescent speckle microscopy (Figure 3A, Video S5). Low doping of fluorescent actin produces fiduciary marks that allow us to measure the mechanical deformations of the network [36]. We tracked five parameters: bead displacement, expansion of the crack, circumferential stretching of the inner shell, circumferential stretching of the outer shell, and radial stretching of the shell (Figure 3B and 3C). When symmetry breaks, the crack opens rapidly and then slows as the shell approaches its final shape. As the shell opens, the outer circumference contracts with kinetics that mirror the crack opening, but the inner shell remains approximately the same circumference, merely reducing its curvature. As the shell opens, it also becomes thicker, with the kinetics of radial expansion mirroring the circumferential contraction and crack opening (magenta and blue lines in the graphs in Figure 3C).
We plotted similar parameters for a simulation run. We measured the 3-D distance between pairs of points approximately 2 µm apart (e.g., in the circumferential direction; Figure 3-D and Videos S6 and S7). The mechanics of the simulations behave like the in vitro experiments, with the crack opening rapidly, the outer circumference of the shell contracting and the shell becoming radially thicker, all with similar kinetics. The values of the Poisson's ratios differ a little, approximately 0.2 for the in vitro shell and approximately 0.3 for the simulation, likely resulting from simplifications in the functional forms for the link and repulsive forces (previous theoretical models have assumed a wide range of Poisson ratios, from 0 to 0.5 [31],[33],[37]). Also, the behavior of the inner shell differs slightly between experiment and simulation, with the circumference transiently expanding slightly (frame 140) before returning to its original length, whereas in vitro, the length remains constant. This most likely reflects transient disequilibrium during the most rapid part of the symmetry breaking, which is equilibrated more quickly in vitro than in the simulations. The current model therefore reproduces the qualitative behavior of the experiments but requires calibration in future work before it would be able to match quantitative measures. (N.B. For convenience, we note that 1 s corresponds to approximately 1.4 frames, but stress that this is not extensively kinetically calibrated.)
Our simulations provide detailed information about the mechanism of symmetry breaking, e.g., the network motion, distribution of forces and ripping of the network (Figure 4A–4D, Video S8). In the left panels (Figure 4A(i)–4D(i)), we colored the regions of the network with red stripes to show the trajectory of the network as it moves away from the bead surface. Initially (frames 1–60), this pattern is radially symmetric—broken links occur randomly around the surface, giving no indication of the future site of symmetry breaking (link breaks are stochastic, see Video S8(ii) and Video S11). By Frame 62 (Figure 4A), the nodes around the future crack site have begun to diverge (Figure 4A(i)), followed by a burst of localized link breaks at the site (Figure 4B(ii)). This weakens the network, causing stress in that region to be distributed over fewer remaining links, leading to more breaks by positive feedback (Figure 4C(ii)), followed by the bead moving off with links breaking primarily at the front (Figure 4D(ii), Video S12).
To determine the force balance that contributes to shell formation and symmetry breaking, we examined the spatial distribution of stresses within the network. The right-hand graphs (Figure 4A(iv)–4D(iv), Video S8(iv)) show how the radial and circumferential tensions vary with distance from the surface of the bead (negative tension corresponds to compression), and the center panels (Figure 4A(iii)–4D(iii)) show the spatial distribution of circumferential tension. These are calculated as sums of the link tension forces (positive) and the node–node repulsion forces (negative), split into radial and circumferential components (individual components are graphed in Video S9; we exclude the data point nearest the bead because of surface artifacts caused by the way we deal with nodes that enter the nucleator, see Video S10 for full data). Both radial and circumferential tensions are negative at the bead surface, i.e., the center of the shell is under compression, the inner compressive forces balancing the outer circumferential tension. For small network distortions (close to the surface), the network equilibrates this compressive force primarily through the isotropic node–node repulsions, so the compression is not restricted to the radial component. Close to the bead surface, circumferential tension is lower (as predicted by the Elastic Gel model), so the compressive force is greater than the tension force (and the overall tensile force is negative).
Circumferential tension increases rapidly with distance from the bead (Figure 4C(iv)), becoming positive at approximately 1.0 µm, with the maximum tension approximately 1.5 µm from the surface, and tailing off at higher distances as the network becomes sparse. This distribution of forces can be clearly seen when the symmetry break begins (Figure 4C(iii)) as a red band of maximal circumferential network tension at approximately 1.5 µm encloses a blue band of maximal network compression at the bead surface. The distribution remains relatively static over time as forces build up (Figure 4A(iv)–4C(iv)), although the magnitudes of the forces change, with the maxima occurring when symmetry breaking begins (Figure 4C(iv)). These data support the Elastic Gel model for symmetry breaking: as the network is pushed out by nucleation at the center, it expands in the circumferential direction like a balloon, creating circumferential tension. Network compression close to the surface provides the balancing force for this circumferential tension—and because the expanding layers of network pull the network apart circumferentially, but not radially, the resulting radial forces are always compressive (negative tension in the graphs in Figure 4A(iv)–4D(iv)). The release of tensile energy upon symmetry breaking can be vividly seen between Figure 4C(iii) and 4D(iii)—the shell opens and pulls back away from the bead, contracting circumferentially and releasing the energy stored in circumferential tension—much of the red region of maximum circumferential tension in Figure 4C(iii) turns blue (compression) in Figure 4D(iii), Video S8.
Small defects in the outer shell have been proposed to establish the site of symmetry breaking [32],[33]. We can determine when the symmetry breaking site is established in our simulations relatively easily. In our simulations, we add new network stochastically at the bead surface—this randomness results in a unique network and symmetry-breaking direction for each run. For each run, we save a complete description of the system at each time point, and can resume the run at any point with a different random seed. To discover the time at which the symmetry-breaking direction is determined, we ran a simulation through to symmetry breaking, then rewound and restarted the same simulation from nine different time points, but with a different random seed. We repeated this set of nine runs five times to calculate the mean and standard deviation of the angle between the new symmetry-breaking direction and the original direction (Figure 4E). This produces a high variance in symmetry-breaking direction before the direction is determined, and both very low variance and a close to zero deviance angle afterwards. We find the symmetry-breaking direction is essentially random until frame 80, at which point the direction becomes the same as the original run. Symmetry-breaking direction is therefore determined between frames 70 and 80, i.e., very late—just before symmetry breaks—rather than being determined early by defects in the initial outer network.
Our simulations also show that the force balance and pattern of link breaks in the outer network before symmetry breaking define the final curvature of the shell after symmetry has broken. Figure 4F shows that halving the spring constant (the FL parameter) causes the shell to double in thickness, and Figure 4G shows that increasing the threshold force for link breakage (the FBL parameter in the simulation) causes the shell to become flat (see also Figures S13 and S12). These results follow from the Elastic Gel model: decreasing the spring constant between links of the network will require that more material be deposited to build up enough circumferential tension for symmetry to break, so the shell is thicker. Also, the final curvature of the shell after recoil is dependent on the number of links that have broken in the outer shell during the earlier stages of shell buildup. Without breaks in the outer shell, the final equilibrium area of the outer shell is still the same as the inner, so the resulting shell is flat. The more links that break in the outer network, the larger its equilibrium area, and the higher the resulting curvature. These parameters and others are more thoroughly explored in Model Robustness, Section S3 of the supporting text (Protocol S1).
Symmetry breaking is a particularly robust behavior of our model. Of the parameters tested, those that do not break symmetry are those that set network link density to extremes (Figures S10, S11, S12, S13, S14, S15, S16, S17, S18, S19, and S20). One extreme creates a very strong network that builds a dense shell that never breaks symmetry, by creating conditions in which the network strength increases faster than the network strain, e.g., when we increase the threshold for link breakage (Figure S12). The other extreme creates a very weak network in which symmetry does not break because chains of links are too short to communicate tension around the bead, so the network remains unpolarized, seen by decreasing the crosslinking probability, or decreasing the link-breaking threshold (Figures S11 and S13).
Our model network is constructed from nodes and links that are short compared to the size of the bead—to transmit force around the bead, there must be enough links to form chains spanning around the bead. The “mesh size” characterizes the length scale of the network formed from these chains of links, referring to the minimum size of a particle that would be trapped by a network made of these chains. In our case, if the mesh size is greater than the size of the bead, the bead would be able to move through the network, so it would not be possible to build up tension in the shell, and there would not be a clean symmetry break. For our purposes, we define network coherency as the bead size divided by the mesh size, i.e., high network coherency means that the bead will see the network as an elastic solid, whereas low coherency means the bead would be able to squeeze through the network.
We find that even a low level of network coherency is sufficient to support symmetry breaking, the key is that tension is transmitted around the bead. This kind of symmetry breaking does not involve a distinct shell that cracks, but rather a gradual oozing of the bead from a network cloud (Figures S11 and S13). This oozing demonstrates a qualitative change in behavior that results from the quantitative change in degree of crosslinking. When a sparsely linked network deforms, it undergoes plastic flow as energy is lost by links breaking independently, whereas when a dense network deforms, it builds up elastic energy, as each link stretches slightly while remaining below its breaking strain. Eventually, this dense network undergoes brittle fracture when many links break at once.
The initial shell shows a gradient of network density increasing from the outer to the inner surface of the shell both in vitro and in silico. This density gradient emerges spontaneously from the APS model as a result of the increasing circumferential tension in the outer shell compressing the inner shell. The initial outer network is sparse because it is not under compression, so the network has a low density of links (since links are formed to nearby nodes, and a sparse network means fewer nodes nearby). This sparse initial outer network is weak and plastic but does provide enough compression on the inner network to cause an increase in density, hence a greater number of links, and a stronger network, which builds by positive feedback. As demonstrated in Figure 4A–4D, which shows a peak in circumferential tension towards the center at around 1.5 µm from the surface, it is this inner brittle network that stores the bulk of the elastic energy, and undergoes brittle fracture during symmetry breaking.
In both our experiments and simulations, the bead continues to move after breaking symmetry. To investigate the motility mechanism, we examined network movement by plotting orthogonal views of the network trajectory for a simulation of smooth motion (Figure 5A). To show the network trajectory, we marked the network with a spatiotemporal grid, coloring it red when it originated at evenly spaced locations around the bead (the parallel lines in the tail), and at even time intervals during the run (the orthogonal shell-like curves). During the smooth motion phase, we see a pattern of parallel lines behind the bead, demonstrating that the network does not contract orthogonally as it moves away from the bead surface, which agrees with previous experimental work showing no orthogonal network contraction for motile beads [17],[38]. So in our simulations, orthogonal contraction of the network does not provide the driving force for motility by squeezing the bead forwards. In Figure 5A, the time-pulse markings highlight regions of network that come from the bead surface within short time windows—in effect demonstrating what happens to the equivalent of “shells” for smooth motion. In the tail, they appear as red lines with curvature much lower than the bead curvature, i.e., even during smooth motion, the high-curvature network produced at the bead is opening up just like the shell during symmetry breaking. The shape of these smooth-motion shells also match well those produced by physically switching the color of the actin during in vitro experiments [17],[38].
Even though the bead in this simulations is not constrained, during smooth motion, the network sweeps around the bead primarily in one plane—Figure 5A shows that the tail is much wider in one axis than the other, similar to the shell during symmetry breaking in Figure 2A and 2B. In three dimensions (Figure 5B and Figure S7), tracking the network trajectory shows ripping in one axis along a sustained straight-line crack at the front of the bead. We confirmed that the trajectories of the network in our simulations match those seen in vitro using fluorescent speckle microscopy (Figure 5C, Video S13). The composite image is produced by coloring and overlaying successive frames from a video of a motile bead in vitro, registered to the motile bead (i.e., lines represent movement relative to the bead). The trajectories in vitro mirror those seen in silico, with network expanding away from the bead as it is swept around and incorporated into the tail, and no convergence of trajectories behind the bead. The effect of this sweeping motion on the circumferential tension in the simulated network can be seen in Figure 5D. The network shows a peripheral zone of circumferential tension (red) at the outer network surface, and a region of network compression (blue) just behind the bead. This tension zone is far from the bead surface except at the thinnest part of the network at the front of the bead.
The opening of the “smooth-motion shells” in Figure 5A is reminiscent of how the shell opens during symmetry breaking, and suggests that the network might contract circumferentially and expand radially, as we saw during symmetry breaking in Figure 3. To test this, we made similar measurements of the network stretching during smooth motion, and because the network is asymmetric during smooth motion, we restricted measurements to the rear of the bead; Figure 5E and 5F show lines used to take circumferential and radial length measurements during the smooth motility phase (shown in Videos S14 and S15). Figure 5G shows how the network behind the bead stretches as the bead moves, confirming that it stretches circumferentially to approximately 120% before relaxing back to approximately 107% of its original length. As it does so, it expands radially to approximately 112%—similar to the radial expansion of the outer shell during symmetry breaking. This relaxation is complete after approximately 150 frames (∼18 µm), consistent with previous in vitro photobleaching data showing the network is still undergoing relaxation at approximately one bead diameter and is complete by approximately four bead diameters [17].
Why do the trajectory lines of the network look parallel (and even diverge slightly) as they move away from the bead? Although the network contracts circumferentially, it also rotates around the bead, i.e., the network on the outer edges of the tail sweeps backwards relative to the inner tail. This rotation allows the points in this smooth-motion equivalent of a shell to contract relative to one another while following the parallel trajectories shown in Figure 5B; i.e., there is circumferential, but not orthogonal, network contraction. The Soap-Squeezing model proposes that orthogonal elastic contraction of the network drives motility. The lack of orthogonal network contraction rules this out, but could circumferential elastic network contraction play a similar role?
To determine whether circumferential elastic contraction is required for motility, we performed in silico experiments to find out what happens when elastic contraction is reduced or eliminated. Changes in these parameters affect both the bead velocity profile and the stretching of the shell. Figure 5H shows the velocity profile of the bead described above, before reducing elastic contraction. The bead is initially at rest, with a distinct spike in velocity upon the original symmetry-breaking event. (Note: the smooth motility regime still has small velocity fluctuations, especially just after symmetry breaking. To clearly distinguish between the two regimes, we define smooth motion as having velocity that varies <25% of the mean velocity, and pulsatile motion as having velocity that varies >100% of the mean velocity.)
We first reduced the elastic contraction by tuning network parameters to produce a less elastic network. We based these parameters (RM = 5.0, FBL = 2.0, FL = 4.0) on the Model Robustness results, Section S3 in the supporting text (Protocol S1). Figure 5I shows this less elastic network expands more and contracts less: the network stretches circumferentially to 133% of its original length before relaxing back to only 128%, with a slight radial expansion, to 105%. The velocity profile under these conditions (Figure 5J) shows smooth motility, but strikingly lacks the initial spike in velocity compared to Figure 5H, and the onset of motility is delayed. For the elastic network, the initial velocity spike corresponds to the symmetry-breaking event, and Figure 5M shows that for the less elastic network, rather than producing a single shell with its buildup of elastic energy and sudden release and contraction that ejects the bead, the network fractures in multiple places, producing three separate tails. Eventually, the bead squeezes out orthogonal to these tails (Figure 5N, Videos S16 and S17), with smooth motion and network trajectories that resemble the bead in Figure 5A. In spite of being less elastic, this network still contracts circumferentially, and observation of network motion suggests this contraction is likely driven by network fractures that opened during expansion being closed by the compression forces of material swept around the bead. To abrogate this contraction, we performed the same experiment but allowed network movement only for nodes within a limited range of the bead, permitting the network to expand, but locking it in place before it could contract. This results in similar smooth motility (and a similar pattern of network tracks) under these conditions, showing that network recoil is not required for smooth motion (Figure 5L and 5K, and Video S18).
What explains smooth directional motility? We propose a “Sustained Rip” model: an extension of the symmetry-breaking mechanism combined with a pressure-induced transition from brittle to plastic network behavior. For smooth motility, as during symmetry breaking, network produced at the bead surface tends to be pushed outward, creating circumferential tension (Figure 5D). During motility, however, the existing shell (or tail) reinforces the network at the rear, forcing circumferential tension to be relieved by stretching and ripping at the front (Figure 5B). The radial compression that balances the circumferential tension presses on the bead from all sides except where there is little network—at the front (Figure 5D). The imbalance of these compressive forces causes the bead to move forwards, driving it through the rip site. Ripping also means that radial compression does not build up enough to compress the network and cause it to become dense and brittle—it remains sparse and plastic. Direction is maintained because contact with the tail (or the original shell) always reinforces the network at the back, leaving tension from the expanding network to be relieved by ripping in the unreinforced zone at the front. The network trajectories in Figure 4D and circumferential tension plot in Figure 5D support this, showing that contact with the original shell restricts the new network from free expansion at the rear—the new network does not expand symmetrically as the original shell did in Figure 4A, but diverges less in the rear region in contact with the shell, and more at the front.
This Sustained Rip model predicts that specific changes in network properties will affect the continuity of motion. For example, after symmetry breaking, motility should be smooth only if the newly forming network is sparse and plastic when uncompressed. If the newly forming network has a high enough link density that it behaves like the brittle inner network of the original shell, we should see pulsatile motion—essentially repeated symmetry breaking as new brittle shells form one after another. Changing the probability of forming network links (PXL) is a simple way to test this prediction by altering the network link density. (Note that this is an alternative to, but does not exclude, friction as a contributor to pulsatile motion [20].)
We ran simulations to see how varying the probability of forming links affects the smoothness of motility. Figure 6A shows the network architecture at regular time intervals, and Figure 6B shows the corresponding bead velocity profiles, for a range of link probability (PXL) values. At very low link probabilities (PXL = 0.125), there are so few links that each part of the network behaves independently rather than forming a single coherent network—and a symmetric cloud of material surrounds a stationary bead. Increasing PXL to 0.375, symmetry breaks and the bead moves off. Under these conditions, the shell is barely coherent—it remains together but does not recoil when symmetry breaks; instead a diffuse cloud of material forms, and the bead gradually oozes from it. There are fluctuations in the velocity, but they remain small (<25% deviation from the mean velocity). As we increase PXL to 0.625, a distinct shell forms, the bead undergoes one pulse after the initial symmetry break, and then the motion becomes smooth (<25% deviation from average velocity). As PXL increases further to 0.875, the shell becomes denser, and the motion becomes very strongly pulsatile (>250% deviation from the mean velocity) and periodic, as strong shells repeatedly undergo largely independent symmetry-breaking events. Bead velocity rises abruptly when the shell breaks, and tails off slowly as the shell relaxes, leading to an asymmetric velocity profile that closely matches experimental measurements of bead velocity during pulsatile motion [20]. This transition from smooth to pulsatile motion supports the Sustained Rip model for motility: as network coherency increases, the stronger shells formed are more immune to the influence of the previous shell, causing them to undergo essentially independent symmetry breaking. The small influence of the previous tail explains the relatively constant direction of motion.
Further supporting the Sustained Rip model, two other parameters of the APS model also control smoothness of motility by affecting the ability of the old network to alter the brittleness of the newly forming network: 1) Increasing the node repulsive force makes the network less compressible, reducing the pressure-dependent density increase, and leading to smooth motion (Figure S15); and 2) lowering the link spring constant FL results in circumferential tension (and radial compression) building up more slowly (i.e., the network has to get bigger before the dense, brittle shell forms) causing a much thicker shell when symmetry breaks, thick enough to be beyond the effect of the initial tail, and immune from the sustained rip effect's ability to induce smooth motion (Figure S14).
Friction may also contribute to pulsatile motion: in vitro, increasing surface ActA concentration (intended to increase the ActA-filament attachment component of friction) causes a transition from smooth to pulsatile motion [20]. We see a similar effect in our simulations: when we increase friction by increasing the strain limit before node–bead links break, we also see a transition from smooth to pulsatile motion (Figure S17; note the transition is less clear-cut than those described above). However, in the APS model, we can show that friction is unnecessary for pulsatile motion. We can set friction to zero by eliminating node-bead links, but still induce the transition from smooth to pulsatile motion by increasing network coherency, e.g., by increasing PXL (Figure S20). We interpret this to mean that the change from smooth to pulsatile motion is directly caused by a change from a plastic to brittle network, and that a dense, brittle network can be caused by increasing its density in two ways, either 1) by increasing the coherency of the outer shell, which puts pressure on the inner shell, or 2) by increasing the network–bead attachment, which increases the density of the inner shell by holding it close to the bead surface.
Our data show how an evenly coated spherical bead can be driven on an actin comet tail, but the original observations of this form of motility were on the intracellular motility of the bacterium, Listeria monocytogenes, which is a different shape (capsule-shaped rather than spherical) and has an asymmetric distribution of the actin nucleation factor, rather than symmetric. How important is this asymmetric distribution to the lengthwise motility of Listeria? To determine the importance of shape and of nucleator distribution on motility, we tested the effect of varying them in silico. When we simulate a capsule-shaped nucleator with nucleation restricted to one half of the capsule, motility is lengthwise and symmetry breaking is unnecessary (Figure 7A–7D). Network tracks with regular spacing and frequency (Figure 7C) and 3-D tracks (Figure 7D, Figure S8, and Video S19) show that the network expands outward from the nucleator, opening up as it moves away from the surface. Similar to the motility of spherical beads, there is no evidence for orthogonal contraction of the network.
When we distribute nucleation uniformly over the capsule surface, however, the direction of motion changes: for both symmetry breaking and motility, the capsule moves sideway, as shown in top and side views in Figure 7E–7H and Video S20. The Elastic Gel model predicts that the higher the surface curvature, the faster the buildup of strain within the network [19]. We therefore anticipated the higher curvature regions at the ends would build up strain faster and that symmetry breaking would occur there (the ends are higher curvature because although the radii are equal, the curvature is 2-D at the ends but only 1D on the linear section). To understand why symmetry breaks sideways, we examined the network motion by plotting network tracks just prior to symmetry breaking (Figure 7I). This shows that as tension builds up, the network on the linear section is drawn towards the ends of the capsule, so relieving the strain and the network tension in this direction remains low (Figure 7J). Around the capsule's cylindrical axis, however, there is no linear section to expand and relieve the strain buildup, so the tension in this direction builds up rapidly (Figure 7K). Symmetry breaking therefore occurs in this direction (causing sideways motion) by a similar mechanism to the spherical beads, and the sideways symmetry breaking and motion of this geometry can be explained by the sustained rip mechanism described above, in which the axis of the rip is defined by the long axis of the capsule.
We checked our prediction of sideways symmetry breaking and motility by stretching spherical beads to make ellipsoids and comparing their in vitro motion with simulations. Figure 7L (Video S21) shows that simulations of ellipsoids produce the same sideways symmetry breaking seen for the capsules (subsequent motion is also sideways like the capsules, Video S22). We performed bead motility experiments as above with a 15.5-µm headspace (i.e., unconstrained), and captured 3-D z-stacks of the beads soon after symmetry breaking. Figure 7M and 7N show a 2-D projection and 3-D reconstruction of such an ellipsoidal bead experiment after sideways symmetry breaking, with two density isosurfaces: the green chosen to show the shell, and the semitransparent grey chosen to outline the void space of the ellipsoidal bead to confirm the bead position and orientation. (Note that it is not possible to determine the direction of motion relative to the bead axis from the 2-D projection in Figure 7M alone.) More examples of sideways symmetry breaking of ellipsoidal beads are shown in Figure S9. For ellipsoid aspect ratios >1.75∶1, we almost always see sideways symmetry breaking (98%, n = 58) and sideways motion (95%, n = 55), though we occasionally see beads changing direction or curved bead paths during the subsequent motion.
In this study, we show that a minimal set of viscoelastic network properties are sufficient to reconstitute actin-based motility in silico. Having gathered data on the behavior of the actin network during in vitro motility experiments and reconstituted this behavior in silico, we explored this in silico system to show how the network properties give rise to the behavior. We also found some novel behaviors, e.g., sideways motion of ellipsoids and shell dents for 3-D symmetry breaking, which we tested by performing more experiments with the in vitro system. Experimentally confirming these novel predictions without having to re-tweak the model suggests that the model is not simply replicating the experimental data fed to it, but has captured the essence of a significant underlying mechanism of actin-based motility.
Our simulations build on the Elastic Gel model of symmetry breaking [19],[31], using an Accumulative Particle-Spring (APS) model to capture the mesoscopic viscoelastic properties of actin networks. The APS model represents these properties using a series of nodes and springs that allow us adjust a simple set of viscoelastic network parameters that correspond to mechanical properties of the in vitro network. For example, the repulsive force between nodes (FR) roughly corresponds to the resistance of the network to compression, and the spring constant (FL) roughly corresponds to the resistance to tension. The APS model also captures some network behavior as emergent properties. For example, as the network stretches circumferentially, links reorient circumferentially to result in strain hardening, and compression of the inner network by the outer network increases the node and spring density, resulting in the more brittle behavior necessary to produce the symmetry breaking and transition from smooth to pulsatile motion seen in silico and in vitro.
The APS model builds the network from spring-node units that correspond to a particular mesoscopic mechanical behavior of crosslinked actin networks. We know a good deal about the viscoelastic behavior of in vitro actin networks from studies that examine the randomly crosslinked networks produced by mixing crosslinking proteins with stabilized actin filaments. For these networks, crosslinking proteins connect adjacent filaments with one another to form chains with a characteristic mesh size that can resist tension across the sample. The chains of nodes and springs in silico approximate the behavior of these chains of filaments, crosslinks, and friction, to transmit tension around the in silico bead. For Arp2/3-built networks to transmit tension around the bead implies significant friction and entanglement. Activated at the bead surface by ActA, Arp2/3 binds to existing filaments and nucleates new filaments from their sides to form a dendritic branched structure [3],[5]. Because only new filaments are crosslinked, each dendritic tree cannot crosslink to any other, so there can be no encircling chains of filaments and crosslinks around the bead that could carry tension. Circumferential tension would simply be dispersed by separation of these independent dendritic networks were it not for friction and entanglement. The node-spring links in our APS model, therefore, also implicitly represent these friction and entanglement links between dendritic trees, and just as friction and entanglement would be expected to increase with network density and pressure, so the density of node-spring links in the APS model increase with density and pressure. We create links only at the surface when nodes form, to mimic in vitro filament entanglement, which can only occur when filaments polymerize and insert through gaps in the existing network, and this occurs only at the bead surface. We keep the polymerization rate constant in our simulations in spite of changes in protein concentrations and pressures at the bead surface during shell growth, because previous data show the in vitro rate of deposition of actin to remain essentially constant over this period of the reaction (Figure S6 from [16]).
In an expanding shell, the actin network continuously stretches as it is displaced outward by assembly of new actin at the surface. The opening of the shell during symmetry breaking is well explained by the basic assumption of the Elastic Gel model that all network layers tend to relax to their equilibrium area, the area of the surface of the bead where they were created. Since this area is the same for all layers, and since connected layers with equal areas and a non-zero thickness would tend to flatten to a plane, the shell tends to flatten towards a plane once symmetry breaks. For most conditions, we do not see a perfectly flat plane, but we do see the shell relax to a flat plane when we increase the link strength. This is because high link strength reduces the number of links that break in the initial outer shell as it is stretched—high link strength means that links only break during the actual symmetry-breaking event. This explains the curvature of the arc of the symmetry-breaking shell: Before symmetry breaking, as the outer shell is stretched, links break irreversibly, expanding the equilibrium area of the outer shell, so the final shell shape is no longer the relaxation of planes of equal equilibrium areas. The larger equilibrium area of the outer network results in a convex shell.
The APS model also shows how the rip that occurs during symmetry breaking brings about the clam-like 3-D geometry of the shell. Since the starting geometry is a sphere, as the shell opens and flattens, large tensile strains occur around the circumference (Figure 8A). Rips relieve these circumferential strains; a single rip will produce a bilobed structure, but multiple cracks are possible (and observed in silico and in vitro) as the network strength is increased. We also often see a crack in the outer network opposite the main symmetry-breaking crack. When the bead is unconstrained, this tends to line up with the dent in both the simulation (Figure 2A and 2C) and experiment (Figure 2G), but can also be present in constrained beads without the dent (Figure 2E), showing that the dent is not the cause of the rip. In line with a previous experimental observation [35], our simulations also show linear cracks (instead of a round-hole opening to release the bead). These are linear rather than circular because positive feedback concentrates the strain to regions of high curvature [39]. The resulting cracked-shell geometry is reminiscent of the Mollweide projection of the globe, in which linear cuts in the map allow a 3-D sphere to be flattened to a plane and reduce stretching distortions at the poles.
Paradoxically, pulsatile motion is relatively simple—it is essentially repeated symmetry breaking—whereas smooth motion is more complex, involving a transition to a different regime. The very same conditions build an initial rigid brittle shell that cleanly and distinctly breaks symmetry and then builds a more plastic tail on which the bead moves smoothly. How does the presence of the old shell cause adjacent new network to behave in the plastic manner that produces smooth motion? Our simulations suggest that this switch to plastic behavior rests on the pressure dependence of network plasticity. By reinforcing one side of the newly forming network, the old shell focuses the circumferential tensile strain on a small region of newly forming, uncompressed, and therefore, plastic network on the other side, which rips. Just like inflating a balloon with duct tape on one side—the duct tape not only prevents that side expanding, but it means the other side is stretched twice as much to accommodate and ruptures sooner. In the bead case, this leads to a rip before pressure has built up—so the network remains sparse and plastic, which in turn leads to continued ripping and steady-state smooth motion. If this pressure dependence is disrupted or reduced, the transition to smooth motion is delayed or abolished. In our simulations, increasing PXL increases the number of links and the coherency of the shell, leading to essentially independent shells and pulsatile motion.
We expect this mechanism to correspond to the physical mechanisms that produce the switch to smooth motion seen in real actin networks, in this case through pressure-dependent increases in entanglement, friction, and filament orientation effects (likely to be significantly affected by pressure, as load-directed filaments stall). Oblique filaments would tend to entangle and reinforce the network while contributing little to the movement of the bead away from the network, and so this may tip the system into a positive feedback of network stiffening that is relieved by symmetry breaking. We predict a significant alignment of filaments orthogonal to the direction of motion for a pulsatile bead, but less orthogonal alignment for a smoothly motile bead.
We can also consider these network behaviors in terms of changes in network mesh size. This refers to the distance between the chains of links that transmit tension through the network, i.e., the mesh size decreases as crosslink density increases but is always greater than the individual link lengths. When the symmetry-breaking shell forms, the pressure produces a tightly crosslinked network with a small mesh size (on the order of the link length). Because the mesh size is very much smaller than the bead size and the shell, the network behaves as an elastic solid. Decreasing crosslink density increases the mesh size and results in a mesh size that is larger than the bead, but smaller than the shell. This means that the shell can still resist tension, but beads can essentially move through the network, resulting in the oozing symmetry breaking seen in Figures 6 and S11. Decreasing crosslink density still further produces a mesh size greater than the bead and the shell, so tension is not communicated around the bead, and symmetry does not break. The switch from brittle to plastic behavior can also be seen in terms of mesh size. Although the pressure buildup in the initial shell produces a dense network with small mesh size and elastic-solid behavior, once symmetry breaks and the rip at the front prevents pressure buildup, the sparse network at the front of the bead essentially has a large mesh size that allows the bead to move through unhindered.
The repeated shell-breaking mechanism we propose for pulsatile motion does not exclude other proposed models; e.g., Listeria and motile vesicles have asymmetric nucleator localization during motility [16],[20],[38],[40]–[42], so are unlikely to build up symmetric shells. This suggests a friction mechanism for pulsatile motion, though pressure buildup still may contribute to periodic variations in friction. In our simulations, we show that a frictionless bead still produces pulsatile motion, suggesting that although friction may contribute to pulsatile motion, it may not be required. In addition to the pulsatile motion whose steps are of the order of the bead size, Listeria can also make steps of approximately 5.4 nm [29],[43]. These “nano-saltations” are very likely to be directly caused by friction because their scale is of the order of actin monomers, much smaller than the characteristic scale of the elastic gel properties of the network.
Our prediction that the shell outer network is more flexible and plastic and the inner network more rigid and brittle has implications for the mechanism of symmetry breaking. The driving force behind symmetry breaking is the circumferential stretching of the network as it moves outward, and we initially expected to see a brittle crack in one region of the outer network that would seed the symmetry break as has been previously proposed [32],[33]. Instead, we find that the symmetry-breaking direction is determined late because the tensile stress is primarily carried, not by the very outer network, but by a dense rigid network relatively close to the bead surface. We stress that this does not mean that the network does not rip at the outside first—it does because this is the most stretched region—but the outer network rips in many places without triggering symmetry breaking; it is the rip of the inner network that determines the symmetry breaking site, and this is not determined by the outer network.
If stochastic variations in the density of the initial (outer) layers of the network were to determine the symmetry-breaking direction, we would expect the direction to be determined early, when this initial network forms. We show that symmetry-breaking direction is determined late in the simulations, just before the rip occurs, implying that there is no existing vulnerability in the outer network that later seeds the crack, but rather that network density and linking are finely balanced up to the critical point when load becomes too great, and failure occurs stochastically. This fits well with the mechanism proposed above for curved versus flat shells: the balanced stochastic breaking of links in the outer network, not only equilibrates the strain, but results in the even-expansion equilibrium area of the outer shell. When symmetry breaks, shell curvature is determined by the balance of the equilibrium areas of the inner and outer shells—when the outer layer equilibrium area expands, we see curved shells, and when the link strength is increased, the even breaking is eliminated, the outer layer equilibrium area does not expand, and we see flat shells.
Our conclusions about site selection are based on our simulations—so do they also hold for the in vitro system? This depends on where tension is carried, which depends on the network rigidity—if the inner network is more rigid than the outer network in vitro, then our conclusions should hold; if the outer network is more rigid than the inner, then they will not. There are several reasons to think the inner network will be more rigid in vitro: First, the inner network is denser in vitro, as shown in Figures 1 and 3. Second, we often observe numerous small cracks in the outer network (Figure 1A and 1B) prior to symmetry breaking that do not predict symmetry-breaking direction, but rather suggest a general stochastic fracture of the outer network similar to the general breakage of links we observe in the simulations. A third reason follows if the Sustained Rip model is valid, since it predicts that under no compression the network will be plastic, not rigid. Since the initial outer shell is formed under no compression, it should be plastic and therefore not carry significant tension.
We show that elastic recoil is not required for smooth motility, but is necessary for the classic “shell-retraction” type of symmetry breaking. At first sight, the lack of orthogonal network contraction during bead motility seems to suggest a lack of elastic recoil during smooth motion, but detailed data from our simulations show elastic retraction circumferentially around the bead and, because of its positive Poisson's ratio, radial expansion, an elastic recoil very similar to symmetry breaking. Although elastic recoil is not required for smooth motility, it is necessary for the shell retraction during symmetry breaking. Without it, the network is unable to expand circumferentially and absorb the energy with elastic stretching, but instead quickly rips, resulting in several tails from which the bead eventually emerges.
During smooth motility, the network motion appears dominated by plastic flow around the bead. In previous work, Paluch et al. [17] describe a model for smooth motility that explains the network motion by treating the actin network as an incompressible gel that flows around the bead. Although this model relies on force generation by soap squeezing, which is contradicted by their photobleaching data, the general model of network motion by flow of an incompressible gel is consistent with our findings that network compressibility and retraction are not required for smooth motility. Lacking experimental data, previous models have varied widely in their assumptions about network compressibility [31],[33],[37], though recent work suggests it is a particularly important determinant of stress buildup [44]. In our simulations, the plastic flow we see during smooth motility approximates an incompressible gel regime, not because the gel itself is less compressible, but because the compressive forces are lower—the front rip prevents pressure building up enough to significantly compress the gel. The results of our simulations show how the two processes can be reconciled in one system: Symmetry-breaking behavior is dominated by network compression and elastic recoil because the shell is elastic and brittle because it is built under high pressure, whereas smooth motility is dominated by plastic flow because the tail is built under lower pressure because of tension release at the rip.
Our conclusions also agree with previous results showing that actin shells from which a solid bead escapes open wide, straighten, and then go on expanding after the bead has moved out [38]. In that paper, Delatour et al. [38] also suggest that evacuation of the gel by elastic recoil is required for movement by evacuating the actin filaments grown in front of the bead to maintain anisotropy in the system. This is based on the observation that during pulsatile motion, the bead periodically slows down and reinitiates the formation of a quasisymmetric actin shell and repeats the initial symmetry-breaking step over and over. The actin shells in this regime are never perfectly symmetrical, but weaker at the front, so the initial direction of the movement (defined by the gap in the first shell) is partially conserved. Our results support Delatour et al.'s interpretation that direction is maintained mechanically by reinforcement by the existing tail, but we differ in our interpretation of the role of elastic recoil. We find that elastic recoil is not necessary for movement (though its absence prevents pulsatile motion); rather, plastic flow evacuates material from the front of the bead. In our model, direction is also maintained by the tail, which reinforces the network at the rear of the bead, but this works by concentrating circumferential tension at the unreinforced zone at the front, leading to a sustained rip.
We find that the Elastic Gel model helps explain the sideways symmetry breaking and motility of capsule-shaped and ellipsoidal nucleators. The network stretches around the long axis to relieve the circumferential tension, so only around the short axis does tension buildup cause symmetry breaking (and motility) in the sideways direction. Our experiments using ellipsoidal beads confirm this behavior in vitro, and support the elastic gel mechanism as the determinant of symmetry breaking and motility behavior.
We show that for lengthwise symmetry breaking and motility, a capsule geometry requires asymmetric nucleation. Wild-type Listeria is capsule-shaped, moves lengthwise, and has such an asymmetric distribution of its ActA nucleation factor [45],[46], but a deletion mutation of ActA has been identified that results in a “skidding” sideways motion of Listeria in vivo [47]. Our data raise the possibility that the effect of this mutation could be to alter the asymmetric distribution of ActA activity.
Simple models such as ours have limited scope—e.g., we do not include filament-specific effects such as filament orientations and elongation by monomer addition—so we cannot evaluate the Brownian ratchet mechanism, nor can we investigate the hollow tails seen for beads coated with VASP [27], or recreate the nano-saltations observed in vitro [43]. The first 3-D computer simulation of actin-based Listeria motility took a detailed approach, simulating the behavior of large numbers of individual actin filaments and branches [29]. The Alberts-Odell model provided an important insight into the connection between the microscale behavior of individual filaments and larger-scale behavior of motile networks, namely how the buildup and breakage of filament-load attachments can produce nano-saltations in motility similar to those observed experimentally [43]. As with our model, the Alberts-Odell model has limited scope. To make their model computationally tractable, Alberts and Odell modeled actin filaments as inflexible rods, fixed rigidly in space soon after nucleation. Thus, the actin network in their model is an inelastic solid and could not be used to study processes involving elastic energy storage, plastic deformation, or mechanical failure: e.g., the Alberts-Odell model could not be used to study mechanical symmetry breaking or the role of elastic recoil in sustained motility. Concentrating on different aspects of the system, the two models complement one another and explain a wider range of behaviors.
Our approach has been to use a simple model with few parameters that confers strong explanatory power at the risk of oversimplifying the physical mechanisms. One potential oversimplification in our model is the constancy of conditions: e.g., we assume no changes in polymerization rate over time or spatially over the bead surface. The concentrations of components change during the reaction, and although this does not affect the rate of actin polymerization in the shell in vitro [16], this does not mean it does not affect more subtle physical characteristics of the network architecture. We also know that Arp2/3-based actin nucleation is autocatalytic [48], which might bias polymerization to the rear of the bead where there is a higher density of existing actin and help maintain directional motion. Our simulations include the code to implement such processes, but we have deliberately not used them in the current study (Ockham's razor). This allows us to show that we can explain the behavior of the system using viscoelastic mechanical effects alone.
The goal of this simulation has been to demonstrate the qualitative mechanisms of symmetry breaking and motility, and we have stressed that our simulations do not produce calibrated physical quantities for force, speed, etc. To do so would require both kinetically tuning the model to a more extensive experimental dataset, and also to include a more sophisticated treatment of internal network friction. The current model treats drag very simply: the system is over-damped, with drag proportional to velocity relative to the reference frame, consistent with a low Reynolds number regime. This explains a significant deviation between our model and our experimental data: that the rapid recoil of the shell in symmetry breaking is slower in our simulations. Would kinetic tuning significantly alter the qualitative behavior of the model? There are two reasons to think not. First, most of the kinetics are close to observed (e.g., the ratio of polymerization rates to rates of shell buildup, relaxation, bead movement, etc. are similar), so adjustments should not be major, and therefore, would be unlikely to affect the qualitative behavior. Second, even during the rapid recoil of the shell when the kinetics are dissimilar, the equilibrium states match well—i.e., the close match in the shapes of the curves shown in Figure 3C and 3E suggest that both the in vitro and in silico systems are relaxing from the same initial to same final states, and therefore, are driven by the same processes.
We have aimed to include as few parameters as possible, and although we make no claims that these parameters correspond to calibrated physical units of the in vitro network, an important question is how their values are chosen and how critical these choices are to the behavior. Essentially, we arrived at values that qualitatively reproduce the behaviors of the in vitro system by systematically exploring the effects of varying the model parameters, e.g., in Model Robustness, Section S3 of the supporting text (Protocol S1). Some behaviors (e.g., symmetry breaking, directional motion) are extremely robust, whereas others, such as smooth motion, are fragile and are disrupted by varying many different parameters.
Working with simulations allows us to refine the hypotheses. Full access to the behavior of the in silico system allows us separate out the gross morphological changes measured in vitro, e.g., the 2-D shape of the final shell, from the underlying components of the motion, e.g., circumferential squeezing, but no orthogonal squeezing, to refine our ideas about the underlying mechanisms. Furthermore, simulations allow us to directly test whether the proposed mechanisms are required for the motion or are epiphenomena, for example, by producing networks in silico that do not have elastic recoil effects and seeing that motion is essentially unchanged.
The APS model demonstrates how the simple viscoelastic properties of the in silico reconstituted actin gel can give rise to the observed dynamics of symmetry breaking and steady and pulsatile motility of spherical, capsule-shaped, and ellipsoidal objects coated with actin-nucleation factors. The model demonstrates both explanatory and predictive power in these areas, e.g., explaining how a pressure-dependent change in gel properties allows for a transition between motility regimes and predicting the 3-D geometry of in vitro shells.
In the future, we plan to refine the model, calibrating it with time, length, and force data to allow quantitative estimates of internal actin network parameters that are not directly measureable. For example, excising a cubic “slab” of a calibrated nodes-and-link network, then performing “computer experiments” by compressing, stretching, and shearing this slab in silico and recording the resulting stresses will allow us to compute the effective macroscopic elastic moduli of the in silico network, including Young modulus and Poisson ratio. More experimental data will also allow refinement of the functional forms of the repulsive and link forces, and to determine the extent that polymerization is regulated by force.
The APS model also offers a general framework to help investigate other physical cell phenomena that may be dominated by similar, relatively simple viscoelastic behaviors, e.g., lamellipodia and pseudopodia extension and cell septation, by including the effects of interactions with cell membranes, and simulating the anisotropic networks and contractile proteins found in vivo.
A brief overview of the model is given here (more details are available in the supporting text (Protocol S1) Sections S1 and S4, S5, S6, S7, S8). We simulate the network using a discrete-element approach, i.e., the actin network is represented as network of nodes in 3-D space held together by links (Figures S1 and S2). This is unlike a finite element approach in which the mesh is a way to reduce the dimensionality of a continuum problem into finite number of equations (elements). Rather, network links and the effective mesh size that results are important properties of the network. Network links also have no direct correspondence to actin filaments, but rather the bulk viscoelastic properties of the network of links and nodes are intended to capture the bulk viscoelastic network properties of the actin network. Under the polymerization conditions used (i.e., in the absence of crosslinking proteins) nodes more properly correspond to entanglement of filaments, and links correspond to the elastic properties of the network. We model these links as simple linear springs with a defined breaking strain and an inverse square repulsive force between nodes that models the compression resistance of the material. We explicitly avoid the unresolved question of how polymerizing filaments behave on a molecular level at the nucleator surface (according to Brownian ratchet or other models [24],[49]), and model polymerization as the stochastic introduction of material (nodes) at constant rate at the nucleator surface. Simulations begin at t = 0 with zero nodes (and links). Once introduced, new nodes form links with their neighbors, with a higher probability of forming links with nearby nodes (linear tail-off with distance, max probability PXL at zero distance), and a limit on the maximum number of links. Nodes at the surface of the bead are also linked to the bead at their last contact point by a link with force proportional to its length. Forces are calculated iteratively (Figures S3 and S21), and since this is a low Reynolds number regime, there is no inertia (i.e., velocity is proportional to force.)
The computational model is implemented in C++, and run times to symmetry breaking are approximately 1–2 h on a typical desktop computer. The code is designed to use multiple threads to enable large-scale problems to be explored across a number of parameter regimes (runs typically involve 105 nodes, 106 links, and 106 iterations per simulation). The code is open source and made freely available under the GNU General Public License to allow the results to be reproduced, to convey the full details of the model, and to encourage further use of the code by other researchers. A snapshot of the source code together with the parameter control file (Protocol S2) and a compiled executable for Mac OS X (Protocol S3) are provided. A detailed explanation of the code and the parameter control file are included in the supporting text (Protocol S1) and in an online wiki at http://www.dayel.com/comet, where the latest version of the code can also be downloaded.
To visualize the results of the simulations in a way comparable to in vitro microscopy images, we calculate the symmetry breaking plane, and create a 2-D projection of the nodes of the network convolved with a Gaussian to represent the point spread function of the microscope. To make visual comparison easier, we rotate the reference frame afterwards so that the bead always appears to move to the right. Measurements of forces in the radial and circumferential directions in Figure 4 are calculated as components in the direction of, or perpendicular to, a vector from the bead center, the magnitudes of which are summed over spherical shells of different radii. “Stretch factor” measures in Figures 4 and 5 are calculated by measuring the distance between particular pairs of nodes over time, normalized to the initial distance then averaged.
Bead motility experiments were carried out as previously described [16], with modifications. Briefly, 5-µm diameter carboxylated polystyrene beads (Bangs Laboratories) were covalently coated with ActA. The motility mix contained 0.5 mM ATP, 1 mM MgCl2, 1 mM EGTA, 15 mM TCEP-HCl, 50 mM KOH (to neutralize TCEP-HCl), 20 mM HEPES (pH 7.0), 125 nM Arp2/3 complex, 100 or 120 nM capping protein, and 3 µM actin. To aid microscopic observation, we included 3 mg/ml BSA (A0281; Sigma-Aldrich) and 0.2% methylcellulose (M0262; Sigma-Aldrich). Initial attempts to define headspace by controlling reaction volume were unsuccessful—the coverslip was not perfectly parallel to the slide, causing the headspace to vary across the sample—so we controlled the headspace by adding 0.1% v/v 5.1-µm or 15.5-µm diameter glass spacer beads (Duke Scientific) prior to starting the reaction. For 3-D reconstructions, reactions were stopped before imaging by adding 50% volume of 15 µM phalloidin and 15 µM Latrunculin B (Sigma-Aldrich). Fluorescent speckle microscopy (Figure 3A) conditions: 7.5 µM actin (1/3,000 TMR-labeled), 3 µM profilin, 40 nM Arp2/3, and 56 nM capping protein.
For the ellipsoidal bead experiments, spherical beads were stretched as previously described [50] with the following modifications: 140 µl of polystyrene bead stock was suspended in 6 ml of 3.8% w/v suspension of polyvinyl alcohol (PVA). The PVA/bead suspension was degassed before casting films in a 4.5×7.0 cm leveled tray. After stretching, the PVA was dissolved by incubating at 90°C for 2 h in distilled water containing 0.1% NP-40. The beads were washed three times in isopropanol and dried in a rotary evaporator. The bead surface was refunctionalized by incubation in 50% (w/v) NaOH for 1 h at 90°C and overnight at 42°C, washed once with 20 mM Tris HCl (pH 8.0) and 0.1% NP-40, and three times with 0.1% NP-40 before coating with ActA.
|
10.1371/journal.pntd.0006991 | Potential inconsistencies in Zika surveillance data and our understanding of risk during pregnancy | A significant increase in microcephaly incidence was reported in Northeast Brazil at the end of 2015, which has since been attributed to an epidemic of Zika virus (ZIKV) infections earlier that year. Further incidence of congenital Zika syndrome (CZS) was expected following waves of ZIKV infection throughout Latin America; however, only modest increases in microcephaly and CZS incidence have since been observed. The quantitative relationship between ZIKV infection, gestational age and congenital outcome remains poorly understood.
We characterised the gestational-age-varying risk of microcephaly given ZIKV infection using publicly available incidence data from multiple locations in Brazil and Colombia. We found that the relative timings and shapes of ZIKV infection and microcephaly incidence curves suggested different gestational risk profiles for different locations, varying in both the duration and magnitude of gestational risk. Data from Northeast Brazil suggested a narrow window of risk during the first trimester, whereas data from Colombia suggested persistent risk throughout pregnancy. We then used the model to estimate which combination of behavioural and reporting changes would have been sufficient to explain the absence of a second microcephaly incidence wave in Bahia, Brazil; a population for which we had two years of data. We found that a 18.9-fold increase in ZIKV infection reporting rate was consistent with observed patterns.
Our study illustrates how surveillance data may be used in principle to answer key questions in the absence of directed epidemiological studies. However, in this case, we suggest that currently available surveillance data are insufficient to accurately estimate the gestational-age-varying risk of microcephaly from ZIKV infection. The methods used here may be of use in future outbreaks and may help to inform improved surveillance and interpretation in countries yet to experience an outbreak of ZIKV infection.
| Zika virus (ZIKV) infection is associated with the rise of microcephaly cases observed in Northeast Brazil at the end of 2015. For women in endemic or at-risk areas, understanding how the relationship between time of infection and microcephaly risk varies through pregnancy is important in informing family planning. However, a relatively modest number of congenital Zika syndrome cases have been observed following subsequent waves of ZIKV infection, limiting our understanding of gestational risk. We used a mathematical model to quantify the shape and magnitude of the gestational-age-varying risk to a fetus. Although the risk profile should be conserved regardless of location, we estimated different profiles when using surveillance data from locations in Northeast Brazil and Colombia. Our results suggest that time-dependent reporting changes likely confound the interpretation of currently available surveillance data. Furthermore, we investigated a range of behavioural and reporting rate changes that could explain two waves of ZIKV infection in Bahia, Brazil despite only one wave of microcephaly. Plausible changes in reporting could explain these data whilst remaining consistent with the hypothesis that ZIKV infection carries a significant risk of microcephaly. Further evidence is needed to disentangle the true risk of congenital Zika syndrome from time-varying reporting changes.
| A substantial body of experimental and clinical evidence implicates Zika virus (ZIKV) infection in the sharp rise in the incidence of microcephaly cases in Brazil at the end of 2015. [1–5] Previous population-level studies investigating the relationship between ZIKV and microcephaly incidence found consistent patterns of high first-trimester risk and lower risk later in pregnancy, which is consistent with early clinical findings for ZIKV-associated microcephaly. [6, 7] However, clinical studies investigating the link between ZIKV infection and a distinctive pattern of congenital abnormalities, collectively termed congenital Zika syndrome (CZS), suggest that adverse outcomes are associated with ZIKV infection throughout pregnancy. [8–10] How these clinical findings link to the complex picture portrayed by the epidemiological data in Brazil is still unclear and leaves a substantial knowledge gap for those counseling pregnant women in ZIKV-affected populations. For example, why did the majority of Latin America demonstrate a relatively small rise in microcephaly incidence rates compared to those seen in Northeast Brazil, and why was the second wave of microcephaly in Brazil much smaller than the first despite two similar waves of Guillain-Barré Syndrome (GBS)? [11]
It is useful to consider the conceptual model in which observed population-level CZS incidence reflects underlying ZIKV transmission dynamics and a gestational-age-varying risk of CZS given infection with ZIKV. If a pregnant woman is infected at some point during gestation, her baby may present as a case of CZS with a probability conditional on the gestational age of her baby when infection occurred. Characterizing this link between underlying gestational risk of CZS and its presentation in epidemiological data has two potential benefits. First, an estimate of the underlying gestational risk profile from surveillance data may provide evidence to inform women of childbearing age to help plan pregnancy and mitigate exposure risk. [12] Second, if the risk profiles differ substantially between populations, those differences could support the study of alternative hypotheses of risk factors for CZS beyond ZIKV infection. For example, prior infection with another arbovirus has been suggested as a potential cofactor for risk of GBS, however prior arbovirus infection has not yet been shown to play a role in increased neurological adverse event risk. [13]
Research on other teratogenic pathogens shows the potential importance of gestational age to CZS risk. [14] For example, prospective cohort studies of pregnant women have shown that infection early in gestation greatly increases the risk of congenital rubella syndrome and cytomegalovirus-associated adverse fetal outcomes relative to infection later in pregnancy. [15, 16] Although a similar pattern seems likely to be the case for CZS, the timing and magnitude of risk throughout pregnancy remains uncertain. However, quantifying this underlying gestational-age-varying risk profile should be possible given reliable data on infection and CZS incidence combined and a robust statistical approach. Here, we demonstrate how a transmission model fitted to reported incidence data can be used to infer the relationship between gestational age at the time of ZIKV infection and the risk of microcephaly. We adapt this model to explore potential explanations, including changes in reporting rates and abortions, for the lack of a second observed wave of microcephaly incidence in Brazil.
We searched the literature, Pan American Health Organization (PAHO), World Health Organization (WHO) and Brazilian state health authority websites for reports of suspected or confirmed ZIKV infection incidence and microcephaly cases in 2015 and early 2016, building on a comprehensive literature search performed in 2016. [17] In particular, we searched www.paho.org, www.who.int, Brazilian state-level ministry of health websites (eg. www.suvisa.ba.gov.br), and PubMed for the terms “zika” and “microcephaly”. Where confirmation status of cases was not recorded (eg. where incidence was only shown as “reported cases”), we classified these data as “notified” cases. Where suspected and confirmed cases were distinguished, we classified the sum of suspected and confirmed cases as “notified” cases. A summary of data included in the analyses can be found in S1 Table.
Coverage, case definitions and protocols for ZIKV and microcephaly surveillance in Brazil and Colombia changed throughout 2015 and 2016, making direct comparison of incidence data between these years difficult. [18] Although the first outbreak of laboratory confirmed ZIKV was reported in Brazil on 07/05/2015, ZIKV reporting only became compulsory through the national notifiable information system (SINAN) on 17/02/2016. [19, 20] Furthermore, laboratory confirmation was only performed on suspected cases from previously unaffected areas and on certain subpopulations of interest (eg. pregnant women, hospitalised patients with neurological complications). Reporting of microcephaly and other congenital abnormalities was through the information on live births system (SINASC) until November 2015, when the new public health events registry (RESP) was implemented to improve surveillance in pregnant women and newborns. [21] In Colombia, passive reporting of ZIKV cases and major congenital abnormalities (including microcephaly) is ongoing through the National Health Institute (INS) surveillance system. [22, 23] Laboratory testing and mandatory reporting of ZIKV infections based on clinical symptoms began on 14/10/2015. [24] However, RT-PCR confirmation was not systematic and, like Brazil, was only used to confirm the presence of ZIKV in municipalities that had not yet detected the virus and in subpopulations at particular risk of complications. Therefore, the incidence of confirmed ZIKV infections gives an indication of the spread of the virus, but not necessarily the true magnitude and dynamics of the epidemic within affected locations.
Some data sources were only available in graphical form, and these numbers were therefore extracted using a web digitiser (https://automeris.io/WebPlotDigitizer/). The results presented in the main text used data from: Northeast Brazil; Colombia; the city of Salvador, Bahia, Brazil; and state-reported incidence from Bahia, Rio Grande do Norte and Pernambuco, Brazil. Numbers of live births were obtained for Brazil from the SINASC/CGIAE/SVS/MS system. [7, 25] For Colombia, live births were obtained from a publication of microcephaly and ZIKV infection incidence in Colombia and from Colombian ministry of health vital statistics. [23, 26]
We developed a two-component model to describe the relationship between the incidence of ZIKV infection and the incidence of microcephaly-affected births, depicted in Fig 1. Full details of the model and sensitivity analyses are available in S1 Text and all code and analyses are available as an R package (https://github.com/jameshay218/zikaInfer). Our aim was to estimate the shape and size of the risk window for developing ZIKV-associated microcephaly given infection during gestation, and to test for differences in inferred risk using data sets from various Brazilian states and Colombia. The first component of the model described the transmission dynamics of ZIKV via the Aedes aegypti mosquito vector based on the Ross-MacDonald model for vector-borne disease. [27] Through estimation of the force of infection over time, we estimated a per capita risk of human infection per unit time, PI(t). The second component of the model described the risk of a fetus developing microcephaly given that the mother was infected in a particular week during pregnancy using a modified gamma function, P m ′ ( t ). The expected proportion of microcephaly-affected births at any time t (Fig 1E) was obtained by multiplying these two components together:
P m ( t ) = ∑ i = t - 40 t P I ( i ) P m ′ ( i - t + 40 ) (1)
where Pm(t) is the probability of a ZIKV-associated microcephaly birth at time t, PI(i) is the probability of an individual becoming infected at time i (and not before), and P m ′ ( i - t + 40 ) is the probability of a fetus developing microcephaly given ZIKV infection at gestational week i − t + 40. Including a baseline microcephaly rate gives the probability of observing any microcephaly case at time t as:
P m i c r o ( t ) = ϕ i [ P m ( t ) + P b - P b P m ( t ) ] (2)
where Pb is the baseline per birth microcephaly incidence rate and ϕi is a multiplicative factor for the number of true cases that were reported in location i (less than one indicates underreporting, greater than one indicates overreporting).
Model parameters governing ZIKV transmission were obtained from the literature as described in S2 Table. Fixed parameters were mainly chosen based on a previously published transmission model resulting in a generation time (ie. the length of time between the time of infection in a human case and the times of infection in the secondary human cases resulting from that case) of approximately 20 days. [17] Given a fixed generation time, the model allowed the shape of the incidence curve for number of infected individuals (ie. the number of individuals entering the I compartment of the SEIR model) to vary depending on the value of the basic reproduction number, R0. As R0 is comprised of multiple correlated parameters, all components of R0 other than the vector density per human were fixed. Vital statistics (human life expectancy and population size) for particular locations are described in S3 Table.
We calculated the combined likelihood of observing ZIKV and microcephaly incidence data conditional on the model parameters, as described in S1 Text. Where full ZIKV infection incidence data were not available (ie. Pernambuco, which only reported 3 weeks of ZIKV infection incidence in the first half of 2015), we fit the two-component model to microcephaly incidence data alone and placed a 4-month-wide window on the time of peak ZIKV infection incidence given by the model as a uniform prior centered on the time of peak reported ZIKV incidence in Pernambuco (ie. the peak of the three reported weeks). We fit the model separately to available notified and confirmed incidence data for each location using a Markov chain Monte Carlo (MCMC) framework written in R and C++ with ordinary differential equations solved used the rlsoda package. [28, 29] By examining the posterior distributions of the parameters that determined the gamma risk profile, we also estimated the following parameters of interest: first week of gestational age where ZIKV infection confers a risk of microcephaly greater than 1 in 1000; last week of gestational age where ZIKV infection confers a risk of microcephaly greater than 1 in 1000; number of gestational weeks spent at risk; mean per-trimester risk; gestational week of greatest risk (gamma mode).
We carried out a sensitivity analysis using seroprevalence data from the city of Salvador in Bahia, Brazil to infer the microcephaly risk profile without the SEIR component of the model. Here, we assumed that reported acute exanthematous illness (AEI) was proportional to the true incidence of ZIKV infection during this time, and scaled the weekly reported incidence to give a final attack rate of between 59.4% and 66.8% in line with ZIKV IgG seroprevalence estimates for Salvador in May 2016 based on an NS1 antigen ELISA. [30, 31]
We included four model parameters to quantify potential changes in behaviour and reporting rates that could explain the two seasons of observed data in Bahia, Brazil, where only one wave of microcephaly incidence was observed despite two waves of ZIKV infection incidence. Time-dependent changes in behaviour and reporting were likely during the epidemic due to media hype and public awareness, demonstrated by changes in Google search behaviour shown in Fig 2A. First, we assumed that microcephaly reporting became 100% accurate from March 2016 (the most recent change in case definition in Brazil) and estimated the relative reporting rate prior to this as a model parameter. [18, 32] Second, we assumed that immediately following the WHO declaration of a Public Health Emergency of International Concern in February 2016, the rate of aborted pregnancies under 24 weeks gestation could have changed. [33–35] Third, we assumed that the number of ZIKV-affected births after this date may have changed, either due to avoided pregnancies or additional precautions taken by pregnant women to avoid infection relative to the rest of the population. [34, 35] Finally, we assumed that ZIKV infection reporting accuracy may have changed after 11/11/2015, when the Brazilian Ministry of Health declared a National Public Health Emergency, and just before WHO/PAHO issued an alert with laboratory detection guidelines for ZIKV. [36, 37]
Our model did not explicitly include seasonality, and the SEIR model was therefore only suitable for the single-season analyses. We therefore did not use the SEIR model component in the multi-season analysis, but rather assumed that the per capita risk of becoming infected with ZIKV was proportional to reported ZIKV infection incidence and that reported incidence of ZIKV infection represented a fraction of true cases scaled by one parameter up to 11/11/2015 and another from 11/11/2015 onwards.
Although our study data were from similar reporting systems, Fig 3 illustrates substantial differences in the key features of both ZIKV infection and microcephaly incidence patterns. Peak timing, width of the incidence curves, maximum per capita incidence and the lag between ZIKV infection and microcephaly incidence peaks differed by location and dataset. Variation in total and maximum per-birth microcephaly incidence indicates location-specific differences in the proportion of pregnant women that were infected with ZIKV or in the probability of developing microcephaly following infection. For example, weekly notified microcephaly incidence peaked at 32.6 cases per 10,000 births in Colombia, which was far lower than peak notified microcephaly incidence in Pernambuco, Brazil at 760 cases per 10,000 births, suggesting that microcephaly risk given infection and/or ZIKV attack rates were higher in Pernambuco. The lag between incidence peaks also varied, ranging from 23 weeks (bootstrapped confidence intervals: 19-32 weeks) for Colombia compared to 31 weeks (bootstrapped confidence intervals: 30-36) from state-level reports for Bahia, Brazil. Given that only a small fraction of the Colombian population is at risk of arbovirus infection compared to the Brazilian population due to differences in vector ecology, it is unsurprising that the absolute per capita incidence of ZIKV infection and microcephaly are lower in Colombia than in Brazil. [22] However, as the population at risk of ZIKV infection is the same population at risk of microcephaly, we would expect the lag and relative magnitudes of ZIKV infection and microcephaly incidence to be the same between Brazil and Colombia. These differences in time lags and relative magnitudes therefore suggest that the time of peak risk during pregnancy may have varied between locations, potentially through differences in additional risk factors in these locations such as prior arbovirus exposure.
Differences in observed incidence patterns may also arise as a result of reporting bias, which may be reduced using confirmed rather than notified case data. Although some confirmed microcephaly case data were available for Brazil, the only available data of confirmed microcephaly cases in Colombia reported number of cumulative confirmed cases. [39] It should also be noted that only a subset of suspected cases were laboratory confirmed in Colombia to test for the presence of ZIKV in municipalities which had yet to confirm ZIKV infection, and case reporting was otherwise based on clinical symptoms. [24] Due to variation in confirmation delays, we were unable to extract the time of birth of these cases and were therefore unable to use these data in model fitting. Between 03/01/2016 (epidemiological week (EW) 1 of 2016) and 05/06/2017 (EW 18 of 2017), approximately 70% of notified microcephaly cases were discarded in Colombia (328 cases confirmed, 874 discarded, 37 under investigation), highlighting that total notified case data likely overestimates true incidence. [40] The proportion of total notified cases that were discarded was similar for Rio Grande do Norte (138 confirmed vs. 475 notified) and somewhat higher for Pernambuco (365 confirmed vs. 2117 notified). Confirmed ZIKV infection incidence were available for Colombia but not Brazil due to the lack of reporting infrastructure during the first wave. [41] The lag between peak ZIKV infection and microcephaly incidence did not change when using notified or confirmed ZIKV infection or microcephaly data.
We inferred interpretable gestational risk profiles for 5 of the 6 datasets used here (Fig 4). Though clear estimates of the gestational-age-varying risk were obtained for each location, substantial differences are apparent in the inferred gestational age of peak risk, duration of the risk period, and maximum absolute risk. The model did not produce a biologically interpretable risk profile using data from Pernambuco, Brazil that was comparable to those inferred using the other 5 datasets. Sensitivity analyses excluding the ZIKV infection incidence data from the model fitting for other locations were able to produce plausible risk profiles, suggesting biases in reported microcephaly incidence data for Pernambuco that could not be explained by the model (see Section 4.3, S1 Text). It is important to note that microcephaly is only one manifestation of CZS, and the risk profile of other adverse outcomes may differ. These risk estimates therefore apply only to the specific outcomes of proportionate and disproportionate microcephaly, which were not distinguished in these data. [8]
The time from the peak of ZIKV infection incidence to the peak of microcephaly incidence indicates the typical gestational age at which microcephaly cases were infected. When using both state and city level notified ZIKV infection and microcephaly incidence from Bahia, Rio Grande do Norte, and Salvador, Brazil, the peak week of gestational-age-varying risk was estimated to be in the middle of the first trimester (Fig 4). When notified ZIKV infection incidence in pregnant women and confirmed microcephaly incidence from Northeast Brazil were used, the estimated peak gestational-age-varying risk was towards the end of the first trimester. Notified case data from Colombia were also suggestive of peak risk in the first trimester. Inference did not change substantially using confirmed as opposed to notified ZIKV infection incidence data for Colombia; however, using confirmed microcephaly incidence data for Rio Grande do Norte resulted in a shift of the risk profile towards the start of the second trimester. When a ZIKV infection incidence peak in a 4-month window around March 2015 was assumed for Pernambuco, the inferred microcephaly risk profile was highly skewed towards the first week of pregnancy, suggesting that these data are incompatible with the other 5 data sets.
There was substantial variation in the inferred window of heightened gestational risk between different populations. The window of heightened gestational risk is estimated from the relative durations of the ZIKV infection and microcephaly incidence curves (using an illustrative threshold of 1 in 1,000 infections leading to microcephaly to define the heightened gestational risk window). A narrow period of ZIKV infection incidence preceding a wide period of microcephaly incidence suggests a wide window of heightened gestational risk. If the period of heightened gestational risk is long, then infections at a particular point in time would present as cases of CZS across a wider interval of birth dates. Inferred risk profiles using notified case data from Rio Grande do Norte, the city of Salvador and Colombia all suggested heightened risk throughout pregnancy (Fig 4).
Conversely, two similarly narrow (or wide) periods of ZIKV infection and microcephaly incidence would suggest a relatively small window of heightened gestational risk, as all ZIKV-affected pregnancies would present as births after a similar delay. A true microcephaly incidence period that is narrower than the ZIKV infection incidence period should not be possible, as the narrowest microcephaly incidence curve would arise when all infected pregnant women give birth after the same delay. Aggregated confirmed case data from Northeast Brazil, state-level notified case data from Bahia and state-level confirmed case data from Rio Grande do Norte all suggested a more limited window of risk during pregnancy, with lower risk suggested towards the end of pregnancy (Fig 4, Northeast Brazil, Bahia and Rio Grande do Norte (confirmed cases)).
Public awareness, media hype, changing criteria for case reporting and variation in laboratory testing capacity likely resulted in changing reporting rates throughout the epidemic. [18, 41, 48, 49] Location-specific time-varying changes in reporting sensitivity and specificity are therefore one potential explanation for differences in the risk profiles inferred using data from Northeast Brazil and Colombia. Given that Colombia was expecting an increase in microcephaly cases during 2016, an increase in notified cases may have been reported before a true increase in confirmed cases, which would falsely suggest some gestational risk late in pregnancy. Time-varying reporting bias may also explain the extremely narrow and early window of risk inferred using data from Pernambuco (Fig 4, Pernambuco, Brazil). The impact of reporting bias is clearly demonstrated by the contrasting results using confirmed or notified microcephaly case data for Rio Grande do Norte, wherein confirmed data suggested a narrower and later risk window than the notified data.
The absolute risk of CZS is more difficult to estimate as it depends on the true incidence of ZIKV infection in pregnant women and CZS cases as a proportion of live births. A high ZIKV infection attack rate with known microcephaly incidence would suggest a lower microcephaly risk per infection to the fetus than a low infection attack rate with the same observed microcephaly incidence. [7] Reported infection incidence data may be subject to under-reporting and over-reporting, potentially through missing asymptomatic or mild cases that might not present to surveillance systems (under-reporting), or misclassifying infections caused by other arboviruses as ZIKV infection, namely dengue and chikungunya virus (CHIKV) (over-reporting). [41, 50] These confounders present identifiability problems in inferring levels of true incidence and therefore microcephaly risk; surveillance data in a scenario of high risk with under-reporting would be similar to a scenario of low risk with over-reporting. For example, during the 2015 wave in Brazil many cases of illness likely caused by ZIKV were misclassified as dengue infection, resulting in under-reporting of ZIKV infection incidence. [41] Over-reporting of microcephaly incidence during the initial wave of cases was also possible, due to changing case definitions, reclassification of suspected cases and increased awareness in surveillance systems. [18, 32] Estimating the proportion of true ZIKV infections that led to observed microcephaly cases is therefore dependent on knowing the true risk of ZIKV infection during the epidemic period. ZIKV IgG seroprevalence was estimated to have reached 63.3% (95% confidence interval, 59.4 to 66.8%) in Salvador, Brazil between 2015 and 2016 despite only 16,986 reported cases of AEI from a population of nearly 3 million (approximately 0.6%), suggesting that under-reporting of ZIKV infection incidence was a key problem in this location. [30, 31]
By assuming that 100% of true microcephaly cases were reported but that reported ZIKV cases represented only a fraction of the true incidence, we inferred the absolute risk of ZIKV-associated microcephaly from each of the datasets (S4 Table). The average first trimester risk of microcephaly given ZIKV infection was estimated to be 2.81% (mean; 95% credible interval (CI): 2.51-3.16%) based on data from Bahia, Brazil, but much lower in the second trimester at 0.365% (mean; 95% CI: 0.0715-0.588%). Conversely, the level of absolute risk estimated using notified case data from Colombia suggested that the risk was lower but consistent throughout gestation at 0.303% (mean; 95% CI: 0.239-0.367%), 0.268% (mean; 95% CI: 0.228-0.322%) and 0.186% (mean; 95% CI: 0.135-0.232%) in the first, second and third trimesters respectively. The former estimate is slightly higher than risk estimates inferred based on seroprevalence data from French Polynesia which suggested a risk of 0.95% (95% confidence interval; 0.34–1.91%) in the first trimester, whereas the latter estimate suggests a lower risk. [6]
We performed a sensitivity analysis with better constraint on the true ZIKV attack rate by taking microcephaly and AEI data from Salvador, Brazil for 2015 scaled by recent ZIKV IgG seroprevalence data, as described in Section 6, S1 Text. [30] Here, we assumed that the true risk of ZIKV infection in Salvador was proportional to the per capita reported incidence of AEI scaled such that the overall attack rate was between 59.4% and 66.8%. [31] Based on the ZIKV infection and microcephaly incidence data from Salvador, Brazil, we estimated the mean first trimester risk of microcephaly given ZIKV infection to be 3.06% (mean, 95% CI: 2.66-3.49%); the mean second trimester risk to be 0.805% (mean, 95% CI: 0.649-0.980%); and the mean third trimester risk to be 0.0833% (mean, 95% CI: 0.0407-0.142%). We did not scale incidence data for any other location due to the lack of seroprevalence data. However, given that the model is powered by the pattern of microcephaly incidence relative to the pattern of ZIKV infection incidence after accounting for differences in infection risk and reporting, these risk estimates may apply to other locations if no additional cofactors affect the risk of microcephaly given infection.
Despite a clear second wave of GBS incidence at the beginning of 2016, no second wave of microcephaly incidence in Northeast Brazil was observed in the latter half of 2016. [11] Similar to [11], Fig 2B illustrates the incidence of microcephaly that would have been expected in Bahia, Brazil using our model framework and based on reported ZIKV infection incidence under the assumption that the underlying gestational-age-varying risk profile and reporting behaviour did not change from 2015 to 2016.
We used the population-level data fitting framework described above to test the hypothesis that plausible changes in behaviour or reporting are sufficient to provide a consistent narrative between the two waves of ZIKV and microcephaly case data. Fig 2A describes the timings of particular events that may have led to these changes. We considered four hypotheses describing changes in behaviour and reporting rates. First, we assumed that microcephaly reporting accuracy may have been different before week 11 of 2016 (13/03/2016, the most recent change in case definition in for microcephaly reported through the Registro de Eventos em Saúde Pública (RESP) database in Brazil) [32, 51] and estimated the relative reporting rate for microcephaly prior to this that would be consistent with the observed data. Second, we assumed that immediately following the National Public Health Emergency announcement by the Brazilian Ministry of Health on 11/11/2015, the frequency of early abortions (up to 24 weeks gestation) due to early detection of CZS may have increased. [36] The earliest date at which targeted abortions would be observed as a drop in birth rate would be 16 weeks after this shift in behaviour (02/03/2016). [34, 35, 52] A reduction in birth rate from delayed pregnancy would also be possible; however, this would only appear approximately 40 weeks after the behavioural shift. Third, we assumed that the number of pregnant women affected by ZIKV after this date may have changed through additional precautions taken to avoid infection relative to the rest of the population. [53] Finally, we assumed that ZIKV reporting itself may have changed on 11/11/2015 before the start of the second wave of ZIKV infection incidence through increased surveillance, increased awareness and/or increased misclassification of other arbovirus infections as ZIKV infection. Over both time periods, we assumed that the per capita risk of becoming infected with ZIKV was proportional to reported ZIKV infection incidence, but that the scale of that proportion changed on 11/11/2015 following the potential change in ZIKV infection reporting.
Based on state-level reports from Bahia, Brazil and assuming that ZIKV infection reporting did not change, our analyses suggest that the lack of a second microcephaly peak could be explained by the combined effect of: a 151% reporting rate of microcephaly cases prior to 13/03/2016 relative to fixed 100% accurate reporting after 13/03/2016; targeted abortions ending 88.4% of microcephaly-affected pregnancies prior to 24 weeks gestation; and a relative decrease in infection probability in pregnant women of 0.60% (values shown are the maximum a posteriori probability (MAP) estimates). It is important to note that many of these parameters are highly correlated, suggesting that these data could be explained by a combination of multiple mechanisms, or by a greater contribution of some mechanisms and a reduced effect from the others (Fig 5). If ZIKV infection reporting accuracy increased substantially between the two waves in addition to the behavioural changes described above, then a smaller increase in the proportion of terminated pregnancies would have been necessary. Similarly, targeted abortions and precautions to avoid infection by pregnant women would present a similar reduction in microcephaly incidence, and these estimates are therefore highly correlated (Fig 5C).
Assuming that there were no targeted abortions, no additional precautions to avoid infection taken by pregnant women, and no change in microcephaly reporting accuracy, we estimated that these data could be explained solely by a 18.9-fold (mean, 95% CI: 10.0-59.1-fold) increase in ZIKV infection reporting after 11/11/2015. Conversely, assuming that targeted abortions after 11/11/2015 were the only change, 92.5% (mean, 95% CI: 89.8-94.9%) of microcephaly-affected births would need to have been aborted to explain the lack of a second peak, corresponding to 1090 (803-1480) aborted pregnancies between 02/03/2016 and 31/12/2016. Fig 5D shows how the total number of aborted microcephaly-affected births, which may be observable, would change with different abortion rates of microcephaly-affected births. If microcephaly reporting accuracy were the only factor to change, then a 601% (mean, 95% CI: 492-726%) reporting rate of microcephaly cases prior to 13/03/2016 relative to fixed 100% accurate reporting after 13/03/2016 would have been necessary. Accurate data on the true number of abortions in this time period and information on the changes in ZIKV and microcephaly reporting would help to clarify the relative contributions of these mechanisms.
Overall, these results highlight the limitations of currently publicly available population-level data in explaining epidemiological trends. Different datasets suggest different risk profiles, some of which contrast with previous population-scale analyses. Whilst data from Bahia, Brazil were suggestive of a risk profile similar to that estimated using data from French Polynesia, data from Colombia and Rio Grande do Norte, Brazil suggest a much longer gestational risk period. [6]
Although reporting bias may explain the differences in inferred microcephaly risk in different locations, heterogeneity in the distribution of additional host risk factors of microcephaly may be important. Interpretation of epidemiological data for dengue infection requires an understanding of pre-existing immunity due to the presence of antibody-dependent enhancement, which may also be relevant to the interpretation of CZS incidence given the potential role of dengue antibodies in ZIKV disease enhancement. [54–56] Observations of increased prior dengue exposure in areas of disproportionately increased microcephaly incidence would support this hypothesis and be of importance for dengue- but not yet ZIKV-affected areas, highlighting the need for comprehensive serological studies. [31, 57] An understanding of other potential host risk factors that may differ between affected areas, such as socioeconomic status or maternal smoking, will further aid the interpretation contrasting incidence data. [32]
A limitation of our model is the aggregation of data into high-level administrative units, which may mask small-scale heterogeneity in infection risk and case reporting. This may be particularly problematic in our analysis for Colombia, as using the entire Colombian population and birth numbers as the susceptible population may underestimate the true risk should only a fraction of the population actually be exposed to ZIKV infection. [58, 59] Similarly, differences in transmission peak times at a small spatial scale coupled with location-specific reporting accuracy may reduce the reliability of the population-wide inferred risk profile. Although we were unable to fit the model at a smaller administrative unit due to the lack of necessary meta-data for Colombia, doing so may reveal a similar risk profile to that estimated using data from Northeast Brazil.
Our estimates suggest that ZIKV infection reporting rates would need to have increased 18.9-fold (mean, 95% CI: 10.0-59.1-fold) to explain the lack of a second microcephaly wave in Bahia, Brazil on its own, which may have been possible if awareness and diagnostic accuracy improved through the epidemic. We note that syndromic ZIKV reports may have included misclassified CHIKV infections which may not have represented an increased risk of ZIKV-associated microcephaly during the second wave in 2016. [60] A 18.9-fold increase in ZIKV reporting as estimated here could therefore mean that ZIKV reporting was a more accurate representation of the true ZIKV attack rate in 2016, or that 18 Chikungunya cases were misclassified as ZIKV for every 1 true reported ZIKV case with no change in the proportion of true ZIKV cases that were reported. [41] However, during the period in which second waves of ZIKV infection occurred, there was sufficient virological testing to justify confidence in the relative specificity of reported ZIKV cases. [61] Furthermore, in Salvador, Brazil, where serological data are available, the increase in CHIKV seropositivity from 2015 to 2016 was far lower than for ZIKV seropositivity. [31] Nonetheless, diagnostic tools with improved sensitivity and specificity in distinguishing these infections would help to clarify the proportion of true ZIKV infection incidence that observed incidence data represent.
We estimated that 1090 (mean, 95% CI: 803-1480) microcephaly-affected births would need to have been aborted between 02/03/2016 and 31/12/2016 to explain the observed data through increased abortions alone. Given that approximately 1000 abortions are reported in Northeast Brazil weekly, it may be possible to identify the true increase in abortion rate during this time period if and when complete data become available (Supplementary Material of [11]). [35]
Estimating the true shape and magnitude of the underlying gestational-age-varying risk profile requires additional data that could either be gathered retrospectively or through surveillance in areas where the first wave of transmission is ongoing or has not yet happened. A key limitation of the epidemiological data gathered in Brazil during 2015 and early 2016 is that surveillance systems were implemented during the epidemic, leading to possible inconsistencies in case definitions and ascertainment rates. Retrospective regional serological surveys have been suggested previously as a means of inferring attack rates, which would constrain estimates for the reporting rate of microcephaly and ZIKV infection and in turn constrain estimates of both the underlying risk and potential changes in behaviour/reporting in the second wave. [60, 62] In particular, community seroprevalence studies of ZIKV antibodies in women of child-bearing age would provide an accurate estimate of the true proportion of ZIKV-infected women during the outbreak irrespective of symptomatic status and time of infection. In terms of future outbreaks, consistent and accurate case definitions for microcephaly and CZS,—such that sensitivity and specificity are high throughout the epidemic period—would greatly increase the utility of clinical surveillance data for population-level analysis.
A key remaining question is whether or not the epidemiological data from Brazil accurately represent the relationship between ZIKV infection and microcephaly, and indeed the wider set of outcomes associated with CZS. Retrospective cohort studies for women of childbearing age to assess whether changes in behaviour regarding conception and infection avoidance occurred in 2016 should clarify whether the second season of ZIKV/microcephaly in Brazil is fully consistent with estimates of gestational-age-varying risk from the first season. [53] If actual reporting rates and behaviour changes are not sufficient to explain the apparent discrepancy between first-wave incidence in Brazil compared to later and elsewhere, the investigation of other potential cofactors, such as prior arbovirus infection, becomes a higher priority. It should then be possible to accurately calculate the risk of CZS based on gestational age at infection and the presence or absence of other possible cofactors.
|
10.1371/journal.pcbi.1004825 | Cellular Growth Arrest and Persistence from Enzyme Saturation | Metabolic efficiency depends on the balance between supply and demand of metabolites, which is sensitive to environmental and physiological fluctuations, or noise, causing shortages or surpluses in the metabolic pipeline. How cells can reliably optimize biomass production in the presence of metabolic fluctuations is a fundamental question that has not been fully answered. Here we use mathematical models to predict that enzyme saturation creates distinct regimes of cellular growth, including a phase of growth arrest resulting from toxicity of the metabolic process. Noise can drive entry of single cells into growth arrest while a fast-growing majority sustains the population. We confirmed these predictions by measuring the growth dynamics of Escherichia coli utilizing lactose as a sole carbon source. The predicted heterogeneous growth emerged at high lactose concentrations, and was associated with cell death and production of antibiotic-tolerant persister cells. These results suggest how metabolic networks may balance costs and benefits, with important implications for drug tolerance.
| In bacteria, changes in gene expression, with resulting changes in protein concentration, can drastically change how fast cells and cellular populations grow. This fact has big implications for how we treat infectious disease, which types of organisms make up our microbiomes, and what patterns of gene regulation have undergone evolutionary selection. Here, we show how, in principle, the expression level of a single enzyme can affect bacterial population growth by creating a threshold where cells grow optimally fast just below it, but rapidly reach a state of no growth just above it because metabolic byproducts build up and halt growth. The narrow margin between these two states makes entering either of them possible for the same bacterium because of intrinsic uncertainty, or "noise", in gene expression. The predicted result is a variety of growth rates in a single population of genetically identical cells, manifested as a mix of fast- and slow-growing cells. We created laboratory conditions that reproduce the effect in the model organism E. coli, and showed that there may be a benefit to having slower growing cells, because they can survive antibiotic exposure for longer.
| Metabolism in single-celled organisms is subject to dynamic regulatory responses balancing cost and benefit [1–11]. Most single-celled organisms are under strong selective pressure to optimize growth in many conditions [12–23]. Nevertheless, some conditions exist where side effects of processing available metabolites slow growth. For example, mutations or rapidly changing conditions can result in toxic metabolic effects, including substrate-accelerated toxicity or even cell death. Examples include buildup of galactose derivatives arising from mutations in the Leloir pathway [24] and the lactose killing effect in bacteria [24–28] and yeast [29–31]. In the latter, byproducts of proton-catabolite symport in Major Facilitator Superfamily (MFS) permeases [32] are toxic [26, 33]. Defects in permease selectivity can result in excess sugar uptake causing excessive intracellular osmotic pressure [34].
Effects of metabolic stress have typically been considered at the population level, but recent findings suggest it may be important to consider the possibility that stress drives non-genetic variation between cells within a population. For example, in bacteria, metabolic starvation stress can induce toxin-antitoxin (TA) systems and resultant formation of non- or slow-growing antibiotic tolerant persister cells [35, 36]. This suggests a metabolic route to regulating cellular toxicity and growth arrest and raises the question of how metabolic cost and benefit affect population growth dynamics in the face of heterogeneity. In yeast, a thresholding effect in metabolism creates coexisting subpopulations of cells with different growth rates [31, 37]. Some signaling [38] and metabolic [39] pathways may have evolved to minimize intracellular noise of relevant protein or metabolite levels. However, single cells cannot control extracellular perturbations, and intracellular noise control is costly [40]. Multiple metabolic pathways in Escherichia coli have been found to operate close to the saturation point of their constituent enzymes [41], near an ultrasensitive threshold [42]. Beyond the threshold, intracellular metabolite concentrations rise sharply [42]. These studies suggest an important effect of intrinsically variable cellular metabolic states on cellular and population growth rates in the face of metabolic toxicity. However, the dynamics of cellular growth around metabolic thresholds is unknown.
In this study we developed mathematical and computational models of simple metabolic pathways. We determined the effects of pathway efficiency, cost, and demand on emergent cellular and population scale growth rates. Our results imply that, past a certain threshold, intracellular metabolite toxicity dominates growth kinetics, causing some cells to enter into a growth-arrested state while the rest of the population maintains fast growth. By growing E. coli cultures in varying levels of lactose, we confirmed cellular growth rate heterogeneity as theory predicts. Populations grown in conditions that produced a growth-arrested fraction of cells also had high frequencies of dead and antibiotic-tolerant persister cells. We propose a conceptual framework for understanding growth heterogeneity among individual cells as a consequence of optimizing population growth.
We consider a generic model for an irreversible metabolic pathway with enzymes A and B producing and consuming an intracellular metabolite M with fluxes V+ and V–, respectively (Fig 1a). M is also degraded by first-order dilution from cellular growth. This model captures key aspects of single metabolite conversion steps. We use the model to examine the effect of metabolic conditions at various time and size scales: short timescales (faster than gene regulation), intermediate timescales on the order of gene regulatory events, and the larger size and timescale of population growth.
To study the consequences of metabolic cost/benefit trade-offs on short timescales, we modeled metabolite levels with constant enzyme concentrations. Enzyme B supports growth via the flux V− relative to a demand, δ, that determines the flux optimizing cell growth. Cost, or toxicity, may arise from the substrate, M, or from metabolic byproducts (e.g., symported protons from a permease protein). Increasing metabolite production flux relative to demand (V+/δ) speeds cellular growth when V+ is not too high (Fig 1a and 1b). However, for excessively high V+ (Equation S8 in S1 Text) the consumption flux V− saturates. In this regime, our theory predicts an increase in the concentration of M (or byproducts of M production) until growth stalls (Fig 1c).
Consequently, we predict three qualitatively distinct cellular growth regimes with this model (Fig 1b and 1c). Regime I corresponds to substrate-limited starvation in which low metabolite production limits growth. In Regime II (satiation), metabolite production meets demand for the pathway while toxicity is low enough not to drive cells into growth arrest. Regime III (surfeit) represents a phase where metabolite production exceeds the demand and metabolic benefit cannot compensate for toxicity, resulting in stalled growth and possibly cell death. Even slight toxicity from pathway activity and saturability of the consumption flux V− are sufficient for the emergence of Regime III (S1 Text; S1 Fig).
For moderate pathway demand, low total flux suffices to meet the demand. Thus, cells remain well below toxic levels of activity and Regime II is wide (Fig 1b). On the other hand, for high pathway demand, the fastest cellular growth rate is near the point where Regimes I, II and III converge (Fig 1b). For demand higher than this convergence point, no level of metabolic flux can offset the costs of toxicity. For demand just below the convergence point, stochastic fluctuations in enzyme levels could drive fast-growing cells to drastically slow the cellular growth rate. Stochastic crossing of the critical surface bordering Regime III soon becomes irreversible at least until other compensating mechanisms, such as toxin efflux, ensue (Fig 1c). This happens because the stabilizing effect of growth- or enzyme-mediated dilution of toxicity is absent in Regime III, causing continued toxic buildup. As long as the probability of entering Regime III is non-zero, biochemical irreversibility causes any particular cell to eventually end up in the growth-arrested state. Yet, if cell division is sufficiently faster than the rate of threshold crossing, the population can still grow even while individual cells arrest their growth.
In the presence of metabolic fluctuations, the existence of a regime characterized by growth arrest suggests two alternative strategies to optimize population growth when demand for the pathway product is high. First, all cells could “play safe”, avoid Regime III and support relatively uniform growth rates across the cell population by ensuring low production or high consumption of the metabolite. However, limiting metabolite production causes starvation and slows growth, while excessive, underutilized downstream metabolic flux capacity incurs a cost from enzyme expression [25]. Alternatively, metabolic pathways could operate close to the critical threshold, maximizing cell population growth rates, but also risking stochastic transitions into Regime III and subsequent growth arrest. We hypothesized that sufficiently high growth rates of a metastable, sub-threshold cell subpopulation can more than offset losses across the threshold. This would lead to faster population growth than the strategy of uniform cellular growth rates achieved by lowering the ratio of metabolite production to consumption. Consequently, we predict that a signature of cell populations near the border of Regime III will be growth rate heterogeneity as a result of noise causing some cells to exceed the threshold (fitness noise [43]). Such populations would consist of a rapidly dividing majority of cells that constantly supply a slower-growing subpopulation prone to entering growth arrest. If the cells move further toward Regime III, population growth rate should slow, giving rise to an optimum slightly below Regime III.
We next computationally tested our hypothesis in models of cellular growth while accounting explicitly for gene expression, as well as biochemical, noise. We used stochastic simulations to capture the stimulatory and inhibitory effects of pathway activity on cellular growth, molecular dilution, and gene transcription. By simulating multiple trajectories, we predicted average cellular growth rates as a representation of population growth rates (Fig 2a–2c; Table 1; parameters ktA and ktB are represented by ktE in the table; see footnote † in Table 1). We therefore relaxed the assumption that metabolite dynamics are faster than gene expression dynamics, but still assumed metabolic cost and benefit to arise quickly from the metabolic intermediate and product, respectively. We scanned the rate of metabolite production by changing the transcription rate of enzyme A and performing N = 10,000 simulations for each value.
Consistent with the simpler models, a peak population growth rate occurs at intermediate metabolite production rates (Fig 2a). Complete growth arrest predicted at high metabolite production rates relates to the lack of population dynamics in stochastic simulations (Methods).
Correlations between variables in the simulation and growth rates show the predicted effect of each component on growth (Fig 2b). Namely, positive correlations imply that the molecule species improves growth, while negative correlations imply the opposite. Below peak population growth, single-trajectory (corresponding to single-cell) growth rate predictions were positively correlated with enzyme A, which increases metabolite levels (blue line, Fig 2b). Above peak population growth, the growth rate predictions correlated negatively with enzyme A, but correlated positively with enzyme B (orange line, Fig 2b). Above starvation levels of metabolite production rate, the amount of metabolite M was negatively correlated with predicted growth rates (red line, Fig 2b). These results confirm the same mechanism driving toxicity in this model as the simpler model presented in Fig 1.
We then compared cellular growth rates to metabolite concentrations in individual cell simulations (Fig 2c). Red regions depict the most frequent simulation outcomes, while blue shows ones. Uncolored areas show regions with no simulation trajectories. Gaps in the density at low metabolite concentration correspond to low metabolite levels (low integer values, such as 1–2 per cell). These plots show that population-level growth at various metabolite production rates arises from individual cells being in qualitatively distinct growth regimes. Most trajectories predict low metabolite levels. At higher metabolite production rates, a fraction of cells enter the growth-arrested phase because of high metabolite concentrations. We analyze the implications of these single-cell predictions on population dynamics in S1 Text and S3 Fig. Together, the single-cell and population-scale models provide a consistent prediction that metastable population dynamics can cause a non-monotonic relationship between metabolite levels and growth rates.
To experimentally test which population-level growth strategy is followed by a fast-growing bacterial population with high metabolite demand, we exploited E. coli lactose catabolism. In this pathway, imbalances between metabolite production and consumption are toxic, and there is the potential for growth modulation via downstream events in glucose and galactose catabolism. Lactose permease LacY (equivalent to enzyme A in Fig 1a), catalyzes metabolite production by importing extracellular lactose. The consumption flux consists of lactose conversion into glucose and galactose by β-galactosidase LacZ (equivalent to enzyme B Fig 1a). Excess lactose in growth media may thus inhibit growth via mechanisms related to lactose killing or buildup of one or more metabolites. Nevertheless, intracellular lactose and subsequent processing of its catabolic products is necessary for cellular growth when no other carbon source is present. Thus, the system has a trade-off with any toxic effects that may arise.
We examined the effect of extracellular lactose concentration on the population growth rate of E. coli in simple settings devoid of feedback regulation. To do so, we grew a lacI− strain derived from REL606 [44] at increasing lactose concentrations (corresponding to increasing flux V+ along the black lines in Fig 1b). Population growth rates were highest at intermediate lactose concentrations (1–5 mg/ml; Fig 3a). A quadratic model (dashed line in Fig 3a; Akaike information criterion [AIC] = -148.674) with a growth rate optimum at intermediate lactose concentration describes the pattern better than a linear model (AIC = -129.603). Parallel growth experiments on the ancestral lacIwt strain confirmed the same trend (S4a and S4b Fig). These results are consistent with our prediction that toxicity of high lactose concentrations will lower population growth rates in cultures acclimatized to those conditions. However, they cannot distinguish between uniform toxicity to the entire population and increased growth heterogeneity with a subset of cells transitioning into growth arrest or death. To distinguish between these two possibilities we used flow cytometry and microscopy to characterize growth properties in single cells.
To determine whether population growth rates were characterized by uniform or heterogeneous cellular growth, we used a chromosomally-integrated PlacO1-GFP reporter (with constitutive expression in the lacI− strain [44]) as a single cell-level growth rate sensor (Fig 3b). To calibrate the sensor, we matched the demonstrated negative dependence of constitutively expressed protein concentrations on bacterial growth rates [45] to our data. That is, we took lower fluorescence readouts to indicate faster growth. Low mean fluorescence values observed at high lactose concentrations indicate fast growth of single cells, yet the population growth rate is low in those conditions. We compared the likelihood that two possible models explain the results. In one, we assumed balanced exponential growth (uniform growth model). In the other, we added a mechanism for lactose-dependent transition into a non-growing state in which the mean fluorescence is determined by the majority of growing cells (heterogeneous growth model). We found that the model incorporating heterogeneous growth explained the pattern of mean fluorescence much better than the model enforcing uniform growth (S5 Fig). Moreover, fluorescence heterogeneity (CV) increased significantly as lactose doses increased (Fig 3b), as did skewness and kurtosis (S6 Fig). These increases indicate both an increase in the relative heterogeneity of growth rates and fatter tailed growth rate distributions at high lactose concentrations.
Cell death may be a feature of cell sub-populations following prolonged time in regime III. To estimate cell death in each condition, we used propidium iodide (PI) staining to identify dead cells following incubation at different lactose concentrations. As predicted, we saw a significant increase in dead (PI+) cells as lactose concentrations increased past the point supporting a maximum growth rate (Fig 3a). This result is consistent with metabolite toxicity concurrent with growth arrest resulting in an increased chance of cellular death. We also observed an elevated frequency of PI+ cells in low lactose, likely attributable to starvation-induced cell death [46] (S4c Fig). To further confirm the pattern of PI staining, we fit the data with a quadratic curve (Fig 3a, dashed line); AIC = -694.25 vs -655.964 for a linear model.
In the above experiments we used shaking microplates that provide less aeration than some other methods, and did not permit time-lapse detection of fluorescence. We therefore next visualized the competing effects of growth stimulation and inhibition on individual cells at different lactose concentrations. We did so with time-lapse fluorescence microscopy of lacI−PlacO1-GFP E. coli cells growing in an incubated microfluidic device with a constant flow of freshly oxygenated air over the cell growth chamber (Fig 3c, S1–S4 Movies). After growth for 18–24 hours, we stained for dead cells using 1 μg/ml PI in the perfused medium. The resulting images show striking qualitative differences in population structure. In low lactose (0.1 mg/ml), cells exhibited uniform green fluorescence (and therefore, growth rates). At 50 mg/ml lactose, clustered subsets of cells exhibiting higher fluorescence emerged. These patterns are consistent with subpopulations once poised on the threshold between fast growth and growth arrest that subsequently stopped growing (Fig 3c). In 50 mg/ml lactose, PI-stained cells co-localized with these islands of high fluorescence, corresponding to an elevated chance of cell death arising from growth-arrested subpopulations.
Taken together, these observations support a model of population growth balancing the costs and benefits of lactose metabolism. There is a limit to the benefit of lactose in the media, and growth is inhibited in a fraction of cells as the concentration of lactose increased. Metabolite buildup may also occur downstream in the pathway, in which case enzymes A and B are downstream as well, or upstream, due to LacY saturation or toxicity from some off-target mechanism. We discuss why the latter possibility is unlikely in S1 Text and Discussion. Because we observed acclimatized cultures in constant growth conditions, the underlying mechanism for growth inhibition at high concentrations is a continuously occurring process, as opposed to a shock caused by changing growth conditions.
The existence of growth heterogeneity in high lactose cultures raises the question of whether some growth-arrested cells function as persisters (slow- or non-growing cells that tolerate antibiotic treatment). Emergence of bacterial persister cells is typically attributed to the action of toxin-antitoxin systems [35, 47]. However, starvation-induced (p)ppGpp signaling [48], entry into stationary phase [49], or cell-cell signaling [50] can also induce persister formation, suggesting that the underlying mechanisms may be diverse. To our knowledge, persister formation from excessive intracellular metabolic activity has not been shown. To determine if persisters arise preferentially in cultures with high lactose concentrations, we measured the kinetics of cellular killing in ampicillin following growth at low (0.1 mg/ml), moderate (1.5 mg/ml) and high (50 mg/ml) lactose concentrations (Fig 4a). To estimate the efficiency of killing, we subsequently (post-treatment) grew cells in the absence of lactose, which should relieve the metabolic burden, allowing surviving culturable cells to seed colonies. Consistent with the presence of persister cells, killing (estimated from colony forming units, CFUs from the media) was biphasic with time. Namely, the first phase of fast killing up to ~6 hours was distinguishable from the second phase of slower killing beyond 6 hours in all conditions. This second phase of killing is caused by persister cells. The second phase of killing in 50 mg/ml lactose had significantly slowed death compared to 1.5 mg/ml. Our results also suggest enrichment of persisters in 0.1 mg/ml lactose, consistent with starvation-induced persister formation [48, 51]. We also grew acclimatized cultures in a range of lactose concentrations with or without the antibiotic doxycycline [52] or ampicillin and counted colony forming units (CFUs) from growth on LB medium as a measure of cell survival after 20 h. Following antibiotic treatment, we found significantly enriched colony formation in populations grown at higher lactose concentrations (Fig 4b), indicating increased antibiotic tolerance. We conclude that conditions that lead to the production of growth-arrested subpopulations also enrich for antibiotic tolerant and persister cells.
Our proposed model for metabolism-induced growth arrest represents a limit where toxic side effects slow cellular growth more quickly than shifts in gene expression or detoxification by other factors, such as LacA [53, 54], can alleviate the toxicity. We now introduce a generalized framework for the effects of molecular-scale events on bacterial growth dynamics. From this framework, we derive criteria for the effects of molecular subnetworks on population growth with different timescales.
Fig 5a illustrates the generalized framework. It contains four states of bacterial growth: (i) balanced, exponential growth; (ii) a transient state of unbalanced growth, with growth rates undergoing a change; (iii) viable growth arrested cells; (iv) dead cells. Characteristic timescales for switching between states are indicated. Various limits on the timescales give well-known types of population growth. For example, as (1/τ1)/(1/τ-1) → 0, we arrive at balanced exponential growth (Fig 5b). For (1/τ2)/(1/τ-2) → ∞, we arrive at the metastable population dynamic underlying our model above as well as typical toxin-antitoxin systems. Other growth conditions that may exist as limits of this framework are stationary phase or biofilms where (1/τ-1) → 0, and diauxic shifts where 1/τ1 is transiently larger than other parameters.
In short, our model of metabolism-driven metastable population dynamics exists as a special-case limit of a larger growth framework. It appears to apply to the E. coli B REL606 strain grown in lactose. How frequently different parameter limits reflect the effects of different gene regulatory networks remains to be determined. Homeostatic feedback by metabolites could also reduce toxicity by raising the growth arrest threshold (e.g. via PTS IIAGlc feedback to LacY [55]) or eliminating it altogether (S1 Text). For instance, the E. coli strain K-12 MG1655 does not exhibit the same pattern of dependence of growth rate on lactose concentration (S7a Fig). Further, high lactose concentrations do not produce detectably more persister cells in K-12 compared to a lower lactose concentration (S7b Fig). Via comparisons of the strains, we identify a number of potential mechanisms that could leave the B REL606 strain prone to this effect (S1 Text, S7 Fig, S1 Data). A lack of oxygen in the growth conditions did not appreciably contribute to the effect, as indicated by qualitatively similar results obtained from selected lactose concentrations grown in well-aerated flasks (S9 Fig). No difference in pH between the growth media of the strains was observed in flask or microplate growth (S10 Fig).
In predicting and characterizing the effects of metabolic costs and benefits on bacterial populations at single-cell resolution, our results indicate that metabolic pathways have physiological effects far beyond their metabolite-processing roles. In particular, specific conditions may allow certain metabolic pathways to determine the uniformity or heterogeneity of cellular growth rates. Our results raise the clinically important possibility that some antibiotic persisters are intrinsic consequences of threshold-crossing metabolic events in growing cells.
The effects we found depend on a set of network characteristics common to most, or perhaps all, metabolic pathways: enzyme saturation, toxicity, and a limitation to the benefit that can be derived from pathway products. Another set of criteria for our results suggest how robustness to heterogeneity may form: benefit and toxicity must take effect faster than compensatory regulatory or enzyme-kinetic mechanisms can rescue cells on the trajectory toward growth arrest (Fig 5), and some types of homeostatic feedback may reduce or eliminate the effect (S1 Text). Our experiments provide strong evidence that metastable growth driven by metabolic thresholds is possible in bacterial populations. We see no reason why this conclusion should be limited to any particular species or pathway, but rather its generality is limited by the types of molecular mechanisms present and the criteria for the effect identified in S1 Text.
Questions about the mechanism of metastable population dynamics remain. What is the direct cause of cellular growth arrest? It is possible that off-target toxicity in our experiments, not directly emerging from intermediate production, changes growth rates after LacY is saturated. Our analysis suggests that such an effect would drive a pattern of uniformly reduced growth rates (S1 Text), unlike what we observed experimentally. We chose to test our phenomenological models of toxicity in lactose because of the well known lactose killing effect [26], assuming that permease-related toxicity would be the cost. However, our models predict the same metastable dynamics even with minor toxicity that could arise from intracellular metabolite buildup. It is thus possible that intracellular lactose or downstream metabolite processing contributes to the cost of excess lactose in the media. A comparison between an E. coli strain subject to the effect (B REL606) with one that is less sensitive to it (K-12 MG1655) explores three possible types of differences that could underlie the effect: amino acid substitutions, changes in gene regulation, and differences in cell size (S1 Text, S7 Fig, S1 Data). Our analysis does not conclusively rule out any of them, but reveals more extensive differences in gene regulatory sequences than amino acid sequences. It also shows that cell size differences could lengthen the timescale of metabolite buildup for a limited range of kinetic parameters. Finally, persister cell enrichment is largely attributed to toxin-antitoxin systems. We cannot rule out that the effects of metabolite excess percolate to TA systems that ultimately induce growth arrest.
With each of these mechanisms, and possibly others, being candidates for driving growth heterogeneity, what is the physiological makeup of the growth arrested pool of cells? We found enrichment of PI+ dead cells and antibiotic persisters in our study. Other phenotypes of non-growing cells may be present as well. For instance, rapid onset of growth arrest may freeze protein concentrations at non-steady-state levels that originated from gene expression noise.
We conclude that single-cell resolution computational and experimental studies are an indispensable tool for understanding how metabolism drives cellular growth. Competing timescales of metabolite kinetics and gene regulation appear to be of central importance for determining population growth dynamics. The relevance of our findings may extend beyond metabolic pathways as well, to other systems governed by supply and demand.
This study used E. coli REL606 lacI−(or its ancestral lacIwt) transformed with Tn7∷PlacO1GFP (KanR) as described previously [44], and E. coli K-12 MG1655 (Coli Genetic Stock Center). Cultures were started in LB medium (BioWorld) from -80°C storage, grown for 12–15 hours in a shaking 37°C incubator (VWR 1575, Sheldon Manufacturing, Inc.), and resuspended 1:100 in Davis Minimal medium (DM; Difco) supplemented with thiamine along with concentrations of lactose as described in the main text. Resuspended cultures were grown in 96-well non-cell culture-treated flat bottom plates (Falcon) for 24 hours to allow them to physiologically acclimate to the prevailing environment. After acclimation, the cultures were resuspended into fresh 96-well plates containing DM with identical sugar concentrations, 1:100 for flow cytometry and 1:1000 for growth measurements in an incubating microplate reader (BioTek Eon). Representative growth curves are shown in S8 Fig.
The flow cytometry cultures were grown for 5–7 hours in a shaking 37°C incubator, diluted 1:5 in plain DM, and run through a Beckman Coulter MoFlo XDP flow cytometer with Hoechst stain to trigger read events. To assay cell death, PI (1 μg/ml) was introduced to cultures, they were incubated for ten minutes, and the fraction of PI+ cells was quantified using a Beckman Coulter MoFlo XDP. Flow cytometry files were converted to plain text format in R [56], and filtered to eliminate cells more than 0.5 standard deviations around the mean forward and side scatter in linear coordinates. The scatter distributions were relatively tight, so this gate had the effect of eliminating outliers. A very small number of additional outlier points were filtered for having extremely high fluorescence (between 0 and 40 points per condition, maximally 0.13% of the data points; the threshold was being greater than 6-fold higher than the mean fluorescence in 0.1 mg/ml lactose). We computed average fluorescence from the distributions by finding the means of the gated data of 3 culture replicates, and then computing mean ± standard error of the means. The flow cytometry files are available from Flow Repository, ID: FR-FCM-ZZLJ.
Flow cytometric analysis of cells in each lactose concentration presented have three biological replicates each (including independent measurements of GFP and PI stain). The overall trend in GFP expression has been observed in three separate sets of experiments and PI staining has been observed in two separate sets of experiments.
Microplate reader cultures were grown for 24 hours at 37°C with continuous shaking and a reading of OD450 every two minutes. We computed growth rates from resulting growth curves in early exponential phase using a log-linear model fit in R, and finding mean ± standard error of the growth replicates. Growth rate experiments have three biological replicates per condition. The overall growth response trend for growth in lactose in E. coli B REL606 has been measured at least three times (with 3–4 biological replicates each). The overall growth response trend for growth in lactose in E. coli K-12 MG1655 has been measured twice (with 3 biological replicates each).
To confirm that low oxygen levels in the plate reader did not qualitatively change our conclusions, we repeated flow cytometry for cultures in representative lactose concentrations (2.5 mg/ml and 50 mg/ml) that were grown in well-aerated 150 ml flasks in 10 ml of culture (S9A Fig). Colony forming units and OD450 readings were taken periodically to determine growth curves and set a time to sample for flow cytometry (S9B and S9C Fig). We selected 4h post-inoculation for flow cytometry, and carried it out as described above.
To rule out the possibility that acidification of the culture media underlies the higher growth heterogeneity in B REL606 than K-12 MG1655, cultures of the two strains were grown in 50 mg/ml lactose in well-aerated 150 ml flasks in 10 ml of culture, and pH of samples measured with a Mettler-Toledo pH meter (S10 Fig). To check for pH differences in the 96-well microplate setup, we carried out a plate reader experiment in 50 mg/ml lactose with both strains out to 7.5 hours, representative of the final time used to fit the growth curves. At that time, we pelleted the cells in a centrifuge, and measured the spectral properties of supernatant from B REL606 and K-12 MG1655 treated with the pH-indicating dye phenol red (30 μg/ml) using an Agilent Cary 60 UV-Vis spectrophotometer with a quartz cuvette. Changes in the pH level are reflected in the spectral peak near 460 nm (S10 Fig). At this wavelength, the absorbance difference between the strains is non-significant (p = 0.27; N = 3 for each strain).
Cultures were acclimatized in 1 ml cultures as described above in 14 ml polypropylene tubes (Falcon), resuspended 1:1000, and grown for an additional 5–6 hours. At this time, the cultures were resuspended to an OD600 of approximately 0.005 and added to an ONIX microfluidic plate (CellAsic model B04A). Cultures were grown for 18–24 hours with a heated manifold maintaining 37°C, a constant flow of fresh air over the culture chamber, and constant perfusion (1 psi) of fresh media on a temperature-stabilized 100x oil immersion objective in a Nikon Ti1000 microscope. Every two minutes, the imaged field was algorithmically autofocused and images taken in bright field, at 460 nm (green fluorescence) and at 565 nm (red fluorescence). At the end of the experiment, DM with the same concentration of lactose containing propidium iodide (1 μg/ml) was perfused for 5–10 minutes to stain dead cells, and images of various locations in the chamber were then captured. Individual color channels in images in Fig 3 were adjusted for brightness; PI stained cells appear as bright red or yellow, depending on the relative PI and GFP levels.
Time-lapse microscopy has been repeated 2–3 times for some of the conditions to confirm the existence of the patterns observed.
E. coli cultures were started and acclimatized in various concentrations of lactose as described above. After stabilization, the cultures were resuspended into identical lactose concentrations containing either 32 μg/ml doxycycline [52], 100 μg/ml ampicillin, or a control treatment without antibiotic. For the antibiotic tolerance assay, these cultures were grown for 20 hours and plated onto LB plates without antibiotic selection at 1x, 10x, 100x, and 1000x dilutions (and up to 107-fold dilution for untreated cultures). For the ampicillin kill curves, samples were taken at the indicated time points and diluted appropriately for accurate CFU counts on LB plates.
In the antibiotic tolerance assay, each 96-well culture plate had a single lactose dose with ampicillin, doxycycline, or no antibiotic controls. Persister timecourses have three biological replicates per time point and the experiment has been done twice for B REL606 and once for K-12 MG1655. Survival ratios after 20h have been measured twice for each strain (3 biological replicates) to confirm the pattern. Survival ratios were calculated from raw data by matching between treated (xi) and untreated (yi) replicates originating from the same culture, and calculating statistics from the ratio log10(xi/yi) calculated for each replicate.
Analysis of the numerical results, mathematical models, and data processing not otherwise specified were done in Mathematica versions 8 or 9 (Wolfram Research).
To determine statistical trends in the data, subsets of data points were subjected to linear or nonlinear regression and the resulting ANOVA statistics reported using the LinearModelFit or NonlinearModelFit function in Mathematica.
To assess the effects of intrinsic biochemical noise on predicted fitness in the simple metabolic pathway, we extended a previously published model of this pathway that included reactions capturing transcription, translation, substrate-enzyme interactions and catalysis, and concentration-dependent molecular dilution at a constant growth rate [39]. Simulation models were constructed in Copasi 4.8 (Build 35) [57]. Stochastic simulations were performed with StochKit 2.0.1 [58] on a computational cluster using detailed mass action propensities for chemical reactions and custom propensities for dilution and degradation reflecting growth feedback effects. The model includes complex propensity functions that capture the effects of changing growth rates on dilution and transcription rates [45] as outlined in Table 1.
For each condition, we ran 10,000 simulations. For each simulation, we set initial conditions equal to the mean field steady state for the given parameter set and ran time trajectories for 8,000 s of simulation time each. Population growth rates were calculated as the mean growth rate across the trajectories. Because the trajectory did not include any contribution from cell division, parameter sets resulting in a high probability of crossing the threshold into growth arrest resulted in virtually all trajectories entering the growth arrest state; as a result, conditions with very high simulated metabolite production rates result in predictions of virtually no growth. Changing the trajectory time would thus have the effect of changing the fraction of cells entering growth arrest. The chosen time of 8000 s demonstrates the principle on a timescale relevant to cell growth and minimizes the effects of the initial condition on the final result. An exemplary file used for stochastic simulations is presented in S2 Data.
To identify possible underlying mechanisms of growth arrest in B REL606 in high lactose, we analyzed genetic differences between that strain and K-12 MG1655, which is less prone to the effect (S1 Text). We first identified 65 proteins involved in lactose processing or events directly downstream, including glycolysis, galactose degradation (the Leloir pathway), and other enzymes or transporters that result in the production or degradation of lactose, glucose, or galactose (listed in S1 Data). For each protein, we retrieved the amino acid sequence for each strain from UniProt, and identified coordinates of regulatory regions for each gene in each strain. Regulatory sequences were identified based on the well-annotated K-12 MG1655 strain, and putative regulatory regions of B REL606 genes were selected to encompass intergenic regions up to the neighboring operon. Databases of E. coli gene and protein sequences were used in the analysis as described in greater detail in S1 Text [59–62].
Each pair of amino acid or nucleotide sequences was aligned with ClustalW2 and the score recorded in S1 Data. Alignment scores were not penalized for having unaligned regions. To determine if the regulatory DNA sequences were more alike than a randomized sequence, we used a resampling procedure to generate 1000 pairs of randomized sequences for each actual sequence, with nucleotides for each sequence randomly sampled with replacement from its corresponding actual sequence. We calculated alignment scores for the randomized pairs, used them to create a numerical score distribution, and calculated from that the probability that the alignment score for each regulatory region was more alike than random.
|
10.1371/journal.pcbi.1004063 | Modulation of Calmodulin Lobes by Different Targets: An Allosteric Model with Hemiconcerted Conformational Transitions | Calmodulin is a calcium-binding protein ubiquitous in eukaryotic cells, involved in numerous calcium-regulated biological phenomena, such as synaptic plasticity, muscle contraction, cell cycle, and circadian rhythms. It exibits a characteristic dumbell shape, with two globular domains (N- and C-terminal lobe) joined by a linker region. Each lobe can take alternative conformations, affected by the binding of calcium and target proteins.
Calmodulin displays considerable functional flexibility due to its capability to bind different targets, often in a tissue-specific fashion. In various specific physiological environments (e.g. skeletal muscle, neuron dendritic spines) several targets compete for the same calmodulin pool, regulating its availability and affinity for calcium. In this work, we sought to understand the general principles underlying calmodulin modulation by different target proteins, and to account for simultaneous effects of multiple competing targets, thus enabling a more realistic simulation of calmodulin-dependent pathways. We built a mechanistic allosteric model of calmodulin, based on an hemiconcerted framework: each calmodulin lobe can exist in two conformations in thermodynamic equilibrium, with different affinities for calcium and different affinities for each target. Each lobe was allowed to switch conformation on its own. The model was parameterised and validated against experimental data from the literature. In spite of its simplicity, a two-state allosteric model was able to satisfactorily represent several sets of experiments, in particular the binding of calcium on intact and truncated calmodulin and the effect of different skMLCK peptides on calmodulin’s saturation curve. The model can also be readily extended to include multiple targets. We show that some targets stabilise the low calcium affinity T state while others stabilise the high affinity R state. Most of the effects produced by calmodulin targets can be explained as modulation of a pre-existing dynamic equilibrium between different conformations of calmodulin’s lobes, in agreement with linkage theory and MWC-type models.
| Calmodulin, the ubiquitous calcium-activated second messenger in eukaryotes, is an extremely versatile molecule involved in many biological processes: muscular contraction, synaptic plasticity, circadian rhythm, and cell cycle, among others. The protein is structurally organised into two globular lobes, joined by a flexible linker. Calcium modulates calmodulin activity by favoring a conformational transition of each lobe from a closed conformation to an open conformation. Most targets have a strong preference for one conformation over the other, and depending on the free calcium concentration in a cell, particular sets of targets will preferentially interact with calmodulin. In turn, targets can increase or decrease the calcium affinity of the calmodulin molecules to which they bind. Interestingly, experiments with the tryptic fragments showed that most targets have a much lower affinity for the N-lobe than for the C-lobe. Hence, the latter predominates in the formation of most calmodulin-target complexes. We showed that a relatively simple allosteric mechanism, based the classic MWC model, can capture the observed modulation of both the isolated C-lobe, and intact calmodulin, by individual targets. Moreover, our model can be naturally extended to study how the calcium affinity of a single pool of calmodulin is modulated by a mixture of competing targets in vivo.
| Calmodulin is a ubiquitous calcium-binding protein involved in many cellular processes, from synaptic plasticity to muscular contraction, cell cycle regulation, and circadian rhythms. Structurally, it is organized in two highly homologous globular domains joined by a flexible linker [1]. Each domain contains two calcium-binding EF-hands that can undergo a transition between closed and open conformations, the latter favored by calcium binding [2, 3]. The transition to the open state results in exposure of hydrophobic residues able to interact with numerous binding partners [4]. However, some targets interact preferably with the closed form [5].
The shift of calmodulin’s calcium saturation curve in the presence of targets was studied experimentally by several groups [6–10]. Targets that markedly increase calmodulin’s calcium affinity include the calcium-calmodulin dependent kinase II (CaMKII), protein phosphatase 2B (PP2B) and skeletal muscle myosin light chain kinase (skMLCK). It was shown that, in general, targets do not bind calmodulin exclusively in either calcium-saturated or calcium-free forms, but rather bind to both forms with different affinities. Numerous biologically relevant targets exhibit this behavior, such as skMLCK, CaMKII, and the NaV1.2 sodium channel [7, 10–13]. The binding domains that preferably interact with calcium-free calmodulin are called IQ motifs, a family of 14-residue sequences named after the two most frequent initial amino acids (usually isoleucine, followed by a highly conserved glutamate) [5].
Calmodulin’s properties, such as cooperativity and affinity modulation by binding targets, can be explained by an MWC allosteric model. The first model of allosteric transitions was introduced in the seminal paper by Monod, Wyman and Changeux in 1965, which dealt with multimeric proteins whose subunits could undergo concerted conformational transitions, and also be modulated by target binding either to the R state, or the T state, in an exclusive fashion [14]. The model was then further extended by Rubin and Changeux to describe molecules that could be modulated by binding partners capable of binding both conformational states, but with different affinites [15]. The hypothesis of the existence of distinct calmodulin conformations in thermodynamic equilibrium, which is crucial to an allosteric model, is supported by evidence of conformational dynamics of the protein in solution, with a time constant on the order of microseconds [16–18], and was also suggested by theoretical and computational analysis of coarse-grained Hamiltonians [19, 20]. Structural studies also supported the hypothesis that a small fraction of calmodulin molecules can exist in a more compact conformation even in the presence of calcium [21].
Experimental studies with constrained mutants showed that the capability to switch conformation is necessary for cooperativity [3, 22, 23], which is compatible with an MWC model. Interestingly, conformational transitions are not common to all EF-hand based proteins, despite their high structural homology in calcium-free conditions. For example, calbindin D9k is structurally nearly identical to a calmodulin N-lobe, but its EF-hand remains closed after binding calcium [24]. It was proposed that the hydrophobic pockets of the open calmodulin lobes are stabilised by the unusual local abundance of the usually rare methionine residues (over 6% for calmodulin, against the 1% of the average protein) which have the highest flexibility, minimum steric hindrance and minimum solvation energy of all hydrophobic amino acids, and can therefore adapt to both segregated and and solvent-exposed conditions [4]. The flexible aliphatic side chains of methionine can also easily establish contacts with very diverse substrates, contributing to calmodulin’s promiscuity [25]. Some examples of calmodulin’s capability to bind very diverse targets is shown in Fig. 1.
Despite their conceptual simplicity, the application of allosteric models can be challenging. The intrinsic calcium affinities of different conformational states are not easily measurable and need to be reverse-engineered, because most experimental results are fitted using the phenomenological Hill or Adair-Klotz models, which do not incorporate conformational transitions. In addition, modelling intact calmodulin, within which each domain can switch its conformation independently, results in a large number chemical species that need to be explicitly enumerated (combinatorial explosion), and a larger number of parameters that need to be fitted simultaneously. A previous allosteric model of calmodulin was developed by Stefan et al. [26], to model the differential calmodulin-dependent activation of calcineurin and CaMKII in synaptic plasticity. However, the model postulated that both lobes would undergo concerted conformational transitions, and also had similar calcium-binding properties. In fact, the two lobes possess a remarkable degree of autonomy [27], and the calcium saturation curve of calmodulin was almost exactly reproduced by superposition of the saturation curves of tryptic fragments TR1C and TR2C, containing respectively the N-terminal or C-terminal lobe only [6, 28, 29]. The four alpha-helices of each globular domains form two EF-hands that work together as one cooperative unit with two calcium-binding sites [30]. Despite of their high level of structural similarity, the two domains also exhibit strikingly different affinities and binding kinetics for calcium ions. The C-lobe has a 10-fold higher affinity, but much slower kinetics, than the N-lobe [31, 32]. The clearly different saturation curves of the lobes, as observed in the intact molecule, are shown in Fig. 2.
The differences in calcium-binding behavior are surprising in domains that are structurally so similar, but several studies showed that the EF-hand motif is tunable over a wide range of affinity and kinetics by modification of few key residues [33, 34]. The incredibly high level of sequence conservation across species suggests that the different properties of the two lobes are in some way crucial to calmodulin’s function. Moreover, even mutations that do not affect calcium-binding properties, but alter calmodulin’s affinity for one or more targets, can be lethal [35]. The sequence of calmodulin’s four EF-loops (12-residue calcium-binding pockets within each EF-hand) is given in Fig. 3.
The C-lobe was reported to play a pivotal role in mediating calmodulin’s calcium-dependent interactions with its targets [36], and experiments with tryptic fragments of calmodulin consistently showed that the affinity of the C-lobe for calmodulin-binding domains is usually much higher than the affinity of the N-lobe [7, 13]. Moreover, targets that bind calcium-free calmodulin seem to interact almost exclusively with the C-lobe [10, 37, 38]. In the light of these facts, we postulated that a major portion of the observed target-induced modulation effects could be explained by the interactions of the targets with the C-lobe only, and we focussed our initial effort on modelling the observed properties of TR2C, (i.e. the isolated C-lobe). Taking advantage of existing experimental data sets, we modelled the behavior of the TR2C tryptic fragment in the presence of peptides WFF and WF10 (peptides that express full-length and truncated versions of the calmodulin-binding domain of skMLCK) and Nav1.2IQp (a peptide based on the calmodulin-binding domain of the calcium-activated NaV1.2 sodium channel). We developed and parameterised a model of the tryptic fragment, in order to reliably reverse-engineer some intrinsic properties of the C-lobe of calmodulin. We then parameterised an MWC model of the isolated N-lobe by postulating that the R-states of both lobes had very similar affinities (as suggested by [6]), thus reducing the number of free parameters to estimate (see Methods).
This simplifying assumption was needed to circumvent the comparative scarcity of experimental data on the isolated N-lobe. The two submodels were then combined into a model of intact calmodulin, under the assumption that the protein behaved as the sum of its parts.
The MWC model explains cooperativity and target-induced affinity modulation as emergent properties, rather than a priori assumptions, as opposed for example to the Adair-Klotz model. The two formulations are however mutually consistent, and MWC and Adair-Klotz models are interconvertible, as shown by Stefan et al. [39], since for any MWC model the corresponding Adair-Klotz parameters can be computed. A notable advantage of a MWC model is that the effects of multiple competing targets are straightforward to incorporate, simply by defining each target’s affinity for the different conformational states of calmodulin. The work described in this paper shows that a carefully parameterised MWC model can indeed give reliable predictions of how individual targets modulate calmodulin affinity, and can also help investigate biologically relevant situations where numerous targets simultaneously modulate (and compete for) the same calmodulin pool.
A diagram of the allosteric model of calmodulin with hemiconcerted conformational transitions is shown in Fig. 4. Our model was built in three steps. First, we parameterised a model of TR2C, (isolated C-lobe), then parameterised a model of TR1C (isolated N-lobe), and finally the two submodels were merged into a model of intact calmodulin. (see Methods). The model of intact calmodulin (in SBML format) was deposited in BioModels Database [40] and assigned the identifier MODEL1405060000.
The first requirement we set for our model was the capability to reproduce the saturation curve of TR2C alone. We modelled TR2C as an MWC protein with two identical binding sites and two conformational states (T and R). In the absence of targets, Equation 5 (see Methods), that expresses the fractional saturation as a function of the concentration of free intracellular calcium, simplifies to the form
Y ¯ = α ( 1 + α ) + L α c ( 1 + α c ) ( 1 + α ) 2 + L ( 1 + α c ) 2 ,
(1)
where:
L =
T
0
/
R
0
(2)
is the allosteric constant in the absence of targets, defined as the ratio between the concentrations of the T and R states, when no ligand is bound),
α = [ C a 2 + ] f r e e / K R
(3)
is the ratio between the ligand affinity of the R state and the concentration of free ligand, and
c = K R / K T
(4)
is the ratio of the ligand affinities of the R and T state. Equation 1 depends on only 3 free parameters (L, KR, KT). We observed that the intrinsic affinity of the T state must lie within the wide, and biologically plausible, 1 μM–1 mM range. Once the value of KT is chosen, the model can be rewritten as a function of the “classical” MWC parameters, L and c [14]. For any choice of KT, the score function of the fitting is a surface on the the (L, c) plane, with the same qualitative features, as shown in Fig. 5.
In particular, regardless of the value chosen for KT, all the generated fitting landscapes showed a flat valley, i.e. a region of parametric space where many different choices of L and c could all fit the saturation curve very well. Although it was technically possible to locate a point of global minimum in this region, its position was not robust to noise, and moved erratically within the valley if random errors were applied to the data sets (data not shown). This portion of parameter space contained an ensemble of possible parameter choices that lie between two limit cases: in the first one, the change of affinity upon conformational transition is relatively small (c ≃ 0.01), and the allosteric constant is also comparatively small(L ≈ 100); in the second one, the change of affinity is much greater (c ≤ 0.001) and so is the allosteric constant (L ≥ 1000). In the first case, (higher c, lower L) the molecule has a high propensity to spontaneously switch conformation to the high-affinity R state, but the R state is stabilised less strongly by the binding of calcium. In the second case (smaller c, greater L), the opposite holds true: the molecule has a very low propensity to spontaneous conformational transitions, but the stabilising effect of calcium is much stronger. It must be stressed that in both cases the model is capable of giving an excellent fit of the experimental saturation curve in target-free conditions. The knowledge of the saturation curve in the absence of targets, alone, is therefore insufficient to discriminate between the two possibilities and univocally identify the parameters. The problem could be solved using the additional information provided by the saturation curves observed in the presence of targets, which allowed us to constrain the parameter space to be sampled during the fitting procedure. The obtained constraints implied that the first of the two limit scenarios mentioned above (with greater c and lesser L) had to be be discarded, because the resulting model would be unable to account for the extent of the target-induced affinity shifts observed in the experiments, as explained in the Methods section.
We tested whether it was legitimate to reverse engineer the intrinsic properties of the C-lobe from those of TR2C. The level of saturation of the C-lobe can be monitored in both intact calmodulin and TR2C fragments by measuring the intrinsic fluorescence of their tyrosine residues [9, 10, 12, 31, 41], It was shown that truncation does not have a strong effect on the secondary structure and tyrosine fluorescence intensity of the C-lobe [42]. In order to verify that the calcium-binding properties of the intact calmodulin’s C-lobe were maintained in the TR2C tryptic fragment, we compared the two saturation curves as presented in Fig. 6. The data shows some scattering, and the uncertainty on the calcium concentration necessary to produce half-saturation is roughly of a factor two. This is probably due to slightly different experimental conditions (see the Discussion section). However all curves were similar and shared the same qualitative behavior, meaning that truncation does not dramatically alter the lobe’s properties.
The model of TR2C was parameterised by fitting the saturation curves of TR2C alone and in the presence of targets. The target ligands were chosen primarily based on the availability of published, quantitative experimental data. Also, the route we took to parameterize the model was in part dictated by the available data (see Methods). The data by Bayley and coworkers was chosen as a starting point, mainly for its thoroughness in quantitatively reporting the effects of the targets on both TR2C and intact calmodulin, both in the absence and presence of calcium [7].
The estimated parameters are summarised in Table 1, with the resulting saturation curves shown in Fig. 7. The simulations of the model were overall in good agreement with experiments.
Once the parameter values were determined, we tested the capability of the model to predict the behavior of TR2C under conditions different from those used for the fitting. The saturation curve of TR2C, as measured by Bayley and coworkers in the presence of WF10 peptide, (a truncated form of WFF with a 10-fold lower affinity for the R-state) was in good agreement with experiments (S1 Fig.). Moreover, the saturation curves by Evans and coworkers [12] were measured with two different concentrations of peptide, and it was found that doubling the concentration of Nav1.2IQp (from 1:1.4 to 1:2.8 TR2C:peptide ratio) did not produce a further shift in the saturation curve, a behavior that our model was able to reproduce (S2 Fig.). A summary of all fitted and predicted saturation curves, plotted against the corresponding experimental data points, is given in Fig. 8.
The high calcium affinity of the C-lobe, further enhanced by interactions within some targets (CaMKII, PP2B, skMLCK), has led to speculation that calmodulin could activate such targets even at resting calcium concentrations, i.e. in the absence of calcium signals in neuronal or muscular cells [6]. We used the model of TR2C to perform a preliminary investigation of the effect of competing targets on calmodulin. At this stage of analysis we can only investigate the part of such interactions that are mediated by the C-lobe. However, as previously mentioned, the C-lobe was shown to be mainly reponsible for mediating calmodulin-target interactions, and targets that bind the T-state with high affinity, in particular those that appear to have very little interaction with the N-lobe [36]. We simulated the steady-state response of a system containing calmodulin and two competing allosteric targets using data taken from the literature, one binding prefentially to the the T state (T-state binding target, TBT) and the other binding preferentially to the R state (R-state binding target, RBT). When both targets are present, the total saturation curve depends on their relative concentrations, as shown in Fig. 9. Shifting the concentration of targets is a potential way to tune the saturation curve of the calmodulin pool.
The model of the isolated N-lobe, in the absence of targets, was based on the model of C-lobe. The resulting parameters are summarised in Table 2. For the detailed procedure, see the Methods section.
The model of intact calmodulin was built by combining the two models of the N and C and lobe. The saturation curve agreed very well with the available experimental data as shown in Fig. 10. A model including calmodulin and binding targets was then automatically generated in SBML format, as described in the Methods section, and simulated in COPASI.
We parameterised the model by fitting it to available experimental data. Then, a posteriori, we showed that the model was consistent with experiment, because it could reproduce additional data that had not been used in the parameter estimation process, as shown in Fig. 11. The effect of a target on the saturation curve of a MWC molecule depends on how the different conformations are stabilised, which in turn depends on the affinity of the target for each conformation. The allosteric models of isolated lobes assume that each lobe can only exist in two states, T and R, which implies that both EF-hands on each lobe always undergo concerted conformational transitions. On the other hand, the model of intact calmodulin contains two lobes but does not constrain them to be the same state. As a result, one needs to take into account “asymmetric” conformations, where the two lobes are in different states. The model of intact calmodulin thus contains 4 possible states, called RR, RT, TR, TT. The first and second capital letter refer to the state of the N and C lobe, respectively. For more details, see the Methods section.
Bayley and coworkers measured calmodulin saturation in the presence of two peptides: WFF, which represent the full-length CaM-binding domain of skMLCK, and WF10, a truncated version of the same domain, corresponding to the portion that interacts with the C-lobe [7]. The saturation curve with WFF was markedly shifted to the left, while that in the presence of WF10 was biphasic.
Crucially, they were careful to measure the affinity of calmodulin and TR2C for these peptides, both in calcium and calcium-saturated conditions, which gave us an excellent initial estimate for the affinity of the peptides for the RR and TT state of intact calmodulin, and for the R and T states of TR2C. The affinity of each target for the isolated C-lobe in the R state was used as an estimate of the affinity of the target for the intact molecule in the TR state (i.e. we assumed that the target interacted with the C-lobe was much more strongly than with the N-lobe, in agreement with available experimental data). The estimated affinities for the four states are summarised in Table 3. The agreement between simulation and experiment was very satisfactory (Fig. 11).
Throughout this work, we have assumed that the affinity of targets for the T and R state could be estimated by the affinity that the targets exhibit for calmodulin in low and high calcium conditions, respectively. The possibility that the binding of target is sufficient to cause structural change (which would invalidate our assumption) was considered but discarded based on the following published data. Several authors consistently found that peptides of calmodulin targets bind calmodulin with different affinities in the absence of calcium, or at saturating calcium concentrations [7, 9, 10, 12]. The fact that calcium induces a conformational transition in calmodulin is also known from experimental results now widely accepted in the field [2, 43, 44]. Targets that bind calmodulin with high affinity in the presence of calcium generally bind the calcium-free form with a measurable but much lower (10 to 100-fold) affinity, as shown by numerous experiments [7, 9, 10]. This strongly suggests that calcium ions are actually the main determinants in driving conformational transitions. We must therefore conclude that the binding of a target alone is insufficient to deterministically produce a conformational transition to the high-affinity form (induced fit). Consequently, in the absence of calcium, targets are interacting with calmodulin molecules that are predominantly in their calcium-free form. Conversely, at the high concentrations of calcium, targets are interacting predominantly with calmodulin molecules in their calcium-bound form. In the absence of direct experimental evidence, such considerations are of course qualitative and must be regarded as assumptions. On the other hand, more dramatic peptide-induced effects, such as folding of previously unstructured portions of the protein, can probably be safely ruled out, given structural evidence that calmodulin is a well-folded protein over a wide range of calcium concentrations, and that its conformational transitions mostly occurs by rearrangement of well-folded secondary structural elements [45, 46]. Moreover, our model explicitly contemplates the possibility of conformational transitions for calmodulin bound to a peptide but not to calcium (see Methods). Whether this T to R transition in the peptide-bound state is triggered by the presence of target peptide, or whether the R state is merely stabilised by the peptide after a transition produced by spontaneous thermodynamic fluctuations, is a very subtle question. The latter scenario however still provides a simple and internally coherent mechanistic explanation.
The work presented in this article deviated from a model previously published by our group, because the conformational transitions of calmodulin’s lobes are no longer assumed to be fully concerted. Concerted transitions were an assumption that greatly simplified the resulting model but, we later found, also conflicted with a published experimental paper, which gave evidence that asymmetric conformational states are possible [26, 47]. Moreover, several authors had previously argued in favour of lobe independence or very high degree of autonomy [28, 29, 48, 49]. Lobe independence also offer a natural explanation as to why targets that interact with only one lobe (such as the WF10 peptide by Bayley et al.) seem to only modulate half of the saturation curve. A fully concerted model would not be suitable, for example, to incorporate the effects of an important neuronal protein such as neurogranin, which mostly interacts only with the C-lobe of calmodulin [37]. Even if the conformational transitions were not fully concerted, there could be a measure of coupling between lobes, i.e. the state of one lobe could affect the probability of conformational transition of the other. Much of our parameter estimation work rests upon the assumption that there is no evidence of strong coupling effect. As for the possibility of weak coupling, there is no clear consensus in the literature. If coupling effects were strong between the lobes, the behaviour of the tryptic fragments containing one calmodulin lobe would show evident divergence from the behaviour of the lobes in the intact protein (which can be monitored separately as shown by [31]). However, several experimental studies show that lobes of intact calmodulin, and tryptic fragments, have very similar calcium binding properties [3, 7, 28, 29]. Given the scattering of experimental data, weak coupling effects, with a magnitude smaller than the noise level, cannot be ruled out with absolute certainty, but including them would have complicated the model with no visible benefit, and therefore we decided to omit them. The two lobes may not be absolutely independent, but they clearly enjoy a remarkable degree of autonomy, and most definitely they are not rigidly coupled. There is also experimental evidence that the actual “cooperative” unit in the EF-hand protein family is the four-helix domain (i.e. in the case of calmodulin, the globular domain rather than the whole molecule) [4, 30]. The lack of lobe coupling, with respect to the conformation selection, implies that the early saturation of the high-affinity C-lobe at lower calcium concentrations does not promote an earlier transition of the low-affinity N-lobe to the R-state. If such coupling were present, the activation of the N-lobe in the intact protein would be “helped” by the higher-affinity C-lobe. The N-lobe in the intact protein would therefore show higher affinity than the isolated N-lobe. Experimental evidence, however, shows that this is not the case, at least within the level of precision allowed by the noise on the experimental data [3, 28].
We have shown that allosteric regulation can explain the modulation of the calcium-binding properties of the TR2C fragment (and hence, of the C-lobe of calmodulin) by several calmodulin-binding peptides. The immediate consequence is that for a given concentration of target, the parameter that determines the extent of the saturation curve shift, at steady-state, is the ratio of the affinities of the target for the R and T states of calmodulin. Plotting the experimental points against the fitted curves in the Hill plane also offer an explanation as to why the MWC and Hill models can both provide a very good fit, despite predicting two quite different qualitative behaviors: in the Hill plane, the Hill curve is a straight line, whilst the MWC model predicts a gradual shift from a lower to a higher asymptote. As shown in Fig. 8, in the range of physiologically plausible calcium concentrations (from nanomolar to tens of micromolar) the MWC model is close to the Hill model. The Hill coefficient predicted by the MWC model also decreases in the presence of targets, in qualitative agreement with experiments (Table 4). The simulations of intact calmodulin in the presence of WF10 shows that biphasic saturation curves can be produced by targets that only stabilise the high-affinity form of one lobe. WF10 is an artificial peptide, but targets that induce differential modulation of calmodulin domains can be found in nature, as in the case of the anthrax edema factor [50].
An allosteric model of calmodulin was previously developed in our group that could satisfactorily reproduce the dose-response curve of wild-type calmodulin and also explain differential activation of PP2B and CaMKII during synaptic tetanic stimulation [26]. The previous model was however based on different premises and some additional simplifying hypotheses, such as concerted conformational transitions of both lobes, and exclusive binding of the target by one conformation. A direct comparison would therefore not be meaningful.
Calmodulin in vivo is always in the presence of a large number of targets, which simultaneously modulate its activity. We used our model to investigate the effect of mixtures of competing targets on calmodulin’s C-lobe. We exploited the seemingly predominant role of the C-lobe in mediating calmodulin-target interactions to test the behavior of our model in the simultaneous presence of two different targets that had higher affinity for the T state and the R state, respectively. As an example scenario we chose peptides of abundant neuronal proteins, whose binding affinities for TR2C were available in the literature. We chose a 1:1:2.5 molar ratio for the three proteins, (a scenario where the concentration of calmodulin-binding proteins is much higher than the concentration of calmodulin [51]), to account for the higher concentration of targets that favour the R-state. In the chosen conditions, the effect of the competing targets was almost cancelled out and the resulting calcium saturation curve was close to that in the absence of targets, as shown in Fig. 9. Regulating the relative abundance of targets can therefore bidirectionally tune the amount of calcium bound to calmodulin’s C-lobe. At any given concentration, the target with the higher affinity is expected to exert the stronger effect.
A two-state model is of course a simplification of something as complex as a protein’s conformational dynamics. It is more plausible that calmodulin is capable of sampling a wider ensemble of conformations, and its high conformational plasticity allows it to regulate downstream protein targets that are structurally very diverse [10, 25, 52]. Moreover, the PEP-19 peptide, expressed in the cerebellum, was shown to regulate mostly calmodulin’s calcium binding kinetics, with little effect on affinity, and could do so even when calmodulin was already associated to CaMKII [53]. These observations imply that the tuning of calmodulin’s affinity and kinetics is highly adaptable, but capturing every possible feature of this mechanism was outside the purpose of the present work. For example, our model in its present form doesn’t allow calmodulin to bind more than one target, which is true for many targets but not for PEP-19 [53], but our primary focus was to present a putative mechanism to explain how a rather large class of targets modulate calmodulin’s affinity, as shown for example by [6]. For the conditions focussed on here, the agreement of a two-state model with experimental results was already satisfactory, and the regulatory properties of the above mentioned PEP-19 peptide (which is cerebellum-specific) seem to be the exception rather than the rule [53]. Therefore, the model in its present form should be applicable to a quite wide range of biological scenarios. On a technical level, parameter estimation is made challenging by the uncertainties and the scattering shown by the available experimental data, but the differences in experimental protocols do not seem so severe to invalidate the interpretation of the results (see S1 Table). Thanks to the fact that calmodulin is perfectly conserved across all known mammals, protein sequences were most definitely identical, and did not represent a potential source of uncertainty in the experimental results. In fact, for reasons of consistency, a substantial amount of published data on paramecium calmodulin, which is only 88% identical to the mammal one [10, 13], was not included in the datasets used in this work. Also, we were reassured by the fact that the vast majority of the reported data was obtained in very similar conditions with regard to pH, temperature, and ionic strength. The most significant difference in the buffer types was the amount of magnesium added to the solution, which varied from 0 to 1mM (see S1 Table). The effects of buffer conditions on calmodulin were reviewed by Ogawa and Tanokura [54], which found that pH and temperature had a measurable but weak effect on affinity and cooperativity. The uniform levels of added salt were relevant because higher ionic strength makes calcium affinity largely independent of pH within the range 6.5–8, as noted by [29], and lowers the effect of added magnesium on calcium affinity of calmodulin [54]. As discussed by [55], the main role of magnesium is most likely the stabilization of the calcium-free form, possibly to prevent proteolytic cleavage, which should not be a problem in vitro.
Ideally, allosteric models would have to be parametrised using data sets produced in highly standardised conditions, designed to be as close to the in vivo conditions as possible. However, different groups performed similar experiments and measured apparent calcium affinities that sometimes differ by up to a factor of two or more. This could be due to a number of reasons, including variable levels of purity of reagents or proteins. As a consequence, pinpointing the exact value of some parameters can be challenging. However, imprecisions in the estimation of parameters do not invalidate the general behaviour of our model. For example, the standard deviation on the fitted value of the allosteric constant is rather large. On the other hand, this is mostly due to the low sensitivity of the saturation function with respect to this parameter: even an estimation error of a factor of two (which is still relatively small, and close to the noise level) would shift the saturation curve by an amount that is comparable to the scattering in the available literature data, as shown in Fig. 12.
The main goal of our mechanistic model was to test the consistency of a plausible physical mechanism with respect to the experimental observations, rather than to achieve a “perfect fit”. To justify this approach, we must first acknowledge that there exist two main types of models. The first type comprises what we call phenomenological models, which fit an analytical curve to the data points, using as many parameters as necessary to represent the data within defined limits. Such models can fit the data well, but do not necessarily provide insights into the underlying physical mechanism. The second type is that of what we call mechanistic models, where we make reasonable assumptions about a (possibly simplified) molecular mechanism that underlies the observed behaviours. A mechanistic model must abide the restriction that its parameters must have well defined physical meaning, associated with an elementary physical phenomenon. A mechanistic models will almost always give a poorer fit, but its contribution to knowledge lies in its explanatory character whereby observed complex behaviour can emerge from simpler physical principles. Given the above mentioned limitations in the precision of the available data, our message is that a major portion of the observed properties of calmodulin can be explained by a model of the second type, with the limitation that at the present state of experimental knowledge, a more precise determination of parameter values is not possible.
We have shown that a classic MWC allosteric model can successfully fit the saturation curves of calmodulin lobes and of the intact protein, and also account for the effects of several biologically relevant targets. This wide applicability was also achieved with a comparative economy of hypotheses (independence of two domains, each in thermodynamic equilibrium between two conformations that have different affinities for the ligand and also different affinities for each target), thus providing a useful conceptual framework upon which further modelling work can be constructed. A crucial point is that the apparent affinity of calmodulin is modulated by simultaneous dynamic equilibria with different targets that exhibit a preference (tighter binding) to one of the possible conformations of calmodulin.
Each isolated lobe of calmodulin exhibits cooperativity between its two calcium binding sites. We chose to model the isolated C-lobe (TR2C tryptic fragment) as a two-subunit, two-state MWC molecule, where the subunits undergo concerted conformational transitions. As a starting point, we assumed that both calcium-binding sites on the C-lobe were identical, as was observed for example for the CaM85/112 mutant [22]. Therefore, in our model, both sites have calcium affinity KT when they are in the T state, and a higher affinity KR when the lobe is in the R state. Targets can bind both conformations of the molecule, but with different affinities, as shown by several experimental studies [7, 12]. A diagram of the resulting model is given in Fig. 13.
Under the simplifying assumption, used throughout this work, that the two calcium-binding sites are identical, the following equations can be derived for the saturation function of TR2C in the presence of an allosteric target A [15]:
Y
¯
=
α (
1 + α
) +
L
′
α c (
1 + α c
)
(
1 + α
)
2
+
L
′
(
1 + α c
)
2
(5)
L
′
= L
(
1 + γ e
1 + γ
)
2
(6)
where:
L = T 0 / R 0
(7)
is the allosteric constant in the absence of targets (defined as the ratio between the concentrations of the T and R states, when no ligand is bound), L′ is the allosteric constant in the presence of targets,
α = [ X ] f r e e / K R
(8)
is the ratio between the ligand affinity of the R state and the concentration of free ligand,
c = K R / K T
(9)
is the ratio of the ligand affinities of the R and T state,
γ = [ A ] f r e e / K R t
(10)
is the ratio between the concentration of free allosteric target, [A]free and the target affinity of the R state, KRt, and
e = K R t / K T t
(11)
is the ratio between the target affinity for the R state KRt, and that for the T state KTt. The above equations allow the calculation of the saturation level of TR2C in the presence of a known concentrations of free ligand (calcium) and allosteric targets (peptides). They also clearly show that the effect of targets is equivalent to a modulation of the allosteric constant, whilst the other parameters remain unaffected. The function Y ¯ describes a surface in the plane (α,γ), as shown in Fig. 14. The ligand saturation of an allosteric molecule, at equilibrium and at a given concentration of free ligand and free target, is entirely determined by the following parameters (note that all affinities are expressed as dissociation constants):
10.1371/journal.pcbi.1004063.g014
Figure 14
Saturation function of an MWC allosteric model of a two-state molecule with 2 binding sites, such as TR2C, in the presence of a target with preference for the R state.
Increasing concentrations of target produce a left shift in the calcium saturation curve.
- the allosteric constant L (or isomerisation constant), defined as the concentration ratio between the R and T states in the absence of ligand.
- the ligand affinity of the binding sites in the T state, KT;
- the affinity change upon transition from T to R state, c = KR/KT;
- the affinity for the target when the molecule is in the T state, KTt;
- the change of affinity for the target upon transition from the T state to the R state, e=KRt/KTt.
A diagram of an MWC model of TR2C is given in Fig. 13. However, most available experimental data were obtained performing calcium titrations of calmodulin the presence of a fixed total concentration of target, and therefore Equation 5 was not directly applicable for fitting purposes. This apparent difficulty can be circumvented analytically, by calculating the corresponding concentrations of free target, as shown for example by Martinez et al. [56]. However, we found convenient to follow a computational route, directly implementing the model in the biochemical pathway simulator COPASI, which supports parameter fitting of steady-state properties when a system is initialised with total concentrations of reagents in a reaction compartment. The analytical model was instead used for a preliminary study on the general behavior of an MWC model for several choices of parameters, and also to discriminate between alternative possible parameterisations, as shown in the Results section.
Several published steady-state titration curves described how calmodulin and TR2C fractional saturation at varying concentrations of free calcium change in the presence of different peptides that mimic the consensus motif of several in vivo binding partners of calmodulin. Published datasets were digitised using the freely available PlotDigitizer program. The datasets of Bayley et al. [7] and Theoharis et al. [12] were of particular interest because they reported the affinity for the peptides both in the absence and at saturating concentrations of calcium. In an allosteric perspective, these data provide estimates of the intrinsic affinity of the peptide for the T-state (predominant in calcium-free calmodulin) and R-state (predominant in calcium-saturated calmodulin). These affinities were thus known quantities in our model. The properties of the peptides used in the above mentioned experiments are summarised in Table 5.
It must be noted that none of the experimental datasets in the literature reported error bars or standard deviations for the measured datapoints, even when the datapoints were averages of measurements performed in triplicate [12]. In reality, all plotted datapoints are affected by uncertainty both on the x (free calcium concentration) and y axis (saturation) although free calcium concentration is usually known to high precision [6]. We estimated the TR2C model parameters by fitting experimental calcium saturation curves of TR2C alone, TR2C+WFF peptide (1:1.4 molar ratio), TR2C+Nav1.2IQp peptide (1:1.4 molar ratio). A fourth and fifth data set, TR2C+WF10 (1:1.4 molar ratio) and TR2C+Nav1.2IQp peptide (1:2.8 molar ratio) were not used for parameter fitting and were kept for validation purposes. Data sets were taken from references [7, 12].
The resulting computational model had many parameters but, crucially, the MWC equations constrained the majority of them to be known functions of a a much smaller number of free parameters. These free parameters were the only ones that needed to be fitted. In the absence of targets, the COPASI model contains 3 independent parameters, (L,KR,KT). In the presence of target t, the saturation curve depends also on the target affinities K R t, K T t, which were both known quantities for WFF. The affinities of the peptide NaV1.2IQp for both the T and R states were in the nanomolar range, and only estimated upper limits for their value were available in the literature (Table 5). We assumed that NaV1.2IQp affinity for the R state was equal to the estimated upper limit, while its affinity for the T state was set as an additional parameter to be fit, which we called K T N a V. In total, the 4 independent parameters KT, KR, K T N a V and L were required to fit simultaneously the three experimental data sets, i.e. the score function minimised in the fitting was the total sum of square residues from all three available datasets. The contributions of each dataset to the total sum of square residues were weighted to account for the fact that not all datasets contained the same number of data points. When choosing the initial conditions for the fitting procedure we assumed that our model must account for a wide range of observed behaviors. An initial exploratory study on different possible parameterizations was performed using the analytical MWC model described in the previous paragraphs, using custom Python scripts. These preliminary results were used to formulate additional constraints on the parametric space that the computational model was allowed to sample, when fitted to the experimental data describing the saturation curve of TR2C in the presence of targets. In the MWC framework, the apparent ligand affinity is produced by a mixture of two distinct populations of molecules, in the R-state and T-state. Target peptides shift the saturation curve by differentially stabilising the R and T state. Therefore, any observed saturation curve, with or without targets, is bound to lie between two limit curves corresponding to populations of 100% R state, and 100% T state. The two fully-stabilised states also exhibit no cooperativity, because no further population shift is possible. These observations provide a simple way to determine lower bounds for KT and upper bounds for KR, which in turn provide an upper bound for c = KR/KT. The saturation curve for TR2C in the presence of Nav1.2IQp showed an apparent affinity of about 10μM and a Hill coefficient greater than 1, indicating that the T-state was not fully stabilised [12]. Therefore the affinity of the pure T-state must be lower than 10μM. On the other hand, data obtained with full-length calmodulin, showed that some peptides can increase affinity and decrease cooperativity, shifting the saturation curve to the left to an extent that implies that the calcium affinity of the pure R state must to be higher than 20nM [6], assuming that the C-lobe is responsible for the higher affinity when CaM is bound to a target, as shown by data published by Shea et al. [9]. Taken together, these observations led to the following constraints on the allowed parameter values:
K
R
≤ 20 n M
(12)
K
T
≥ 10 μ M
(13)
c =
K
R
K
T
≤ 0.002.
(14)
The fitting was performed using the built-in parameter estimation functions of COPASI, using 1000 iteration of a genetic algorithm with stochastic ranking [57], with a population size of 20, and with the constraints on the parameter values described in the previous paragraph. A genetic algorithm is a non-local optimisation method and is therefore less prone to converging to local minima in comparision to gradient-based methods.
The Hill coefficients of calmodulin and calmodulin in the presence of targets were calculated as described in Ref. [58], as the slope of the saturation function in the Hill plane:
n
H
=
d (
log
Y
¯
1 −
Y
¯
)
d (
log α
)
(15)
where α = [Ca2+]free/KR is the free ligand concentration normalised by the ligand affinity of the R state, and the fractional saturation function, Y ¯, is defined as the ration between the number of occupied binding sites and the total number of binding sites, at a given concentration of free ligand, and can be calculated as:
Y ¯ = [ C a 2 + ] b o u n d 2 · [ T R 2 C ]
(16)
For an MWC model with two identical subunits, the Hill coefficient as a function of ligand concentration describes a symmetric bell-shaped curve, with its maximum at the point of half-saturation.
The model of isolated N-lobe (corresponding to the TR1C tryptic fragment) was formally identical to the TR2C fragment described in the previous paragraphs. However, the parameter estimation required a different approach. We could not directly fit the model to saturation curves in the absence and presence of targets, because the affinity of the isolated N-lobe for targets is very low [7], and the resulting shift of the saturation curve very weak [12]. Moreover, targets that bind calmodulin in calcium-free conditions have little or no interaction with the N-lobe, and their effect on its saturation curve was negligible. First we assumed, for the sake of simplicity, that the two calcium-binding sites of the N-lobe are identical (as we did for the C-lobe). Based on the experimental evidence [6], that the two lobes have very similar affinity when bound to high-affinity targets, we made the additional assumption that, when in the R state, both lobes had the same calcium affinity. The affinity of the C-lobe in the R state was already known from the parameterization of the TR2C model. We estimated the calcium affinity of the N-lobe in the T state by fitting the saturation curve by Grabarek and coworkers for the NCaM41/75 mutant (where the N-lobe was constrained in a closed conformation by a disulfide bond) with a non-cooperative model with two identical binding sites. Knowing the affinity of both the R and T state, the allosteric parameter c = KR/KT was readily calculated. The only remaining free parameter in the model was the allosteric constant L, which was estimated by fitting and analytical MWC model onto the available experimental points for the saturation curve of the N-lobe [3, 9, 31].
|
10.1371/journal.pgen.1007833 | Amino acid residues in five separate HLA genes can explain most of the known associations between the MHC and primary biliary cholangitis | Primary Biliary Cholangitis (PBC) is a chronic autoimmune liver disease characterised by progressive destruction of intrahepatic bile ducts. The strongest genetic association is with HLA-DQA1*04:01, but at least three additional independent HLA haplotypes contribute to susceptibility. We used dense single nucleotide polymorphism (SNP) data in 2861 PBC cases and 8514 controls to impute classical HLA alleles and amino acid polymorphisms using state-of-the-art methodologies. We then demonstrated through stepwise regression that association in the HLA region can be largely explained by variation at five separate amino acid positions. Three-dimensional modelling of protein structures and calculation of electrostatic potentials for the implicated HLA alleles/amino acid substitutions demonstrated a correlation between the electrostatic potential of pocket P6 in HLA-DP molecules and the HLA-DPB1 alleles/amino acid substitutions conferring PBC susceptibility/protection, highlighting potential new avenues for future functional investigation.
| Primary Biliary Cholangitis (PBC) is a chronic autoimmune liver disease that exhibits strong genetic associations, especially with variants in the human leukocyte antigen (HLA) gene region. Here we use dense single nucleotide polymorphism (SNP) data from the largest PBC study to date (2861 cases, 8514 controls) to investigate the likely underlying causes of this association, via performing imputation of HLA classical alleles and amino acids. We show that the HLA association can be largely explained by variation at five separate amino acid positions, one of which shows functional relevance to electrostatic potentials of HLA-DP molecules.
| Primary Biliary Cholangitis (PBC; formerly known as Primary Biliary Cirrhosis) is a rare cholestatic liver disease characterized by progressive auto-immune destruction of intrahepatic bile ducts resulting in cholangitis, liver fibrosis and, eventually, cirrhosis. Candidate gene studies have consistently demonstrated association with polymorphisms in the human leukocyte antigen (HLA) region [1–8]. Genome-wide studies [9–12] have confirmed these HLA associations and have also identified 27 non-HLA risk loci. The MHC region, however, remains by far the strongest genetic contributor to disease susceptibility, with HLA haplotypes containing the HLA-DQA1*04:01 allele conferring an approximately threefold increased disease risk [13].
To help understand the mechanisms underlying these HLA associations, and to identify functional, potentially causal, variants within the HLA region, we used previously-generated dense SNP genotype data from 2861 UK PBC cases and 8514 UK controls [13] to impute classical HLA alleles and amino acid polymorphisms within these 11375 individuals, using current state-of-the-art methods implemented in the software packages HLA*IMP:03 [14], HLA*IMP:02 [15], HIBAG [16] and SNP2HLA [17]. Previous interrogation of the UK PBC case/control data set risk [13] using classical HLA alleles imputed using the package HLA*IMP:01 [18, 19] had revealed four haplotypes showing independent disease associations: the well-established association at HLA-DQA1*04:01 (which forms a haplotype with HLA-DQB1*04:02 and HLA-DRB1*08:01), two previously identified protective effects marked by alleles HLA-DQB1*06:02 [20] and HLA-DQB1*03:01 [1], and a novel association marked by the haplotype HLA-DRB1*04:04/HLA-DQB1*03:02. Similar associations were also observed in application of HLA*IMP:01 to a smaller separate European data set [21]. Our updated analysis of the UK PBC data set, reported here, confirms these previously-observed associations, suggests potential additional independent associations, and suggests that the majority of the SNP and classical allele association in the HLA region can largely be explained by variation at five separate amino acid positions.
Various software packages have been developed for the imputation of classical HLA alleles (and, in some cases, amino acid substitutions) using dense SNP data; here we used the current state-of-the-art packages HLA*IMP:03 [14], HLA*IMP:02 [15], HIBAG [16] and SNP2HLA [17], and compared the results obtained for classical HLA alleles with those previously obtained [13] using HLA*IMP:01 [18, 19]. Our rationale for using four different software packages was the fact that the precise methodology implemented varies across the different packages, as do the reference sets used to inform the imputation. Thus, we were interested in examining the sensitivity of our findings to the software implementation used, with concordance of findings seen across different software implementations providing a greater degree of confidence in the results obtained.
Analysis of the UK PBC data set using these packages confirmed the previously-observed associations [13, 21] seen with classical HLA alleles (Table 1, haplogroups 1–4) and suggested potential additional novel independent associations at HLA-DPB1 (HLA-DPB1*03:01, HLA-DPB1*06:01, HLA-DPB1*04:01, HLA-DPB1*10:01 and HLA-DPB1*17:01), HLA-C (HLA-C*04:01) and HLA-DPA (HLA-DPA*02:01) (Table 1, haplogroups 5–10). Results were largely concordant across different HLA imputation programs whenever the same alleles were interrogated. Our understanding is that HLA-DPB1 was not included in the reference set used by HLA*IMP:01 and so could not be assessed in previous analyses [13, 21] using this software. Further associations with HLA-DQA1*03:01 identified in our current analysis (Table 1, haplogroup 4) but not reported in previous analysis of these data [13] appear to be part of the previously identified HLA-DRB1*04:04/HLA-DQB1*03:02 haplogroup.
Our detected association at HLA-DPB1*03:01 is consistent with results from HLA imputation in a medium-sized Italian PBC data set (676 cases and 1440 controls) in which Invernizzi et al. [22] used the Beagle software [23], in conjunction with the T1DGC HLA reference set, to demonstrate association of HLA-DPB1*03:01 with disease [22]. Previous much smaller studies had generated somewhat contradictory results, with some showing [24] and others not showing [8] association between HLA-DPB1*03:01 and PBC.
In addition to examining the effects of individual classical HLA alleles, we also used the HIBAG imputed dosages to perform multi degree-of-freedom (df) omnibus gene-based tests, examining the effects of all alleles (with frequency > 0.5%) at a gene simultaneously, although we note that the large (and differing) numbers of alleles at each HLA gene makes this procedure arguably less powerful and interpretable than the testing of individual alleles. All genes showed highly significant marginal association (S1 Table), with all genes except HLA-DQB1 retaining some level of support even conditional on the other genes (i.e. when all non-rare alleles at all other genes were included in the model).
For our primary analyses, we did not consider it necessary to include additional covariates such as gender or principal component scores (PCs) in the regression model to account for possible population stratification (see S1 Text); an investigation of the sensitivity of the results to inclusion or not of these covariates (S2 Table) suggested that the odds ratios (ORs) and P-values achieved were largely unaffected by the inclusion or not of the top 10 PCs (calculated from a pruned set of SNPs—with SNPs in the extended HLA region removed) and were only slightly altered by the additional inclusion of gender as a covariate. We also found that the marginal associations of the lead allele from each haplogroup shown in Table 1 were largely reproduced when modelled as part of a 9 variable model (with all lead alleles included simultaneously) (S3 Table).
Given that classical HLA alleles encode combinations of amino acid substitutions at specific positions, and given the a priori functional relevance of amino acid substitutions, we next focussed our attention on the variables encoding their effects. To determine whether the association seen between PBC and HLA SNPS and/or classical alleles could be explained by the amino acid substitutions encoded by the associated HLA alleles, we tested the imputed dosage of each amino acid residue at each amino acid position for association with PBC using logistic regression (Fig 1A, Table 2, S1 Spreadsheet). The marginal association results obtained using classical alleles imputed from HIBAG were seen to be highly concordant with those obtained directly from SNP2HLA; we focus here on the results from HIBAG which allows more complex subsequent modelling via regression and stepwise regression approaches (on account of outputting a posterior probability for each possible genotype, in contrast to the estimated dosage output by SNP2HLA, see Methods).
The strongest association (P = 6.64x10-59) was seen with residue L (the substitution leucine for glycine) at position 11 of HLA-DPβ1; an equivalent association was seen with residue G (reflecting the fact that at this position there are only two possible substitutions, and so the test of L versus G is equivalent to the test of G versus L). Once the variable encoding this effect had been included as a covariate in the regression model, the next most associated amino acid (P = 1.73x10-39) was residue L at position 74 of HLA-DRβ1. HLA-DRβ1 74L was previously identified as significantly associated with PBC by Donaldson et al. [1] and Invernizzi et al. [22], as well as by an earlier small Japanese study (53 PBC patients and 60 controls) [5]. This substitution occurs both on classical alleles HLA-DRB1*08:01 (which is strongly associated with PBC in Europeans) and on HLA-DRB1*08:03 (which is known to be associated with PBC in Japanese/Chinese populations [4, 5, 25] and thus offers a potential explanation for the HLA-DRB1*08 associations seen in these different populations. Our identification of amino acids in HLA-DPβ1 and HLA-DRβ1 as the top contributors to HLA-induced PBC risk is consistent with the results of Invernizzi et al. [22] who found in their (much smaller) Italian data set that conditioning on residue L at position 11 of HLA-DPβ1 largely removed the signal at HLA-DPB1, and who noted that, considered together, HLA-DRB1*08 and HLA-DPB1*03:01 accounted for the majority of the signal in the HLA region.
Using a stepwise regression approach similar to that used in previous studies [26–28], we continued adding amino acid residues into the regression model in a stepwise fashion to account for their effects [29] until none reached significance level < P = 4.87x10-5 (representing a Bonferroni-corrected threshold of 0.05, allowing for 1028 amino acids tested); this resulted in a final model that included nine amino acids (Table 2). Use of a more stringent stopping threshold of P = 1.0x10-8 resulted in a final model that included five amino acids in five separate genes (Table 2, Fig 1A–1F, S1 Spreadsheet). None of the top nine or top five amino acids dropped out of the model (all P>4.87x10-5) when allowing a backward stepwise step (S4 Table). Stepwise inclusion of the top five amino acids in association analyses carried out with respect to individual SNPs (Fig 2, S2 Spreadsheet) or classical HLA alleles (Fig 3, S3 Spreadsheet) indicated that these five amino acids could account for the majority of the HLA association seen at the level of SNPs or classical alleles; once these five amino acids had been included, the minimum significance levels achieved were P = 7.05x10-9 for SNPs and P = 1.98x10-7 for classical alleles. Thus, although some residual association remains, inclusion of the top five amino acids is sufficient to remove the strongest disease associations observed. Interestingly, in spite of the strong linkage disequilibrium (LD) across the HLA region (resulting in haplogroups spanning multiple genes, see Table 1), visual inspection of Fig 3 suggests that each of the five implicated amino acid residues accounts primarily for the disease association observed with classical alleles of its own gene, although HLA-DRβ1 74L does partly account for association seen at HLA-DQA1 and HLA-DQB1, and HLA-DQβ1 57D in turn partly accounts for association seen at HLA-DQA1 and HLA-DRB1, probably due to the long-range correlations (due to extensive LD) between alleles (and thus between amino acid substitutions and classical alleles) at different genes.
We investigated the sensitivity of our results to the inclusion of the top 10 principal component scores (calculated from a pruned set of SNPs—with SNPs in the extended HLA region removed) in order to account for possible population stratification, and also to the inclusion of gender, as covariates in the regression equation, but found (as expected from theoretical arguments, see Methods), that this had little impact on the results obtained, either with respect to the main effects of amino acids (S1 Fig, S4 Spreadsheet) or with respect to the stepwise entry of predictors. With principal component scores included, exactly the same top five amino acids entered the model, while, with principal component scores and gender included, four out of the top five identified amino acids (all still with P<1.0x10-8) remained the same, with HLA-DQβ1 87F entering the model in preference to the marginally less significant HLA-DQβ1 57D, and with the order of entry slightly altered as follows: (1) HLA-DPβ1 11L/G (P = 5.70x10-53), (2) HLA-DRβ1 74L (P = 4.67x10-33), (3) HLA-DQα1 -13A (P = 2.34x10-20), (4) HLA-DQβ1 87F (P = 1.18x10-13), and (5) HLA-C 156R (P = 2.23x10-10).
We additionally investigated the consistency/stability of the predictors identified by the stepwise selection procedure using a resampling approach (see Methods), and found the approach to be highly stable in terms of the top amino acid predictors identified. In 1000 bootstrap replicates (each containing 2/3 of our PBC cases and 2/3 of our controls, see Methods) the top amino acid HLA-DPβ1 11L/G entered as the most significant predictor in 86.7% of replicates (and entered as second in the remaining 13.3% of replicates). The second amino acid to enter was HLA-DRβ1 74L in 83.1% of replicates. The third amino acid to enter was HLA-DQβ1 57D in 51% of replicates (the closest competitor was HLA-DRβ1 67L which entered third in 25.8% of replicates). The fourth amino acid to enter was HLA-C 156R in 33.7% of replicates (the closest competitor was HLA-C 152A which entered fourth in only 10.3% of replicates). The fifth amino acid to enter was HLA-DQα1 -13A in 27.9% of replicates (the closest competitor was HLA-DRβ1 58A which entered fifth in only 10.5% of replicates).
The association seen between PBC and HLA-C 156R is intriguing, as HLA-C is known to be not very potent in antigen presentation. Given that HLA-C has a significant role in interaction with killer-cell immunoglobulin-like receptors (KIRs), this raises the question of whether the HLA-C association is related to T-cell interaction, or is rather about presentation to KIRs. We therefore used the software package KIR*IMP [30] to examine the association between PBC and genes on chromosome 19q13.4 that encode for KIRs. In contrast to the detected HLA associations, however, analysis of PBC association with imputed KIR haplotypes and copy number variation detected no significant associations (minimum observed P value = 0.07) between PBC and KIR variation. We therefore chose not to focus any further attention on the KIR gene region at this current time.
We continued our investigation by examining the association between PBC and amino acid residues, SNPs or classical HLA alleles simultaneously, by allowing either amino acid residues and/or SNPs and/or classical HLA alleles to enter the stepwise regression model at each step. At each of steps 1–2, an amino acid residue (HLA-DPβ1 11L at step 1 and HLA-DRβ1 74L at step 2) entered the model preferentially in comparison to a SNP or a classical HLA allele. This ability of amino acid substitutions to explain the association of the MHC to PBC in a more parsimonious way than is achieved by classical HLA alleles contrasts with results previously found using stepwise regression in inflammatory bowel disease (IBD) [27], where classical HLA alleles (specifically HLA-DRB1*01:03) entered the model first and better explained the association than models based on amino acid substitutions, leading the investigators in that study to focus their subsequent efforts on an HLA-DRB1 centric model. Our results here are more akin to those found using stepwise regression in rheumatoid arthritis [28], where amino acid substitutions entered the model first and were found to provide a better fit, and a more parsimonious explanation for the observed association, than models based on either two- or four- digit classical alleles.
Continuing the stepwise regression procedure, our results at subsequent steps (S1 Text) illustrated the difficulty of disentangling “causal” from “hitchhiking” effects amongst highly correlated variables such as the amino acid residues, classical alleles and SNPs considered here—although it is noteworthy that in each of steps 1–4 an amino acid always entered the model in preference to a classical allele. In most cases, the difference in model fit between including the top SNP and the top amino acid or classical allele was relatively small. Given the a priori potential functional role of amino acid substitutions, we found it most natural to focus primarily on variables directly encoding these effects. The fact that, in some instances, inclusion of a SNP provided a slightly better model fit could indicate that the SNP itself is having a functional role (perhaps through a mechanism such as modulation of gene expression) but, equally, could arise from the phenomenon whereby a SNP tags the combined effects of several functional amino acids. In terms of accounting for the overall association in the region, we found the model that included the top 5 amino acids (S2 Fig, left hand panels) performed similarly to the model that included the top 5 variables of any type (S2 Fig, right hand panels, S5 Spreadsheet, S6 Spreadsheet, S7 Spreadsheet).
To explore further the degree to which amino acid substitutions could account for the effects of classical alleles, and to investigate whether such results could occur by chance by tagging classical alleles of differential risk, we used the permutation approach employed in IBD [27]. Specifically, we repeatedly reassigned (permuted) the amino acid sequence assigned to each of the classical HLA alleles, creating a null hypothesis distribution whereby the relationship between classical HLA alleles and disease was retained, but the relationship between amino acid substitutions and classical alleles was permuted. The results (Fig 4, S8 Spreadsheet) indicated that the model deviance accounted for by the top 1–5 amino acids generally fell in the tail of the empirical null distribution, suggesting that the observed amino acid associations were unlikely to have arisen through chance tagging of classical HLA alleles.
To investigate whether alternative amino acid substitutions could provide an equally good explanation for the top amino acid associations, we examined the correlations between our top five residues and other amino acid residues (S5 Table). Substitutions L and G at position 11 of HLA-DPβ1 were perfectly correlated as previously noted; no other amino acid substitution reached r2 > 0.8 with this substitution. Similarly substitutions A and T at position -13 of HLA-DQα1 were almost perfectly correlated; no other substitution reached r2 > 0.8 with this substitution. Five alternative residues showed r2 >0.8 with HLA-DRβ1 74L and thus might be considered plausible alternative causal explanations for the association seen at this residue. No alternative residues reached r2 > 0.8 with the amino acid residues implicated at HLA-DQβ1 57D and HLA-C 156R, suggesting that these effects are unlikely to be attributable to alternative substitutions.
Similar to previous studies conducted in IBD [27], we additionally fitted multi-df omnibus models that included predictor variables encoding the effects of all amino acid substitutions at a position simultaneously (the maximum number of such amino acid variants at a position was 8). This analysis strategy investigates the combined effects seen at a particular position of the amino acid sequence, rather than the effects of individual specific amino acid residues. The results (S1 Text, S6 Table, S3 Fig, S4 Fig, S5 Fig, S9 Spreadsheet, S10 Spreadsheet, S11 Spreadsheet, S12 Spreadsheet) showed reasonable (albeit not perfect) concordance with the results seen when considering individual amino acid substitutions, while incurring the expense of a larger number of df and an arguably less interpretable model. Given that the five individual amino acid residues previously identified do almost as well at accounting for the association as do the multi-df models, overall we tend to prefer the five amino acid model identified through stepwise regression as representing the most parsimonious solution.
Similar to previous studies conducted in rheumatoid arthritis [28], we also performed an exhaustive search for all pairs and all triplets of amino acid residues and amino acid positions, in order to determine the best pairwise or three-way combination associated with PBC. The best pairwise combination of amino acids was HLA-DPβ1 11L and HLA-DRβ1 74L, which corresponds to the top two residues identified through stepwise analysis (Table 2). The best (multi-df) pairwise combination of amino acid positions was position 11 of HLA-DPβ1 and position 27 of HLA-DQβ1, again the same as the top two positions identified through stepwise analysis (S6 Table). The best three-way combination of amino acids was HLA-DPβ1 11L, HLA-DQα1 -13A and HLA-DQα1 53G, which includes the first and 5th amino acids identified through stepwise analysis; this combination was only very marginally better (AIC 12311.42) than the combination DPβ1 11L, HLA-DRβ1 74L and HLA-DQβ1 57D (AIC 12311.51) which corresponds to the top three amino acids identified through stepwise analysis. The best (multi-df) three-way combination of amino acid positions was position 11 of HLA-DPβ1, and positions 58 and 13 of HLA-DRβ1, corresponding to the first, 3rd and 7th positions identified through stepwise analysis (S6 Table).
To move beyond three-way combinations of predictors in an exhaustive search is computationally challenging. We therefore used the FINEMAP [31] and GUESSFM [32] programs which implement (slightly different) Bayesian stochastic search algorithms for selecting important predictors within a densely genotyped candidate region. Preliminary analyses with FINEMAP generated many equivalently-fitting models; we circumvented this issue by filtering out highly correlated amino acid variables, retaining an index set of 396 amino acids with pairwise correlation values less than 0.98 for analysis. The equivalent strategy in GUESSFM was achieved through setting the user-defined input parameter “tag.r2” (the r2 threshold for grouping predictors together into LD groups) as 0.9604 (= 0.982); this resulted in 354 tag groups once monomorphic amino acid positions had been discarded. Following model fitting, the “expand.tags” function within GUESSFM package was then used to expand the set of models considered by GUESSFM to consider all predictors (rather than using a single “tag” amino acid as a surrogate for the other amino acids in its LD group) and the “snp.picker” function used to pick out the resulting amino acids that had the highest posterior probability of inclusion.
The top models and amino acids identified by FINEMAP and GUESSFM respectively are shown in S7 Table and S8 Table. We found the results from FINEMAP and GUESSFM to be somewhat sensitive to the choice of user-defined input parameters, particularly the “nexp” parameter (the expected number of causal variants) in GUESSFM and the maximum number of causal variants in FINEMAP. The results from FINEMAP (S7 Table) were relatively concordant with those from stepwise regression, strongly implicating four out of the top five amino acids from stepwise regression (HLA-DPβ1 11L/G, HLA-DRβ1 74L, HLA-DQβ1 57D and HLA-C 156R), but also providing some support for additional predictors HLA-DQβ1 -4V, HLA-DQβ1 71T, HLA-DQα1 175E and HLA-B 45T (or their correlates). The results from GUESSFM (S8 Table) were more variable and, in general, GUESSFM generated final models that involved a relatively large number of predictors compared to stepwise regression. However, three out of the top five amino acids from stepwise regression (HLA-DPβ1 11L/G, HLA-DRβ1 74L and HLA-C 156R) retained strong levels of support, which was maintained following application of GUESSFM’s snp.picker algorithm (S9 Table).
Given the strong LD in the HLA region, it is perhaps not surprising that GUESSFM ended up preferring models with large numbers of predictors which can better capture subtle haplotype effects. Our comparison between these different analysis approaches again illustrates the difficulty of statistically identifying true causal variants (as opposed to good markers of causal variants) in regions of high LD such as the HLA region. S6 Fig, S7 Fig, S8 Fig, S9 Fig, S10 Fig and S11 Fig (see also S13 Spreadsheet, S14 Spreadsheet, S15 Spreadsheet) illustrate the degree to which the top predictors implicated by FINEMAP and GUESSFM can account for the amino acid, SNP and classical allele associations observed in the region. Although FINEMAP and GUESSFM perform well when larger numbers of predictors are included, they did not generally outperform stepwise regression when limited to 5 predictors. The fact that the five amino acid model identified through stepwise regression performs well at explaining the observed SNP and classical allele association again motivates the five amino acid model as representing arguably the most parsimonious solution.
Given that non-multiplicative effects at HLA have been observed in other autoimmune diseases [33], for each of the classical HLA alleles (Table 1) and amino acid substitutions (Table 2) identified using the 1df multiplicative allelic model, we additionally explored models that allowed the effects to operate via dominant, recessive, genotypic or interaction effects. However, we found little compelling support for such models from the data (S2 Text, S10 Table, S11 Table). In most cases there was little difference in fit between the multiplicative and dominant models, suggesting insufficient data (in particular insufficient observations with two copies of the allele in question) as to be able to distinguish between these two scenarios. We additionally performed pairwise interaction analysis to investigate whether particular combinations of classical HLA alleles or amino acid residues led to increased or reduced risks (over and above their individual multiplicative effects) (S2 Text) but found no evidence of any significant interactions, once Bonferroni correction had been made for the number of tests performed.
To explore the potential functional consequences of changes at the key PBC-associated amino acid residues identified, we followed an approach previously used in primary sclerosing cholangitis [34]. HLA alleles carrying and not carrying the associated residues were three-dimensionally modelled using the program MODELLER 9.14 [35], electrostatic potentials around the resulting 3D structures were calculated using DelPhi 6v2 [36], and the surface of the modelled molecules were coloured according to charge using Chimera [37]. We focussed on the top three amino acid residues (HLA-DPβ1 11L, HLA-DRβ1 74L, HLA-DQβ1 57D) identified through stepwise regression (Table 2), all of which showed strong marginal association with PBC, and we modelled HLA molecules corresponding to alleles showing significant marginal association with PBC (Table 1) that either carried or did not carry the associated amino acid residue. We also investigated the electrostatic potential of residues 56, 70 and 71 in HLA-DQβ1 on account of their strong correlation with the second top amino acid HLA-DRβ1 74L (S5 Table). In relation to the top two residues, it has previously been suggested that polymorphism at position 11 of HLA-DPβ1 has the potential to influence properties of binding pocket P9, while polymorphism at position 74 of HLA-DRβ1 may influence properties of binding pocket P4 [38]. HLA-DQβ1 57D, the critical residue that enters third in the stepwise regression procedure, is known to be associated with protection from type 1 diabetes [39]. Its carboxylate group forms a salt bridge with a conserved arginine at position 76 of HLA-DQα1 that stabilises the heterodimer and may affect peptide binding [38, 40].
Results for HLA-DPβ1 11L (Fig 5) showed a remarkable correlation between the electrostatic potential of pocket P6 in HLA-DP molecules and the HLA-DPB1 alleles/amino acid substitutions conferring PBC susceptibility or protection. The PBC-associated HLA alleles (Fig 5B, top right) all contain an L at position 11 of the amino acid sequence, and show negative potentials, while the protective allele (Fig 5C, bottom right) contains a G at position 11, and shows neutral or slightly positive. Results for the other modelled alleles (S12 Fig, S13 Fig, S14 Fig) were less compelling in terms of demonstrating clear-cut correlations between electrostatic potentials and alleles/amino acid substitutions conferring PBC susceptibility or protection, suggesting that the mechanisms underlying these detected associations may be more complex than can be accounted for by simple amino acid substitution.
In this, the largest such study carried out to date in PBC, we present results from an investigation of the association between classical HLA alleles, the amino acids they encode, and PBC, through an analysis of pre-existing immunochip data using a variety of state-of-the-art HLA imputation software packages. Previous HLA imputation-based studies in PBC [22] have sought to determine whether the amino acid associations observed with PBC could be explained by classical HLA alleles; here we took the view, in common with earlier studies of PBC [1, 8] and of other diseases [8, 28, 34, 39, 41] that, given the functional relevance of amino acid substitutions, a more natural question is whether the association with classical alleles (and with SNPs in the HLA region) can be explained by the amino acid substitutions themselves.
We found that the majority of the strong association between PBC and SNPS in the HLA region and/or classical HLA alleles could indeed be explained by variation at five separate amino acid substitutions. These included two substitutions (HLA-DPβ1 11L/G and HLA-DRβ1 74L) that were previously implicated (at much lower levels of significance) by smaller earlier studies, and three substitutions that represent (to our knowledge) novel findings; once these five effects had been accounted for, there remained some residual association, but this was nowhere near as strong as the genome-wide levels of significance observed in marginal analysis. Given the a priori functional relevance of amino acid substitutions, we thus considered these amino acids (or other residues highly correlated with them) as good candidates for being causal; we note that this viewpoint—which places a strong prior on amino acids accounting for association at classical alleles being a more convincing explanation than classical alleles accounting for association of amino acids (given the fact that classical alleles can effectively be considered as specific combinations of amino acid substitutions)—is, to some extent, borne out by that fact that, in steps 1–4 of our stepwise analysis, an amino acid always entered the model in preference to a classical allele.
In our primary analyses, we focussed on examining the effects of individual amino acid substitutions and classical HLA alleles, reserving for secondary consideration omnibus tests that encode the effects of all amino acid substitutions at a position, or all alleles at a gene, simultaneously. This strategy differs from that used in previous studies of IBD [27] and RA [28], in which the investigators focussed first on omnibus tests, reserving tests of individual amino acid substitutions and classical HLA alleles for secondary consideration. Which analysis strategy is to be preferred is perhaps debatable. We considered the individual testing strategy to be a priori the more appealing and interpretable of the two approaches, on account of having fewer df (one df per residue/allele), which provides higher power than testing all residues/alleles at a position/gene simultaneously. For example, if only one residue out of 8 possible residues at a particular position was actually associated with the disease, then this would be clearly visible when testing individual amino acid substitutions, but the effect might well be drowned out when constructing a combined test of all 8 substitutions (one of which is associated, and 7 of which are not). A further reason for preferring the test of individual residues/alleles is the fact that individual residues/alleles may themselves have specific functional effects, separate from the effects of any other residues/alleles at the same position/gene. In application to our PBC data set, we did not find the multi-df omnibus approach to add substantial significance compared to testing individual predictors, perhaps justifying retrospectively our choice of analysis strategy.
We took forward the implicated amino acid residues from our analyses for three dimensional predictive modelling and calculation of electrostatic potentials, following an approach that was previously used in primary sclerosing cholangitis [34] but which has not, to our knowledge, been previously used in PBC. This analysis demonstrated a correlation between the electrostatic potential of pocket P6 in HLA-DP molecules and the HLA-DPB1 alleles/amino acid substitutions conferring PBC susceptibility/protection, highlighting a potential mechanistic explanation for the observed association that warrants further investigation.
A previous study in IBD [27] addressed this question of electrostatic potentials in a slightly different way, motivated by their finding that classical HLA alleles (specifically HLA-DRB1*01:03) better explained the IBD association than did models based on amino acid substitutions, leading the investigators to focus their subsequent efforts on an HLA-DRB1 centric model. In that study, Goyette et al. [27] followed three dimensional predictive modelling with an analysis of the electrostatic properties around the seven peptide residues (at peptide positions 1, 2, 3, 4, 5, 7 and 9) that are known to make contact within the binding grove, and then clustered electrostatically similar HLA-DR molecules together. (A similar analysis was also carried out focussing only on electrostatic properties in the region affected by amino acid positions 67, 70 and 71, which had shown significant association with IBD). This analysis identified four clusters of HLA-DR molecules sharing similar electrostatic properties with respect to the seven peptide regions (or two clusters sharing similar electrostatic properties with respect to the region encompassing amino acid positions 67, 70 and 71). The resulting clustering of HLA-DRB1 alleles (encoding these molecules) showed that alleles associated with increased risk of IBD generally fell into different clusters than alleles associated with decreased risk of IBD, suggesting that the HLA-DR molecules associated with increased risk of IBD exhibited structural and electrostatic properties within the peptide-binding grove that were largely distinct from those associated with decreased risk.
Given the strong associations of HLA-DRB1 alleles with PBC risk (Table 1), and the fact that our second most associated amino acid (P = 1.73x10-39) is residue L at position 74 of HLA-DRβ1, we investigated how the HLA-DRB1 alleles associated with either increased or decreased PBC risk in our study mapped onto the four clusters of electrostatically similar HLA-DR molecules that had been identified by Goyette et al. In contrast to the results seen in IBD, we did not find a consistent pattern of alleles associated with increased risk of PBC falling into different clusters than alleles associated with decreased risk of PBC, suggesting that the mechanisms underlying the HLA-DRB1 associations with PBC may be more specific than is captured by this analysis.
Although the amino acid substitutions highlighted here represent the most compelling disease-causing factors implicated by our study, we note that high LD in the region and the ability of alternative, more complex, models (such as those examined in our multi-df, FINEMAP and GUESSFM analyses) to account for the disease associations observed means that we cannot definitively rule out the contribution of other factors whose effects are statistically intertwined with the substitutions we have identified; follow-up functional studies will be required to further investigate this question. It has previously been demonstrated that HLA-DRB1*08:01 (DR0801), but not HLA-DRB1*11:01 (DR1101), can bind functional epitopes derived from the dominant autoantigen pyruvate dehydrogenase complex-E2 (PDC-E2) [42]. We hypothesise that epitope analysis (using resources such as the Immune Epitope Database and Analysis Resource) may suggest that the identified significant DQ residues are likely to be equally significant immunologically, with risk, but not protective alleles being permissive for binding of binding of epitopes derived from the immuno-dominant inner lipoyl binding domain of PDC-E2.
The association seen between PBC and HLA-C 156R, implicated by all analysis methods, is intriguing, as HLA-C is known to be not very potent in antigen presentation. However, HLA-C does have a significant role in interaction with killer-cell immunoglobulin-like receptors (KIRs), and this therefore raises the question of whether the HLA-C association is related to T-cell interaction, or is rather about presentation to KIRs. The fact that, in our data set, we observed no association between PBC and imputed KIR haplotypes and copy number variation makes this explanation seem less plausible, but it remains an interesting topic for future investigation.
This study was approved by the Research and Development Departments of all National Health Service (NHS) Trusts participating in this study and by the Oxford Research Ethics Committee C (Oxford REC C reference 07/H0606/96).
The cases, controls and genotype data used here have been described previously [11, 13]. In brief, a total of 2981 PBC cases were contributed by the UK PBC Consortium, a consortium operating within 142 NHS trusts, including all UK liver transplant centres. All cases were of self-declared British or Irish ancestry, over 18 years old with probable or certain PBC. 8,970 controls of self-declared British or Irish ancestry were provided by the 1958 British Birth Cohort and the National Blood Service. Samples were genotyped on an Illumina iSelect HD custom genotyping array at either the Wellcome Trust Sanger Institute (2981 cases and 4537 controls) or the Center for Public Health Genomics at the University of Virginia (4433 controls). Following sample and SNP quality control, we retained 2861 cases and 8514 controls that passed previously-derived quality control checks [13], genotyped at 143,006 SNPs, of which 7848 fell within the extended MHC region [43] on chromosome 6 (ranging from 25,650,000 to 33,426,000 base pairs, Build36). One SNP within the region (rs2394173) showing apparent association with disease status was subsequently excluded from analysis following visual inspection of its cluster plots.
A variety of different software packages have been developed for the imputation of classical HLA alleles (and, in some cases, amino acid substitutions) using dense SNP data; we used the current state-of-the-art methods implemented in the software packages HLA*IMP:03 [14], HLA*IMP:02 [15], HIBAG [16] and SNP2HLA [17], and we compared the results obtained for classical HLA alleles with those previously obtained [13] using HLA*IMP:01 [18, 19]. For a detailed description of the HLA imputation and subsequent association analysis performed using the different software packages, see S3 Text. Motyer et al. [14] recently compared the performance of HLA*IMP:03, HLA*IMP:02, HIBAG and SNP2HLA and found that HLA*IMP:03 and HIBAG gave generally the best (similar) levels of performance, achieving high accuracy (in the range ~90–99%, depending on HLA locus). Earlier studies had shown that HLA*IMP:02 and HIBAG performed well in comparison to HLA*IMP:01 [15, 16] while SNP2HLA performed similarly to HLA*IMP:01 [17].
Although based on similar methodological approaches (and, in our investigation, producing largely concordant results, see Results), the output from the four packages that we considered is not fully comparable. HLA*IMP:03 and SNP2HLA provide the posterior probability for each best-imputed allele (although SNP2HLA does not output this quantity directly but rather converts it to a dosage), while HLA*IMP:02 and HIBAG produce a posterior probability for each possible genotype (i.e. for each combination of two alleles, including combinations that have lower probabilities than the best combination). We found this genotype-based output most convenient for averaging over the possible genotype combinations (while allowing appropriately for imputation uncertainty) and for subsequent imputation of amino acid substitutions. Neither HIBAG nor HLA*IMP:02 directly impute amino acid substitutions (as is done by SNP2HLA) but this can be done manually using the peptide sequences of classical alleles available in the IMGT/HLA database (see below). According to Motyer et al. [14], HLA*IMP:03 can also impute amino acid substitutions, but this feature was not enabled in the development version of HLA*IMP:03 to which we were given access. Given (a) the similar performance of the four packages in our data set with respect to classical HLA allele imputation (see Results), (b) the superior performance of HLA*IMP:03 and HIBAG seen by Motyer et al. [14], and (c) the greater convenience of the output from HLA*IMP:02 and HIBAG, for all subsequent detailed modelling of amino acid substitutions we used the imputations derived from HIBAG.
We also used the software package KIR*IMP [30] to examine the association between PBC and genes on chromosome 19q13.4 that encode for killer-cell immunoglobulin-like receptors (KIRs). In a previous evaluation of KIR*IMP, Vukcevic et al. [30] showed that KIR imputation using the high-density Illumina Immunochip array is extremely accurate, achieving > 98% accuracy for the majority of loci, at least 95% accuracy for half the remaining loci, and > 90% for the rest. For distinguishing the broad A and B haplotype groups, KIR*IMP achieves ~98.5% accuracy. KIR imputation in our data set was informed by 241 genotyped SNPs in the KIR region that matched the 301 SNPs available in the KIR*IMP training data set. The resulting imputed KIR allele and haplotype frequencies in our PBC data set were found to be extremely close to those seen in the KIR training set. Case-control association analysis was carried out using the Unphased package [44] and via logistic regression in R, with predictors corresponding to KIR*IMP’s “best-guess” genotypes (provided that both inferred alleles had posterior probability > 0.8). Analysis of each individual KIR allele (or haplotype) was performed by using the dosage of each allele (or haplotype) as a single predictor variable. Additionally, a multi-allelic (multi-df) omnibus analysis was also carried out by including predictor variables encoding the multiplicative effects of all alleles at a position (or all haplotypes at a set of positions) in the regression model simultaneously.
Stepwise logistic regression [29] was used to assess the importance of variables while accounting for the effects of other, previously detected, effects. Predictor variables encoding an individual’s estimated dosage of the relevant SNP, HLA classical allele or amino acid substitution were included as predictors in a logistic regression equation in a forward stepwise fashion (and were subsequently considered for removal from the model in a backward stepwise fashion). For full details, see S1 Text.
For our primary analyses, we did not consider it necessary to include additional predictors such as principal component scores in the regression model to account for possible population stratification, as prior analysis of this data set [11, 13, 45] has shown little evidence of population stratification (once appropriate QC has been performed to remove outlying individuals) in this UK-based sample. Similarly we did not consider it necessary to include gender as a covariate–even though gender is known to be important in PBC (the disease is more prevalent in women than in men)–as theoretical arguments dictate that inclusion of gender should not bias the results of association tests between disease and genetic factors (outside of the X/Y chromosomes) as gender is not a confounder (it is associated with the disease outcome, but not with the genetic predictors). Indeed, it has been shown [46] that inclusion of known covariates such as gender can even reduce power to detect genetic effects in case-control studies. We subsequently investigated the sensitivity of our results to the inclusion (or not) of principal component scores and gender in the regression model, using the top 10 principal component scores calculated from a pruned (by LD) set of SNPs with SNPs in the extended HLA region removed.
We also investigated the stability of the stepwise selection procedure through a resampling approach motivated by the stability selection procedure of Meinshausen and Buhlmann (2010) [47]. In each of 1000 bootstrap replicates, we randomly selected 2/3 of our cases and 2/3 of our controls to form a new case/control data set and applied stepwise regression to select the top 20 amino acid predictors, noting the order and significance of entry of each predictor in each replicate.
The atomic coordinates (3 dimensional structures) of any HLA molecules carrying or not carrying a significantly associated amino acid residue were determined using comparative protein structure modelling by satisfaction of spatial restraints as implemented by the MODELLER 9.14 computer algorithm [35, 48–50]. HLA proteins of known structure suitable as modelling templates were identified in the Protein Data Bank. The peptide sequences of the target classical alleles were downloaded from the IMGT/HLA database. Sequence alignment was performed with Clustal Omega and manually corrected where necessary. The stereochemical qualities of the modelled structures was verified using the COOT program [51] by assessment of Ramachandran plots and by calculating the least square mean deviation between the template molecule and computed model.
|
10.1371/journal.pntd.0002360 | An Epidemic of Dengue-1 in a Remote Village in Rural Laos | In the Lao PDR (Laos), urban dengue is an increasingly recognised public health problem. We describe a dengue-1 virus outbreak in a rural northwestern Lao forest village during the cool season of 2008. The isolated strain was genotypically “endemic” and not “sylvatic,” belonging to the genotype 1, Asia 3 clade. Phylogenetic analyses of 37 other dengue-1 sequences from diverse areas of Laos between 2007 and 2010 showed that the geographic distribution of some strains remained focal overtime while others were dispersed throughout the country. Evidence that dengue viruses have broad circulation in the region, crossing country borders, was also obtained. Whether the outbreak arose from dengue importation from an urban centre into a dengue-naïve community or crossed into the village from a forest cycle is unknown. More epidemiological and entomological investigations are required to understand dengue epidemiology and the importance of rural and forest dengue dynamics in Laos.
| Dengue disease is caused by a virus transmitted by mosquitoes. In Southeast Asia, where it is endemic, it represents a very important public health problem. Major outbreaks, including severe cases and death, occur every year. Two distinct transmission cycles have been described. Most common is the human-mosquito-human cycle observed throughout most tropical regions of the world, often associated with urban locations and always human habitations, often producing explosive outbreaks, whereas “sylvatic” dengue, genetically different, circulates in forest wild animals and has been reported to be able to infect humans. In the Lao PDR, a developing country where dengue is endemic, data on this disease are sparse. This study reports an unusual outbreak of dengue that occurred during the cold season in a village in a forested area. It also is the first extensive analysis of dengue virus nucleotide sequences, from 39 patients across the country, from Laos. Results suggest three patterns of dengue circulation in Laos: local transmission, transmission over the whole country, and transmission implicating bordering countries. The dengue virus isolated from patients in the forest village outbreak proved to be genetically similar to those found in urbanized areas throughout the country. More investigations are needed to understand the relationships between dengue in forested and urban areas.
| Dengue is endemic in more than 100 countries in Asia, Africa and the Americas, but 70% of those currently at risk live in South-East Asia and the Western Pacific. WHO estimates that 50–100 million people are infected by dengue globally every year [1]. Dengue infections may be asymptomatic or symptomatic, classified as dengue fever (DF), dengue haemorrhagic fever (DHF), and dengue shock syndrome (DSS) [2] or more recently as dengue, dengue with warning signs and severe dengue [3]. Dengue fever is characterized by a sudden onset of high-grade fever with non-specific symptoms and most cases resolve without specific treatment. However, DHF, caused by increased vascular permeability, may progress to hypovolaemic shock and to potentially lethal DSS [2]. Dengue viruses (DENV) are single stranded RNA viruses from the family Flaviviridae, transmitted by Aedes spp. mosquitoes, which are predominantly urban. Sylvatic dengue has also been described in humans in SE Asian and West African forests [4], [5], [6], [7], [8], [9], [10], [11] but has only been associated with one outbreak [5]. Although there is evidence of interactions between urban and sylvatic dengue, their importance for dengue epidemiology and public health is not well understood [11].
In the Lao PDR (Laos), dengue is endemic with re-occurring epidemics during the monsoon season [12], [13], [14], [15]. Diagnosis of dengue in Laos is usually based on clinical symptoms [15], biological confirmation being only occasional. This is a major public health issue as many other pathogens have clinical manifestation similar to those of dengue. There is little information on the eco-epidemiology of dengue in the country or whether rural populations are affected.
Here, we report an epidemic of dengue in a rural village of the Xayabury Province, north-west Laos in 2008, with the first dengue molecular epidemiology data from Laos, including dengue strains from across Laos from 2007 to 2010, and discuss their public health significance.
In November and December 2008, an outbreak of unexplained fever occurred in Latsavang Village (18.222°N, 101.322°E, altitude ∼312 m above msl), Paklai District, Xayabury Province, NW Laos (Figure 1). Latsavang village is on the bank of the Namyang River, 60 km, by forest track, from the nearest town of Pak Lai, 10 km to the E and 14 km to the W of the Lao/Thai border (Muang Chet Ton in Thailand). The 1,526 inhabitants, living in 298 households, are predominantly maize farmers.
For the outbreak investigation, performed by the Government of the Lao PDR and WHO, oral informed consent was obtained from all patients (or their guardians/parents if they were aged <15 years). As this was an urgent public health outbreak investigation for diagnosis of the epidemic, ethical approval was not requested and oral consent was not formally documented. Retro-active ethical approval was not requested for the outbreak investigation.
All the patients seen at Mahosot Hospital, Luang Namtha Provincial Hospital and Salavan Provincial Hospital, with unexplained fever, gave informed written consent. For children, their guardians/parents gave informed written consent. These studies were granted ethical approval by the Lao National Ethics Committee for Health Research (124/NECHR, 134/NECHR) and the Oxford Tropical Research Ethics Committee (025-02, 006-07, 015-10)
DNA was extracted from 200 µl of acute serum using the QIAamp DNA minikit (Qiagen), following manufacturer's instruction, and eluted with 100 µl of elution buffer. Viral RNA was extracted using EZ1 Virus v2.0 Kit and EZ1 Biorobot (Qiagen) from 200 µl of acute serum following manufacturer's instruction and eluted with 90 µl of elution buffer. Internal phage controls were added to all sera before extraction to monitor the amplification processing and check for the presence of PCR inhibitors [16].
Diagnosis of dengue relied on the detection of either (i) NS1 dengue antigen (Dengue Early ELISA (Cat no. E-DEN01P), PanBio Ltd., Sinnamon Park, Queensland, Australia), (ii) specific anti-dengue IgM antibodies (Japanese encephalitis-dengue IgM Combo ELISA (Cat no. E-JED01C), PanBio), (iii) high level specific anti-dengue IgG antibodies associated with acute secondary dengue infection (Dengue IgG capture ELISA (Cat no. E-DEN02G), PanBio), (iv) and/or dengue PCR. ELISA tests were performed following manufacturer's instruction. During secondary dengue infection, anti-dengue IgM antibodies may be produced at low or undetectable levels, whereas anti-dengue IgG reach levels above those found in primary or past infection [2]. The Dengue IgG capture ELISA kit (Panbio) permits specific detection of these high levels of anti-dengue IgG antibodies present in acute secondary infection, increasing sensitivity for detecting dengue, when IgM and/or NS1 are not detected.
Random Reverse Transcription (RT), using hexamer primer, was performed using 10 µl of RNA extracts and the TaqMan Reverse Transcription Reagents (Applied Biosystems) following the manufacturer's instruction in a final volume of 50 µl.
The ‘Dengue All’ TaqMan real-time PCR (that detects all 4 dengue serotypes) was performed on 10 µl of RT product, following Leparc-Goffart et al. [17] and using the Platinum Quantitative PCR SuperMix-UDG kit (Invitrogen). Samples positive with this PCR were further characterised by using serotype-specific real time PCR systems [17].
The Panflavivirus SYBR Green real-time PCR [18], that detects all viruses belonging to the genus Flavivirus (family Flaviviridae), was performed using 3 µl of RT product and the QuantiMix Easy SYG kit (Biotools, Madrid, Spain). Amplicons (272 bp in the NS5 gene) were sequenced (Macrogen Inc., Seoul, Korea) and the corresponding sequences were BLASTed on the NCBI website (blastn) for identification. All negative primary Panflavivirus PCR reactions underwent a hemi-nested PCR using 3 µl of the primary PCR product, the same reverse primer, a distinct degenerate forward primer [19] and the same amplification protocol as in the primary PCR. Amplicons (200 bp) were sequenced and BLASTed for identification.
In a Biosafety Level 3 laboratory, 200 µl of patients' sera were inoculated onto confluent Vero cells in a 12 well plate format. After 1 week at 37°C in a 5% CO2 incubator, cells were scraped off and centrifuged. The clarified supernatant (SN) was collected and 1 ml was passaged onto fresh Vero cells using 12.5 cm2 flasks. To check for dengue virus growth for all patient sera, viral RNA was extracted from 0.2 ml of SN at passages 0 (inoculation on plates) and 1 (12.5 cm2 flasks) using the QIAamp MinElute spin kit (Qiagen) following manufacturer's instructions with an elution volume of 90 µl, and tested using the dengue 1-specific real time PCR system and Panflavivirus real-time PCR, as described above [17], [18].
Eighteen PCR amplification systems (Table S1), distributed along the complete dengue 1 (DENV-1) genome (from bases 17 to 10,737), were designed from the alignment of complete genome sequences of 36 DENV-1 strains isolated worldwide (GenBank accession numbers in supporting material Text S1). They were used to produce overlapping PCRs along the viral genome from random hexamer RT reactions. The PCRs were performed using 3 µl of RT with 1.25 unit of AmpliTaq Gold DNA polymerase (Applied Biosystems), 2 mM of MgCl2, 200 µM of dNTPs and 400 nM of each primer. The 3′ end fragment of the genome was amplified using the QuantiTect SYBR Green RT-PCR Kit (Qiagen) directly from 5 µl of RNA extract. The corresponding PCR products were sequenced by Macrogen Inc.
Nearly full-length DENV-1 genome sequences were produced directly from the sera of 7 patients: two patients from Latsavang, 2 patients from Northern Laos (Luang Namtha, October 2008 and September 2009), 2 patients from Southern Laos (Salavan, September and October 2008) and one patient from Vientiane (December 2008).
For sera from 32 additional DENV-1 patients from Luang Namtha, Salavan and Vientiane, between 2007 and 2010, 1,723 bases covering the dengue 1 envelope gene were amplified and sequenced using 3 pairs of primers (Table S1).
All 2,160 DENV-1 envelope sequences available in the Hemorrhagic Fever Viruses (HFV) database (http://hfv.lanl.gov) were downloaded and aligned, using ClustalX2.1 [20], with the 39 Lao DENV-1 envelope sequences obtained in this study. A Neighbour-Joining tree was constructed using Mega 5.05 software with Kimura-2 model [21]. Reliability of nodes was assessed by bootstrap resampling with 500 replicates. Another tree using the Maximun Likelihood method was constructed using the same alignment.
In addition to this ‘exhaustive’ reconstruction, another Neighbour-Joining tree was constructed from the alignment of the 39 Lao envelope sequences with the 1,318 complete or nearly complete DENV-1 genomic sequences (10,000 bases or more) available in HFVdb, to produce a tree with a robust phylogenetic backbone.
The 1,318 complete (or nearly complete) DENV-1 genomic sequences were aligned, using ClustalX2.1, to the 7 Lao DENV-1 genome sequences obtained in this study. Indication for the presence of molecular recombinations was investigated using the Recombination Detection Program (RDP) version 3 software [22]. RDP, GENCONV and MAXCHI methods were used for primary screening and the BOOTSCAN and SISCAN methods for automatic checking of the recombination signals [23], [24], [25], [26], [27]. For optimal recombination detection, we selected the RDP3 automask procedure that allows automatic masking of all sequences except one within groups of similar sequences. This means that within each group of similar sequences, only one sequence was examined during the exploratory search for recombination signals. When the program finds a recombination signal in an unmasked sequence it then examines all masked sequences from the same group (RDP user manual, http://darwin.uvigo.es/rdp/rdp.html).
Recombination events of more than 400 bases with a multiple comparison corrected p-value lower than e-10 were selected for downstream phylogenetic analyses. Neighbour-Joining trees using Mega 5.05 software with Kimura-2 model were produced from the alignment of the sequences remaining after the automask process. For each recombination event, two trees were produced and compared: one using sequences located between the putative recombination breakpoint positions and another excluding the putative recombinant region.
Between 19th November and 18th December 2008, seventy people met our case definition: any person residing in Latsavang Village and experiencing fever and headache with one or more of nausea and vomiting, myalgia and weakness in the previous three weeks, and any children less than on year old presenting with failure to thrive or neurological symptoms in the past three weeks. The median age was 18 years (range, 2 months-64 years); 28 patients (40%) were aged <15 years and 43 (61%) were female. The median (range) duration of illness was 2 (1–7) days. Of the 70 patients, 65 (93%) had fever, 65 (93%) headache, 22 (31%) nausea, 24 (34%) dizziness, 16 (23%) vomiting, 7 (10%) facial erythema, 2 (3%) myalgia, 3 (4%) cyanosis, 3 (4%) convulsions, 2 (3%) conjunctivitis and 1 (1%) had a rash. Of the 42 adult patients, 35 (83%) were farmers.
Six (13.6%) patients died during the outbreak. All were <20 years old and three were infants (1.6–3 months). These infants had no documented fever and presented with central and peripheral cyanosis with convulsions.
The 11 patients who had blood samples taken presented with 1–7 days of fever, headache, nausea and vomiting (Table 1). All of them survived.
Amongst the 11 patients with samples, 8 had laboratory evidence of acute dengue infection (Table 2): four tested positive for NS1 antigen and/or by real time PCR and four had serological evidence of recent infection (IgM or high level IgG antibodies). Serotype specific PCR and sequencing confirmed the presence of DENV-1 for all four NS1 and/or PCR positive patients. In addition, for six patients (XB998, XB1000, XB1009, XB1012, XB1013, XB1014) for whom remaining sera were available and placed in cell culture, dengue virus was isolated from the serum of one patient (XB998, NS1+/IgM+/PCR+) and was associated with obvious CPE in Vero cells at day 3 post inoculation and positive PCR detection of DENV-1. No CPE was observed after 7 days for the remaining 5 patients and panflavivirus PCR on cell culture extracts were negative.
The 18 overlapping DENV-1 PCRs yielded an almost full sequence of the genome of the DENV-1 virus infecting two patients from the outbreak, XB998 (10,629 bases) and XB1011 (9,823 bases), in addition to two patients from Luang Namtha, Northern Lao PDR, (9,823 bases each), two patients from Salavan, Southern Lao PDR, (9,823 and 10,648 bases respectively) and one patient from Vientiane (10,641 bases) (GenBank accession numbers in Table S2). All sequences were obtained from direct processing of patients' sera and shared a high level of identity at the nucleotide level (>97%, Table S3).
The tree constructed with 2,199 DENV-1 envelope sequences, including 2 sequences from the Latsavang outbreak, 5 from Luang Namtha, 13 from Salavan and 19 from Vientiane between 2007 and 2010 (GenBank accession number in Table S2), shows (Figure 2) that all Lao strains belong to genotype I.
Two subtrees, including only genotype I strains, were extracted from the tree constructed with 2,199 DENV-1 envelope sequences and from the tree constructed with the 39 Lao envelope sequences aligned with 1,318 complete sequences. The two subtrees are presented in Supporting Material Figure S1 and Figure 3, respectively. Seven clusters supported by high bootstrap value (>90) can be distinguished in Figure 3. The sequences from the Latsavang outbreak group with a strain from Vientiane of the same period (cluster 4), whereas most of the other Lao strains belong to six other defined clusters. Cluster 1: 7 strains from Salavan 2008, 2009 and 2010. Cluster 2: one strain from Luang Namtha 2008 and one strain from Salavan 2009. Cluster 3: 8 strains from Vientiane 2007 and 2008 and one strain from Vietnam 2007. Cluster 5: one strain from Vientiane 2008 and strains from Cambodia and Vietnam between 2003 and 2007. Cluster 6: 2 strains from Salavan 2010 and strains from Vietnam and Cambodia between 2005 and 2009. Cluster 7: 7 strains from Vientiane 2007 and 2010, 3 strains from Luang Namtha 2010, one strain from Salavan 2010, one strain from China 2010 and one strain from South Korea 2007. Two strains from Vientiane 2007, 2 strains from Salavan 2008 and one strain from Luang Namtha 2009 do not belong to well supported clusters.
Similar topologies were obtained with the Maximum Likelihood method (complete tree and genotype 1 subtree are presented in Supporting Material Figures S2 and S3).
An alignment of 1,325 DENV-1 sequences of 9,756 nucleotides (starting at ORF position 21) was submitted to RDP3 software. The recombination event signals of more than 400 bases with multiple comparison (MC) corrected p-values lower than e-10, obtained by this bioinformatic analysis are displayed in Table S4. Lao dengue strains characterised in this study were not found to be implicated in recombination events. Some putative recombinant strains from China (numbers 1, 4–9, 11,12, and 14 in Table S4) were identified, as described by Wu et al. [33]. The other recombination events (numbers 2, 3, 10, and 13) were further investigated by phylogenetic analysis presented in Supporting Material Figure S4.
Anti-JEV IgM ELISA, HRP-2 P. falciparum RDT and PCR assays for O. tsutsugamushi, and R. typhi were negative for all samples tested from Latsavang. Sera from 8 patients were tested by IFA for anti-O. tsutsugamushi antibodies in acute and, for 6 patients in convalescence samples. All samples had IgM and IgG titres >1∶400 and titres were very high, between 1∶1600 and 1∶3200 (Table S5). No sera were left for IFAs for three patients. There were no significant differences between the acute and convalescent anti-O. tsutsugamushi antibody titres. IFA results from 5/6 patients with acute and convalescent sera demonstrated anti-R. typhi IgM and IgG titres >1∶400. Two further patients with only acute samples had IgM and IgG titres >1∶400. Titres were generally high, but lower than those against O. tsutsugamushi. One patient had acute and convalescent anti-R. typhi IgM and IgG titres <1∶400. No sera were left for IFAs for three patients. There were no significant differences between the acute and convalescent anti-R. typhi antibody titres.
High IgM and IgG titres against both O. tsutsugamushi and R. typhi can be detected in healthy rural Lao farmers, presumably reflecting repeated infections in endemic areas (LOMWRU, unpublished data). Therefore, detection of such high titres may be a consequence of repeated or recent rickettsial infections rather than suggesting that O. tsutsugamushi and/or R. typhi were responsible for the outbreak. This is supported by the lack of change in serological titres (albeit the interval between early and convalescent sera was only 5 days) and the negative PCR for these bacteria (albeit that sera and not buffy coat, were used). Therefore, although rickettsial diseases cannot be excluded as contributing to the outbreak, this aetiology seems unlikely.
In contrast, the evidence that dengue was the predominant pathogen causing the outbreak was supported by the evidence of acute or recent infection for eight of the eleven patients tested. The median (range) patients' age was 15 (5–52) years (n = 11). Of note, all those PCR dengue positive were ≤15 years of age, and all were infected by dengue serotype 1. The clinical presentation was consistent with dengue, but also with many other infections including typhus. The disease was of short duration (1–7 days), consistent with dengue, but no shock or haemorrhage, characteristic of severe dengue, was reported. Nausea and vomiting were frequent. Although not conventionally thought to be major clinical features of dengue, amongst 170 Lao adults with serologically confirmed dengue 44% had nausea and 28% had vomiting [15]. The clinical features of those with and without samples were very similar.
That three infants who died did not have documented fever suggests the possibility they had thiamin deficiency (beriberi), which is not uncommon amongst breastfed Lao infants and is associated with traditional post-partum maternal food avoidance [34]. It is likely that infections may exacerbate thiamin deficiency and it is possible that dengue infection could precipitate fatal thiamin-deficiency heart failure [35]. However no biological confirmation could be provided to support this hypothesis.
Dengue was not suspected to be the responsible pathogen during the outbreak, as in Laos dengue is regarded as an urban disease and incidence is low during the cooler months of November and December. Latsavang is a river bank village of 1,526 inhabitants living in 298 households among maize and paddy fields and surrounded by tropical forest, 120 km as the crow flies from the nearest large town (Xayabury), that has ∼68,000 inhabitants. During the outbreak there were no reports of high dengue incidence in adjacent parts of Laos or Thailand. Daily rainfall data collected in Xayabury town does not show significantly higher rainfall or temperature in November or December 2008 in comparison to the same months in 2009. The mean minimum/maximum monthly temperatures were 17.8/27.9°C and 14.3/26.7°C in November and December, respectively. However, there were many uncovered still water containers and pools in and around the village.
This investigation highlights the difficulty in investigating epidemics of unknown etiology in remote areas. Accessing remote villages, delay in launching investigations, incomplete collection of medical and epidemiological data and of relevant biological samples, shipment conditions, limited size of the population sample studied and absence of negative controls hinder efforts to elucidate aetiology. The results demonstrate the importance of field epidemiology human and financial capacity and standardised sample collection and data recording protocols in developing countries.
This study provided the opportunity to perform the first analyses of the molecular epidemiology of dengue in Laos. Despite endemic dengue infection in Laos, only four partial genomic sequences of dengue virus were available in public databases prior to the current study (including one DENV-1 from 1996). Here, (nearly) complete genomic sequences were established for two patients from the 2008 Latsavang outbreak, but also for two patients from Luang Namtha (2008–2009), two patients from Salavan (2008), and one patient from Vientiane (2008). In addition, complete envelope sequences were produced for a number of Lao DENV-1 positive sera sampled between 2007 and 2010 (Luang Namtha: 3; Salavan: 11; Vientiane: 18).
The tree constructed with 2,199 DENV-1 envelope sequences, shows (Figure 2) the five DENV-1 genotypes previously described [36]. Genotype I includes strains from Southeast Asia, China and East Africa. Genotype II includes Thailand strains from the 1950s and 1960s. Genotype III contains strains from Malaysia, isolated in 1972 from a sentinel monkey and from a dengue patient in 2005 [37]. Genotype III represents spillback from the human into the sylvatic transmission cycle [5]. Genotype IV includes strains from West Pacific Island and Australia. Genotype V includes strains from America, West Africa and limited Asian strains. Within the genotype I, 4 clades have been described [33]: (i) a majority of all genotype I strains belongs to the ‘Asia 3’ clade. (ii) the ‘Asia 1’ clade includes strains isolated in the 1940's in Japan, Hawai and the USA. (iii) the ‘Asia 2’ clade includes mostly Thai (1980–1994), Chinese (2010), Saudi Arabian (2004–2006), Burma (Myanmar) strains (1998–2002), and the only Lao strain (1996) available in GenBank before those sequenced for this study. (iv) the ‘Asia 4’ clade includes Chinese strains (1997–1999).
Molecular characterisation of Latsavang DENV-1 outbreak strain shows that it belongs to the ‘Asia 3’ clade of Genotype I, with a genomic sequence very similar to other strains that have caused human infections in the region. Therefore, although the outbreak was in a forested area, it was not caused by a genotypically ‘sylvatic’ dengue virus in the conventional use of the term. Sylvatic dengue strains have been considered to be ecologically and genetically distinct from endemic strains [38], [4], circulating in the forest between non-human primates and arboreal Aedes mosquitoes, in contrast to strains responsible for human infections by domestic Aedes aegypti and peridomestic Aedes albopictus mosquitoes [39], [40]. However, many cases of sylvatic dengue infecting human have been reported [4], [5], [6], [7], [8], [9], [10], [11] and genetically ‘non-sylvatic’, dengue viruses have been detected in wild vertebrates in tropical forests in French Guiana [41]. Little is known about the interplay between anthropogenic and sylvatic transmission cycles and how this may contribute to dengue evolution and maintenance. Using the term ‘sylvatic’, an ecological term, for specific evolutionary groups of strains could therefore be confusing since dengue strains that do not belong to these genotypes may circulate within a sylvatic cycle. Further discussion of the terminology is required with separation of genetic and ecological terms.
Isolating a dengue strain from patients in a rural forest environment during a non-epidemic period raises questions. We identified a closely related strain in Vientiane in late December 2008 (Figure 3) and both strains appear to be descended from viruses described circulating in Thailand, and constitute a lineage that has not been identified anywhere else. The virus may have been brought to Latsavang by someone from Vientiane or elsewhere, provoking an outbreak in a population with limited prior exposure against DENV-1, as suggested by the significant proportion of adults affected (there are no data on incidence of dengue in the Latsavang area before the outbreak). The village is connected to Xayabury via a forest motorable track and the Mekong River is also a transport artery connecting urban centres in north-western Laos. Alternatively, the virus could have been maintained in the Latsavang area through a forest cycle and then transmitted to humans. A major limitation of this investigation was the lack of entomological investigations to determine which mosquito species (Aedes albopictus, Aedes aegypti, or arboreal Aedes species) were dengue vectors in Latsavang. Other limitations include the fact that only 10% of patients were sampled, that there was not a long interval between acute and convalescent samples and that buffy coat blood, for the PCR detection O. tsutsugamushi and R. typhi, was not available.
Phylogenetic analyses provided the very first clues regarding the molecular epidemiology, dynamics and dispersal of dengue in Laos. First, there is strong evidence that a given strain can be maintained and circulate during consecutive years in specific geographically limited evolutionary clusters. The well-defined phylogenetic cluster 1 (Figure 3), for example, only includes strains from Salavan that were however identified during three consecutive years (2008, 2009, 2010). Similarly, cluster 3 only includes strains from Vientiane in 2007 and 2008. Second, in contrast, it was also observed that a cluster of closely related sequences (cluster 7, Figure 3) was identified in Vientiane, Salavan and Luang Namtha during the same period in 2010. This demonstrates wide geographical viral dispersal, presumably associated with infected patients travelling within Laos. A similar observation is suggested by cluster 4 which includes 2008 Latsavang and Vientiane viruses (see above), and cluster 2 which includes 2008–2009 Salavan and Luang Namtha viruses. This unexplained contrasting epidemiology of clusters deserves attention, suggesting that, despite the rural nature of large tracts of Laos, human activity allows the efficient dispersal of Aedes-borne pathogens such as dengue. This illustrated what Stoddard et al. recently reported [42], [43] concerning human movement being a key component in determining dengue virus spread. As they suggested, studying individual spatiotemporal movements at fine and broad scales could permit better understanding of the different patterns of dengue transmission reported here and to propose options for dengue control adapted to Laos.
Third, beyond the above-mentioned patterns of local maintenance and dispersal within Lao territory, there is evidence of broad circulation of DENV-1 in the region, crossing national borders. In cluster 5 (Figure 3), a strain from Vientiane 2008 groups with strains from Cambodia and Vietnam isolated between 2003 and 2007. In cluster 6, two strains from Salavan 2010 group with strains from Cambodia and Vietnam isolated between 2005 and 2009. Fourth, recombination analysis, using only bioinformatic tools, did not identify recombinant events amongst DENV-1 Lao strains.
Altogether, these results provide a preliminary, complex picture of dengue epidemiological dynamics in a country in rapid socioeconomic transition. Cornerstones of future investigations may aim for (i) a more complete understanding of the entomological and ecological aspects of dengue transmission, dispersal and maintenance; (ii) a nationwide collection of population-based sero-epidemiological data, of confirmed infection cases and of viral sequences and (iii) investigation of dengue epidemiology of dengue in rural Laos. It is expected that such information associated with the rapid improvement of diagnostic procedures would allow for a better public health management of dengue in Laos.
|
10.1371/journal.pcbi.1000278 | Prediction of Protein–Protein Interaction Sites in Sequences and 3D Structures by Random Forests | Identifying interaction sites in proteins provides important clues to the function of a protein and is becoming increasingly relevant in topics such as systems biology and drug discovery. Although there are numerous papers on the prediction of interaction sites using information derived from structure, there are only a few case reports on the prediction of interaction residues based solely on protein sequence. Here, a sliding window approach is combined with the Random Forests method to predict protein interaction sites using (i) a combination of sequence- and structure-derived parameters and (ii) sequence information alone. For sequence-based prediction we achieved a precision of 84% with a 26% recall and an F-measure of 40%. When combined with structural information, the prediction performance increases to a precision of 76% and a recall of 38% with an F-measure of 51%. We also present an attempt to rationalize the sliding window size and demonstrate that a nine-residue window is the most suitable for predictor construction. Finally, we demonstrate the applicability of our prediction methods by modeling the Ras–Raf complex using predicted interaction sites as target binding interfaces. Our results suggest that it is possible to predict protein interaction sites with quite a high accuracy using only sequence information.
| In their active state, proteins—the workhorses of a living cell—need to have a defined 3D structure. The majority of functions in the living cell are performed through protein interactions that occur through specific, often unknown, residues on their surfaces. We can study protein interactions either qualitatively (interaction: yes/no) using large-scale, high-throughput experiments or determine specific interaction sites by using biophysical techniques, such as, for example, X-ray crystallography, that are much more laborious and yet unable to provide us with a complete interaction map within the cell. This paper presents the machine learning classification method termed “Random Forests” in its application to predicting interaction sites. We use interaction data from available experimental evidence to train the classifier and predict the interacting residues on proteins with unknown 3D structures. Using this approach, we are able to predict many more interactions in greater detail (i.e., to accurately predict most of the binding site) and with that to infer knowledge about the functions of unknown proteins.
| Most proteins in a living cell interact in order to fulfil their function. Protein interactions occur through the formation of complexes, either transient or more long lasting, as a result of a balance between different molecular properties: sequence, shape, charge distribution, entropy and dynamics. Proteins often interact through multiple components, with examples like the replisome, RNA polymerases, the spliceosome, the ribosome, chaperonins and the various complexes formed along signal transduction pathways and during enzyme catalysis and inhibition. Knowledge of protein interactions is sometimes crucial in elucidating their functional roles. 3D structures of protein complexes have been the basis for detailed understanding of protein function; however, it is much more technically demanding to determine the structure of a complex as opposed to solving a structure of a single protein or even a fragment of the whole protein—a protein domain. This is the reason why in the current release of the Protein Data Bank (http://www.pdb.org) [1], 3D structures of protein complexes are poorly represented. In addition, the number of protein sequences deposited in the UniprotKB/Swiss-Prot database (http://www.uniprot.org) [2] outstrips the number of known 3D structures by around 7 times—a fact that further demonstrates the restricted effective size of the structural sample set available for studying protein interactions. On the other hand, experimental methods for detection of protein interaction residues from proteins without a known 3D structure are based on mutation and deletion studies. These methods are expensive, laborious and, most importantly, poorly applicable on a large scale.
The abundance of information that can be extracted from a 3D structure and sequence, the increase in computer power and the invention of novel classification methods have triggered development of computer based methods for prediction of protein interfaces. Since the pioneering work of Jones and Thornton [3] and their attempt to predict surface patches that overlap with interaction interfaces, several papers presenting different methods have been published. Methods presented therein can be roughly divided into three groups based on the choice of features used for prediction. The first group consists of methods based solely on sequence information that predict protein interfaces [4]–[8]. Methods in the second group [9]–[11] use structural information to refine sequence sets that are then used to construct predictors. Methods of the third group use 3D structure information exclusively or a combination of 3D structure and sequence for prediction [3], [12]–[17]. Selection of the classification method used also varies across different prediction tools: scoring functions [15], SVM (support vector machines) with radial kernel [6],[7][9][10][14] and neural networks [5],[8],[11].
In this paper we present two methods for prediction of interaction sites of protein heterocomplexes using only a) sequence information; and b) information obtained from a combination of sequence and 3D structure features. Both new methods are based on the random forest algorithm [18] and linear classifier combinations. Our classification features are derived from sequence and spatial information, from sliding windows of nine residues in width. For the first time we rationalized this most commonly used window size through entropy analysis and demonstrated that it contains the highest amount structural information per sequence length.
Proteins commonly have many more residues that do not participate in an interaction than interacting residues, which creates an effect of imbalance between positive and negative datasets and must be dealt with in the process of classification. One drawback of imbalanced datasets is that some of the classification methods (especially the SVM) work with impaired performance and may introduce a bias in the resulting classification. Another consequence of working with an imbalanced dataset is that some commonly used evaluation measures, such as accuracy, are not appropriate because they favour the majority class [19]. Instead, we used the precision-recall graph, F-measure [20] and AUC (area under the ROC curve) [21], commonly used in the Information Retrieval sciences. It was demonstrated [22] that the classification method based on Random Forests achieves good results with unbalanced data. In addition we employed a classifier combination approach, which further improved predictions made from unbalanced data.
Performance comparison between different methods is rather difficult owing to the (i) lack of a good interaction benchmark set; (ii) different definitions of interaction sites; and (iii) different evaluation measures. Nevertheless, performance of our method in terms of structure-based prediction produces results comparable with best results obtained by other authors and we believe that our method outperforms others in the prediction of protein interacting residues based on sequence information alone.
For testing purposes, we built a Ras–Raf complex whose 3D structure has not been experimentally determined.
The first step in our investigation was to determine the optimal sliding window length. We used a method (See Methods) based on the entropy difference between the occurrence of a particular number of interacting residues within a window length of N residues and the uniform occurrence distribution. We investigated only windows with a central interacting residue present. The result of the analysis is presented in Figure 1. Although the results for different window lengths are similar, it is evident that for the window length of 9 the entropy has the maximum difference.
The most challenging part of our work was to construct a predictor of interacting residues using only sequence information. The input feature vector consisted of names of nine consecutive residues in a sequence. The class label of an instance was defined positive if at least N residues, including the central one, were marked as ‘interacting’. We classified data for values of N from 1 to 9. The evaluation of results is presented in Table 1. All of the presented values of measures, except the AUC, were calculated using a majority vote rule. The threshold for distinction between positive and negative output classes was 0.5. Table 2 shows the confusion matrix for a threshold of 1 interacting residue in a window. For error estimation 10-fold cross validation was used. It can be seen that the precision for almost all N's was over 80% with recall around 25%. When we further combined classifiers (See methods) the recall, F-measure and AUC increased, while the precision decreased. Using combining classifiers at a precision of 84%, we achieved a recall of about 26%. The F-measure, a harmonic mean of precision and recall obtained by combining classifiers was 40%, with the AUC at 74.7%. It is important to notice how accuracy increased as the ratio between positive labelled and negative labelled instances decreased. At the same time the precision was decreasing. If we further decreased the ratio between positive and negative labelled classes, accuracy would converge to the accuracy of the majority class, while precision would decrease to zero. Apparently, accuracy itself is not a good measure for evaluating method performance on an unbalanced dataset.
Figure 2 shows the precision-recall graph for combined classifiers. The results obtained by randomization testing (see Methods) are also presented. It can be seen that our method significantly outperforms random results.
In order to improve our results we introduced class weights. The Random Forests method uses different class weights for positive and negative classes in an effort to improve results of imbalanced data classification [22]. The results achieved using different weights are presented in Table 3. As can be noticed, the introduction of weights resulted in an increase in recall and F-measure, but with a decrease in precision. If we compare these values to those on the precision-recall graph it can be seen that the weighted classifiers are on or slightly above the curve. Random Forests is a discrete classifier so its output is represented with one point on the precision-recall curve. However, we can move along the curve to the desired values of precision or recall using different class weights.
Figures 3 and 4 show histograms of recall values for protein complexes and chains obtained for overall precision at 48% and recall at 53%. For these values our method correctly predicted at least one interaction site in 99.7% of the proteins and 99% of the chains.
For precision at 76% and recall at 31% we correctly predicted at least one interaction site in over 90% of the proteins and in over 80% of the chains.
We analysed and used an exhaustive set of 3D structure based attributes (see Methods): accessible surface area (ASA) [23], depth index (DPX) [24], protrusion index (CX) [25], hydrophobicity as well as protein secondary structure. We used all 3D structure information available from PSAIA [26] with the addition of secondary structure. As the first step we performed training and prediction with all available sequence and 3D structure (a total of 26) features. The random forest algorithm has the capability to estimate the importance of a particular feature (an equivalent of the principal component analysis), so we employed it in the process of input parameter set reduction. Figure 5 shows the importance of particular features and their contribution to the overall prediction quality. It is evident that the information obtained from sequence has the highest importance. In addition, we also selected five best ranked structural features: non-polar ASA, maximum depth index, relative non-polar ASA, average DPX and minimum CX. With this reduced set of descriptors we obtained only slightly inferior results then by the entire dataset and therefore it was used in all subsequent analyses.
We defined the central residue in the sliding window as an interacting residue if at least N of the residues (including the central residue) in the window are in contact with another chain. We tested threshold values of N in the range of 1 to 9. The evaluation of results is presented in Table 4. For precision at 78%, we achieved a recall of about 35%. When combining classifiers at the precision of 76%, we achieved a recall of about 38%. It can be seen that results obtained using combining classifiers are better. From the precision-recall graph (Figure 6) it is evident that prediction using structural information in combination with sequence information is better, especially in the central region, the most important part of the curve. Results with different class weights are presented in Table 5.
Similarly to predictions of interaction sites using only sequence information we evaluated the results per protein complex and chain. For precision at 75% and a recall of 40% our method correctly predicted at least one interaction site in over 97% of the proteins and over 90% of the chains. In addition for precision at 61% and a recall of 59% we correctly predicted at least one interaction site in 100% of the proteins and in over 99% of the chains.
Reliability of our method was tested at the RBD (Ras Binding Domain) of C-Raf1 (PDB::1C1Y) and the wild type Ras (PDB::121P). Although the 3D structures of both C-Raf1 and Ras were solved, the structure of their complex has not been determined experimentally yet.
Using information from sequence and structure the following residues were predicted as potentially interacting: Ile21, Gln25, His27, Glu31, Asp33, Pro34, Thr35, Ile36, Glu37, Asp38, Ser39, Tyr40, Arg41, Lys42 and Ser65 (Ras protein) and Arg67, Val70, Val88, Glu104, Gly107, Lys108, Leu112 and Asp113 (C- Raf protein).
The complex was built by AutoDock, version 4.0, [27] using the Ras Protein as a receptor and by setting the centre of the grid to Ras Asp38, the central residue of the largest predicted interaction region. Docking simulations were carried out with an initial population of 200 individuals, and a maximum number of 2 500 000 energy evaluations. The model with the amino acids residues predicted as possible interacting sites labelled, is displayed in Figure 7. Majority of the predicted residues are part of the modeled complex interface and their importance for complex formation was found by the experiments [28]–[30] as well. Exceptions are Raf amino-acids Glu104, Lys108, Leu112 and Asp113 on the opposite side of its Ras binding interface. Although they have not been described in the literature as interacting residues, they might present interaction sites for some currently unknown interaction partner.
Stability of the complex was tested during 700 ps of molecular dynamics (MD) simulation performed by AMBER9 (http://amber.scripps.edu/) [31]. The proteins' conformations, as well as their mutual position, did not change significantly during the simulation: RMSD for the main chain atoms was 1.66 Å (RMSD for Raf was 1.31 Å, and for Ras 1.02 Å, final structure), and its plateau was achieved within 200 ps of unrestrained simulation, i.e., during the last 500 ps of simulations RMSD for the main chain was within 0.5 Å
The main aim of this paper was to improve the prediction of interaction residues solely in protein sequence. In order to facilitate comparison, we used the same dataset and definitions of interacting sites as Ofran and Rost [5]. Because they divided sets of protein complexes into three subsets, we did the same for comparison of results. Figure 8 shows results obtained using 3-fold cross validation. For precision between 60 and 70%, Ofran and Rost achieved a recall of about 10% [8], while for the same precision we obtained a recall level of about 30%. The results can also be compared by the precision-recall (P-R) graphs where the P-R curve obtained by our methods shows better results, for recall at less than 50%. For higher recall values curves are similar. However, it is important to emphasise that the P-R curve obtained with randomization testing by our method has a lower value than theirs (27% compared to 30%). Because of that for recall values at more than 50% the curve obtained by our method is more distant from a random curve. Res et al. [7], did not present a precision-recall curve so we could compare only single point results. For the level of recall at 57.5% Res et al. [7] achieved a precision of 27.3%. This result is inferior to ours, i.e., for a recall of 57.5% we obtained precision above 40%.
Since authors [3], [13], [15]–[17] in the field of predicting interacting residues using 3D structure information used different estimation measures, datasets and definitions of interacting sites, it is difficult to objectively compare results.
Docking of the Ras-Raf complex is an example of how our proposed methods can help with practical problems. Information on possible interacting residues can significantly help and speed up determination of reliable complex conformation. Similarly, prediction based on sequence information only, can help in the determination of possible deletion or mutation residues in experiments when 3D structure is unknown. Using different class weights a compromise can be made between expected prediction and recall of achieved results.
Finally, one of the results of this paper is the confirmation that a widow of nine concatenating residues contains the highest information content for prediction of interacting residues.
We believe that the results can be further improved in the ways explained below. For example, a bigger non redundant test set should be defined. The dataset we have used dates back to 2003. Since then, the number of experimentally determined 3D structures has increased from 20 000 to 50 000. In addition, methods proposed in this paper do not use information that other authors found valuable like evolutionary information [7],[8],[11],[16], electrostatics [13],[15] and desolvation [15]. Furthermore, it is evident that some 3D structure data like ASA improve prediction of interaction sites compared to sequence-only predictions. Hence, prediction of these 3D structure features from sequence through the usage of existing methods [32]–[34] or newly developed ones could further improve results.
Finally, the aggregation of interacting residues noticed by Ofran and Rost [5] might also be the beneficial approach. One way of using this information is described in the paper by Yan [9].
For training and testing, we used a dataset of transient hetero interactions derived by Ofran and Rost [5]. The dataset consists of 1134 chains in 333 complexes. A residue was defined to be involved in a protein–protein interaction if any of its atoms were within 6 Å of any atom in a neighbouring non homologous chain. In our work, we used the PSAIA application [26] for the extraction of interacting residues. The main reason why we used the same dataset and the same method for definition of interacting residues as the above mentioned authors was for the purpose of comparing our results since their results are currently the best achievement in the field of proteins' interaction prediction from sequence alone.
The input vector of features was defined on a sliding window of nine residues. The window was defined as positive, if the central residue and at least N−1 other residues were interacting residues. We used a value for N in a range of 1 to 9. For determination of true negatives we used a method similar to the one of Ofran and Rost. We made an alignment of all homologous chains (at least 90 percent of sequence similarity) in the 3D structure of a complex. If all aligned chains at a particular site had the same nine residues in the window and none of them had a central residue in contact with a neighbouring non homologous chain we defined this window as a true negative.
The input vector consists of nine residues' names, and min, max or average values for features that belong to residues in the window. In this paper we used the following features:
The average value of a particular feature k was calculated as:where i was the ordered number of a residue in the window of N = 9 residues. The secondary structure information was extracted by DSSP [36], while for extraction of all other features PSAIA was used [26].
The length of the sliding window can influence the classification of results. For determination we used a method based on entropy. First we defined the interacting residues for all proteins in the datasets. Secondly, we calculated the number of interacting residues using sliding windows of different lengths. Only the windows with a central interacting residue were taken in consideration. Finally, the entropies for different window lengths were calculated, and subtracted from entropies calculated for a uniform distribution of numbers of interacting residues in the window. As the best result we defined the one with a highest calculated entropy difference. The calculation can be shown as following:where N is the length of a window, pi is the frequency appearance of i interacting residues in a window of N residues, given a central interacting residue.
The uniform distribution of a particular set has the highest entropy, so data that has the highest difference from that value has more structure than others and it is easiest to describe.
The results reported in this paper concern the evaluation of residue classification based on the following quantities: the number of true positives (TP) (residues correctly classified as interacting), the number of true negatives (TN) (residues correctly classified as non-interacting), the number of false positives (FP) (non-interacting residues incorrectly classified as interacting), and the number of false negatives (FN) (interacting residues incorrectly classified as non-interacting). These values are usually presented in a confusion matrix. We use the following measures of performance:In addition we used a precision-recall graph and area under the ROC curve (AUC) [21] for comparison of the results of our method with a random classifier. Although we believe that accuracy is not an appropriate measure in the event of imbalanced data we used it as a directly comparable measure to results of other prediction methods. In Random Forests there is no need for cross-validation or a separate test set to get an unbiased estimate of the test set error. It is estimated internally with the out of bag error estimate [18]. When estimating the test error using both methods we achieved a very slight difference in results, but for comparison to other results we present results obtained by 10-fold cross validation. For cross validation we divided a set of 333 proteins into 10 subsets, so that we would not use data from the same proteins in test and training sets. Performances of classifiers were estimated by the ROCR R package [37].
Classification methods are sensitive to over-fitting so it was important to measure the significance of obtained AUC values and precision-recall graphs. Randomisation testing has been found to be very effective at assessing over-fitting [38],[39]. Here, the original training set was copied and class labels were replaced with random class labels. The ratio between positive and negative class labels was preserved. Then the Random Forests were trained with these data using the same methodology that was used with the original data.
Random Forests [18] is an ensemble method that combines several individual classification trees in the following way: from the original sample several bootstrap samples are drawn, and an unpruned classification tree is fitted to each bootstrap sample. The feature selection for each split in the classification tree is conducted from a small random subset of predictor variables (features). From the complete forest the status of the response variable is predicted as an average or majority vote of the predictions of all trees. Random Forests is often used when we have very large training datasets and a very large number of input features (hundreds or even thousands of input features). A random forest model is typically made up of tens or hundreds of decision trees. In this paper we used 200 trees. As part of the algorithm, Random Forests returns few measures of feature importance. The most reliable measure is based on the decrease of classification accuracy when values of a feature in a node of a tree are permuted randomly and this is the measure of feature importance that we used in this paper. PARF (parallel Random Forests) [40] implementation of the random forest method and the randomForest R package [41] were used for classification.
Random Forest method is a discrete classifier. When such a classifier is applied to a test set, it yields a single confusion matrix, which in turn corresponds to a single point on a ROC curve. However, it is possible to use, as output, the percentage of votes for a particular class. Using different threshold values for producing positive or negative response variables it is possible to produce ROC and precision-recall curves.
A simple method for combining classifiers was used in this paper. The method takes output of the first stage classifiers as input values for the second stage. The output of the second stage is a positive class if at least one of the input values was positive. This method achieves good results with imbalanced data (Šikić and Jeren, manuscript in preparation). In this paper we labelled a sliding window instance with a positive class if it contained equal or more interacting residues than a value defined by the threshold and if the central residue was an interacting residue. We made classifications with threshold values from 1 to 9. The outputs of all of these classifiers were combined as explained above.
To better describe this method let us take one example. We start the prediction process by selecting a sliding window of nine residues. First we make a prediction using the classifier for threshold 1. This classifier is trained to predict interaction sites if at least the central residue is in an interaction. Second we make a prediction using the classifier for threshold 2. This classifier is trained to predict interaction sites if the central residue and at least one other residue inside the window are in an interaction. Using same method we make a prediction for thresholds 3 to 9. It can be easily seen that instances labelled positive (windows) for classifiers that use thresholds 2 to 9 are subsets of positive instances for classifiers that use threshold one. If we assume that it is possible that some classifiers can in some cases more accurately predict subsets than the original set we can combine classifiers using the OR rule. Hence, the output class label would be positive if at least one classifier labelled that instance as positive.
|
10.1371/journal.pcbi.1007188 | Primacy coding facilitates effective odor discrimination when receptor sensitivities are tuned | The olfactory system faces the difficult task of identifying an enormous variety of odors independent of their intensity. Primacy coding, where the odor identity is encoded by the receptor types that respond earliest, might provide a compact and informative representation that can be interpreted efficiently by the brain. In this paper, we analyze the information transmitted by a simple model of primacy coding using numerical simulations and statistical descriptions. We show that the encoded information depends strongly on the number of receptor types included in the primacy representation, but only weakly on the size of the receptor repertoire. The representation is independent of the odor intensity and the transmitted information is useful to perform typical olfactory tasks with close to experimentally measured performance. Interestingly, we find situations in which a smaller receptor repertoire is advantageous for discriminating odors. The model also suggests that overly sensitive receptor types could dominate the entire response and make the whole array useless, which allows us to predict how receptor arrays need to adapt to stay useful during environmental changes. Taken together, we show that primacy coding is more useful than simple binary and normalized coding, essentially because the sparsity of the odor representations is independent of the odor statistics, in contrast to the alternatives. Primacy coding thus provides an efficient odor representation that is independent of the odor intensity and might thus help to identify odors in the olfactory cortex.
| Humans can identify odors independent of their intensity. Experimental data suggest that this is accomplished by representing the odor identity by the earliest responding receptor types. Using theoretical modeling, we here show that such a primacy code outperforms alternative encodings and allows discriminating odors with close to experimentally measured performance. This performance depends strongly on the number of receptors considered in the primacy code, but the receptor repertoire size is less important. The model also suggests a strong evolutionary pressure on the receptor sensitivities, which could explain observed receptor copy number adaptations. By predicting psycho-physical experiments, the model will thus contribute to our understanding of the olfactory system.
| The olfactory system identifies and discriminates odors for solving vital tasks like navigating the environment, identifying food, and engaging in social interactions. These tasks are complicated by the enormous variety of odors, which vary in composition and in the concentrations of their individual molecules. In particular, the olfactory system needs to separately recognize the odor identity (what is there?) and the odor intensity (how much is there?). For instance, the identity is required to decide whether to approach or avoid an odor source, whereas the intensity information is important for localizing it. It is not understood how these two odor properties are separated and how odors are discriminated reliably.
Odors are comprised of chemicals that bind to and excite olfactory receptors in the nose in mammals and on antenna in insects. Each receptor responds to a wide range of odors and each odor activates many receptor types. The resulting combinatorial code allows to distinguish odor identities [1–3], but also depends on the odor intensity, since receptors respond stronger to more concentrated molecules [4]. To separate these two properties, the neural signals are processed in the olfactory bulb (antennal lobe in insects) and then forwarded to the olfactory cortex, where odors are identified by comparing to memorized patterns. Indeed, experiments indicate that the olfactory cortex receives a concentration-invariant code [5, 6], which allows to identify odors irrespective of their intensity. Consequently, the olfactory bulb can be thought of as a signal processor that removes statistical redundancies in the input to provide a more useful representation to the olfactory cortex. However, so far it is not clear what processing the olfactory bulb performs and how this affects odor representations.
The olfactory bulb contains neural clusters called glomeruli, which each receive input from a specific receptor type [7–9]. Each glomerulus excites associated projection neurons, which project into the olfactory cortex. Additionally, the glomeruli are connected by local neurons [10, 11], which inhibit the projection neurons [12–18]. These local neurons could mediate a global normalization resulting in an intensity-invariant representation of the odor identity [19, 20]. However, we showed that simple normalized representations still depend strongly on the number of ligands in a mixture and might thus not be optimal for solving olfactory tasks [21]. An alternative to these normalized representations is rank coding, where the order in which the receptors are excited is used to encode the odor identity robustly and independently of the odor intensity [22]. Indeed, experiments suggests that odors are encoded robustly by the receptor types that respond within a given time window after sniff onset [23–25]. In particular, the odor identity could be robustly encoded by a fixed number of the receptors that respond first, which is known as primacy coding [24, 26]. So far, it is unclear how efficient and useful primacy coding is and how it compares to alternative schemes.
In this paper, we consider a simple model of primacy coding and investigate how well it represents complex odors. In particular, we identify how much information is transmitted to the cortex and how well this information can be used to perform typical olfactory tasks, like identifying the addition of a target odor to a background or discriminating odor mixtures. Our statistical approach allows linking parameters of the primacy code to results from typical psychophysical experiments. We show that primacy coding provides a robust and compact representation of the odor identity over a wide range of odors, independent of the odor intensity, and that it outperforms other simple coding schemes. However, this good performance of the olfactory system hinges on tuned receptor sensitivities, which suggests that there is a strong selective pressure to adjust the sensitivities on evolutionary and shorter timescales.
We describe odors by concentration vectors c = ( c 1 , c 2 , … , c N L ), which determine the concentrations ci ≥ 0 of all ligands that can be detected by the olfactory receptors. The number NL of possible ligands is at least NL = 2300 [27] although the actual number is likely much larger [28]. Typical odors contain only tens to hundreds of ligands, implying that most ci are zero.
In experiments, the olfactory system is typically characterized by presenting odors with particular statistics, e.g., by choosing mixtures from a given ligand library. Although such experiments allow to characterize the olfactory system in a part of odor space, we ultimately want to understand how the system performs in its natural environment. Unfortunately, the statistics of natural odors are difficult to measure [29], so we here consider a broad class of odor distributions to approximate natural odor statistics [30]. In particular, we consider a situation in which each ligand i has a probability pi to appear in an odor. For simplicity, we neglect correlations in their appearance, so the mean number s of ligands in an odor is s = ∑i pi. To model the broad distribution of ligand concentrations, we choose the concentration ci of ligand i from a log-normal distribution with mean μi and standard deviation σi if the ligand is present. Consequently, the mean concentration of a ligand in any odor reads 〈ci〉 = pi μi and the associated variance is var ( c i ) = ( p i - p i 2 ) μ i 2 + p i σ i 2. For simplicity, we consider ligands with equal statistics in this paper, so the distribution Penv(c) of odors is characterized by the three parameters pi = p, μi = μ, and σi = σ. Using these broad odor statistics and more specific ones will allow us to analyze the performance of olfactory models in natural environments and in typical psycho-physical experiments, respectively.
Odors are detected by an array of receptors in the nasal cavity in mammals and on the antenna in insects. The receptor array consists of NR different receptor types, which each are expressed many times. Typical numbers are NR ≈ 50 in flies [7], NR ≈ 300 in humans [31], and NR ≈ 1000 in mice [32]. The excitations of all receptors of the same type are accumulated in an associated glomerulus in the olfactory bulb in mammals and the antennal lobe in insects [33]. Since this convergence of the neural information mainly improves the signal-to-noise ratio, we here capture the excitation of the receptors at the level of glomeruli; see Fig 1A. The excitation en of glomerulus n can be approximated by a linear function of the odor c [4, 34],
e n = ∑ i = 1 N L S n i c i , (1)
where Sni denotes the effective sensitivity of glomerulus n to ligand i. Note that Sni is proportional to the copy number of receptor type n if the response from all individual receptors is summed [30].
The sensitivity matrix Sni could in principle be determined by measuring the response of each glomerulus to each possible ligand. However, because the numbers of receptor and ligand types are large, this is challenging and only parts of the sensitivity matrix have been measured, e.g., in humans [35] and flies [36]. Using these data, we showed that the measured matrix elements are well described by a log-normal distribution with a standard deviation λ ≈ 1 of the underlying normal distribution [30]. Motivated by these observations, we here consider random sensitivity matrices, where each element Sni is chosen independently from the same log-normal distribution, which is parameterized by its mean 〈 S n i 〉 = S ¯ and variance var ( S n i ) = S ¯ 2 ( e λ 2 - 1 ) with λ = 1 [30]. Since these receptor sensitivities are broadly distributed, they might not include specific receptors related to innate behavior [37], but they can collectively discriminate concentration differences of several orders of magnitude [30].
The odor representation on the level of glomeruli excitations en depends strongly on the odor intensity, which is quantified by the total concentration ctot = ∑i ci. This dependency complicates the extraction of the odor identity, which is determined by the ligands which are present and their relative concentrations. A concentration-invariant representation could be achieved by normalizing the excitations by their mean [16], which leads to an efficient neural representation on the level of projection neurons [21]. However, recent experimental data suggest an alternative encoding based on the timing of the glomeruli excitation [24]. The key idea of this primacy coding is that the set of receptor types that are excited first is independent of the total concentration ctot and thus provides a concentration-invariant representation.
In our simple model of primacy coding, odors are encoded by the identity of the NC glomeruli that respond first. For simplicity, we here neglect the order in which they respond, in contrast to rank coding [22], and we also consider the simple situation where ligands binding to receptors only affect the magnitude of the receptor output, but not the signaling dynamics. In this case, receptors that respond first are also the ones with the largest excitation, so that the primacy code is given by the identity of the NC glomeruli with the largest excitation, which is known as the primacy set [38].
The primacy set can be represented by a binary activity vector a = ( a 1 , a 2 , … , a N R ), where an = 1 implies that glomerulus n belongs to the primacy set and is active, while an = 0 denotes an inactive glomerulus not belonging to the primacy set. Since the active glomeruli have the highest excitation, they can be identified using an excitation threshold γ; see Fig 1B. Consequently, the activities are given by
a n = { 1 e n > γ ( e ) 0 e n ≤ γ ( e ) . (2)
Physiologically, the activities an could be encoded by projection neurons in insects and mitral and tufted cells in mammals. These neurons receive excitatory input from one glomerulus [39] and are inhibited by a local network of granule cells [20, 33]. These granule cells basically integrate the activity of all glomeruli [40] and could inhibit the glomeruli once a threshold is reached. Taken together, this would implement primacy coding since only the glomeruli that respond earliest would be activated. For simplicity, we consider the case where the number NC of active glomeruli is fixed and does not depend on the odor c. The associated constraint
N C = ∑ n = 1 N R a n (3)
determines the threshold γ. Note that the activity pattern a is sparse since only a fraction NC/NR of all glomeruli is activated. Moreover, a is independent of S ¯ and ctot, implying concentration-invariance. This is because multiplying the concentration vector c by a factor changes the excitations en and the threshold γ by the same factor, so that a given by Eq (2) is unaffected. In essence, only relative excitations are relevant for our model of primacy coding.
In the binary representation given by Eq (2), each receptor type can at most contribute 1 bit of information to the odor representation. This worst-case scenario corresponds to large processing noise, such that intermediate excitations cannot be identified in the downstream processing. In fact, the concentration range over which receptors are sensitive is typically small compared to the expected range of odorant concentrations [41]. Consequently, receptors will be activated either very little or very strongly for natural odors, suggesting a binary picture. Moreover, there is evidence that the identity of active neurons can robustly encode odor identity and is actually used in the olfactory system [5].
To see whether primacy coding encodes odor information efficiently [42], we quantify the amount of information I that can be learned about the odor c by observing the binary activity pattern a with given sparsity NC/NR. Since our model is deterministic, I is given by the entropy
I = - ∑ a P ( a ) log 2 P ( a ) , (4)
where the probability P(a) of observing an output a depends on the odor environment Penv(c) as well as the properties of the olfactory system, which in our model are quantified by NC, NR, and λ. Since further processing in the downstream regions of the brain can only reduce the amount of information, Eq (4) provides an upper bound for the information that the brain receives about odors when primacy coding as described by Eqs (1)–(3) is used.
In an optimal receptor array, each output a occurs with equal probability when encountering odors distributed according to Penv(c) [30]. In our model, only outputs with exactly NC active receptor types are permissible. The resulting representations would be optimal if each receptor type was activated with a probability 〈an 〉 = NC/NR and all types were uncorrelated, cov(an, am) = 0 for n ≠ m. The associated information
I max ( N C , N R ) = log 2 ( N R N C ) ≈ N C - 1 ln 2 + N C log 2 N R N C (5)
provides an upper bound for I given by Eq (4). Here, the approximation on the right hand side is obtained using Stirling’s formula for large receptor repertoires (NR ≫ NC).
We start by analyzing the information I transmitted by the primacy code using numerical ensemble averages of Eqs (1)–(4); see Methods and Models. Fig 2A shows that I is very close to the maximal information Imax given by Eq (5), which is obtained when all receptor types have equal activity and are uncorrelated [30]. This indicates that the primacy code uses the different receptor types with similar frequency and that correlations between them are negligible. The expression for Imax implies that the information grows linearly with the primacy dimension NC, but only logarithmically with the number NR of receptor types. Consequently, the number of distinguishable signals, NS = 2I, grows strongly with NC, but the dependence on the repertoire size is weaker; see Fig 2B. Given equal NC, our model thus predicts that the transmitted information in mice is only twice that of flies, although mice possess about 20 times as many receptor types. However, the number of discriminable signals changes by many orders of magnitudes, since it scales exponentially with I.
The logarithmic scaling of the transmitted information I with the receptor repertoire size NR could explain why the ability of rats to discriminate odors is not significantly affected when half the olfactory bulb is removed in lesion experiments [43, 44]. If this operation removes half the receptor types, our model implies that the transmitted information I is lowered by NC bits; see Eq (5). This corresponds to a reduction of I by about 10% in rats where NR ≈ 1000; see Fig 2C. Conversely, the transmitted information decreases by almost 50% in flies, which have a much smaller receptor repertoire of NR ≈ 50. Our model thus predicts that lesion experiments have a much more severe effect on the performance of animals with smaller receptor repertoires.
Taken together, this first analysis already suggests that the primacy code provides a robust odor representation, which is sparse, concentration-invariant, and depends only weakly on the details of the receptor array. However, for this representation to be useful to the animal, it needs to allow solving typical olfactory tasks.
Typical olfactory tasks include detecting a ligand in a distracting background, detecting the addition of a ligand to a mixture, as well as discriminating similar mixtures. All these tasks involve discriminating odors with common ligands, implying that the associated primacy sets are correlated. This correlation can be quantified by the expected Hamming distance d between the primacy sets, which counts the number of glomeruli with different activities. The probability η that this distance is larger than 0, so that the two odor representations can be discriminated in principle, is given by
η ( d ) ≈ 1 - ( 1 - d 2 N C ) N C ; (6)
see Methods and Models. In particular, discriminating similar odors will be impossible (η = 0) if their primacy sets are identical (d = 0).
We showed that primacy coding contains sufficient information to perform typical olfactory tasks with experimentally measured accuracy. Although this provides some support for primacy coding, alternative encoding schemes might also be consistent with experimental data. To elucidate this, we next compare primacy coding to two alternatives, which are also based on the simple model described by Eqs (1) and (2). The first alternative is binary coding, where glomeruli become active when their excitation exceeds a constant threshold γ [30, 51]. The second alternative is normalized coding, where the threshold is proportional to the mean excitation, γ = α N R - 1 ∑ n e n, and the inhibition strength α determines how many glomeruli are active on average [21].
To see how binary and normalized coding compare to primacy coding, we calculate the probability η that adding a ligand to a mixture of s ligands can be detected; see Fig 6A. The binary code strongly depends on the overall concentration of the presented odor (or, equivalently, the imposed threshold γ). This implies that there is only a narrow region of mixture sizes s where the binary code allows detecting the addition of a ligand. Conversely, the normalized code is concentration-invariant and could thus in principle discriminate odors at all intensities. However, we showed in Ref. [21] that the encoded information and the discriminability still depend strongly on the mixtures size s in this model. Consequently, normalized codes can only discriminate mixtures of realistic sizes when the inhibition strength α is very low and thus many glomeruli get activated on average.
The example of the normalized code shows that it is not sufficient to study how well different coding schemes can solve olfactory tasks, but one also needs to consider how useful this code is to the downstream decoder. Without modeling the decoder in detail, we here just propose that sparser codes are preferable since they imply fewer firing neurons, which saves energy and simplifies the downstream processing. In fact, sparse coding is typical for sensory information [19, 52]. In our model, the sparsity is given by the fraction 〈an〉 of activated glomeruli and Fig 6B shows this quantity as a function of the mixture size s. Since larger mixtures imply a higher odor intensity, this number increases quickly in binary coding and makes this code inefficient. Conversely, the number of active glomeruli decreases strongly in normalized coding [21], which explains the poor discriminatory performance for large mixtures. In contrast, primacy coding has a constant sparsity, because it is directly controlled by the primacy count NC. Taken together, primacy coding outperforms both binary coding and normalized coding essentially because the sparsity of the representation is independent of the presented odors and can thus be adjusted to be useful and efficient over the whole range of possible odors.
The three models discussed here differ in how the statistics of the odor c affect the statistics of the output a. In the binary model, the odor intensity given by the mean concentration ctot affects the mean excitation 〈en〉 and therefore the sparsity 〈an〉. This clearly prevents the response from being useful over a wide concentration range. This dependence on the odor intensity is removed in normalized coding, but the variance of the excitations en still depends on the odor statistics, e.g. larger mixtures imply smaller variations in en. This is problematic since it implies the excitations of fewer glomeruli exceed the fixed threshold in normalized coding, so the sparsity 〈an〉 and the usefulness decrease [21]. In primacy coding, however, the mean activity 〈 a n 〉 = N C N R is independent of the odor statistics, so the system is useful in all situations. In fact, primacy coding can be interpreted as normalized coding with an inhibition strength α that depends on the non-dimensional width of the concentration distribution; see Methods and Models. Primacy coding is thus an example for global inhibition with instantaneous adaptation, which displays better performance than a simple fixed threshold γ. Taken together, this simple model comparison indicates that the mean response of the olfactory systems needs to be controlled and that simple normalization is not sufficient for this.
So far, we calculated the transmitted information and tested the discrimination performance of primacy coding under the assumption that all receptor types behave similarly. In fact, we established that the maximal information is achieved when all receptor types are activated with equal probability NC/NR. However, neither the receptor sensitivities nor the odors themselves are distributed equally in realistic situations. Variations in these quantities affect the transmitted information and thus the usefulness of the primacy code. For instance, the transmitted information decreases if a single receptor is activated less often than all the others; see Fig 7A. This effect is small, since in the worst case the receptor is never active and the transmitted information thus corresponds to an array with this receptor removed. Conversely, having a receptor that is active more often than all others can have a much more severe effect; see Fig 7A. In fact, if the receptor type is more than three times as active, the transmitted information I is lower than if the receptor type was remove completely; see Methods and Models. This indicates that receptors can shadow the response of other receptors and thus be detrimental to the overall array when they are overly sensitive.
The effect of varying receptor sensitivities can be studied in our model of primacy coding by discussing more general sensitivities matrices. We consider S n i = ξ n S n i iid, where each receptor type can have a different sensitivity factor ξn, which modulates the uniform sensitivity matrix S n i iid where each entry is independently chosen from the same log-normal distribution. The case of homogeneous sensitivities that we discusses so far thus corresponds to ξn = 1 for n = 1, 2, …, NR.
To investigate the effect of heterogeneous sensitivities, we start by varying the sensitivity factor of one receptor type while keeping all others untouched, i.e., we change ξ1 while keeping ξn = 1 for n ≥ 2. There are three simple limits that we can discuss immediately. For ξ1 = 0, the first receptor type will never become active, the array behaves as if this type was not present, and the transmitted information is approximately Imax(NC, NR − 1). This value is lower than the maximally transmitted information Imax(NC, NR) reached for the symmetric case ξ1 = 1. However, the associated information loss ΔI = Imax(NC, NR) − Imax(NC, NR − 1) ≈ NC/(NR ln 2) is relatively small in large receptor arrays (NR ≫ NC); see Fig 7B. Conversely, the transmitted information can be affected much more severely if the sensitivity of the first receptor type is increased beyond ξ1 = 1 and the receptors will thus be active more often than the others. In the extreme case of ξ1 → ∞, the first receptor type will always be active and thus not contribute any information. Since this receptor type would always be part of the primacy set, the information transmitted by the remaining receptor types is approximately Imax(NC − 1, NR − 1), which is smaller than Imax(NC, NR − 1) in the typical case NR ≫ NC. Consequently, an overly active receptor type can be worse than not having this type at all under primacy coding.
The fact that overly sensitive receptors are detrimental to the transmitted information is also visible in numerical simulations. Fig 7B shows ensemble averages of the information I transmitted by receptor arrays as a function of the sensitivity factor ξ1. As qualitatively argued above, I is maximal for ξ1 = 1 and it is slightly lower for smaller ξ1 since the receptor type is active less often. In contrast, for ξ1 > 1, I decreases dramatically and falls below the value of ξ1 = 0 for ξ1 ≳ 1.5. These data suggest that it would be better to remove receptor types that exhibit a 50% higher sensitivity than the other types.
To see whether overly sensitive receptor types are also detrimental when all types have varying sensitivities, we next considering sensitivity factors ξn distributed according to a lognormal distribution. Numerical results shown in Fig 7C indicate that the transmitted information indeed decreases with increasing variance var(ξn) of the sensitivity factors. In fact, a variation of var(ξn)/〈ξn〉2 = 0.5 already implies a reduction of the transmitted information by almost 50% for small concentration variations σ/μ = 1. If the odor concentrations vary more, the information degradation is less severe, but the same trend is visible. Interestingly, rescaling the information by the maximal information Imax given in Eq (5) collapses the curves for all dimensions NC and NR, suggesting that this analysis also holds for realistic receptor repertoire sizes. Note that the reduced transmitted information also implies poorer odor discrimination performance; see Fig 7D. Taken together, this provides a strong selective pressure to limit the variability of the receptor sensitivities so overly sensitive receptors do not dominate the whole array.
We analyzed a simple model of neural representations of olfactory stimuli, where odors are identified by the NC strongest responding receptor types. This version of primacy coding provides a sparse representation of the odor identity, which is independent of the odor intensity. We showed using numerical simulations and a statistical model that the primacy dimension NC strongly affects the transmitted information and the discriminability of odors. Interestingly, already for small values of NC ≲ 10, the typical olfactory discrimination tasks can be carried out with performances close to experimentally measured ones. Conversely, the number NR of receptor types does not strongly affect the coding capacity and the discriminability of similar odors, in accordance with lesion experiments. Our model even indicates that lowering NR can improve the identification of a target ligand in a background.
The advantage of our simple model is that we can analyze its behavior in depth and explicitly link the statistical properties of the olfactory system to data from psycho-physical experiments. In particular, we predict how likely two different odors drawn from a particular statistics can be distinguished. For instance, our model implies that target odors are easier to detect if disturbing backgrounds consist of many ligands. We generally find that representations are sensitive to the relative concentration of ligands in mixtures and that dilute components are basically completely shadowed. Conversely, for fixed ligand concentrations mixtures can typically be discriminated very well. However, identifying the individual ligands in mixture is only possible for mixtures with few components. In any case, our results suggest that the primacy code formed in the olfactory bulb is more useful to identify odors in the subsequent olfactory cortex than simple alternatives, essentially because the statistics of the representations are independent of the odor statistics.
Our model predicts that receptors are only useful if their likelihood to respond to incoming odors is similar. This is because receptor types that are overly sensitive and respond strongly to many odors could dominate other types and thus degrade the total information. In fact, having a receptor type that is 50% more sensitive than others, and thus responds about three times as often, can lead to less transmitted information than when this type is absent. This observation is related to the primacy hull discussed in [38], which also predicts strong restrictions on the receptor sensitivities stemming from primacy coding. Various strategies could play a role in keeping the activity of the receptor types similar [53]: On timescales as short as a single sniff, the inhibition strength could be adjusted to regulate the relative importance of receptor excitations [54]. On longer timescales of several weeks, there are changes of the receptor copy number that directly affect the sensitivity of the glomeruli [55–57] and the processing neurons in the olfactory bulb [58, 59]. Receptor copy number adaptations influence the signal-to-noise ratio at the receptor level, so the copy number could be increased to improve the detection of frequently appearing odors [60]. In contrast, we predict a decrease of the copy number of overly sensitive receptor types that respond often. Combining the two alternatives, receptor copy numbers could be controlled such that noise is suppressed sufficiently while ensuring that single receptor types do not dominate the array. Finally, receptor sensitivities can also be adjusted by genetic modifications on evolutionary timescales [61, 62]. Moreover, direct feedback from higher regions of the brain could modify the processing of olfactory signals, e.g., in response to the behavioral state [7]. Although our work shows that the activities of the receptors need to be balanced, the actual distribution of the sensitivities matters much less. For instance, log-uniform distributions, which have been suggested to describe realistic receptor arrays [51, 63], lead to similar odor discriminability as log-normally distributed sensitivities; see S2 Fig.
Our results raise the question why mice have 20 times as many receptor types than flies although the transmitted information under primacy coding is only increased by a factor of 2 (see Eq (5)) and the odor discriminability is hardly affected by the receptor repertoire size (see Fig 4). The apparent usefulness of large receptor repertoires hints at roles of the olfactory system beyond transmitting the maximal information and discriminating average odors. For instance, having many receptor types might help to hardwire innate olfactory behavior when receptors are narrowly tuned to odors. In this case, our model would only apply to the fraction of the receptor types that are broadly tuned and are not connected to innate behavior. Alternatively, having many receptor types might be advantageous to discriminate very similar odor mixtures, to cover a larger dynamic range in concentrations of individual ligands, or to allow for a larger variation in average sensitivities, enabling quick adaptation to new environments. Finally, biophysical constraints of the receptor structure might imply that many receptors are required to cover a large part of chemical space.
We discussed the simplest version of primacy coding with a minimal receptor model and a constant primacy dimension NC implemented by a hard threshold. This model neglects the complex interactions of ligands at the olfactory receptors, which can affect perception [64]. In particular, antagonistic effects can already provide some normalization at the level of receptors [65]. Generally, it is likely that many mechanisms contribute to the overall normalization of the receptor response [66]. A more realistic model of primacy coding might also consider a softer threshold, where receptor types with larger excitation are given higher weight in the downstream interpretation, which is related to rank coding [22]. In this case, information from fewer glomeruli might be sufficient to identify odors, since the rank carries additional information. Realistic olfactory systems could also use a timing code, taking into account more and more receptor types (with decreasing excitation) until an odor is identified confidently. Such a system could explain that the response dynamics in experiments depend on the task [67, 68]. Generally, a better understanding of the temporal structure of the olfactory code [8, 69–73] might allow to derive more detailed models. These could rely on attractor dynamics that are guided by the excitations and thus respond stronger to the early and large excitations [74, 75].
All numerical simulations are based on ensemble averages over sensitivity matrices Sni. The elements of Sni are drawn independently from a log-normal distribution with var ( S n i ) / S ¯ 2 = 1 . 72 corresponding to λ = 1. In Figs 2A and 7B–7D, an additional ensemble average over odors c is performed using the distribution Penv(c). Here, odors c are chosen by first determining which of the NL ligands are present using a Bernoulli distribution with probability p = s/NL and then independently drawing their concentration from a log-normal distribution with mean μ and standard deviation σ. In all simulations the primacy set a corresponding to c is given by the NC receptors with the highest excitation calculated from Eq (1). Statistics of a and the transmitted information I given by Eq (4) are determined by repeating this procedure 105 and 107 times, respectively.
In order to obtain deeper insights into the numerical results, we also develop analytical approximations using a statistical description of all involved quantities, which is based on accounting for the means and variances of the respective distributions. For instance, the statistics of the output a given by Eqs (1)–(3) can be estimated using ensemble averages of sensitivity matrices for different odors c, similar to our treatment presented in [21] and [30]. In particular, Eq (1) implies that the effects of different ligands are additive. Since the log-normal distribution describing the sensitivities is narrow (λ = 1), the excitations en are also well approximated by a log-normal distribution with mean 〈 e n 〉 S = S ¯ ∑ i c i and variance and var S ( e n ) = var ( S n i ) ∑ i c i 2 [76], whereas correlations are negligible [21]. The probability that the excitation en exceeds the threshold γ, and the associated receptor type is thus part of the primacy set, reads
⟨ a n ⟩ S = 1 - G ( γ ( c ) ⟨ e n ⟩ S ; ζ ( c ) ) (7)
with
G ( x ; ζ ) = 1 2 + 1 2 erf ( ζ + log ( x ) 2 ζ 1 2 ) (8)
being the cumulative density function of a log-normal distribution with 〈x〉 = 1 and var(x) = exp(2ζ) − 1. The width of the distribution is determined by the positive parameter ζ = 1 2 ln ( 1 + var ( e n ) / 〈 e n 〉 2 ), which reads
ζ ( c ) = 1 2 ln [ 1 + ( e λ 2 - 1 ) ∑ i c i 2 ( ∑ i c i ) 2 ] (9)
for an ensemble average over sensitivities. Note that ζ is concentration-invariant, since it does not change when the concentration vector c is multiplied by a constant factor. In the simple case of ligands that are distributed according to Penv(c), we find 〈 ( ∑ i c i 2 ) ( ∑ i c i ) - 2 〉 c = s - 1 ( 1 + σ 2 / μ 2 ). Consequently, the distribution width ζ is large for broadly distributed sensitivities (large λ), few ligands in an odor (small s), and wide concentration distributions (large σ/μ).
The constraint Eq (3) implies 〈an〉 = NC/NR, so that the mean threshold reads
⟨ γ ⟩ = ⟨ e n ⟩ S · G - 1 ( 1 - N C N R ; ζ ) , (10)
where G−1 is the inverse function of G defined in Eq (8). Using this expression as an estimate for γ in Eq (7) results in concentration-invariant activities an, since 〈γ〉 is proportional to the excitation 〈en〉. This situation is comparable to simple normalized representations resulting from the threshold γ = α〈en〉, where α is a constant inhibition strength [21]. In fact, primacy coding can be interpreted as global inhibition with an inhibition threshold depending on the width of the excitation distribution, α = G - 1 ( 1 - N C N R - 1 ; ζ ).
|
10.1371/journal.ppat.1001031 | Bim Nuclear Translocation and Inactivation by Viral Interferon Regulatory Factor | Viral replication efficiency is in large part governed by the ability of viruses to counteract pro-apoptotic signals induced by infection of the host cell. Human herpesvirus 8 (HHV-8) uses several strategies to block the host's innate antiviral defenses via interference with interferon and apoptotic signaling. Contributors include the four viral interferon regulatory factors (vIRFs 1–4), which function in dominant negative fashion to block cellular IRF activities in addition to targeting IRF signaling-induced proteins such as p53 and inhibiting other inducers of apoptosis such as TGFβ receptor-activated Smad transcription factors. Here we identify direct targeting by vIRF-1 of BH3-only pro-apoptotic Bcl-2 family member Bim, a key negative regulator of HHV-8 replication, to effect its inactivation via nuclear translocation. vIRF-1-mediated relocalization of Bim was identified in transfected cells, by both immunofluorescence assay and western analysis of fractionated cell extracts. Also, co-localization of vIRF-1 and Bim was detected in nuclei of lytically infected endothelial cells. In vitro co-precipitation assays using purified vIRF-1 and Bim revealed direct interaction between the proteins, and Bim-binding residues of vIRF-1 were mapped by deletion and point mutagenesis. Generation and experimental utilization of Bim-refractory vIRF-1 variants revealed the importance of vIRF-1:Bim interaction, specifically, in pro-replication and anti-apoptotic activity of vIRF-1. Furthermore, blocking of the interaction with cell-permeable peptide corresponding to the Bim-binding region of vIRF-1 confirmed the relevance of vIRF-1:Bim association to vIRF-1 pro-replication activity. To our knowledge, this is the first report of an IRF protein that interacts with a Bcl-2 family member and of nuclear sequestration of Bim or any other member of the family as a means of inactivation. The data presented reveal a novel mechanism utilized by a virus to control replication-induced apoptosis and suggest that inhibitory targeting of vIRF-1:Bim interaction may provide an effective antiviral strategy.
| Human herpesvirus 8 (HHV-8) is a pathogen associated with cancers Kaposi's sarcoma (KS), an endothelial cell disease, and B cell malignancies primary effusion lymphoma and multicentric Castleman's disease. KS is particularly prevalent amongst HIV-positive populations in Africa and is a major health concern. Virus productive replication, in addition to latency, is important for maintaining viral load within the host and also for KS pathogenesis. Essential to HHV-8 and other virus replication is the control of innate host defenses, which comprise stress-sensing cellular signaling pathways that result ultimately in programmed cell death (apoptosis). Here we identify a novel mechanism whereby a viral protein, viral interferon regulatory factor-1 (vIRF-1), mediates inhibition of a stress sensor and initiator of apoptosis, Bim, by inducing its translocation to the cell nucleus and thereby sequestration away from the cytoplasmic compartment where it exerts its pro-death activity. We show that vIRF-1:Bim interaction is necessary for efficient HHV-8 productive replication and that it can be blocked using a cell-permeable antagonist of vIRF-1:Bim binding. Our data not only identify previously unsuspected mechanisms of Bim inactivation and vIRF-1 function, but suggest that inhibitory targeting of vIRF-1 interaction with Bim may be of therapeutic benefit.
| Human herpesvirus 8 (HHV-8) is associated with the endothelial tumor Kaposi's sarcoma in addition to the B cell malignancies primary effusion lymphoma (PEL) and multicentric Castleman's disease [1]–[3]. Several genes, including vIRF-1, have been noted to have oncogenic capacity in culture and in in vivo models [4], [5]. However, most of these genes are expressed during productive, lytic replication, suggesting that they do not play direct roles in malignant pathogenesis, but rather serve to enhance virus production. Oncogenic properties such as promotion of proliferative signaling pathways and cell survival are indeed consistent with putative roles in establishing conditions that are conducive to virus productive replication. For example, the viral IRFs function to block innate cellular responses of cell cycle arrest and apoptosis that would be induced by virus replication [6]–[8], and these properties, functioning normally to promote virus replication, could also be pro-oncogenic in experimental systems. Of note is that vIRF-1 can bind to and inhibit interferon-activated apoptotic effector proteins such as p53 and GRIM19 [(gene for) retinoid-IFN-induced mortality 19] in addition to p53-activating ATM [9]–[11]. In contrast to investigations of pro-survival and pro-tumorigenic activities of the vIRFs, studies of the functions of these proteins in normal virus biology, and in particular their roles during lytic replication, are lacking, although it is speculated that they do indeed function to enhance virus production by countering innate cellular defenses.
Previous studies from this laboratory noted the importance of the pro-apoptotic BH3-only protein Bim in negative regulation of HHV-8 productive replication in endothelial cells [12]. The viral chemokines vCCL-1 and vCCL-2 were found to induce signal transduction in endothelial cells leading to the repression of Bim induction following starvation-mediated stress and to promote virus replication, effected via both endogenously produced and exogenously added v-chemokines. The central relevance of Bim to productive replication of HHV-8 was indicated more directly by the demonstration that HHV-8 production was massively increased in cells depleted of Bim via shRNA transduction [12]. In this system, the positive effects of vCCL-1 and vCCL-2 were abrogated, suggesting that these viral chemokines exert pro-replication effects via control of lytic cycle-induced Bim expression, thereby acting to inhibit apoptosis and allow a window for virus production.
Bim is induced by a number of stress factors, such as nutrient deprivation, growth factor withdrawal, U.V. irradiation and anti-tumor drugs, in addition to stress induced by virus replication [13]. The classical model of Bim activation is via JNK-mediated phosphorylation, of long (BimL) and extra-long (BimEL) splice-isoforms of Bim, and consequent release from dynein-motor complexes, allowing translocation of the BH3-only protein to mitochondria [14], [15]. Disruption of Bim-cytoskeletal sequestration can also be mediated via induction of Gadd45, regulated by p53 and a mediator of cell cycle arrest and apoptosis [16]; Gadd45 may promote Bim release in part through activation of JNK kinase MEKK4. At mitochondria, Bim triggers apoptosis via interactions with anti-apoptotic Bcl-2 proteins to relieve suppression of Bax/Bak apoptotic effector oligomerization and pore formation in mitochondrial membranes. In addition to cytoplasmic sequestration via dynein motor association, negative regulation of BimEL can be effected by AKT and ERK phosphorylation, leading to 14-3-3 cytosolic sequestration and proteasomal degradation of Bim, respectively [17]–[19]. The short isoform of Bim, BimS, lacks all phosphorylation sites (JNK, AKT and ERK targets) and therefore cannot be regulated like its larger counterparts. However, in contrast to these proteins, expression of BimS appears to be highly restricted in vivo, and its precise role remains unclear [20], [21].
Here we identify a novel mechanism of Bim inactivation, via nuclear sequestration, mediated by HHV-8 vIRF-1. The data presented suggest that disruption of such regulation could provide a unique and useful means of controlling HHV-8 productive replication.
Whilst we previously observed that HHV-8 encoded chemokine signaling led to diminished Bim expression [12], immunofluorescence analysis of Bim expression in HHV-8 lytically infected endothelial cells revealed that detectable Bim was largely sequestered in the nuclei, rather than in the cytoplasm where it is normally localized and known to function. Thus, co-staining for Bim and K8.1 late lytic antigen, to identify cells supporting lytic reactivation in HHV-8+ telomerase-immortalized endothelial (TIME) cells [22], enabled correlation of lytic infection with Bim nuclear localization (Fig. 1A). Co-staining for early (vIRF-1, ORF59) lytic antigens, in addition to K8.1, again demonstrated Bim nuclear localization specifically in cells supporting lytic reactivation, providing verification of this phenomenon (Fig. 1B). Also revealed in this experiment was correspondence of nuclear staining patterns of Bim and vIRF-1, suggesting the possible involvement of vIRF-1 in Bim nuclear localization. It is worth noting that while some cytoplasmic Bim staining was occasionally detected in cells co-staining for lytic antigen, nuclear Bim staining was always predominant, and nuclear-localized Bim was never detected in mock-infected cultures (+TPA).
To investigate the potential role of vIRF-1 in mediating nuclear localization of Bim, cotransfection assays (in HEK293T cells) were employed. Expression vectors for BimEL [extra-large isoform [23], Flag-tagged] and vIRF-1, in addition to other HHV-8 nuclear proteins or GFP negative control, were utilized in these experiments. Cytoplasmic versus nuclear distribution of Bim in the absence and presence of the co-expressed viral proteins was determined by immunoblotting of the respective fractions. The results showed that vIRF-1, specifically, induced nuclear translocation of Bim (Fig. 2A). This was verified in intact cells by immunofluorescence assay (IFA); vIRF-1, co-expressed with BimEL in transfected cells, was able to induce nuclear translocation of the BH3-only protein (Fig. 2B, top), consistent with the western data. This effect was not seen with Puma, another pan-Bcl-2-binding BH3-only protein, demonstrating specificity of the effects seen with Bim (Fig. 2B, bottom). Whether vIRF-1 was able to interact (directly or indirectly) with Bim was tested by utilizing glutathione-S-transferase (GST)-fused bacterially-derived recombinant vIRF-1 in co-precipitation assays. GST-vIRF-1 was added to lysates of Flag-BimEL transfected HEK293T cells or to lysates of BCBL-1 (PEL) cells [24], naturally expressing high levels of Bim, and glutathione bead-precipitated material analyzed by immunoblotting. Bim was co-precipitated in a vIRF-1-dependent manner (Fig. 2C). Evidence of vIRF-1:Bim interaction was obtained also from immunoprecipitations from lysates of cells cotransfected with Flag-BimEL and vIRF-1 expression vectors (Fig. 2D). Direct interaction between vIRF-1 and Bim was demonstrated by co-precipitation assays utilizing bacterially-expressed and purified proteins, fused to GST and chitin-binding domain (CBD) sequences, respectively. vIRF-1 could be co-precipitated with CBD-fused Bim (short, long and extra-long isoforms), but not with CBD alone, following sedimentation with chitin beads (Fig. 2E). Together, these data provided evidence of vIRF-1:Bim association and of vIRF-1-mediated Bim nuclear translocation. That vIRF-1:Bim co-localization was seen in cells lytically infected with HHV-8 (Fig. 1B) suggested biological relevance of vIRF-1:Bim interaction.
As vIRF-1 is known to interact with several cellular proteins, such as IRFs, p53, ATM, GRIM19 and Smads [25], [9]–[11], [26], precise mapping of its interaction with Bim was necessary to enable experimental assessment of the functional significance of vIRF-1:Bim interaction, specifically. To this end, a series of successively refined deletion variants of vIRF-1 were generated as bacterially expressed GST-fusion proteins for use in in vitro co-precipitation assays, along with CBD-fused BimEL. Full-length vIRF-1-GST and derivatives containing the central region (residues 80–256) could be co-precipitated with BimEL-CBD using chitin beads, demonstrating involvement of these sequences in binding (Fig. 3A, top). Further deletion analysis revealed that the central portion of this region was sufficient for binding (Fig. 3A, middle). Based on this result, sequences coding for overlapping 18-mer peptides derived from this central portion were cloned to further map the Bim-binding sequences. The region corresponding to residues 170–187 (peptide-4) was sufficient for association with BimEL in this assay (Fig. 3A, bottom). Mutations within this putative amphipathic α-helical region were introduced to fine-map the Bim-binding region of vIRF-1; mutation of the central residues (174–181) abrogated interaction (Fig. 3B). Therefore, a region of vIRF-1 sufficient for direct interaction with BimEL was mapped to vIRF-1 residues 170–187, the core amino acids 174–181 of this Bim-binding domain (BBD) being required for binding. This region of vIRF-1 is divergent from collinear regions of other IRFs, both viral and cellular.
Next, the relevance of the BBD region to vIRF-1-mediated nuclear localization of Bim was examined. First, the 170–187 BBD of vIRF-1 was tested for its ability to bind to and effect nuclear localization of Bim when linked to a nuclear localization signal (NLS). Sequences encoding BBD and NLS were fused to the GFP open reading frame in a eukaryotic expression vector; a vector specifying GFP-NLS, lacking the 18-mer BBD coding sequence, was made to provide a control. The former, specifically, was able to induce nuclear translocation of BimEL in appropriately transfected cells, as determined by IFA (Fig. 4A), demonstrating sufficiency of the mapped sequences for interacting with BimEL intracellularly and enabling its nuclear translocation (directed by NLS). The requirement of the 174–181 region and core residues 178/179 of vIRF-1 BBD for Bim nuclear translocation in the context of full-length vIRF-1 was demonstrated in analogous experiments utilizing wild-type and BBD core-deleted or -mutated vIRF-1 (Fig. 4B). The requirement of these residues for Bim interaction was determined directly by co-precipitation assay using lysates of HEK293T cells transfected with expression vectors for wild-type or BBD-mutated/deleted vIRF-1 proteins and Flag-tagged BimEL (Fig. 4C). Although disrupting Bim interaction and nuclear localization by vIRF-1, the substitution and deletion mutations had no significant effects on functional interactions of vIRF-1 with p53, Smad3 and IRF-1 having potentially overlapping interactions with the central region of vIRF-1 [25], [27], [10], [11] (Fig. 4D). Thus, the introduced mutations disrupted vIRF-1 interaction with Bim specifically.
The functional consequence of vIRF-1 expression on cell viability and apoptosis in response to BimEL was examined using GFP- and TUNEL-based assays. In the former, GFP expression and fluorescence was diminished as a function of BimEL plasmid transfection and therefore the proportion of GFP+ cells in the population, determined by counting of cells under UV microscopy (and using co-staining with Hoechst to visualize nuclei), provided a measure of cell viability. Both apoptosis (identified by TUNEL staining) and cell viability were inhibited significantly (>50% under the conditions used) by vIRF-1 co-expression (Fig. 5A). Using the GFP viability assay and quantifying GFP fluorescence by fluorometry, we found that in contrast to wild-type vIRF-1, the Bim-refractory vIRF-1 deletion and point variants, vIRF-1(Δ174-181) and vIRF-1(GK179AA), were unable to inhibit of BimEL-induced cell death (Fig. 5B). That nuclear translocation of Bim induced by vIRF-1 could theoretically account for the observed decrease in BimEL-induced apoptosis was verified utilizing an NLS-fused version of BimEL, which accumulated predominantly in the nuclei of transfected cells (Fig. 5C, left). This construction was greatly impaired relative to native BimEL with respect to apoptotic induction, although expressed equivalently (Fig. 5C, middle and right panels), confirming the inhibitory effect of nuclear sequestration of the normally cytoplasmic protein. Combined, these data suggest that pro-apoptotic functions of Bim are lost upon nuclear localization and that inhibition of Bim activity by vIRF-1 is mediated via direct interaction that enables cytoplasmic-to-nuclear translocation of the BH3-only protein by vIRF-1.
To address biological significance, vIRF-1 function and vIRF-1:Bim interaction in the context of HHV-8 lytic replication were examined. First, lentiviral vectors were generated specifying vIRF-1-targeted shRNAs for vIRF-1 depletion, or non-silencing (NS) control shRNA, to determine whether vIRF-1 contributed detectably to virus productive replication in culture. TPA-induced virus production from HHV-8 latently infected telomerase immortalized endothelial (TIME) cells was markedly reduced in vIRF-1-depleted relative to control (NS shRNA-transduced) cultures, as determined by qPCR applied to DNaseI-pretreated and -resistant (encapsidated) viral DNA (Fig. 6A).
Having identified pro-replication activity of endogenously produced vIRF-1, we next generated TIME cell cultures expressing Dox-inducible vIRF-1, vIRF-1(Δ174-181) or vIRF-1(GK179AA) to assess the relative abilities of the wild-type and Bim-refractory vIRF-1 proteins to enhance HHV-8 productive replication. While all proteins enhanced virus production in Dox-treated cells, the Bim-refractory vIRF-1 variants were significantly less active, demonstrating the contribution and importance of Bim interaction, specifically, to pro-replication activity (Fig. 6B, top). Western analysis of Bim nuclear localization in these cultures revealed vIRF-1-enhanced nuclear localization of Bim, relative to empty vector (EV) control, and apparent reductions of nuclear Bim in cultures expressing the BBD variants of vIRF-1. The latter suggests possible dominant negative activity of these Bim-refractory proteins, although the mechanism that might be involved is not clear. Nonetheless, these data demonstrate that vIRF-1 is able to induce nuclear localization of Bim in the context of virus productive replication. In parallel experiments, apoptosis induced in infected cells (positive for latency-associated nuclear antigen, LANA) upon TPA treatment was substantially reduced in vIRF-1 overexpressing cultures (+Dox), but the BBD-mutated vIRF-1 proteins displayed little or no activity (Fig. 6C). These data implicate vIRF-1:Bim interactions as centrally important for vIRF-1-mediated protection from lytic cycle-induced apoptosis. In shRNA-depleted cultures, rates of apoptosis upon TPA treatment of HHV-8 infected TIME cultures were induced ∼2.5-fold relative to control NS shRNA-transduced cultures, revealing the significant contribution of endogenously expressed vIRF-1 to suppression of lytic cycle-induced apoptosis (Fig. 6D). Together, these data indicate that vIRF-1 and vIRF-1:Bim interaction, specifically, are effective mediators of apoptotic inhibition during lytic replication and, in combination with the replication experiments (Fig. 6B), that this is important for establishing conditions conducive to efficient virus production.
The role of vIRF-1:Bim interaction, specifically, independent of vIRF-1 interactions with other cellular proteins, was further demonstrated by addition in replication experiments of Tat-fused (cell-permeable) peptides corresponding to Bim-interacting vIRF-1 residues 170–187 or Bim-refractory GK179AA-mutated equivalent. The wild-type peptide, specifically, led to reduced virus titers (Fig. 7). Again, these data indicate that vIRF-1:Bim interaction and inhibition of Bim pro-apoptotic activity are important for virus productive replication. The results further suggest that disruption of vIRF-1:Bim, perhaps through the use of small molecule inhibitors, could potentially provide a means to inhibit virus replication specifically and for therapeutic benefit.
HHV-8 encoded viral interferon regulatory factors have been noted previously to interfere with innate immune responses of cells, in particular via their inhibition of activities of cellular IRFs, necessary for interferon induction and consequent cell cycle arrest and apoptosis [6], [8]. vIRF-1 has been reported to interact with a broad range of proteins in addition to cellular IRFs; these other proteins include p53, ATM, TGFβ-activated Smad transcription factors, and GRIM19 [25], [9], [10], [26], all of which play roles in mediating innate immune functions via regulation of cell cycle and apoptosis. These activities, induced by infection, presumably must be controlled sufficiently by the virus to allow productive replication in the face of stress signaling. Bim, induced strongly and rapidly following HHV-8 lytic reactivation in latently infected endothelial cells, is a powerful inhibitor of HHV-8 production in this cell type and thus represents a biologically relevant target of vIRF-1 [12]. The data presented here indicate that vIRF-1:Bim interaction is indeed important in the context of virus replication, being necessary for substantial or major proportions of replication-enhancing and anti-apoptotic activities specified by vIRF-1 in lytically reactivated endothelial cells. Thus, vIRF-1 regulation of Bim pro-apoptotic function represents a critical component of vIRF-1 activity and one which is essential for normal virus productive replication, at least in this cell type.
To our knowledge, this is the first demonstration of interaction between an IRF homologue and a member of the Bcl-2 family, and the first report of nuclear translocation of Bim or any other Bcl-2 related protein as a means of functional inactivation. Whether physiological conditions unconnected with viral infection can promote such nuclear localization and inactivation is unclear; none has been reported to date. On the other hand, several previous investigations have noted the nuclear localization and functions of Bcl-2 family members. Bcl-2 nuclear localization can occur during oxidative stress, Bcl-2 overexpression or loss of interaction with mitochondrial-localizing protein FKBP38, with promotion of apoptosis via blocking of nuclear trafficking of transcription factors [28]–[31]. Mcl-1 has been reported to inhibit cell proliferation via inhibitory interactions with proliferating cell nuclear antigen and cyclin-dependent kinase 1 [32]–[34], whereas the Bcl-2-related protein Bok, which cannot heterodimerize with Bcl-2 or Bcl-xL, can localize to and function in the nucleus to promote apoptosis [35]. A splice variant of Bfl-1, Bfl-1S, in which mitochondrial-localizing hydrophobic C-terminal sequences are replaced with a basic nuclear-localization signal, may mediate anti-apoptotic activity via nuclear sequestration of components of the apoptotic cascade [36]. Nuclear-localized pro-death activity of apoptotic effector Bax has been proposed due to correlation of Bax nuclear translocation in response to alkylating agent (BCNU) with glioma cell sensitivity to the apoptotic inducer [37]. Thus, while there is prior evidence of nuclear localization of Bcl-2 family proteins, both pro- and anti-apoptotic, along with evidence of function in this compartment, the induced nuclear localization and inactivation of BH3-only protein Bim identified here appears to be unique and the first example of viral control of apoptosis via nuclear sequestration of a Bcl-2 family member. That a viral IRF homologue mediates this effect is also a novel finding. Our demonstration that vIRF-1:Bim interaction is both important for virus productive replication and can be inhibited via peptide-mediated disruption suggests that targeting vIRF-1:Bim interaction may provide a useful antiviral strategy.
Two short hairpin RNAs (shRNA) for vIRF-1 were cloned into pYNC352/puro (a derivative of pYTF [38]) using BamHI and MluI enzyme sites; target sequence of the shRNAs correspond to 5′- AGCCGGACACGACAACTAAGA -3′ (sh1) and 5′-ATCAAGGATTGGATAGTATGT-3′ (sh2). Sequences specifying wild-type or mutated forms of vIRF-1 were cloned into lentivirus vector pYNC352/SV40/puro using MluI and BamHI cloning sites. BimEL cDNA sequences linked to Flag were cloned between the BamHI and EcoRI sites of pcDNA3.1 (Invitrogen; Carlesbad, CA), for expression in transfected cells. Coding sequence for the nuclear localization signal (NLS) of SV40 large T antigen was inserted between the HindIII and BamHI sites of pcDNA3.1-flag-BimEL to generate a eukaryotic expression vector encoding NLS-flag-BimEL. Bacterial expression plasmids for BimEL and vIRF-1 were generated by cloning of the respective coding sequences into pTYB4 (New England Biolabs; Ipswich, MA) and pGEX-4T-1 (GE Life Sciences; Piscataway, NJ), using NcoI and SmaI sites and BamHI and EcoRI sites, respectively. The BimEL and vIRF-1 proteins were fused to intein/chitin-binding-domain (CBD) and GST, respectively, used for precipitation and purification via chitin- and glutathione-bead capture. Fas ligand promoter sequences encompassing 1.2-kb upstream of the initiator codon [39] were amplified from BCBL-1 cell DNA by PCR and cloned between the XhoI and HindIII sites of pGL3/basic to provide a reporter construction responsive to IRF-1. The NLS and BBD coding sequences were cloned between KpnI and AgeI and BsrGI and XbaI sites, respectively, of pEGFP-N1 (Clontech Laboratories; Mountainview, CA) to generate nuclear-directed GFP and GFP-BBD proteins. Coding sequences for wild-type or mutated vIRF-1 proteins were cloned between the BamHI and MluI sites of pRetroX-Tight-Pur (Clontech Laboratories) to construct viral vectors for the generation of TIME cultures conditionally expressing the proteins (+Dox).
TIME cells [22] were maintained in EGM-2 MV medium (Lonza, Walkersville, MD) containing 5% fetal bovine serum (FBS) and cytokine supplements. HEK293 and HEK293T cells were grown in Dulbecco's modified Eagle's medium supplemented with 10% FBS and gentamicin. BCBL-1 cells [24] were cultured in RPMI 1640 supplemented with 20% heat-inactivated FBS and gentamicin. For lentivirus production, HEK293T cells were transiently transfected with virus vector and gag/pol-encoding packaging plasmids using standard calcium-phosphate precipitation method and virus harvested after 48 h by centrifugation at 49,000×g, essentially as described previously [12]. Other transfections were performed using Lipofectamine 2000 (Invitrogen). Stable transduction of shRNA or cDNA into TIME cells using lentivirus vectors was performed under puromycin selection.
For whole cell extracts, cells were lysed in lysis buffer (50 mM Tris-HCl [pH 8.0], 150 mM NaCl, 1 mM EDTA, 1% IGEPAL CA-630, 0.25% sodium deoxycholate, and protease inhibitor cocktail). For nucleo-cytoplasmic fractionation, cells were homogenized in buffer A (10 mM HEPES [pH 8.0], 1.5 mM MgCl2, 10 mM KCl, 0.5 mM DTT, and protease inhibitor) using a Dounce homogenizer. After centrifugation of the homogenate at 1,500×g, the supernatant was used as the cytoplasmic fraction and the pellet, after resuspended in buffer B (20 mM HEPES [pH 8.0], 1.5 mM MgCl2, 420 mM NaCl, and 0.2 mM EDTA), was used as the nuclear fraction. For immunoblotting, proteins were size fractionated by sodium dodecyl sulfate-polyacrylamide gel electrophoresis (SDS-PAGE) and transferred to a nitrocellulose membrane. Immunoreactive bands were detected with enhanced chemiluminescence solution (GE Healthcare, Piscataway, NJ) and visualized on X-ray film. For immunofluorescence assay (IFA), cells on a 0.1% gelatin-coated coverglass were fixed and permeabilized in chilled methanol. Following incubation with superblock blocking buffer (Thermo Scientific Inc., Rockford, IL), coverslips were incubated with primary antibody, washed with phosphate-buffered saline (PBS), and then incubated with appropriate fluorescent dye-conjugated secondary antibody. Coverslips were mounted in 90% glycerol in PBS containing 10 mg/ml p-phenylenediamine. Nuclei were visualized by staining with Hoechst 33342.
Infectious HHV-8 was obtained by inducing BCBL-1 cells with phorbol 12-myristate 13-acetate (PMA/TPA; 20 ng/ml) and calcium ionophore (A23187; 500 ng/ml). After 20 h, cells were pelleted and resuspended in fresh medium without TPA and A23187. After four days, virions were pelleted from culture media by centrifugation at 27,000×g for 2 h in an SW41 rotor and resuspended in basal EGM-2 MV medium. For HHV-8 infection, TIME cells were centrifuged at 1,000×g for 1 h in the presence of HHV-8 virions, and then cultured in fresh complete medium for 7 days to allow establishment of latency in the absence of ongoing lytic replication. Lytic replication of HHV-8 in TIME cells was induced by treatment with TPA. For determination of encapsidated HHV-8 genome copy number, viral DNA was extracted using guanidinium thiocyanate (6M) and silica gel following pre-treatment of virus suspensions with DNaseI for 20 min. at 37°C. Excess HHV-8 bacmid DNA was treated with DNaseI and processed identically to control for DNaseI efficacy. All qPCRs were performed in a 96-well microplate using an ABI Prism 7500 detection system (Applied Biosystems; Foster City, CA) with SYBR green/ROX master mix (SuperArray Bioscience Corp.; Frederick, MD). Reactions were performed in a total volume of 25 µl, containing viral DNA sample and 250 nM of each primer. To calculate copy number of viral DNA, BAC-cloned HHV-8 genomic DNA was used as a standard. PCR conditions included an initial incubation step of 2 min. at 50°C, and enzyme heat activation step of 10 min at 95°C, followed by 45 cycles of 15 seconds at 95°C for denaturing and 1 min at 60°C for annealing and extension.
HEK293T cells were transiently transfected with plasmids expressing vIRF-1 and p53, Smad3 or IRF-1 along with reporter plasmids, PG13-luc (Addgene; Cambridge, MA), SBEx4-luc (Addgene), or Fas ligand promoter-luc (see above), respectively, for 24 h and then lysed with passive lysis buffer (Promega, Madison, WI). Luciferase activity was measured by standard methods using D-luciferin and luminometry.
Bim-induced cell death of HEK293 cells was monitored by cotransfection of pEGFP-N1 (Clontech laboratories, Mountain view, CA) and measuring fluorescence by fluorometry. For terminal deoxynucleotidyltransferase (TdT)-mediated dUTP-biotin nick end labeling (TUNEL), cells were fixed in chilled methanol for 5 min and preincubated in TdT reaction buffer (25 mM Tris-HCl [pH 6.6], 200 mM sodium cacodylate, 0.25 mg/ml BSA, and 1 mM CoCl2) for 10 min. TUNEL reactions were carried out at 42°C for 2 h in TdT reaction buffer containing TdT and biotin-dUTP (Roche, Indianapolis, IN) and terminated with stop solution (300 mM NaCl and 30 mM sodium citrate). TUNEL-positive cells were visualized by staining with FITC-avidin. LANA IFA was performed after TUNEL reactions, using LANA monoclonal antibody (Advanced Biotechnologies Inc; Columbia, MD) and staining with Cy-3-conjugated secondary antibody.
Glutathione-S-transferase (GST)-fusion proteins were purified by standard methods. Proteins from BimEL-transfected HEK293T or BCBL-1 cells were incubated with purified GST or GST-vIRF-1 proteins immobilized on glutathione beads. After washing with lysis buffer, the bead-precipitated material was subjected to SDS-PAGE and analyzed by immunoblotting using Bim- or GST-specific antibodies. For binding-site mapping, BimEL (lacking the C-terminal 18 residues, affecting solubility) was fused to intein-chitin binding domain (CBD) in pTYB4 and purified according to the manufacturer's protocol. GST or a series of GST-vIRF-1 fusions were incubated with the purified Bim protein immobilized on chitin beads. After washing with lysis buffer, bead-associated proteins were size-fractionated by SDS-PAGE and analyzed by immunoblotting using GST- or CBD-specific antibodies. For peptide competition assays, peptides (35-fold molar excess) were pre-incubated with BimEL-intein-CBD for 1 h before addition of GST-vIRF-1. Peptide sequences (the first 11 residues comprising Tat basic region) were: BBD-WT, YGRKKRRQRRRGGGRTRGLQEIGKGISQDGHH; BBD-Mut, YGRKKRRQRRRGGGRTRGLQEIAAGISQDGHH (altered residues underlined). For immunoprecipitation, HEK293T cells transfected with plasmids expressing Flag-BimEL or vIRF-1 were lysed in lysis buffer and cell extracts were incubated with anti-Flag antibody (M2) and immune-complexes precipitated with protein A/G-agarose. After washing with lysis buffer, immune-complexes were subjected to SDS-PAGE and analyzed by immunoblotting using vIRF-1 antiserum, Bim antibody or biotinylated Flag antibody and secondary, detection reagents comprising HRP-conjugated anti-Ig antibody or sptreptavidin-HRP.
K8.1 and LANA antibodies were purchased from Advanced Biotechnologies Inc (Columbia, MD). Antisera directed to vIRF-1 and ORF59 were provided by Drs. Gary Hayward and Bala Chandran, respectively. Actin and Flag antibodies were purchased from Sigma (St. Louis, MO), HDAC1 and GST antibodies from Santa Cruz Biotechnologies, Inc. (Santa Cruz, CA), Bim antibody from Cell Signaling Technologies, Inc. (Beverly, MA), GFP antibody from Epitomics, Inc. (Burlingame, CA), and CBD antibody from New England Biolabs, Inc. (Ipswich, MA).
|
10.1371/journal.ppat.1003347 | Challenges in Detecting HIV Persistence during Potentially Curative Interventions: A Study of the Berlin Patient | There is intense interest in developing curative interventions for HIV. How such a cure will be quantified and defined is not known. We applied a series of measurements of HIV persistence to the study of an HIV-infected adult who has exhibited evidence of cure after allogeneic hematopoietic stem cell transplant from a homozygous CCR5Δ32 donor. Samples from blood, spinal fluid, lymph node, and gut were analyzed in multiple laboratories using different approaches. No HIV DNA or RNA was detected in peripheral blood mononuclear cells (PBMC), spinal fluid, lymph node, or terminal ileum, and no replication-competent virus could be cultured from PBMCs. However, HIV RNA was detected in plasma (2 laboratories) and HIV DNA was detected in the rectum (1 laboratory) at levels considerably lower than those expected in ART-suppressed patients. It was not possible to obtain sequence data from plasma or gut, while an X4 sequence from PBMC did not match the pre-transplant sequence. HIV antibody levels were readily detectable but declined over time; T cell responses were largely absent. The occasional, low-level PCR signals raise the possibility that some HIV nucleic acid might persist, although they could also be false positives. Since HIV levels in well-treated individuals are near the limits of detection of current assays, more sensitive assays need to be developed and validated. The absence of recrudescent HIV replication and waning HIV-specific immune responses five years after withdrawal of treatment provide proof of a clinical cure.
| There is intense interest in developing a cure for HIV. How such a cure will be quantified and defined is not known. We applied a series of measurements of HIV persistence to the study of an HIV+ adult who has exhibited evidence of cure after a stem cell transplant. Samples from blood, spinal fluid, lymph node, and gut were analyzed in multiple laboratories using different approaches. No HIV was detected in blood cells, spinal fluid, lymph node, or small intestine, and no infectious virus was recovered from blood. However, HIV was detected in plasma (2 laboratories) and HIV DNA was detected in the rectum (1 laboratory) at levels considerably lower than those expected in antiretroviral treated patients. The occasional, low-level HIV signals might be due to persistent HIV or might reflect false positives. The sensitivity of the current generation of assays to detect HIV RNA, HIV DNA, and infectious virus are close to the limits of detection. Improvements in these tests will be needed for future curative studies. The lack of rebounding virus after five years without therapy, the failure to isolate infectious virus, and the waning HIV-specific immune responses all indicate that the Berlin Patient has been effectively cured.
| Given the well-recognized limitations of antiretroviral therapy (ART)—which include side effects, costs, and difficulties delivering complex regimens to a global population for decades—there is intense interest in curative interventions [1], [2]. This interest in curative strategies is also driven by a single case report in which a cure was apparently achieved [3]. In 2007, an HIV-infected adult living in Berlin developed acute myelogenous leukemia (AML), for which he was treated with an allogeneic hematopoietic stem cell transplant from a donor who was homozygous for the CCR5Δ32 deletion [3], which confers resistance to infection with CCR5-utilizing virus. The patient interrupted ART soon after the transplant and has had no detectable plasma HIV RNA for over five years [3], [4]. Previous studies reported that: 1) he lacked HIV RNA in cerebrospinal fluid (CSF) [4]; 2) he had no detectable HIV DNA in PBMC, bone marrow, brain, or colon [3], [4]; 3) HIV-specific T cell responses decreased after the transplant [3]; and 4) he lost antibodies to Pol and Gag but not Env [3]. Although CCR5-expressing cells were detected in the colon at 5.5 months post-transplant [3], no CCR5-expressing cells were detected in the colon at later time points or in the liver or the brain [4].
Despite the unquestioned success of the transplant, theoretical reasons suggest that HIV could survive the transplant. These include: 1) the possible presence of X4-tropic virus prior to transplant [3], [5]; 2) the detection of rare CCR5+ macrophages 5.5 months after transplant [3]; and 3) the possibility of long lived nonhematopoietic cell reservoirs [6] that could produce virus even if the ability to replicate were constrained by lack of CCR5-expressing hematopoietic cells.
In most ART-suppressed patients, the level of persistent HIV is very low. Even with single copy assays, some patients have essentially no detectable HIV RNA in plasma (i.e., <0.1–1 copy/mL, depending on volume) [7], and only one in approximately a million circulating CD4+ T cells contains replication-competent virus [8], [9], [10], [11], [12]. The burden of HIV may be higher in the lymphoid tissues [8], [13], [14], [15] and gut [16], [17], [18]. As most long-term treated adults have HIV burdens that are near or at the sensitivity of current assays, it is unclear as to whether such assays will be amenable to monitoring virologic responses to future curative interventions.
In 2011, the Berlin Patient transferred his care to San Francisco and consented to a series of studies aimed at determining if HIV persisted. Multiple samples were obtained from a number of different sites, including lymphoid tissues, and relatively large biological inputs were analyzed using the most sensitive assays available. Our objectives were to determine whether the transplant resulted in a sterilizing cure and to assess the capacity of currently available assays to detect and quantify possible low-level persistence.
The subject was enrolled in the UCSF-based SCOPE cohort and had multiple study visits over two years (Figure 1). Plasma, serum, and PBMC were obtained at each visit. The subject also consented to separate procedures at UCSF, including leukapheresis, lumbar puncture, and flexible sigmoidoscopy with rectal biopsies. He was also seen at the University of Minnesota, where he underwent a lymph node biopsy and a colonoscopy with ileal and rectal biopsies. In order to compare HIV antibody levels in the subject with those from HIV-infected and HIV-uninfected subjects, blood samples were analyzed from participants in another UCSF based pathogenesis cohort, the Options study. All procedures were approved by the Institutional Review Boards at UCSF and the University of Minnesota.
Samples (plasma, PBMC, CSF, lymph node, gut biopsies) were analyzed in a number of different laboratories with expertise in detecting extremely low levels of virus, including the Blood Systems Research Institute (BSRI); Johns Hopkins; the National Institutes of Health (NIH); the Karolinska Institutet in Solna, Sweden; the University of California San Diego (UCSD); the University of California, San Francisco (UCSF); and the University of Minnesota.
Plasma HIV RNA was measured in 4 different laboratories using 4 different techniques: 1) Roche Ampliprep assay [19]; 2) single copy assay [7], [20] (SCA); 3) Transcription-Mediated Amplification (TMA) [21], [22]; and 4) a modified Abbott assay that involves pelleting virus from 30 ml of plasma [23] (Table 1). Cell-associated HIV DNA in PBMC was measured in 4 laboratories using qPCR [19], [22], [24], [25] or digital droplet PCR [26] (Table 2). Cell-associated HIV RNA in PBMC was measured in 3 laboratories using the Roche Ampliprep [19], qRT-PCR [24] or TMA [22]. Latently-infected peripheral CD4+ T cells (in infectious units per million cells, IUPM) were measured in 2 laboratories by co-culture [27], [28].
Lumbar puncture was performed once (Table 3). The CSF was spun down to yield 8,400 cells, which were tested for HIV DNA in one laboratory [19]. The supernatant CSF was tested for HIV RNA in two laboratories [7], [19].
Thirty rectal biopsies were obtained by sigmoidoscopy. Six biopsies were snap frozen, while 24 biopsies were digested with collagenase [18] to yield total rectal cells, which were counted, aliquoted, and frozen. Total DNA and RNA were isolated from each aliquot of rectal biopsies and rectal cells using Trireagent (Molecular Research Center). RNA was further treated with RNase-free DNase (QIAgen) and purified using QIAgen RNEasy kits. DNA and RNA from each aliquot of rectal biopsies and rectal cells were tested for HIV in one laboratory using qPCR and qRT-PCR for the LTR [24] and normalized to cell equivalents by DNA or RNA mass, as determined by Nanodrop.
Colonoscopy and inguinal LN biopsy were performed at Minnesota. Tissue sections from LN, ileum, and rectum were assessed for HIV RNA by in situ hybridization (ISH) [13]. Sections were taken at 20μ intervals, and two individuals examined each section. The portion of LN that was not placed in fixative was immediately suspended in liquid nitrogen, while the remaining gut biopsies were digested with collagenase to obtain total gut cells. LN tissue and cryopreserved cells from blood and gut were then sent to the NIH, where subsequent procedures included cell sorting (blood and gut), whole transcriptome sequencing, and fluorescence-assisted clonal amplification of HIV env.
PBMC and gut cell suspensions were stained with LIVE/DEAD violet stain (Molecular Probes) and the following fluorescently-conjugated antibodies: CCR5-Cy7PE (Pharmingen); CD45RO-PE-Texas Red (Coulter); CD14-PE (Pharmingen); CD11c-PE (Pharmingen); CD3-H7APC (BD); T cell receptor-γδ-APC (Pharmingen); CD20-APC (BD); CD56-APC (Pharmingen); CD45-Qdot 800 (Invitrogen); CD8-Qdot 655 (Invitrogen); and CD4-Qdot 605 (Invitrogen). Cells were sorted on a FACS Aria into four subsets for each tissue, as shown in Figure 2. Viable single cells from PBMC from two time points were sorted into CCR5+ cells, CCR5- non-T cells, CCR5- T cells that were not CD4+, and CCR5- CD4+T cells. All gut cells from each site were sorted into one of 4 different populations: non-hematopoietic cells (CD45−); hematopoietic cells (CD45+) that were not T cells; T cells that were not CD4+; and CD4+T cells.
Snap-frozen lymph node tissue and sorted cells were homogenized in RNAzol RT (Molecular Research Center, Inc.), and DNA and total RNA were extracted separately. DNA samples were diluted to a concentration of ≤5,000 cell equivalents per well in 384-well plates and amplified with HIV env-specific primers ESf2(CCGGCTGGTTTTGCGATTCTAAARTG) and ESr1b(AGGAGCTGTTGATCCTTTAGGTATCTTTC) in PCR reactions containing SYBR Green (Invitrogen). CD4+T cell genomic DNA from a viremic HIV+ subject was amplified as a positive control. HIV env amplification was detected by SYBR Green fluorescence and specificity was confirmed by melt curve analysis.
mRNA was purified from the sorted samples above as well as PBMC from two HIV- donors, CD4+T cells from one viremic HIV+ control, and stably-infected ACH-2 cells. Messenger RNA was fragmented, and barcoded libraries were constructed and clustered on an Illumina Truseq Paired-End Flowcell v3. The flow cell was sequenced on an Illumina HiSeq 2000 in a 150bp single-read indexed run. Reads were aligned to the human genome (hg19, NCBI Build 37) and transcriptome using TopHat. Unaligned reads were then aligned to the patient's pre-transplant HIV sequences using Novoalign. Reads that did not align to these sequences were then aligned against all molecular HIV clones on the Los Alamos National Laboratories HIV database. In addition, all unfiltered reads were independently aligned to the patient's pre-transplant sequences, the HIV HXB.2 reference sequence, and all molecular HIV clones on the Los Alamos HIV database. Lastly, unfiltered reads were aligned to a wild-type (NM_001100168) and Δ32 CCR5 sequence.
Western blot for HIV was performed on blood from one time point (ARUP labs). Blood from four other time points was analyzed for HIV-specific antibody (Ab) at the BSRI using the HIV-1/2 VITROS assay (Ortho Clinical Diagnostics, Rochester, NY), a detuned version of the HIV-1 VITROS assay [29], and the Limiting Antigen Avidity assay (Sedia, Portland, OR) [30]. To assess how these antibody levels compared with levels in typical HIV infection, we compared the detuned VITROS Ab results with those for participants in the UCSF Options study, whose samples were also tested at BSRI. Results were included from subjects who were HIV- uninfected (negative for HIV antibody and HIV RNA); from subjects who were HIV infected for at least one year and ART-naïve; and from subjects who were HIV infected on suppressive ART for at least 6 months.
Antibodies to other infectious diseases, including CMV, EBV, HSV, VZV, hepatitis B, measles, mumps, rubella, and toxoplasmosis, were measured in post-transplant samples from 5/2011 to 2/2012 using Bio-Rad Bioplex 2200 multiplex and MONOLISA microtiter clinical assays. T-cell responses to HIV and CMV were measured once by flow cytometry for intracellular cytokine staining [31].
Plasma HIV RNA levels were quantified using several methods. HIV RNA was detected in 2 of 4 laboratories in plasma samples from 3 different time points (Table 1). In one laboratory, HIV RNA was detected by the Roche Ampliprep in 3 of 10 1.0 ml replicates from one time point (each with 1–2 copies, for an average of 0.4 copy/ml), 1 of 24 replicates from a second time point (at 70 copies, for an average of 2.9 copy/ml), and zero of 24 replicates from a third time point. In a separate laboratory, plasma from a fourth time point (17 ml) was positive in 2 of 6 replicates, each with 4–5 copies, for an average of <1 copy/ml. At this level of amplification, false positive signals at the level of 1–2 copies/ml can occasionally be seen, although these controls were negative when this assay was run. No HIV RNA was detected in 2 other laboratories using samples from other independent time points.
Unfractionated PBMC samples, each from a different time point, were studied in 4 laboratories using HIV sequence-specific PCR or TMA. No HIV DNA or RNA was detected (Table 2). In a fifth laboratory, PBMC samples from two different time points were sorted into CCR5− CD3− (non-T) cells, CCR5− CD4+T cells, all other CCR5− T cells and, using an inclusive gating strategy to collect all possible events, a cell population that appeared to be CCR5+ by staining (Figure 2a–g). Messenger RNA from each of these sorted populations was negative for HIV by whole transcriptome sequencing, which has been found to yield several hundred HIV-derived reads per sample in comparable experiments on sorted CD4+ T cells from ART-naïve subjects (E. Boritz and D. Douek, unpublished observations). DNA from each of these populations was negative for HIV by fluorescence-assisted clonal amplification of env, which detects individual HIV DNA genomes with a sensitivity equivalent to the widely-used quantitative PCR assay for HIV gag (E. Boritz and D. Douek, unpublished observations).
Co-cultures were performed in two laboratories using peripheral CD4+T cells from leukapheresis and from venous phlebotomy at a second time point. No replication-competent virus was detected (Table 2).
Lumbar puncture was performed once. No HIV was detected in the CSF supernatant (2 labs) or cells (1 lab) (Table 3).
One inguinal lymph node was studied in two laboratories (Table 3). No HIV RNA was detected by in situ hybridization or by whole transcriptome sequencing. DNA was negative for HIV by fluorescence-assisted clonal amplification of env.
Whole rectal biopsies and total rectal cells from sigmoidoscopic biopsies were tested for HIV DNA and RNA (Table 3). HIV DNA was detected in DNA isolated from both intact biopsies (2 of 15 wells positive, at 1 and 3 copies) and total rectal cells (1 of 10 wells positive at 1 copy) but not negative controls, while HIV RNA was not detected in either the rectal cells or rectal biopsies.
Biopsies from the ileum and rectum were obtained by colonoscopy at a second time point and analyzed in 2 other laboratories. No HIV RNA was detected in ileal or rectal tissue by ISH or by whole transcriptome sequencing of mRNA from sorted populations of CD45− (non-hematopoietic) cells, CD45+CD3− (non-T) cells, CD4+T cells, or other T cells from ileum and rectum. DNA from each of these populations was negative for HIV by fluorescence-assisted clonal amplification of env.
No sequence data are available from the samples where HIV was detected in the quantitative assays (plasma and rectum). Cloning and sequencing was not attempted from plasma, as no plasma remained from the three time points that were positive for HIV RNA. Likewise, very little rectal DNA remained after repetitive testing by qPCR, and limited attempts to amplify the env proved unsuccessful.
Although no HIV was detected in the PBMC by quantitative assays, two labs attempted to amplify and clone env from PBMC. Sequencing from one laboratory revealed a subtype B HIV-1 envelope (predicted to be X4) that does not match any known sequence but is also highly divergent from the pre-transplant consensus R5 sequence (16.5% genetic distance). Sequences obtained from PBMC in a second laboratory were consistent with a lab contaminant.
Whole transcriptome sequencing reads from lymph node and sorted subsets of PBMC, ileal cells, and rectal cells were aligned against the CCR5 reference. All 548 reads aligning to the area of the CCR5Δ32 deletion were consistent with the Δ32 sequence of CCR5. Of particular interest were the CCR5 sequences transcribed by the PBMC populations that appeared to express CCR5 protein by flow cytometry. Though these populations were quite sparse (for example, see Figure 2d), they were sorted separately to ensure directed molecular analysis of any potential CCR5-expressing cells. As a result, whole transcriptome sequencing of these populations would have been expected to show abundant CCR5 mRNA had they truly been CCR5+. Instead, CCR5 mRNA was detected at very low levels in these “CCR5+” populations (1.0 and 2.8 CCR5 reads per million total reads in the two sorted PBMC populations, compared to 53.5 and 37.4 CCR5 reads per million total reads in CD4+ T cell populations from ileum and rectum). This suggests that the cells were either not expressing the gene or that the gene transcripts were unstable. Furthermore, all CCR5 reads from the “CCR5+” populations contained the CCR5Δ32 deletion, suggesting that the CCR5 labeling of these cells reflected a fluorescence staining artifact rather true protein expression.
Western blot was 2+ (strongly positive) for gp160, +/− for p24, and negative for other bands (inconclusive, possibly infected). Blood from four other time points was analyzed using 3 different serological assays. Using the undiluted HIV-1/2 VITROS assay, HIV Ab was readily detectable and relatively stable at all 4 time points, with levels considerably above the cutoff for HIV negative individuals (Figure 3A) but beyond the dynamic range of the assay. In the detuned HIV-1 VITROS assay, HIV specific Ab were detectable at levels that were above those seen in HIV-negative individuals but below those seen in HIV-infected individuals before and after long term suppressive ART (Figure 3D). Moreover, HIV Ab levels tended to decrease over time by both the detuned VITROS and the Limiting Antigen Avidity assays (Figure 3B–C). We screened post-transplant samples for chronic infections using several multiplex commercial assays and identified three viruses (EBV, hepatitis B, and measles) against which the subject had serologic responses that were positive but not above the dynamic range of the assays. In contrast with HIV Ab levels, Ab levels for these viruses remained stable over a 9 month period post-transplant (Figure 3E).
PBMC were stimulated with CMV pp65 or HIV Gag peptide pools, and flow cytometry was used to measure the frequencies of CD4+ and CD8+T cells with intracellular staining for cytokines. HIV Gag-specific responses from the subject were compared to responses in at-risk HIV-uninfected adults, chronically HIV-infected adults on long term ART with undetectable plasma HIV RNA levels, and chronically infected adults controlling HIV in the absence of therapy (“elite” controllers). The frequencies of T cells expressing or producing CD107, interferon-γ, IL-2 and TNF-α were generally consistent with the levels seen in HIV-uninfected adults (Figures 4A–B), lower than those observed in HIV-infected adults on long-term combination antiretroviral therapy (Figures 4C-D), and much lower than those observed in elite controllers (Figures 4E–F). In contrast, CMV-specific responses were robust and higher than those observed in HIV-uninfected adults (data not shown) [30].
Multiple measurements of HIV persistence were applied to the study of the “Berlin Patient,” who has achieved an apparent cure after a hematopoietic stem cell transplant from a homozygous CCR5Δ32 donor. The vast majority of assays revealed no evidence for persistent HIV, confirming prior reports [3], [4], although 3 independent laboratories detected low level PCR signals in 4 different samples from 2 sites (plasma and rectum). The calculated HIV levels were much lower than typical ART-treated patient patients (Table 4) and close to the detection limit of the most sensitive assays. Replication-competent virus was not detected in PBMC, and HIV-specific antibody (as measured by detuned assays) tended to decline over the 18 months of observation. HIV-specific T cell responses were not detected.
The positive signals can be interpreted in two ways. The first possibility is that all 4 signals could be false positives, perhaps due to nonspecific amplification (e.g., from human endogenous retroviruses) or from contamination of from other clinical samples or lab viruses. The relatively large inputs of biological material and large numbers of replicates may have afforded more opportunities for nonspecific amplification or contamination, especially at copy numbers that are close to the limits of assay detection. Negative controls were run alongside the patient samples for all but the automated Roche assay, and they remained uniformly negative. However, the number of negative controls was generally less than the number of patient wells. For future studies, it may be better to have blinded samples in which the test samples are mixed with an equal or greater number of negative controls.
To better approximate the rate of false positives in a larger number of samples, we used data from other experiments to determine the number of positives among all wells containing samples from HIV-uninfected donors that had been extracted and run along with positive samples. Depending on the number of samples tested, false positives could be documented for most assays (Tables 1–3). In one assay for plasma HIV RNA, the false positive rate of 22% was not much lower than the fraction of positive wells observed for the Berlin Patient (33%), increasing the probability that these may have been false positives. For other assays, the false positive rate was considerably lower than the fraction of positive wells observed for the Berlin Patient, although it is also possible that false positives reactions could occur above the “predetermined” rate.
The second possibility is that one or more of these signals were true positives. If HIV persists at extremely low levels or in nonhomogeneous foci, the detection could be inconsistent and could depend on sampling variability, sample volume, number of replicates, or differences in sensitivities of the various assays. In addition, some assays did not rigorously control for inhibitors of PCR, which can contribute to false negatives. However, even if some HIV DNA or RNA sequence persists, the significance of this finding is uncertain, as most proviral DNA is transcriptionally silent [18], [32] and/or defective [33], [34], and most HIV virions are not infectious [35]. One recent study suggests that the number of resting CD4+ T cells with defective proviruses may be an average of 300 fold higher than the number that can produce replication-competent virus following cellular activation [36].
Despite the PCR signals from plasma, no replication-competent virus could be cultured from circulating blood cells. The vast majority of ART-treated subjects have detectable replication-competent virus using current co-culture assays, particularly if large number of cells are studied, as was the case in this study. Virus isolation is often more challenging in “elite” controllers, however, suggesting these assays have limited sensitivity to detect rare events. Also, it is worth noting that co-cultures would have only assessed for infectious virus in the peripheral CD4+ T cells, which were negative for HIV DNA by PCR, rather than the plasma or gut, where HIV was detected.
The low level of HIV-specific T cell responses, the low and declining levels of HIV-specific Ab in detuned and limiting antigen avidity assays, and the partial seroreversion (based on an indeterminate western blot) are all different from the usual course in antiretroviral-treated adults and “elite” controllers, suggesting very low to absent levels of HIV antigen. Indeed, the levels of residual HIV-specific Ab and T cells detected in this patient were considerably lower than those seen in chronically–infected individuals who have been on suppressive ART for many years, although low or undetectable HIV-specific immune responses have been observed in some patients treated early in the course of HIV infection [37], [38], [39]. Complicating this interpretation is the likelihood that the transplant procedure caused the loss or irreversible impairment of host B and T cells. In HIV-uninfected individuals who have undergone allogeneic stem cell transplant, serum antibodies to tetanus decay at variable rates, but remain detectable in half of patients for a year or more [40]. On the other hand, antibody levels to other infectious diseases, including some associated with latent viral infection, were present at typical levels and stable in the Berlin Patient over a 9 month period post-transplant. Based on these findings we propose that measurement of HIV-specific immune responses may prove to be a sensitive and accurate means to diagnosis an effective cure.
Given the multiple negative signals and the waning HIV-specific immunologic responses, it is possible that all of the positive signals were false positives. True verification of HIV persistence requires sequence confirmation. Unfortunately, the same factors that limit the ability to detect rare, low-level foci of HIV also make it difficult to amplify or clone rare sequences. No sequences are available from the sites where HIV was detected in the quantitative assays (plasma and gut), while the one available sequence comes from the PBMC, where HIV could not be detected in the quantitative assays, and differs from the pre-transplant consensus sequence. Although this sequence does not match any published sequence, it is possible that it is a contaminant from another clinical isolate. Opportunities for contamination may be higher during the nested PCRs and multiple steps used for env cloning than they may be in a one step, closed reaction for a quantitative real time PCR.
If one or more PCR signals is a true positive, one must ask why the HIV was not detected previously, and why is it is now detected intermittently. It is possible that HIV persists at extremely low levels, perhaps in rare foci that are frequently missed in sampling. Another possibility is superinfection, especially since the PBMC sequence is predicted to be X4 tropic. Arguing against this possibility is that the vast majority of transmission occurs by one or a limited number of CCR5-utilizing virions [41], [42], HIV infection is extremely rare in homozygous CCR5Δ32 individuals [43], [44], [45], [46], [47], [48], [49], and transmission of X4 virus would be expected to result in sustained replication, which has not occurred.
If HIV persists, one must also question why it has failed to spread, and where it persists. The HIV could be R5-tropic, replication-defective, or unable to overcome his immune defenses. Anatomic sources could include the gut, brain, male genital tract, or lymphoid tissues such as the mesenteric lymph nodes and spleen. Cellular sources could include rare host hematopoietic cells, donor immune cells infected in a non-R5 dependent fashion [50], [51], nonhematopoietic cells (such as neural progenitor cells, astrocytes, neurons, and oligodendrocytes [6]), or cell-associated extracellular virus on the follicular dendritic network [13], [14], [52], [53], [54].
In addition to well-validated PCR, TMA, and in situ hybridization assays, we used whole transcriptome sequencing to search for HIV mRNA in the patient samples. Deep sequencing offers several advantages in HIV eradication studies. Because mRNA libraries incorporate all transcripts from each sample, the method's sensitivity is not limited by differences between the sequences of synthetic primers or probes and a patient's autologous viral sequences. In addition, all viral transcripts are quantified simultaneously, allowing characterization of differences in viral gene expression between different cell types, tissues, or individuals. Finally, concurrent sequencing of cellular transcripts allows discernment of host factors, such as CCR5 genotype. It is important to acknowledge that the whole transcriptome sequencing performed here would be expected to detect cell-associated HIV RNA down to a level approximately two orders of magnitude lower than in viremic patients. Therefore, the resulting data do not by themselves rule out a very low frequency of residual HIV RNA-expressing cells in the patient. In future eradication studies, however, greater levels of sensitivity may be achieved by sequencing greater proportions of the mRNA libraries produced. Advances in library preparation and increases in data yield per unit cost may soon make this depth of sequencing routine, allowing detection of single HIV-infected cells within very large cell and tissue samples.
Despite the possibility of intermittently detectable, very low levels of HIV, the Berlin Patient has remained off of ART for 5 years, has no detectable viremia using standard assays, has waning HIV antibody levels, has limited to undetectable HIV-specific T cell responses, and has no evidence of HIV-related immunologic progression. The patient certainly meets any clinical definition for having achieved a long-term remission, and may even have had a sterilizing cure. Even the most extraordinary “elite” controllers described in the literature have more robust evidence for persistent infection [55]. Given that all assays are susceptible to false positive results, and given our inability to provide sequence confirmation, we cannot robustly document HIV persistence and hence conclude he has been cured. This inability to prove a negative is not unexpected, and will likely prove to be problem in future studies on curative interventions. Similar problems are emerging with regard to detection of persistence in other settings, including post-allogeneic stem cell transplants and potentially cured infants. A key priority for the field is the development of robust, well-validated assays that can readily detect very low levels of virus. Future studies of eradication will ultimately require interruption of treatment and subsequent long-term observation as the gold standard for defining a successful outcome.
|
10.1371/journal.pcbi.1004529 | Analyzing and Quantifying the Gain-of-Function Enhancement of IP3 Receptor Gating by Familial Alzheimer’s Disease-Causing Mutants in Presenilins | Familial Alzheimer’s disease (FAD)-causing mutant presenilins (PS) interact with inositol 1,4,5-trisphosphate (IP3) receptor (IP3R) Ca2+ release channels resulting in enhanced IP3R channel gating in an amyloid beta (Aβ) production-independent manner. This gain-of-function enhancement of IP3R activity is considered to be the main reason behind the upregulation of intracellular Ca2+ signaling in the presence of optimal and suboptimal stimuli and spontaneous Ca2+ signals observed in cells expressing mutant PS. In this paper, we employed computational modeling of single IP3R channel activity records obtained under optimal Ca2+ and multiple IP3 concentrations to gain deeper insights into the enhancement of IP3R function. We found that in addition to the high occupancy of the high-activity (H) mode and the low occupancy of the low-activity (L) mode, IP3R in FAD-causing mutant PS-expressing cells exhibits significantly longer mean life-time for the H mode and shorter life-time for the L mode, leading to shorter mean close-time and hence high open probability of the channel in comparison to IP3R in cells expressing wild-type PS. The model is then used to extrapolate the behavior of the channel to a wide range of IP3 and Ca2+ concentrations and quantify the sensitivity of IP3R to its two ligands. We show that the gain-of-function enhancement is sensitive to both IP3 and Ca2+ and that very small amount of IP3 is required to stimulate IP3R channels in the presence of FAD-causing mutant PS to the same level of activity as channels in control cells stimulated by significantly higher IP3 concentrations. We further demonstrate with simulations that the relatively longer time spent by IP3R in the H mode leads to the observed higher frequency of local Ca2+ signals, which can account for the more frequent global Ca2+ signals observed, while the enhanced activity of the channel at extremely low ligand concentrations will lead to spontaneous Ca2+ signals in cells expressing FAD-causing mutant PS.
| Aberrant Ca2+ signaling caused by IP3R gating dysregulation is implicated in many neurodegenerative diseases such as Alzheimer’s, Huntington’s, Spinocerebellar ataxias, and endoplasmic reticulum stress-induced brain damage. Thus understanding IP3R dysfunction is important for the etiology of these diseases. It was previously shown that FAD-causing mutant PS interacts with the IP3R, leading to its gain-of-function enhancement in optimal Ca2+ and sub-saturating IP3 concentrations. Here, we use data-driven modeling to provide deeper insights into the upregulation of IP3R gating in a wide range of ligand concentrations and quantify the sensitivity of the channel to its ligands in the presence of mutant PS. Our simulations demonstrate that these changes can alter the statistics of local Ca2+ events and we speculate that they lead to Ca2+ signaling dysregulations at the whole cell level observed in FAD cells. These models will provide the foundation for future data-driven computational framework for local and global Ca2+ signals that will be used to judiciously isolate the primary pathways causing Ca2+ dysregulation in FAD from those that are downstream, and to study the effects of upregulation of IP3R activity on cell function.
| Alzheimer’s disease (AD) is a fatal neurodegenerative disease that leads to cognitive, memory, and behavioral impairments followed by progressive cell death. The symptoms of AD include the extracellular deposition of amyloid β (Aβ) plaques and intracellular neurofibrillary tangles—aggregates of microtubule-associated protein τ [1]. According to the amyloid hypothesis, the accumulation of Aβ oligomers or plaques due to the imbalance between synthesis and clearance of Aβ is the driving force for AD pathogenesis [2]. However, whether τ and Aβ aggregates are the causes or symptoms of AD remains a matter of debate [3].
Aβ is produced by the cleavage of amyloid precursor protein (APP), an integral membrane protein. APP is cleaved sequentially by β and γ-secretases to generate Aβ monomers that are released to the extracellular space and form oligomers. The γ-secretase complex contains four different proteins including presenilins (PS) that are synthesized and localized in the ER [1]. Mutations in PS alter the APP processing, thus leading to Aβ oligomarization either through higher production or relatively higher proportion of amyloidogenic Aβ types [4]. It is well established that mutations in PS and APP are the main causes of Familial AD (FAD) [5]. What is not clear is how PS mutations and Aβ accumulation lead to the impairment of brain function and neurodegeneration. The Ca2+ hypothesis of AD, which is based on the enhanced intracellular Ca2+ signaling during AD, accounts for early memory loss and subsequent cell death [6, 7, 8].
There is strong evidence in favor of intracellular Ca2+ signal exaggeration by FAD-causing PS mutations as an early phenotype that could contribute to the pathogenesis of the disease [9]. The exaggerated cytosolic Ca2+ signals are ascribed mainly to the enhanced release of Ca2+ from intracellular endoplasmic reticulum (ER) store due to overloading of ER lumen by up-regulated sarco-endoplasmic reticulum Ca2+-ATPase (SERCA) pump [10]; disruption of ER-membrane Ca2+ leak channels [11]; or enhanced gating of IP3R [12, 13, 14, 15], the ubiquitous ER-localized Ca2+ release channel crucial for the generation and modulation of intracellular Ca2+ signals in animal cells [16]. Single channel studies in multiple cell lines show that the sensitivity of IP3R to its agonist IP3 increases significantly in the presence of FAD-causing mutant PS [14, 15], leading to a several fold increase in the open probability (Po) of IP3R channel in subsaturating IP3. These studies were performed in the absence of Aβ, suggesting that the modulation of IP3R is a major mechanism for intracellular Ca2+ signal dysregulation in cells expressing FAD-causing mutant PS. Furthermore, altered IP3R-mediated Ca2+ release has been suggested as the fundamental defect and highly predictive diagnostic feature in AD [17]. Suppression of IP3R-mediated Ca2+ signaling was recently shown to restore normal cell function and memory tasks in M146V (FAD-causing PS mutation) knock-in [18] and triple-transgenic [19] mice models of FAD [20].
IP3R gates in three distinct modes: an “L” mode with very low Po, in which brief openings are separated by quiescent periods with long mean closed channel durations (τc); an “I” mode with intermediate Po, in which the channel opens and closes rapidly with short τc and mean open channel durations (τo); and an “H” mode with high Po, in which the channel gates in bursts [21]. All three modes are observed under all conditions in which the channel gates, and the channel spontaneously switches among all three modes even under constant ligand conditions. The Po of the channel remains remarkably consistent in each gating mode in all ligand conditions so the ligand dependencies of overall channel Po, τo and τc come from the ligand dependencies of the relative prevalence (normalized occupancy) πM of the gating modes (M can be L, I, and H) [21].
Due to the significant role of IP3R-mediated Ca2+ signaling dysregulation in AD, a comprehensive understanding of the IP3R function is important for both the etiology of the disease and designing effective therapeutic reagents. It was discovered that IP3R channels in cells expressing FAD-causing mutant PS exhibit relatively higher πH and lower πL in comparison to IP3R in cells expressing wild-type PS [15, 14] in non-optimal ligand conditions. πI, on the other hand, remains largely the same. The switch in the prevalence of H and L modes causes the increase in Po of IP3R in the presence of FAD-causing mutant PS.
In this paper, we employ a data-driven modeling approach to gain further insights into the gating behavior of IP3R in the presence of wild-type and FAD-causing PS. We focus on the channel gating behaviors of endogenous IP3R in the presence of human wild-type (PS1-WT) and FAD-causing mutant (PS1-M146L) PS expressed in the Sf9 cells, an insect cell line derived from the moth Spodoptera frugiperda. Other FAD-causing mutant PS1 (PS1-L116P, PS1-G384A) and PS2 (PS2-N141I) have similar effects on IP3R channel gating as PS1-M146L. On the other hand, non-FAD-associated mutant PS1 (PS1-L113P and PS1-G183V), wild-type PS2 and EVER1 (an irrelevant ER transmembrane protein) have little to no effects on IP3R channel gating, like PS1-WT [15, 14]. Therefore, the conclusions from studying IP3R channel in the presence of PS1-WT (IP3RPS1WT) and IP3R channel in the presence of PS1-M146L (IP3RPS1M146L) can be generalized to other FAD-causing mutations as well. We used the data-driven kinetic model developed to describe all observed gating behaviors of the endogenous IP3R channels in Sf9 cells: channel Po, τo and τc distributions in various steady Ca2+ and IP3 concentrations (𝓒 and 𝓘, respectively); modal gating behaviors in various steady 𝓒 and 𝓘; and kinetic response of IP3R channels to abrupt changes in 𝓘 and/or 𝓒 [22] as the starting point of our approach. By modifying a minimum number of model parameters, the data-driven model was applied to fit observed gating behaviors: channel Po, τo, τc, and modal prevalence, in optimal (1 μM) 𝓒 and sub-saturating (100 nM) 𝓘, of the Sf9 IP3RPS1M146L and IP3RPS1WT (used as control) [15] as well as Po, τo, and τc of IP3RPS1WT in 33 nM and 10μM 𝓘 at 𝓒 = 1μM. In addition to elucidating the kinetics and factors contributing to the gain-of-function enhancement of IP3R activity [14, 15], we extrapolate, using our modified model, the gating behavior of IP3RPS1WT and IP3RPS1M146L for a wide range of 𝓒 and 𝓘. We also quantify and compare the ligand sensitivities of IP3RPS1M146L and IP3RPS1WT. Simulations of local Ca2+-release events based on the results of the data-driven model demonstrate that the gain-of-function enhancement of IP3R activity leads to larger, longer, and more frequent local Ca2+ releases events in cells expressing FAD-causing PS mutants. The models derived here will provide the foundation for developing future data-driven computational framework for global intracellular Ca2+ signals that will be used to judiciously isolate the primary factors causing Ca2+ signaling dysregulation in FAD from those that are downstream, and to study the effects of upregulation of IP3R activity on cell functions such as ATP production.
The main experimental data used in this paper for fitting the models were previously published elsewhere [15]. Basic experimental data (Po, τo, and τc) at 𝓘 = 33nM, and 10μM for both IP3RPS1WT and IP3RPS1M146L [14] were also used to generate our model. The full details of experimental methods are given in [15] and summarized below.
Two Sf9 cell lines expressing recombinant PS1-WT and PS1-M146L, respectively, were generated and maintained as described in [14].
Nuclei isolated from transfected Sf9 cells [23, 14] were used for nuclear patch clamp experiments in on-nucleus configuration at room temperature [24]. All experimental solutions contained 140 mM KCl and 10 mM HEPES (pH 7.3). Bath solution contained 90 nM free Ca2+ (buffered by 0.5 mM BAPTA (1,2-bis(2-aminophenoxy)ethane-N,N,N’,N’-tetraacetic acid). Pipette solution contained 1 μM free Ca2+ (buffered by 0.5 mM 5, 5′-dibromo-BAPTA), 0.5 mM Na2ATP and sub-saturating 100 nM IP3.
Segments of current records exhibiting current levels for a single IP3R channel were idealized with QuB software (University of Buffalo) using SKM algorithm [25, 26]. The idealized current traces were further analyzed as described in [21] to characterized the modal gating behaviors of IP3R channels. Short closing events, presumably caused by ligand-independent transitions [27], were removed by burst analysis. Gating modes were assigned according to durations of channel burst (Tb) and burst-terminating gaps (Tg) [21], using a critical Tb of 100 ms and a critical Tg of 200 ms to detect modal transitions.
We fit the twelve-state model previously developed for IP3R in Sf9 cells [22] (Fig 1) to the channel gating data of Sf9 IP3RPS1WT and IP3RPS1M146L at 𝓒 = 1μM and 𝓘 = 100 nM [15], following the procedure in [22, 28, 29], which we describe below. Although the scheme in Fig 1 seems to suggest simultaneous binding/unbinding of multiple ligands as the channel goes from one state to another (for instance C 00L to C 20L, or C24H to C 20L), in reality, there is no such direct transition in our model. Each one of such transitions actually involves one or more intermediate states that are not explicitly shown in order to keep the scheme simple. As discussed in detail in [22], the simplifying approximations were made by considering the fact that the intermediate states have relatively low occupancy and therefore can be aggregated into the main states. The rates for the composite transitions between those main states explicitly shown in the scheme were actually derived with the intermediate low-occupancy states carefully taken into consideration.
Representative time-series traces of the gating behavior of IP3RPS1WT and IP3RPS1M146L are shown in Fig 2A. Occupancy parameters for the twelve states in the model obtained by fitting the Po (Figs 2B, 3A) and prevalence (Fig 2C) data from IP3RPS1WT and IP3RPS1M146L are given in Table 1. Notice that some of the parameters for IP3RPS1WT are different from those in [22] because IP3RPS1WT behaves somewhat differently from the IP3R channel in wild type untransfected Sf9 cells (IP3RnoPS1), despite the general similarity in the gating of the two (Fig 3A–3C, triangles for IP3RnoPS1, squares and solid lines for IP3RPS1WT), especially Po at 𝓘 = 100nM and 400 nM < 𝓒 ≤ 1μM. It is remarkable that the occupancy of only 4 states changes in the presence of PS1-M146L as compared to PS1-WT. The four states are K O 24 H , K O 24 I , K O 14 I, and K C 32 L whose occupancies change by a factor of 4.587, 1.284, 1.718, and 0.018 respectively. Thus IP3RPS1M146L spends relatively more time in the states K O 24 H , K O 24 I, and K O 14 I and less time in K C 32 L as compared to IP3RPS1WT. This is consistent with the prevalence data where there is a significant increase in πH (0.345 vs 0.836) at the cost of πL (0.515 vs 0.059) in IP3RPS1M146L as compared to IP3RPS1WT. πI on the other hand does not change significantly (0.14 vs 0.104) (Fig 2C). Thus the increase in Po is mainly due to the significantly less time spent by IP3RPS1M146L in C 32 L and more time spent in O 24 H (Fig 2B) as compared to IP3RPS1WT.
To derive the probability flux parameters used in the transition rates between different gating states (Table 2), we fit the model to idealized current traces recording openings and closings of IP3RPS1WT and IP3RPS1M146L by minimizing the likelihood score Eq (6) of the data. The flux parameters from the fits are given in Table 3. Only two of the eighteen flux parameters of IP3RPS1M146L are different from IP3RPS1WT (shown in bold). These two parameters are involved in the C 04 I ↔ O 14 I and C 24 H ↔ O 24 H transitions and are higher for IP3RPS1M146L as compared to IP3RPS1WT.
The mean gating properties of the channel as a function of 𝓒 at different 𝓘 values are shown in Fig 3. Theoretical values of the Po of IP3RPS1WT (Fig 3A, solid lines) and IP3RPS1M146L (Fig 3A, dashed lines) were generated by using Eq (2) and the occupancy parameters in Table 1. The Po of IP3RPS1M146L at 𝓘 = 33nM (turquoise dashed line and turquoise circle) and 100nM (blue dashed line and blue circle) are significantly larger than Po of IP3RPS1WT at 𝓘 = 100nM (solid blue line and blue square). In fact, the Po of IP3RPS1M146L at 𝓘 = 33nM is comparable to Po of IP3RPS1WT (magenta square and solid line) or IP3RnoPS1 (magenta triangles from [23]) at saturating 𝓘 = 10μM. The Po data for IP3RnoPS1 at saturating 𝓘 = 10μM is shown for comparison to emphasize that IP3RPS1M146L is already maximally activated at a significantly lower 𝓘 of 100nM. This clearly indicates that IP3R in the presence of FAD-causing mutant PS is highly sensitized to activation by IP3, and is maximally activated at significantly lower 𝓘 than IP3RPS1WT or IP3RnoPS1. For 𝓒 > 800nM, Po of IP3RPS1M146L at 𝓘 = 10μM (magenta dashed line and magenta circle) is higher than that of IP3RnoPS1 or IP3RPS1WT at the corresponding 𝓘 and 𝓒, mostly due to the higher saturating Po of IP3RPS1M146L (0.86 ± 0.03) as compared to those of IP3RnoPS1 (0.72 ± 0.03) [23] and IP3RPS1WT (0.72 ± 0.05) [15]. A close examination of τc (Fig 3B) and τo (Fig 3C) reveals that the increase in Po of IP3RPS1M146L is mostly due to the substantial shortening of τc with relatively modest increase in τo. Theoretical values of τo and τc were calculated using Eq (7) and the parameters in Tables 1 and 3. Gating properties (Po, τo, and τc) of IP3RnoPS1 in 𝓘 = 100nM and 10μM observed in [23] (triangles in Fig 3A–3C) are similar to theoretical values calculated for IP3RPS1WT, since IP3RPS1WT gating is generally similar to that of IP3RnoPS1, as observed in [14], albeit with some noticeable differences (Fig 3A–3C). The open triangles in Fig 3A–3C representing data from [23] of IP3RnoPS1 in wild type Sf9 cells is shown here to demonstrate that the model for IP3RPS1WT can replicate reasonably well the Po, τo, and τc of IP3R channel gating in the absence of PS1.
Next, we calculated with our modified data-driven model the prevalence of the three gating modes using Eq (3) and occupancy parameters given in Table 1 (Fig 3D and 3E). Both experimental (symbols) and theoretical (lines) results show a significant increase in πH (green), and decrease in πL (red) for IP3RPS1M146L (dashed lines and circles) as compared to IP3RPS1WT (solid lines and squares) (Fig 3D) at 𝓒 = 1μM and 𝓘 = 100 nM. πI (orange), on the other hand, remains largely unchanged. Comparison with the prevalence data from IP3R in untransfected Sf9 cells at saturating 𝓘 = 10μM (triangles) from [21] confirms the saturating activation of IP3RPS1M146L at relatively low 𝓘. The mean life-times of the three gating modes follow a similar trend as seen in their prevalences. τH is longer, τL is shorter, while τI remains unchanged for IP3RPS1M146L as compared to IP3RPS1WT (Fig 3E). The theoretical modal mean life-times were calculated from our modified model by using Eq (9) and parameters in Tables 1 and 3.
The open and closed dwell-time distributions were calculated from the model as described in the Dwell-Time Distributions section Eqs (14) and (15) using occupancy and flux parameters in Tables 1 and 3 respectively. As shown in Fig 4, the model (red lines) fits the observed dwell-time distributions (gray bars) very well. Consistent with the τo observed (Fig 3C), there is a minor right-shift in the open dwell-time distribution of IP3RPS1M146L (Fig 4C) as compared to IP3RPS1WT (Fig 4A). Thus, PS1-M146L does not have significant effect on τo of the IP3R channel. The close dwell-time distribution of IP3RPS1M146L (Fig 4D), on the other hand, shows significant shift to the left when compared to IP3RPS1WT (Fig 4B), leading to the shorter τc and therefore the higher Po observed.
The Dwell-Time Distributions section also describes the derivation of the open and closed dwell-time distributions in H and I modes from the model. The dwell-time distributions from the model for 𝓒 = 1μM and 𝓘 = 100nM in the H (Fig 5A–5D) and I (Fig 5E–5H) modes calculated by using Eqs (18) and (19) and parameters in Tables 1 and 3 are given by red lines. The experimental data are presented by the gray bars for comparison. A close inspection of the modal open and closed dwell-time distributions of IP3RPS1WT and IP3RPS1M146L provides useful insight into the modal behavior of the channel. In line with the over-all open dwell-time distribution, the open dwell-time distributions in the H (Fig 5A, 5C) and I (Fig 5E, 5F) modes do not change significantly. The closed dwell-time distributions in the two modes in IP3RPS1M146L (Fig 5D, 5H) on the other hand, shift significantly to the left as compared to IP3RPS1WT (Fig 5B, 5G). Furthermore, the shift in the closed dwell-time distribution in the H mode is more significant (Fig 5B, 5D). This suggests that the relatively shorter time spent by IP3RPS1M146L in the H mode’s closed state plays a major role in the shortening of τc and hence enhancement of Po of the channel in the presence of FAD-causing mutation as compared to wild-type PS1.
To gain further insights and quantify the extent of IP3R sensitization due to PS1-M146L, we extrapolate the Po, τc, and the mean modal properties of IP3R at different values of 𝓒 and 𝓘 from our model using parameters tabulated in Tables 1 and 3. Fig 6A shows Po of IP3RPS1WT (black solid lines) and IP3RPS1M146L (blue dashed lines) as a function of 𝓒 at different 𝓘, calculated using Eq (2). Even at 𝓘 = 8nM, Po of IP3RPS1M146L is already higher than that of IP3RPS1WT at 𝓘 = 100nM (thick solid black line) for all 𝓒. Thus, whereas IP3RPS1WT in 8 nM IP3 are minimally active (Po ∼ 0.005) in resting 𝓒 (∼ 70 nM), IP3RPS1M146L under the same ligand conditions can have sufficient activity (Po > 0.04) to initiate intracellular Ca2+ signals.
The τc of the channels from simulation Eq (7) in Fig 6B correlate well with the Po values (notice that at fixed 𝓘, Po decreases as τc increases and vice versa). This close correlation between the Po and τc is the consequence of the lack of dependence of the τo on 𝓘 (see Eq (7) and Fig 3C). Thus for 𝓘 > 8 nM, τc of IP3RPS1M146L (blue dashed lines) is shorter than that of IP3RPS1WT at 𝓘 = 100nM (black solid line) for all physiological 𝓒 values. Whereas the strong Ca2+ activation of IP3RPS1M146L between 𝓒 = 0.01 and 0.5 μM remains the same as 𝓘 is raised from 16 to 100 nM, channel Po for higher 𝓒 does increase as 𝓘 increases from 16 nM to 1 μM, confirming the higher saturating Po of IP3RPS1M146L as described above (Fig 3A) and is in line with observations (Table S1 in [14]).
To quantitatively compare the sensitivity of IP3RPS1WT and IP3RPS1M146L to activation by IP3, we plot in Fig 6C the Po of IP3RPS1M146L as a function of 𝓘 at fixed 𝓒 = 1μM (black line) and the experimentally observed Po of IP3RPS1WT at 𝓒 = 1μM and 𝓘 = 100 nM (black square). The plot reveals that Po of IP3RPS1M146L at 𝓘 = 8 nM already exceeds that of IP3RPS1WT at 𝓘 = 100 nM (see black dotted lines). In contrast, Po of IP3RPS1WT is negligible at 𝓘 = 8nM and 𝓒 = 1μM (Fig 6A, black thin solid line). Correspondingly, τc of IP3RPS1M146L (red line in Fig 6C) becomes shorter than the observed τc of IP3RPS1WT (red square) at 𝓘 = 100nM as 𝓘 is raised beyond 8nM.
In Fig 6D, we show the ratio of simulated Po of IP3RPS1M146L to that of IP3RPS1WT as a function of 𝓘 at 𝓒 = 70nM, 100nM, 250nM, 500nM, 1μM, and 2μM. For 𝓘 < 200nM, IP3RPS1M146L is more than twice as active as IP3RPS1WT for all 𝓘 and 𝓒 values. For physiological resting 𝓒 = 70nM, IP3R channel activity is enhanced by 265% in cells expressing PS1-M146L relative to that in PS1-WT expressing cells. For optimal 𝓒, IP3RPS1M146L exhibits a gain-of-function enhancement by over 100 folds as compared to IP3RPS1WT, with maximum enhancement occurring around 𝓘 = 7–8 nM IP3. The drop in the Po ratio for 𝓘 > 8nM is due to the fact that Po of IP3RPS1M146L peaks much faster than IP3RPS1WT as a function of 𝓘.
At fixed 𝓒 = 1μM, the theoretical πL Eq (3) of IP3RPS1M146L decreases for 2nM < 𝓘 < 100nM and plateaus outside this window (Fig 6E, dashed red line). πH of IP3RPS1M146L (dashed green line) changes in the opposite direction for 2nM < 𝓘 < 100nM. πI of IP3RPS1M146L (dashed orange line), on the other hand, remains largely constant. Around 𝓘 = 8 nM, both the πL and πH curves of IP3RPS1M146L crosses the observed πL (red square) and πH (green square) levels, respectively, of IP3RPS1WT measured at 𝓘 = 100nM and 𝓒 = 1μM. πI of IP3RPS1M146L gets close to but does not exceed that of IP3RPS1WT observed at 𝓘 = 100nM and 𝓒 = 1μM (orange square). This indicates that the increase in Po of IP3R in the presence of PS1-M146L is mainly due the switching of the channel from L to H mode. Interestingly, the 𝓘 value (∼ 10nM) where πL and πH cross each other is almost the same as where Po reaches half (0.41) of its peak value (0.82) (see black line in Fig 6C). This is in line with the observations that mode switching is the major mechanism of ligand regulation of IP3R [21]. Our results confirm that the mechanism of ligand regulation of IP3R is mainly due to the switching of the channel between L and H modes with minimal contributions from I mode [21]. Furthermore, switching of IP3RPS1M146L from L to H mode in comparison to IP3RPS1WT translates into the gain-of-function enhancement of IP3R gating. Fig 6E also shows how theoretical values of the prevalences of the three gating modes for IP3RPS1WT vary with 𝓘 at 𝓒 = 1μM (solid lines).
Theoretical values of the mean life-times of I (orange) and H (green) mode calculated using Eq (9) at 𝓒 = 1μM remain largely unchanged for all values of 𝓘 > 3nM (Fig 6F). In that range of 𝓘, τH of IP3RPS1M146L (dashed green line) is shorter than τH of IP3RPS1WT (solid green line), while τI of IP3RPS1M146L (dashed orange line) remains almost the same as τI of IP3RPS1WT (solid orange line). For 𝓘 < 4nM, τI of IP3RPS1M146L drops below the observed value for IP3RPS1WT at 𝓘 = 100nM. Simulated τL of IP3RPS1M146L (dashed red line), on the other hand, decreases significantly as 𝓘 increases. As 𝓘 increases beyond 15 nM, simulated τL of IP3RPS1M146L becomes shorter than that of IP3RPS1WT measured at 𝓘 = 100nM (red square). Relatively speaking, this does not correlate very closely with the 𝓘 = 8nM where the Po of IP3RPS1M146L crosses the observed Po of IP3RPS1WT (Fig 6A) when compared to the prevalences where this critical value of 𝓘 is 8nM. Nevertheless, the shorter τL will increase the Po of IP3RPS1M146L. Thus the shorter τL and longer τH contribute to its increased sensitivity to IP3 as compared to IP3RPS1WT.
To investigate how the remodeling of single IP3R channel gating kinetics in the presence of mutant PS1-M146L affects the dynamics of IP3R-mediated Ca2+ release events, we simulated such events at a Ca2+-release site consisting of a cluster of ten IP3Rs as described in the Methods Section. 400 s-long records of Ca2+ blips and puffs (Ca2+-release events involving just one, or multiple IP3R channels in the cluster, respectively) from the IP3R cluster were generated, and statistics about these Ca2+-release events were derived as described previously [28]. In line with observations reported in [35, 12], the site produces significantly potentiated puffs in the presence of FAD-causing PS1. As shown in Fig 7, the behavior of puffs and blips arising from a cluster of IP3RPS1M146L (dashed lines with circles) is significantly different from that of a IP3RPS1WT cluster (solid line with squares). Puffs from a IP3RPS1M146L cluster have significantly larger amplitudes (Fig 7A), longer life times (Fig 7B), and take longer to terminate (Fig 7C). Furthermore, puff frequency in cells expressing PS1-M146L is significantly higher than that in PS1-WT-expressing cells (10.33/sec versus 1.83/sec) (Fig 7D). Interestingly, the statistics of Ca2+ blips are very similar in both case. Although the life times of blips in PS1-M146L-expressing cells is slightly longer than that in PS1-WT-expressing cells (Fig 7E), frequencies of the blips are almost the same (10.67/sec versus 9.99/sec) (Fig 7F). Thus the higher number of single-channel events caused by higher sensitivity of IP3RPS1M146L to activation by IP3 translates into triggering more frequent and longer puffs.
Accumulation of Aβ aggregates and intracellular neurofibrillary tangles of τ protein are the main symptoms of AD [1]. However, most drugs focused on restricting Aβ production and accumulation or enhancing its clearance from the brain have yielded disappointing results [36]. This could be due to the fact that these drugs target the late stage features of the disease, i.e. plaque formation—whereas there is poor correlation between Aβ deposits and the progressive memory loss and cognitive decline observed [37]. Compelling evidence suggests that FAD-causing mutant PS disrupt intracellular Ca2+ signaling before Aβ deposition, pointing towards the up-regulated Ca2+ signaling as a proximal event that could be involved in disease pathogenesis. Previously, we showed that FAD-causing mutant PS interact with ER-localized IP3Rs, leading to their gain-of-function enhanced channel activity [14, 15]. However, further analysis is needed to elucidate the modified gating properties leading to the gain-of-function enhancement of IP3R channel activity and to quantify their sensitization due to FAD-causing PS mutations.
Our modeling results show that there is a significant increase in the H mode prevalence, and a corresponding significant reduction in the L mode prevalence of IP3R channels in the presence of PS1-M146L. This change in the prevalence of L and H modes arises mainly from the shorter mean life-time of the L mode and longer mean-life-time of the H mode. On the other hand, both the life-time and prevalence of I mode remain constant. Our model predicts that the Po of IP3RPS1M146L saturates at significantly lower IP3 concentrations and its peak value is higher compared to that of IP3RPS1WT. Furthermore, our model predicts that the channel’s gain-of-function enhancement is sensitive to both IP3 and Ca2+ (Fig 6A). Interestingly, our model predicts that the higher Po of IP3RPS1M146L as compared to IP3RPS1WT is mainly due to the lower occupancy of C 32 L state and higher occupancy of the O 24 H state (with small change in the occupancies of O 24 I and O 14 I). This means that PS1-M146L interacts with IP3R in such a way that the closed configuration of the channel with 3 Ca2+ ions and 2 IP3 bound is less likely and the open configuration in the H mode with 2 Ca2+ ions and 4 IP3 molecules bound more easily attainable. The changes in the occupancies of states O 24 H and C 32 L lead to the increase in the prevalence of H mode and the corresponding decrease in the prevalence of L mode, resulting in the gain-of-function enhancement of IP3R gating in the presence of PS1-M146L.
Our goal in this study was to use all the data at our disposal to model the gating kinetics of IP3RPS1WT and IP3RPS1M146L; and use resulting models to simulate the behavior of IP3RPS1WT and IP3RPS1M146L in a wide range of ligand concentrations to gain a better understanding of how the gain-of-function enhancement of IP3RPS1M146L alters the characteristics of local Ca2+ release events at IP3R clusters.
Because of the considerable technical difficulties in obtaining single-channel current records of IP3R channels in their native membrane milieu by nuclear patch clamp electrophysiology, the most comprehensive set of such single-channel data (including steady-state gating records over a broad range of combinations of cytoplasmic IP3 and Ca2+ [23], long gating records in multiple constant ligand conditions suitable for modal gating analysis [21], records of response kinetics of IP3R channel to rapid changes in cytoplasmic ligand conditions [38]) was obtained in the study of the endogenous IP3R from insect Sf9 cells. This set of data was used to develop the twelve-state kinetic model [22] that provides the basis of the models we develop here to simulate the behavior of the endogenous Sf9 IP3R in the presence of exogenous recombinant human PS1 (WT and mutant). Until a comparable or more comprehensive set of single-channel data for a mammalian IP3R becomes available, our approach is the best that can be achieved to simulate gating behaviors of IP3R interacting with WT and mutant PS1. Although Sf9 IP3R does not interact with human PS1 in its natural environment, the study in [15] showed that endogenous IP3R in human lymphoblasts in presence of PS1-M146L exhibits very similar changes in its gating characteristics (increase in Po and τo with corresponding reduction in τc) and modal gating behavior (rise in πH with simultaneous drop in πL) when compared to IP3R in the presence of PS1-WT. Therefore, we have reason to be confident that simulation results from our modeling effort do reflect the gating behaviors of IP3R naturally interacting with WT and mutant PS1, and that the insights our effort provides about the effects of mutant PS1 on IP3R single-channel gating and local intracellular Ca2+ release events can improve our understanding of the pathophysiology of FAD.
The higher prevalence and longer life-time of H mode and shorter life-time of L mode for IP3RPS1M146L may have important implications for cell physiology. Since the open time in the H mode is significantly higher than that in the L mode, in which the channel has near-zero mean open time, IP3R gating in the H mode will have higher probability of activating neighboring channels through Ca2+-induced-Ca2+ release (CICR), thus leading to higher frequency of Ca2+ puffs (Fig 7D). The higher sensitivity of IP3RPS1M146L to activation by Ca2+ and IP3 also increases the number of open channels in an IP3R channel cluster during a puff, leading to bigger (Fig 7A) and longer (Fig 7B) puffs, which in turn will cause more frequent global Ca2+ waves and oscillations at the cellular level, in line with observations [35, 12]. The higher prevalence of IP3R channel being in the H mode resulting in the higher probability of inducing global Ca2+ events can account for the higher frequency of Ca2+ oscillations in B lymphoblasts from FAD patients and DT40 cells expressing FAD-causing mutant PS [14, 15].
Our model reveals that even at resting level of 𝓘 (8 nM), IP3RPS1M146L exhibits significant Po of 0.35 (40% of the Po when the channel is in saturating 𝓘) at 𝓒 = 1μM whereas the Po of IP3RPS1WT in the same ligand condition is negligible. This will lead to stronger CICR among channels in the same cluster and between IP3R channel clusters, thereby generating more global spontaneous Ca2+ oscillations in cells expressing FAD-causing mutant PS as observed experimentally [14, 15].
To conclude, our study provides insights into the gating modulation of IP3R that leads to the gain-of-function enhancement due to FAD-causing mutations in PS. Furthermore, significant activity exhibited by IP3R at resting IP3 concentration in cells expressing FAD-causing mutant can explain the spontaneous global Ca2+ signals observed in those cells. The models developed here for single-channel IP3R channel gating and local Ca2+-release events can provide part of the foundation for building whole-cell models to judiciously separate the key pathways leading to the global Ca2+ signaling dysregulation in AD from those that are by-products due to the CICR nature of Ca2+ signaling. For example, what are the relative contributions of gain-of-function enhancement and over-expression of ryanodine receptors [39, 20] to Ca2+ signaling dysregulation in AD? Does the down-regulation of Ca2+ buffers such as calbindin [40, 41] and higher resting Ca2+ concentration [42] play a role (through CICR mechanism) in the exaggerated Ca2+ signals? What are the conditions or factors that would reverse the exaggerated Ca2+ signaling back to normal state? These and many other interesting questions are the focus of our future research and the models developed here will play a key role in addressing them.
|
10.1371/journal.pcbi.1005797 | Unified thalamic model generates multiple distinct oscillations with state-dependent entrainment by stimulation | The thalamus plays a critical role in the genesis of thalamocortical oscillations, yet the underlying mechanisms remain elusive. To understand whether the isolated thalamus can generate multiple distinct oscillations, we developed a biophysical thalamic model to test the hypothesis that generation of and transition between distinct thalamic oscillations can be explained as a function of neuromodulation by acetylcholine (ACh) and norepinephrine (NE) and afferent synaptic excitation. Indeed, the model exhibited four distinct thalamic rhythms (delta, sleep spindle, alpha and gamma oscillations) that span the physiological states corresponding to different arousal levels from deep sleep to focused attention. Our simulation results indicate that generation of these distinct thalamic oscillations is a result of both intrinsic oscillatory cellular properties and specific network connectivity patterns. We then systematically varied the ACh/NE and input levels to generate a complete map of the different oscillatory states and their transitions. Lastly, we applied periodic stimulation to the thalamic network and found that entrainment of thalamic oscillations is highly state-dependent. Our results support the hypothesis that ACh/NE modulation and afferent excitation define thalamic oscillatory states and their response to brain stimulation. Our model proposes a broader and more central role of the thalamus in the genesis of multiple distinct thalamo-cortical rhythms than previously assumed.
| Computational modeling has served as an important tool to understand the cellular and circuit mechanisms of thalamocortical oscillations. However, most of the existing thalamic models focus on only one particular oscillatory pattern such as alpha or spindle oscillations. Thus, it remains unclear whether the same thalamic circuitry on its own could generate all major oscillatory patterns and if so what mechanisms underlie the transition among these distinct states. Here we present a unified model of the thalamus that is capable of independently generating multiple distinct oscillations corresponding to different physiological conditions. We then mapped out the different thalamic oscillations by varying the ACh/NE modulatory level and input level systematically. Our simulation results offer a mechanistic understanding of thalamic oscillations and support the long standing notion of a thalamic “pacemaker”. It also suggests that pathological oscillations associated with neurological and psychiatric disorders may stem from malfunction of the thalamic circuitry.
| The thalamocortical network plays a central role in cerebral rhythmic oscillations [1–4] and abnormal thalamocortical rhythms have been associated with disorders such as depression, schizophrenia and Alzheimer’s disease [5–7]. Understanding the cellular and circuit mechanisms of thalamocortical oscillations thus constitutes a crucial first step to comprehend the network impairments underlying neurological and psychiatric disorders. However, the mechanisms by which the thalamocortical network generates distinct states of oscillatory patterns remain highly debated [8–11]. One important question is whether the thalamus, originally believed to be the “pacemaker” of thalamocortical oscillations [7, 8, 12], is indeed able to independently generate multiple distinct brain rhythms or whether the thalamus requires interaction with the cortex [13–16]. Answering this question not only provides the basis for a mechanistic understanding of brain oscillations, but also will provide important insights in the design of effective mechanism-based brain stimulation techniques that specifically target abnormal thalamocortical dynamics.
Experimental evidence suggests that the thalamus is capable of independently generating multiple, distinct oscillatory states. In the cat lateral geniculate nucleus (LGN), in vitro and in vivo studies have identified a subset of thalamocortical cells (TCs) that generate high-threshold bursting at theta (θ) and alpha (α) frequency bands and thus may mediate the cellular mechanism of both θ and α oscillations [12, 17, 18]. Coupled with gap junctions [12, 17], high-threshold bursting TC cells provide synchronized excitatory inputs to local interneurons and reticular cells that entrain the majority of TC cells (i.e., non-high threshold bursting TC cells) into the α rhythm via feed-forward and feedback inhibition [18]. Besides θ/α oscillations, the thalamus is also critically involved in the genesis of the slow delta rhythm and spindle oscillations that appear at different stages of non-rapid eye movement (NREM) sleep [2, 19, 20]. Moreover, the thalamus is able to produce high frequency oscillations in both β and γ bands (20–60 Hz) in the neonatal rat whisker sensory system [21], during attentional processing in cats [22] and during cognitive tasks in humans [23]. Consistently, experimental data indicated that fast rhythms (30–40 Hz) could be synchronized with an intrathalamic mechanism [24]. It is not known whether the same neural substrate and circuitry for θ and α oscillations could also mediate other oscillatory patterns and what controls the transition among these oscillatory states.
Thalamic processing is subject to the action of modulatory neurotransmitters including acetylcholine (ACh), norepinephrine (NE), serotonin (5-HT), histamine (HA) and dopamine (DA) [25]. Of these neurotransmitters, cholinergic and noradrenergic modulation plays the key and best understood role in shaping the oscillatory state of the thalamocortical network [26]. Consistently, the α rhythm is both induced by muscarinic cholinergic receptor activation in slices of cat LGN [17] and supported by cholinergic innervation in vivo [18]. Fast rhythmic activities in the β/γ frequency band (20–60 Hz) are observed in cat thalamocortical cells [11] and promoted by cholinergic projection from the brainstem [27]. Besides, slow δ oscillations are believed to be mediated by hyperpolarization of TC neurons with diminished activation of both the cholinergic and noradrenergic systems [28, 29]. In addition to neuromodulation, the specific type of thalamic oscillation is also dependent on the level of afferent excitation. For example, the amplitude of the α oscillation is maximal when the eyes are closed and therefore input form the retina is low [30]. Hence, generation and transition of distinct thalamic oscillations depend critically on neuromodulation and afferent excitation, yet a unified model has been lacking so far.
To close this gap, we developed a biophysical, conductance-based model of the thalamic network constrained by extensive experimental data to test the hypothesis that generation and transition of distinct thalamic oscillations are functions of both ACh/NE neuromodulation and afferent excitation under various physiological conditions. By varying only these two model parameters, ACh/NE neuromodulation and afferent excitation, we demonstrated that the thalamic network is capable of generating multiple distinct oscillatory states (δ, spindle, α/θ and γ/β oscillations) in absence of cortical input. We elucidated the cellular and circuit mechanisms for each oscillatory state by manipulating the network connectivity and key model parameters. Simulation results suggest that generation of distinct thalamic oscillations is a result of both intrinsic oscillatory cellular properties and specific network connectivity patterns. The manifestation of multiple distinct oscillations in one unified biophysical thalamic model enabled us to examine the impact of rhythmic stimulation on thalamic network dynamics. By applying periodic stimulation to the thalamic model during three major oscillatory states (δ, α and γ oscillations), we observed that entrainment of thalamic oscillations is highly state-dependent in that the same stimulation induced much stronger and more prominent entrainment during γ oscillations than δ and α oscillations due to the different oscillatory mechanisms. Our findings emphasize the importance of considering the rich role of endogenous oscillations in thalamus for the study of thalamo-cortical rhythms and highlight the need to consider the network state when modulating brain oscillations with periodic stimulation waveforms.
We developed a biophysical conductance-based thalamic network model containing both the lateral geniculate nucleus (LGN) and the reticular nucleus (TRN) of the thalamus (Fig 1A; see Methods). We first constructed single cell models of high-threshold bursting TC cells (HTCs), relay-mode TC cells (RTCs), local interneurons (INs) and thalamic reticular cells (REs) that replicated experimentally observed firing patterns both in control condition and in case of modulation by ACh/NE (S1 and S2 Figs; S1 Text). By connecting the four types of neurons into a thalamic network (Fig 1A), we tested the central hypothesis that generation and transition of distinct thalamic oscillations are functions of both ACh/NE modulation and afferent excitation (Fig 1B).
By varying the potassium leak conductance modulated by ACh/NE and the maximal input conductance corresponding to different levels of afferent excitation in all four types of thalamic neurons (Table 1), we were able to generate four distinct oscillatory states (δ, spindle, α, and γ oscillations) in the thalamic network. In the model, the increasing level of ACh/NE during the transition from deep sleep to wakefulness corresponded to lower potassium leak conductance in HTC, RTC and RE cells, but higher potassium leak conductance in INs (Table 1; also see Methods). All four types of thalamic neurons received random Poisson distributed inputs mediated by AMPA receptors and the maximal input conductance was a fixed constant value associated with the AMPA synaptic channels.
To quantify the network activity during the four oscillatory states, we calculated the average firing rates and synchronization index of four different groups of thalamic neurons (HTC, RTC, IN & RE, Fig 3A and 3B) along with the correlation between different neuronal populations (Fig 3C and 3D). Overall, the average firing rates of RTC cells were lower than HTC cells for all oscillatory states, because RTC cells were less excitable than HTC cells due to a much smaller high-threshold T-type Ca2+ current (ICa/HT, S1 Table). In addition, RTC cells received inhibition from both INs and REs, while HTC cells received inhibition from REs only (Fig 1A). The firing rates of both HTC and RTC cells increased from δ to spindle oscillations and decreased during α oscillations followed by a large increase during γ oscillations (Fig 3A). This is because the bursting frequency of TC cells increased from δ to spindle oscillations (δ: 3.7 Hz; spindle: 7.9 Hz) and during the transition from spindle to α oscillations, the oscillation frequency was similar (spindle: 7.9 Hz; α: 9.2 Hz), but RTC cells switched from bursting to single spiking while the number of spikes per burst reduced in HTC cells (spindle: 3.2; α: 2.1). The TC firing rates were highest during γ oscillations among all four oscillatory states since TC cells received strong afferent drive (Table 1). During the transition from γ to α oscillations, the firing rates of INs increased (γ: 9.8 Hz; α: 16.6 Hz; Fig 3A) while the firing rates of REs reduced (γ: 26.6 Hz; α: 8.2 Hz; Fig 3A). This is consistent with the experimental observation that LGN interneurons in cats exhibited an increase in firing rate during α oscillations (compared to non-α state; presumably γ state in our model), whereas TRN neurons showed a decrease in firing rate [18]. Our simulations suggest that such differential effects resulted from the following three factors: (1) INs were less excitable than RE cells in the high ACh/NE state owing to a larger potassium leak conductance (Table 1); (2) INs received inputs from HTCs only while REs received inputs from both HTC and RTC cells (Fig 1A). Consequently, the large increase in RTC firing during γ oscillations led to substantial increase in RE spikes; (3) HTC cells switched from HTBs during α oscillations to tonic spiking during γ oscillations, which reduced the effectiveness of excitatory drive on INs, as mentioned earlier. Combined with STD, INs switched from strong bursting to irregular mix of bursting and tonic spiking (compare Fig 2C1 with Fig 2D1, upper middle). As a result, the IN firing rate reduced during γ oscillations compared with α oscillations.
The generation of distinct oscillations in the thalamic network depends on synchronization of different neuronal populations. To evaluate the degree of neuronal synchrony in the thalamic network, we calculated the synchronization index (SI) of the four neuronal populations during different oscillatory states (Fig 3B). The SI of both HTC and RTC cells maintained at relatively high level during δ, spindle and α oscillations (> 0.87; Fig 3B). During γ oscillations, the SI of RTC cells substantially decreased to 0.09, while that of HTC cells only slightly reduced to 0.76 (Fig 3B) because of gap junctions. The SI of INs decreased moderately from δ to spindle/α oscillations and reduced greatly during γ oscillations (δ: 0.98; spindle: 0.76; α: 0.78; γ: 0.03; Fig 3B). Similarly, the SI of REs showed a decreasing trend from δ to γ oscillations, but the reduction during γ oscillations was smaller (δ: 0.97; spindle: 0.90; α: 0.69; γ: 0.23; Fig 3B) due to the inter-RE gap junctions and inhibition. These results indicate that the thalamic network had the highest level of synchrony during δ oscillations followed by spindle and α oscillations and the synchronization level was the lowest during γ oscillations.
We next computed the cross-correlation between different groups of neuronal populations during the four oscillatory states (Fig 3C). First, the correlation was consistently highest during δ oscillations and lowest during γ oscillations for all six neuronal population pairs. Second, the correlation during spindle oscillations was either comparable to (e.g., HTC-RTC) or moderately lower than (e.g., RTC-RE) α oscillations (Fig 3C). Lastly, the correlation between HTC cells and RE neurons during γ oscillations was substantially higher than other population pairs (Fig 3C), consistent with higher level of synchrony of these two neuronal assembles during γ oscillations (Fig 3B). On average, the neuronal correlation was largest during δ oscillations followed by α and spindle oscillations and the correlation was lowest during γ oscillations (Fig 3D).
Overall, as the ACh/NE modulation level and synaptic input increased, the thalamic network switched from low frequency oscillations (δ oscillations) to higher frequency oscillations (spindle and α oscillations), and to fast frequency oscillations (γ oscillations). Correspondingly, the membrane potentials of TC neurons gradually depolarized and the firing patterns of HTC cells switched from LTBs to rebound LTBs to HTBs and to tonic spiking, while those of RTC neurons changed from LTBs to rebound LTBs and to tonic spiking. Hence, with minimal change of parameters (the potassium leak conductance and the synaptic input strength; Table 1), the thalamic network was able to generate multiple distinct and stable oscillatory states that appear under different behavioral and cognitive conditions [2, 18, 19, 22, 23]. Next, we dissected the cellular and circuit mechanisms for each oscillatory state by manipulating the major network connectivity and varying key model parameters.
To understand how δ oscillations were generated in the thalamic network, we first removed the gap junction connections among HTC cells. Removal of gap junctions led to large variation in HTC burst timing as the LTB frequencies of individual cells differed from each other because of intrinsic heterogeneity (i.e., different leak conductance) and external noise input (Fig 4A, top). It also reduced RTC synchrony through the HTC-RTC gap junctions (compare Fig 4A with Fig 2A2, lower middle). Consequently, after a few partially synchronized δ cycles, both HTC and RTC cells broke into two subpopulations separated by the IN and RE inhibition. In one cycle, a majority of HTC cells burst with a small percentage of RTC cells while in the next cycle, a majority of RTC cells burst with a small percentage of HTC cells (Fig 4A). As a result, the network oscillation frequency doubled to about 6.7 Hz (S11 Fig). Interestingly, if the HTC-RTC gap junctions were additionally removed, the frequency doubling effect was not observed (Fig 4B) and the network oscillation frequency just slightly increased to about 4 Hz (S11 Fig). Although the δ rhythm was maintained without any TC gap junctions, the degree of network synchrony was substantially reduced compared with the control condition (compare Fig 4B with Fig 2A2).
To examine the role of synaptic inhibition in thalamic network synchronization, we first blocked the HTC→IN projections to remove the feedforward IN inhibition on RTC cells (gap junctions intact). We found that RTC cells remained well synchronized (Fig 4C). We next blocked the TC→RE connections to eliminate the RE feedback inhibition on TC cells. Similarly, RTC bursting was still well phase-locked to the δ rhythm (Fig 4D). If, however, both the HTC→IN and TC→RE connections were blocked, RTC bursts became largely desynchronized, except the subset of RTC cells that formed gap junctions with HTC cells (Fig 4E). This suggests that either IN or RE inhibition is required for the synchronization of RTC cells. Lastly, when the HTC-RTC gap junctions were additionally removed (besides IN and RE inhibition), RTC bursting became completely desynchronized (Fig 4F). Therefore, thalamic δ oscillations are generated intrinsically by TC cells and are synchronized by both gap junctions and synaptic inhibition.
In the thalamic network, spindle oscillations were triggered by transient synchronized RE burst firing which produced rebound LTBs in TC cells. Besides RE inhibition, RTC cells also received IN inhibition which could elicit rebound LTBs. To differentiate the role of IN and RE inhibition in spindle oscillations, we blocked the HTC→IN and TC→RE projections, respectively, to eliminate IN and RE bursts. When IN burst firing was removed, HTC and RTC cells fired only three and one bursts respectively (Fig 5A), indicating that IN inhibition contributed to spindle oscillations by hyperpolarizing RTC cells. Nevertheless, the lack of IN inhibition could be easily compensated by increased inhibition from RE neurons. In the absence of IN inhibition, when the inhibitory RE→TC synaptic weight increased only 33% (from 3 nS to 4 nS), spindle oscillations persisted for about 4 seconds, twice the duration of the control case (compare Fig 5B with Fig 2B1). Thus, with sufficient RE inhibition, the TC-RE feedback loop was able to create and sustain spindle oscillations for a few seconds. On the other hand, when the TC→RE synapses were blocked, the TC-IN network generated four cycles of spindle oscillations before termination (Fig 5C). This was because the IN bursts hyperpolarized RTC cells resulting in rebound RTC bursts. At the same time, RTC hyperpolarization facilitated rebound bursts in HTC cells through the HTC-RTC gap junctions, which further drove IN bursting. Nevertheless, in the absence of RE inhibition, the inhibitory IN→RTC synaptic weight needed to increase fourfold (from 3 nS to 12 nS) in order to sustain spindles for about 1.5 seconds (Fig 5D). This contrasted to a much more prominent increase in spindle duration induced by a much smaller increase of the RE→TC synaptic weight (compare Fig 5D with Fig 5B). Such difference suggests that RE inhibition plays a more major role than IN inhibition in generating spindle oscillations, consistent with experimental data [3, 20, 38].
Similar to the effect of increased RE→TC synaptic weight, when the TC→RE synaptic weight increased 50% (AMPA: from 4 nS to 6 nS; NMDA: from 2 nS to 3 nS), RE neurons responded to TC inputs with more burst spikes and the duration of spindle oscillations increased substantially from about 2 seconds to 3.8 seconds (S12 Fig). Hence, spindle duration was fine-tuned by the strength of RE inhibition. As RE inhibition gradually decreased over the course of spindles due to short-term depression (STD), we hypothesized that removing STD at the TC→RE or RE→TC synapses could significantly extend or even sustain spindles indefinitely. Indeed, when STD at the TC→RE synapses was removed, spindle oscillations were sustained for about 4 seconds (Fig 5E), whereas they continued beyond 7 seconds when STD at the RE→TC synapses was abolished (Fig 5F). Thus in our model, spindle oscillations are terminated mainly by STD at the inhibitory RE→TC synapses, as suggested previously [39]. Our results are consistent with a recent experimental study showing that the duration of spindle oscillations depends critically on the inhibitory strength of RE neurons on TC cells [38]. Moreover, it suggests that the variable spindle duration observed in experiments may be a result of difference in RE excitability and heterogeneous synaptic strength between TC and RE neurons.
Experimental data has identified the HTC cell as an important neuronal substrate for thalamic θ and α oscillations [12, 17, 18]. To test the importance of HTBs in generating α oscillations, we varied the maximal conductance density of the high-threshold T-type Ca2+ current (gCa/HT) in HTC cells from 1 mS/cm2 to 5 mS/cm2 (default: 3 mS/cm2). When gCa/HT was reduced to 1 mS/cm2, HTC cells failed to produce HTBs (Fig 6A1, top). With a lack of excitation from HTC cells, IN neurons were mostly silent (Fig 6A1, upper middle). In the absence of feedforward inhibition from IN neurons, RTC cells fired random spontaneous activities at about 3 Hz without synchronization (Fig 6A1, lower middle). As a result, the α rhythm disappeared. On the other hand, when gCa/HT was increased to 5 mS/cm2, depolarized HTC cells started to transition from the bursting mode to the tonic spiking mode and became desynchronized because of strong heterogeneous bursting dynamics (Fig 6A2, top). Consequently, INs fired random bursts (Fig 6A2, upper middle) and suppressed the activity of RTC cells (Fig 6A2, lower middle). As a result, no synchronized α oscillations developed. These simulation results demonstrated that synchronized HTBs of HTC cells were essential for the generation of thalamic α oscillations, consistent with experimental observations [12, 17]. Fig 6A3 plots the network oscillation frequency (blue) and the spectral peak power (red) as a function of gCa/HT. When gCa/HT increased from 1.5 to 4.5 mS/cm2, the oscillation frequency moderately increased from 7.3 Hz to 9.8 Hz while the oscillation power remained relatively stable. At the values of 1 and 5 mS/cm2, there was a large drop in oscillation power because no synchronized α oscillations occurred. Hence, thalamic α oscillations are limited to a relatively narrow frequency band (8–10 Hz) where HTC cells fired robust high-threshold bursting.
The HTB frequency of HTC cells is also dependent on the depolarization level (S1 Fig; [7, 17]). As such, changing the external drive to the network would alter the frequency of α oscillations. Indeed, when the random afferent inputs were removed from the thalamic network, HTC cells still fired spontaneous synchronized HTBs, but at a lower frequency (6 Hz; Fig 6B1, top). Subsequently, the whole network was synchronized at about 6 Hz (Fig 6B1), which was within the θ frequency band. One the other hand, when the maximal synaptic input conductance to the whole network increased to 4 nS (default: 1.5 nS), the HTB frequency of HTC cells increased to about 12 Hz, but there was only one spike per burst (Fig 6B2, top) due to inactivation of the ICa/HT current. The random activity of both IN and RE neurons increased substantially because of reduced synchronized HTC excitation and increased random input drive, which substantially suppressed RTC firing (compare Fig 6B2 with Fig 2C2, lower middle). As a result, the α oscillation power was significantly reduced (Fig 6B3). The network oscillation frequency (blue) and power (red) as a function of the maximal input conductance (ginput) are shown in Fig 6B3. As ginput increased from 0 to 4 nS, the oscillation frequency increased monotonically from 6.1 Hz to 11.6 Hz. By comparison, the oscillation power increased initially from 0 nS to 0.5 nS and stayed in the same level for values up to 2 nS before decreasing considerably for larger input strength. Therefore, reducing the excitatory drive to the thalamic network switched α oscillations to θ oscillations while increasing the excitation level moved it to the upper α frequency band, consistent with experimental observation that HTC cells could underlie both α and θ oscillations dependent on the depolarization level of HTC cells [7, 12]. Nevertheless, although stronger depolarization of HTC cells increased α frequency, it reduced the α power by switching high-threshold bursting to tonic spiking (Fig 6A2 and 6B2). Consequently, thalamic α oscillations exhibit optimal power between 8–10 Hz (Fig 6A3 and 6B3). Our results thus explain why α oscillations decrease with more afferent excitation (e.g., eyes opening; [30]).
The key for γ oscillations was that HTC cells maintained a high level of synchrony even in the presence of strong random afferent inputs (Fig 2D2, top). The HTC rhythmicity was propagated to RE neurons via excitatory projection and to RTC cells via gap junctions. RE synchrony was boosted by inter-RE gap junctions and inhibition, which moderately constrained RTC firing through the RE→RTC inhibitory synapses. If the gap junctions between HTC cells were removed during γ oscillations, HTC cells were completely desynchronized (Fig 7A1, top) leading to unconstrained RE and RTC firing (Fig 7A1, lower middle and bottom). As a result, the sLFP spectral peak of either HTC or RTC cells was eliminated (Fig 7A2). Thus, gap junctions may play a critical role in synchronizing γ oscillations in the thalamus, as in the hippocampus [40] and neocortex [41].
Besides gap junctions, the negative feedback loop between excitatory (e.g., pyramidal cells) and inhibitory neurons also plays an important role in fast oscillation synchronization [42, 43]. To examine whether the feedback interaction between TC and RE cells could sustain γ oscillations as well, we increased the maximal RE→TC synaptic conductance fourfold (default: 3 nS; 4-fold increase: 12 nS) in the absence of HTC gap junction coupling. With much stronger RE inhibition, moderate degree of population rhythmicity emerged from the HTC, RTC and RE neurons (Fig 7B1). The sLFP frequency spectra revealed peaks of similar amplitude as controls, but at a lower frequency (controls: 30.5 Hz; 4-fold increase of RE inhibitory strength: 23.2 Hz; Fig 7B2) because of increased RE inhibition. Also, without HTC gap junctions, the amplitude of spectral peaks was similar for both HTC and RTC cells since they had similar degree of spike synchronization (Fig 7B1, top and lower middle). This was in contrast to the control case where the spectral peak of HTC-sLFP was much higher than that of RTC-sLFP (Fig 7B2). As a slight increase of the RE inhibitory synaptic weight significantly prolongs spindle oscillations (Fig 5B), such strong RE feedback inhibition (i.e., 4-fold increase of synaptic strength) seems unlikely in the thalamic network. Therefore, we conclude that HTC gap junctions are required for the synchronization of thalamic γ oscillations.
So far we have demonstrated that the thalamic network was able to generate multiple distinct oscillations dependent on ACh/NE modulation and afferent excitation. To further examine how the thalamic network transitions from one state to the other on a continuous basis, we divided the ACh/NE modulation into 11 levels evenly ranging from 0% to 100% and varied the maximal input conductance to TC cells from 0 up to 20 nS with a 0.5 nS step, which resulted in 451 different parameter combinations (see Methods section). We then simulated the thalamic network with all possible combinations of ACh/NE and input levels and identified the oscillatory state of the network across the entire parameter space. The network oscillation frequency and spectral power heat maps are shown in Fig 8A and 8B respectively. The networks with higher oscillation frequency (> 15 Hz) were predominantly located above the principal diagonal of the 2D parameter space (Fig 8A) indicating that the network oscillated faster with higher levels of ACh/NE modulation and afferent excitation. The highest oscillation frequency with the maximal level of ACh/NE (100%) and input (20 nS) was 38 Hz (note some oscillation frequencies close to 40 Hz right above the principal diagonal, but these network states were classified as non-oscillatory due to low oscillation power; see below). By comparison, the networks with higher oscillation power were primarily located below the principal diagonal with the maximal power residing in the bottom of the 2D space, corresponding to weak afferent excitation (Fig 8B). This suggests that oscillations driven minimally by the afferent input (e.g., spindle oscillations) are stronger than oscillations driven mostly by afferent excitation (e.g., γ oscillations). In addition, the oscillation power was minimal (< 1) above the principal diagonal except for high levels of ACh/NE modulation (> 70%).
Next, we determined the oscillatory state of the thalamic network under all examined combinations of ACh/NE and input levels and plotted the prominent oscillatory regimes across the entire parameter space (Fig 8C). Consistent with our modeling hypothesis (Fig 1B) and pervious analysis, we observed slow δ oscillations with low ACh/NE modulation (< 30%) and minimal afferent excitation (< 1 nS), spindle oscillations with medium ACh/NE modulation (20%-80%) and relatively weak afferent input (< 5 nS), and α, β and γ oscillations under high ACh/NE modulation (> 70%) with weak, moderate and strong inputs respectively (Fig 8C). The spindle oscillations under medium ACh/NE modulation were induced by a transient input to RE neurons as mentioned above (e.g., Fig 8E1) and the spindle regime moved to lower input area as the ACh/NE level increased (Fig 8C). For example, spindle appeared at 3 nS with 20% ACh/NE compared with 0 nS with 50% ACh/NE. This was because with low ACh/NE, TC cells fired spontaneous low-threshold bursts (LTBs) so higher afferent input was required to inactivate the low-threshold Ca2+ current to generate the state that permitted the occurrence of spindles. By comparison, for higher ACh/NE levels (50%-90%), TC cells no longer generated spontaneous LTBs and the network was relatively quiet without or with little afferent input (e.g., Fig 2B1). In addition to the externally-induced spindles (e.g., triggered by a transient input to RE neurons), we observed spontaneous spindle oscillations under low ACh/NE modulation (< 30%) and with higher level of afferent input (4.5–6.5 nS; Fig 8C). Such spontaneous spindle oscillations were induced spontaneously by the random afferent input and usually lasted for a few cycles (S13 Fig). Also, besides α oscillations at high ACh/NE modulation (100%), there were prominent α oscillations under low or medium ACh/NE modulation (< 60%) and with relatively large afferent input (> 8 nS). This was because large random afferent input activated the high-threshold bursting dynamics in HTC cells (see below). The region of such low-modulation α oscillations shifted to lower input area and became narrower as the ACh/NE level increased (Fig 8C; pink area in the left side). There were also sporadic α oscillations right above the region of spindles. During the transition from δ to spindles and from spindles to α oscillations, there existed two prominent θ oscillation regions (light green area): the first one appeared in the lower left corner while the second one started at higher input levels (7–12 nS) with 0% ACh/NE and gradually decreased to 0 nS with 100% ACh/NE, indicating lower afferent input was needed to drive the network to θ oscillations as ACh/NE increased. Besides the prime oscillatory regimes, there was also non-oscillatory area in the parameter map (light blue area in Fig 8C) characterized by very low oscillation power (Fig 8B). The first major non-oscillatory region was relatively narrow and appeared between the low-input θ oscillations and spontaneous spindle oscillations under low ACh/NE modulation (< 30%). By comparison, the second major non-oscillatory region was much wider which occurred above the principal diagonal extending from 0% up to 80% ACh/NE. One important characteristic of this non-oscillatory region was that it started at the highest input level (20 nS) without ACh/NE modulation (i.e., %0 ACh/NE) and expanded into lower input area as the ACh/NE modulation increased (Fig 8C, upper light blue area).
To understand the cellular basis of such oscillatory state transitions, we plotted the voltage traces of representative thalamic neurons with different afferent input drive at three levels of ACh/NE modulation (0%, 50% and 100%; Fig 8D–8F). As shown earlier, the thalamic network generated slow δ oscillations (1–4 Hz) with minimal input (e.g., 0.1 nS) in the low ACh/NE condition (Fig 2A). When the afferent input slightly increased (e.g., 2 nS), the LTB frequency of TC cells increased to the θ band (4–8 Hz; Fig 8D1), which underlay θ oscillations. Thus, model simulation suggested that both δ and θ oscillations could be mediated by the LTBs of TC cells depending on the input. With further increase of the afferent drive (e.g., 5 nS), the network entered the non-oscillatory state where HTC cells stopped bursting and RTC cells burst sparsely (Fig 8D2). This was because as the membrane depolarization increased, the low-threshold T-type Ca2+ current started to inactivate, reducing the intrinsic bursting dynamics of TC cells. Consequently, the spontaneous HTC bursts driven by random Poisson inputs were not able to induce synchronous bursts (only burstlets) but after-hyperpolarization in neighboring cells through gap junctions. As a result, the coupled HTC cells eventually became silent. Indeed, when the gap junctions were blocked, HTC cells were able to burst randomly (S14 Fig). When the afferent drive increased beyond 6.5 nS (e.g., 10 nS), HTC cells started to burst synchronously again owing to the activation of the high-threshold T-type Ca2+ current and the bursting frequency went into the θ band again (7.3 Hz, Fig 8D3). As the afferent input increased further (e.g., 15 nS), the burst frequency of HTC cells rose to the α band (9.2 Hz, Fig 8D4). At the highest afferent input tested (20 nS), HTC cells switched from HTBs to spare tonic spiking because of inactivation of the high-threshold T-type Ca2+ current and the network became desynchronized in the presence of strong random inputs (S15 Fig). To recapitulate, at the lowest level of ACh/NE (0%), as the afferent input increased, the thalamic network switched from slow δ oscillations to slow θ oscillations both mediated by the LTBs of TC cells. After a brief non-oscillatory state, the network exhibited spontaneous spindle oscillations followed by θ and α oscillations mediated by HTBs of HTC cells. The network eventually became desynchronized with strong afferent drive.
With medium level of ACh/NE modulation (50%), the thalamic network exhibited externally- induced spindle activity for relatively weak input (0–3 nS; Fig 8E1, also Fig 2B). The duration of spindles reduced with afferent excitation (S16 Fig) because depolarization increased the inactivation of the low-threshold Ca2+ current in TC cells making rebound LTBs less effective. Increasing the afferent input to 3.5 nS generated intermittent θ or spontaneous spindle-like oscillations with an oscillation frequency of 7.9 Hz (Fig 8E2). Specifically, the network driven by random Poisson inputs burst spontaneously for about 500 ms and stopped for about 500 ms before bursting again. Since the inter-burst interval (~500 ms) was relatively short, such network behavior was classified as intermittent or transient θ/α oscillations. When the afferent input further increased to 5 nS, the network generated continuous oscillations in the θ band (~ 6 Hz) which was mediated by HTBs of HTC cells (Fig 8E3). With large afferent drive (e.g., 15 nS), HTC cells switched from periodic HTBs to random sparse firing and the network became desynchronized (Fig 8E4). Similar to the non-oscillatory state in the low ACh/NE and high input condition (0% ACh/NE, 20 nS input; S15 Fig), the network desynchronization was caused by the inactivation of the high-threshold T-type Ca2+ current in the presence of random input drive (i.e., the gap junctions were no longer able to synchronize HTC cells with sparse random spikes). Notably, with medium ACh/NE modulation, the network entered the desynchronizing state at a much lower input level (0% ACh/NE: 20 nS; 50% ACh/NE: 9.5 nS; Fig 8C). This was due to the fact that with higher ACh/NE, the high-threshold T-type Ca2+ current started to inactivate at smaller input intensity. As a result, the non-oscillatory state expanded into lower input region when the ACh/NE modulation increased (Fig 8C, upper light blue area).
Under the condition of high ACh/NE modulation (100%), the thalamic network generated θ oscillations (6.7 Hz) without any input (0 nS, Fig 8F1) due to spontaneous HTBs of HTC cells. The bursting frequency increased to α band (8–14 Hz) when the afferent input slightly increased (e.g., 1.5 nS, Fig 8F2; similar to Fig 2C1). With further increase of the afferent input (e.g., 10 nS), HTC cells switched from HTBs to tonic spiking and the network synchronized at the β frequency band (23.8 Hz; Fig 8F3). With strong afferent excitation (e.g., 15 nS), the oscillation frequency of the thalamic network increased to the γ band (31.7 Hz, Fig 8F4; similar to Fig 2D1), which gave rise to potential γ oscillations. Note that INs had no spiking activities for medium or large afferent inputs (Fig 8F3 and 8F4, upper middle), different from default γ oscillations (Fig 2D1, upper middle). This was because INs did not receive direct afferent inputs in producing the oscillation map, while INs received weak afferent inputs (1.5 nS, Table 1) in the default γ simulation. In addition, different from the desynchronizing state under the low and medium ACh/NE conditions (S15 Fig and Fig 8E4), strong afferent drive did not disrupt network synchrony in the high ACh/NE condition (>80%; Fig 8F3 and 8F4). This was because HTC cells were much more excitable with high ACh/NE modulation and fired rhythmic spiking activities that were synchronized by gap junctions even when the high-threshold T-type Ca2+ current was inactivated. Thus, the model suggests that fast β/γ oscillations only occur at high ACh/NE level in the thalamic network.
Lastly, we applied periodic stimuli to the thalamic network to examine how distinct thalamic oscillations are modulated by rhythmic perturbations. The phasic pulsatile stimuli were introduced to the LGN and we assumed that all TC cells and IN neurons received the same stimulation input. To analyze the impact of stimulation on thalamic oscillation dynamics, we plotted the normalized color-coded frequency power spectrum of the sLFP (frequency spectrum heat map) in response to ascending stimulation (1–50 Hz) with a fixed stimulation amplitude (0.2 nA) for three major oscillatory states (δ, α and γ oscillations, Fig 9A1–9C1). In addition, to examine how the dominant oscillatory dynamics varied with the stimulation frequency, we plotted the dominant network oscillation frequency with normalized spectral peak as a function of the stimulation frequency for all three oscillatory states in Fig 9A2–9C2. Consistent with a recent computational study of an abstract cortical model [44], stimulation induced entrainment and resonance in the thalamic network model in all three oscillatory states. Importantly, we found that the occurrence of these phenomena was state-dependent. Entrainment, a response pattern where the intrinsic oscillations are locked to the simulation, was reflected by the highlighted spectral power along the diagonal with multiple harmonic and/or subharmonic components (above and below the diagonal) in the frequency spectrum heat map (Fig 9A1–9C1). Entrainment was much more prominent during stimulation of γ oscillations than δ and α oscillations indicated by the longer and higher diagonal power during γ oscillations (compare Fig 9C1 with Fig 9A1 and 9B1). The state-dependent entrainment effect was also evident in the dominant frequency plot where the entrained frequency range during primary entrainment (1:1 entrainment) was much wider during stimulation of γ oscillations than δ and α oscillations (compare Fig 9C2 with Fig 9A2 and 9B2, top; enclosed by red ellipses). In addition, we observed discontinuous entrainment where the thalamic network switched between the entrained and unentrained states during stimulation of α oscillations, but not during the other two oscillatory states (compare Fig 9B2 with Fig 9A2 and 9C2, top; enclosed by red ellipses). The entrainment behavior of the thalamic network during δ oscillations is illustrated in Fig 9A3. In response to 6 Hz stimulation, all four types of thalamic neurons fired highly synchronized bursts at the same stimulation frequency (6 Hz) that were tightly phase-locked to the stimulation pulses. We also note that primary entrainment occurred when the stimulation frequency was close to but mostly higher than the endogenous frequency for all three oscillatory states (Fig 9A2–9C2, top; enclosed by red ellipses), suggesting that stimulation favored higher frequency entrainment. For example, primary entrainment occurred at 3–10 Hz during stimulation of δ oscillations (endogenous frequency: 3.7 Hz) and took place at 19–47 Hz during stimulation of γ oscillations (endogenous frequency: 30.5 Hz). Besides primary entrainment, thalamic network oscillations were also shaped by subharmonic as well as harmonic entrainment (indicated by blue, cyan and magenta ellipses, Fig 9A2–9C2, top). The thalamic network activity during subharmonic entrainment of α oscillations is illustrated in Fig 9B3. In response to 22 Hz stimulation, HTC cells burst at half of the stimulation frequency (11 Hz) and the whole network synchronized at 11 Hz. By comparison, during (the first) harmonic entrainment of γ oscillations, HTC cells oscillated at twice the stimulation frequency (S17 Fig).
Stimulation of the LGN also induced resonance, an enhancement of oscillation power when the stimulation frequency was close to the endogenous frequency, its harmonics and/or subharmonics. The thalamic neuronal activity during primary resonance is exemplified in Fig 9C3. When the LGN was stimulated at 35 Hz, close to the endogenous frequency (30.5 Hz), the spiking activities of all four types of neurons became more synchronized and rhythmic than without stimulation (compare Fig 9C3 with Fig 2D2). Resonance occurred in all three oscillatory states (Fig 9A2–9C2, bottom), but with substantial differences among the states. During stimulation of γ oscillations, the primary resonance peak was much wider than that during stimulation of δ or α oscillations (compare Fig 9C2 with Fig 9A2 and 9B2, bottom; enclosed by red ellipses), in line with much wider primary entrained frequency range during stimulation of γ oscillations. Notably, similar to entrainment, resonance was asymmetric with respect to the endogenous frequency and favored higher frequency stimulation. For instance, the primary resonance peak occurred at 5 and 6 Hz during stimulation of δ oscillations (Fig 9A2, bottom; enclosed by red ellipse) and 10 and 11 Hz during stimulation of α oscillations (Fig 9B2, bottom; enclosed by red ellipse), both of which were higher than their respective endogenous frequencies (3.7 Hz and 9.2 Hz respectively). Such asymmetry was most prominent during stimulation of γ oscillations. Although the endogenous γ frequency was 30.5 Hz, stimulation induced maximal resonance (> 60%) between 35 and 41 Hz (Fig 9C2, bottom; enclosed by red ellipse). The strong bias towards higher frequency in both entrainment and resonance suggests that stimulation increases the endogenous frequency of thalamic oscillations by enhancing the excitability of TC neurons. In addition to resonance enhancement, we observed substantial power suppression during high frequency (> 25 Hz) stimulation of α oscillations (Fig 9B2, bottom; indicated by black arrow), but not during the other two oscillatory states. This was because high frequency stimulation switched HTC bursting to tonic spiking and suppressed RTC activities by increasing IN firing (S18 Fig). Overall, stimulation of the thalamic network induces state-dependent entrainment and resonance, which are stronger during γ oscillations than δ and α oscillations.
Given the dominant focus on cortical circuits for the mechanistic study of brain oscillations, the thalamus has been somewhat overlooked, despite repeated suggestions that the thalamus may serve as a “pacemaker” of brain oscillations [7, 8, 45, 46]. Indeed, recent experimental findings indicate that the thalamus plays a key role in controlling cortical states and functioning [32, 47, 48] and regulating the information transmission between different cortical areas [49, 50]. However, it is not clear whether the thalamus can independently generate different types of brain oscillations without interaction with the cortex. Answering this question will help to determine the validity of a thalamic pacemaker model. Using a biophysically realistic thalamic network model, we provided for the first time a full account of how multiple distinct oscillations could arise from the cellular and network properties of the thalamic circuitry and how thalamic oscillation transitions from one state to the another. This model then enabled us to investigate the state-dependent response to brain stimulation.
The key for the genesis of multiple distinct oscillations is that TC cells exhibit multiple oscillatory bursting/spiking patterns depending on neuromodulation and afferent excitation level (S1 Fig). In the low ACh/NE modulation state, TC cells are sufficiently hyperpolarized to generate low-threshold bursts (LTBs) in the δ frequency band mediated by the low-threshold T-type Ca2+ current [51, 52]. Such low-frequency LTBs form the neuronal basis of δ oscillations (Fig 10). In the medium ACh/NE modulation state, although TC cells are not spontaneously bursting, hyperpolarization of TC cells by afferent input or inhibition evoked rebound LTBs after the release of hyperpolarization [53, 54]. Such hyperpolarization-induced rebound LTBs are necessary for spindle oscillations (Fig 10). In the high ACh/NE modulation state, HTC cells fire high-threshold bursts (HTBs) in the θ/α frequency band that mediate θ or α oscillations depending on afferent excitation level (Fig 10; [7, 12]). Stronger depolarization of HTC cells by afferent input switches HTBs into high frequency tonic spiking enabling γ oscillations (Fig 10). Such high frequency tonic spiking of TC cells was supported by experimental data that TC neurons exhibited fast oscillatory discharge in the γ frequency band in LGN (~50 Hz; [11]) and in the ventroanterior-ventrolateral (VA-VL) complex (20–40 Hz; [27]). Thus, the rhythmic burst/spiking properties of TC neurons provide the cellular mechanism underlying multiple distinct thalamic oscillations [55].
To generate oscillations at the population level, the rhythmic burst firing or tonic spiking of TC neurons has to be synchronized. We found that the specific connectivity of the thalamic network endows it with the capability to synchronize under different neuromodulatory states and over a wide range of frequencies. First, the existence of gap junctions plays a crucial role in the synchronization of network activity. In particular, the gap junctions between HTC cells [7, 12, 17, 18] enable them to serve as a “pacemaker” or synchronization “engine” of the circuit that remain synchronized even in the presence of strong random noisy inputs. We showed that thalamic oscillations were either impaired or eliminated if the HTC gap junctions were removed, consistent with experimental observation [12, 35]. Besides HTC gap junctions, the weaker gap junctions between HTC and RTC cells and the inter-RE gap junctions also contribute to the synchronization of the network. Second, the feedforward IN inhibition and feedback RE inhibition lead to synchronization of RTC cells. Different from HTC cells, there are no gap junctions among RTC neurons [7, 17]. Nevertheless, synchronized HTC activity propagates to both IN and RE neurons, which constrain RTC firing via feedforward inhibition. As RTC cells also project excitatory inputs on RE neurons, they receive feedback RE inhibition as well which enhances the synchronization. During low frequency oscillations (δ, spindle and α), we observed that either IN or RE inhibition was sufficient to synchronize RTC activity, while during fast frequency oscillations (β/γ), simultaneous rhythmic IN and RE inhibition was needed for high level of RTC synchrony due to strong random noisy inputs (S9 Fig). Thus, the presence of both IN and RE inhibition strengthens the synchronization of TC neurons. Third, multiple mechanisms contribute to the generation, synchronization and stability of thalamic oscillations. The specific connectivity of the thalamic circuit allows for the synergy of multiple mechanisms in the generation of synchronized oscillations. For example, δ oscillations are synchronized by both gap junctions and inhibition (Fig 4). During spindle oscillations, both the feedforward IN and feedback RE inhibition contribute to the rebound bursting of TC cells. In particular, the feedforward HTC→IN→RTC inhibition could also lead to feedback inhibition on HTC cells via the HTC-RTC gap junctions. Similarly, during fast β/γ oscillations, while the HTC gap junctions play a major role, the feedback RE inhibition could also contribute to fast oscillations if the inhibitory strength is sufficiently strong. We hypothesize that the existence of multiple synchronizing mechanisms coupled with the strong intrinsic oscillatory properties of TC cells enables the thalamic network to serve as a pacemaker during thalamocortical oscillations, consistent with experimental observation [56].
To further test the hypothesis that generation and transition of distinct thalamic oscillations are functions of both ACh/NE neuromodulation and afferent excitation, we varied the ACh/NE and input levels systematically to identify the prominent oscillatory regimes across the entire parameter space. The oscillation transition map (Fig 8C) confirmed the existence of multiple distinct oscillations under physiological conditions (Fig 2B): δ oscillations with low ACh/NE modulation and minimal input (deep sleep); spindle oscillations with medium ACh/NE modulation and slight to weak inputs (light sleep); α/θ oscillations with high ACh/NE modulation and weak input (relaxed wakefulness) and γ/β oscillations with high ACh/NE modulation and strong input (arousal and attention). One interesting and surprising finding from the oscillation transition map is that it reveals prominent α oscillations under low ACh/NE modulation and with large afferent input thus extending the original model hypothesis where α oscillations occur under high ACh/NE modulation and with low afferent input (compare Fig 8C with Fig 1B). Note that α oscillations arising from these two different regions have different physiological implications. While α oscillations under high ACh/NE modulation correspond to the state of relaxed wakefulness, α oscillations under low ACh/NE modulation represent a non-physiological state in deep sleep where the thalamus receives strong afferent drive.
In addition, the oscillation transition map reveals several important insights and predictions as to the distribution and transition of thalamic oscillations. First, there exist two separate regions for the generation of θ oscillations. The first region locates between δ and spindle oscillations and is mediated by the LTBs of TC cells, while the second region resides between spindle and α oscillations and is mediated by HTBs of HTC cells. Thus, the model predicts the existence of θ oscillations during the transition from spindle to δ oscillations. Also, persistent and coherent θ oscillations mediated by LTBs are definitive signatures of a number of neurological or psychiatric disorders including Parkinson’s disease, major depressive disorder (MDD) and schizophrenia [6, 7]. Hence, the θ oscillation region mediated by LTBs could potentially correspond to a pathological state. Second, spontaneous spindle oscillations can be generated under relatively low ACh/NE modulation but with moderate level afferent input. Spindle oscillations can be initiated by a number of mechanisms including spontaneous firing from a more excitable region of the thalamus, spontaneous oscillating TC cells or cortical stimulation through excitation of RE cells [57]. In the default simulation (Fig 2B), we used a brief input to RE cells to induce spindle oscillations. Here we showed that spindle oscillations could also be initiated by spontaneous firing of TC cells given that the random background inputs are relatively high, consistent with previous hypotheses [57]. Third, there exist transient θ/α oscillations under medium ACh/NE modulation and with moderate level of afferent input. During the transition from spindle oscillations to persistent θ oscillations under medium level ACh/NE (e.g., 50%), we observed transient θ/α oscillations which lasted for a few hundred milliseconds (e.g., Fig 8E2). Such transient θ/α oscillations were induced by a surge of background inputs, sustained a few cycles through the TC-RE interaction and terminated because of the inactivation of the low-threshold Ca2+ current (due to afferent input; returning back the resting state), similar to spontaneous spindle oscillations. Interestingly, recent experimental studies discovered transient or intermittent α/β events in the awake mammalian neocortex [58, 59]. Combined experimental and computational evidence showed that such transient oscillations emerged from the integration of synchronous bursts of excitatory synaptic drive targeting proximal and distal dendrites of pyramidal neurons [59]. Thus, though both the thalamus and neocortex are capable of generating transient rhythms, their underlying mechanisms may be different. Fourth, the effect of afferent input is somewhat equivalent to ACh/NE modulation. For example, without afferent input (0 nS), as the ACh/NE modulation increased, the thalamic oscillatory state switched from δ oscillations to θ oscillations and to spindle oscillations and to θ oscillations again (Fig 8C). Similarly, for 0% ACh/NE, as the afferent input increased, the thalamic oscillation switched from δ oscillations to θ oscillations and to spindle oscillations after a brief quiescent state followed by θ oscillations again (Fig 8C). Lastly, fast frequency β or γ oscillations in the thalamus can only be generated under high ACh/NE modulation. This is because under low and medium ACh/NE modulation, large (random) afferent input desynchronizes the network when the HTBs of HTC cells switch to sparse tonic spiking. This is consistent with the experimental data that fast γ oscillations in the thalamus are supported by cholinergic projection from the brainstem [27] and acetylcholine release contributes to gamma oscillations in prefrontal cortex during attention [60].
We used a brief depolarizing input (to RE neurons) to trigger spindle oscillations in the default simulation (Fig 2B). Such brief depolarization could arise from the cortex; for example, the initial portion of the cortical up state during slow oscillations (<1 Hz) triggers thalamic spindles [61, 62] and cortical stimulation induces spindle oscillations in the thalamus [29]. Alternatively, such transient depolarization may result from a synchronized surge of random background inputs to TC or RE neurons, as discussed above. In either case, spindle oscillations are generated internally from the thalamic circuit in the absence of cortical modulation, as observed experimentally [63, 64]. Besides, substantial depolarization of TC cells is required to generate γ rhythms in the thalamic network [27, 37]. Such strong depolarization could result from a combination of sensory stimulus [11], elevated cholinergic neuromodulation [27] and increased corticothalamic projection to the thalamus [24]. Independent of the source of depolarization, it should be emphasized that we modeled the TC depolarization (in the high ACh/NE modulation state) by increasing the synaptic strength of the random Poisson inputs to TC cells which did not contain any rhythmic structure. Hence, although the cortex may provide necessary inputs to trigger spindles and depolarize TC neurons during γ oscillations, it is through the intra-thalamic mechanisms that the thalamic model produces distinct states of oscillatory patterns, in agreement with experimental findings [2, 24, 27].
Our modeling results suggest that the thalamus could be a driving force for thalamocortical oscillations, consistent with experimental observations [7, 8, 46]. Nevertheless, this prediction does not preclude the existence of cortically-originated oscillations in the thalamocortical systems. Indeed, δ oscillations are found to contain both thalamic and cortical components [37] and the cortical δ waves persist in the absence of the thalamus [65]. Similarly, α oscillations could be of cortical origin mediated by layer 5 pyramidal cells [10]. Also, corticothalamic feedback strongly modulates spindle oscillations [33] and synchronizes spindle waves to widespread cortical areas [63]. Moreover, γ oscillations in both the visual and motor cortex could arise from intracortical mechanisms [9, 66]. Thus, it is possible that independent neural generators exist for distinct thalamic and cortical oscillations and the corticothalamic feedback synchronizes the thalamic and cortical rhythms into coherent thalamocortical oscillations [3, 37, 67]. Indeed, it has been proposed that three cardinal oscillators (one cortical oscillator and two thalamic oscillators) underlie the generation of the slow (< 1 Hz) electroencephalogram (EEG) rhythm of NREM sleep through intricate dynamic interactions [68]. Future studies are needed to examine how independent cortical and thalamic oscillators interact dynamically to create coherent rhythms in the thalamocortical system.
Rhythmic stimulation has become an important and promising technique to study the causal role of oscillations in brain function [69, 70] and as therapeutic intervention to treat neurological and psychiatric disorders [71, 72]. As such, the responses of neuronal circuits to periodic perturbation have been examined in a number of experimental [73, 74] and computational/theoretical studies [44, 75–79]. Most of the existing theoretical models focused on the nonlinear dynamics of neurons using simplified neuronal models (e.g., integrate-and-fire neurons [75–77]) and did not consider the interaction between rhythmic input and intrinsic neuronal dynamics. One notable exception was a cortical network model developed by Tiesinga [78] that investigated the dependence of LFP resonance on different biophysical time scales of the neuronal circuit. Since the Tiesinga model focused on pyramidal interneuron network gamma (PING) in the cortex, it remains unclear how the stimulation interacts with endogenous neural dynamics and depends on the physiological state of the thalamus. The manifestation of multiple distinct oscillations in one unified biophysical thalamic model enabled us to examine the impact of rhythmic stimulation on thalamic network dynamics. When subject to periodic stimulation, the thalamic network displayed two most prominent response patterns, entrainment and resonance, as shown previously in more abstract neuronal models [44, 75–77]. Importantly, such response patterns were highly state dependent in that stimulation of γ oscillations induced much stronger entrainment and resonance than δ and α oscillations (Fig 9). We hypothesize that this is because γ oscillations are mainly driven by afferent excitation and the network synchrony or endogenous oscillation power is much lower compared with δ or α oscillations which are driven by intrinsic mechanisms (i.e., neuronal bursting; Fig 2). Thus, during fast γ oscillations, the thalamus can more effectively relay sensory inputs than slow δ and α oscillations (i.e., easier to be entrained). Our simulation results are consistent with experimental observations that reduced α oscillation power enables entrainment [80] and fast γ band oscillations facilitate visual information processing [73, 81]. In addition to entrainment and resonance, high frequency stimulation induced strong suppression on α oscillations, which was not observed during stimulation of δ and γ oscillations for the same stimulation amplitude (Fig 9). Taken together, our model is the first to demonstrate how the response patterns of the thalamic network to periodic stimulation depend on the physiological state or intrinsic dynamics of constituent neurons. As the particular oscillation state of the thalamus is set by both ACh/NE modulation and afferent excitation, our model thus suggests that ACh/NE and afferent excitation also define the thalamic response to brain stimulation.
Thalamic oscillations have been modeled by a number of previous studies, either in isolated thalamus [57, 82–84] or integrated thalamocortical network [15, 85–93]. Several novel features distinguish our model from the existing thalamic models. First, our model incorporated a newly identified, special subclass of TC cells, high-threshold bursting TC cells (HTCs) and the gap junctions among HTCs [12, 17, 18]. In addition, the model contained gap junctions between HTCs and relay-mode TC cells (RTCs) [17] and gap junctions among reticular neurons [94, 95], which enhanced the thalamic network synchrony. The only existing model that consisted of HTCs and gap junctions was developed by Vijayan and Kopell [84] to study α oscillation and its role in stimulus processing. However, the Vijayan and Kopell model did not model INs explicitly and did not include the gap junctions between HTCs and RTCs and among RE cells. Also, the Vijayan and Kopell model did not consider the action of NE on α oscillations. As a result, the RE neurons were silent due to ACh inhibition during the muscarinic ACh receptor- induced α activity (Fig 1B of [84]). By comparison, our model considered the combined action of ACh and NE so RE cells fired tonic spiking during α oscillations (Fig 2C).
Second, our thalamic model integrates multiple distinct oscillations (δ, spindle, α/θ, and γ/β) into one unified framework, in a similar spirit to a previous model of carbachol-induced δ, θ and γ oscillations in the hippocampus [96]. Notably, all four neuronal models (HTC, RTC, IN and RE) in the thalamic network were parameterized carefully to produce experimentally observed firing patterns under different neuromodulatory states (S1 Text, S1 Fig and S2 Fig), enabling it to generate distinct, neuromodulation-dependent oscillation patterns. In contrast, most of the existing thalamic models focused on only one or two specific oscillatory patterns such as α oscillations [84], spindle oscillations [15, 57, 83, 89, 90], δ and spindle transition [82, 85], spindle and gamma oscillations [88] or spindle and slow (< 1 Hz) oscillations [92]. Other thalamocortical models have studied the transition from slow sleep oscillations (< 1 Hz) to asynchronous waking state [86, 87]. It should be noted that a recent thalamocortical model from the Bazhenov group also investigated the generation and transition of multiple distinct oscillations during the sleep stages: spindle in NREM 2, slow δ oscillations during NREM 3 and α or mu-like rhythm during REM sleep [93]. There are several differences between our model and the Krishan et al. model. First, the Krishan et al. model focused δ band activity mostly on slow oscillations (0.5–1 Hz) generated from the cortex, while our model considered regular δ oscillations (1–4 Hz) originated from thalamic TC cells. Second, the α or mu-like rhythm in the Krishan et al. model was generated by sparse synchronized single spiking in the cortex, while the α oscillation in our model was produced by synchronized high-threshold bursting in thalamic HTC cells. Third, our thalamic model was able to produce fast frequency γ/β oscillations, while the Krishan et al. model did not consider fast frequency oscillations. Last, we only varied two model parameters (potassium leak conductance and synaptic input conductance) to generate multiple oscillations, while more parameter change was needed to switch oscillation from one state to the other in the Krishan et al. model.
Lastly, our model is the first to consider the co-regulation of neuromodulation and afferent input in thalamic oscillatory state transition. The effects of neuromodulation on thalamic oscillatory state transition have been examined in a number of models (e.g., [87, 92, 93]), but these previous models mostly focused on the sole action of neuromodulation. By comparison, we investigated the combined effect of neuromodulation and afferent input and demonstrated that the generation and transition of thalamic oscillations are functions of both neuromodulation and afferent excitation.
As for any scientific study, there are several limitations to our work. First, as mentioned above, in addition to acetylcholine and norepinephrine, thalamic processing is subject to other neuromodulator action such as histamine (HA), serotonin (5-HT) and adenosine [25]. Similar to ACh and NE, most of these neuromodulators target the potassium leak current (IKL) in TC and RE neurons [25]. For example, application of HA leads to a slow depolarization in TC cells by decreasing the IKL [97]. Also, application of 5-HT in vitro strongly depolarize RE neurons via reduction of the IKL, an effect similar to noradrenergic modulation [98]. Thus, the net effect of these neuromodulators seems to be a depolarization of both TC and RE neurons and we hypothesize that activation of multiple neuromodulatory systems strengthens the neuromodulatory control of thalamic oscillatory state transition. Second, cholinergic and noradrenergic inputs modulate other ionic currents in thalamic cells other than IKL. For example, the hyperpolarization-activated cation current IH in TC cells is modulated by NE [99] and variation of IH conductance density was able to switch the oscillatory pattern between δ and spindle-like oscillations in a TC cell model [82]. Also, ACh may influence the muscarinic current IM and affect excitatory synaptic strength in the thalamus, similar to its modulatory effect in the cortex [100]. We only considered the modulation of IKL since it is the major target of cholinergic and noradrenergic inputs [25, 26, 101] and we attempt to model the transition of multiple oscillatory states with minimal change of parameters. Thus, our model can be considered as a minimal model of thalamic oscillation transition and inclusion of other ionic currents modulated by ACh/NE can be studied in future modeling studies. Third, in our model, δ oscillations occur during deep sleep stage, which corresponds to low ACh/NE modulation and minimal afferent excitation (Fig 8C), consistent with experimental observation [29, 31]. A recent experimental work in nonhuman primates demonstrates that in primary auditory cortex δ and γ oscillations co-occur during attentive processing while α and β oscillations occur during periods of inattention [102]. The existence of δ oscillations in the study could be due to the fact that both the auditory and visual stimuli are presented in the δ frequency band (1.6 Hz and 1.8 Hz respectively). In addition, our modeling results agree with the experimental data [102] that α oscillations occur during inattention while γ oscillations occur during attention. Future study is needed to examine δ and γ phase coupling during attention in a thalamocortical model subject to rhythmic δ band stimulation. Lastly, as the major goal of this study is to determine whether an isolated thalamus is capable of generating multiple distinct oscillation patterns, it presently does not include the cortex. The absence of the cortex prevents us modeling certain oscillatory pattern that is of cortical origin such as slow oscillations (< 1 Hz; [86, 87, 103]). Nevertheless, a deeper mechanistic understanding of thalamic oscillations enables the systematic investigation of the cellular and circuit mechanisms of thalamocortical rhythms in the future.
The thalamic network consisted of both the lateral geniculate nucleus (LGN) and the reticular nucleus (TRN) (Fig 1A). The LGN contained two major cell types: thalamocortical (TC) cells and local interneurons (INs), and the TRN contained reticular (RE) cells. The TC cells were further divided into high-threshold bursting TC (HTC) cells and relay-mode TC (RTC) cells, based on whether TC cells can generate high-threshold bursting or not [12, 18]. In the cat LGN, HTC cells account for about 25%-30% of the whole TC population [12, 17], while INs constitute about 25% of the total neuronal population in all dorsal thalamic nuclei of cats and primates [104]. Accordingly, the thalamic network contained 49 (7×7) HTC cells, 144 (12×12) RTC cells, 64 (8×8) INs and 100 (10×10) RE neurons, all placed in a two-dimensional grid. The major modeling results were robust to the network size when the synaptic weight and connectivity density were scaled accordingly. The network connectivity between the four types of neurons was illustrated in Fig 1A. HTC cells were connected with gap junctions [12, 17] and provided feedforward excitation to INs, which in turn delivered feedforward inhibition to RTC cells [18]. A small percentage (20%) of RTC cells were also connected with HTC cells via gap junctions [17]. Both HTC and RTC cells sent glutamatergic synapses to RE neurons, while receiving GABAergic feedback inhibition from the RE population [20, 105]. RE neurons were connected with both gap junctions [94, 95] and GABAergic synapses [106, 107]. Lastly, a small percentage (~10%) of RE neurons project GABAergic synapses to local interneurons [108].
In the model, all gap junction connections were local and the maximal distance between two electrically coupled neurons was two units (the distance between two adjacent neurons in horizontal or vertical direction was assumed to be one unit). Each HTC cell formed gap junctions with neighboring HTC cells with a random connection probability. Also, 20% of RTC cells (randomly selected) formed gap junctions with neighboring HTC cells and 20% of RE cells (randomly selected) formed gap junctions with neighboring RE cells. For all gap junction connections, the connection probability was taken to be 30% within the local region. By comparison, all chemical synapses were global in this relatively small network. The connection probability (0.3) was higher for the TC-IN connections (including HTC→IN and IN→RTC projections) than that (0.2) for the TC-RE connections (including HTC→RE, RTC→RE, RE→HTC and RE→RTC projections) because TC cells show higher correlation with INs than with RE cells during α oscillations in cats [18]. A connection probability of 0.2 was used for the RE→RE synapses, while a much smaller probability (0.05) was used for the RE→IN synapses according to experimental data [108].
Following previous “point” models of thalamic cells [57, 82–84, 86, 109], all single cell models in the thalamic network contained one single compartment and multiple ionic currents described by the Hodgkin-Huxley formulism [110]. The current balance equation was given by:
CmdVdt=−gL(V−EL)−gKL(V−EKL)−∑Iint−10−3∑IsynA+10−3IappA
(1)
where Cm = 1μF/cm2 is the membrane capacitance for all four types of neurons, gL is the leakage conductance (nominal value: gL = 0.01 mS/cm2 for all four types of cells) and gKL is the potassium leak conductance modulated by both ACh and NE (see Table 1 and below for details). EL is the leakage reversal potential (EL = −70 mV for both HTC and RTC cells; EL = −60 mV for both IN and RE neurons), and EKL is the reversal potential for the potassium leak current (EKL = −90 mV for all neurons). Iint and Isyn are the intrinsic ionic currents (in μA/cm2) and synaptic currents (in nA) respectively and Iapp is the externally applied current injection (in nA). The following total membrane area (A) was used to normalize the synaptic and externally applied currents in Eq (1): HTC and RTC cells: 2.9×10−4 cm2 [109]; INs: 1.7×10−4 cm2 [111]; RE cell: 1.43×10−4 cm2 [57, 86].
All active ionic conductances were modeled using the Hodgkin-Huxley formalism [110]. Specifically, the ionic current for channel i, Ii, was modeled as Ii = gimphq(V − Ei), where gi was its maximal conductance density, m its activation variable (with exponent p), h its inactivation variable (with exponent q), and Ei its reversal potential. The ICAN current utilized a slightly modified equation [114]: ICAN = gCANM([Ca]i)m(V − ECAN), where M([Ca]i) = [Ca]i/(0.2 + [Ca]i) is a Michaelis-Menten function and [Ca]i denotes the intracellular calcium concentration. The kinetic equation for the gating variable x (m or h) satisfied a first-order kinetic model:
dxdt=ϕxx∞(V,[Ca]i)−xτx(V,[Ca]i)
(2)
where ϕx is a temperature-dependent factor, x∞ is the voltage or Ca2+- dependent steady state and τx is the voltage or Ca2+- dependent time constant in msec. Equivalently, Eq (2) can be written as:
dxdt=ϕx(αx(V,[Ca]i)(1−x)−βx(V,[Ca]i)x)
(3)
where αx and βx are the voltage or Ca2+- dependent rate constants with dimension of msec-1. The maximal conductance densities of all ionic currents and the kinetic parameters of all gating variables for all four types of neurons are listed in S1 and S2 Tables, respectively. The sodium reversal potential was set to ENa = 50 mV and the potassium to EK = -90 mV. The reversal potentials for IH and ICAN currents were EH = -43 mV [113] and ECAN = 10 mV [114] respectively. The calcium reversal potential (ECa) was dynamically determined by the Nernst equation in all cell types in the model [118]:
ECa=RT2Flog([Ca2+]o[Ca2+]i)
(4)
where R = 8.31441 J/(mol°K), T = 309.15°K, F = 96,489 C/mol, and [Ca2+]o = 2 mM.
Intracellular calcium was regulated by a simple first-order differential equation of the form [118, 119]:
d[Ca2+]idt=−ICazFw+[Ca2+]rest−[Ca2+]iτCa
(5)
where ICa is the summation of all Ca2+ currents, w is the thickness of the perimembrane “shell” in which calcium is able to affect membrane properties (0.5 μm), z = 2 is the valence of the Ca2+ ion, F is the Faraday constant, and τCa is the Ca2+ removal rate (10 ms for HTC, RTC and IN cells; 100 ms for RE cells). The resting Ca2+ concentration was set to be [Ca2+]rest = .05 μM.
The gap junction current was computed as Igap = (Vpost − Vpre)/Rg, where Vpre and Vpost are the membrane potentials of the presynaptic and postsynaptic neurons respectively. Gap junctional coupling was stronger among HTC cells than between HTC and RTC cells [17]. Accordingly, the gap junction resistance Rg was smaller for the HTC-HTC synapses (100 MΩ) than for the HTC-RTC synapses (300 MΩ). The coupling strength between RE cells was set to be the same as that between HTC and RTC cells (Rg = 300 MΩ). These gap junction resistance values were selected to match the experimental data [12, 17, 94]. In the model, glutamatergic synaptic current was mediated by both AMPA and NMDA receptors, while GABAergic synaptic current was mediated by GABAA receptors. The synaptic current was calculated by the following equation [119, 120]:
Isyn=gsynsB(V)(V−Esyn)
(6)
where gsyn is the maximal synaptic conductance and Esyn is the synaptic reversal potential. The default maximal conductances were: gAMPA = 6 nS and gNMDA = 3 nS for HTC→IN synapses, and gAMPA = 4 nS and gNMDA = 2 nS for the TC→RE synapses. The synaptic strength from inhibitory neurons (INs and REs) to TC cells was assumed to be higher than that among inhibitory neurons: gGABA = 3 nS for IN→RTC and RE→TC synapses while gGABA = 1 nS for both RE→IN and RE→RE synapses. Esyn = 0 mV for AMPA and NMDA currents, and Esyn = -80 mV for GABAA receptors in TC cells, while Esyn = -70 mV for GABAA receptors in RE neurons [106, 121]. The function B(V), which implements the Mg2+ block for NMDA currents, was defined as B(V) = 1/(1 + exp(−(V + 25)/12.5) [86, 112]. For AMPA and GABAA currents, B(V) = 1. The gating variable s represents the fraction of open synaptic ion channels and obeys a first-order kinetics [82, 83, 122]:
dsdt=α[T](1−s)−βs
(7)
where [T] is the concentration of neurotransmitter in the synapse and α and β are forward and backward binding rates. The neurotransmitter is assumed to be a brief pulse that has duration of 0.3 ms and amplitude of 0.5 mM following an action potential in the presynaptic neuron [57]. The channel opening rate constants (α and β) are given as: α = 0.94 ms-1, β = 0.18 ms-1 for AMPA receptor current, α = 1 ms-1, β = 0.0067 ms-1 for NMDA receptor current and α = 10.5 ms-1, β = 0.166 ms-1 for GABAA receptor current. These values were taken from previous modeling studies [57, 86, 118]. A synaptic delay of 2 ms was introduced in all chemical synapses.
Short-term synaptic depression was implemented in all chemical synapses and was modeled by scaling the maximal conductance of a given synaptic channel by a depression variable D, which represented the amount of available “synaptic vesicles” [86, 87]. The variable D was updated according to a simple phenomenological rule [86, 123]:
D=1−(1−Di(1−U))exp(−t−tiτ)
(8)
where U = 0.07 is the fraction of resources used per action potential, τ = 700 ms is the time constant of recovery of the synaptic vesicles. Di is the value of D immediately before the ith presynaptic spike and ti is the timing of the ith spike event.
All neurons in the thalamic network received independent Poisson-distributed spike inputs at an average rate of 100 Hz (results maintained unchanged if higher input rates were used when also scaling down the maximal synaptic input conductance). These random inputs represented both extrinsic sources of background noise and asynchronous visual input. This input was exclusively mediated by AMPA receptors modeled as an instantaneous step followed by an exponential decay with a time constant of 5 ms [124]. The synaptic input weights (i.e., maximal synaptic conductance) for all neuronal types during different oscillatory states are given in Table 1. Spindle oscillations were triggered by a transient input (100 ms × 100 pA) injected into RE neurons which represented a cortical UP state or a surge of synchronized background inputs.
To introduce heterogeneity into the model neurons, the leakage conductance (gL) of all neurons in the network was drawn from a uniform distribution within ±25% of the default value (i.e., 0.0075–0.0125 mS/cm2). This leak conductance variation, random synaptic connectivity, and random external inputs constituted the model noise in the thalamic network.
Our central modeling hypothesis is that the generation and transition of distinct thalamic oscillatory states are functions of both ACh/NE neuromodulation and afferent excitation level (Fig 1B). Main motivations were the fact that different oscillatory states appear under different behavioral conditions in the sleep-wakefulness cycle (Table 1) and that the transition from sleep to wakefulness is controlled mainly by activation of both cholinergic and noradrenergic neuromodulatory systems [25, 26, 125]. The thalamic oscillatory state transition is also a function of afferent excitation since thalamic neurons receive stronger afferent inputs during wakefulness than sleep due to activation of the sensory systems. Accordingly, we modeled four distinct oscillations (δ, spindle, α and γ) under three different ACh/NE modulation states (low, medium and high) corresponding to deep sleep, light sleep and awake conditions (Table 1). Specifically, δ oscillations were modeled in low ACh/NE modulation state with minimal afferent excitation; spindle oscillations were modeled in the medium ACh/NE modulation state with slight afferent excitation; and α and γ oscillations were modeled in the high ACh/NE modulation state with weak and strong afferent excitation respectively (Table 1). The two levels of afferent excitation in the high ACh/NE modulation state modeled two different behavioral conditions: awake with eyes closed (where α oscillations are maximal) and awake with eyes open and with attention. The effect of ACh/NE modulation was modeled by varying the potassium leak conductance in all four types of thalamic neurons while different afferent excitation was modeled by changing the maximal synaptic conductance of the random Poisson inputs to thalamic cells (Table 1). The rationale and selection of specific parameter values during each oscillatory state for both the potassium leak conductance and synaptic input conductance are described below.
Acetylcholine (ACh) and norepinephrine (NE) alter the intrinsic excitability of thalamic neurons mainly by modulating the potassium leak current [25, 26, 101, 126–128]. Both ACh and NE directly depolarize TC cells via blocking the potassium leak current [25, 101, 127]. By comparison, ACh inhibits LGN local interneurons and RE cells by activating the potassium leak current via muscarinic receptor activation [126, 128]. In contrast, application of NE or stimulation of the locus coeruleus enhances the excitability of RE neurons by reducing the potassium leak current [25, 26, 127]. The combined effect of ACh and NE on RE cells is inferred from experimental data showing that a progressive hyperpolarization occurred in RE neurons during the transition from arousal to quite wakefulness and to deeper states of EEG-synchronized sleep [61, 129]. We thus assumed that the excitatory effect of NE dominated the inhibitory effect of ACh on RE neurons so that the potassium leak current was decreased during transition from sleep to wakefulness. Also, since the action of NE on LGN interneurons remains unknown [25, 26], the ACh/NE neuromodulatory effect on interneurons was assumed to be mediated by cholinergic action only.
We selected a periodic pulsatile stimulus that conceptually resembles the waveform of deep brain stimulation (DBS) or repetitive transcranial magnetic stimulation (rTMS) [44]. The stimulation was assumed to be global: all neurons in the LGN (TC & IN cells) received the same stimulus pattern. Stimulation consisted of a train of 10 ms brief square current pulse applied at different frequencies ranging from 1 Hz to 50 Hz with a 1 Hz/step increment. The stimulation amplitude was fixed at 0.2 nA. Stimulation was performed on three major oscillatory states (δ, α and γ oscillations) and stimulation at each frequency lasted for 1 second.
The thalamic network model was coded with C++. All simulation were performed using the fourth-order Runge-Kutta (RK4) method with a fixed time step of 0.02 ms. Shorter simulation step did not change the results. The major simulation results were validated in a separate model implementation using the Brian simulator [131]. Simulations were run on a Dell Linux workstation under Ubuntu. The model source codes are available in the ModelDB database (https://senselab.med.yale.edu/modeldb/).
|
10.1371/journal.pbio.1001651 | Modulation of Global Low-Frequency Motions Underlies Allosteric Regulation: Demonstration in CRP/FNR Family Transcription Factors | Allostery is a fundamental process by which ligand binding to a protein alters its activity at a distinct site. There is growing evidence that allosteric cooperativity can be communicated by modulation of protein dynamics without conformational change. The mechanisms, however, for communicating dynamic fluctuations between sites are debated. We provide a foundational theory for how allostery can occur as a function of low-frequency dynamics without a change in structure. We have generated coarse-grained models that describe the protein backbone motions of the CRP/FNR family transcription factors, CAP of Escherichia coli and GlxR of Corynebacterium glutamicum. The latter we demonstrate as a new exemplar for allostery without conformation change. We observe that binding the first molecule of cAMP ligand is correlated with modulation of the global normal modes and negative cooperativity for binding the second cAMP ligand without a change in mean structure. The theory makes key experimental predictions that are tested through an analysis of variant proteins by structural biology and isothermal calorimetry. Quantifying allostery as a free energy landscape revealed a protein “design space” that identified the inter- and intramolecular regulatory parameters that frame CRP/FNR family allostery. Furthermore, through analyzing CAP variants from diverse species, we demonstrate an evolutionary selection pressure to conserve residues crucial for allosteric control. This finding provides a link between the position of CRP/FNR transcription factors within the allosteric free energy landscapes and evolutionary selection pressures. Our study therefore reveals significant features of the mechanistic basis for allostery. Changes in low-frequency dynamics correlate with allosteric effects on ligand binding without the requirement for a defined spatial pathway. In addition to evolving suitable three-dimensional structures, CRP/FNR family transcription factors have been selected to occupy a dynamic space that fine-tunes biological activity and thus establishes the means to engineer allosteric mechanisms driven by low-frequency dynamics.
| Allostery is a process by which a molecule binding to one site of a protein alters the activity of the protein at another site. Allostery is typically thought to occur through a change in protein structure, but there is now clear evidence that the dynamic properties of a protein can also regulate allostery without a change in overall conformation. Here we examine two members of a large family of bacterial transcription factors and provide a mechanism to describe the allosteric binding of their activating ligands. We demonstrate, in these systems, that allostery arises as a natural consequence of changes in global low-frequency protein fluctuations on ligand binding. We further demonstrate that the higher dimensional parameter space that describes all potential variant transcription factors can be reduced to a two-dimensional free energy landscape that determines the key molecular parameters that predominantly regulate allostery. We additionally show that the amino acids we determine as contributing sensitively to allosteric control tend to be conserved in diverse bacteria; thus we identify a link between residues that contribute to low-frequency fluctuations and evolutionary selection pressures.
| Small regulatory molecules frequently bind proteins at regions remote from the active site. These allosteric events can switch proteins between inactive and active states [1]. Knowledge of the molecular basis of allostery derives from a wealth of theoretical and experimental studies and traditionally describes the process in terms of conformational change within the protein [2],[3]. Combinations of X-ray crystallography and NMR have permitted analysis of the ligand binding sites, intermolecular interactions, and conformational fluctuations that underpin diverse allosteric systems [4],[5]. There is also considerable evidence that allosteric cooperativity can be communicated between distant sites on proteins through modulation of their dynamic properties, even in cases where that are no structural changes between the ligand bound (holo) and unbound (apo) forms [6]–[12]. Since the original identification, by Cooper and Dryden [4], of this alternative route of “allostery without conformational change,” there has been considerable debate over the mechanisms by which dynamic fluctuations are communicated between allosterically coupled sites on proteins.
One hypothesis for fluctuation-induced allostery is that binding modifies the structure of the thermally excited global normal modes and thence the coupling interaction between cooperative elements. This in turn affects the structural ensemble of the distant sites and so the free energy of binding [13]–[15]. Another view maintains that physically connected pathways of excited or repressed dynamics, coupled along their trajectories, connect allosteric sites [16]–[18]. Here we propose the hypothesis that the normal modes of protein structural motion, large-scale motions dispersed across the entire protein, are important carriers of the allosteric signal and act without requiring structural change. Previous studies of the normal modes have demonstrated that conformational transitions in proteins, including those that underpin allosteric regulation dependent on conformational change, are well described by one or a few low-frequency modes [19]–[25]. The normal modes, however, can also be used to describe the whole spectrum of internal fluctuations of a protein around a mean structure. The low-frequency global modes, in particular, can involve entire protein domains. Alteration of the normal modes might therefore be communicated to distant sites of a protein as a change in the degree of motion around a mean structure without overall conformational change. Global low-frequency fluctuation therefore represents an alternative theoretical approach to allosteric communication that does not depend upon conformational change. An important consequence of this alternative mechanism of allosteric communication is that it can be captured by coarse-grained representations and models, such as the elastic network model (ENM). Here we develop this theory, and the validity of a coarse-grained model approach, through a computational and experimental study of the homodimeric CRP/FNR family transcription factors Catabolite Activator Protein (CAP) of Escherichia coli and GlxR of Corynebacterium glutamicum.
CAP is a 210-amino-acid transcription factor that binds cAMP generated by adenylyl cyclase in response to the phosphorylated form of Enzyme IIAGlc (phosphorylated in response to the phosphoenolpyruvate-carbohydrate phosphotransferase system) [26],[27]. cAMP-bound CAP regulates the transcription of over 100 genes crucial for carbon utilization through its binding to a specific promoter region and recruitment of RNA polymerase [28]. Previous studies of the ligand binding domain of CAP demonstrated negative cooperativity between cAMP binding sites in the absence of structural change within this domain [10]. The observed negative cooperativity in this isolated domain occurs through a conformational entropic penalty for binding the second molecule of cAMP, but there is no mechanistic description for how such a phenomenon can occur in the full-length protein. Seven of eight CAP mutants previously examined showed a direct correlation between ΔΔG and the adiabatic compressibility (βs°) where proteins with a higher βs° (reflecting increased structural flexibility in solution) demonstrated enhanced negative cooperativity [29]. While it is therefore reasonable to hypothesize a role for protein dynamics in allostery in CAP, there is no conceptual framework to describe how these changes in motion might arise, how they contribute to allostery, and how a resulting theory might translate to related molecules. CAP is therefore a suitable model system for a theoretical and experimental investigation of the contribution of the normal modes to allostery.
Here we propose that changes to global low-frequency protein backbone fluctuations are carriers of an allosteric signal in CAP and present this in the context of a significant new quantitative theory for allosteric coupling. We produce coarse-grained models that describe global low-frequency protein backbone motions of CAP and show a strong correlation between negative cooperativity for cAMP and modulation of the delocalised normal modes on ligand binding without a requirement for a spatially distinct physical pathway or conformational change. We demonstrate experimentally that altered connectivity between backbone elements in CAP can give predictable alterations to cooperativity for cAMP binding through altered mode amplitudes. We further demonstrate a broader applicability for this theory using an additional CRP/FNR family transcription factor, GlxR of C. glutamicum. We unite our findings for CAP and GlxR to determine the extent to which key inter- and intramolecular parameters contribute to negative cooperativity in CRP/FNR family transcription factors. We further demonstrate that amino acids that contribute significantly to allosteric control are more likely to be conserved in variant proteins from diverse species. The theoretical and experimental work and associated data analysis provide both a significant advance in our understanding of the mechanisms that underpin the dynamic regulation of allostery and also a means for informed rational engineering of cooperativity in proteins.
To computationally address cases of allostery that arise from fluctuation-modification, without conformational change, requires a very different approach from those corresponding to the classic Monod-Wyman-Changeaux case of conformational switching. On the one hand, fully atomistic simulations are not capable of attaining, in most cases, the long dynamical time scales explored by the slow, global dynamic modes whose thermodynamics are essential for the effect. On the other hand, because these modes by their nature integrate many local interactions into their effective geometries and potentials, coarser-grained models of protein structure can possibly provide sufficiently accurate calculations of the relevant dynamics, while allowing the computation of dynamics to the necessary timescales. Models that represent protein structures by Cα-atom positions alone reproduce low-frequency modes well in comparison to experimental data [21],[30]. We therefore used the co-ordinates from a high-resolution crystal structure determination of the full-length cAMP bound CAP homodimer to construct an ENM [31] for the apoprotein as well as single and double ligand bound holoprotein states (Figure S1). Free energies, ΔG, were calculated using the full harmonic solution, and the negatively cooperative binding of cAMP to wild-type full-length CAP confirmed by calculating a positive value for ΔΔG = (ΔGholo2−ΔGholo1)−(ΔGholo1−ΔGapo) = 179 cal mol−1 consistent with experimentally obtained values (Table S2) [32]–[35]. To confirm that the total motion within the ENM is not an artefact of coarse-graining, we also carried out molecular dynamics simulations [36] with full atomistic detail, including an explicit water model, and performed principle component analysis (PCA) on the generated trajectories [37]. B factors represent the convolution of static and dynamic disorder in the crystal. Dynamic disorder can be attributed to local motions of individual atoms, whereas static disorder represents different atomic positions in the individual protein molecules. The experimental B factors, albeit constrained by crystal packing, therefore represent a reasonable approximation of the local motions in solution [38]. ENMs and atomistic PCAs represent overall unconstrained dynamic motions and hence show much larger deviations in the termini and the flexible loop regions (for example, residues 150–175 of Figure S2). The crystallographic B factor data were qualitatively well represented at either scale of coarse-graining (Figure S2a) and the distribution of the normal mode frequencies agreed well between ENM and PCA (Figure S2b). The total predicted motion within the ENM, at least at the level of B factors and low-frequency mode structure, is therefore similar to other methods of analysis and not an arbitrary feature of the model. Since the fluctuation-induced allosteric effect arises from the low-frequency structure of the protein dynamics, the ENM level of analysis applies to the experimental phenomena studied here.
We hypothesized that if side-chain replacement on amino acids at sites distinct from the cAMP binding site of CAP do not cause conformational rearrangement, yet increase or decrease amino acid side chain hydrophobic or electrostatic forces in their local environments, the normal modes of protein motion would be altered without significant structural changes. If these changes to the normal modes have sufficiently global effects, they will in turn modify cooperativity between the cAMP binding sites through an entropic contribution to the binding free energy. Amino acid side chain replacement can therefore act as a sensitive probe of the contribution of side chain connectivity to cooperativity and the underlying mechanism for allostery within the elastic structure of the protein. The change in allosteric free energy (ΔΔG) as a function of altering the entire primary amino acid sequence (one residue at a time) can therefore be viewed as a quantitative map of the contribution of the normal modes to cooperativity. Such a quantitative map can be constructed either by simulation or experiment; in practice, it is convenient, as we demonstrate below, to use simulation of the entire allosteric map to guide mutagenesis for experimental study. We therefore performed a scanning computational mutagenesis of the entire CAP protein to investigate the influence of side chain connectivity on cooperativity via their influence on the normal modes.
Changing the effective elastic potential between protein backbone carbon atoms in the neighbourhood of each residue of the ENM in turn and calculating effects on ΔΔG was used to determine the scanning computational mutagenesis map. The increase and decrease in elastic potential in the ENM was hypothesized to simulate the effects of local strengthening and weakening of side chain interactions in CAP. A color-coded map corresponding to altered cooperativity with changing local interaction strength is plotted graphically by amino acid residue (Figure 1a) and in real space (Figure 1b). The global map for the ENM (Figure 1a) demonstrates large regions where cooperativity is susceptible to control by altering side chain connectivity. It is important to note that these control regions are not necessarily adjacent to the cAMP-binding site. For example, regions corresponding to amino acids 127–137 (at the interface between the two monomers) and 150–162 (within the DNA binding domain, far from both the dimer interface and cAMP binding regions) appear to exercise considerable control over cooperativity without contributing to a spatially distinct dynamic pathway and without direct interference with the cAMP binding site.
To experimentally test the model and demonstrate rational engineering and control of allostery, we selected the residues of CAP highlighted in Figure 1b. We examined amino acids predicted to show altered (V132, H160) or neutral (V140) responses to altered amino acid side chain interactions (Table 1). The removal (V132A) or addition (V132L) of a side chain methyl group of V132 was engineered to decrease and increase, respectively, the strength of hydrophobic interaction across the dimer interface. Computation predicted that these mutations would result in more negative and positive cooperativity in CAP, respectively (Figure 2a) and that the most important contacts contributing to this effect were with L62 and V132 of the opposing monomer (Figure S3b). High-resolution X-ray crystal structures of CAP mutants V132A and V132L demonstrated that these variants possessed decreased and increased hydrophobic interactions across the dimer interface, respectively (Figure 2b). Comparison of variant crystal structures with wild-type demonstrated that there was no statistically significant change in structure (Figure S4, Table S1). Cooperativity for cAMP binding was studied by isothermal titration calorimetry (ITC) for wild-type, V132A, and V132L proteins to examine whether the experimentally observed changes in cooperativity matched computational predictions (Figure 2c–e, Table 1). The ITC data were well-described by a three-site model, with two major and one minor cAMP binding site (Figure S5) [39] and allowed derivation of the thermodynamic parameters for all proteins (Table S2). The qualitative computational prediction for altered cAMP cooperativity was tested experimentally including a significant controlled inversion of the sign of the cAMP cooperativity (V132L). The thermodynamic parameters for wild-type CAP demonstrated an overall favourable entropy change and unfavourable enthalpy change on binding the second molecule of cAMP consistent with a previous report [39]. A previous study of the truncated CAP ligand-binding domain demonstrated that binding of the second molecule of cAMP was entropically unfavoured [10]. The difference in thermodynamics between our experiments (Table S2) and previous experiments using the ligand-binding domain alone [10] is therefore likely due to the contribution of motions of the DNA binding domain [40]. This interpretation is supported by previous analysis that has calculated the thermodynamic contribution of the DNA binding domains in the switch to the active conformation [41]. Previous calculations and experiments anticipate that, while the contribution of the normal modes to allostery is entropically controlled (in terms of the net allosteric free energy), coupling of the low-frequency modes to side-chain motion generically gives rise to additional, but compensating, contributions to enthalpy and entropy and this is observed in our thermodynamic data (Table S2) [9]. It is notable that, due to this self-cancelling of the contribution of local fast modes within the total free energy, the entropically driven ENM is able to predict qualitative changes to experimental cooperativity despite the local mode contribution of enthalpy to overall thermodynamics.
The ENM calculations predicted a reduction in the negative cooperativity of CAP in response to a reduction in the strength of the local interactions of residue H160 (Figure 3a). In particular, H160 was predicted to form interactions that contribute to allostery with D162 and Q165 (Figure S3a). The mutation H160L was predicted to break these interactions while maintaining side chain bulk; this was confirmed by X-ray crystallography of the H160L CAP protein (Figure 3b). No overall change in H160L protein structure was evident compared to wild-type (Figure S4, Table S1). ITC experiments (Figure 3c) demonstrated that cooperativity for cAMP became less negative as predicted by computation (Table 1). This crucial experiment demonstrates that altering low-frequency motions at a site distant from both the ligand binding site as well as the dimer interface, and from any presumed physical pathway of structural change connecting these sites, can nonetheless give predictable effects on cooperativity.
Altering local interactions associated with V140 was predicted by the ENM to have minimal effects on cooperativity (Figure 4a) despite significant local hydrophobic interactions; we therefore examined the effect of decreased and increased local hydrophobic interactions in V140A and V140L variants as a control experiment. The V140L mutant protein had no discernible effect on protein structure (Figure S4). As predicted by the ENM mutagenesis, measurement of cooperativity for cAMP in V140L by ITC (Figure 4c) showed no differences when compared to wild-type (Table 1). Interestingly, although V140A protein showed no global change in structure (Figure S4), there is, in this mutation, a significant local conformational change evident in the crystal structure where the mutated V140A residue formed a new hydrophobic contact with the rotated side chain of C179 that is not present in the wild-type or V140L proteins (Figure 4b). When included in the model, simulated as kC179/k = 4, this new contact revealed new interactions within the monomer (Figure S3a) that drove CAP towards positive cooperativity on simulation (Table 1). ITC experiments (Figure 4d) demonstrated that this CAP variant with the identified side chain rearrangement was positively cooperative, thus supporting the qualitative prediction of the model.
A bar graph for the calculated and observed values for K2/K1 revealed the agreement in the direction of the change of cooperativity on simulation and experiment (Figure S6a). A plot of the experimentally observed value for K2/K1 against that predicted from the ENM demonstrated a correlated relationship where observed increases to K2/K1 are associated with similar changes to K2/K1 by the ENM (Figure S6b). The consistency in prediction by the ENM and the quantitative correlation between predicted and observed changes do not support the notion that the agreement between experiment and the ENM is due to a chance occurrence.
The ENM can provide further insight into the mechanism by which allosteric control is associated with alterations to the normal modes. No global structural changes were induced in the ENM simulations or were evident from crystal structures of variant proteins; only the pattern of coupled low-frequency fluctuations was modified by the simulated side-chain mutations. This appearance of “control at a distance” in the CAP homodimer is explained, through contributions to binding entropy, if there are correlations in the low-frequency motions between cAMP binding sites and if ligand binding or side chain mutation modifies this correlation [42]. As all fluctuating systems dominated by locally harmonic interactions possess a structure of normal modes, with just such distant correlations, they suggest the mechanism for allostery in CAP. To examine whether the mutations studied here can have such distant effects, we calculated the change to local Cα flexibility in the case of tightening and loosening side chain interactions at V132 at the dimer interface (Figure 5a). Modifications to simulated backbone flexibility are present throughout CAP with varying amplitude and furthermore follow opposite signs at kV132/k = 0.25 (V132A) and kV132/k = 4 (V132L). For example, kV132/k = 4 shows significant tightening of the protein (compare Figure 5a and Figure S3b). An examination of the effect of simulated mutations at V140 and H160 on nonlocal Cα flexibility reinforces this finding (Figure S7). The predominantly neutral mutation, V140L, simulated as kV140/k = 4 has little effect on protein backbone flexibility, except at sites where V140 has calculated interactions, consistent with the absence of any effect on allostery on both simulation and experiment. In the case of H160 (kH160/k = 0.25; at a surface loop distant from both the cAMP binding site and dimer interface) and V140A (kC179/k = 4, kV140/k = 0.25), the simulated mutations create a uniform decrease in flexibility throughout the monomer except for the straightforward loosening/tightening at the site of the mutations. There is a general trend, therefore, for those simulated mutations that decrease negative cooperativity to be associated with decreased protein backbone motion nonlocally.
A specific requirement of global low-frequency motion as an underpinning mechanism for allostery at a distance is a coupling between protein motion and the behaviour of the cAMP-binding site. We find that the loosening and tightening effects of simulated mutations is correlated with significant modulation of backbone flexibility in the region of the cAMP-binding site (amino acids 71–74, 83–85, and 121) (Figure 5b). The figure shows that, in general, changes in root-mean-square deviation (rmsd) at the ligand-binding site induced by mutation correlate (in this case, kR/k = 0.25) with cooperativity. Mutations that increase motion at the ligand bind site are associated with an increase in the extent of negative cooperativity and vice versa. This is entirely consistent with the controlling entropic allosteric mechanism in these cases, providing that cAMP binding has the effect of increasing local rigidity. This interaction between the heightened local motions following the first cAMP-binding event creates an entropic contribution to negative cooperativity in ΔΔG [9]. Heightened fluctuation at the second binding site (on binding the first molecule of cAMP) is a general mechanism for negative cooperativity without conformational change [6]. Positive cooperativity without conformational change can be induced by reducing the fluctuation amplitude (for example, the MetJ transcription factor of E. coli [9]).
Studies using CAP have successfully demonstrated that changes to global low-frequency protein dynamics are associated with allostery. We investigated another protein to explore the more general applicability of the mechanism. GlxR of C. glutamicum is a cAMP binding homodimeric transcription factor of the CRP/FNR family that activates genes required for aerobic respiration, glycolysis, and ATP synthesis [43],[44]. We solved the X-ray crystal structure of the GlxR apoprotein to produce an ENM for the non-cAMP bound state [45]. Coordinates from an available crystal structure determination of the full-length cAMP bound GlxR homodimer allowed us to construct an ENM for the single and double ligand bound holoprotein states. Examination of the structures for GlxR in the apo and holo forms revealed no significant difference in structure. GlxR therefore represents a new exemplar for allostery in the absence of conformation change. Free energies, calculated from ENMs for GlxR, predicted considerably greater negative cooperative binding of cAMP (K2/K1 = 2.37; ΔΔG = 513 cal mol−1) than for CAP (K2/K1 = 1.35; ΔΔG = 179 cal mol−1). This prediction of enhanced negative cooperativity was confirmed on experiment with an observed value for K2/K1 of 19.47 (Table 2). A computational scanning mutagenesis map was produced for GlxR, as done previously for CAP, and altered cooperativity with changing local interaction strength is plotted graphically by amino acid residue (Figure 6a) and in real space (Figure 6b). Both local tightening and loosening across the dimer interface, depending on the residue, was predicted to reduce negative cooperativity and therefore provides a robust experimental test of the model. We generated dimer interface loosening (kL134/k = 0.25; L134V; Figure 7a) and tightening (kA131/k = 4; A131V; Figure 7b) GlxR variants and compared simulated and experimental values for cooperativity in these proteins. Both L134V and A131V showed a clear reduction in negative cooperativity, as predicted, when compared to wild-type (Table 2) by ITC (Figure 7c–e), despite the fact that the mutants have opposing effects on hydrophobic interactions across the dimer interface. Allostery is therefore correlated with global low-frequency dynamics in an additional CRP/FNR family transcription factor.
Our findings indicate general biophysical principles that describe the emergence of negative cooperativity in CRP/FNR family transcription factors through the allosteric modulation of normal modes. The property that allosteric effects are carried in general by the more globally distributed, and so typically longer wavelength, normal modes motivated the exploration of the underlying physics by coarse-graining the CAP and GlxR representations even further into rotational-translational block representations [46]. Two coarse-grained blocks per monomer (one is the entire DNA-binding region, coupled only to the other block of its own monomer) emerged naturally from the many residue–residue couplings internal to and between monomers at the molecular level. Figure 8a and 8b display the block structure and the corresponding “super-coarse-grained” model. A single representative internal mode within each dynamically tight block and the coupling strengths between the blocks (including coupling across the dimer interface) were investigated as “design parameters” for a general class of cooperative homodimer. Figure 8c (CAP) and 8d (GlxR) show allosteric cooperativity, calculated at this high level of coarse-graining, as a function of the integrated coupling strengths within the ligand binding domain (k1) and between monomers (k12). Points below and above the z = 0 plane correspond to positive and negative cooperativity, respectively. The wild-type proteins for both CAP and GlxR are offset from the maxima of anti-cooperative ridges in the two-dimensional free energy landscapes that emerge. At this position, loosening coupling internal to monomers (k1) moves the system into a basin of less negative cooperativity (GlxR) or positive cooperativity (CAP), while loosening in the coupling region (k12) moves the system for both CAP and GlxR to the top of the ridge (red) to increase negative cooperativity. Further analysis demonstrated consistency in the negative cooperativity arising through the normal modes in the ENM and in the super-coarse-grained model. For example, the simulated loosening (kV132/k = 0.25; V132A) and tightening (kV132/k = 4; V132L) mutations of the CAP ENM and the tightening (kA131/k = 4; A131V) mutation of GlxR alter cooperativity through generating effective changes in k12 at the super-coarse-grained level. The super-coarse-grained model therefore effectively reveals the critical intra- and intermolecular parameters that associate with cooperativity and how these parameters can be altered to move within the allosteric free energy landscape.
If cooperativity confers a selective advantage on the organism, then the allosteric free energy landscape can also be viewed as evolutionary landscape. In this case, the position of a protein within the landscape depends upon selection pressures that impact upon k1 and k12. This general hypothesis can be used to make an additional significant experimental prediction. If the similar position of CAP and GlxR within their respective free energy landscapes is the result of a selection pressure, then we predict that amino acids that contribute significantly to quantitative allosteric control (Figure 1a and 6d) will be more invariant in related proteins from different species. We therefore examined 163 CAP variants from diverse bacterial species and plotted the frequency of mutation of each amino acid residue against the contribution of that amino acid to allostery (defined as absolute change (Δ) in K2/K1 for that amino acid in the canonical CAP ENM at kR/k = 0.25). We found evidence that the rate at which an amino acid mutates is negatively related to ΔK2/K1 (LRT, G2 = 33.7, p<0.001; Figure 9). The coefficient quantifying this decrease, β1, was significantly different from zero [95% CI = (−3.34,−1.49)]. Amino acids of CAP that contribute to allostery through regulation of low-frequency protein dynamics are therefore more likely to be conserved in CAP variants through their contribution to protein function. Note that a test for overdispersion was significant, even after allostery had been accounted for (LRT, G1 = 1,663.9, p<0.001), suggesting that other variables also have an influence on mutation rates.
Here we demonstrate that negative allostery in CRP/FNR family transcription factors is correlated with modulation of the normal modes of protein motion on ligand binding in the absence of conformational change. The model makes key predictions that we test at select sites of the CAP and GlxR proteins, the latter identified as an important new exemplar for allostery in the absence of conformation change. The alterations in protein flexibility that are a signature for allostery in CRP/FNR family transcription factors are a consequence of the global nature of those normal modes responsible and mutations that predictably alter cooperativity do so by influencing protein backbone flexibility. Our theory describes how allostery can arise from changes to low-frequency dynamics in the absence of any mean structural change. The theory is particularly significant as it describes allostery as a natural consequence of the dynamic properties of a protein without a requirement for spatially localised dynamic pathways between allosteric sites. The allostery observed is unlikely to have microheterogeneity as an alternative explanation as all CAP proteins crystallised as a single superimposable structure. Any form of heterogeneity reduces the likelihood of forming ordered crystals [47]. Microheterogeneity is therefore not supported as a molecular cause for allostery in CAP.
The possibility of a direct interaction between cAMP binding sites might also be considered as a mechanism to explain the allostery observed. The closest distance between the two cAMP molecules in the CAP dimer is 9.5 Å (the distance between the N6 atoms of the adenine ring). Although it is impossible to conclusively eliminate small local changes that binding of the first molecule of cAMP has at the second site, no conformational changes have been reported in this region in previous NMR studies, making this explanation unlikely. The possibility of a direct interaction is made even more unlikely as, similar as to that described above, any invoked direct interaction between cAMP binding sites would have to consistently match not only the qualitative aspects of the computational predictions for the role of the global modes, but also their quantitative correlation with the observed experimental values. Analysis of the relationship between Cartesian distance and protein motions demonstrated strongly correlated motions between allosteric sites at distances of <10–20 Å [48] and the global normal modes are a suitable candidate to mediate such correlations in CRP/FNR family transcription factors.
The range of available sites for side chain mutagenesis of CRP/FNR family transcription factors do not constitute as large a set of separate and independent control parameters as at first seems, but in a good approximation explore a lower dimensional space (i.e., reducing the very high dimensional parameter-space of the entire number of residues, just one slice of which is represented in Figures 1a and 6d, to the two-dimensional parameter spaces of Figure 8c–d). We hypothesize that this two-dimensional parameter space is, in turn, related to an evolutionary landscape for a protein. In the case of CAP and GlxR, our analysis reveals that evolutionary selection has resulted in the location of the proteins in a region close to maximizing negative cooperativity. The extent of negative cooperativity in CAP is generally small (ΔΔG = 0.3 kcal mol−1). However, the scale of biologically relevant cooperative effects is set by the thermal energy (RT≈0.6 kcal mol−1). The values of ΔΔG observed and manipulated experimentally are those that modulate the concentration range of cAMP to which the system is sensitive by an order of 1. Engineering of cooperativity is therefore possible by manipulating ΔΔG, as described here, with the caveat that it is likely only possible over a thermodynamic range to which the protein is responsive.
We find that there is a selection pressure against mutation of residues that contribute to allostery in CAP variants. A significant question that arises, therefore, is that of the selective advantage provided through negative cooperativity in CAP. In general, the advantages conferred by negative cooperativity in biological systems are not well resolved [49]. It is proposed that negative cooperativity reduces the sensitivity of a system and extends the concentration range over which a response can be observed [50]. In metabolism, recent modelling suggests that there is a significant overall advantage for metabolic pathway flux with components showing negative cooperativity [51],[52]. In transcriptional regulation, negative cooperativity in the binding of D-camphor to the CamR repressor of Pseudomonas putida is proposed to enable coupling of high specificity for D-camphor with a physiological response to high concentrations of the metabolite [53]. Against this framework, it is reasonable to conjecture that negative cooperativity in CAP offers a selective advantage by increasing the concentration range over which a transcriptional response can be generated [54]. The decreased sensitivity of the response to cAMP in negative cooperativity might result in a selective advantage through resource conservation when compared to amplifying effect of a signal response in positive cooperativity [50]. The position within the effective parameter space can also allow CAP variants to further tune cooperativity in either direction without a potentially disastrous influence on protein structure and therefore function. Future experiments to experimentally validate the selective advantage provided by negative cooperativity will therefore be crucial and might typically combine high throughput sequencing of extensive mutational libraries of CAP, after selection in E. coli, with the simulated mutational map of this study [55].
The super-coarse-graining and finer-grained tools we have developed and tested in this work suggest a route to artificial protein design through modification of protein low-frequency fluctuations without compromise of structure. The mechanism also reflects an important balance between phenomena at different length scales within molecular biology. The role of the global normal modes in conveying allosteric signals requires a similarly coarse-grained picture of the protein to identify and discuss the mechanism. On the other hand, the exquisite specificity to local biochemistry is preserved in the mechanism; a set of single residues, themselves spatially distant from either binding site, exercise significant control on the size (and sign) of the underlying allosteric signal. The delicate interactions of effects at different length scales are missed without such a multiscale approach to the physics of protein dynamics. Changes to the normal modes are presented as an important new theory to describe how allostery can arise in the absence of structural change and provide an important theoretical context within which to frame global issues of allostery in proteins.
The open reading frame corresponding to the full-length CAP protein was cloned into the BamHI and HindIII sites of pQE30 and mutant variants constructed by site-directed mutagenesis. Wild-type and mutant recombinant protein was expressed from E. coli M182 ΔCAP F− Δ(lacIPOZY)X74 galE15 galK16 rpsL thi+ lambda− [pREP4] for 2 h at 37°C with 1 mM IPTG. Protein was purified using sequential nickel-chelated sepharose affinity and Superdex 75 16/60 size exclusion columns (GE Healthcare). Protein concentration was calculated using the Beer-Lambert Law and a molar extinction coefficient of 20,065 M−1 cm−1 at 280 nm. Full-length GlxR protein was expressed and purified as previously described [56].
Protein was dialyzed against 100 mM KPO4 pH 7.8, 200 mM KCl, 2 mM 1-thioglycerol at 4°C. Protein and buffer were degassed under vacuum and degassed buffer used to dilute cAMP ligand. cAMP concentration was calculated using the Beer-Lambert Law and a molar extinction coefficient of 14,650 M−1 cm−1 at 259 nm. Data were generated using an iTC200 (MicroCal) by typically 40 sequential 1 µL injections of 4–6 mM cAMP into 202 µL 130–400 µM protein. Data for the first injection was routinely discarded as this is affected by diffusion between the syringe and the protein solution during equilibration prior to data collection.
Ligand binding for cAMP to CAP was described by a sequential three-site model (two major and one minor binding site [39]). The presence of three cAMP binding sites in CAP was further confirmed in the crystal structures from this study (Figure S5). A sequential two-site model described ligand binding for cAMP to GlxR. The free ligand concentration, [L], was calculated for each injection using the bisection method, which allowed calculation of the fraction of the protein in each bound state, Fi:Comparing the calculated heat content, Q, to the experimental value allowed calculation of the best fit of the binding constants, Ki, and the binding enthalpies, ΔHi, using the solver plug-in for Excel:
ITC and ENM data for mutant proteins was compared to the wild-type by a comparison of means by one-way ANOVA. Normal distribution of the data was confirmed by the Shapiro-Wilk test. Homogeneity of variances was rejected for ITC data and confirmed for ENM data using the Levene test. ITC data were therefore examined using a Dunnett's T3 post hoc test for pairwise comparisons with unequal variances and ENM data examined using a two-sided Dunnett's post hoc test for pairwise comparisons with equal variances.
CAP crystals were produced at pH 6.5 with 7–10% (w/v) polyethylene glycol 3350 and 15–20% (v/v) 2-methyl-2,4-pentanediol with 2 mM cAMP in 24-well hanging-drop vapour diffusion plates. Crystals were cryoprotected using mother liquor containing 30% (v/v) glycerol and flash cooled in liquid nitrogen [57]. Diffraction data for the wild-type protein were collected in-house using a Bruker MicroStar rotating anode and processed with SAINT [58]. All CAP mutant data were collected at the Diamond Light Source beams I-04 and I-24 and processed using Mosflm [59] and Scala [60]. CAP structures were solved using molecular replacement with Phaser [61] using CAP (PDB 1I5Z). Model building and refinement were accomplished iteratively using COOT [62] and Refmac5 [63] in CCP4 [59]. CAP structures from crystals produced at pH 6.5 were indistinguishable from those previously produced at pH 7.5 [64]. Structural and refinement statistics are provided in Table S4. Full details of GlxR crystallography and analysis of the structures will be reported elsewhere [45]. Members of the CAP family often crystallise with more than one protein chain in the asymmetric unit. In these cases the functional protein dimer is either generated by the crystallographic 2-fold axis on each of the protein chains or by noncrystallographic symmetry leading to a varying degree of asymmetry [65],[66]. Significantly different conformations for each monomer have been observed in some homodimeric bacterial regulator proteins, most notably Mt-CRP [67]. The structures presented here contain one dimer (wild-type CAP in space group P21), two dimers (wild-type in space group P1), and three dimers (V140A CAP in space group I2) (see Table S4). In all cases the dimers are symmetric with no significant differences between the two protein chains than for the functional dimer.
ENM simulations were performed using our own code based on the regular implementation [31],[68]. The spring constants were set to a constant value of 1 kcal mol−1 Å−2 with a cutoff radius of 8 Å, and only the Cα atoms in the protein were considered. The presence of cAMP effector at the binding site was treated by the addition of one node at the mass weighted average coordinate for each ligand. Varying the spring constant of any springs attached to a single residue of the protein was used to represent side chain mutations. The allosteric free energy was calculated by summing over modes 1 to n. n was determined by examining where values K2/K1 converged (Figure S8). The final results quoted used the converged value of K2/K1. PDB files for constructing CAP ENMs were 1CGP, 1G6N, 1HW5, 1I5Z, 1I6X, 1J59, 1O3T, 1RUN, 1RUO, 1ZRC, 1ZRD, 1ZRF, 2GZW, 4HZF (this work), and an additional in-house file isostructural to 2GZW. The PDB file for constructing the GlxR ENM was 3R6S.
The CAP and GlxR proteins were modelled as two blocks for each monomer, one for the ligand binding domain and one for the DNA binding domain. We assigned one internal breathing mode to each subunit and allowed each subunit to move, producing seven degrees of freedom. For the apo-protein the internal subunit coupling strengths are characterized by k1 though k4 and the intersubunit couplings by k12, k13, and k24 (Figure 4b). The effect of one ligand binding was included by modifying k1 by a factor β, k12 by α, and k12 by γ. The second ligand binding was therefore represented by further modifying k2 by β, k12 by a further factor of α, and k24 by γ. The allosteric free energy was determined from the determinant of the interaction matrix [69]. The couplings were defined from PCA analysis of 300 ns molecular dynamics simulations for the three states. In each case the protein was divided into the four zones by performing a rotational-translational-block approximation (Figure 8a) [46],[70]. Examination of the couplings calculated for each of the three states allowed calculation of the apo values and the ligand binding factors. Varying the values of k1, k2, and k12 represents mutations in residues affecting the intra- and the interblock interactions. Wild-type values for CAP are: k1 = k2 = 13.70, k12 = 27.08, k3 = k4 = 3.98, k13 = k24 = 5.19 kcal mol−1 Å−2, α = 1.30, β = 0.560, and γ = 0.901. Wild-type values for GlxR are: k1 = k2 = 12.85, k12 = 24.67, k3 = k4 = 3.98, k13 = k24 = 4.21 kcal mol−1 Å−2, α = 1.40, β = 0.71, and γ = 0.99.
Molecular dynamics (MD) simulations employed the harmonic force field equations used in the ff99SB and GAFF force fields within the AMBER simulation program [71].The simulations employed the ff99SB force field for the CAP protein and the GAFF force field (v. 1.4) for cAMP. ff99SB force field is used as the energetic interactions of side chains, which are reasonably represented by this force field [72], and outperforms the ff03 force field [73]. MD calculations used a short-range cutoff of 10 Å, with the long-range portion of the Coulomb potential represented by an Ewald summation, and employed a time step of 2 fs. The bond lengths were constrained by the SHAKE algorithm. The initial starting structures were obtained directly from X-ray diffraction. These structures were then solvated in TIP3P water and energy minimized prior to simulation [74]. The system was heated to 300 K over a period of 20 ps and further equilibrated for 40 ns. Production runs at 300 K were carried out over 200 ns. PCA was performed by diagonalising the mass weighted covariance matrix of the atomistic simulations. The eigenvectors represent the shape of the atomistic motion and the corresponding eigenvalues the extent of the motion.
To determine if ΔK2/K1, hereon denoted x, is associated with the mutation rate of amino acids, we first estimated the relative amino acid mutation rate using the sequence data for CAP variants and we then statistically tested for an effect of x on this rate. Relative mutation rate was estimated by finding the minimum number of amino acid mutations needed to generate the observed variations in the sequence data, which we denote N. For each of the 165 proteins we found the protein having the smallest number of amino acid differences. The sum of these differences gave N. When summing differences, if more than one protein had the minimum difference, we included all the proteins having the minimum. We then determined the number of these mutations that were associated with each of the 210 amino acids, which we denote ni. Thus, ni estimates the relative mutation rate of amino acid i, and these estimates account for the evolutionary history of the proteins. If all amino acids had an equal mutation rate, then we would expect the ni to all be approximated by N/210. We assumed that the true relative rate of mutation was related to x according to the logistic function: μ(x) = , where β0, β1, and β2 are constants. To account for overdispersion among the ni, which might be due to unmeasured covariates associated with the proteins, we assumed that the variation between the ni could be described by the beta-binomial distribution. Under these assumptions, the log-likelihood of the model described by the set of parameters θ = {β0,β1,β2,φ}, is given by:where BB(n|N,μ,φ) is the beta-binomial distribution, which describes the probability of observing n successes from N trials when, on average, successes occur with probability μ and variation in this probability among replicates is described by the beta-distribution with variance μ(1−μ)φ/(1+φ) [75].
Evidence that mutation rate was related to x was found by applying a likelihood ratio test (LRT) comparing the fit of the full model with the model that ignored x (i.e., when β1 = β2 = 0). Let LL1 and LL0 be the maximum log-likelihood of the full model and the simpler model, respectively. Under the null hypothesis that x is not associated with mutation rate, the test statistic G = 2[LL1−LL0] is chi-square distributed with two degrees of freedom, as the more complex model has two additional free parameters: β1 and β2. A LRT was also used to test for overdispersion by comparing the fit from the full model described above with the model that assumed variation had a binomial distribution (φ is vanishingly small). This latter test, if significant, justifies the use of the beta-binomial distribution rather than the binomial. Confidence intervals for model parameters were estimated using the likelihood profile approach.
The genome accession numbers analysed are: NP_232242.1, NP_246094.1, NP_249343.1, NP_439118.1, NP_458435.1, NP_462369.1, NP_671249.1, NP_716257.1, NP_760245.1, NP_799172.1, NP_873260.1, NP_927748.1, YP_052151.1, YP_089126.1, YP_128534.1, YP_152459.1, YP_205663.1, YP_237645.1, YP_262678.1, YP_272974.1, YP_455981.1, YP_492074.1, YP_526229.1, YP_564189.1, YP_588978.1, YP_606222.1, YP_690711.1, YP_693743.1, YP_718344.1, YP_751967.1, YP_855526.1, YP_928876.1, YP_941848.1, YP_960806.1, YP_001048976.1, YP_001092716.1, YP_001143048.1, YP_001178491.1, YP_001189422.1, YP_001218107.1, YP_001343325.1, YP_001440391.1, YP_001443362.1, YP_001464812.1, YP_001475605.1, YP_001503357.1, YP_001675803.1, YP_001759053.1, YP_001909102.1, YP_002152521.1, YP_002228709.1, YP_002294894.1, YP_002476451.1, YP_002650381.1, YP_002801694.1, YP_002875051.1, YP_002893931.1, YP_002923696.1, YP_002986005.1, YP_003002662.1, YP_003008634.1, YP_003039145.1, YP_003074496.1, YP_003255073.1, YP_003261368.1, YP_003469961.1, YP_003532766.1, YP_003555253.1, YP_003812150.1, YP_003914673.1, YP_004117516.1, YP_004211044.1, YP_004382110.1, YP_004391469.1, YP_004419866.1, YP_004472683.1, YP_004565203.1, YP_004713013.1, YP_004821770.1, YP_005091541.1, YP_005334361.1, YP_005458526.1, YP_005817463.1, YP_006006755.1, YP_006238931.1, YP_006286710.1, YP_006326252.1, YP_006459298.1, YP_006523113.1, YP_006588319.1, ZP_00134303.1, ZP_00991497.1, ZP_01161654.1, ZP_01215522.1, ZP_01815379.1, ZP_01894180.1, ZP_01898714.1, ZP_02478644.1, ZP_02958582.1, ZP_03319669.1, ZP_03611762.1, ZP_03825776.1, ZP_04636540.1, ZP_04640765.1, ZP_04752629.1, ZP_04977551.1, ZP_05043634.1, ZP_05637197.1, ZP_05774479.1, ZP_05849758.1, ZP_05879825.1, ZP_05880998.1, ZP_05919259.1, ZP_05972068.1, ZP_05990699.1, ZP_06018230.1, ZP_06051220.1, ZP_06126446.1, ZP_06542208.1, ZP_06637662.1, ZP_07161146.1, ZP_07222409.1, ZP_07266238.1, ZP_07379670.1, ZP_07395486.1, ZP_07528968.1, ZP_07744420.1, ZP_07777878.1, ZP_07888842.1, ZP_08039455.1, ZP_08068248.1, ZP_08079426.1, ZP_08100561.1, ZP_08148040.1, ZP_08310711.1, ZP_08519301.1, ZP_08725568.1, ZP_08731411.1, ZP_08745737.1, ZP_08754750.1, ZP_09013912.1, ZP_09039716.1, ZP_09185001.1, ZP_09505069.1, ZP_09557915.1, ZP_09710329.1, ZP_09778630.1, ZP_09972449.1, ZP_10075284.1, ZP_10125383.1, ZP_10128956.1, ZP_10135899.1, ZP_10142323.1, ZP_10146384.1, ZP_10426764.1, ZP_10438900.1, ZP_10622342.1, ZP_10628430.1, ZP_10630449.1, ZP_10643899.1, ZP_10655392.1, ZP_10677933.1, ZP_10700164.1, and ZP_10763153.1.
|
10.1371/journal.pcbi.1004451 | Quantifying Stochastic Noise in Cultured Circadian Reporter Cells | Stochastic noise at the cellular level has been shown to play a fundamental role in circadian oscillations, influencing how groups of cells entrain to external cues and likely serving as the mechanism by which cell-autonomous rhythms are generated. Despite this importance, few studies have investigated how clock perturbations affect stochastic noise—even as increasing numbers of high-throughput screens categorize how gene knockdowns or small molecules can change clock period and amplitude. This absence is likely due to the difficulty associated with measuring cell-autonomous stochastic noise directly, which currently requires the careful collection and processing of single-cell data. In this study, we show that the damping rate of population-level bioluminescence recordings can serve as an accurate measure of overall stochastic noise, and one that can be applied to future and existing high-throughput circadian screens. Using cell-autonomous fibroblast data, we first show directly that higher noise at the single-cell results in faster damping at the population level. Next, we show that the damping rate of cultured cells can be changed in a dose-dependent fashion by small molecule modulators, and confirm that such a change can be explained by single-cell noise using a mathematical model. We further demonstrate the insights that can be gained by applying our method to a genome-wide siRNA screen, revealing that stochastic noise is altered independently from period, amplitude, and phase. Finally, we hypothesize that the unperturbed clock is highly optimized for robust rhythms, as very few gene perturbations are capable of simultaneously increasing amplitude and lowering stochastic noise. Ultimately, this study demonstrates the importance of considering the effect of circadian perturbations on stochastic noise, particularly with regard to the development of small-molecule circadian therapeutics.
| As most organisms exist in an environment that changes predictably with a 24-hour period, highly optimized genetic circuits turn on and off the production of key regulatory proteins to anticipate the day/night cycle. In humans, the demands of a modern society have required that we deviate from this evolutionarily prescribed sleep and feeding schedule, resulting in increased long-term risks of metabolic disease. There is therefore a desire to find pharmacological treatments that would restore the normal functioning of our circadian clock despite irregular behavioral schedules. One aspect of these treatments that is often overlooked in searching for candidate drugs is how these treatments might affect the accuracy of the circadian timing system. Recording the time of each cell is possible but difficult; as a result single-cell approaches cannot be scaled up to high-throughput searches. In this paper, we show that it is possible to estimate how much the noise of a system has changed by looking only at the averaged protein production of an entire population of cells. Such an approach allows us to analyze prior data from high-throughput screens, and show that the natural clock has been highly optimized to be both accurate and high amplitude.
| Circadian rhythms are daily changes in gene expression and physiology that persist even in the absence of external environmental cues [1]. In mammals, such rhythms are organized in a hierarchical fashion: at the tissue-level, the brain’s suprachiasmatic nucleus (SCN) serves as the master pacemaker and keeps circadian oscillations in peripheral tissues in phase with the light-dark cycle. In the SCN, cell-to-cell coupling keeps individual cells in tight synchrony [2], while coupling between circadian oscillations in peripheral tissues in vivo or cultured reporter cells in vitro is thought to be very weak or absent entirely [3, 4]. Within each tissue, cellular-level rhythms in gene transcription are generated by a large network of interacting gene regulatory elements, in which time-delayed transcription-translation negative feedback gives rise to sustained oscillations [5]. The robust oscillation of circadian factors has been linked to metabolic health [6], since rhythms compromised by gene knockout [7] or irregular feeding schedules [8] result in an increased risk of metabolic disease. Additionally, as the amplitude of circadian transcription can be affected by lifestyle variables such as diet, age, or work schedule, there has been recent interest in developing pharmacological strategies for increasing the amplitude of circadian cycles in metabolic tissues [9].
A detailed understanding of the underlying transcriptional mechanisms is essential for the development of circadian therapeutics to be successful. The functional roles of different genes in circadian regulation have traditionally been studied using behavioral-level data and genetic knockout experiments [10]. Bioluminescence-based cellular circadian reporters offer a more direct view of the gene regulatory network [11] and are amenable to high-throughput screens, allowing genome-wide exploration into factors that affect circadian rhythmicity [12]. Additionally, cultured circadian reporter cells allow the change in transcriptional amplitude following a perturbation to be quantified. This additional parameter has proven useful in differentiating between perturbations with the same effect on period [13] and has aided the search for small-molecule therapeutics to boost clock amplitude [9].
Bioluminescence rhythms at the cell culture or tissue-level are the result of the collective behavior of thousands of individual cells. Transcription at the single-cell level is strongly affected by intrinsic cellular noise, caused by the low molecular counts of the mRNA and protein species involved. As a result, bioluminescence traces of individual cells are stochastic, with significant variability in both amplitude and period length from cycle to cycle [14]. In addition to intrinsic noise, circadian oscillations are also affected by extrinsic noise sources. Extrinsic noise results from heterogeneity between cells, such as differences in size or physical environment, leading to differences on a cell-to-cell basis in expected period and amplitude. The effects of noise in biological systems has been well-studied, and relative amounts of intrinsic and extrinsic noise can be identified through carefully designed single-cell experiments [15]. For circadian systems, intrinsic noise has been suspected to play a larger role: a single cell’s variability in period from cycle-to-cycle is larger than the variability in mean period length between cells [2]. However, both sources of noise have an effect on population-level rhythms: in cell cultures that lack cell-to-cell coupling, it has been shown that stochastic noise is manifested in damped oscillations at the population-level as individual oscillators gradually drift out of phase [14, 16]. This type of behavior has also been seen in other experimental systems, such as NF-κB signaling or yeast glycolytic oscillations [17, 18]. The amount of noise in system is therefore linked to the ability of tissue-level clocks to maintain high-amplitude rhythms.
Despite the averaging that occurs at the population-level, cell-autonomous stochastic noise plays an important role in determining the function of the circadian oscillator. Noise in circadian rhythms has long been considered an important factor in how circadian rhythms have evolved [19]. A recent study showed that stochasticity is critical to the population-level response to a neuropeptide and forms the basis for how the SCN entrains to light-mediated cues [20]. Additional studies have suggested that the basis of single-cell rhythmicity may depend on stochastic noise, as models of deterministically damped oscillators, when simulated stochastically, capture the noise characteristics seen in single-cell fibroblast data equally well as limit-cycle oscillators [21]. Despite the importance of single-cell stochasticity in circadian rhythms, measuring stochastic noise currently requires careful preparation, recording, and image processing of individual cells [22]. As a result, while circadian perturbations have been postulated to affect single-cell stochasticity [23], no study has experimentally quantified changes to stochastic noise as a result of a small molecule or genetic perturbation.
In this study, we demonstrate that changes to stochastic noise can be reliably inferred from the changes to the damping rate of population-level bioluminescence recordings of cultured circadian reporters. Our method assumes that oscillations in individual cells are both
which have been shown to hold experimentally for cultured fibroblast cells. We demonstrate the validity and usefulness of such an approach on several types of circadian data. First, we show using single-cell fibroblast data that intrinsic stochastic noise is directly related to population-level damping. Next, we show that a small-molecule modulator is able to change damping rate in a dose-dependent fashion, and verify using a mathematical model that changes to intrinsic stochastic noise is a likely mechanism. Finally, we calculate the genome-wide effects of siRNA knockdown on overall stochastic noise, and demonstrate that population-level damping rate is independent of other circadian parameters, such as period or amplitude. Using this additional information, we show that circadian rhythms have likely evolved to an optimal point of high amplitude and low stochastic noise. Our results should prove especially important in the future search for small molecule circadian therapeutics, as it allows the effect of candidate drugs on stochastic noise to be quantified in a high-throughput manner.
Raw experimental data x(ti), i ∈ {0, …, N − 1} are first detrended using Hodrick-Prescott filter with a smoothing parameter γ = 0 . 05 ( 24hrs T s ) 4, in which Ts is the sampling rate (in hours) [24]. The detrended data are then filtered using a low-pass filter to remove high-frequency noise (forward-backward Butterworth filter with n = 5, wc = 0.1). We denote the detrended and filtered experimental data by y(ti). A damped sinusoid, specified by:
y ^ ( t ) = A e - d t sin ( 2 π t T + θ )
is then fit to the filtered data. For numerical efficiency, the period, T, and damping rate, d, parameters are fit first using a matrix pencil method [25], reviewed in [26]. Amplitude, A, and phase, θ, parameters are subsequently fit using a linear least-squares regression. Note that in this manuscript we use amplitude to denote the initial rhythm strength, and damping rate to denote the rate at which this strength decays with time. Overall R2 values for the regression were calculated from the residual error between the detrended data and fitted sinusoid:
R 2 = 1 - ∑ i = 0 N - 1 ( y ( t i ) - y ^ ( t i ) ) 2 ∑ i = 0 N - 1 ( y ( t i ) - y ¯ ( t i ) ) 2
in which y ‾ ( t i ) represents the mean of the detrended data.
Single-cell bioluminescence data for 79 cells was obtained from Leise et al., 2012 [22]. As was done in the original study, a discrete wavelet transform (using PyWavelets, http://www.pybytes.com/pywavelets) was performed to detrend and remove noise. A discrete wavelet transform decomposes the signal into multiple frequency bands [27]. By only considering frequency bands close to the circadian frequency, high-frequency noise and low-frequency baseline oscillations can be removed.
Bioluminescence traces (Ts = 1.67, N = 71) with increasing small molecule concentrations were fit with a damped sinusoid using the method described in a previous section. Because the small molecules were toxic at very high concentrations, experiments were removed from further analysis where R2 < 0.80 (S1 Fig).
A previously published mathematical model of circadian rhythms [28] was used to predict the effects on population damping rate from the dose-dependent small molecule experiments. The parameters used to capture the effects of each small molecule were the same as described previously [13]. The model was converted to a stochastic biochemical system and subsequently simulated using StochKit2 [29] (via GillesPy, https://github.com/JohnAbel/gillespy). Population-level rhythms were found by taking the average of 1,000 noninteracting oscillators, starting at identical initial conditions. The only parameter left unspecified by the deterministic model was the cell volume, Ω, which controlled the amount of noise in the system. For each Ω, a damped sinusoid was fit to the population-averaged state trajectory. An R2 value was calculated for each fit, taking into account all eight state variables.
We analyzed the data and annotations for the 111,743 wells (Ts = 2, N = 72) in the Zhang et al., 2009 screen [12]. Fits for which the R2 < 0.80 were discarded. The natural logarithm of the amplitude was used, since it more closely resembled a normal distribution and was on a similar scale to the damping rate. Plate to plate variation, as shown in S3 Fig, was more severe than variation on a well-to-well basis, S4 Fig. Parameters were therefore normalized on a plate-by-plate basis using a robust z-score [30]:
z R , i = p i - M ( p i ) M ( | p i - M ( p i ) | ) , i ∈ { 0 , ⋯ , P - 1 }
where M(⋅) denotes the median of a vector, and pi contains all the points in one plate, and 𝒫 is the number of plates. We removed outlier points prior to calculating the moments of the distributions, Pearson correlation coefficients, and performing the multivariable linear regression. Outliers were defined as points that contained a z-metric (in either period, phase, amplitude, or damping rate) with an absolute value greater than eight. We chose the “control” wells to be those that contained no siRNA, as these proved to be more numerous than those containing reference siRNA perturbations and were clustered similarly to the highest-density regions of the perturbed fits.
All computations were performed using Python. Code used to perform the analysis and produce the figures in this manuscript can be found online at https://github.com/pstjohn/decay-methods.
While both intrinsic and extrinsic noise sources can contribute to population-level damping, intrinsic noise is thought to play a more significant role in circadian systems [2]. We therefore first sought to determine whether changes to intrinsic stochastic noise alone are sufficient to explain population-level changes in damping rate. To do this, we calculated noise characteristics from experimental data on individual PER2::LUC fibroblast cells [22]. Cells were sorted into two groups, a low-noise group and a high-noise group, based on the relative high-frequency noise, period variability, and amplitude variability present in each trace. Example rhythms from cells in both groups are shown in Fig 1A. Because the cells were not synchronized at the start of the recording, this effect is replicated in silico by shifting each series in time to align their start phases. Population-level bioluminescence traces were then found by averaging the cellular PER2::LUC signal in each group. Both populations displayed averaged rhythms that resembled a damped sinusoid, similar to those seen in bioluminescence recordings of entire cell cultures. Fitting the averaged expression of each group with a damped sinusoid revealed that the low-noise group also had a lower damping rate (Fig 1B). The significance of this difference was confirmed via a bootstrap analysis (Fig 1C), where cells were randomly assigned in each bootstrap trial to either the low-noise or the high-noise group.
Unlike from using single-cell imaging, inferring stochastic noise from the desynchronization rate of population-level recordings can be applied to existing and future high-throughput circadian screens. We demonstrate the insights that can be gained from such an approach by analyzing the publicly available genome-wide siRNA screen from Zhang et al., 2009 [12]. The results of fitting a damped sinusoid to each of the 111,743 bioluminescence trajectories is shown in Fig 4, in which 86% of fits had an R2 > 0.8. Since sinusoidal parameters can only be confidently inferred for fits in which the R2 was sufficiently high, wells were removed from further analysis if R2 < 0.8. Additionally, of the fits with a high R2 value, only a small minority (0.1%) had a negative damping rate. This trend supports the assumption that intercellular synchronization is unlikely in cultured U2OS cells.
We next checked how parameters varied on a plate-to-plate and well-to-well basis. Well-to-well variation was relatively absent, aside from expected variation due to long- and short-period controls (S4 Fig). Fits were normalized to remove plate-to-plate variation (S3 Fig) using a robust z-score [30]. Additionally, we separated wells into a “perturbed” category and “control” category, depending on whether or not the well contained an siRNA perturbation. As we show in Fig 5, all fitted parameters displayed normal-like distributions, in which the control distributions showed much tighter clustering around the most likely values. Quantifications of the mean, variance, skewness, and kurtosis for each distribution are shown in Table 1.
In the preceding sections, we have demonstrated that changes to single-cell intrinsic noise are sufficient to explain the observed changes in population-level damping. However, experimental work has shown that cell-autonomous fibroblast cells have a distribution of mean free-running periods [22]. Indeed, prior to the availability of single-cell data, studies explored the possibility that differences in mean periods served as the mechanism behind population-level damping [31]. While it is true that the dephasing of a group of oscillators can be caused by both variance in the mean period as well as cycle-to-cycle variability, intrinsic stochastic noise may play a more significant role. We show that there is greater variance in period on a cycle-to-cycle basis than on a cell-to-cell basis in cultured fibroblast cells (S5 Fig): individual cell period lengths have an average inner quartile range (IQR) of 3.18 hrs, while cell-to-cell average period has an IQR of 1.55 hrs. These results are independently confirmed by a previous study using Bayesian modeling, which found a standard deviation of periods within cells of 1.43 hrs, and 0.89 hrs across the population [32]. A similar result has also been observed in dispersed SCN neurons, suggesting that while both intrinsic and extrinsic period heterogeneity likely contribute to the dephasing kinetics, cell-to-cell differences are less severe than cell-autonomous noise [2].
It is also possible that damping rate changes due to siRNA or small molecule perturbation could be manifested through altering the system’s extrinsic noise. Such a change could be caused by an unequal uptake of siRNA or drug on a cell-to-cell basis, as has been demonstrated by a distribution of single-cell knockdown efficiency through flow cytometry [33]. This effect would increase cell heterogeneity and lead to faster dephasing kinetics. While differentiating between intrinsic and extrinsic noise sources from population-level data is possible in theory (S6 Fig), these differences are likely not identifiable from typical population-level data (S7 Fig). This distinction would likely be possible with single-cell level data, as has been done in other experimental systems [15]. However, such an experiment would likely not be amenable to high-throughput methods. Differences in damping rates are therefore best viewed as representative of changes to overall stochastic noise from both intrinsic and extrinsic factors. Since both types of noise are important to determining the overall function of population-level rhythms, damping rates are still a valuable method of quantifying stochastic noise.
In this study we have shown that the damping rate of population-level circadian oscillations can be changed by genetic or pharmacological perturbations. As populations-level rhythms are determined by the coherence of many individual cells, desynchronization due to stochastic noise is a likely explanation for population-level damping. Using single-cell data, we showed that population-level damping rate is proportional to single-cell noise. Furthermore, we used a computational model to predict the changes in damping rate from two small molecules, demonstrating that changes to intrinsic stochastic noise are sufficient to explain the observed damping rate changes.
We have described a method by which changes to stochastic noise can be estimated from population-level circadian bioluminescence recordings. An overview of the computational steps involved in our method are outlined in S8 Fig. While the method can be applied to existing experimental data, there are also practical considerations for the design of future experiments. Because the damping rate must be inferred from the time-varying amplitude, collecting bioluminescence data for longer time periods yields more accurate results and reduces the potential impact of initial transient regions. Additionally, a sampling rate that is high enough to confidently capture the peaks and troughs of gene expression is required—in this study, the 2 hr sampling window of the siRNA screen proved sufficient. While achieving such a rate is typically not difficult for bioluminescence experiments, it may limit the method’s applicability in experiments where samples need to be analyzed at each time point. We also note that there are many available tools for detrending and regressing time-series data. While we prioritized computationally efficient methods (which could be scaled to genome-wide screens), the best methods for any particular application may vary depending on the data.
As high amplitude circadian oscillations are important for maintaining metabolic health, many studies have sought to find small molecule candidates that increase oscillatory amplitude [9]. In the search for circadian clock therapeutics, high-throughput methods are frequently used to screen for such drugs, often neglecting potential effects on cell-autonomous noise. While we have demonstrated a method by which the effects of small molecules on noise can be inferred from high-throughput methods, we have also shown that the potential improvement of clock robustness may be limited. Amplitudes of circadian rhythms may therefore be best increased by small molecule therapies that act transiently to synchronize peripheral oscillators. While such a method would require accurate alignment of drug administration to the correct circadian phase, a recent in silico study has demonstrated the potential effectiveness of such an approach in improving amplitudes in peripheral tissues [34]. Finally, the ability to extract an additional biologically relevant parameter from existing datasets will likely prove useful for many studies, as it allows further differentiation between perturbations that might otherwise have identical effects on bioluminescence rhythms.
|
10.1371/journal.pcbi.1004023 | The Thalidomide-Binding Domain of Cereblon Defines the CULT Domain Family and Is a New Member of the β-Tent Fold | Despite having caused one of the greatest medical catastrophies of the last century through its teratogenic side-effects, thalidomide continues to be an important agent in the treatment of leprosy and cancer. The protein cereblon, which forms an E3 ubiquitin ligase compex together with damaged DNA-binding protein 1 (DDB1) and cullin 4A, has been recently indentified as a primary target of thalidomide and its C-terminal part as responsible for binding thalidomide within a domain carrying several invariant cysteine and tryptophan residues. This domain, which we name CULT (cereblon domain of unknown activity, binding cellular ligands and thalidomide), is also found in a family of secreted proteins from animals and in a family of bacterial proteins occurring primarily in δ-proteobacteria. Its nearest relatives are yippee, a highly conserved eukaryotic protein of unknown function, and Mis18, a protein involved in the priming of centromeres for recruitment of CENP-A. Searches for distant homologs point to an evolutionary relationship of CULT, yippee, and Mis18 to proteins sharing a common fold, which consists of two four-stranded β-meanders packing at a roughly right angle and coordinating a zinc ion at their apex. A β-hairpin inserted into the first β-meander extends across the bottom of the structure towards the C-terminal edge of the second β-meander, with which it forms a cradle-shaped binding site that is topologically conserved in all members of this fold. We name this the β-tent fold for the striking arrangement of its constituent β-sheets. The fold has internal pseudosymmetry, raising the possibility that it arose by duplication of a subdomain-sized fragment.
| In the public perception, thalidomide mainly evokes children with stunted limbs. Less known is that thalidomide continues to be a very useful drug, licensed in most countries for the treatment of multiple myelomas and leprosy. Aside from its catastrophic effect on human embryonal development, it has a manageable spectrum of side-effects and a broad range of potential indications. Interest in its further pharmacological development thus remains high, but is hindered by our limited knowledge of the reasons for its worst side-effect – teratogenicity. For half a century, even the main protein target of thalidomide in the human body remained unknown, until a seminal study showed in 2010 that this was cereblon. Further progress towards a mechanistic understanding has however been limited by the difficulties in using cereblon for biochemical studies. Here we show that the thalidomide-binding region of cereblon is contained within a domain also present in other protein families, and that this domain is related to several domains with known functions. Where established experimentally, all these domains are seen to form their main substrate-binding sites at the same location in the common fold, often using residues in equivalent positions. Our findings offer the possibility to develop model systems for the study of specific aspects of cereblon activity.
| Thalidomide was provided to pregnant women as an anti-nausea and sedative drug from 1957 to 1962, and was available over the counter in many countries. It was withdrawn after it became apparent that it had caused a range of birth defects in many newborns, with over 10,000 cases reported from more than 46 countries [1], [2]. Soon after its ban, however, it was reintroduced as an agent against a complication of leprosy [2]–[4], due to its anti-inflammatory and immunomodulatory activity, and has since then also been evaluated for treatment of, among others, AIDS and Crohn's Disease [5]. In 1994, evidence of its antiangiogenic activity led to its consideration for cancer therapy [6] and it is one of the main drugs available today against multiple myeloma [7], [8]. Despite strict controls on its use, its value in the treatment of leprosy leads to the ongoing birth of babies with thalidomide-induced malformations in developing countries [2], [9].
Because of its antiangiogenic and immunomodulatory activities, pharmacological interest in thalidomide continues to be very high [5], but until recently, its molecular mechanism of action – both with respect to its positive and its negative effects – remained unclear due to a lack of known targets. In 2010, Handa and co-workers showed in a landmark study that cereblon, a protein originally identified in a screen for mutations causing mild mental retardation [10], is a major target of thalidomide and is responsible for the teratogenic effects of the drug [11]. Cereblon owes its name to its involvment in brain development and to its central LON domain, which led to its initial annotation as an ATP-dependent Lon protease [10]. In the 2010 study, Handa and co-workers showed that cereblon is a cofactor of damaged DNA-binding protein 1 (DDB1), which acts as the central component of an E3 ubiquitin ligase complex and regulates the selective degradation of key proteins in DNA repair, replication and transcription [12]. Binding of thalidomide to a C-terminal region in cereblon inhibits the E3 ubiquitin ligase activity of the complex and leads to developmental limb defects in chicks and zebrafish [11]. Point mutations in this region, which abolish thalidomide binding, but allow the continued formation of the E3 complex, restore ubiquitination and prevent the teratogenic activities of thalidomide.
Despite these advances, there has been little progress in understanding the mechanism of thalidomide binding and teratogenicity, due to the difficulties in preparing cereblon protein for biochemical and biophysical studies. Such progress would however be important for further pharmacological development, given that current thalidomide derivatives, such as pomalidomide and lenalidomide, appear to have inherited its teratogenicity (see e.g. [13]). In search of a better understanding of cereblon, we decided to subject the protein to a detailed bioinformatic analysis, with a particular focus on its thalidomide-binding domain. Here we show that this domain is present in several protein families of eukaryotes and bacteria, is related to the highly conserved yippee and Mis18 proteins of eukaryotes, and has a homologous origin with methionine sulfoxide reductase B, the regulatory domain of RIG-I helicase, and glutathione-dependent formaldehyde-activating enzyme. Our findings place the domain into a broad evolutionary context and show that the development of model systems is possible in order to study specific aspects of cereblon activity.
Cereblon proteins occur throughout eukaryotes, however not in fungi. They are typically 400–600 residues long (442 in the case of human cereblon) and their genes occur in single copy per genome. Their most salient feature is the presence of a central LON domain (residues 80–317 in human cereblon, Fig. 1). As defined in the Pfam and SMART databases, the LON domain actually comprises two domains, an N-terminal pseudo-barrel of six β-strands, closed off on one side by a helix, (LON-N) and a helical bundle of four to five helices – four in the case of cereblon, to judge by secondary structure prediction and length of the domain (LON-C).
The two LON domains are connected by an unstructured loop of typically around 10 residues, which however is much longer in cereblon, at about 60 residues. Handa and co-workers identified this region by deletion analysis as responsible for DDB1 binding (Fig. 1) [11]. Since an α-helical motif – the H-box – has been found to be a crucial structural element used by both viral and cellular substrate receptors to bind to DDB1 [14], we surmised that an H-box exists in cereblon as well, in the segment connecting LON-N with LON-C. H-box sequences are however very divergent and the H-box is thus primarily defined by its helical propensity and a general pattern of hydrophilic and hydrophobic residues [14]. We tried to identify an H-box in cereblon, and particularly in the connector between the two LON domains, by searching against a profile HMM generated from the H-box sequences listed by Li et al. [14], but did not obtain statistically significant matches. The connector is poorly conserved across phyla, but contains a highly conserved motif WPxWxYxxYD immediately prior to the start of the LON-C domain (Fig. 1). Since this motif coincides with a region of elevated helical potential, we considered this the best guess for the site of DDB1 interaction. After submission of this manuscript, two studies presented the structure of cereblon in complex with DDB1 [15], [16], which showed that the interaction between the two proteins is topologically novel and not mediated by an H-box. All three regions with elevated helical propensity in the connector (Fig. 1) are indeed helical, albeit the first one as a 310 helix. The main interactions are formed between the first region and DDB1 β-propeller A, and between the third region (containing the conserved motif) and DDB1 β-propeller C.
N-terminally to the LON domain, cereblon proteins have an extension of typically 50–100 residues (79 in the case of human cereblon), of which the front part is not conserved across phyla and generally predicted as intrinsically unstructured (Fig. 1). Starting about 35 residues prior to the beginning of the LON domain, the extension becomes well-conserved and particularly a motif FDxxLPxxHxYLG is recognizable in most cereblon homologs, from humans and plants to basal eukaryotes. Since the extension is adjacent to the connector between LON-N and LON-C in the LON domain model (Fig. 1) and experimental evidence suggested the binding interactions with DDB1 to be bipartite [14], we considered that this motif could contribute to DDB1 binding. The cereblon-DDB1 complex structures however show that the extension interacts with the C-terminal region of cereblon via the conserved motif [15], [16]. Indeed, since it runs alongside the thalidomide-binding site, it may relay information on the occupancy of the site to the LON-N domain. We therefore conjecture that deletion of the extension could uncouple the C-terminal region from the rest of the E3 ligase complex, alleviating or even entirely abolishing the effects of thalidomide binding.
The C-terminal region of cereblon, comprising about 100–130 residues (125 in the case of human cereblon), represents the best-conserved part. It has multiple invariant residues and encompasses completely the part of the protein identified through deletion analysis as responsible for thalidomide binding [11]. Sequence similarity searches show that this region also occurs in much shorter proteins than cereblon, where it essentially covers the entire length of the protein, identifying it as a domain. We name this domain CULT, for cereblon domain of unknown activity, binding cellular ligands and thalidomide.
With the CULT domain of human cereblon as a starting point, PSI-Blast searches of the non-redundant protein database at NCBI converge in three iterations. Analysis of the results, for example using clustering by pairwise sequence similarity in CLANS (Fig. 2), shows that most of the search space consists of cereblon sequences, recognizable by their LON domain, but that several other groups of CULT domain-containing proteins are identifiable. The two main groups are: (I) prokaryotic proteins, mainly from δ-proteobacteria, but with a few representatives from α- and γ-proteobacteria and one sequence from a spirochete; these proteins consist entirely of the CULT domain; and (II) animal proteins from placozoans to vertebrates, but not occurring beyond fishes; these proteins also consist of the CULT domain, but carry an N-terminal secretion signal sequence. Indeed, the homolog from the sand fly Phlebotomus arabicus has been identified experimentally as a salivary protein [17]. Two further, more divergent groups are also apparent: one from oomycetes, with an N-terminal signal sequence followed by a CULT domain and ending with a carbohydrate-binding domain (SCOP: b.64); and the second from kinetoplastids, with an N-terminal CULT domain followed by a C-terminal region that cannot be assigned to a known domain family at present. The remaining few sequences from the PSI-Blast search do not recognizably belong to any of these groups; they are mainly from green algae and can all be confirmed by reverse PSI-Blast searches to contain a CULT domain.
Multiple alignment of the sequences identified in the search (Fig. 3A) show that several residues are highly conserved in the CULT domain. Conservation of these residues is highest in the CULT core group (cereblon, secreted eukaryotic sequences, bacterial sequences) and declines towards the periphery. Particularly conspicuous are three tryptophan residues, marked by arrows in Fig. 3A, which are seen to form the binding site for thalidomide and cellular ligands in the crystal structure of the bacterial CULT protein MGR_0879 from Magnetospirillum gryphiswaldense ([18]; PDB ID:4V2Y; Figs. 3B,C). The crystal structure, which we determined after this bioinformatic study, shows that the other highly conserved residues group around this binding site (S1 Fig.), their conservation being rationalized by an influence on substrate recognition and discrimination. The one exception to this are two CxxC cysteine motifs, which we took from the beginning of this project to be indicative of a zinc binding site and thus present for structural reasons.
We developed the bacterial model system to study the CULT domain because we found eukaryotic cereblon proteins very difficult to express in useful amounts and even more difficult to purify in a soluble state. In contrast, the bacterial protein could be produced and purified in a straight-forward way [18]. We reasoned that, at 36% sequence identity and with almost all well-conserved positions similar or the same between the CULT domains of humans and Magnetospirillum, the bacterial system should represent an accurate model for the eukaryotic domain. The structures of eukaryotic CULT domains from human, mouse and chicken [15], [16] now show that the expectation is true to an astonishing extent, with a root-mean-square deviation (r.m.s.d.) of around 0.9 Å over 100 Cα positions between the bacterial and human proteins (Fig. 3B). The bacterial domain is thus structurally almost as similar to the eukaryotic domains as these are to each other (Table. S1).
We searched for remote homologs of the CULT domain using profile Hidden Markov Model (HMM) comparisons in HHpred and obtained matches at probabilities better than 90% (E values <1e-6) for multiple protein families, several of which have members of known structure (Fig. 4). The best matches were to a protein family found throughout eukaryotes, yippee [19]. Two of the five yippee paralogs in mammals have been implicated in signal transduction [20] and tumor suppression [21], respectively, but the actual mechanism of these proteins remains unknown. In searching against Pfam we noticed that, in this database, the profile for yippee (PF03226) was generated jointly with another protein, Mis18, which in our analyses is not particularly close to yippee and indeed seems about as remote from yippee as cereblon is (Fig. 5). Mis18 proteins are broadly represented in eukaryotes, except plants, and appear to be involved in centromere assembly [22], [23], although their actual mechanism remains unknown. The reason for merging Mis18 with yippee in Pfam is unclear to us.
Two protein families of known structure related at a similar level to the CULT domain as yippee and Mis18 are methionine sulfoxide reductase B (MsrB or SelR) and the regulatory domain of retinoic acid-induced gene-1 (RIG-I). MsrB is the most widely distributed protein in this study and is universal to all cellular life. It protects cells from oxidative stress by reducing methionine-R-sulfoxide residues (for a review see e.g. [24]). RIG-I has a more limited phylogenetic spectrum, being detectable only in animals. It is an RNA helicase that, upon binding viral RNA, activates the host innate immune system (for a review see e.g. [25]). The regulatory domain is the RNA 5′-triphosphate sensor of RIG-I, activating the ATPase activity of the protein by RNA-dependent dimerization [26]. These four proteins, yippee, Mis18, MsrB, and RIG-I, are sufficiently close to the CULT domain in sequence space that they usually show up in the non-significant part of sequence similarity searches, between E values of 0.005 and 10, and are occasionally included in the significant part as well. Thus, for example, PSI-Blast searches of the nr database with our bacterial model protein, Magnetospirillum MGR_0879, include the first yippee and RIG-I sequences in the second iteration and the first Mis18 and MsrB sequences in the fifth. These proteins appear roughly equidistant from cereblon in sequence space (Fig. 5). More distantly related, but showing up with fair regularity in our searches is glutathione-dependent formaldehyde-activating enzyme (GFA), a protein found in bacteria and most eukaryotes, except plants. GFA catalyzes the first step in the detoxification of formaldehyde [27].
All these proteins share a common fold, formed by two four-stranded, antiparallel β-sheets that are oriented at approximately a right angle and pinned together at their tip by a zinc ion (Fig. 6). The two sheets are connected covalently across the top on both sides by loops, due to circular permutation. Thus, the last strand of the domain is topologically the first strand of the first sheet, yielding the strand order β8-β1-β2-(β3) for the first sheet and (β4)-β5-β6-β7 for the second (β3 and β4 are shown in brackets as, in some structures, they have lost their β-strand character). Because of the striking arrangement of these β-sheets we have named this fold the β-tent.
A conserved feature of all proteins with a β-tent fold is an insertion between strands β2 and β3, which usually has a β-hairpin stem and reaches across the bottom of the tent to extend the second β-sheet at its C-terminal edge. Due to the curvature of the β-sheet and the sizable nature of the loops connecting β4 to β5 on one side and the strands of the insertion on the other, all β-tent proteins contain a cradle-shaped groove at this location, which hosts the binding site (Fig. 7).
The residues giving the binding site its specificity in the individual proteins are frequently found in equivalent positions. This is particularly conspicuous when comparing the binding sites of CULT and MsrB (Fig. 8). Of the four residues forming the thalidomide-binding site in Magnetospirillum CULT (4V2Y: W79, W85, W99, and Y101), the last three have equivalents in homologous positions in the methionine sulfoxide-binding site of MsrB (3HCI: R97, H111, F113); the first, W79, is also a tryptophan in the MsrB binding site, but from an analogous position in the insert loop (W73), due to a shift in the position of the site caused by the shape difference between W85 of CULT and R97 of MsrB. This shift places the ligand above β7 in MsrB, rather than above β6, allowing the positioning of a further residue into the active site, which is the catalytic cysteine; conversely, there appears to be no need for a catalytic residue in CULT. We note that the homology of the conserved aromatic residues in CULT to the residues of the binding site in MsrB can be readily seen from the HHpred alignment.
Extending these observations to yippee, which has a similar distribution of conserved residues as CULT (S1 Fig.), we predict that the binding site of this protein is also an aromatic cage, comprising the highly conserved Y43, F45, W82, and Y84, as numbered in D. melanogaster yippee isoform B, ABC67182.1 (Fig. 7). Of these, W82 and Y84 are in homologous positions to W99 and Y101 of CULT and H111 and F113 of MsrB, whereas Y43 and F45 are at the same position as W73 in MsrB, but not recognizably homologous.
A striking property of the β-tent fold is that, in several of the proteins, the two sheets have considerable structural symmetry, such as for example in the MsrB structure 3HCJ, where superposition of the two 43-residue halves yields an r.m.s.d. of 1 Å over the Cα positions of the core 30 residues (Fig. 6). This raises the possibility that the fold originated by duplication of a subdomain-sized fragment, but we note that no similarity is detectable between the two halves by sequence comparisons.
Searches in structure space for other proteins with the β-tent fold yielded three more proteins of known structure (Figs. 6, S2), which share the fold with the same topology of secondary structure elements, including the β-hairpin extension between strands β2 and β3, but have no significant sequence similarity to the other proteins in this study, or to each other (Fig. 4). These are MSS4, a guanine exchange factor and nucleotide-free chaperone for the Rab GTPase [28], [29], TCTP, a pleiotropic protein involved in malignant transformation and regulation of apoptosis [30]–[32], and DUF427, a domain of unknown function. Whereas MSS4 and TCTP are eukaryotic proteins, TCTP being present universally and MSS4 broadly, but not in plants, DUF427 is seen mainly in bacteria and fungi, with a small number of archaea presumably having acquired this domain by lateral transfer. Of these proteins, only MSS4 has the zinc binding site (Fig. 6). In the SCOP database, MSS4 and TCTP are grouped together with MsrB and GFA as families within the MSS4-like superfamily, which is the sole representative of the MSS4-like fold (b.88).
A fundamental issue in understanding the biological role of CULT domains, not directly illuminated by their homology to other proteins, is the identification of their physiological ligand(s). The only ligands known today, thalidomide and its derivatives, are clearly non-physiological. Given that the clustering of the invariant tryptophans into a cage-like arrangement was already suggested at the modeling stage (see above), we searched PDB for ligands bound in aromatic cages, loosely defined. For this we allowed the aromatic residues to be Phe and Tyr, as well as Trp, and provided only a very general requirement for cage-like geometry, in order to gain as broad a view as possible (see Methods). We obtained 1098 distinct ligands, which could be grouped approximately into five classes, corresponding to heterocyclic rings, hydrocarbon rings, hydrocarbon chains with and without heteroatoms, and ammonium-based cations (Fig. 9, Table S2). Upon inspection, many of the “cages” identified indeed turned out to be only approximately cage-like and for 46 ligands, all binding sites turned out to be geometrically too divergent to be considered further.
Half of the identified ligands belonged to the largest class, comprising heterocyclic rings. Many of these were enzyme inhibitors, both of natural and synthetic origin, such as indole-2,3-diones (4KWG), aryl hydrazines (4MQQ), or non-nucleoside reverse transcriptase inhibitors (1S9G). Thalidomide, which is bound in the aromatic cage of CULT domains via its glutarimide ring, belongs to this class. Among the natural compounds, we found pyrimidines and their nucleosides of particular interest, as these resemble the glutarimide ring of thalidomide [18] and are bound in similar cages. For example, the transcription factor RutR (3LOC; Fig. 9B) can bind both uracil and thymine in its aromatic cage and acts as the master regulator of genes involved in the synthesis and degradation of pyrimidines [33]. An experimental screen against Magnetospirillum MGR_0879 found that, of the nucleobases, uracil and its nucleoside (uridine) were indeed bound and their relevance for eukaryotic cereblon could be established in vivo in zebrafish. It is attractive to consider that they might also be the physiologically relevant ligand, given that DDB1 is an integrator of cellular information on DNA damage and incorporation of uracil into DNA represents a mutagenic lesion [18].
Another type of ligand that we found to be of particular interest in this analysis comprises amino acid sidechains, modified and unmodified. These include the heterocyclic rings of His, Pro, and Trp, the hydrocarbon rings of Phe (Fig. 9C) and Tyr, the hydrocarbon chains of Ile, Leu (Fig. 9D), Met, and Val, and the cationic sidechains of metylated and unmethylated Lys (Fig. 9G) and Arg (Figs. 9H, S3). Particularly the latter occur prominently in the tails of histones and are recognized by aromatic cages in a range of different domains, including bromodomains, chromodomains (Fig. 9G), and Tudor domains (Fig. 9H). Given that the DDB1-Cul4A E3 ubiquitin ligase complex is known to bind and ubiquitinate histones (see e.g. [34]), an activity of cereblon in recognizing histone tail modifications within a linear sequence motif and providing the target specificity for the ligase complex appears fully plausible [18].
Other sidechain interactions, particularly in the context of linear sequence motifs, also appear entirely possible. Thus, the homeobox transcription factor MEIS2, which is implicated in various aspects of human development, was recently identified as a cereblon interactor [15]. Its binding was exclusive with thalidomide and its derivatives, suggesting that it is recognized via the same binding site. We note that MEIS2 and its paralogs contain two folded domains, one being the homeobox domain and the other uncharacterized at present, flanked by extended regions predicted to be unstructured and with low sequence conservation. The N-terminal approximately 10 residues are however very highly conserved and contain sidechains (Arg, Tyr, His) that could easily be envisaged as the ligands of an aromatic cage. We therefore consider this region to be the most attractive first candidate for exploring the MEIS2-cereblon interaction. This said, the very high similarity between the bacterial and eukaryotic CULT domains, particularly in the area of the aromatic cage, points to a wide-spread ligand, present also outside the cell, rather than to a linear sequence motif. By similarity to metyllysine, one might envisage choline, carnitine, betaine, and related compounds, but none of these could so far be seen to interact with the CULT domain in our model system.
In this article we have presented evidence that the thalidomide-binding region of cereblon is a conserved domain, CULT, present in several other proteins of eukaryotes and bacteria. The CULT domain is recognizably homologous to at least five other domain families, which share - where known - a common fold and a shared mechanism of ligand binding. The fold is also recognizable in three further domain families, which however do not have detectable sequence similarity to any of the other proteins, or to each other, and whose evolutionary relationship thus remains unclear (the SCOP database, however, clearly considers them homologous, as it groups them into the same superfamily). We have named the common fold of these proteins the β-tent, due to the orientation of its two constituent β-sheets.
The widely differing activities of proteins with a β-tent fold, as well as the absence of invariant residues across the domains, suggest that the β-tent is a structural scaffold, which mounts a binding site at a specific location. The binding site is formed by a cradle-shaped groove, whose sides are provided by loops connecting strands β4 to β5, and βI1 to βI2 of the common fold; the bottom is formed by strands β5, β6, and β7. The elaboration of this site in the individual families is tailored to their specific function, but appears to follow common principles, particularly in families binding small-molecule ligands. Here, binding residues are mainly located on the two loops and strand β6, while catalytic residues appear to be located on strand β7. For families whose binding site is at present unknown, this can therefore be reasonably predicted by mapping their conservation pattern onto homology models of the relevant region.
Members of the β-tent fold show, to varying degrees, a twofold rotational symmetry around a central axis passing through the apical zinc ion (where present). The symmetry is most pronounced in MsrB and this domain also has the broadest phylogenetic spectrum, being the only one with a universal representation in all cellular life forms. It therefore seems attractive to surmise that it is the ancestral representative of this fold, from which the others evolved by duplication and differentiation, and that it itself originated by duplication of a four-stranded β-meander. We have previously argued for an origin of folded proteins from subdomain-sized peptides [35], [36]. But for the apparent lack of internal sequence symmetry to support this inference, the β-tent would seem an attractive candidate for such a scenario.
The absence of statistically significant sequence similarity between MSS4, TCTP, DUF427 and the other proteins of this fold raises the possibility of a convergent origin. We note however that MsrB and RIG-I also do not share statistically significant sequence similarity between each other (Fig. 4) and are only connected conclusively in sequence space via CULT and yippee (Fig. 5). The homology of all proteins with a β-tent fold thus remains a clear possibility, which may become substantiated by new domain families found in hitherto poorly explored parts of the tree of life.
Sequence similarity searches were carried out at the National Institute for Biotechnology Information (NCBI; http://blast.ncbi.nlm.nih.gov/) and in the MPI Bioinformatics Toolkit (http://toolkit.tuebingen.mpg.de; [37]). PSI-Blast [38] at NCBI was run on the non-redundant protein sequence database (nr) with an E-value threshold of 0.005. CS-Blast [39] in the MPI Toolkit was run on a version of nr clustered at 70% sequence identity (nr70), also with a threshold of E = 0.005. The sequence relationships of proteins identified in these searches were explored by clustering them according their pairwise Blast P-values [40] in CLANS [41]. Clustering was done in default settings (attract = 10, repulse = 5, exponents = 1), with other settings as given in the figure legends.
Searches for more distant homologs were made with HHpred [42] and HHsenser [43] on the databases pdb70 (sequences of protein databank structures, as available in April 2014, clustered at 70% sequence identity), CDD (conserved domain database from NCBI, as of February 2014), pfamA release 27.0, SCOP release 1.75, and profile HMM databases of all human and all Drosophila proteins built locally and available through the MPI Toolkit.
Secondary structure was predicted in the MPI Toolkit, using the meta-tool Quick2D.
Structure similarity searches were carried out on the Dali server (ekhidna.biocenter.helsinki.fi/dali_server; [44]). Molecular models were built using Modeller [45], and manipulated in Swiss PDB Viewer [46] and PyMol (Schrödinger, LLC). Sequence conservation patterns were visualized with ProtSkin (http://www.mcgnmr.mcgill.ca/ProtSkin/; [47]).
Aromatic cage-like conformations containing ligands in PDB structures were detected by applying a set of geometric criteria (see Figure 9A). First, in each PDB structure, all non-water molecules in the HETATM record were identified and regarded as ligands. Only aromatic residues (phenylalaline, tryptophan and tyrosine) within 6.0 Å distance to these ligands were considered in further analysis. Then, we defined a set of at least three aromatic residues from the same polypeptide chain to form a cage-like conformation interacting with a ligand if: a) all pairwise distances between their side chain mass centers (MCSC) were less than 10.0 Å; b) the angle between and was less than 60° for at least three of the aromatic residues, where is the normal vector of the aromatic ring, is the vector connecting MCSC and the mass center of all side chain heavy atoms (MCALL); and c) at least two ligand atoms were within 3.0 Å distance to MCALL. The program was implemented in Python using BioPython [48], SciPy [49] and NetworkX [50] libraries.
We applied these geometric rules to scan 102,886 PDB files downloaded from the PDB (24 Aug 2014). In total, 6,144 putative aromatic cage-like conformations were detected with 1,098 different ligands binding to them. We grouped the cages according to the ligands they interact with. In each group, redundant cage-like conformations were removed (two cage-like conformations were considered identical if the composite residue names and numbers were the same). Subsequently, we manually examined at least one cage-like conformation in each of the 1,098 groups. Based on the ligand moiety within the aromatic cage, we further classified the 1,098 groups into different categories (S2 Table).
|
10.1371/journal.pntd.0006549 | Consensus criteria for the diagnosis of scabies: A Delphi study of international experts | Scabies was added to the WHO Neglected Tropical Diseases portfolio in 2017, and further understanding of the disease burden is now required. There are no uniformly accepted test methods or examination procedures for diagnosis, which limits the interpretation of research and epidemiological findings. The International Alliance for the Control of Scabies (IACS) designated harmonization of diagnostic procedures as a priority for the development of a global control strategy. Therefore, we aimed to develop consensus criteria for the diagnosis of scabies.
We conducted an iterative, consensus (Delphi) study involving international experts in the diagnosis of scabies. Panel members were recruited through expression of interest and targeted invitation of experts. The Delphi study consisted of four rounds of anonymous surveys. Rounds 1 and 2 involved generation and ranking an extensive list of possible features. In Rounds 3 and 4, participants were presented results from previous rounds and indicated agreement with a series of draft criteria. Panel participants (n = 34, range per Round 28–30) were predominantly highly experienced clinicians, representing a range of clinical expertise and all inhabited continents. Based on initial rounds, a draft set of criteria were developed, incorporating three levels of diagnostic certainty–Confirmed Scabies, Clinical Scabies and Suspected Scabies. Consensus was reached in Round 4, with a very high level of agreement (> 89%) for all levels of criteria and subcategories. Adoption of the criteria was supported by 96% of panel members.
Consensus criteria for scabies diagnosis were established with very high agreement. The 2018 IACS Criteria for the Diagnosis of Scabies can be implemented for scabies research and mapping projects, and for surveillance after control interventions. Validation of the criteria is required.
| Scabies causes rash and severe itch and predisposes to serious infection and chronic diseases of the heart and kidneys. Despite scabies being an ancient disease found in all parts of the world, we currently lack reliable laboratory tests. Clinicians generally make an assessment based on history and skin examination. Variation in diagnosis causes problems when trying to determine the prevalence of scabies in a region, or when trying to investigate the effectiveness of a treatment or control strategy. We aimed to establish consensus among experts on how to diagnose scabies. Thirty-four international experts responded to anonymous questionnaires on how tests and clinical features of scabies should be used to form a diagnosis. Draft criteria were developed and refined, incorporating three levels of diagnostic certainty–Confirmed Scabies, Clinical Scabies and Suspected Scabies. After three rounds of surveys, 82% of participants supported the criteria. After four rounds, consensus was reached with very high agreement for all criteria (Confirmed Scabies, 96%; Clinical Scabies, 93%; Suspected Scabies, 100%) and 96% supported adoption of the criteria. It is hoped the 2018 IACS Criteria for the Diagnosis of Scabies will allow harmonization of diagnosis and reporting, and comparison between studies of scabies burden and treatments.
| Scabies is a common skin condition, caused by the ectoparasite Sarcoptes scabiei var. hominis. The Global Burden of Disease study estimates more than 200 million people are affected at any one time [1]. The direct effects of scabies are estimated to account for 0.21% of DALYs caused by all conditions [2], but the broader impact, incorporating complications of bacterial skin infection, invasive bacterial disease and auto-immune kidney and heart disease is likely to be substantially greater [3].
The World Health Organization has recently designated scabies as a neglected tropical disease (NTD) for large-scale disease control action [4]. Whilst there has been a very high prevalence reported in some population-based studies, there is a paucity of disease burden data from most regions [5]. There are numerous opportunities to integrate population-based surveys for scabies with mapping and surveillance for other NTDs and health programs [6, 7], but few successful examples thus far [8]. Therefore, standardized methods are needed for further epidemiological mapping, and to monitor the effectiveness of control interventions.
The need to improve and standardize the diagnosis of scabies was identified as a priority at the inaugural meeting of the International Alliance for the Control of Scabies (IACS) in Atlanta, 2012 [3]. Microscopy of skin samples has limited sensitivity and utility in field settings, and there are no laboratory tests available. A systematic review of diagnostic methods used in therapeutic trials for scabies found wide variation with no uniformity [9], similar to results of a previous systematic review [10]. Given the limitations in available evidence for diagnostic methods, we aimed to develop consensus criteria for the diagnosis of human scabies in a range of epidemiological and research environments, using a formal consensus process.
The Delphi method is an established, iterative, multi-stage process for developing consensus using at least two rounds of anonymous surveys [11]. The Delphi method has been used for establishment of diagnostic and/or treatment frameworks where there is a lack of scientific evidence to make recommendations, including for several skin conditions [12–14]. After identification of the research issue and formation of the panel of participants, initial survey rounds generate a range of ideas and opinions through open-ended questions. In subsequent rounds, participants are provided a summary of the responses of the whole panel, and then given the opportunity to revise their own response. The advantages of the method include the ability to involve a large group of international participants, relatively low cost, anonymity, and reducing the likelihood of dominance by certain participants. Disadvantages include arbitrary cut-offs for consensus, potential for influence of moderators, and length of time required [11, 15].
The moderator group was comprised of a pediatrician, pediatric infectious disease physician and dermatologist, all experienced in the diagnosis and management of scabies and other tropical skin diseases, and all members of the IACS Steering Committee.
Recruitment to the panel was via an expression of interest circulated to the IACS membership, as well as targeted invitations to experts in scabies and tropical dermatology known to the working group or panel members. Moderators did not participate in the survey rounds. A panel of 10–15 members is usually recommended [15]. We aimed to have at least 20 respondents for each round.
Agreement was defined as the percentage of positive responses (agree or strongly agree) divided by the total completed responses (excluding responses outside areas of expertise). Consensus was defined a priori as 65–79% agreement indicating moderate consensus and ≥ 80% agreement indicating strong consensus.
The Delphi process consisted of four rounds. In Round 1, participants were asked to contribute all possible features relating to the diagnosis of scabies that could be considered for inclusion in the diagnostic criteria, including features on history, physical examination and investigation methods. Open-ended questions were used. Participants were asked to consider diagnosis in both resource-limited / field settings and high income / office-based settings. This generated a comprehensive list of possible features.
In Round 2, participants were asked to rank the importance of this long list of features for diagnosis using a visual analogue scale (from ‘not at all’ to ‘very important’). Further questions regarding clinical examination, investigations and dermatoscopy were assessed using categorical responses with four possible options (essential, important but not essential, not important, not required). Participants were also encouraged to provide further detail about the proposed list of features, possible combinations of features and the structure of the diagnostic criteria using free text responses. Visual analogue scales were converted to numerical values from 0 to 100. Features with a median value < 60 were omitted, those with value > 80 were retained and those between 60–79 were discussed by the moderator group, alongside all comments, taking into account the known evidence base. The narrowed list of features was developed into a draft set of criteria.
In Round 3, the draft set of criteria was presented, along with the analysis from Round 2. Participants indicated agreement to each aspect of the criteria, as well as the overall structure, using four-point Likert-type questions (strongly disagree, disagree, agree, strongly agree) or had the option of declining to respond if the item was outside their specific area of expertise (for example, questions related to mite visualization techniques). Participants provided additional free-text comments wherever there was disagreement to a criterion, or to provide suggestions about structure, syntax, clarity of phrasing and terminology. The working group further revised the draft criteria based on the analysis of the Round 3 agreement and free text commentary.
In Round 4, the revised criteria were presented, along with the agreement results of Round 3 and explanations for any revisions. Participants were again asked to state their agreement with each aspect of the criteria and the criteria structure, using the same Likert-type questions. Demographic information of participants was also collected in Round 4.
Study data were collected and managed using REDCap electronic data capture tools hosted at Murdoch Children’s Research Institute [16]. Up to three reminders were sent to participants who had not responded. All responses and comments were anonymous. Comments were compiled, grouped by theme and discussed by the moderator group after each round.
Thirty-four panel participants completed at least one survey. Participants were predominantly highly experienced dermatologists with practice across a range of settings and representing all major geographic regions. The characteristics of the Round 4 participants are shown in Table 1.
Responses were received from 30 participants for both Round 1 and Round 2. Participants expressed that criteria should only be used for diagnosis of common scabies, and not crusted or other atypical clinical variants. Based on responses, a set of proposed criteria were developed that could be used in a variety of settings, by incorporating three levels of diagnostic certainty–Confirmed Scabies, Clinical Scabies and Suspected Scabies.
Twenty-eight responses were received for Round 3. Agreement on the structure and contents of the draft criteria was high, with 23 respondents (82%) supporting the adoption of the draft criteria (Table 2). Extensive commentary was also received about content and terminology, leading to further revision of the draft criteria.
Twenty-eight responses were received for Round 4. There was a high level of agreement on all levels of criteria and individual subcategories (Table 3). Twenty-seven participants (96.4%) supported the adoption of the proposed criteria. Further commentary from panelists resulted in minor editing of terminology to improve clarity.
We established consensus on criteria for the diagnosis of scabies through a four-round Delphi process (Box 1). The final product is a single set of criteria with three levels and eight subcategories, representing a spectrum of diagnostic certainty. The agreement for each level and subcategory was very high, with at least 89% agreement for all subcategories, and 96% support for the adoption of the criteria.
The advantage of the levels and categories is that the criteria may be applied in a range of situations and settings. For example, a clinical trial of new therapies may choose to only enrol participants with Confirmed Scabies (level A) as this would be the most specific and least sensitive level. Alternately, for field mapping, Clinical and Suspected Scabies (levels B and C) may be used, acknowledging these would be less specific but more sensitive. Where examination of genital areas is not feasible or appropriate, due to a lack of visual privacy or other dignity considerations, sub-category B2 could be omitted. Investigators using the criteria are advised to list the distribution of diagnoses by sub-category, which will further allow comparison between survey findings.
Whilst these criteria are not intended to replace the clinical acumen and assessment of clinicians treating individual patients, we anticipate there may be clinical applications, particularly as a reference in settings where non-expert health workers are required to make a diagnosis. Clinical practice requires consideration of other factors outside of these criteria, such as history and progression over time, and the benefits and risks of treating or not treating individuals. Therefore, the threshold for diagnosis and treatment may be higher or lower in some cases.
Although this Delphi process resulted in high agreement for the diagnostic criteria, there are several limitations. In the absence of an appropriate reference standard test, or adequate data on the accuracy of clinical criteria, this process is reliant on expert opinion, which may not correlate with diagnostic accuracy. Therefore, an important next step is the validation of the criteria in a range of settings, including different geographic regions, areas of differing prevalence, and where other important differential diagnoses may be prevalent.
These criteria should be viewed as a starting point, allowing consistency and comparison in collected data. The criteria are likely to require revision, for example, if accurate laboratory or point-of-care tests for scabies infection become available. Although the criteria allow for diagnosis in mapping scenarios using a quick skin examination, the inclusion of history features, particularly relating to contact history, is not ideal for rapid assessment. Therefore, further research into the accuracy and utility of more simplified approaches to assessment are needed. It may be that more simplified assessment (for example, a limited skin examination of the limbs) could be used in field surveys, particularly if results can be correlated back to a comprehensive skin examination including assessment of itch and contact history.
Application of the criteria, summarized in Box 1, requires reference to the evidence-based explanatory notes and definitions that are being prepared in parallel. These notes will clearly define and explain each feature and subcategory of the criteria and will include recommended techniques for examination, microscopy and visualization, as well as details of important differential diagnoses. To enable use of the criteria for epidemiological mapping, a training methodology will be required for non-expert health workers, similar to that developed for trachoma graders [17].
Strong consensus has been established for criteria to diagnose scabies in a variety of settings. These criteria will facilitate diagnosis and comparison of findings across studies. The 2018 criteria require validation in a range of epidemiological settings.
|
10.1371/journal.ppat.1007382 | The seven transmembrane domain protein MoRgs7 functions in surface perception and undergoes coronin MoCrn1-dependent endocytosis in complex with Gα subunit MoMagA to promote cAMP signaling and appressorium formation in Magnaporthe oryzae | Regulator of G-protein signaling (RGS) proteins primarily function as GTPase-accelerating proteins (GAPs) to promote GTP hydrolysis of Gα subunits, thereby regulating G-protein mediated signal transduction. RGS proteins could also contain additional domains such as GoLoco to inhibit GDP dissociation. The rice blast fungus Magnaporthe oryzae encodes eight RGS and RGS-like proteins (MoRgs1 to MoRgs8) that have shared and distinct functions in growth, appressorium formation and pathogenicity. Interestingly, MoRgs7 and MoRgs8 contain a C-terminal seven-transmembrane domain (7-TM) motif typical of G-protein coupled receptor (GPCR) proteins, in addition to the conserved RGS domain. We found that MoRgs7, but not MoRgs8, couples with Gα MoMagA to undergo endocytic transport from the plasma membrane to the endosome upon sensing of surface hydrophobicity. We also found that MoRgs7 can interact with hydrophobic surfaces via a hydrophobic interaction, leading to the perception of environmental hydrophobiccues. Moreover, we found that MoRgs7-MoMagA endocytosis is regulated by actin patch-associated protein MoCrn1, linking it to cAMP signaling. Our studies provided evidence suggesting that MoRgs7 could also function in a GPCR-like manner to sense environmental signals and it, together with additional proteins of diverse functions, promotes cAMP signaling required for developmental processes underlying appressorium function and pathogenicity.
| The 7-TM domain is considered the hallmark of GPCR proteins, which activate G proteins upon ligand binding and undergo endocytosis for regeneration or recycling. Among eight RGS and RGS-like proteins of M. oryzae, MoRgs7 and MoRgs8 contain a 7-TM domain in addition to the RGS domain. We found that MoRgs7 can form hydrophobic interactions with the hydrophobic surface. This interaction is important in sensing hydrophobic cues by the fungus. We also found that, in response to surface hydrophobicity, MoRgs7 couples with Gα subunit MoMagA to undergo endocytosis leading to the activation of cAMP signaling. Moreover, we found that such an endocytic event requires functions of the actin-binding protein MoCrn1. Our results revealed that MoRgs7 also functions as a GPCR-like receptor protein to sense surface cues and activate signaling required for pathogenesis, providing new insights into G-protein regulatory mechanisms in this and other pathogenic fungi.
| In the rice blast fungus Magnaporthe oryzae, the appressorium is a special infection structure produced by the fungus to penetrate the host plant. Appressorium formation and function depend on signal transduction pathways, including G protein-coupled receptors (GPCRs)/G protein-mediated cAMP signaling [1, 2]. Once extracellular surface cues are sensed by GPCRs, such as the non-canonical GPCR Pth11 at the plasma membrane (PM), the GPCR stimulates the specific G-protein Gα subunit for activating the cAMP signaling pathway [2]. M. oryzae contains three distinct Gα subunits (MoMagA, MoMagB, and MoMagC) [3, 4] and other conserved pathway components, such as adenylate cyclase MoMac1, cAMP-dependent protein kinase A catalytic subunits MoCpkA, and MoCpk2 [1, 5–7]. Together, they regulate not only growth but also appressorium formation and pathogenesis.
In addition, M. oryzae contains at least eight RGS (regulator of G-protein signaling) and RGS-like proteins (MoRgs1 to MoRgs8). Previous studies found that all these RGS proteins have certain regulatory functions in various aspects of growth and pathogenicity with MoRgs1, MoRgs2, MoRgs3, MoRgs4, MoRgs6, and MoRgs7 being mainly involved in appressorium formation and MoRgs1, MoRgs3, MoRgs4, and MoRgs7 in full virulence [3, 8]. Despite such understandings, detailed mechanisms associated with specific RGS proteins remain not fully understood. In particular, RGS-like MoRgs7 and MoRgs8 proteins that also contain a seven-transmembrane domain (7-TM)‚which is a hallmark of GPCRs important in signal perception and transduction. In eukaryotes, GPCRs are well known for their role as heterotrimer ligand-regulated guanine-nucleotide exchange factors (GEFs) [9]. A ligand/agonist binding to a GPCR activates GPCR and promotes GPCR to mediate the exchange of GTP on the Gα subunit of the heterotrimer, leading to Gα dissociation from the Gβ-Gγ and activation of G protein-mediated signal transduction pathway, including the cAMP signaling pathway. In plant pathogenic fungi and oomycetes, it is generally considered that GPCRs have functions in perception of the environmental cues. This function enables plant pathogens to coordinate their metabolism with environment and to develop infection structures [10–12], although how these GPCRs detect environmental cues remains not clear.
Previous studies have found that RGS proteins such as human Rgs14 contain a C-terminal GoLoco/G protein regulatory motif that exhibits an in vitro GDP-dissociation inhibitor for Gα(i) [13]. Since MoRgs7 or MoRgs8 contain the 7-TM domain, we were interested in revealing whether MoRgs7 or MoRgs8 has additional functions mimicking a GPCR. Here we found that MoRgs7, but not MoRgs8, is involved in a distinct regulating mechanism. MoRgs7 couples with MoMagA to undergo endocytosis that is triggered by sensing surface hydrophobicity. Interestingly, MoRgs7 can sense environmental hydrophobic cues through interacting with the hydrophobic surface. In addition, MoRgs7 endocytosis depends on the actin-binding coronin homologue protein MoCrn1. Together, they contribute to G-protein/cAMP signaling required for appressorium function and pathogenicity.
Despite containing a relatively conserved RGS/RGS-like domain, 8 RGS proteins of the blast fungus are structurally divergent [8]. MoRgs7 and MoRgs8, in particular, contain a long C-terminus domain that was analyzed by transmembrane domain prediction systems (http://mendel.imp.univie.ac.at/sat/DAS/DAS and www.cbs.dtu.dk/services/TMHMM) to have a GPCR-like 7-TM motif (S1 and S2A Figs). MoRgs7 was demonstrated to have a role in appressorium function and pathogenicity, and this role is dependent on the 7-TM domain [8, 14].
To dissect the roles of MoRgs7 domains, the RGS domain was deleted (S2A Fig) and the mutant allele containing the 7-TM was fused to GFP and expressed in the ΔMorgs7 mutant. The fusion proteins MoRgs7Δ7-TM:GFP and MoRgs7:GFP [14] were also expressed in the ΔMorgs7 mutant as a control. Analytical results showed that the ΔMorgs7 mutant expressing 7-TM:GFP still remained a relatively high cAMP concentration, similar to the ΔMorgs7 mutant [8] but not the wild-type strain (S2B Fig). In hydrophobic surfaces, about 8% of ΔMorgs7 conidia improperly generated two appressoria (S2C Fig), which could also be observed in the ΔMorgs7/7-TM strain (S2C Fig). The ΔMorgs7/7-TM strain was also attenuated in virulence, similar to the ΔMorgs7 mutant (S2D and S2E Fig). In contrast, the expression of MoRgs7-GFP was able to suppress most of the defects in the ΔMorgs7 strain. These tests showed that the 7-TM and RGS domains are important for MoRgs7 function in cAMP and virulence. However, the test failed to establish an independent role of the 7-TM.
MoMagA plays a major role in cAMP signaling, appressorium formation and pathogenesis in M. oryzae and it is also one of the three Gα subunits demonstrated to interact with MoRgs7 [8]. To investigate functional mechanisms of MoRgs7-MoMagA interaction, we first validated the interaction through co-immunoprecipitation (co-IP). In addition to MoMagA, the constitutively active form of MoMagA, MoMagAG187S [3] was also included in the test. The result showed that MoRgs7 can interact with both MoMagA and MoMagAG187S and that the 7-TM and the RGS domain both can interact with MoMagA (Fig 1A and 1B).
Since GPCRs undergo endocytosis for receptor recycling [15], and both of MoRgs7 and MoMagA were localized to late endosomes that are the main components of the endocytic pathway, we hypothesized that MoRgs7 and MoMagA may also undergo actin-dependent endocytosis. To test this, we employed actin polymerization inhibitor latrunculin B (LatB) to disrupt endocytosis as previously described [16, 17]. At 3 h post-inoculation, MoRgs7:RFP and MoMagA:RFP signals remained very strong at the PM of the germ tube, in contrast to DMSO control (Fig 1C and 1D). Given that 4-bromobenzaldehyde N-2,6-dimethylphenyl (EGA) inhibits early to late endosome transport [18], it was applied and that led to an appearance of MoRgs7 and MoMagARFP signals in MoRab5:GFP-labeled early endosomes in germ tubes, in contrast to DMSO control (Fig 1E and 1F). Without EGA treatment, MoRgs7:RFP was predominantly localized to Rab7:GFP-labeled late endosomes (Fig 1E). These co-localizations of proteins with endosomes were corroborated by Pearson correlation coefficient statistical analysis. Taken together, MoRgs7 and MoMagA movement follows the common endocytic pathway.
To further validate MoRgs7 and MoMagA endocytosis, we photobleached MoRgs7 and MoMagA fluorescence in late endosomes of the germ tubes on hydrophobic surfaces and examined the fluorescence recovery dynamic using Fluorescence Recovery After Photobleaching (FRAP) at 3 h post-inoculation. In addition, we applied the microtubule-destabilizing benomyl to inhibit endosome trafficking via microtubuleand cycloheximide to inhibit newly synthesized fluorescent proteins moving into endosomes [19]. We found that endocytosis promotes recovery of RFP fluorescence of MoRgs7 and MoMagA in late endosomes within 90 sec (Fig 2A and 2B). Furthermore, we used FRAP to bleach the fluorescence in endosomes in the germ tube on the hydrophilic surfaces at 3 h post-inoculation. The recovery of fluorescence of MoRgs7:RFP and MoMagA:RFP in the endosomes was rarely detected (Fig 2A and 2B), suggesting that MoRgs7 and MoMagA are rarely internalized through endocytosis upon the perception of the hydrophilic surface.
Intriguingly, the absence of MoRgs7 and MoMagA endocytosis on the hydrophilic surface did not couple with accumulation of MoRgs7:RFP or MoMagA:RFP signals at the PM of the germ tubes (Fig 2C and 2D). As treating germinated conidia with LatB on hydrophilic surfaces for 1 h still could not cause accumulation of RFP signals at the PM (Fig 2C and 2D), we thus reasoned that in response to exposure to hydrophilic cues MoRgs7 and MoMagA were rarely sent to the PM from intracellular systems.
MoRgs8 also contains a 7-TM domain. To examine whether MoRgs8 undergoes similar endocytosis, we expressed MoRgs8:GFP in Guy11 and observed MoRgs8 localization during appressorium development on the hydrophobic surface. However, MoRgs8:GFP was found to be evenly distributed in the cytoplasm of germ tubes (Fig 3A). When compared with MoRgs7:GFP (Fig 3B), MoRgs8:GFP did not display any obvious endosome-localization patterns in the germ tubes. Further, LatB failed to cause any effects to MoRgs8:GFP distribution (Fig 3A). In contrast, the MoRgs7:GFP signal was enhanced at the PM in response to LatB (Fig 3B). These results revealed that MoRgs8 may function differently from MoRgs7.
Since the above results showed that the hydrophobic surface, not the hydrophilic surface, induces the PM localization of MoRgs7 in germ tubes during appressorium development, we hypothesized that MoRgs7 is possibly involved in sensing hydrophobic surfaces and the 7-TM may have a role in this process. We hypothesized that MoRgs7 at the PM may attach to hydrophobic surfaces in a hydrophobic interactive manner, and formation of such interactions by PM proteins including MoRgs7 is a step in the perception of hydrophobic cues. To test this hypothesis, we first examined whether MoRgs7 has the ability to bind to hydrophobic materials by performing an affinity precipitation assay with phenyl-agarose gel beads. The phenyl groups attached to the beads are highly hydrophobic. The beads were then incubated with MoRgs7:GFP and the GFP protein (a negative control), respectively, in a high concentration of salt solution containing 1.5 M NaCl and 1.5 M MgSO4. This allowed proteins to bind to the beads, as at high salt concentration non-polar side chains on the surface upon protein can interact with the hydrophobic groups [20]. Then we washed the beads to remove unbound proteins using a series of aqueous solutions with different salt concentrations. If an intense hydrophobic interaction between the protein and phenyl groups was formed, the protein will be hardly removable from beads even by low salt concentration solution containing 0.3 M NaCl and 0.3 MMgSO4, or containing 0.2 M NaCl and 0.2 M MgSO4. After washing, we used Western-blot analysis to detect the amount of MoRgs7:GFP or GFP that remained bound to beads. The results indicated that MoRgs7:GFP, but not GFP, remained in the elution (Fig 4A). This suggested that MoRgs7 has astrong ability to interact with hydrophobic materials and this ability may allow MoRgs7 to mediate a hydrophobic interaction between the pathogen and the hydrophobic surface. In addition, the hydrophobicity of 7-TM of MoRgs7 was tested and found to have hydrophobicity as full-length MoRgs7 (Fig 4A), and deletion of 7-TM decreased the level of hydrophobicity of MoRgs7, suggesting that this 7-TM is critical for hydrophobicity of MoRgs7. The results also indicated that MoRgs8 has a weak hydrophobicity compared to MoRgs7 despite having a 7-TM, suggesting that the characteristic of 7-TM is varied from MoRgs7 to MoRgs8.
We then investigated whether MoRgs7 forming a hydrophobic interaction with hydrophobic surfaces is an approach of M. oryzae to detect hydrophobic cues. Given that urea and ethylene glycol can interrupt hydrophobic interactions by causing a disordering of water molecules on hydrophobic regions [21, 22], they were applied to germinating conidia on hydrophobic surfaces at 1 h-post inoculation when most of conidia only germinated with a germ tube. In the presence of 0.5 M ethylene glycol or 0.1 M urea, appressorium formation was about 50% lower than that of water treatment at 4 h post-inoculation, even though almost 80% of conidia developed appressorium at 10 h post-inoculation (Fig 4B and 4C). Moreover, in the presence of 1 M ethylene glycol or 0.8 M urea, less than 20% of conidia developed appressoria even at 10 h post-inoculation. Most of conidia only germinated germ tubes with or without swelling at terminals. These results implied that a successful hydrophobic interaction formation is a critical step in hydrophobic surface recognition by M. oryzae.
To examine the nature of MoRgs7-MoMagA endocytosis and whether MoRgs7 internalization is dependent on MoMagA, we determined the rate of MoRgs7 internalization in the wild-type strain Guy11 and the ΔMomagA mutant using FRAP analysis. We found that MoRgs7 internalization had a normal rate in the ΔMomagA as that in Guy11 (S3A and S3B Fig). In addition, the internalization rate of MoMagA was also the same in Guy11 and the ΔMorgs7 strain (S3C and S3D Fig). These results suggested that MoRgs7 and MoMagA do not depend on each other in internalization.
To further understand the endocytosis process of MoRgs7, we searched for additional protein partners of MoRgs7 through a yeast two-hybrid (Y2H) screening and identified a coronin protein homolog, MoCrn1, as two polypeptides of 148 and 273 amino acids, from a cDNA library in the pGADT7 vector. MoRgs7 cDNA was inserted into pGBKT7 as bait. The interaction between MoCrn1 and MoRgs7 was specific, as an interaction between MoCrn1 and other RGS proteins, including MoRgs1, MoRgs3 and MoRgs4, cannot be established (Fig 5A). The interaction was further validated by co-IP and bimolecular fluorescence complementation (BiFC). The co-IP assay indicated that both the 7-TM and the RGS domains could interact with MoCrn1, independently (Fig 5B and 5C). BiFC revealed that MoCrn1 interacts with MoRgs7 during appressorium development (Fig 5D). The YFP signal could be detected at the PM of germ tubes while some weak signals appeared in the cytoplasm (Fig 5D), suggesting that the interaction is more often to occur at the PM. To investigate the interaction at the PM of germ tubes, we further conducted co-localization assay with co-expression of MoCrn1:GFP and MoRgs7:RFP in Guy11. And their co-localization at the PM in germ tubes of conidia on the hydrophobic surface was examined at 3 h post-inoculation (S6A Fig). However, we failed to detect the co-localization at the PM in germ tubes because the small amount of PM-localized MoRgs7:RFP is present (S6A Fig), as shown above (Fig 1C). MoCrn1:GFP formed actin patches-like structures, as described for coronin in Neurospora crassa [23]. But, from a small number of conidia, slightly obvious PM-localized MoRgs7:RFP was observed in germ tubes and partly co-localized with MoCrn1 patches (S6A Fig). The co-localization result suggested that a small amount of PM-localized MoRgs7 indeed has the opportunity to interact with MoCrn1.
In the eukaryotic cells, coronin proteins act as F-actin binding proteins and regulate actin-related processes such as membrane trafficking [24]. We tested whether MoCrn1 associates with actin in M. oryzae using Lifeact, a living cell actin marker described previously [17, 25, 26]. The MoCrn1:GFP and Lifeact:RFP were co-expressed in Guy11 and co-localization of MoCrn1:GFP and Lifeact:RFP was examined under a confocal microscope. We observed that MoCrn1:GFP and actin were dispersed in nascent appressoria after 6 h of incubation (S5A Fig), and that MoCrn1 punctate patches were localized to the membrane. However, MoCrn1:GFP formed ring-like structures in mature appressoria, which were highly co-localized with the F-actin network at the center of mature appressoria (S5A Fig). We also observed that MoCrn1:GFP were co-localized with actin at sub-apical collar region of hyphae andactin patches in hyphae andconidia (S5A Fig). The interaction between MoCrn1 and F-actin was again demonstrated through Y2H and binding assays (S5B and S5C Fig).
We next investigated whether MoCrn1 affects the actin organization by generating a ΔMocrn1 mutant, in which MoCRN1 gene knock-out was validated by Southern-blot (S4 Fig), and expressing Lifeact:RFP in the ΔMocrn1 mutant and Guy11. In Guy11, the hyphal tip regions were occupied with many actin patches that are associated with the PM (S5D Fig). However, about 20% of the hyphae formed some abnormal, enlarged actin patches in the cytoplasm of ΔMocrn1 (S5D Fig). Also, the enlarged actin patches could be found in over 10% of ΔMocrn1 conidia (S5E Fig), likely due to actin aggregation. Moreover, we found that Guy11 formed normal ring-like actin structure at the base of 80% appressoria, compared to 72% in ΔMocrn1 that displayed a disorganized actin network. This observation was confirmed by line-scan analysis (S5F Fig). Thus, we concluded that MoCrn1 regulates actin assembly and the ΔMocrn1 mutant displays minor defects in actin structures.
In the budding yeast Saccharomyces cerevisiae, Crn1 interacts with the microtubule [27]. The Δcrn1 mutant cells as well as cells overexpressing Crn1 showed microtubule defects and the mutant Δcrn1 is more sensitive than wild type strains to benomyl [28]. To determine whether MoCrn1 also affects the microtubule, the pYES2 construct containing the full-length MoCrn1 cDNA was expressed in the yeast Δcrn1 mutant. On SD plates containing 10, 20, and 30 μg/ml benomyl, Δcrn1 exhibited most significant inhibition in growth compared to the wild type strain BY4741 (S5G Fig). However, there was no significant difference between the Δcrn1 strain expressing MoCRN1 and BY4741. Further, we examined Guy11, the ΔMocrn1 mutant, and the complemented strain for benomyl resistance. On CM plates with 10, 20 and 30 μg/ml benomyl, we found that ΔMocrn1 was less sensitive to benomyl than Guy11 and the complemented strain (S5H Fig). Together, these results suggested that MoCrn1 has conserved microtubule-related functions.
As MoCrn1 interacts with MoRgs7 and is localized to the PM associated actin patches that represent endocytic pits [29], we hypothesized that MoCrn1 may function as an adaptor protein to direct MoRgs7 to endocytic pits/vesicles for internalization during appressorium development. To prove this, we investigated whether MoCrn1 affects MoRgs7 endocytosis by observing the spatial distribution of MoRgs7:RFP in germinated conidiaonthe hydrophobic surface at 3 h post-inoculation. Despite of that endosome-localized MoRgs7 was found in both the ΔMocrn1 mutant and Guy11, the ΔMocrn1 mutant displayed a higher concentration of MoRgs7:RFP at the PM of the germ tube than Guy11 did (Fig 6A). FRAP analysis indicated the fluorescence recovery of MoRgs7:RFP in ΔMocrn1 was evidently delayed than that in Guy11 (Fig 6C), suggesting that the diffusion of MoRgs7:RFP fluorescence into endosomes was impaired. This is consistent with our hypothesis that MoCrn1 is implicated in MoRgs7 internalization during appressorium development.
Since MoRgs7 and MoMagA are both internalized via endocytosis, we also examined if MoCrn1 has a role in the MoMagA internalization through a protein-protein interaction. We first validated the interaction between MoCrn1 and MoMagA. In Y2H, we found that MoCrn1 interacts with MoMagA and this interaction was specific, since MoCrn1 was not found to interact with MoMagB and MoMagC (Fig 5E). In addition, MoCrn1 did not interact with MoMagAG187S and MoMagAQ208L (Fig 5E), the twoactive forms of MoMagA [3]. The interaction between MoCrn1 and MoMagA was again confirmed by co-IP (Fig5F) and BiFC assays (Fig5G). In BiFC assay YFP is observed at the PM of germ tubes, revealing that MoCrn1 can interact with MoMagA at the PM of germ tubes during appressorium development. Moreover, similar to the co-localization of MoCrn1 with MoRgs7, from a small number of conidia the partial co-localizaiton between MoCrn1 and MoMagA was also found at the PM in germ tubes (S6B Fig).
We next tested whether MoCrn1 affects the MoMagA distribution during appressorium development on hydrophobic surfaces. In Guy11, we have observed that MoMagA:RFP displayed the endosome localization pattern in germ tubes and conidia. In ΔMocrn1, we could still observe MoMagA:RFP on late endosomes, but there was a significant increase in the membrane localization of MoMagA:RFP (Fig 6B). We again employed the FRAP assay to determine MoMagA internalization and found that the recovery of fluorescence of MoMagA:RFP in endosomes was slower in ΔMocrn1 than that in Guy11 (Fig 6D). These results confirmed that MoCrn1 is important for MoMagA internalization during appressorium development.
MoCrn1 is co-localized with F-actin so that MoCrn1 is similar to the adenylate cyclase associated protein MoCap1 that functions in cAMP signaling [6]. To examine whether MoCrn1 is required for MoCap1 localization, we expressed MoCap1:GFP in ΔMocrn1 and observed that the actin-like localization pattern of MoCap1 was completely disrupted in appressoria, conidia and hyphae of ΔMocrn1 (S8 Fig). Strikingly, MoCap1 preferred to form cytoplasmic aggregations. Additionally, we found that MoCrn1 interacts with MoCap1 by performing a co-IP assay (Fig 8F), in which the strain co-expressing MoCrn1:GFP and MoCap1:S was used. These results led us to conclude that MoCrn1 has a crucial role in recruiting MoCap1 to actin patches.
MoCrn1 has been associated with MoRgs7, MoMagA, and MoCap1 that all have a role in cAMP signaling. Indeed, we found that the ΔMocrn1 mutant also showed attenuated cAMP levels (S7A Fig) and a delay in appressorium formation (S9 Fig). At 4 h post-germination, nearly 40% of ΔMocrn1 conidia formed appressorium on a hydrophobic surface compared with 80% of Guy11 did. However, over 80% of ΔMocrn1 conidia could still form the appressorium at 6 h post-germination (S9 Fig). An incipient collapse assay indicated that MoCrn1 contributes to full turgor generation, since the collapse rate of the appressorium was significantly higher in ΔMocrn1 than in Guy11 and the complemented strains (S7B Fig).
Intracellular cAMP levels regulate the degradation of glycogen and lipid that are required for proper turgor generation in the appressorium [5, 30]. We thus compared the degradation of glycogen and lipid betweentheΔMocrn1mutant and Guy11 strains. Conidia were allowed to germinate on hydrophobic surfaces and iodine and Neil Red were used to stain glycogen and lipid, respectively [31]. At 6 h post-inoculation, glycogen appeared in the early appressorium (S7C Fig), and it broke down in 68.4% of the Guy11 appressoria after 16 h and 87% after 24 h, in comparison to 22.4% of ΔMocrn1 appressoria after 16 h and 53% after 24 h (S7E Fig). Resembling to the glycogen, lipid degradation in ΔMocrn1 appressoria was slower than Guy11. Lipid bodies disappeared in 44% of ΔMocrn1 appressoria at 16 h, compared to 86.4% of Guy11 appressoria (S7D and S7F Fig). These results indicated that MoCrn1 is indispensable for an efficient degradation of glycogen and lipid necessary for the appressorial turgor generation.
We further evaluated the ΔMocrn1 mutant for pathogenicity on rice. The conidial suspensions from Guy11, ΔMocrn1, and the complemented strain were sprayed onto the susceptible rice cultivar CO-39. ΔMocrn1 produced fewer lesions than Guy11 and the complemented strain, which were confirmed by lesion quantification (Fig 7A). We also performed rice sheath penetration assays by observing 100 appressoria each strain and classifying invasive hyphae (IH) types as previously described [17]. We observed that over 40% of ΔMocrn1 appressoria were defective in penetration and 55.6% of appressoria that penetrated and formed less extended IH. In contrast, 90% of Guy11 appressoria successfully penetrated rice cells and about 50% of that produced strong IH (Fig 7B).
To explore whether MoCrn1 regulates turgor generation involving the process of cAMP signaling, the incipient collapse assay was performed. We found that exogenous 8-Br-cAMP could suppress the defect of ΔMocrn1 in turgor generation (S7B Fig). The numbers of the collapsed appressoria in the ΔMocrn1mutant were reduced by 20% and 10% with 1 and 2 mM cAMP, respectively, compared to those without 8-Br-cAMP. In addition, the ΔMocrn1 mutant appressorium underwent successful glycogen and lipid breakdown following 8 and 16 h, respectively, following treatment with 5 mM 8-Br-cAMP (S7E and S7F Fig). Furthermore, 1 or 2 mM 8-Br-cAMP addition to the conidia suspensions in the inoculation of detached barley leaves could suppress the defect of ΔMocrn1 in infection to some degree (Fig 7C). This result was also confirmed by the penetration assay, in which 8-Br-cAMP treatment restored the penetration defect to almost 80% of the ΔMocrn1 appressoria in comparison to 43 ± 4.9% of ΔMocrn1 without cAMP (Fig 7B). This is similar to the effect of the ΔMocrn1mutant that expresses the constitutively active form of MoMagA, MoMagAG187S (S7G and S7H Fig and Fig 7A).
To examine the ability of MoCrn1 in binding multiple proteins, we identified putative actin binding domains and characterized their function. Human coronin Arg29and Arg30 are thought to be important for the interaction with F-actin [32, 33]. The alignment showed that a majority of coronins contain a conserved basic amino acid at these two positions (Fig 8A). In addition, the C-terminal coiled-coil (CC) domain is important for coronins to interact with the actin nucleation complex Arp2/3 [34]. Accordingly, we mutated His29 to Asp29 and deleted the CC domain of MoCrn1, and fused the mutant proteins with GFP (Fig 8B). We found that MoCrn1H29D and MoCrn1ΔCC mutants had completely altered actin-like localization patterns (Fig 8C). To further analyze the effects of these mutant alleles, we performed the co-IP assay and found that MoCrn1H29D and MoCrn1ΔCC mutants failed to interact with MoRgs7, MoMagA, and MoCap1 (Fig 8D, 8E and 8F).
We also expressed MoCrn1H29D and MoCrn1ΔCC mutants in ΔMocrn1. FRAP analysis showed that the expression of MoCrn1H29D and MoCrn1ΔCC caused no effect on delayed endocytosis of MoRgs7 and MoMagA in ΔMocrn1 (Fig 9). HPLC analysis revealed cAMP levels of the strain expressing MoCrn1H29D or MoCrn1ΔCC comparable to that of the ΔMocrn1 mutant (Fig 7D). Moreover, virulence and the degradation of appressorial glycogen and lipid in the MoCrn1H29D and MoCrn1ΔCC strains were also indistinguishable from those of the ΔMocrn1 mutant (Fig 7E, 7F and 7G). Taken together, these results suggested that MoCrn1 function is dependent on its ability to interact with F-actin, MoRgs7, MoMagA, and MoCap1.
We here investigated the distinct functional mechanism of RGS and 7-TM-containing protein MoRgs7 beyond its RGS functions.We found that MoRgs7 has a GPCR-like endocytosis pattern and is predominantly localized to late endosomes similar to other signaling proteins, including MoRgs1, MoMagA, and MoMac1. Such late endosome localizations of signaling proteins are critical to GPCR function and for cAMP signal transduction. Our results further showed that MoRgs7 couples with MoMagA to undergo endocytosis. Interestingly, by inhibiting endocytosis, we could observe increased PM localization of MoRgs7 and MoMagA. And by inhibiting trafficking from the early endosomes to the late endosomes, we could observe the early endosome localization of MoRgs7 and MoMagA.
Understanding how pathogen receptors recognizethe plant surface signal has a beneficial effect on the controlling rice disease at early stages. Our results provided evidences that MoRgs7 serves as a GPCR-like receptor to detect environmental hydrophobic cues. The affinity precipitation assay with phenyl-agarose gel beads indicates that MoRgs7 has strong ability to form hydrophobic interaction with hydrophobic materials, revealing that MoRgs7 can form interaction with hydrophobic surface when MoRgs7 is localized to the PM. Importantly, disruption of such hydrophobic interaction during M. oryzae germinating on the hydrophobic surface led to the aberrant appressorium formation. We also noted that the ΔMorgs7 mutant developed defective appressoria, even though no decrease in appressorium formation frequency. Based on these studies, we concluded that forming hydrophobic interactions with hydrophobic surface by MoRgs7 and other membrane proteins is a critical step in recognizing hydrophobic surface cues.
We reasoned that MoRgs7 may undergo a functional process similar to mammalian GPCRs. In mammalian cells, when a ligand binds to a GPCR, ligand can activate GPCR by inducing conformational changes in GPCR, subsequently the active GPCRs can activate the Gα proteins and are transported by endocytosis to sustain downstream signaling, be recycled, or be degraded from endosomes [35]. Considering our studies and previous findings by others in mammalian cells, we proposed a functional model of MoRgs7 (Fig 10). In this model, MoRgs7 acts as a GPCR during appressorium development to interact with the hydrophobic surface. Subsequently, this interaction induces MoRgs7-MoMagA endocytosis that is regulated by MoCrn1. MoRgs7 facilitates activating cAMP signaling from endosomes along with MoMagA. Conversely, MoRgs7 may elevate its GAP activity to regulate MoMagA when cAMP signaling is fully activated. Thus, MoRgs7 has dual roles in regulating signal transduction.
How MoRgs8 that also contains 7-TM domain but lacks sensory functions is not understood. MoRgs8 was distributed in the cytoplasm of germ tubes but did not undergo endocytosis. MoRgs8 could be involved in a mechanism distinct from MoRgs7 and future studies are needed to address such distinct mechanism(s).
There was precedence that endocytosis of RGS proteins plays a role in promoting Gα-mediated signaling. In Arabidopsis thaliana, in response to glucose, RGS protein AtRgs1 internalizes via endocytosis to uncouple itself from Gα protein AtGPA1 anchored in the PM, leading to AtGPA1 sustaining activation. And this process is required for both G-protein-mediated sugar signaling and cell proliferation [36]. However, other details including the initiation of MoRgs7-MoMagA complex disassembly following endocytosis remain not understood. We recently reported a distinct mechanism of how M. oryzae might negatively regulate the GAP activity of MoRgs7. This mechanism implicates the MoMip11 protein that interacts with MoRgs7 and the GDP bound MoMagA, but not the GTP bound MoMagA (Fig 10) [14]. MoMip11 prevents MoRgs7 from interacting with the GTP bound MoMagA, therefore interfering with MoRgs7 GAP function by sustaining MoMagA activation [14].
During endocytic vesicle formation, a series of adaptor proteins in cytoplasm can accumulate at endocytic sites. Those adaptors serve to select endocytic cargos and specifically bind to cargos, recruiting their cargos to endocytic pits/vesicles [29, 37]. Since we found that endocytosis of MoRgs7 and MoMagA are independent of each other, we considered that the adaptor protein(s) for the two proteins can anchor MoRgs7 or MoMagA to endocytic pits even though MoRgs7 and MoMagA do not interact with each other. Despite of that, the MoRgs7-MoMagA interaction is still important for MoRgs7 to regulate MoMagA activity. To further investigate the physiological function of MoRgs7 and MoMagA endocytosis, coronin protein MoCrn1 that emerges as an adaptor protein for MoRgs7 and MoMagA was identified and characterized. MoCrn1 is localized to actin patches that represent endocytic sites, interacting with MoRgs7 and MoMagA and regulating their endocytosis. Disruption of MoCrn1 by gene deletion or point mutations (H29D mutation and CC domain deletion) not only attenuated MoRgs7 and MoMagA endocytosis, but also led to a decreased cAMP level that is lower than the threshold for proper appressorium development. Our results support that MoRgs7 and MoMagA endocytosis regulated by MoCrn1 facilitates initiating cAMP signaling and appressorium development. However, for BiFC assays to test MoCrn1-MoRgs7 and MoCrn1-MoMagA interactions we queried why the YFP fluorescence evenly distributes at the PM of germ tubes, not just at the actin patches. A possible explanation is that, before MoCrn1 accumulates at actin patches, the cytoplasm-localized MoCrn1 has already bound to the cytoplasmic peptides of PM-localized MoRgs7 and MoMagA. At later stage, these interactions enable MoCrn1 to direct MoRgs7 and MoMagA to endocytic pits/vesicles.
Coronin proteins are known as regulators of the cytoskeleton and membrane trafficking in a number of species including yeast, Neurospora crassa, Dictyostelium discoideum, Drosophila, and human [23, 24]. In D.discoideum and mammalian cells, coronins have evolved to be modulators of signal transduction. Those coronins are critical for Rac1 GTPase activation and Rac1-dependent signaling [33, 38]. Additionally, upon cell surface stimulation coronin 1 interacts with and activates Gα to stimulate cAMP/PKA pathway in neuronal cell, even though how coronin 1 activates Gα is less clear [39]. Compared to those studies, our work revealed that MoCrn1 is involved in a distinct mechanism to facilitate Gα-cAMP signaling. MoCrn1 has an adaptor protein-like function by directing MoRgs7 and MoMagA to endocytic pits to promote their internalization. This function, thereby allows MoCrn1 to have a role in facilitating cAMP signaling. However, the function was not found yet for other eukaryote coronins, thus it is not known whether coronin is generally required for endocytosis of RGS and Gα proteins in eukaryotes except M. oryzae. Interestingly, MoCrn1 also interacts with MoCap1 that is thought as one of activators of MoMac1 [6]. Based on the above, we proposed that MoCrn1 is likely to be a hub or organizing protein of the network of MoRgs7-MoMagA-MoCap1.
The M. oryzae Guy11 strain was used as wild type for transformation in this study. For vegetative growth, small agar blocks were taken from the edge of 7-day-old cultures and cultured in liquid CM medium for 48 h. For conidiation, strains were cultured on SDC plates at 28°C for 7 days in the dark, followed by constant illumination for 3 days [8, 17, 31, 40–43].
The MoCRN1 deletion mutant was generated using the standard one-step gene replacement strategy [44]. First, two approximate 1.0 kb of sequences flanking of MoCRN1 (MGG_06389) were amplified with two primer pairs MoCRN1-F1/MoCRN1-R1, MoCRN1-F2/MoCRN1-R2, the resulting PCR products ligated with the HPH cassette released from pCX62. The protoplasts of wild type Guy11 were transformed with the vectors for targeted gene deletion by inserting the hygromycin resistance HPH marker gene cassette into the two flanking sequences of the MoCRN1 gene. For selecting hygromycin-resistant transformants, CM plates were supplemented with 250 μg/ml hygromycin B (Roche, USA).
To generate complementary construct pYF11-MoCRN1, the gene sequence containing the MoCRN1 gene and 1.0 kb native promoter was amplified with MoCRN1-comF/ MoCRN1-comR. Yeast strain XK1-25 was co-transformed with this sequence and XhoI-digested pYF11 plasmid. Then the resulting yeast plasmid was expressed in E. coli. To generate the complementary strain, the pYF11-MoCRN1 construct was introduced into the ΔMocrn1 mutant and pYF11 contains the bleomycin-resistant gene for M. oryzae transformants screen [31, 44].
EcoRV was used to digest the genomic DNA from Guy11 and the ΔMocrn1 mutant. The digest products were separated in 0.8% agar gel and were hybridized with the MoCRN1 gene probe. The probe was designed according to the disruption strategy and was amplified from Guy11 genomic DNA using primers MoCRN1-InterF/MoCRN1-InterR. To confirm MoCRN1 replacements, labeled MoCRN1 probe was used to hybridize the EcoRV-digested genomic DNA from the ΔMocrn1 mutant and wild-type Guy11. The copy number of the HPH gene in the ΔMocrn1 mutant was detected using labeled HPH fragments that amplified from the plasmid of pCB1003 with primers FL1111/FL1112. The whole hybridization was carried out according to the manufacturer’s instruction for DIG-High Prime.
The conidia were suspended in a 0.2% (w/v) gelatin solution (5×104 spores/ml), then the solutions were sprayed onto 2-week-old seedling of susceptible rice (Oryza sativa cv. CO-39) and also inoculated into 3-week-old rice CO-39 as described. Then the plants were incubated at 25°C with 90% humidity in the dark for the first 24 h, followed by a 12h/12h light/dark cycle. Lesions were observed after 7 days of incubation [41]. For pathogenicity assay with detached barley leaves [40], three 20 μl droplets of the conidia suspensions (1×105, 1×104, 1×103 spores/ml, respectively) added cAMP solution or not, were placed onto the upper side of the 7-day-old barley (cv. Four-arris) leaves. Then the leaves were incubated at 25°C with 90% humidity and in the dark for the first 24 h, followed by a 12h/12h light/dark cycle. Lesions were observed after 5 days of incubation.
To visualize glycogen, the samples were stained by iodine solution containing 60 mg/ml KI and 10 mg/ml I2 for 1 min. Nile red solution consisting of 50 mM Tris/maleate buffer (pH 7.5) and 2.5 mg/ml Nile red (9-diethylamino-5H-benzo-a-phenoxazine-5-one, Sigma), was used to treat the samples for 3 min, then the samples were examined under a fluorescence microscope with RFP channel [17, 26, 30].
The DNA fragments for expressing GFP fusion proteins were respectively inserted into the pYF11 construct that contains bleomycin resistant gene and G418 resistancegene, and the DNA fragments for expressing S-tag fusion proteins were respectively inserted into the pXY203 construct hat contains hygromycin gene. Then the constructs for expressing GFP and S-tag fusion proteins were co-transformed into wild-type strain Guy11, and the transformants resistant to hygromycin and bleomycin or G418 were isolated. The total protein of the transformants was extracted from mycelium using protein lysis buffer [1 M Tris-Cl (pH7.4), 1 M NaCl, 0.5 M EDTA, 1% Triton×100] and incubated with anti-GFP agarose beads (GFP-Trap, Chromotek, Martinsried, Germany) for 4 h, followed by washing beads with washing buffer (50 mM Tris HCl, 150 mM NaCl, pH 7.4) for 4 times. The proteins that bind to the beads were eluted by 0.1 M glycine HCl (pH 3.5) and were probed by anti-GFP and anti-S antibodies.
MoCRN1 and MoACT1 full-length cDNAs were cloned and inserted into pGEX4T-2 and pET32a, respectively. These constructs were transformed into E. coli strain BL21 for expressing proteins. Bacterial lysate containing GST:MoCrn1 protein was incubated with 30 μl GST agarose beads for 2 h. Then the beads were washed by washing buffer for 4 times and incubated with His:MoAct1 protein for 2 h, followed by washing beads with using washing buffer (50 mM Tris HCl, 150 mM NaCl, pH7.4) for 4 times again. The beads were boiled to elute proteins, and eluted proteins (output) were probed with anti-GST and anti-His antibodies.
Constructs of BD:MoMagA, BD:MoMagB and BD:MoMagC were used in previous experiments and kept in our lab. Full-length cDNAs of MoCRN1 was cloned and inserted into pGADT7 (AD) vector. Full-length cDNAs of MoCAP1, MoMagAG187S, MoMagAQ208L and MoACT1 genes were inserted into pGBKT7 (BD) vector. To examine the interaction of proteins, the AD and BD constructs were co-transformed into yeast strain AH109 and the transformants were grown on SD-Trp-Leu medium. Then the Trp+ and Leu+ transformants were isolated and assayed for growth on SD-Trp-Leu-His-Ade medium added X-α-Gal.
The germinated conidia with 3 h of incubation on hydrophobic or hydrophilic surfaces were treated with cycloheximide and benomyl as described [17]. FRAP were performed using a fluorescence microscope Zeiss LSM710. Regions containing MoRgs7:RFP and MoMagA:RFP in germ tube were selected for photo-bleaching. Photobleaching was carried out using an Argon-multiline laser at a wavelength of 561 nm with 80% laser power and 80 iterations in ROI. Images were acquired with 2% laser power at a wavelength of 555 nm every 5 sec. For quantitative analyses, fluorescence intensity was measured using the ZEISS ZEN blue software and fluorescence recovery curves were fitted using following formula: F(t) = Fmin + (Fmax − Fmin)(1-exp−kt), where F(t) is the intensity offluorescence at time t, Fmin is the intensity of fluorescence immediately post-bleaching, Fmax is the intensity of fluorescence following complete recovery, and k is the rate constant of the exponential recovery [45]. Mobile Fraction was calculated as the following formula: Mf = (Fend − F0)/(Fpre − F0), where Fend is the stable fluorescent intensity of the punctae after sufficient recovery, F0 is the fluorescent intensity immediately after bleaching, and Fpre is the fluorescent intensity before bleaching [46].
LatrunculinB (Cayman, USA) is dissolved in DMSO at a concentration of 25 mg/ml. Conidia incubated on the coverslips with hydrophobic surface were treated with LatB (final concentration 0.1 μg/ml) for 30 min, while the controls were treated with 5% DMSO. Then samples were washed with distilled water. Cycloheximide (MedChemExpress, USA) was solved in distilled water and the germinated conidia were treated with a final concentration 10 μg/ml for 10 min. Then samples were washed with distilled water. Benomyl (Aladdin, Shanghai, China) was solved in 0.1% DMSO and added to germinated conidia with a final concentration 1μg/ml. Then the samples were washed with distilled water. EGA (Merck, USA) was solved in 5% DMSO and was applied to samples with concentration 5 μg/ml for 1 h.
The total proteins were extracted from the Guy11 strain expressing MoRgs7:GFP or GFP, respectively, and were incubated with 100 mg of Phenyl-agarose beads (Senhui Microsphere Tech, Suzhou, China) in 1.5 ml microcentrifuge tubes at 10°C for 16 h. After incubation, the tubes were centrifuged (13000 g, 5 min) to remove the suspension. The beads were then gently washed with a series of aqueous solutions with different concentrations of NaCl and MgSO4 (1.5/1.0/0.8/0.5/0.3/0.2/0.1 M NaCl and MgSO4, 10 mM HEPES, pH 7.0), respectively, for 3 times to remove the unbound proteins. 100 μl of 1% SDS solution was added to the washed beads, followed by boiling the SDS solution and beads for 10 min to obtain elution, which was examined by western-blot using anti-GFP antibody.
All the samples were observed under a confocal fluorescence microscope (Zeiss LSM710, 63× oil). The filter cube sets: GFP (excitation spectra: 488 nm, emission spectra: 510 nm), RFP (excitation spectra: 555 nm, emission spectra: 584 nm). Exposure time: 800 ms. ImageJ software was applied to calculate Pearson correlation coefficient for analyzing co-localization of GFP fusion protein with RFP fusion protein. One area of interest was photographed with GFP and RFP channels respectively and photographs were opened using ImageJ software. Picture type was set to 8 bits. The “colocalization finder” in “plugin” section was applied to the pictures and Pearson correlation coefficient was calculated.
All of the strains were cultured on CM medium at 28°C, were cut into 1×1 mm squares, and were cultured in liquid CM for another 2 days. Filtering to collect mycelium and quickly ground into powder in liquid N2. 1 mg of mycelium was mixed with 20 μl of 6% TCA solution. Samples were centrifuged (1,377 × g, 15 min), the top layers were collected and were washed twice with five times the volume of anhydrous ether. The pellet was collected for HPLC. HPLC analysis was done with a programmable Agilent Technology Zorbax 1200 series liquid chromatograph. The solvent system consisted of methanol (90%)and water (10%), at a flow rate of 1 ml per minute; 0.1 mg of cAMP solution per milliliter was eluted through the column (SBC18, 5 μl, 4.6 × 250 mm) and was detected at 259 nm UV. Each sample was eluted through the column in turn and peak values were detected with the same time as the standard [47].
For construction of pHZ65:MoMagA vector used to express MoMagA-N’YFP, the N’YFP sequence was inserted into the alphaB-alphaC loop of MoMagA as described [48], then the MoMagA sequence containing N’YFP and the native promoter was fused with the pYF11 plasmid. For construction of vector used to express MoMagA:RFP, the RFP sequence was also inserted into the alphaB-alphaC loop of MoMagA. Then the MoMagA sequence containing the native promoter was fused with pYF11 plasmid. For construction of other vectors used to express proteins tagged with RFP or GFP, RFP or GFP was fused to protein sequence C-terminals, then protein sequences containing their native promoters were fused with the pYF11 plasmid.
MoRGS7 (MGG_11693), MoRGS8 (MGG_13926), MoMagA (MGG_01818), MoCRN1 (MGG_06389), MoCAP1 (MGG_01722)
|
10.1371/journal.ppat.1004647 | Porphyromonas gingivalis Evasion of Autophagy and Intracellular Killing by Human Myeloid Dendritic Cells Involves DC-SIGN-TLR2 Crosstalk | Signaling via pattern recognition receptors (PRRs) expressed on professional antigen presenting cells, such as dendritic cells (DCs), is crucial to the fate of engulfed microbes. Among the many PRRs expressed by DCs are Toll-like receptors (TLRs) and C-type lectins such as DC-SIGN. DC-SIGN is targeted by several major human pathogens for immune-evasion, although its role in intracellular routing of pathogens to autophagosomes is poorly understood. Here we examined the role of DC-SIGN and TLRs in evasion of autophagy and survival of Porphyromonas gingivalis in human monocyte-derived DCs (MoDCs). We employed a panel of P. gingivalis isogenic fimbriae deficient strains with defined defects in Mfa-1 fimbriae, a DC-SIGN ligand, and FimA fimbriae, a TLR2 agonist. Our results show that DC-SIGN dependent uptake of Mfa1+P. gingivalis strains by MoDCs resulted in lower intracellular killing and higher intracellular content of P. gingivalis. Moreover, Mfa1+P. gingivalis was mostly contained within single membrane vesicles, where it survived intracellularly. Survival was decreased by activation of TLR2 and/or autophagy. Mfa1+P. gingivalis strain did not induce significant levels of Rab5, LC3-II, and LAMP1. In contrast, P. gingivalis uptake through a DC-SIGN independent manner was associated with early endosomal routing through Rab5, increased LC3-II and LAMP-1, as well as the formation of double membrane intracellular phagophores, a characteristic feature of autophagy. These results suggest that selective engagement of DC-SIGN by Mfa-1+P. gingivalis promotes evasion of antibacterial autophagy and lysosome fusion, resulting in intracellular persistence in myeloid DCs; however TLR2 activation can overcome autophagy evasion and pathogen persistence in DCs.
| Among the most successful of human microbes are intracellular pathogens. By entering the intracellular milieu, these pathogens are protected from harsh environmental factors in the host, including the humoral and cellular immune responses. Porphyromonas gingivalis is an opportunistic pathogen that colonizes the oral mucosa and accesses the bloodstream and distant sites such as the blood vessel walls, brain, placenta and other organs. Still unclear is how P. gingivalis traverses from oral mucosa to these distant sites. Dendritic cells are highly migratory antigen presenting cells that “patrol” the blood, skin, mucosa and all the major organ systems. Capture of microbes by dendritic cells activates a tightly regulated series of events, including directed migration towards the secondary lymphoid organs, where processed antigens are ostensibly presented to T cells. Autophagy is now recognized as an integral component of microbial clearance, antigen processing and presentation by dendritic cells. We report here that P. gingivalis is able to subvert autophagic destruction within dendritic cells. This occurs through its glycoprotein fimbriae, called Mfa-1, which targets the C-type lectin DC-SIGN on dendritic cells. The other major fimbriae on P. gingivalis, FimA, targets TLR2, which promotes autophagic destruction of P. gingivalis. We conclude that DC-SIGN-TLR2 crosstalk determines the intracellular fate of this pathogen within dendritic cells, and may have profound implications for the treatment of many chronic diseases involving low-grade infections.
| Antimicrobial autophagy or xenophagy plays an important role in controlling bacterial infection and promoting innate immunity. Recent evidence has revealed critical roles for autophagy in the ability of immune cells to recognize and selectively target microbes for elimination. [1–4][1]. Dendritic cells (DCs) are innate immune cells that serve as a bridge to the adaptive immune response. DCs capture a wide variety of microbes in the peripheral tissues for which they are equipped with broad spectrum of pattern recognition receptors (PRRs). The major classes of PRRs expressed by DCs include Toll-like receptors (TLRs), NOD-like family receptors, CARD helicases and C-type lectin receptors [5,6]. Many of the PRRs come equipped with unique phagocytic machinery evolved for efficient antigen processing and presentation [7–9]. Phagocytosis relies on a network of endocytic vesicles such as early endosomes and/or autophagosomes, which fuse with lysosomes for degradation [10]. Intracellular vesicle maturation does not necessarily proceed through similar steps in different phagocytic cells [11]. Moreover, different phagosomal maturation pathways have been reported in the same cell type [11,12]. These different pathways are dictated primarily by the initial recognition step by PRRs and by the cargo contained in the vesicle [13]. Hence the immune cell type, the PRRs engaged and the properties of the microbe seem to be crucial for microbial clearance by autophagy.
DC-SIGN (DC specific ICAM-3 grabbing non-integrin) is a C-type lectin receptor involved in pathogen uptake, signaling and antigen presentation in DCs [14–16]. For uptake DC-SIGN contains internalizing motifs in its cytoplasmic tail [17]. Interestingly, DC-SIGN has been implicated in immune suppression and regulation in certain contexts [17–19]. Most notably DC-SIGN is targeted for immune escape by several pathogens such as HIV, hepatitis C virus, herpesvirus 8 (HHV-8), Mycobacterium tuberculosis, Helicobacter pylori and Streptococcus pneumonia [17,20–22]. Recently, we reported that DC-SIGN engagement by the minor fimbriae (Mfa1) of Porphyromonas gingivalis yields weak DC maturation and an immunosuppressive cytokine profile. In the absence of Mfa1, P. gingivalis yields a very different DC response with high levels of IL-23 and IL-6 as well as induction of a Th1/Th17 type response [14,23]. Furthermore, this study demonstrated that the anaerobe P. gingivalis survives within DCs in an aerobic atmosphere, while it dies rapidly in the absence of DCs [14]. Early study of the relationship of fimbrial strain differences to alveolar bone loss, showed that this Mfa1+Pg strain (DPG3) induced higher bone loss than Pg381 strain in a periodontitis mouse model [24]. The destruction induced by Mfa1+Pg was similar to wild type strain P. gingivalis ATCC 53977 that has been reported to be invasive in the abscess model [25].
P. gingivalis expresses a number of virulence factors that bind to and signal through PRRs. The adhesion proteins, known as fimbriae, on P. gingivalis signal through PRRs, and facilitate invasion of host cells. P. gingivalis expresses both minor (MFa1) and major fimbriae (FimA) which are highly regulated depending on growth conditions [26,27]. We have previously shown that expression of Mfa1 is involved in targeting DC-SIGN while other studies have shown expression of FimA targets a non-DC-SIGN route, mostly through TLR2 [28,29]. The engagement of DC-SIGN and TLRs activates distinct signaling pathways [6,30] and we propose that differential signaling through these distinct PRRs results in differential intracellular routing and processing of P. gingivalis within DCs. TLRs are essential for phagosome maturation and subsequent bacterial clearance [31,32]. TLR signaling is also involved in the maturation of autophagosomes [33]. The ability of P. gingivalis to manipulate DC-SIGN and TLR signaling through differential fimbrial expression [26,34], could have profound effects on bacterial survival[26]. However, the role of P. gingivalis major and minor fimbriae in DC-SIGN-TLR2 crosstalk and its influence on survival of P. gingivalis within DCs has not been examined.
In the present study, a combination of approaches was used to address the role of DC-SIGN and TLRs in intracellular routing and survival of P. gingivalis, including blocking PRRs and autophagy, siRNA gene silencing and activation of autophagy in monocyte derived DCs (MoDCs), To address the role of fimbriae in this regard we utilized defined bacterial mutants, that solely express minor fimbriae (Mfa1+Pg), major fimbriae (FimA+Pg) or are deficient in both fimbriae (MFB) [35] (Table 1). Our results indicate that engagement of DC-SIGN by MFA-1 allows P. gingivalis to evade autophagy and lysosome fusion, resulting in pathogen persistence and survival within DCs. In contrast, activation of autophagy or of TLR2 by P. gingivalis expressing FimA results in autophagy mediated killing of this pathogen within DCs. Collectively, our studies reveal a novel mechanism that enables this pathogen to evade host detection and clearance and which could have profound implications for the treatment of other diseases involving low-grade chronic infection.
At 2, 12 and 24h after bacterial co-culture with MoDCs, the MoDCs were imaged for intracellular P. gingivalis by epifluorescence microscopy and transmission electron microscopy (TEM). All P. gingivalis strains except the double fimbriae negative P. gingivalis strain MFB (Table 1) were taken up by MoDCs. There were marked differences in the P. gingivalis content of MoDCs at 2, 12 and 24 hours, particularly when comparing double fimbriae positive strain Pg381 to Mfa1+Pg (Fig. 1A). We observed a higher number of Mfa1+Pg within MoDCs (Fig. 1A) (S1 Fig.) This difference was most apparent after 24 hours, with large numbers of intra-and extra-cellular bacteria present. In contrast, MoDCs infected with Pg381 showed minimal bacterial content after 24 hours. Survival of intracellular bacteria was then assessed quantitatively by lysing MoDCs and growing bacteria in broth cultures and on anaerobic blood agar plates. Mfa1+Pg was recovered at higher numbers from MoDCs lysates in broth and on blood agar compared to Pg381 Fig. 1B). No significant difference was detected in the growth or death patterns of all strains in the media under anaerobic conditions in the absence of DCs (Fig. 1C).
To determine whether expression of DC-SIGN [14] was altered by P. gingivalis infection, MoDCs were infected with all the strains at different multiplicities of infection (MOIs) and gene expression of DC-SIGN was quantified at 2, 6, 12 and 24 hours (Fig. 2A) (S2 Fig.). At 12 hours, a distinct pattern of DC-SIGN expression was detected in MoDCs infected with Mfa1+Pg compared to Pg381 and FimA+Pg. Infection with Mfa1+Pg up-regulated DC-SIGN mRNA at 1 and 10 MOIs in a dose dependent manner (p<0.01) (Fig. 2A) (Table 2). In contrast, we observed decreased expression of DC-SIGN when MoDCs were incubated with Pg381 (MOI-10) down-regulated DC-SIGN mRNA expression significantly (p<0.05) at 12 hours (-4.55 fold). Fold regulations were calculated relative to un-infected MoDCs (Fig. 2A) (Table 2).
We also examined DC-SIGN expression on MoDCs by immunoelectron microscopy (Fig. 2B) and flow cytometry (Fig. 2C). The results confirm a difference in DC-SIGN expression in MoDCs as a function of P. gingivalis strain. Mfa1+Pg induced higher positive immuno-labeling for DC-SIGN in MoDCs relative to Pg381 (Fig. 2B). These results correlated well with our initial results from the mRNA analysis. DC-SIGN was detected on the membrane but also in the cytoplasm of MoDCs infected with Mfa1+Pg. The presence of cytoplasmic staining is consistent with previous evidence for Mfa1+Pg localization to DC-SIGN positive intracellular compartments [14] and the possibility of receptor recycling to the cell membrane after the phagocytic process. The cells infected with Pg381 showed minimal staining for DC-SIGN at the cell membrane and no cytoplasmic staining was detected at any of the time points (Fig. 2B). Stably transfected DC-SIGN positive and negative Raji cells served as positive and negative controls respectively for DC-SIGN expression by immuno-electron microscopy (S3 Fig.).
We quantitatively assessed the increase of DC-SIGN in MoDCs infected with MFa1+Pg by flow cytometry analysis (Fig. 2C) (S4 Fig.). We also monitored the expression of other C-type lectins and TLRs on the infected MoDCs by flow cytometry. Although P. gingivalis up-regulated the expression of TLR2 and CXCR4, we did not observe strain specific differences in expression of these receptors (S4 Fig.). Moreover, there were no changes in the expression of DCIR and mannose receptor (MMR) upon P. gingivalis infection. We did observe increased expression of Dectin receptors, but only on Mfa1+Pg infected MoDCs (S4 Fig.).
To analyze the phagosomal machinery involved in uptake and routing of different P. gingivalis strains by MoDCs, we assessed levels of the GTPase Rab family of proteins at early (2 hours) through late (24hours) stages of infection. Furthermore, we monitored co localization of the P. gingivalis strains with Rab proteins at these time points. Although both Pg381 and Mfa1+Pg were taken up at 2 hours, only Pg381 was associated with Rab5 at significantly higher levels as compared to Mfa1+Pg (Table 3) (S5 Fig. A and B). Association of Pg381 with Rab5 within MoDCs was detected up to 12 hours. After this point, detectable bacteria and Rab5 signals significantly decreased at 24 hours. The Rab5 signal was weak at all time points in MoDCs infected with Mfa1+Pg. In addition, Mfa1+Pg were more apparent than within MoDCs at 24 hours (S5 Fig. A and B).
MoDCs generally showed weak staining for Rab7 following infection either with Pg381 or Mfa1+Pg up to 24 hours (S5 Fig. C and D). Since Rab7 was not detected during intracellular processing of any of the P. gingivalis strains examined, we investigated whether anti-bacterial autophagy may be involved in killing of Pg381 but not Mfa1+Pg.
To investigate the role of autophagy as a putative survival mechanism utilized by Mfa1+Pg, the viability of P. gingivalis strains in MoDCs was monitored after induction of autophagy by the mTOR inhibitor, Rapamycin. Initial studies established that the viability of MoDCs and of P. gingivalis alone were not Rapamycin-sensitive at the concentrations used. Entry into MoDCs resulted in a significant increase in survival of Mfa1+Pg at 24 hours (p<0.001) (Fig. 3A). Rapamycin treatment of Mfa1+Pg-infected MoDCs (Mfa1+Pg+Rapa) significantly decreased P. gingivalis survival by ~48% (p <0.001) (Fig. 3A). Increased autophagy induction was confirmed by immuno-labeling of LC3-II in MoDCs treated with Rapamycin for 11 hours (1 hour after P. gingivalis infections). Rapamycin treatment increased the LC3-II signal in cells infected with all fimbriated strains as well as in un-infected (Cont.) (Fig. 3B and C).
To determine the kinetics of survival of P. gingivalis within MoDCs, as well as the involvement of autophagy, intracellular bacteria were monitored after 6, 12, 24 and 48 hours of incubation with MoDCs with or without Rapamycin (Fig. 4). The levels of P. gingivalis Mfa1+Pg within MoDCs (Mfa1+Pg+DC) were the highest at all time points except at 6 hours when we observed similar levels to that observed with Pg381+DCs (Fig. 4A). Levels of Pg381 and FimA+Pg within MoDCs (Pg381+DCs and FimA+Pg+DCs) were nonetheless significantly higher than bacteria without MoDCs, until 12 hours, at which point we observed a significant decrease in survival of both strains at 24 and 48 hours. Moreover, the highest level of Mfa1+Pg was observed at 24 hours, with the numbers of P. gingivalis increasing within MoDCs at 6, 12 and 24 hours (Fig. 4A). Activation of autophagy with Rapamycin significantly inhibited Mfa1+Pg survival at 6, 12 and 24 hours within MoDCs (Mfa1+Pg+DCs+Rapa) (Fig. 4B). However, the numbers of Mfa1+Pg detected within MoDCs treated with Rapamycin (Mfa1+Pg+DCs+Rapa) were still higher than bacteria alone (Mfa1+Pg) at 12 and 24 hours (Fig. 4B). For Pg381, rapamycin also significantly decreased intracellular survival, with no significant differences detected relative to bacteria alone except at 12 hours (Fig. 4C). FimA+Pg exhibited significant intracellular survival (FimA+Pg+DCs) only at 12 hours, which was significantly inhibited with rapamycin treatment (Fig. 4D). The analysis of the data using three-factor repeated measures ANOVA showed that both time and intracellular environment were significant factors (P<0.05) in P. gingivalis survival, with rapamycin significantly impairing P. gingivalis survival with MoDCs.
Due the marked difference in Rab5 induction in MoDCs infected with Pg381 and Mfa1+Pg, and the important role of Rab5 in regulation of subsequent autophagy [36,37] we tracked the LC3-II signal in MoDCs infected with labeled P. gingivalis strains between 2 to 24 hours (S6 Fig. A and B). LC3-II is the active form of cytosolic LC3 that associates with the autophagosome until cargo degradation [38]. Pg381 infection resulted in significant increases of LC3-II within MoDCs at 2, 6 and 12 hours (p = 0.0317, 0.008 and < 0.001, respectively). In contrast, LC3-II remains unchanged in MoDCs infected with Mfa1+Pg during 24 hours of infection. The highest level of LC3-II within Pg381-infected MoDCs was evident at 12 hours (mean fluorescent intensity = 2360.06 ±251.72) (S6 Fig. A and B). Hence, further analysis of the uptake of different P. gingivalis strains by MoDCs and the expression of LC3-II were carried out at the 12 hr time point (Fig. 5A and B). Infection with Pg381 and FimA+ Pg, but not Mfa1+Pg increased LC3-II levels in MoDCs during the first 12 hours. (Fig. 5A) (S6 Fig. A and B). Although there was generally a low level of co-localization between all P. gingivalis strains with LC3-II, strains Pg381 and FimA+Pg showed higher Pearson’s correlation than Mfa1+Pg with LC3-II. Moreover, quantification of the LC3-II signals in infected MoDCs revealed significant increases in cells infected with Pg381 and FimA+Pg (P<0.001 for both strains compared to Mfa1+Pg). Cells infected with Mfa1+Pg, in contrast, showed decreased levels of LC3-II compared to un-infected cells (Cont.) (Fig. 5B). Uptake of Pg381 and all mutant strains except MFB was confirmed by intensity quantification of CFSE (Fig. 5C). The highest uptake was detected in cells infected with Mfa1+Pg yet these cells had the lowest level of LC3-II signal (Fig. 5B and 5C).
Infection of MoDCs by Mfa1+Pg has previously been shown to depend on engagement of DC-SIGN [14]. We confirmed this result by blocking DC-SIGN with HIV gp120, prior to infection (Fig. 6A and 6B). Blocking DC-SIGN increased the LC3-II signal in MoDCs prior to addition of Mfa1+Pg (Fig. 6B). Moreover, blocking DC-SIGN restored the basal expression of LC3-II in MoDCs (Fig. 6B). To confirm the HIV gp120 blocking experiments, we additionally knocked down DC-SIGN using siRNA. DC-SIGN knockdown inhibited uptake of Mfa1+Pg but not Pg381 (Fig. 6C and 6D) and restored LC3-II signals in MoDCs (Fig. 6C and 6D). Furthermore, DC-SIGN knockdown significantly decreased survival of Mfa1+Pg in MoDCs (Fig. 6E). A scrambled sequence control did not inhibit uptake or effect LC3-II signal. To confirm the contribution of actin-mediated endocytic trafficking in LC3-II induction, MoDCs were treated with cytochalasin-D (CytD) prior to infection. CytD significantly inhibited intracellular localization of both Pg381 and Mfa1+Pg by MoDCs and restored the LC3-II signals to the basal level in MoDCs (S7 Fig. C and D).
We further confirmed LC3-II conversion in MoDCs by Western blot analysis (Fig. 7). Pg381 and FimA+Pg increased LC3-II expression in MoDCs. In contrast, Mfa1+Pg-infected MoDCs showed no significant difference in LC3-II compared to the uninfected control (Cont.) or MFB treated MoDCs (Fig. 7A and 7B). To determine if increased LC3-II signal in MoDCs infected with Pg381 was indeed due to increased induction, as opposed to accumulation from lack of autophagosomal—lysosome fusion, the latter was inhibited by Bafilomycin in the flux test as reported [39]. Dose response of Bafilomycin was confirmed in MoDCs, with the highest LC3-II accumulation observed between 3–4nM (S7 Fig. A and B). Bafilomycin treatment further increased LC3-II by Pg381, indicating an increase in autophagy by Pg381 rather than a block in autophagosome-lysosomal fusion (Fig. 7C and 7D).
To establish the role of TLR signaling in autophagy induction leading to DC maturation and intracellular killing of P. gingivalis, we activated TLR pathways using specific agonists for TLR1/2 (Pam3csk4) and TLR4 (E. coli LPS). The results demonstrate that TLR1/2 activation was highly potent in stimulating CD83 expression and intracellular LC3-II within MoDCs infected with P. gingivalis (Fig. 8A and 8B). In addition, we observed a significant reduction in the Mfa1+Pg counts within MoDCs treated with Pam3csk4 after 24 hours of infection (Fig. 8C).
The initial interaction of P. gingivalis with the outer membrane of MoDCs was visualized by scanning electron microscopy (SEM) after 2 hours of infection (Fig. 9A and 9B). Both strains (Pg381 and Mfa1+Pg) were able to engage the MoDC surface at 2 hours (Fig. 9A and 9B). To directly visualize formation of double membrane autophagosomes, TEM analysis was performed in MoDCs infected with all fimbriated strains. Tracking of Pg381 and FimA+Pg within MoDCs after 12 hours of infections demonstrated that the majority of bacteria were contained in double membrane structures. In contrast, greater numbers of Mfa1+Pg were consistently detected within single membrane vesicles in the cytoplasm of MoDCs (Fig. 9C). Quantification of the double membrane structures that contained bacteria in were carried out in three randomly selected EM grids of each sample. The ratio of Mfa1+Pg trapped in double membrane relative to the total intracellular bacteria was significantly lower than Pg381 and FimA+Pg (Fig. 9D).
Due to the important role of lysosome fusion in autophagosome maturation [40], CFSE-labeled P. gingivalis and LAMP1+ lysosomes were tracked in MoDCs by epifluorescence microscopy and P.gingivalis-LAMP1 co-localization was quantified between 2 and 24 hours. The uptake of all three fimbriated P. gingivalis strains by MoDCs at 2, 12 and 24 hours relative to uninfected controls was confirmed. (Fig. 10A and 10B) (S8 Fig.). At 24 hours, Pg381 was mostly undetectable compared to Mfa1+Pg within MoDCs (S8 Fig.). Uptake of Pg381 was associated with increased LAMP1 at 2 hours, while minimal LAMP1 was detected in cells infected with Mfa1+Pg at all time points (S8 Fig.). The highest LAMP1 signal in Pg381 and FimA+Pg infected MoDCs was observed at 12 hours, yet there was no increase in LAMP1 signal within Mfa1+Pg infected MoDCs (Fig. 10A). Quantification of LAMP1 intensities within Pg381 and FimA+Pg -infected MoDCs was higher than Mfa1+Pg-infected and un-infected MoDCs (p<0.001) (Fig. 10B). In addition, co-localization was higher in Pg381 and FimA+Pg with LAMP1 at 12 hours (Fig. 10C) (Table 4 and 5).
Our results indicate that canonical autophagosomal and lysosomal clearance of P. gingivalis within DCs is dependent on initial routing to early endosomes, followed by autophagosomal and lysosomal routing. We show that P. gingivalis is able to survive within DCs by subversion of this canonical pathway via its Mfa-1 fimbriae. DCs are unlike macrophages and neutrophils in that the lysosomal machinery is specialized for efficient epitope preservation, rather total degradation [41]. Nevertheless, DCs must have mechanisms to control and inhibit the growth of intracellular pathogens; otherwise DCs could serve as a significant niche for pathogen dissemination to distant organs as recently suggested [42]. Autophagy has been widely recognized as an antibacterial lysosomal mechanism with an immune regulatory component [2,4,43][1]. Although the configuration of autophagy apparently serves DCs well in most situations, very little is known about its role in elimination of specific intracellular pathogens. More specifically, it is unclear how engagement of different PRRs on human myeloid DCs influences the induction of and maturation of antimicrobial autophagosomes.
The minor Mfa1 fimbriae are 67-kDa glycoproteins that target DC-SIGN for entry into MoDCs (and Raji cells) [14]. As this strain does not express FimA, a TLR2 agonist [28] expressed by all the other fimbriated strains, the present work represents a unique opportunity to study the role of PRRs on DCs in the context of antimicrobial autophagy. DC-SIGN is a type II transmembrane receptor that recognizes a wide range of pathogens through internal mannose branched structures and terminal di-mannose n-oligosaccharides [44]. Three internalization motifs (di-leucine motif, tri-acidic cluster, ITAM motif) on its cytoplasmic tail facilitate pathogen uptake [45]. Despite the diverse immune functions and regulations mediated by DC-SIGN [15,23,46,47], the immune escape mechanisms associated with DC-SIGN engagement by pathogens remain largely “unexplained” [17,18,20,21].
Our results are consistent with a role for DC-SIGN in routing P. gingivalis into distinct intracellular vesicles that escape early autophagosomal recognition and subsequent vesicle maturation and degradation. This DC-SIGN dependent re-routing seems to begin at the stage of early phagosomal formation, whereas DC-SIGN independent routing ‘traps’ P. gingivalis within Rab5 rich vesicles immediately after uptake. Rab5 has been suggested as an early stage initiator of autophagy and facilitator of subsequent lysosomal fusion [37][40]. In addition, Rab5 is reported to enhance autophagy by inhibition of mTORC1 (mTOR complex). The results of the current study suggest that when TLR signaling is weak (e.g. in the absence of FimA on the Mfa1+Pg strain) engagement of DC-SIGN dominates the response. This was clarified by adding exogenous TLR1/2 agonist, which inhibited Mfa1+Pg survival in parallel with higher LC3-II signals and stronger DC maturation. We surmise that lacking a strong signal for autophagy activation such as a TLR2 agonist, pathogens internalized by DC—SIGN are preferably routed to non-autophagosomal, non-lysosomal compartments where they survive in DCs.
One of the striking observations of the present work was the preferential routing of Mfa1+Pg to single-membrane intracellular vesicles, while the other strains were predominantly contained within characteristic double membrane phagophores. Classically, autophagy has been identified by the formation of double membrane vesicles that ultimately fuse with the lysosome for degradation of its intracellular components [48]. However, recently an alternative pathway known as non-canonical autophagy or LC3 associated phagocytosis (LAP) has been described [49], which involves the formation of single membrane vesicles. Here, DC-SIGN engagement appears to facilitate containment within single membrane vesicles; thus allowing P. gingivalis to evading lysosomal fusion. Some reports suggest that certain pathogens use (single membrane) autophagosomes for replication by blocking autophagosomal maturation [2,50,51]. Whether P. gingivalis utilizes DC-SIGN to enter a non-canonical pathway of autophagy for bacterial survival in DCs remains to be determined. One of the earliest studies describing P. gingivalis evasion of the conventional endocytic pathway and rerouting to (canonical) autophagosomes was in human aortic endothelial cells in [50]. Endothelial cells have been reported to express DC-SIGN [52] but this was not addressed in the previous study [50]. DCs, in contrast to endothelial cells, are highly migratory professional antigen presenting cells. Most reports of autophagy as an anti-microbial mechanism are from studies of immune cells [1,2,50,51,53]. In the present study of myeloid DCs, P. gingivalis appears to use an alternative tactic to evade autophagic capture through early engagement with DC-SIGN receptor and through non-canonical autophagy.
Analyses of intracellular survival of the different P. gingivalis strains suggest that DCs can use autophagy and lysosomal fusion to clear this intracellular pathogen. Pg381, which expresses both Mfa1 and FimA, engages both DC-SIGN and TLR2. It appears that DC-SIGN dependent routing is overruled by a strong TLR signal since Pg381 activates the autophagic process and is subsequently cleared, although this strain still survives longer inside MoDCs than outside in an aerobic atmosphere. Indeed, TLR signaling is a well-accepted inducer of autophagy as well as of phagosomal maturation [31–33]. However, the majority of autophagy studies that addressed TLRs were conducted in macrophages and neutrophils, both of which are naturally equipped with strong intracellular killing arsenal and higher expression of TLRs.
DC-SIGN expression is considered restricted to immature DCs with few exceptions in endothelium and specific macrophage subpopulations [54,55]. Our current data shows that DC-SIGN engagement may cause a positive feedback loop that increases the receptor expression resulting in subsequent bacterial uptake and survival. Increases in DC-SIGN+DCs in tissues of patients with periodontitis has previously been reported [56]. This could further enhance the pathogenicity of P. gingivalis and other DC-SIGN targeting pathogens. In the case of P. gingivalis, its ability to regulate fimbrial expression in different environmental cues such as pH, temperature and hemin content [26,27,34] may aid in its pathogenicity during chronic inflammation by regulation of DC-SIGN engagement (and expression). Current efforts in our laboratory are directed to identifying the levels of expression of Mfa1 and FimA in clinical samples of disease vs. healthy patients.
Phagosomes are known to have a high degree heterogeneity and individuality depending on host cell types, microbe captured and pattern recognition receptors engaged [11]. When Staphylococcus aureus and Salmonella typhimurium engage TLRs, for example, they are delivered to lysosomes at an inducible rate manifest by increased clearance and phagolysosomal fusion. However, slower phagolysosme maturation results from engagement with members of C-type lectin family and scavenger receptors [57]. DCs infected through DC-SIGN maintain an immature state and are more resistant to apoptosis [58]. Typically, immature DCs have a short life and active apoptosis is initiated shortly after maturation to avoid immune overstimulation [59,60]. Dominant DC-SIGN engagement alters such homeostatic balance and hinders the intracellular resistance to such infection. The DC-SIGN route may be a hallmark of chronic inflammation in response to low grade infection as it provides a protective niche for microbial persistence within the host. It is also influential in depolarizing the immune effector response [14]. Several lines of evidence have emerged linking autophagy and chronic inflammatory diseases [1,61,62]. In immature primary DCs, autophagy induction by NOD2 was essential for routing bacteria to lysosome and MHC presentation [61]. Early recognition of the microbe, here by DC-SIGN, could be crucial in driving ‘normal’ versus ‘up-normal’[1] autophagy and affect the fate of inflammatory process in chronic periodontitis.
The relevance of this work from a clinical standpoint comes from evidence that Mfa1+ P. gingivalis strains infect myeloid DCs in oral mucosal tissues and in blood of humans with periodontitis, wherein it is disseminated to distant sites of angiogenesis [42]. We conclude that this intracellular pathogen can survive within DCs by evasion of autophagy through coordinated regulation of Mfa1 and FimA expression. This may also facilitate its dissemination [42]. Potential therapeutic tactics for resolving chronic inflammatory diseases by forced autophagy to activate a strong immune response is also suggested by these studies.
Human monocytes were isolated from mononuclear fractions of peripheral human blood by Human monocyte enrichment technique. After incubating the blood with the enrichment kit (RossetteSep, Cat. no. 15028) for 20 minutes, monocyte separation was carried out using medium density Ficoll (GE Healthcare, Cat. no. 17–1440–03). Cells were seeded in the presence of GM-CSF (1000 unit/ml, Gemini Bio-Product, Cat. no. 300–124P) and IL-4 (1000 unit/ml, Gemini Bio-Product, Cat. no. 300–154P) at a concentration (3–4 x 105 cells/ml) for 5–6 days. Flow cytometry analyses were carried out to verify the immature DC phenotype (CD1a+, CD83-, CD14-, DC-SIGN+). Cell surface markers of DCs were evaluated by four-color immunofluorescence staining with the following antibodies: CD1a-PE (Miltenyi, Cat. no. 120–000–889), DC-SIGN-FITC (Miltenyi, Cat. no. 130–092–873), CD14-PerCP (Miltenyi, Cat. no. 130–094–969) and CD83-APC (Miltenyi, Cat. no. 130–094–186). After 30 min at 4°C and washing with staining buffer (PBS pH 7.2, 2 mM EDTA, and 2% FBS), cells were fixed in 1% paraformaldehyde. Positive marker expression was calculated as a percentage of total DCs by forward scatter and side scatter characteristics [14,23].
To corroborate the immunoelectron microscopy staining of MoDCs for DC-SIGN, stably transfected DC-SIGN-positive (Raji-DCs) and negative Raji cells (Raji) were be obtained and the phenotype verified by flow cytometry[47]. The cells were cultured in 10% heat-inactivated FBS (Gemini, Cat. no. 100–500), RPMI 1640 with L-glutamine, and NaHCO3 (Cornning, Cat. no. 10041CM) in a 5% CO2 incubator at 37°C. Cells were centrifuged into a pellet and prepared for transmission electron microscopy sectioning.
MoDCs were pre-incubated with HIV-1 gp120 Chiang Mai (CM) envelope protein (GP120) for 30 min at 37°C. GP120 protein was obtained through the National Institutes of Health AIDS Research and Reference Reagent Program, Division of AIDS, National Institute of Allergy and Infectious Diseases, National Institutes of Health (Cat. no. 2968) For actin polymerization inhibition, MoDCs were treated with cytochalasin D at 0.5 μM, the minimal concentration needed to arrest cytoskeletal rearrangements in Raji cells [14]. Cells were then washed 2 times with PBS and co-cultured with CFSE- stained P. gingivalis for 2, 12 and 24 hours at 37°C. Cells were fixed with 1% paraformaldehyde and prepared for immunofluorescence staining and epifluorescence microscopy.
Cells were incubated with predesigned Siliencer Select siRNAs for DC-SIGN (Cat. No. 4392420 Ambion) for 24 hours at 10 nM concentration. 12 ul of lipofectamine 2000 Reagent (Cat. No. 11668–500 invitrogen) were used with Opti-MEM medium (Cat. No. 11058–021 LifeTechnologies) were used for siRNA delivery. Flow cytometry analysis was performed in control and infected MoDCs to confirm inhibition of DC-SIGN.
MoDCs were incubated with Rapamycin (Cat. no. Tlrl-rap, InvivoGen, San Diego, CA) at 200nM one hour after P. gingivalis infections. Induction of autophagy was confirmed by fluorescence staining of LC3-II within MoDCs after 2 and 6 hours.
E. coli 026:B6 LPS (L2654, 2.5% protein, 1,500,000 EU/mg LPS, Sigma-Aldrich, St. Louis, MO). For stimulations, cells were treated with LPS at 1000 u/ml (200 ng/ml). For TLR1/2 stimulation Pam3CSK4 (Synthetic triacylated lipoprotein) (Cat. No. tlrl-pms, InvivoGen, San Diego, CA) were used at 1ug/ml.
Four P. gingivalis strains were used in this study; 1) Pg381, which expresses both minor (Mfa1) and major (FimA) fimbriae, 2) isogenic minor fimbria-deficient mutant (FimA+Pg), which expresses only the major fimbriae, 3) isogenic major fimbria-deficient mutant (Mfa1+Pg), which expresses only the minor fimbriae and 4) the double fimbriae mutant (MFB) (Table 1). P. gingivalis strains were maintained anaerobically in (10% H2, 10% CO2, and 80% N2) in a Forma Scientific anaerobic system glove box model 1025/1029 at 37°C in Difco anaerobe broth MIC [63]. Mutant strains were maintained using erythromycin (5 μg/ml) for mutant Mfa1+Pg, tetracycline (2 μg/ml) for mutant FimA+Pg and both erythromycin and tetracycline for double fimbriae mutant MFB.
Bacteria suspensions were washed five times in PBS and re-suspended for spectrophotometer reading at OD 660 nm of 0.11, which previously determined to be equal to 5 x 107 CFU [64]. For bacterial CFSE staining, the suspension were washed (3 times) and re-suspended in 5μM of CFSE in PBS. The bacteria were incubated for 30 min at 37°C in the dark [14,65]. MoDCs were pulsed with Pg381, Mfa1+Pg, FimA+Pg and MFB at 0.1, 1 and 10 MOI and incubated with the MoDCs for 2, 6, 12 and 24 hours and each experimental condition were performed in triplicate.
After 24 hours of MoDCs infection with P. gingivalis strains, cells were washed three times in PBS and re-suspended in sterile water on ice for 20 min to lyse the cells. Lysates were re-suspended in anaerobe broth for 3 days. After broth incubation, bacterial suspensions were washed three times in PBS and re-suspended for spectrophotometer reading at OD 660 in triplicate. Viable counts (CFU) were calculated based on a plate count serial dilution versus OD readings. For confirming the identity of the P. gingivalis (black pigmented Gram negative coccobacilli) suspensions were cultured on 5% blood agar plates in triplicate under anaerobic conditions (10% H2, 5% CO2 in nitrogen). Plates were incubated in anaerobic conditions at 35°C for 14 days until black colonies were detected and select colonies gram-stained.
After MoDCs were infected with the P. gingivalis strains for 2, 6, 12, 24 and 48 hours, cells were washed three times in PBS and re-suspended in sterile water on ice for 20 min to lyse the cells. Bacterial suspensions were washed three times in PBS and re-suspended for spectrophotometer reading at OD 660 in triplicate. Corresponding CFU counts were calculated based on a linear regression of plate count in serial dilution versus OD readings. Black colonies were confirmed in blood agar plate under anaerobic conditions (10% H2, 5% CO2 in nitrogen).
After MoDCs fixation, the procedures were carried out at the Electron Microscopy and Histology Core, Department of Cellular Biology and Anatomy, Georgia Regents University. The cells were fixed in 2% glutaraldehyde in 0.1 M sodium cacodylate (NaCac) buffer, pH 7.4, postfixed in 2% osmium tetroxide in 0.1 M NaCac, stained en bloc with 2% uranyl acetate, dehydrated with a graded ethanol series and embedded in Epon-Araldite resin. Thin sections were cut with a diamond knife and stained with uranyl acetate and lead citrate. Cells were observed in transmission electron microscope (JEM 1230—JEOL USA Inc.) at 110 kV and imaged with a CCD camera and first light digital camera controller (Gatan Inc.).
The procedures were carried out at the Electron Microscopy and Histology Core, Department of Cellular Biology and Anatomy, Georgia Regents University. Staining for DC-SIGN using anti-CD209 (mouse monoclonal, R&D Systems, # MAB161), was carried out to identify the P. gingivalis-containing vesicles. After DCs were pulsed with different P. gingivalis strains for the 2, 12 and 24 hours, cells were centrifuged into a pellet. Cells were be fixed in 4% formaldehyde 0.2% glutaraldehyde in 0.1 M sodium cacodylate (NaCac) buffer, pH 7.4, dehydrated with a graded ethanol series through 95% and embedded in LR White resin. Thin sections were cut with a diamond knife on a Leica EM UC6 ultramicrotome (Leica Microsystems Inc.) and collected on nickel grids. Sections were incubated in blocking buffer (5% BSA, 3% normal serum, 0.05% Tween-20 in Tris-buffered saline, pH 7.4) at room temperature in a humid chamber for 2 hours and with primary antibody diluted in blocking buffer overnight at 4°C. Grids were washed and incubated with gold-labeled secondary antibody for 2 hours at room temperature then washed and stained with 2% alcoholic uranyl acetate and 0.08% alkaline bismuth subnitrate. Cells were observed in a JEM 1230 transmission electron microscope (JEOL USA Inc.) at 110 kV and were imaged with an UltraScan 4000 CCD camera & First Light Digital Camera Controller (Gatan Inc.)
For RNA isolation, direct lysis of the cell suspensions were achieved by RNeasy kit (Cat. no. 74104, Qiagen) by adding 300 μl of Qiagen’s buffer RLT per sample. The lysates were collected and pipetted directly into the Qiashredder spin column. Ethanol (70%) was added and then samples were transferred to RNeasy spin columns. The samples were washed with buffer RW1, RBE, then the RNA samples were collected and stored at -80°C. RNA quantity and integrity were tested and only ratios of absorbance at 260 and 280 nm of 1.8–2.0, were included in the study. One-step qrt-PCR were performed using Express qPCR SuperMix (Cat. no. A10312, Invitrogen). Pre-formulated individual TaqMan gene expression primers (Applied Biosystems) were used for DC-SIGN mRNA detection (Hs.01588349_m1). For qrt-PCR reactions, 5μl of the RNA sample, 25μl PCR master mix (2x) and 2.5μl TaqMan gene expression assay were used per reaction. All PCRs were performed in triplicate and were carried out on a real-time PCR, StepOne (Applied Biosystems). For calculations and statistical analysis, fold changes were calculated using (2-ΔΔCT) method in the experimental samples [66]. Statistical analysis for gene expression was performed using the one sample t-test, which estimates the calculated difference (in fold-regulation) between experimental and control samples. A p value of <0.05 is the cut-off for significant differences.
Cells were fixed with 4% paraformaldehyde, blocked and counterstained with fluorescent-labeled antibodies against Anti-Human CD206 (MMR) (Cat. No. 53–2069–41), Anti-Human DCIR (Clec4A) (Cat. No. 12–9875–41), Anti-Human CD209 (DC-SIGN) (Cat. No. 45–2099–41), Anti-Human Dectin-1 (Cat. No. 46–9856–41), Anti-Human CD284 (TLR4) (Cat. No. 12–9917–41), Anti-Human CD184 (CXCR4) (Cat. No. 15–9999–41), Anti-Human CD282 (TLR2) (Cat. No. 17–9922–41), Anti-Human CD286 (TLR6) (Cat. No. 13–9069–80) (all ebioscience, USA). All markers were measured against isotype controls. The markers were measured as MFI using the Accuri C6 Flow Cytometry system.
MoDCs were infected with P. gingivalis prelabeled with carboxyfluorescein succinimidyl ester (CSFE) fluorescence. Cells were fixed with 1% paraformaldehyde, washed with PBS twice and permeabalized with 0.5% saponin. MoDCs were incubated with LC3-II antibody (ab51520) for 2 hours and then washed with PBS. Pellets were res-suspended in cytospin fluid (Cat. no. 6768315, Shandon) centrifuged at 400 rpm for 4 minutes. Slides were mounted with anti-fade reagent (Invitrogen, P36931) and dried for microscopic analysis. Microscopic images were obtained with epifluorescence microscope (Nikon E600) then analyzed by image enhanced fluorescence microscopy aided by deconvolution analysis. Quantifications of the fluorescent intensities and co-localization within infected cells were done by NIS-Elements BR and AR software. Three randomly selected regions of interest were selected for each field to quantify fluorescence dye intensities.
Cells were centrifuged and washed twice with PBS. After washing, cells were lysed by addition of cell lysis buffer (Cell signaling Cat. no. 9803S) and incubated for 20 minutes on ice. Samples were centrifuged and the supernatant was collected and stored at -80°C. Proteins were denatured at 70°C for 10 minutes immediately prior to loading. For immunoblotting, 50 μg of total cellular protein per lane were separated by blot 4% to 12% Bis Tris Plus gradient gel and transferred to PVDF (polyvinylidine difluoride) membranes using iBlotting dry transfer system (Lifetechnologies, Cat. no. IB1001). The membranes were incubated with primary antibody LC3B (Abcam, Cat. No. ab48394) or GAPDH (Meridian life science, Cat. No. H86504M) and secondary antibody peroxidase-conjugated goat anti-rabbit or goat anti-mouse IgGs in iBind solution for 2.5 hours (iBind western system, life technologies). The specific protein signals were visualized using chemiluminescent peroxidase substrate and exposing the membranes to the high performance chemiluminescene film for detection. Protein loading was verified by detection of GAPDH using mouse anti-GAPDH monoclonal antibody.
Monocytes were transduced with CellLight BacMam 2.0 (lifetechnologies) to visualize lysosomal marker LAMP1 (C10504), early endosome Rab5 (C10587) and late endosome Rab7 (C10589). Transduction was performed simultaneously with differentiation of MoDCs at the 5th day. Cells were transfected with 30 PPC (particle per cell) for 24 hours. At the 6th day MoDCs were harvested and immature phenotype (CD1c+DC-SIGN+CD83-CD14-) by flow cytometry. Quantifications of the fluorescent intensities and co-localization within infected cells were done by NIS-Elements BR and AR software.
These studies were determined by the Human Assurance Committee at Georgia Regents University to be human subject exempt, due to the use of anonymized peripheral blood samples for monocytes.
|
10.1371/journal.pgen.1002845 | Rates of Gyrase Supercoiling and Transcription Elongation Control Supercoil Density in a Bacterial Chromosome | Gyrase catalyzes negative supercoiling of DNA in an ATP-dependent reaction that helps condense bacterial chromosomes into a compact interwound “nucleoid.” The supercoil density (σ) of prokaryotic DNA occurs in two forms. Diffusible supercoil density (σD) moves freely around the chromosome in 10 kb domains, and constrained supercoil density (σC) results from binding abundant proteins that bend, loop, or unwind DNA at many sites. Diffusible and constrained supercoils contribute roughly equally to the total in vivo negative supercoil density of WT cells, so σ = σC+σD. Unexpectedly, Escherichia coli chromosomes have a 15% higher level of σ compared to Salmonella enterica. To decipher critical mechanisms that can change diffusible supercoil density of chromosomes, we analyzed strains of Salmonella using a 9 kb “supercoil sensor” inserted at ten positions around the genome. The sensor contains a complete Lac operon flanked by directly repeated resolvase binding sites, and the sensor can monitor both supercoil density and transcription elongation rates in WT and mutant strains. RNA transcription caused (−) supercoiling to increase upstream and decrease downstream of highly expressed genes. Excess upstream supercoiling was relaxed by Topo I, and gyrase replenished downstream supercoil losses to maintain an equilibrium state. Strains with TS gyrase mutations growing at permissive temperature exhibited significant supercoil losses varying from 30% of WT levels to a total loss of σD at most chromosome locations. Supercoil losses were influenced by transcription because addition of rifampicin (Rif) caused supercoil density to rebound throughout the chromosome. Gyrase mutants that caused dramatic supercoil losses also reduced the transcription elongation rates throughout the genome. The observed link between RNA polymerase elongation speed and gyrase turnover suggests that bacteria with fast growth rates may generate higher supercoil densities than slow growing species.
| A 9-kb module called the “supercoil sensor” was used to measure supercoil density at 10 positions in the 4.8-Mb Salmonella Typhimurium chromosome. The sensor includes a Lac operon flanked by a pair of directly repeated DNA–binding sites for the γδ recombinase. Measurements of chromosomal supercoil levels and the RNA polymerase elongation rates were made at various positions within the 6 potential macrodomains of the chromosome. Transcription and gyrase catalytic rates were mechanistically linked. Gyrase mutants with impaired activity caused the loss of from 30% to >95% of the diffusible supercoiling throughout most of the chromosome, while treatment with rifampicin that temporarily blocked transcription restored most of the lost supercoils in gyrase mutants. A gyrase defect also caused transcription elongation rates to decrease across the chromosome, and a mutation that reduced RNA polymerase efficiency increased average chromosome supercoiling levels. A model in which topoisomerases act close to highly transcribed operons to equilibrate the supercoil flux generated by transcription suggests that matched rates of gyrase turnover and transcription elongation speed determine the average supercoil density in bacterial chromosomes.
| Negative supercoiling in bacterial DNA is generated by gyrase, which is composed of GyrA and GyrB proteins organized as A2B2 tetramers [1]. The average supercoil density of large bacterial chromosomes and small plasmid DNA is influenced by mutations in gyrase and two other topoisomerases. Topo I is a type Ia topoisomerase that breaks and rejoins DNA with a one-strand mechanism [2]. The enzyme is encoded by the essential gene topA [3] and it removes negative supercoils in a cofactor-independent reaction to protect chromosomes from toxic R-loops that can form at sites of high transcription [4]. Topo IV is a hetero-tetramer of ParC and ParE proteins in the form C2E2 [5]. With extensive homology to gyrase, Topo IV breaks both DNA strands simultaneously during the reaction cycle [2] and relaxes both positive and negative supercoils in steps of two supercoils per cycle in ATP-dependent reactions. Although Topo IV influences the supercoil density of chromosomal and plasmid DNA [6], its primary function is thought to be decatenation of sister chromosomes during final stages of chromosome segregation [7].
Changing the average supercoil density (σ) alters the efficiency and phenotype of many proteins involved in DNA replication [8], chromosome segregation [9]–[10], RNA transcription [11]–[13], homologous and site-specific recombination [14], and transposition [15]. Supercoil levels vary with growth conditions, and topoisomerase mutations arise as evolutionary adaptations in bacterial populations undergoing long-term growth on a monotonous carbon source [16]–[17]. Other than topoisomerases, our understanding of the roles of enzymes that contribute to the average supercoil density is poor, in part, because measuring supercoil density at specific locations of a 4 Mb chromosome is technically challenging.
Classical techniques used to measure chromosome supercoiling, like the ethidium bromide titration of nucleoids in sucrose gradients [18], give only an average supercoil density of the entire chromosome. The most common alternative method infers an average chromosomal supercoil density from the linking number of small plasmids in the same cell [19]. We developed techniques to monitor the supercoil-dependent movement of chromosomal DNA strands in vivo [8], [20]–[21]. The γδ site-specific recombination system uses supercoil diffusion to drive the assembly of a precise 3-node synapse of directly repeated Res sites (Figure 1A) [22]–[23]. Once a synapse forms, phosphodiester bond exchange leads to deletion of the intervening DNA segment without any accessory factors from E. coli [24]. The interwound DNA strands synapse by slithering and branching (Figure 1B). Slithering displaces two opposing strands along the axis of interwound loops. Branching rearranges the structure with new loops that grow and ebb laterally. If branching and slithering is unobstructed, resolution efficiency increases as the level of diffusible negative supercoiling increases, and deletions form rapidly and efficiently in vitro [25] and in vivo [26].
To analyze supercoiling at multiple locations, a 9 kb module called the “supercoil sensor” was developed [8]. It contains an entire Lac operon (lacIZYA) plus a selectable gentamycin resistance gene (Gn) flanked by directly repeated Res sites (Figure S1). The ends of the module are directly repeated Frt sites, which can be used to insert or extract sensors at unique chromosomal loci using the yeast 2 µ Flp recombinase (see Figure S1). The deletion efficiency of a LacI-repressed supercoil sensor is 50-fold more sensitive than a gyrB-lacZ promoter fusion, which varies by only 2-fold and has been used in many studies of chromosome topology in E. coli [12], [27]–[28].
The graph in Figure 1C illustrates the resolution response to negative supercoiling. Solid squares represent in vitro recombination rates (left Y axis) for endpoint assays carried out with plasmid DNAs with different supercoil densities (X axis). In vivo, about half of chromosomal σ is constrained (σC) and half is diffusible (σD) so that when σ = −0.060, σD≥0.030. The scale on the right Y axis shows the resolution response to σD in vivo. Calculations of the apparent supercoil densities use the bottom half of the curve because after X-ray-induced relaxation of diffusible supercoiling, the E. coli chromosome retained a constrained supercoiling value of σD = 0.030 [29]. Resolution efficiency at σD = −0.030 is about 50% (blue arrow). When resolution efficiency approaches 100%, the σD≥−0.040 (red arrow). We assume that in vivo reactions fall to 0 at σD≤0.004.
Three regions near the Salmonella ribosomal rrnG operon have different supercoil properties during exponential growth in rich medium [13]. Recombination between γδ Res sites flanking the 5 kb rrnG operon was less than 1% because the presence of 60–80 RNA polymerases in the transcribed track blocked supercoil branching and slithering required for synapse. RNA polymerase unwinds a segment of the template strand at the active site, which represents −1.7 constrained supercoils per enzyme [30]; the accumulated supercoil density within the rrnG operon approaches σC = −0.290. When σC increased, σD decreased, but temporary interruption of RNA transcription by addition of rifampicin increased resolution efficiency 60 to 100-fold [13]. This result confirmed our earlier finding that highly transcribed genes are barriers to supercoil diffusion in the chromosome [31]–[32].
In 1987, Liu and Wang proposed that RNA polymerase generates two supercoiling domains during transcription [33]. The rationale was that rather than RNA polymerase rotating around DNA, the DNA duplex rotates (relative to the cytoplasm) due to the large inertial mass of polymerase, its associated transcription factors, and ribosomes that bind and translate the nascent mRNA during transcription elongation [34]. This model predicts a supercoil density difference with increased (−) supercoiling in DNA upstream and a loss of (−) supercoils downstream from expressed operons. We tested this model by placing a sensor upstream of the rrnG promoter and downstream of the transcription terminator. The upstream sensor had a 75% resolution efficiency compared to 28% for the downstream sensor [13], confirming the twin domain model and indicating a differential supercoiling value of Δ σD = +0.014 [13].
Previously, we measured what happens to twin domain supercoiling in strains with a mutant of Topo I (topA217) and TS gyrA205 and gyrB1820 mutants [13]. Each mutant caused the supercoil differential to increase in regions flanking the rrnG operon. In cells with the topA217 mutation, upstream resolution efficiencies rose to 97% compared to 38% downstream. Conversely, a gyrA205 mutant caused downstream resolution to fall to 9% compared to 60% resolution upstream. Most strikingly, a gyrB1820 mutation caused downstream resolution to fall to 1% while recombination efficiency was 11% in the upstream domain. Local supercoiling levels were able to rise and fall dramatically at opposite ends of a highly transcribed operon in cells growing at permissive temperatures.
Here, we measured Salmonella chromosome supercoiling levels and transcription elongation rates using supercoil sensors at multiple positions covering the 6 macrodomains of E. coli. Our results show that rates of gyrase supercoiling and transcription elongation are linked. Temperature sensitive mutations in gyrase and Topo IV caused significant changes in genome-wide negative supercoil levels, even when cells were grown at a permissive temperature (30°). Transcription played a causal role in the supercoil losses because supercoiling rebounded after addition of rifampicin (Rif), which blocked transcription initiation. Our model is that transcription kinetics determine the optimal catalytic speed for gyrase, and the average chromosome supercoil density is an integral function of topoisomerases and RNA polymerase working in tempo together.
During DNA synthesis, gyrase and Topo IV collaborate to remove (+) supercoils generated by fork movement [35]. However, their contribution to the dynamics of transcription has remained largely untested. We showed previously that some TS gyrase mutants cause a decline in (−) supercoiling at permissive growth temperatures in twin domains of the rrnG operon [13]. To study the general impact of transcription on chromosomal supercoil density, we evaluated 6 TS topoisomerase mutants for their influence on supercoil density near the origin of replication. Strains with TS alleles of gyrase and topo IV were constructed with a supercoil sensor placed between gidB and atpI (Figure S1). The Atp operon encodes a group of 9 highly expressed membrane proteins that generate ATP using the energy of the proton motive force across the cytoplasmic membrane. Each strain also carries the plasmid pJBRes 30′, which expresses a form of resolvase with a 30 min cell half-life.
All 4 subunits of gyrase and Topo IV were tested. GyrA contains the catalytic tyrosine residue that carries out DNA cleavage and re-ligation during the supercoiling reaction (Figure S2 A). NH6016 carries the gyrA213TS allele (R358-H), which has a mutation located in the DNA-binding and cleavage domain [36]. Cultures were grown at 30° and doubling times were measured for each strain during mid log before resolution assays were carried out (Table 1 and Table 2). The complete derivation and genetic structure of each strain used in this manuscript is listed in Table S1.
Strain NH6016 had the same doubling time as WT (39±1 min) but the resolution efficiency fell from 81±3% for WT to 58±1%, representing a 28% loss in recombination efficiency. To compare alleles, we define a term Mutant Impact Factor (MIF) to be the resolution efficiency of the WT strain divided by the resolution efficiency of an isogenic mutant. A large MIF indicates a dramatic change in supercoiling. NH6016 had a significant MIF of 1.4. A gyrA209TS allele (G597-D) in NH6019 alters the second ß-propeller of GyrA, which contributes to DNA-looping that forms a chiral (+) node [37]. The gyrA209 doubling time increased from 39±1 min to 45±3 min, which is a 15% decrease in growth rate. The resolution efficiency in this strain fell to 30±12%, resulting in a MIF of 2.7. Thus, two gyrATS mutants had reduced resolution efficiency, which indicates a loss of (−) supercoiling even at the permissive temperature of 30°.
The GyrB subunit encodes the ATP-binding domain of gyrase, which couples ATP binding to a large conformation shift that drives negative supercoiling reactions (Figure S2 B). Two GyrB mutants were tested. NH6028 contains the gyr652 mutation (R436-S), which alters a Mg++ binding domain that coordinates structural conformation changes during DNA cleavage [38]. This enzyme has a low kcat relative to WT gyrase [8], and the doubling time of strains with this mutation increased by 36% to 53±3 min. The resolution rate in NH6028 was 7±3% (Table 2), and the resulting MIF of 12 indicates dramatic loss of supercoil density. The scale in Figure 1 predicts a change of +0.027 in the supercoil sensor for an apparent σD≤−0.012. A second allele, gyrB1820TS (C56-Y) alters the ATP binding domain that dimerizes and then hydrolyzes two ATPs during the catalytic cycle [39]. This mutant is the most severe allele in our Salmonella gyrase collection. The doubling time at 30° increased by almost 50% to 58±4 min. NH6037 had a resolution efficiency of 8±6%, resulting in a MIF of 10. Thus, the supercoil sensor shows that two TS gyrA and two TS gyrB alleles cause significant losses in (−) chromosome supercoiling at the ATP operon location.
Topo IV has a subunit structure, catalytic mechanism, and sensitivity to drugs novobiocin and fluoroquinolones that is similar to gyrase [28]. Although its primary function is decatenation and untangling of sister chromosomes prior to segregation and cell division [7], Topo IV relaxes both (−) and (+) supercoils in vitro and contributes to the dissipation of (+) supercoils during DNA replication in vivo [35]. Therefore, we tested the impact of TS Topo IV alleles on chromosomal supercoiling. The ParC subunit catalyzes DNA breakage/reunion during strand passage reactions, and NH6040 has the parC281TS (P556-L) mutation, which resides in a region with no known function. This mutant showed no difference in growth rate from the WT (39±1 min) and resolution efficiency was 76±1%, which is close to the WT (81±4%) with a MIF of 1.1. ParE functions like GyrB, binding and hydrolyzing ATP to fuel cycles of strand transfer. The parE206 (V67-M) mutation in strain NH6043 is in the ATP binding domain of Topo IV, and the doubling time at 30° increased by 33% to 52±2 min. The resolution efficiency was 59±4% (Table 1), yielding a MIF of 1.4. Therefore, the defect in the ParE206 subunit of Topo IV caused a supercoil loss comparable to GyrA213.
The E. coli chromosome appears to have multiple levels of organization. In addition to 10 kb domains that restrict supercoil diffusion [40], a long range order called macrodomains has been proposed [41]. Macrodomains represent segments of 0.6 to 1 Mb that may coalesce in the folded chromosome. The first indication of macrodomain structure came from fluorescent in situ hybridization (FISH) with the Ori and Ter regions occupying distinct positions near the opposing cell poles in newborn cells [42]. The Boccard laboratory extended the E. coli framework to include three additional segments and two less structured regions by measuring the interaction frequencies of pairs of λ attachment sites distributed across the chromosome [41], [43]–[44]. Although the efficiency of λ site-specific recombination shows variation at specific points in Salmonella [45], the macrodomains proposed for E. coli may or may not be conserved along with gene order that is shared between these species.
Supercoil levels in all potential macrodomains were measured by introducing sensors into 7 more sites in Salmonella to include at least one measurement in each E. coli macrodomain (Figure 2). E. coli chromosome map coordinates are notated in minutes that reflect the HFR transfer time of each genetic region during a standard mating experiment (1–100 min). The Salmonella chromosome has a gene order that is highly congruent with E. coli, but with numerous inserted gene islands, the genome size is 5% larger. To compensate, map coordinates in Salmonella are described in units of 100 centisomes (Cs) with the same starting position as in E. coli. The largest E. coli macrodomain is Ori, which spans 930 kb of DNA. The corresponding segment in Salmonella extends from Cs 81 to Cs 1 in Figure 2 (the green arc). Ori includes 4 of the 7 ribosomal RNA operons and many highly transcribed genes involved in transcription and translation. 70% of the RNA polymerase in rapidly dividing cells is confined to this chromosome sector. The module at Cs 85 (Table 1) is near the left edge of the Ori macrodomain in replichore 2; it had a recombination efficiency of 81±4%. In NH6008, resolution efficiency was tested in another segment of the Ori domain in replichore 1. The sensor disrupts the Salmonella gene STM4442, which encodes a small putative “cytoplasmic protein” at Cs 96. NH6008 matched NH6000 with a resolution efficiency of 85±3% (Table 1).
Two domains reside exclusively in replichore 1. The Right Unstructured region is shown in black (Figure 2) clockwise of oriC. The smallest macrodomain in E. coli (560 kb), it extends from Cs 1 to Cs 13 in Salmonella. A sensor was inserted at Cs 9 in NH6007 between ampH, which encodes a beta-lactam binding protein, and sbmA, a gene encoding an inner membrane ABC transporter. NH6007 had a resolution efficiency of 73±12%. The Right macrodomain of E. coli spans 600 kb from Cs 13 to Cs 26. In NH6006, a sensor was inserted at Cs 21. This is the only position in which a reporter lies between two divergently transcribed genes. These genes are STM0951, which encodes a “cytoplasmic protein” transcribed in the counterclockwise direction, and STM0952, which is a transcription regulatory protein transcribed in the clockwise direction. The recombination efficiency in NH6006 was the highest measured at 92±2%.
Two E. coli macrodomains reside entirely in replichore 2. The Left Unstructured region is a 550 kb sector. The comparable region of Salmonella is shown in Figure 2 as a black arc counterclockwise of oriC running from Cs 81 to Cs 62. A sensor inserted at Cs 71 lies between STM3261, which encodes a galacticol-1-phosphate dehydrogenase, and STM3262, a putative repressor in strain NH6001. The resolution efficiency was 82±2%. (Table 2, Figure 2). The Left macrodomain in E. coli is an 892 kb region extending from Cs 62 to Cs 43, shown as a blue arc. Two modules were placed in this segment of Salmonella. In NH6002, a module resides at Cs 58 between smpB, which makes a small protein that may bind the SsrA subunit of the SsrA/SsrB two-component regulatory complex [46], and pseudogene STM2689. A second module in this sector is integrated between STM2135, which encodes an inner membrane protein, and the protease-encoding gene yegQ at Cs 45. The deletion efficiencies of NH6002 and NH6003 were 80±3% and 73±6%, respectively.
The macrodomain that lies across from Ori in E. coli is the Ter domain (purple arc), which is a 780 kb region of E. coli. Ter has 24 copies of a unique 14 bp site called matS that is found uniquely in this segment. The matS sites bind MatP, which may organize them into a single focus in cells with a chromosomal MatP-GFP fusion. One model is that 23 Ter domain loops are formed with a central hub of MatP protein [47]. In Salmonella, the Ter domain may be a smaller 560 kb region with only 14 predicted matS sites [47] (black lines in Figure 2). In NH6005, a sensor was inserted at Cs 33 between the pseudogene STM1553 and STM1554, which encodes a putative “coiled coil protein.” The resolution efficiency here was lower than any other site tested in the survey, 45±6%.
The cumulative average resolution efficiency of sensors located at 7 regions (excluding the Ter domain) was 81±7% and the apparent σD = −0.038±.002. There was no statistically significant variation in supercoil levels from the Ori to the terminus. At Cs 33, the resolution efficiency of 45% is roughly half that measured at the other 7 sites. At this location, a Res site is only 470 bp from dif. We believe that resolvase binding to the site nearest dif may be occluded by DNA-binding proteins unique to the region. These proteins include matS-MatP complexes [47], the FtsK DNA translocation motor complex [48], a high affinity site for Topo IV [49], and XerC/D proteins that bind dif to catalyze complex topological reactions that untangle and separate sister chromosomes [50].
With the exception of the Ter macrodomain, the genes and mechanisms that organize the Ori, Right, and Left macrodomains are undefined. Supercoiling is an important factor in bacterial DNA condensation, so we tested the impact of topoisomerase mutations in all domains using the supercoil sensor. In the strain series NH6019-NH6027, each strain has the gyrA209TS allele, which showed a MIF of 2.7 at the Cs 85 position (Table 2). The resolution efficiency measured at 7 positions (excluding position Cs 33) showed more variability than the WT set (Table 2 and Figure 2, green characters). The average recombination efficiency was 28±9%. This drop corresponds to a mean MIF of 3. The estimated change in σD relative to WT at 7 positions was +0.013. Previous work from other laboratories showed that growth of E. coli cells stopped when supercoiling dropped to this level [3], [51]–[52].
Supercoil losses at 7 locations in gyrB652TS mutants were larger than those measured at Cs 85 (Table 2). The resolution efficiency at Cs 71 – NH6029, Cs 45 – NH6031, Cs 33 – NH6033, Cs 21 – NH6034, and Cs 9 – NH6035 were all less than 1% (Figure 2, purple characters). The estimated value of σD dropped from −0.038 to an apparent σD = <−0.004. Resolution efficiency was near the detection limit at Cs 58 – NH6030 (2±1%) and at Cs 96 – NH6036 (3±1%). Averaging across 7 points on the chromosome, the mean recombination efficiency was 2±2% and the MIF was 40. Surprisingly, this strain with a greatly relaxed chromosome has a doubling time only 36% longer than WT (53±3 vs. 39±1).
To see if Topo IV has a related genome-wide supercoil phenotype, the parE206TS allele of Topo IV was tested (Figure 2, blue characters). The resolution efficiency at Cs 85 (59±4%) was similar to results at positions Cs 71, 66±4%; Cs 58, 64±3%; Cs 45, 60±6%; Cs 21, 58±3%; Cs 9, 51±4%; and Cs 96, 63±4% (Figure 2, blue characters). The mean recombination efficiency at 7 chromosomal positions fell to 60±5% for a MIF of 1.4. Again, the Ter macrodomain at Cs 33 showed lower resolution efficiency than all other locations. NH6048 recombined at 27±2% compared to 45±6% in WT.
Replication and transcription generate positive supercoils in regions downstream of replisomes and highly expressed operons, respectively. To understand the reason a TS GyrB mutant loses most of the detectable diffusible chromosomal supercoiling, we tested the role of transcription. Like E. coli, WT Salmonella is organized into 400–500 domains that limit supercoil diffusion [21]. Topo I relaxes negative supercoils generated upstream of highly transcribed regions. If gyrase can't supercoil DNA at rates matching the rotation speeds downstream of the 7 ribosomal RNA operons, the multiple tRNA genes, and 30 highly transcribed protein-encoding genes that are spread out over the chromosome, then transcription could run down reservoirs of stored supercoils in low transcribed regions. Supercoil depletion might also be a consequence of having all highly transcribed genes oriented in the same direction as replication, presumably to mitigate effects of head on replisome-RNAP collisions [53].
To test the role of transcription in supercoil regulation, a strain set carrying the severe gyrB1820TS mutation was constructed (Table 3, NH6037-NH6114). Similar to cultures with the gyrB652 mutation, the resolution efficiency was at the detection limit (≥1%) at all locations other than Cs 85 (Table 3, Figure 3, black numbers). The average recombination value at 7 sites was 1.6±3% (entering values of 0.5% for measurements <1%) and the MIF mean was 50.
We added Rif to aliquots of each culture immediately after the 10 min resolvase induction period. Rif blocks transcription initiation, but elongation and termination occurs normally; no cell death was associated with drug treatment. After 30 min of incubation, the drug was washed out and the recombination efficiencies were measured after cells doubled more than twice, to allow chromosome segregation. Rif had a dramatic impact on resolution efficiency (Table 3, Figure 3, black numbers). At Cs 85 - NH6037, resolution was 8±6% with a MIF of 10. Rif treatment increased resolution 7-fold to 56±5% and the MIF dropped to 1.4 (Table 3 and Figure 3, purple numbers). Dramatic results were also observed at 6 other locations. In strains with modules at Cs 9, Cs 71, and Cs 96, resolution rose at least 10-fold from ≤1% to 11±2%, 11±6%, and 9±5% respectively (Figure 4, Table 3). At Cs 45, resolution increased from <1% to 22±7%. The largest improvement was observed in NH6109 at Cs 58 where resolution increased >60 fold from <1% to 57±2%. Like the gyrB652TS strain set (Figure 2), resolution efficiency in the Ter domain at Cs 33 was low and remained low after Rif addition, rising only from <1% to 2±1%. Overall, excluding the Ter domain, Rif addition reduced the MIF mean from 50 to 3.
The increase in resolution after Rif addition supports the hypothesis that general transcription can deplete supercoil levels when gyrase is impaired. But what would happen to supercoiling in strains with a WT complement of topoisomerases and a slow RNA polymerase? If catalytic rates of transcription and supercoiling are under selection to match, would such cells experience a general supercoil increase? Deletion of 6 amino acids in the ß′ subunit (RpoC Δ Δ215–220) makes a form of RNA polymerase with a constitutive low transcription rate for stable RNA, including the 7 ribosomal RNA operons [54]. This mutation was introduced to Salmonella strain set (NH6206-NH6215), which included sensors upstream and downstream of rrnG, increasing the number of test locations to 10 (Table 4). The doubling time of the mutant growing at 30° increased by 28% from 39±1 min in WT to 50±2 min. Remarkably, the resolution efficiency increased throughout the mutant chromosome, except for one position at Cs 21, which was within experimental error of matching the highest efficiency in the WT RNA polymerase strain (92±2 - 86±7, Table 4, Figure 4). The WT mean resolution efficiency at 10 positions was 74±18%, whereas the RpocΔ215–220 average was 85±8% with a MIF of 0.87. A 13% increase in resolution represents an apparent mean change of ΔσD = −0.004. Interestingly, the impact of the rpoC mutation was greatest at positions where the WT resolution levels were lowest. For sensors adjacent to the rrnG operon at Cs 57.64 and Cs 57.65, the upstream sensor increased from 75±6% resolution to 83±4% and the downstream location changed from 28±3% to 69±5% resolution. The downstream location had a MIF of 0.41, proving that locations where gyrase worked the hardest benefited the most from reduced transcription rates.
In 1973 Pato, Bennett, and von Meyenberg discovered that the rates of transcription elongation and translation were closely matched for most genes in E. coli [34]. Could the transcription rate include a role for gyrase? We measured the coupled lacZ transcription/translation kinetics at 8 locations in WT and gyrB1820 mutants. The method is outlined in Figure 5 A. Cultures grown in minimal medium plus glucose were sampled at 10 sec intervals and placed on ice in lysis buffer [55]. The first three samples established a baseline, then IPTG was added to each culture at a final concentration of 1.5 mM, and 10 sec sampling was continued. After all samples were collected, the chromogenic substrate ONPG was added to timed reactions that ran at 37° for 1.5 to 3 h. The transcription rate in nucleotides per second (nt/sec) is calculated as the length of the LacZ transcript (3072 nt) divided by the lag time to the start of a linear increase in enzyme activity (Figure 5A). Each strain was tested in triplicate using different colonies, and the transcription rates with one standard deviation are shown for WT (red) and GyrB1820 mutants (black) in Figure 5B.
Unexpectedly, coupled transcription/translation rates varied at different positions in the Salmonella genome. The fastest transcription speeds were 69±9 nt/sec at Cs 85 and 62±10 nt/sec at Cs 58. These sites were 45% faster than the 38±1 nt/sec rate measured at Cs 9. The average elongation rate in WT cells across all positions was 52±10 nt/sec. The impact of a gyrB1820TS mutation was tested in strain set NH6222-NH6229. Elongation at 7 positions fell to a uniform mean of 32±6 nt/sec, which is 40% slower than the average of these positions in WT. Again, the Ter domain at Cs 33 was different. Transcription/translation rates at dif fell from 56±6 nt/sec to 16±2 nt/sec in gyrB1820. These results together with the experiments using Rif suggest to us that unique factors influence resolution efficiency and transcription near dif. Nonetheless, throughout most of the genome, and in at least 5 macrodomains, transcription/translation rates and gyrase supercoiling efficiency were covariant.
Three results show that the mean supercoil density of Salmonella DNA is determined by a mechanism that links the catalytic efficiency of gyrase to the elongation rate of transcription. First, TS alleles of GyrB caused a broad and dramatic depletion of (−) supercoiling throughout the Salmonella genome (Figure 2). This effect was largely reversed by temporarily blocking transcription with Rif (Figure 3). Second, supercoil densities rose above the WT level in cells carrying a mutant ß′ subunit (RpocD215–220) (Figure 4). Third, the GyrB1820 mutation caused the rates of coupled LacZ transcription/translation to decrease from the WT mean of 52±10 to 32±6 nt/sec over most of the genome (Figure 5).
The impact of TS mutations in both GyrA and GyrB on resolution efficiencies for cells growing exponentially at a permissive temperature of 30° was unexpected (Table 1, Table 2, and Figure 2). There are three plausible explanations for this reduction in recombination rates: 1) When the catalytic rate of gyrase was slowed by mutation, the loss of negative supercoiling downstream of highly transcribed genes was spread across the genome. 2) The slow growth rate in gyrase mutants caused a drop in resolvase expression that limited recombination. 3) A slow growth rate induced increased expression or rearrangement of nucleoid-associated-proteins (NAPs) that constrained (−) supercoiling [56]–[58] and/or occluded resolvase binding to Res sites.
A change in the resolvase expression level does not explain the γδ recombination results for two reasons. First, we analyzed resolvase in WT and mutant strains using Western blots. The resolvase band at 21 KDa appeared after thermo-induction in all strains tested (Figure S3.) The expressed resolvase contains an SsrA degradation tag appended as the terminal 11 amino acids, and this tag limited the in vivo protein half-life to under 30 min [21]. Resolvase disappeared during a 30 min incubation at 30° following the 42° incubation, including the cells treated with Rif [21]. Whereas the resolvase band intensity varied somewhat between different strains, the band variation did not correlate with the ratios of WT to mutant catalytic resolution efficiency. These results agree with our earlier finding that a 5–10 fold decrease in resolvase expression seen in stationary phase cells does not limit resolution [20].
Second, a much more compelling argument comes from the Rif experiment shown in Figure 3. When transcription was unobstructed, the resolution efficiencies in gyrB1820 strains were at the detection limit of 1% at 7of 8 genome locations. But resolution increased at Cs 85 and Cs 58 to 56% when Rif was added to the cultures after downshift to 30°. An interruption in transcription restored 70% of the resolution efficiency at these two sites compared to that seen in WT cells. The length of time that cells were exposed to resolvase was previously shown to be a factor in resolution efficiency [21], and after transcription inhibition, resolvase exposure dropped from about 40 min (induction plus incubation time) to 25 min or less. Yet, at 7 chromosome locations covering 5 macrodomains (excluding Ter), the mean resolution efficiency for GyrB1820 without Rif was 1.6±3% with a MIF = 50 (81%/1.6%), and after Rif, the mean efficiency rose to 26%±21% with the MIF shrinking to 3 (81%/26%). The variation in resolution around the chromosome after the addition of Rif could mean that extra factors contributed (i.e. increased constrained structure by H-NS). But Rif would not be expected to lower the abundance of NAPs around the genome. The parsimonious explanation is that supercoil density was restored once transcription was reduced by Rif during the incubation at 30°. In our view, non-supercoil factors might account for 2–3 fold of a 50-fold impact that a gyrB1820 mutation exerts on resolution. An independent measure of chromosomal supercoil structure at specific locations would be a very useful tool to help resolve the issue.
Various theories for regulating chromosomal supercoiling have been proposed since gyrase was discovered and the importance of negative supercoiling was revealed [59]. One model is that cells maintain a uniform level of supercoiling throughout the chromosome by varying levels of gyrase and Topo I [60]. When chromosomes experience a significant supercoil decline, the change is sensed by the promoter of GyrB, which increases expression by about 2-fold along with about 100 other ORFs [12]. When excessive levels of (−) supercoiling accumulate, transcription from the Topo I promoter [61] increases along with about 200 other ORFs [11], [62]. This system clearly modulates expression of GyrB and Topo I. However, because supercoil levels do not respond to changes in enzyme levels in a dramatic way, we view this as a fine tuning system [63]. For example, increasing or decreasing the abundance of Topo I or gyrase by 10% resulted in only a 1.3% change in DNA supercoil density [64], which would be equivalent to a MIF of 1.05 or 0.95 for each 10% difference in enzyme level. By contrast, when the Salmonella GyrB protein was expressed at 10% of normal levels in WT E. coli, the toxic effect caused the disappearance of most cells containing the plasmid [10]. We speculate that 50 slow or uncoordinated chimeric gyrase variants working in a WT background may cause sporadic supercoil disruptions with toxic consequences, perhaps by promoting RNAP blockades to the fast moving replisomes.
A second model proposes that a long range supercoil gradient exists within the bacterial chromosome [65]. The origin of replication was proposed to have the highest supercoiling level with σ = −0.068, and the terminus was predicted have the lowest σ = −0.043 with a smooth transition along the genome [65]. If constrained and diffusible supercoiling densities partition equally along the gradient, this model predicts resolution efficiencies of the supercoil sensor to decline from 75–80% near oriC to 15–20% at the terminus. However, our data disagree with this model. Our resolution assays showed equal recombination in 5 different macrodomains. Moreover, previous investigators used supercoil-responsive promoter fusions to lacZ and luxAB to test supercoiling levels in both the E. coli and Salmonella chromosomes [66]–[67]. They both found uniform levels of supercoiling along the genome, although none of their test positions were located close to highly transcribed genes or the dif site.
Our data suggest that the Liu and Wang model of twin domains of local opposite-handed supercoiling is a dominant force near the 30–50 highly transcribed genes [13]. Although the impact of transcription may be limited to a 10 kb zone from the point of origin, like transpositions immunity in Mu [68], a persistent loss of supercoil density during transcription can spread, causing slight or dramatic relaxation of chromosome DNA structure, depending on how an allele modifies gyrase supercoiling efficiency. We propose that the impact of RNA polymerase on global supercoil density is linked to transcription speed. At 30° in WT Salmonella, the elongation rate ranges from 45–60 nt/sec at different points around the genome (Figure 5). This causes DNA rotations of 4–6 supercoils per second. WT gyrase processively supercoils DNA at 4–5 sc/sec at 30° (Rovinskiy and Higgins, manuscript submitted) and Topo I removes negative supercoils at this rate in single molecule studies. Any condition that reduces gyrase supercoiling without directly reducing transcription kinetics or Topo I activity would cause supercoil density to decline across the genome.
We were surprised that a TS mutant of Topo IV also lost significant negative supercoiling at the permissive growth temperature. The common wisdom is that Topo IV functions primarily at the end of replication to decatenate sister chromosomes and allow complete segregation [7]. However, recent work shows that the C-terminal domain of Topo IV interacts with the hinge region of the MukB condensin [68], implicating Topo IV in processes occurring near the fork. Perhaps, in conjunction with DNA compaction, Topo IV removes (+) supercoils of transcription to prevent disruptive interactions between replisomes and RNA polymerase [69]–[70].
The third piece of experimental evidence supporting the mechanistic linkage between rates of gyrase and RNAP catalysis is the decreased rate of coupled transcription/translation throughout the chromosome in cells carrying a gyrB1820 mutation (Figure 5.) The mean WT transcription rate was 52±11 nt/sec for the 7 sites, which fell to 32±6 nt/sec in GyrB1820 strains. We propose that there is a strong selection for matching catalytic rates of gyrase supercoiling with transcription elongation. When cells have a sluggish gyrase, transcription/translation slows down. In cells with reduced transcription efficiencies, like the rpoC Δ215–220, WT gyrase boosted supercoiling above the level in WT cells (Figure 4). Excess supercoiling wastes ATP, increases the likelihood that cells form toxic R-loops at locations of high transcription, and increases the susceptibility of chromosomal DNA to oxidative damage by free radicals that attack single stranded regions more efficiently than double stranded DNA.
Many components are now known to contribute to the transcription/translation enterprise. The list of factors includes DksA, NusA, NusG, MFD, Rho, RfhA, GreA, GreB, RNAP, ppGpp, tmRNA, Topo I, cAMP, cyclic GMP, and ribosomes. Interestingly, when the Cozzarelli lab set up a genetic screen to identify genes that might encode “domainins,” i.e. proteins controlling supercoil density, they uncovered a surprise gene, dksA, in addition to genes for the expected NAPs [28]. DksA mediates the stringent response by binding to RNA polymerase and placing ppGpp near the catalytic active site [71]. DksA makes sense in our model, because it changes transcription rates under stringent conditions. We recently tested deletions in GreA and GreB, which are proteins that promote processive transcription and salvage polymerases that have stalled in mid-stream. Mutants of both subunits raised the average supercoil density of Salmonella (Chesnokova and Higgins, unpublished results). These observations, along with older experiments showing that mutations in RNA polymerase influence cellular resistance to the gyrase inhibitors novobiocin and nalidixic acid [72], increase our confidence that RNA polymerase and its associated factors play a central role with gyrase in controlling the global supercoiling average.
DNA supercoiling has generally been studied as a mechanism to control gene expression by modulating promoter activity around the chromosome [11], [73]. In vivo, 300 E. coli genes are reported to change expression within 5 min after DNA relaxation by drug treatment [12]. However, three problems with studying supercoil regulation of transcription are often ignored or not considered important. First, transcription increases upstream and decreases downstream supercoil levels respectively, so the act of transcription would put the promoter in a zone of increased supercoil density. The increase in supercoiling is substantial [13], so after 200 E. coli genes are induced by increased (−) supercoiling, a different mechanism would be needed to turn them off.
Second, most of the chromosomal changes in transcription detected after DNA relaxation are 2-fold differences. The transcription rate of the Lac operon, which varies several hundred fold from the uninduced state to maximum expression, also changes elongation rate by 2-fold, according to its chromosome position (Figure 5). This is not a result of genetic adaptation, but is dictated by local differences in chromosome dynamics. Dissecting 2-fold changes in gene expression is a daunting task and can be the result of 3 different 1.3-fold causes. It is unclear how many of the 300 E. coli supercoil responsive genes actually improve fitness and how many represent regulatory noise that is insignificant from a physiological perspective.
Third, many investigators rely on plasmids to estimate chromosomal supercoiling and to gauge the effects of mutations on chromosomal structure. But plasmids can be misleading. By assuming that pUC19 was a good reporter of chromosome supercoil density, we completely missed the impact of transcription and the large supercoiling change in a Salmonella gyrB652 mutant chromosome [26]. pUC19 lacks strong promoters and when plasmids from WT and gyrB652 strains were compared, they differed by only 1 topoisomer. One cause for these chromosome/plasmid differences is that plasmids are single domain elements, except when they have an anchoring element or two active transcription units moving in opposing polarities [33]. In single domain plasmids, positive and negative supercoils cancel out by diffusing around the circle [74]. For plasmids with strong promoters, the primary topological effects are changes in constrained supercoil density associated with each added RNA polymerases [75]–[76].
E. coli and Salmonella have a 15% supercoil difference that changes the phenotype of multiple proteins contributing to chromosome dynamics [10]. Could species-specific amino acid substitutions in gyrase orthologs fine-tune supercoil densities in these closely related organisms? Recent work suggests this could be the case. One difference between the E. coli and Salmonella GyrA proteins is the amino acid sequence and length of the acidic amino acid-rich C-terminal tail (Figure S2). This C-terminal segment controls DNA looping of the pinwheel domain and establishes the supercoil reverse point for E. coli gyrase [77]. Moreover, when the E. coli GyrA ortholog was compared to M. tuberculosis (M. tb.) GyrA, the latter protein lacked C-terminal features present in E. coli [78]. In vitro supercoiling tests confirmed that both the speed and endpoint of M. Tb gyrase supercoiling are lower than those measured for the E. coli enzyme. Secondly, the GyrB subunit has the ATP binding site that fuels the supercoiling reaction. Whereas Salmonella GyrB protein is toxic in E. coli, the reverse is not true. Salmonella tolerates the E. coli GyrB chromosomal substitution, and the average supercoiling level of this strain increased at multiple chromosomal locations, including the region immediately downstream of rrnG (Rovinskii and Higgins, unpublished data.) Therefore, both gyrase subunits contribute to the enzyme vmax and supercoil endpoint.
Three untested issues related to these finding are worth mentioning. They involve current limitations on the fluxuation of supercoil density that we can monitor, the implications of the Salmonella/E. coli comparison for other bacterial species, and the relevance of this work to gene expression in eukaryotes. First, our data represent the mean values of an ensemble of cells in different states of the cell cycle during rapid division in rich medium and during slower growth in medium containing a defined carbon sources. Many investigators assume that supercoil density in a bacterium is maintained at a static modulus so that small changes in supercoil density can be used to modulate gene expression. Our view is more dynamic. The constant thrust from highly transcribed genes causes local gradients of supercoil density to arise throughout the genome. If topoisomerase efficiency is changed by metabolism or mutation, supercoil change spreads across the genome. However, little is currently known about single cell metabolism. For example, yeast cells go through an ultradian cycle that oscillates between periods of reductive reactions of the TCA cycle followed by an oxidative phosphorylation phase that increases ATP concentration [79]–[80]. Expression of most yeast genes increases then declines in either the oxidative period (a few genes) or the reductive phase (most genes). This cycle is shorter than the cell cycle and it is usually studied in carbon- or phosphate-limited chemostats, where yeast self-synchronize with the acetate flux. But non-synchronized cells show the same periodic variations of gene expression [81]. Bacteria could have similar behavior because they share with yeast a mechanism to increase or decrease acetate metabolism using a sirtuin-dependent acetylation/deacetylation of acyl-CoA synthase [82]. A test of cyclical transcription and negative supercoiling pulses for periods shorter than a bacterial cell cycle is challenging and would require different approaches.
The second interesting issue is the relationship between optimum growth rates and supercoil levels in different bacterial species. A significant supercoiling difference exists between E. coli and Salmonella [10]. WT E. coli cells grow faster than Salmonella and they double every 25 min at 37° in rich medium where transcription elongation rates top out at 90 nt/sec [83]. The fastest doubling time for Caulobacter crescentus is 2 h [84], presumably because rapid growth rates are not important for life in open ocean water. M. tb. has a doubling time of 14 h, contains only 1 ribosomal RNA operon, and encodes a gyrase with significantly slower vmax and a lower supercoiling endpoint than E. coli [78]. To understand chromosome structure in prokaryotes other than E. coli and Salmonella, methods will be needed to measure in vivo transcription/translation rates and to define supercoil density at multiple chromosome locations.
Third, might a pattern of covariant tempos of transcription elongation and topoisomerase turnover apply to eukaryotes? The short answer seems to be yes. In yeast, a type 1B topoisomerase (Topo I) relaxes both positive and negative supercoils and is active during transcription. Single molecule studies [85] showed that yeast Topo I relaxes both (+) and (−) supercoils at ≥4 sc/sec, matching the yeast Pol II transcription elongation speed of 30–40 nt/sec [86]. Camptothecin caused little change in the Topo I-dependent relaxation of (−) supercoils, but the rate of (+) supercoil removal fell 40-fold in the presence of drug. When the topology of in vivo transcribed DNA was analyzed, camptothecin treated yeast cells produced highly (+) supercoiled DNA, and further transcription was impeded [85], [87]. This is similar to the behavior we see in gyrase mutants (Figure 2). Therefore, tuning transcription machinery to topoisomerase catalytic rates may be necessary for efficient gene expression in yeast as well as in other eukaryotes.
All strains in this work are derivatives of S. Typhimurium LT2, and their genotypes are listed in Table 1. Insertion mutagenesis was done using the λ Red recombineering method and the plasmids pSIM5 or pSIM6 [88]–[89]. The PCR amplification of drug modules for insertion into the chromosome and the electroporation conditions used to introduce DNA for homologous recombination were carried out as described previously [31].
Chromosomal recombinants were selected as antibiotic-resistant colonies on LB medium or as Lac+ colonies on minimal lactose medium. In each case the expected recombinant genotype was verified by PCR analysis using flanking PCR primers. Each recombinant was tested and shown to contain a cassette-modified allele with no WT allele present. Transduction crosses were performed as described previously using P22 HT105/1 int-201, a high-efficiency transducing variant of bacteriophage P22 [20].
The growth rate of individual strains was measured in early-mid log phase and calculated from the log slope of change over time of the OD650 between 0.01 and 0.4. Each strain was tested, starting from three independent colonies grown overnight and diluted 100 fold in fresh LB at 30°. Results are reported ±1 standard deviation from the average.
Plasmid pJB γδ 30′ was used to induce the expression of a modified form of γδ resolvase. In this plasmid, resolvase is controlled from the λPL promoter using the TS cI857 repressor [21]. In pJB γδ 30′, 11 residues were incorporated at the natural C-terminus of resolvase that makes an SsrA degradation tag, which targets the protein to degradation by the ClpXP proteosome. At position 9 of the 11 amino acid SsrA tag, a L9D substitution gives the protein a 30 min half-life in exponentially growing E. coli and S. Typhimurium cells [21].
Log-phase cultures growing in LB at 30° were sampled at a density of 50 Klett units. A 0.1 ml aliquot of each culture was placed in a 42°C shaking water bath for 10 min to induce Resolvase expression. The induced cells were immediately diluted with 2 ml of LB+Cm and incubated overnight at 30°C. On the following day, 100 µl aliquots of 10−6 dilutions of each culture were plated on LB medium or on NCE glucose minimal medium containing chloramphenicol and 5-bromo-4-chloro-3-indolyl β-D-galactoside (X-gal) plus 200 µM IPTG [13]. Plates were incubated for 2–3 days at 30°, and deletion frequencies were scored by counting the number of white colonies that reveal the loss of lacZ [20]. Each data point represents the average ±1 standard deviation of at least three independent experiments in which ≥200 colonies were counted for drug sensitivity or loss of lacZ expression.
To measure elongation rates in Salmonella, the β-Gal method described by Vogel was used [55]. A flask with 20 ml of minimal AB medium supplemented with glucose [90] was inoculated with 2 ml of a fresh overnight culture, and growth was carried out at 30° or 37°. Samples (0.5 ml) of each culture having an OD600 between 0.2–0.4 were added to 500 ul ZS buffer (chilled at 4°C) containing 200 ug/ml chloramphenicol. Three samples were taken at 10 s intervals for the background measurement, and IPTG at a final concentration 1.5 mM was added to each culture. Aliquots of 500 ul were withdrawn every 10 sec and mixed with 500 µl ZS buffer for about 4 min. 100 µl chloroform was added to each sample followed by 200 ul ONPG which initiated a timed enzyme reaction. Reactions incubated at 30° for 1.5–4 hrs to allow development of appropriate levels of color were stopped by the addition of 500 ul Na2CO3. The OD420 and OD550 values were taken, and standard Miller Units were calculated as described [91]. Lag times and the transcription rates were determined using three independent colonies for each strain with results reported as the average value ±1 SD of the mean.
To make a 6 amino acid deletion (ΔKKLTKR) in the rpoC gene of Salmonella, we used the method described by Sharan et al. [92]. Four primers were designed with a 20 bp overlap of N- and C- terminal segments of RpoC. The primer pair of Rpoc fwd (CGCGAAGATGGGGGCGGAAG) and RpoC rev del (aaggcttccagcagtttgat acgcttggtttcggagttgg) and RpoC frd del2 (ccaactccgaaaccaagcgt atcaaactgctggaagcctt) and Rpoc rev (CCATCCAGCGGAACCAGCGG) both make 130 bp PCR products carrying the upstream and downstream region of RpoC with a deletion in both fragments. These products were combined and amplified with the RpoC fwd and RpoC rev primers to generate a 220 bp PCR with the 6 amino acid deletion at the center. The PCR DNA was introduced by electroporation of WT LT2, which had been pre-induced for recombineering function encoded on the pSIM6 plasmid. After incubation for 2 hrs in LB to allow recovery, 200–400 cells were plated onto 5 LB plates and incubated at 30° for 2 days. Small colonies were observed in both WT and fis mutant plates at a frequency of 1/500 to 1/1000. Three colonies were picked and subjected to PCR sequencing using the outside primers and DNA template from a negative control. Every small colony we tested carried the deletion called RpoC (ß′ Δ115–220) by Bartlett et al. [54] and gave no WT RpoC sequence.
|
10.1371/journal.pntd.0006529 | Minimum requirements and optimal testing strategies of a diagnostic test for leprosy as a tool towards zero transmission: A modeling study | The availability of a diagnostic test to detect subclinical leprosy cases is crucial to interrupt the transmission of M. leprae. In this study we assessed the minimum sensitivity level of such a (hypothetical) diagnostic test and the optimal testing strategy in order to effectively reduce the new case detection rate (NCDR) of leprosy.
We used the individual-based model SIMCOLEP, and based it on previous quantification using COLEP data, a cohort study of leprosy cases in Bangladesh. The baseline consisted of treatment with Multidrug therapy of clinically diagnosed leprosy cases, passive case detection and household contact tracing. We examined the use of a leprosy diagnostic test for subclinical leprosy in four strategies: testing in 1) household contacts, 2) household contacts with a 3-year follow-up, 3) a population survey with coverage 50%, and 4) a population survey (100%). For each strategy, we varied the test sensitivity between 50% and 100%. All analyses were conducted for a high, medium, and low (i.e. 25, 5 and 1 per 100,000) endemic setting over a period of 50 years.
In all strategies, the use of a diagnostic test further reduces the NCDR of leprosy compared to the no test strategy. A substantial reduction could already be achieved at a test sensitivity as low as 50%. In a high endemic setting, a NCDR of 10 per 100,000 could be reached within 8–10 years in household contact testing, and 2–6 years in a population testing. Testing in a population survey could also yield the highest number of prevented new cases, but requires a large number needed to test and treat. In contrast, household contact testing has a smaller impact on the NCDR but requires a substantially lower number needed to test and treat.
A diagnostic test for subclinical leprosy with a sensitivity of at least 50% could substantially reduce M. leprae transmission. To effectively reduce NCDR in the short run, a population survey is preferred over household contact tracing. However, this is only favorable in high endemic settings.
| The annual number of new leprosy cases has been stable in the past decade, indicating that transmission has not been yet been interrupted. As current control seems to be insufficient to bring down the number of cases, there is a need for novel tools to interrupt transmission. A diagnostic that permitted diagnosis of subclinical cases will likely be fundamental to achieve elimination and ultimately eradication. In this study we assessed the minimum sensitivity level of such a (hypothetical) diagnostic test and the optimal testing strategy in order to effectively reduce the new case detection rate (NCDR) of leprosy. We showed that a diagnostic test for subclinical leprosy could substantially reduce the NCDR in a high, medium and low endemic population. A significant impact could already be achieved at a test sensitivity level of 50%. To effectively reduce the NCDR in the short run, a population survey is preferred over household contact tracing. However, this is only favorable in high endemic settings, as in medium and low endemic settings testing in a population survey requires many more people to be tested and treated to prevent one new leprosy case.
| Leprosy is an infectious disease caused by Mycobacterium leprae, affecting the skin, peripheral nerves, the mucosa of the upper respiratory tract and the eyes [1]. The most likely route of transmission of M. leprae is via the aerosolic route [2]. Individuals, who have close and frequent contact to a patient with leprosy, in particular within households, have the highest risk of acquiring the infection and developing leprosy [3, 4]. Currently, the main strategy to control leprosy, as recommended by the World Health Organization (WHO), is early detection of cases and treatment with multidrug therapy (MDT) [5]. Leprosy is diagnosed by clinicians based on the clinical signs and symptoms, along with the use of slit-skin smears and biopsies to respectively detect the presence of acid-fast bacteria and determine type of leprosy histologically [6].
Although the prevalence of leprosy has dropped immensely in the last 30 years, worldwide still more than 200,000 new cases of leprosy are detected annually [5]. This number has remained fairly stable over the last decade, indicating that transmission has not yet been interrupted. Global elimination has been a target since 1991, and more recently the target for leprosy has been set to achieve zero transmission [7, 8]. However, it is clear that the current strategy is not sufficient to achieve the goals within a reasonable time frame [9, 10]. For this reason, alternative strategies should be considered. Previous modeling studies have shown that treating people during the subclinical stage of leprosy has a larger impact on the new case detection rate (NCDR) than early (clinical) diagnosis and treatment [11]. Therefore, interventions such as the provision of chemoprophylaxis (antibiotics) or immunoprophylaxis (vaccination) to contacts of leprosy cases could substantially further reduce the NCDR [12–14]. Nevertheless, these are not yet routinely available nor accepted. A more efficient approach would be the use of a diagnostic test that allows identification of M. leprae infected individuals who are at risk of developing leprosy and constitute the major source of transmission.
The identification and validation of new sensitive biomarkers for M. leprae infection and (subclinical) leprosy is currently investigated in several leprosy endemic areas [15–19]. Host immunity after M. leprae infection is determined by host genetics, leading to a complex immuno-pathological spectrum associated with either dominant cellular or humoral immunity. The immune-mediated pathological leprosy spectrum compels detection of M. leprae infection to be based on multiple, diverse biomarkers specific for cellular as well as humoral immunity [15, 19]. The use of serological proteomics can help to unravel the biological pathway in the immunomodulation of leprosy for diagnostic purposes [20]. In addition, pathogen-based approaches identifying the presence of M. leprae in skin smears of contacts and patients may offer additional tools for prophylactic targeting [21, 22]. The availability of a specific and robust diagnostic test to detect infected individuals lacking clinical symptoms, would not only be beneficial to identify and treat cases at early stages before irreversible damage occurs but may also be crucial for breaking the transmission chain. As the prevalence of leprosy decreases, leprosy health care has been integrated into general health care causing decreased clinical expertise for diagnosing leprosy, leading to extended delays of diagnosis and as a result maintenance or re-emergence of infection. Implementation in general health care of a user- and field-friendly diagnostic test specific for leprosy may accommodate for the lack of leprologists in the field.
In this study we aim to identify the minimum requirements of an as yet hypothetical diagnostic test for subclinical leprosy (i.e. identifying infected individuals who will progress to disease) and its potential impact on the NCDR. In order to predict the added value, non-linearity of infectious disease patterns in the transmission process needs to be taken into account. For this purpose, we will use the individual-based model SIMCOLEP, which models M. leprae transmission and control of leprosy in a population structured by households. The model has previously been used to estimate future NCDR trends in Bangladesh, India, Brazil, and Indonesia, and to test the impact of various interventions targeting household contacts [9–11, 23, 24]. We use SIMCOLEP to assess the impact of a diagnostic test on NCDR under various assumptions of sensitivity, ranging from 50 to 100%. Furthermore, we investigate the optimal strategy of using such a test: household contact testing or a population survey. As the impact of test strategies might be dependent on the endemicity level, all analysis were conducted in a high (25 per 100,000), medium (5 per 100,000) as well as low (1 per 100,000) endemic setting [25].
We used the individual-based model SIMCOLEP that simulates the spread of M. leprae in a population structured in households. It models life-histories of individuals, which are born and placed into households. Over time, individuals can create their own household or move to another household after marriage, during adolescence or after becoming a widow(er). Deaths of individuals are determined by death rates at birth [23, 26].
In the model, M. leprae transmission occurs when a susceptible individual has contact with an infectious individual. In the model, susceptibility of an individual to leprosy is randomly assigned. We assumed that 20% of the population is susceptible, implying that 80% will not develop leprosy, although the proportion of susceptibles is likely to be lower [1]. This assumption was made because a previous modeling study showed that assuming 20% susceptibles provided the best fit and that results did not significantly differed from assuming 5% or 10% susceptibles [23]. Two transmission processes are modeled separately: transmission in the general population and within-household transmission. The latter can be regarded as an additional probability of acquiring the infection if household contacts are infected. Infectivity is determined by the product of the contact rate, both in the general population and within households, and the probability of infection during a contact. An infected individual develops either paucibacillary (PB) or multibacillary (MB) leprosy, which is randomly determined based on the distribution of the type of leprosy. We assumed that only MB leprosy is infectious. After infection, an individual enters the asymptomatic state, which on average lasts 4.2 years for PB and 11.1 years for MB. Afterwards the individual proceeds to the symptomatic state, which lasts on average 5 years for PB leprosy after which self-healing occurs. An MB leprosy case remains symptomatic until treatment or death. The natural history of leprosy is modeled following Meima et al. [27]. The model also replicates control measures including treatment with multidrug therapy (MDT), passive case detection, and household contact tracing. All detected cases receive MDT treatment, and will not be infectious from then on. Relapses occurred with a rate of 0.001 per year. 90% relapses to MB and 10% to PB. A full description of the model can be found in Fischer et al. 2010 and Blok et al. 2015 [9, 23].
We used previously published quantifications of the model based on the population and leprosy epidemiology in Nilphamari and Rangpur districts in Bangladesh. A full description of the data used to parameterize the birth, household movements and deaths of individuals, and leprosy epidemiology can be found in Fischer et al. 2010 [23].
Transmission of M. leprae was previously calibrated using data from the COLEP study [4]. The COLEP study population consisted of contacts of 1037 consecutively found new patients with leprosy in Nilphamari and Rangpur districts in Bangladesh. The data contain information about the genetic distance (i.e. kinship) of almost 22,000 contacts and were used to fit the prevalence of cases among contacts of different household sizes, the prevalence of cases among different types of relatives in 2003. Quantifications were not updated to more recent numbers, because a trial with chemoprophylaxis started afterwards. Our main focus was to estimate the impact of a diagnostic test for detection and subsequent effective treatment of M. leprae infected individuals progressing to disease, in addition to the mainstay control program, as recommended by the WHO [25]. The transmission contact rate in the general population and within household was set to 1.33 and 0.98, respectively [23]. The baseline NCDR of Nilphamari and Rangpur was 27 per 100,000 and the MB/PB ratio 20/80 in 2003 [4].
The leprosy control program in the modeled area was well-organized [28]. It consisted of passive case detection with an annual detection delay of on average of 2 years (standard deviation 1.4 years) [29] and continuous household contact tracing with a coverage of 90% [23]. During contact tracing only clinical leprosy cases can be diagnosed. We further included the protective effect of BCG vaccination prior to the infection, which was set to 60% [13]. In order to reflect a high, medium and low endemic setting, we ran the model given the current control until it reached a NCDR level of 25 per 100,000, 5 per 100,000 and 1 per 100,000. This was obtained after 2, 18 and 40 years, respectively.
In different scenarios, we evaluated the impact of using a diagnostic test for subclinical leprosy on the NCDR trend under various sensitivity levels, ranging from 50% to 100%. We neglected lower levels of sensitivity, because it is not expected that such a diagnostic test will come on the market. Individuals testing positive will be treated with MDT [30].
We evaluated four testing strategies with a diagnostic test for subclinical leprosy in addition to the current leprosy control:
All scenarios were compared to the no-testing strategy which only consisted of the current leprosy control program. Leprosy cases under treatment were excluded from testing. Testing compliance in individuals was assumed to be 100%. All strategies were evaluated in terms of its impact on the NCDR and its benefit and costs. The benefit of a testing strategy was measured by the number of prevented new leprosy cases calculated as the difference between the new cases (detected and undetected) of a testing strategy and the no-testing strategy. The costs of a testing strategy were measured in terms of number needed to test and treat. As tests that are not completely predictive may produce false positives, we assessed the number needed to treat given a test specificity of 100%, 95% and 85%. The impact on NCDR, and the benefits and costs of all strategies were evaluated in a high, medium and low endemic setting. The modeled time span was a 50-year period.
In a sensitivity analysis, we varied the passive case detection delay and MB/PB ratio. First, we increased the passive case detection delay to six years, representing an area with a less well-organized leprosy control program in place. Second, we increased the MB/PB ratio to 65/35, representing an area with relatively more MB than PB cases. The latter is relevant because we assume that only MB leprosy is contagious.
Fig 1 presents the impact of a diagnostic test used in household contacts without follow-up and with a 3-year follow-up, and a one-time population survey with a coverage of 50% and 100% on the NCDR under various assumptions of test sensitivities. In all strategies, the use of a diagnostic test further reduces the NCDR of leprosy compared to the no-testing scenario. Household contact testing without follow-up decreases the NCDR gradually over time with slightly larger effects at higher levels of test sensitivities. If household contacts are additionally followed-up for 3 years, the impact of lower test sensitivities increases to the level of the highest test sensitivity. In a population survey of 50% and 100%, a substantial reduction can be seen in the short run followed by a gradual decline afterwards. The impact varies with sensitivity levels of the test: a higher sensitivity level of the test corresponds with a larger impact on the NCDR. In the short run, both population survey testing strategies result in a larger decrease in the NCDR than household contact testing. In a population survey with a coverage of 100%, the NCDR could be 40% lower within 10 years with a test sensitivity of at least 60% (See S1 Fig). In the long run, the impact of household contact testing on NCDR may be larger than that of a population survey, depending on the coverage and test sensitivity. In a high endemic setting, a NCDR of 10 per 100,000 could be reached within ten and eight years in household contact testing without and with follow-up, respectively. In a population survey testing strategy with a coverage of 50% and 100% this level could be reached within six and two years, respectively. The relative impact clearly varies with endemicity level with the high endemic setting showing the largest relative decrease, followed by medium and low endemic setting.
The number needed to test in order to prevent one new leprosy case decreases with level of sensitivity and increases with the endemicity level of the setting (see Fig 2). Household contact testing without and with follow-up require substantially fewer individuals to be tested to prevent one leprosy case compared to both population survey strategies. In a one-time population survey, testing in a high endemic setting requires the least number of people to be tested to prevent one new leprosy case. In a medium and low endemic setting this number is up to four and twenty-three times higher, respectively.
Table 1 summarizes the benefits and costs of all testing strategies with an assumed test sensitivity of 70% in a population of 1 million after 10 years. The numbers of prevented cases are approximately up to ten times higher in a population survey than in household contact testing. The highest number of prevented cases could be achieved in a high endemic setting. However, testing in a population survey would require substantially more people to be tested and treated than household contact testing. If the test is not completely specific, the number needed to treat further increases as a result of an increased number of false positives. This increase is much larger in a population survey than household contact testing, as more people are tested in a population survey. The costs of household contact testing decrease with the level of endemicity, whereas the costs of a population survey do not differ much across endemic settings.
S1 Table provides the results of our sensitivity analysis. The number of prevented new leprosy cases is smaller in a setting with a less well-organized leprosy control program compared to a well-organized setting (i.e. Nilphamari and Rangpur districts in Bangladesh). In a setting with a high MB/PB ratio, more new leprosy cases can be prevented in a 10-year period. The number needed to test and treat is to a large extent comparable across all leprosy settings.
This paper assessed the impact of the use of a (hypothetical) diagnostic test for subclinical leprosy cases with a sensitivity that ranged between 50% and 100% on the NCDR in household contact testing and one-time population survey strategies. All strategies showed an additional reduction in the NCDR over time compared to the current control for all levels of sensitivity in a high, medium and low endemic setting. Testing in a population survey yields a higher impact on the NCDR in the short run compared to household contact tracing. In terms of prevented cases, a population survey is preferred over household contact testing. However, a population survey requires much more people to be tested and treated, especially in medium and low endemic setting, compared to household contact testing.
Our findings indicate that a test with a sensitivity as low as 50% could already result in a significant reduction of the NCDR. This suggests that to reduce transmission the availability of a diagnostic test for subclinical cases is more important than the level of sensitivity, which is very promising. Moreover, the impact of a test with a low test sensitivity could be increased by repeating the test in individuals testing negative. We showed that a test with a sensitivity of 50% used in a household contact testing strategy with a 3-year follow-up could reach a similar impact as a test with a sensitivity of 100%.
Testing in a population survey is most favorable to achieve short term reductions of NCDR in a high, medium and low endemic setting. The lower impact of household contact tracing compared to a population survey, especially in the short run, is primarily due to the limited exposure of the test that is confined to a household with an average size between 4 to 5 persons. A population survey would also result in a higher number of prevented new leprosy cases. However, it requires testing of many individuals and thereby commitment by the national leprosy control programs. Such an approach is only favorable and feasible in high endemic settings. Our results show that the numbers needed to test to prevent one new leprosy case in a medium and low endemic setting are approximately fourfold and twenty-threefold that of a high endemic setting, respectively.
Although tests specificity is of less importance if we aim to reduce transmission of M. leprae, it is relevant for determining the optimal testing strategy. If a test is not completely specific, the number needed to treat may increase dramatically. As the number of people tested is higher in a population survey than in household contact testing, it would also produce many more false positives (See Table 1). A solution to reduce the number of false positive is to apply a two-step approach, whereby a second test with high specificity is added [30].
Our study also highlights that the target of zero transmission might be difficult to achieve, even with a well-organized control program and the use of a diagnostic test for subclinical cases. Over a 50-year period, our model predicted that the NCDR in the high, medium and low endemic setting would be at most reduced to very low levels: 0.10–0.50, 0.01–0.10, and 0.005–0.025 per 100,000 in a high, medium and low endemic setting, respectively (Fig 1). This is mainly the result of the combination of a relative long incubation period of leprosy and detection delay. Areas that reach very low levels of NCDR require continuous monitoring for many more years to prevent maintenance or even re-emergence of M. leprae.
In this study we did not consider the impact of a test for asymptomatic infection for future NCDR, because such a test is only relevant if we assume that all infected individuals are infectious (i.e. can infect other people). In our model, we assumed that only infected individuals who progress to MB leprosy are infectious, implying those who progress to PB leprosy and those who do not develop leprosy do not contribute to the transmission. This assumption was made because it is still poorly understood whether and to which extent asymptomatic infected are infectious.
This study used previously published quantifications of the model based on the population and leprosy epidemiology in Nilphamari and Rangpur districts in Bangladesh. The advantage of using this quantification is that it was based on the COLEP study, which includes a large amount of very detailed information on this population, including data on the prevalence of cases among contacts of different household sizes and the prevalence of cases among different types of relatives [4]. The downside is that quantifications are based on 2003 data, which does not truly reflect current leprosy epidemiology. However, for the purpose of assessing potential impact of a diagnostic test this is less of an issue.
The primary concern of this study is about the extent to which our results are generalizable to other regions or countries with leprosy in the world. The leprosy control program in the Nilphamari and Rangpur districts of Bangladesh is more extensive than usual. The relative short detection delay of two years and active household contact tracing with a coverage of 90% is not common in leprosy control programs in other regions or countries. For that reason, we conducted a sensitivity analysis in which we increased the passive case detection delay to six years. Results show that fewer leprosy cases were prevented (S1 Table). The qualitative findings of our main results (the importance of sensitivity and the impact and efficiency of household contact tracing versus population survey) did not differ when implemented in a setting with a less well-organized control program.
Another concern with respect to generalizability is the distribution of the MB/PB leprosy. In Bangladesh the MB/PB ratio is approximately 20/80. Since we assumed in our model that only MB cases are infectious, the impact of a diagnostic test might differ in areas with strikingly different MB/PB ratios, such as in Brazil (65/35) or Indonesia (80/20) [5]. In a sensitivity analysis, we showed that in a setting with a MB/PB ratio of 65/35, the benefit of a diagnostic test is larger compared to our main results (S1 Table). This is because of the earlier detection of relative more MB cases.
Finally, this study did not look into combining the use of a diagnostic test for subclinical leprosy with additional novel strategies, such as a chemoprophylaxis, as this is beyond the scope of this study. Earlier modeling studies have shown that providing contacts with a single-dose of rifampicin (SDR) chemoprophylaxis can reduce the transmission of leprosy over time [11, 24]. It can be expected that adding chemoprophylaxis to our testing strategies would further reduce the transmission of leprosy, especially if the test sensitivity is not optimal. In that case, the contacts that were false negative might benefit from SDR.
We showed that a diagnostic test for subclinical leprosy could substantially reduce transmission in a high, medium and low endemic population. The test sensitivity influences the impact on transmission in a population survey, but even with levels as low as 50% a substantial reduction could be achieved. To effectively reduce the NCDR in the short run, a population survey is preferred over household contact tracing. However, this is only favorable in high endemic settings, as in medium and low endemic settings testing in a population survey requires many more people to be tested and treated to prevent one new leprosy case.
|
10.1371/journal.ppat.1006834 | A role for domain I of the hepatitis C virus NS5A protein in virus assembly | The NS5A protein of hepatitis C virus (HCV) plays roles in both virus genome replication and assembly. NS5A comprises three domains, of these domain I is believed to be involved exclusively in genome replication. In contrast, domains II and III are required for the production of infectious virus particles and are largely dispensable for genome replication. Domain I is highly conserved between HCV and related hepaciviruses, and is highly structured, exhibiting different dimeric conformations. To investigate the functions of domain I in more detail, we conducted a mutagenic study of 12 absolutely conserved and surface-exposed residues within the context of a JFH-1-derived sub-genomic replicon and infectious virus. Whilst most of these abrogated genome replication, three mutants (P35A, V67A and P145A) retained the ability to replicate but showed defects in virus assembly. P35A exhibited a modest reduction in infectivity, however V67A and P145A produced no infectious virus. Using a combination of density gradient fractionation, biochemical analysis and high resolution confocal microscopy we demonstrate that V67A and P145A disrupted the localisation of NS5A to lipid droplets. In addition, the localisation and size of lipid droplets in cells infected with these two mutants were perturbed compared to wildtype HCV. Biophysical analysis revealed that V67A and P145A abrogated the ability of purified domain I to dimerize and resulted in an increased affinity of binding to HCV 3’UTR RNA. Taken together, we propose that domain I of NS5A plays multiple roles in assembly, binding nascent genomic RNA and transporting it to lipid droplets where it is transferred to Core. Domain I also contributes to a change in lipid droplet morphology, increasing their size. This study reveals novel functions of NS5A domain I in assembly of infectious HCV and provides new perspectives on the virus lifecycle.
| Hepatitis C virus infects 170 million people worldwide, causing long term liver disease. Recently new therapies comprising direct-acting antivirals (DAAs), small molecule inhibitors of virus proteins, have revolutionised treatment for infected patients. Despite this, we have a limited understanding of how the virus replicates in infected liver cells. Here we identify a previously uncharacterised function of the NS5A protein–a target for one class of DAAs. NS5A is comprised of three domains–we show that the first of these (domain I) plays a role in the production of new, infectious virus particles. Previously it was thought that domain I was only involved in replicating the virus genome. Mutations in domain I perturb dimer formation, enhanced binding to the 3’ end of the virus RNA genome and prevented NS5A from interacting with lipid droplets, cellular lipid storage organelles that are required for assembly of new viruses. We propose that domain I of NS5A plays multiple roles in virus assembly. As domain I is the putative target for one class of DAAs, our observations may have implications for the as yet undefined mode of action of these compounds.
| Hepatitis C virus (HCV) is a member of the Flaviviridae family of enveloped, positive-strand RNA viruses [1]. It is estimated to infect up to 170 million individuals globally [2]. HCV causes inflammation and fibrosis in the liver via damage to hepatocytes. Over time, chronic infection progresses to significant fibrosis and may lead to cirrhosis with a risk for decompensation and hepatocellular carcinoma (HCC) [3].
The HCV genome is approximately 9,600 nucleotides in length and comprises 5’ and 3’ untranslated regions (UTRs) flanking a single open reading frame encoding a 3,000-residue polyprotein precursor [4,5]. Co- and post-translational proteolytic cleavage of this precursor by cellular and viral enzymes yields the structural proteins: Core, envelope glycoproteins E1 and E2, and the p7 ion channel, which are involved in viral assembly, along with non-structural (NS) proteins NS2, NS3, NS4A, NS4B, NS5A and NS5B. With the exception of NS2, which is dispensable for RNA replication and may control virus assembly, the other 5 NS proteins (NS3-NS5B) are necessary and sufficient for membrane-associated RNA replication [6]. By definition, NS proteins are expressed in virus-infected cells but are not incorporated into virus particles; although directly involved in RNA synthesis, they also play roles in modulation of host defence mechanisms and virus assembly [7,8]. In addition to NS5A, whose roles are detailed below, recent studies have provided evidence for the involvement of NS3, NS4B and NS5B in the later stages of the virus lifecycle–namely virus assembly and release [9–13].
Over the past few years there have been extraordinary advances in the therapy for HCV infection–the standard IFN and ribavirin therapy has been rapidly superseded by combination therapy with a range of direct-acting antivirals (DAAs) targeting the NS3/4A protease, NS5A, and the NS5B RNA-dependent RNA polymerase. As one important target of DAAs, NS5A is a ~450 amino acid multi-functional phosphoprotein that has essential roles throughout the virus life cycle. It is composed of three domains (I, II and III) linked by low complexity sequences (S1A Fig), although in recent years domains II and III have been increasingly defined as a single, unstructured domain. The protein is anchored to phospholipid membranes by an N-terminal amphipathic helix (residues 1–33) in a manner essential for replication [14]. The structure of domain I has been solved by three independent groups using X-ray crystallography. These studies revealed four different dimeric forms of domain I from genotype 1a and 1b with the same monomeric unit, but different dimeric arrangements [15–17]. By primary sequence comparison, domain I of NS5A shares a high sequence homology among all hepaciviruses, while domain II and III exhibit a lower level of homology [18–22]. These observations suggest that domain I has critical and well conserved functions that are common to all hepaciviruses, whereas the functions of the other two domains may be specific to individual viruses. In this regard, it is generally accepted that the function(s) of domain I are required exclusively for genome replication [23], many culture-adaptive mutations map to this domain, and the majority of domain II together with all of domain III are dispensable for replication [24–26].
In HCV infected cells, NS5A localizes to the endoplasmic reticulum (ER), virus–induced multiple-membrane vesicles (MMV) that host RNA replication complexes (also called the membranous web), and to lipid droplets. The MMV contain the NS proteins NS3-NS5B and virus RNA and represent sites of active genome replication [27–30]. The precise role of NS5A in genome replication remains obscure, however it is widely accepted that this is mediated by binding to viral RNA [31,32], other NS proteins and interactions with various cellular factors, including vesicle-associated membrane protein-associated proteins A and B (VAP-A, VAP-B), cyclophilin A (CypA) and phosphatidylinositol-4-kinase IIIα (PI4KIIIα), which are required for HCV replication [33–37].
Following RNA replication, nascent viral genomes need to be transported from the sites of RNA replication to distinct, as yet poorly characterised, sites of virus assembly. Here infectious virus particles are generated, bringing together the structural proteins and the viral genome to be packaged in a temporally and spatially organized manner [8,38]. An increasing body of evidence points to a role of NS5A in coordinating this process, possibly by transporting the genome RNA to assembly sites and delivering it to the Core protein for encapsidation. A further level of complexity arises from the fact that, compared to other enveloped positive-strand viruses, a key feature of infectious HCV particles is that they exhibit unusually low buoyant densities, while particles with higher buoyant densities are less infectious [39–43]. Indeed highly purified HCV particles are rich in lipids and cholesterol resembling very-low density lipoproteins (VLDL) [44,45]. This property requires that cellular lipid droplets (LDs), lipid storage organelles surrounded by a phospholipid monolayer, are involved in HCV assembly.
Both Core and NS5A are targeted to lipid droplets, and this recruitment is essential for virus assembly. Mutations that block either Core or NS5A localization to LDs inhibit virus production, suggesting that LDs are intimately involved in virus particle assembly [46–48]. The function of NS5A in virus assembly has been mapped to domain III. Mutations close to the C-terminus of domain III disrupt the ability of NS5A to interact with Core, abrogate infectious particle formation and lead to an enhanced accumulation of Core on the surface of LDs [49]. In addition, a number of cellular NS5A-interacting partners have been implicated in LD function/targeting and virus assembly. These include Apolipoprotein E (ApoE), diacylglycerol acyltransferase-1 (DGAT-1), Annexin A2 and Rab18 [50–55]. Of note, both DGAT-1 and Rab18 have been reported to recruit NS5A on to LDs and are proposed to play roles in transporting NS5A (and most likely genome RNA) between replication sites and LDs/assembly sites [52,55]. Although virus encapsidation could occur at the LD, it is noteworthy that LDs are only surrounded by a phospholipid monolayer, therefore the virions cannot obtain their lipid envelope from them. Assembly of an infectious enveloped HCV virion particle must ultimately require that Core and virion RNA are transported from LDs [29] to a membranous compartment, possibly involving the ESCRT and/or endosomal pathways [56–58].
In this study, we present evidence that domain I of NS5A also plays a key role in the assembly of infectious virus. We identify two key surface exposed, conserved residues that, when substituted with alanine, retain genome replicative capacity but block the production of infectious virus. We show that these mutations inhibit the ability of HCV to perturb LD structure and distribution and disrupt the recruitment of NS5A to LDs. They also impair the dimerization of domain I and enhance the binding of domain I to the HCV 3’UTR RNA, revealing a role for these NS5A attributes in virus assembly.
In comparison with domain II and domain III, domain I of NS5A is highly conserved throughout all HCV isolates, and is also well conserved in related viruses such as GB virus type B (GBV-B) and the novel hepaciviruses that have recently been identified in a variety of species (S1B Fig). In addition, the structure of domain I has been determined by three independent groups [15–17]–all three studies agree on the monomer structure but show these monomers assembling into dimers with different monomer orientations and dimer interfaces (S1C Fig). In this study we initially set out to define residues in domain I that were required for viral genome replication. To this end, we first aligned amino acid sequences from 29 isolates representing all 7 HCV genotypes, together with 10 related viruses such as bat hepacivirus (BHV), GB virus-B (GBV-B), guereza hepacivirus (GHV), non-primate hepacivirus (NPHV) and rodent hepacivirus (RHV) (S1 Table). This analysis revealed 24 absolutely conserved residues (S2 Table). We then mapped these conserved residues on to the two genotype 1b structures (PDB 1ZH1 and 3FQM) of domain I to identify surface exposed residues, particularly those that are charged. This analysis identified 11 residues that were then targeted for alanine scanning mutagenesis and subsequent profiling in the context of the JFH-1 sub-genomic replicon (SGR) and infectious virus. In addition, a conserved surface exposed cluster (residues 153 to 158) was mutated collectively to alanine as these residues were located in close proximity on the tertiary structure (S2 Table).
To investigate the role of the selected conserved residues in domain I, the mutants were cloned into a previously described JFH-1 derived SGR (mSGR-luc-JFH-1) [25] in which the NS5A coding sequence was flanked by unique restriction sites generated by mutagenesis to facilitate sub-cloning. Importantly, these modifications did not alter the coding capacity of the polyprotein and had no effect on replication of the SGR [25]. RNAs transcribed from the mutant panel were electroporated into Huh7 cells and luciferase activity was measured at 4, 24, 48 and 72 h post electroporation (h.p.e.). The luciferase activity at 4 h.p.e. correlates with translation of input transcripts prior to onset of replication and subsequent time points were normalized to the 4 h.p.e. signal to account for electroporation efficiency. As a negative control an inactive mutant of the NS5B polymerase was used (GND) [59].
Nine of the mutations (Y43A, G45A, W47A, G51A, C59A, G60A, G96A, T134A and 153-158A) were shown to completely disrupt the ability of the mSGR-luc-JFH-1 to replicate in Huh7 cells (Fig 1A), being indistinguishable from the GND negative control. However, three mutants (P35A, V67A and P145A) were able to replicate, albeit at levels lower than wild type (WT). P35A exhibited a modest but non-significant defect, in contrast V67A and P145A replicated at significantly lower levels than WT (p<0.05) (Fig 1A). All mutants showed broadly comparable luciferase activity at 4 h.p.e., demonstrating that the replication phenotypes observed were not due to differences in electroporation efficiency (S2 Fig).
We then assessed whether the replication defects exhibited by these mutants could be due to the low permissibility of Huh7 cells for HCV replication, rather than a lack of replicative capacity. To test this we evaluated the mutation panel in Huh7.5 cells which were derived from Huh7 cells, and are highly permissive for HCV genome replication [60]. As shown in Fig 1B, those mutants that were unable to replicate in Huh7 cells (Y43A, G45A, W47A, G51A, C59A, G60A, G96A, T134A and 153-158A) exhibited the same phenotype in Huh7.5 cells, confirming that these residues are absolutely required for the function of NS5A in genome replication. However, the three mutants that were able to replicate in Huh7 cells, albeit at a lower level than WT, (P35A, V67A and P145A) were able to replicate more efficiently in Huh7.5 cells, reaching levels almost equivalent to the WT with modest but non-significant impairment (Fig 1B). However, it was important to confirm that this permissiveness in Huh7.5 cells was not a phenomenon that was specific for domain I. To this end, an SGR containing a mutation (D329A) within NS5A domain II [61], which we previously reported replicated approximately 5-fold lower than WT, was electroporated into both Huh7 and Huh7.5 cells. As shown in S3A Fig, D329A was also able to replicate more efficiently in Huh7.5, demonstrating that this effect was not specific for domain I.
We proceeded to confirm that the replication phenotypes observed resulted from the loss (or disruption) of a specific function of NS5A, rather than a defect at the level of polyprotein translation or proteolytic processing. To this end, all 12 mutations were cloned into a plasmid in which the expression of the NS3-5B proteins of JFH-1 was driven by the human cytomegalovirus (CMV) promoter (pCMV10-NS3-5B), thus allowing replication–independent expression of these replicase proteins (S3B Fig). These plasmids were transfected into Huh7.5 cells and cell lysates were analysed for protein expression by western blot at 48 h post transfection (hpt), using HCV NS3 as a polyprotein processing control. All 12 mutants expressed levels of NS5A and NS3 comparable to WT (p ≥0.1) (S3B Fig). This confirmed that the replication phenotypes of these mutants were not the result of effects on NS5A translation, stability and/or polyprotein cleavage.
To determine whether the attenuation of genome replication for P35A, V67A and P145A in Huh7 cells was also observed in the context of infectious virus, these mutations were sub-cloned into the full-length mJFH-1 infectious clone. This construct contains the same unique restriction sites flanking NS5A as mSGR-luc-JFH-1, and the nucleotide sequence changes did not affect the levels of virus assembly and release [25] Following electroporation of full-length virus transcripts into Huh7 cells we determined virus genome replication activity by quantification of the number of NS5A positive cells using the IncuCyte ZOOM at 48 h.p.e. as previously described [62]. As expected, replication of P35A, V67A and P145A in the context of infectious virus (Fig 2A) was consistent with the observation in SGRs (Fig 1A). P35A exhibited a modest reduction which was not significant, whereas V67A and P145A showed a ~100-fold reduction in replication and were indistinguishable from the GND negative control. Consistent with this replication phenotype, neither V67A nor P145A produced any infectious virus particles, either within the cells (intracellular virus), or released into the supernatant (extracellular virus) (Fig 2B). A different picture emerged when these mutant virus RNAs were electroporated into Huh7.5 cells. As shown in Fig 2C, replication of P35A was indistinguishable from WT, whereas both V67A and P145A showed only a modest defect. This result was confirmed by western blot analysis for NS5A and Core expression (Fig 2E). However, despite the restoration of genome replication to WT levels, V67A and P145A were unable to produce any infectious virus (Fig 2D). This phenotype mirrored that of the additional control used in this experiment, ΔE1-E2 (a deletion within the envelope glycoprotein coding region previously shown to be unable to assemble infectious virus)[25,49]. As noted previously [62], although the IncuCyte ZOOM allows for rapid automated quantification of virus titres, the sensitivity of the instrument does result in a high background. However, visual inspection of samples (for example see S4 Fig) confirmed the absence of infectivity for V67A, P145A and negative controls.
We conclude from these data that the two residues V67 and P145 are partially required for genome replication, as mutations of these residues resulted in a reduction of replication that could be rescued by the increased permissibility of Huh7.5 cells. In contrast these two residues are absolutely required for the assembly of infectious HCV particles. This result was surprising, as it is widely accepted that domain I of NS5A is exclusively involved in genome replication. The one exception to this is the report 10 years ago showing that alanine scanning mutagenesis of residues 99–101 or 102–104 had no effect on genome replication, but blocked release of infectious virus from Huh7.5 cells [44], although whether these mutants affected assembly of intracellular infectious virus was not determined. We reasoned that the ability of V67A and P145A to replicate to near WT levels in Huh7.5 cells offered the opportunity to assess the role of domain I in virus assembly, without any confounding replication defect that would make interpretation of the data difficult.
However, before analysing the phenotype of V67A and P145A in more detail, we confirmed that the phenotypes of these mutants were not due to the acquisition of an additional compensatory mutation during the cloning process. To do this, we generated revertant viruses in which the WT NS5A coding sequence was sub-cloned back into the V67A and P145A virus backbones. As shown in S5A Fig, following electroporation of revertant RNA into Huh7.5 cells, both genome replication and production of both intracellular and extracellular virus was restored to WT levels.
We considered that the failure of V67A and P145A to produce infectious virus was either due to a gross assembly defect such that no virus particles were generated, or that virus particles were assembled but were non-infectious. Such non-infectious particles might be empty, lacking the genome, or could exhibit some other more subtle defect such as a failure to associate with lipids. To test this hypothesis, culture medium from Huh7.5 cells electroporated with JFH-1 WT, P35A, V67A and P145A RNA was concentrated and fractionated by iodixanol density-gradient centrifugation. As controls, cells were electroporated with GND and ΔE1/E2 RNAs. Each fraction was analysed by quantitative RT-PCR (Fig 3A) to determine the presence of genomic RNA, and infectivity was measured using the Incucyte ZOOM as described [62] (Fig 3B). As expected JFH-1 WT showed a broad peak of infectivity at a low density (1.064 g/ml) that coincided with a genomic RNA peak, a second larger RNA peak at a higher density (1.1005 g/ml) was less infectious, consistent with previous reports [44]. P35A also showed two coincident peaks of infectivity and RNA, although the majority of the viral RNA was associated with the higher density fraction which exhibited less infectivity. In contrast, no genomic RNA or infectivity could be detected for either V67A or P145A, these two mutants were indistinguishable from the two negative controls (GND and ΔE1/E2). Gradient fractions were concentrated by methanol precipitation prior to analysis for the presence of Core by western blot. This analysis (Fig 3C) revealed a complete lack of any Core protein in fractions from either V67A or P145A, again in common with the negative controls. In contrast both WT and P35A exhibited Core protein correlating with the peaks of infectivity and virus RNA. We conclude that both V67A and P145A mutations block the assembly of infectious virus particles at an early stage. Of note, unlike the replicase function of domain I [63], the assembly function was unable to be trans-complemented by wildtype NS5A: following co-electroporation of V67A or P145A mutant JFH-1 RNA with a wildtype SGR no infectious virus was produced (S5B Fig). This is consistent with a recent study revealing that the assembly function of NS5A domain III was refractory to trans-complementation [64].
To shed light on the phenotype of the V67A and P145A mutations, we applied an imaging approach, using high resolution confocal microscopy (Airyscan) to assess the distribution of both viral and cellular factors during infection [65,66]. In this regard, lipid droplets (LD) are important organelles for the assembly of infectious HCV particles, although their precise role remains to be elucidated [44]. Both Core and NS5A have been shown to localise with LDs and infection with HCV results in dramatic changes to the distribution and size of LDs. This is demonstrated in Fig 4: Huh7.5 cells were electroporated with JFH-1 WT RNA and analysed by Airyscan confocal microscopy for the distribution of LD, Core and NS5A at various time-points up to 72 h.p.e. (Fig 4A). The number (Fig 4B), and total area of LDs (Fig 4C), together with their distance from the nuclear membrane (Fig 4D), were determined. During the first 12 h the number of LDs declined slightly, but then increased at 24 h, followed by a further dramatic decline by 48/72 h. Importantly however, the total area of LDs within the cytoplasm (a measure of the amount of lipids stored in LDs) increased significantly at 48/72 h, indicative of an increase in the size of LDs. There were more subtle changes to the distribution of LDs: at early times (12/24 h)—they scattered throughout the cytoplasm, whereas later the distribution was more restricted to the perinuclear area (48 h) and exhibited a clustering (72 h). As previously documented, both Core and NS5A were associated with LDs at later time points. Core can be seen to completely coat the surface of LDs whereas NS5A is restricted to punctate areas on the surface. We observed the same pattern of changes in cells infected with JFH-1 (S6A Fig).
We then examined the distribution of LDs, Core and NS5A at 72 h.p.e. in Huh7.5 cells electroporated with RNA for the three domain I mutants, P35A, V67A and P145A (Fig 5). Airyscan imaging of these cells revealed some striking differences: P35A was largely indistinguishable from WT but V67A and P145A exhibited distinct phenotypes. The most notable difference was that for V67A and P145A the size of the LDs was dramatically reduced compared to WT and P35A. Quantification confirmed this visual conclusion (Fig 6A), in WT and P35A infected cells the majority of LDs had an area of between 0.2–0.6 μm2, whereas for V67A and P145A infected cells, and uninfected controls, the majority were below 0.2 μm2 (Fig 6B). In addition, there were some other differences between WT/P35A and V67A/P145A: in particular the amount of NS5A localised at the surface of lipid droplets appears to be much less for the latter two mutants. This was confirmed by quantitative analysis (Fig 7A), the percentage of NS5A fluorescence that co-localised with LD was significantly reduced. However the reciprocal analysis (percentage of LD that co-localised with NS5A) showed no differences. This suggested that the proportion of LDs that were associated with NS5A was no different to WT. However, compared to WT, the majority of NS5A did not associate with LDs. Quantitative analysis of the NS5A:Core co-localisation revealed a similar trend whereby the percentage of NS5A co-localised with Core was significantly less for V67A and P145A (Fig 7B). In contrast, although the percentage of Core that co-localised with LD was significantly reduced for V67A and P145A, the reduction was much less dramatic (Fig 7C). Lastly, we observed that there were differences in the distribution of LDs: for both V67A and P145A the LDs were significantly closer to the nucleus, albeit not as close as in either GND-electroporated or mock control cells (Fig 7D).
As the colocalisation of NS5A with Core and LDs was reduced for V67A and P145A, we also investigated the colocalisation with another replicase component, NS3. This analysis revealed a high level of colocalisation of NS5A and NS3 (Fig 8), In this analysis we also included a mutant within domain III of NS5A (S452A/454A), previously shown by us to exhibit a 100-fold reduction in production of infectious virus [25]. Interestingly, this showed a distinct phenotype with large puncta positive for both NS5A and NS3, and LDs comparable to WT/P35A. Quantification (S6B Fig) revealed that in fact V67A and P145A exhibited a modest but significant reduction in NS5A:NS3 colocalisation, suggesting that these mutations disrupt the interactions between NS5A and both the assembly machinery (Core and LDs), but also to a lesser extent the replicase components.
We complemented this imaging analysis by investigating the biochemical composition of LDs. LDs were purified from electroporated cells by density gradient centrifugation and analysed by western blot for NS5A and Core, using antibody to the LD-associated adipose differentiation-related protein (ADRP, also known as adipophilin or perilipin 2) [44,67] as a marker for LDs. The integrity of the LDs and lack of contamination with other cellular components was demonstrated by the absence of GADPH [68]. As shown in Fig 9, ADRP was exclusively present in the LD fraction (not in the cytosolic or membrane fractions). Both NS5A and Core were also detected in the LD fractions, however the relative distribution and amounts of these two viral proteins differed between the mutants and WT. Both V67A and P145A showed significantly less NS5A in the LD fraction (Fig 9B), consistent with the fluorescence data (Figs 5 and 7A). In contrast the amount of Core in the LD fraction of V67A and P145A was increased (Fig 9C). We also used qRT-PCR to quantify the amount of viral RNA in the LD fractions. This analysis revealed that for both V67A and P145A there was a significant reduction in genomic RNA associated with LDs (Fig 9D), consistent with a scenario whereby NS5A transports nascent genomes to LDs where it is transferred to the Core protein for subsequent movement to assembly sites.
Implicit in the above scenario is the specific interaction of NS5A with genomic RNA. In this context, domain I has been shown by us, and others [31,32,69], to bind specifically to the HCV 3’UTR RNA. We therefore asked whether the three mutations affected this binding capacity. To address this, we expressed domain I WT and the three mutants as His-SUMO fusion proteins in E.coli. The fusion proteins were purified and cleaved to release the untagged domain I (S7 Fig). The RNA binding capacity of the WT and mutant domain I proteins was determined by RNA filter binding assay utilizing 32P-labelled HCV 3’UTR RNA (Fig 10A). Surprisingly, we found that V67A and P145A showed strong binding affinity to HCV 3’UTR RNA, exhibiting a 10–20 fold increase compared to WT or P35A. For WT and P35A the Kd values were 246.3 ± 77.19 nM and 245.7 ± 70.09 nM respectively. However for V67A and P145A, the values were 12.89 ± 6.25 nM and 22.35 ± 9.58 nM respectively.
To validate this in vitro data, we immunoprecipitated NS5A from Huh7.5 cells electroporated with either JFH-1 WT or the three mutants and assessed the amount of viral RNA in the immunoprecipitates by qRT-PCR. Consistent with the in vitro RNA filter binding assay data, both V67A and P145A bound more viral RNA compared to WT and P35A (Fig 10B). In contrast, a similar analysis of Core immunoprecipitates revealed significant reductions in the amount of genomic RNA bound to Core for V67A and P145A (Fig 10C). Taken together, these data suggest that NS5A binds specifically to the nascent genomic RNA but that during the assembly process this must be released to Core. By increasing the affinity of NS5A for the 3’UTR RNA, these mutations are preventing this transfer.
NS5A has also been reported to dimerize, both in the published crystal structures [15–17] and in biochemical analyses [70]. Examination of the different dimer structures revealed that P35 was located in the dimer interface of the ‘open’ conformation [15,71,72]. P145 was located in the interface of the ‘closed’ conformation [15–17,72]. In contrast V67 was distal to the dimer interfaces in both conformations (S8 Fig). To test the effects of the three mutations on dimerization, we conducted GST pulldown assays using GST-tagged domain I as bait to precipitate His-tagged domain I (input levels of proteins shown in Fig 11A). We observed that GST-domain I (WT), but not GST alone, precipitated His-domain I (WT) (Fig 11B). GST-domain I (P35A) was also able to precipitate His-domain I (P35A) with a modest but non-significant reduction in binding. In contrast, both V67A and P145A mutant GST-domain I proteins failed to precipitate the cognate His-domain I proteins (Fig 11B), indicating that these two residues are required for dimerization of domain I and implicating a role for NS5A dimerization in virus assembly.
This study identified three residues in NS5A domain I for which alanine substitution had a modest effect on genome replication, but significant defects in the assembly of infectious virus particles. These residues were chosen for their conservation–P35 and P145 are 100% conserved throughout all hepaciviruses, V67 is conserved in all HCV genotypes apart from genotype 4 where it is generally an isoleucine. Structural analyses of domain I also predicted that they are all surface exposed. In particular we focussed our attention on two of these, V67A and P145A, which completely abrogated virus assembly. Previously, domain I has been assumed only to function during genome replication, and to our knowledge this is the first detailed analysis of a role for domain I in virus assembly.
Both V67A and P145A mutants failed to produce intracellular infectious virus and consequently failed to release any virus particles, as judged by the lack of virus RNA or Core protein in cell culture supernatants. This was not due to a lack of genome replication or Core protein within the cells, as levels of both were similar to WT (S4 Fig and Fig 2). In cells infected with V67A or P145A mutant viruses there were defects in LD production. Compared to WT, LD were smaller, closer to the nucleus and NS5A recruitment to LDs was impaired. Lastly, these two mutants enhanced binding of domain I to the HCV 3’UTR RNA and inhibited dimerization.
What are the implications of these data? Firstly, they imply that domain I of NS5A plays multiple roles in virus assembly. It is required both for the association of NS5A with LD as well as the increase in LD size and altered distribution (movement away from the nuclear membrane) that is seen during HCV infection. Taken together with the in vitro data, these support a model in which domain I of NS5A binds to the 3’UTR of nascent genomes and transports them from sites of replication to LD. Here, analogous to the handing on of a baton in a relay race, the RNA is transferred to Core and then subsequently transported to assembly sites. The latter remain to be unambiguously defined but may be endosomal membrane compartments [73,74]. The enhanced binding of V67A or P145A to the 3’UTR RNA may prevent the release of RNA for transfer to Core. The LD distribution in cells infected with V67A or P145A at 72 h.p.e. ressembles that in wildtype at 12/24 h.p.e., suggesting that these mutations might block the transition from genome replication to virus assembly. Furthermore, the loss of dimerization by these two mutants implies that, in contrast to the accepted model of an open NS5A dimer revealing a basic RNA-binding groove, monomeric NS5A is able to bind RNA. However, we cannot rule out the possibility that in the intact protein, domains II and III influence both dimerization and RNA binding by domain I. In this regard we note that our attempts to detect NS5A dimerization within intact cells have so far been unsuccessful, despite testing a variety of experimental protocols (see S9A Fig). Despite this, it is tempting to speculate that monomeric NS5A might transport nascent RNA to LDs, then dimerizes and releases the RNA to Core.
Our data are consistent with previous studies into the role of NS5A during virus assembly which support a model whereby NS5A orchestrates the processes of genome replication and virus assembly. However, these studies have exclusively focused on the role of domain III [49], and it has been widely accepted that the determinants of virus assembly within NS5A lie entirely within domain III. For example, a serine near the C-terminus of domain III is implicated in the interaction between NS5A and Core, and it has been proposed that phosphorylation of this residue by casein kinase II is required for virus assembly [75]. More recently, mutations of a basic cluster at the N-terminus of domain III resulted in modest impairment of Core-RNA and NS5A-RNA interactions and virus particle envelopment, leading to a 100-fold reduction in released virus titres [76]. Our data extend these observations, providing evidence that domain I also makes a major contribution to virus assembly.
Other implications of our study concern the modifications to LD morphology that occur during HCV infection. As illustrated in Fig 4, at late stages (48 h onwards), increases in LD size and total volume most likely reflect the coalescence of smaller LDs into larger structures. Our data indicate that domain I of NS5A plays a role in this process, as V67A and P145A do not exhibit this increase (Figs 5 and 6). NS5A is recruited to LDs, in most cases to discrete punctate locations on the surface, in contrast to the complete coating of LDs with Core.
One apparent discrepancy in our data relates to the co-localisation of Core with LDs. Specifically, the imaging data (Figs 5 and 7) showed a modest reduction in Core:LD co-localisation for V67A and P145A, whereas these mutants showed higher levels of Core co-purified with LDs (Fig 9). Two factors may help to explain this discrepancy: firstly, it is possible that in the case of V67A and P145A, Core associates more strongly with LDs, possibly because it has not been displaced by NS5A. Secondly, V67A and P145A infected cells exhibit larger numbers of smaller LDs, thus the available LD surface area for interaction with Core is also likely to be larger, allowing more Core to associate. In addition, it is important to note that the data in Fig 7C refer to the percentage of total Core associated with LD, and do not take into account the absolute amounts of Core.
Whether the increase in LD size is a direct consequence of recruitment of NS5A, or indirectly driven by NS5A-mediated effects on lipid metabolism, remains unclear. In this context, NS5A has previously been shown to interact with a number of LD-associated proteins, including DGAT-1 [77] and Rab18 [78]. However, the phenotype of V67A or P145A cannot be explained by a lack of binding to these proteins–as shown in S9B Fig, both DGAT-1 and (to a lesser extent) Rab18 precipitated with both WT and the three mutant NS5As. We are currently extending this analysis, using a proteomic approach to determine the interactome of the three mutant NS5As in comparison to WT.
In contrast to V67A and P145A, P35A exhibited a moderate virus assembly phenotype with only a small (less than 10-fold) reduction in virus titre. Nevertheless some important observations can be made: firstly, in the density gradient analysis (Fig 3) the peak of infectivity for P35A resolved at a lower buoyant density than WT (1.0475 g/ml compared with 1.064 g/ml). In contrast the second peak of infectivity with higher buoyant density for P35A was associated with more genome RNA and Core than WT. These data imply subtle differences in the association of virus particles with VLDL or other lipids. In all other analyses (LD size and distribution, NS5A recruitment to LD, dimerization and 3’UTR binding), P35A was not statistically significantly different from WT.
Lastly, it is important to consider our results in the context of the class of potent DAAs that are defined as NS5A inhibitors, exemplified by daclatasvir (DCV). Although initially developed as inhibitors of genome replication [79], it has become clear that DCV also has an independent effect on virus assembly. Treatment of infected cells with DCV resulted in a rapid (2 h) block to virus assembly, preceding the inhibition of genome replication which was only apparent at later time points (24 h) [80]. More recently, it has been shown that DCV treatment prevented the transfer of genomic RNA to assembly sites [81]. DCV has been reported to target domain I, as judged by the location of DCV-resistance mutations (eg L31M and Y93H). It is important to note that none of the 3 mutations analysed in this study exhibited any effect on the activity of DCV measured against HCV genome replication (S10 Fig). However, our observation that domain I is directly implicated in virus assembly does provide a rationale for the rapid effect of DCV on this process, and may therefore help to explain the extraordinary potency of DCV and related compounds.
DNA constructs of luciferase reporter sub-genomic replicon (mSGR-luc-JFH-1), infectious mJFH-1 virus and sub-genomic replicon with NS5A containing the One-Strep-tag (OST) (pSGR-Neo-JFH1-5A-OST) were maintained in our laboratory [82]. pcDNA3.1(+) was used as the vector to subclone the BamHI-HindIII JFH-1 NS5A fragment for site-directed mutagenesis. NS5A fragments with mutations were then cloned into either mSGR-luc-JFH-1 or mJFH-1 via flanking BamHI/AfeI restriction sites. The pCMV10-NS3-5B plasmid was constructed [61], and the NS5A domain I fragments with mutations were then inserted into this wild type vector by cloning the NsiI–RsrII fragment containing the mutations from the corresponding mJFH-1 constructs. NS5A-OST with mutations from pSGR-Neo-JFH1-5A-OST were cloned back into mJFH1 viruses via NsiI and BsrGI restriction sites to generate mJFH1-5A-OST constructs. Primer sequences available upon request.
The following antibodies were used: sheep anti-NS5A (in house polyclonal antiserum) [83], mouse anti-NS5A (9E10) (kind gift from Tim Tellinghuisen, Scripps Florida), mouse anti-NS3 (kind gift from Thomas Pietschmann, TWINCORE, Hannover), rabbit anti-Core (polyclonal serum R4210) and sheep anti-ADRP (kind gifts from John McLauchlan, Centre for Virus Research, Glasgow), sheep anti-GST (in-house), mouse anti-DGAT1 (Santa Cruz), mouse anti-Rab18, anti-Actin and anti-His (Sigma Aldrich).
Huh7 and Huh7.5 cells that are highly permissive for HCV RNA replication were used for electroporation [60]. Cells were washed twice in cold phosphate-buffered saline (PBS) before electroporating 4x106 cells in cold PBS with 2 μg of RNA at 975 μF and 260 V. Cells were resuspended in complete media before being seeded into either 96-well plates (n = 6) at 3x104 cells/well, or 6-well plates (n = 2) at 3x105 cells/well, both plates incubated under cell culture conditions. 4, 24, 48 and 72 h post-electroporation (h.p.e.), cells were harvested by lysis with 30 μl or 200 μl passive lysis buffer (PLB; Promega) from 96- and 6-well respectively. Luciferase activity was determined from 96-well samples on a BMG plate reader by automated addition of 50 μl luciferase assay reagent (Promega) and total light emission was monitored.
Cells were washed twice with PBS, lysed by resuspension in Glasgow lysis buffer (GLB) [1% Triton X-100, 120 mM KCl, 30 mM NaCl, 5 mM MgCl2, 10% glycerol (v/v), and 10 mM piperazine-N,N’-bis (2-ethanesulfonic acid) (PIPES)-NaOH, pH 7.2] supplemented with protease inhibitors and phosphatase inhibitors (Roche Diagnostics), and incubated on ice for 15 min. Following separation by SDS-PAGE, proteins were transferred to a polyvinylidene fluoride (PVDF) membrane and blocked in 50% (v/v) Odyssey blocking buffer (LiCor) in Tris-buffered saline (TBS) [50 mM Tris, 150 mM NaCl, pH 7.4]. The membrane was incubated with primary antibody in 25% (v/v) Odyssey blocking buffer overnight at 4°C, then incubated with fluorescently labelled anti-sheep (800nm), anti-rabbit (800nm) or anti-mouse (700 nm) secondary antibodies for 2 h at room temperature (RT) before imaging on a LiCor Odyssey Sa fluorescent imager.
Huh7.5 cells were washed twice in cold PBS before electroporating 2x107 cells in cold PBS with 10μg viral RNA at 975 μF and 260 V. Cells were resuspended in complete medium and seeded into 6-well plates and T175 flasks for virus replication and virus titration analysis.
48 h.p.e., cells were washed in PBS and fixed in 4% paraformaldehyde (PFA) for 20 min and staining with NS5A-specific sheep polyclonal antiserum as primary antibody (dilution 1:2000) and Alexa Fluor-594 conjugated donkey anti-sheep (Invitrogen) as a secondary antibody (dilution 1:750) for IncuCyte counting (see details in Use of the Incucyte ZOOM).
Culture supernatants in T175 flasks were harvested at 72 h.p.e., and extracellular virus titres were determined. Intracellular infectivity was determined for freeze–thaw lysates of electroporated cells 72 h.p.e. using the protocol reported previously [84]. Naïve Huh-7.5 cells were seeded into 96 well plates (8.0x103 cells/well, 100 μL total volume) and allowed to adhere for 6 h. Clarified virus was serially diluted two-fold into the existing media (final volume 100 μL per well). Cells were incubated for 48h post infection (hpi) before the detection of viral antigens by indirect immunofluorescence. Virus-positive cells were counted using IncuCyte and the titre (IU/mL) was calculated from the wells of multiple virus dilutions [31].
Following immunofluorescence staining for viral antigens, with an Alexa Fluor 594-conjugated (“red”) secondary antibody, fixed microtitre plates were imaged with the IncuCyte ZOOM (Essen BioScience) [62] to determine the total number of virus-positive cells/well. Viral titres were obtained by multiplying the number of virus-positive cells/well by the reciprocal of the corresponding dilution factor, corrected for input volume. As this method measures the absolute number of infected cells, rather than the number of foci of infected cells, the titre is represented as infectious units per mL (IU/mL).
Culture medium from JFH-1 infected cells was concentrated 100-fold using 10% PEG 8000 (w/v) (Fisher Scientific) and centrifugation at 3000 g for 30 min. The pellet was resuspended in 1ml of PBS and overlaid over a 1 ml cushion (20% sucrose, w/v, in PBS), followed by ultracentrifugation at 150,000 g for 3 h at 4°C in an S55S rotor. The resulting pellet was resuspended in 200 μl PBS and then loaded on a 10–40% gradient iodixanol in 2.2 mL tubes followed by centrifugation at 150,000 g for 4 h at 4°C. The gradient was fractionated into 12 fractions of 180 μl each. Each fraction was used for virus titration as well as RNA extraction for qRT-RCR analysis, the remainder of each fraction was mixed with ice-cold methanol (1:3) and proteins precipitated at -80°C overnight. Precipitated proteins were recovered by centrifugation at 13,000 rpm for 30 min at 4°C, and pellets were resuspended in 25 μl SDS-PAGE loading buffer, prior to western blot analysis.
To quantify the number of HCV genomes, RNA from each fraction after gradient centrifugation of extracellular virus was extracted using TRIzol following the manufacturer’s instructions (Invitrogen). Extracted cellular RNA was analysed by qRT-PCR using a one-step qRT-PCR Taqman-based kit as directed by the manufacturer (Eurogentec). Amplifications were conducted in triplicate using the following primers and 6FAM- and TAMRA- labelled probes designed to detect the HCV JFH-1 5’UTR: 5’UTR Taqman probe 83–108: 5’- 6FAM-CATGGCGTTAGTATGAGTGTCGTACA-TAMRA-3’; 5’UTR Forward-57: 5’-CTGTCTTCACGCAGAAAGCG-3’; 5’UTR Reverse-312: 5’-CACTCGCAAGCGCCCTATCA-3’.
Virus RNA electroporated cells were seeded onto 19 mm glass coverslips in 12 well plates, 72 h.p.e. cells were fixed in 4% PFA and permeabilised with 0.1% (v/v) Triton X-100 (Sigma-Aldrich) in PBS for 7 min. Coverslips were washed twice in PBS and the primary antibody applied at the relevant dilution in 10% (v/v) FBS in PBS and incubated for 2 h at RT. To remove any unbound primary antibody, cells were washed three times in PBS before the application of the relevant Alexa Fluor-488, 594 or 647 conjugated secondary antibodies diluted 1:750 in 10% (v/v) FBS in PBS followed by 2 h incubation at RT in the dark. Lipid droplets were stained using BODIPY (558/568)-C12 dye at 1:1000 (Life Technology). The coverslips were washed three times in PBS before the nucleus was stained by the addition of 4’,6’-diamidino-2-phenylindole dihydrochloride (DAPI) diluted 1:10 000 in PBS for 30 min at RT in the dark. Coverslips were washed three times in PBS and mounted on a glass microscope slide in ProLong Gold antifade regents (Invitrogen, Molecular Probes) and sealed with nail varnish. Slides were stored at 4°C in the dark until required and examined. Confocal microscopy images were acquired on a Zeiss LSM880 microscope with Airyscan, post-acquisition analysis was conducted using Zen software (Zen version 2015 black edition 2.3, Zeiss) or Fiji (v1.49) software [85].
For co-localisation analysis, Manders' overlap coefficient was calculated using Fuji ImageJ software with Just Another Co-localisation Plugin (JACoP) (National Institutes of Health) [73]. Coefficient M1 reports the fraction of the LD signal that overlaps either the anti-NS5A or anti-Core signal or the fraction of anti-Core signal that overlaps the anti-NS5A signal. Coefficient M2 reports the fraction of either the anti-NS5A or anti-Core signal that overlaps the LD signal or the fraction of anti-NS5A that overlaps the anti-Core signal. Coefficient values range from 0 to 1, corresponding to non-overlapping images and 100% co-localization images, respectively. Co-localisation calculations were performed on >10 cells from at least two independent experiments.
For the quantification of LD spatial arrangement, images were acquired with the same acquisition parameters, but with variable gain to ensure correct exposure. The two-dimensional coordinates of the centroids of LDs were calculated using the Analyze Particles module of Fiji (ImageJ). The distance of each particle to the edge of the nucleus, visualised using DAPI stain, was looked up using a Euclidean distance map computed with the Distance Transform module of Fiji and exported as a list of distance measurements via the Analyze Particle function. Box and whisker plots of these distance measurements were constructed using GraphPad Prism and compared between samples using a one-way ANOVA and Bonferroni-corrected post-hoc t-tests. Two-dimensional areas of the LDs were also measured using the Analyze Particles function in Fiji. Lists of the area measurements were used for constructing frequency histograms using a custom-written programme implemented in IDL. The shapes of these histograms were compared using a chi-squared test, implemented in IDL.
Four 10 cm dishes of Huh7.5 cells electroporated with mJFH-1 virus RNA (80% confluent) were scraped into 10 mL of PBS at 72 h.p.e.. The cells were pelleted by centrifugation at 1,500 rpm for 5 min and then resuspended with 500 μL buffer A (20mM Tricine, 250mM sucrose, pH 7.8) supplemented with protease and phosphatase inhibitors and kept on ice for 20 min. The suspension was homogenized with a plastic tissue grinder homogenizer. Samples after homogenization were centrifuged at 3000g for 10 min at 4°C to remove nuclei and the post nuclear supernatant (PNS) was collected, transferred into 2.2 mL tubes and overlaid with 1 mL of buffer B (20 mM HEPES, 100 mM KCl and 2 mM MgCl2 pH 7.4) plus protease inhibitors. Tubes were centrifuged in a S55S rotor at 100,000g for 1h at 4°C. After centrifugation, the LD fraction on the top of the gradient was recovered in buffer B and washed twice by centrifugation at 20,000g for 5 min at 4°C to separate the LDs from the buffer. Underlying solution was removed and discarded. Proteins and lipids in LD samples were separated with 2 volumes of ice-cold acetone and chloroform (1:1) to precipitate proteins. RNA in lipid droplet fractions were extracted using TRIzol for qRT-PCR. The collected LD fraction was dissolved in 50μL of SDS sample loading buffer for western blot.
Construction and purification of domain I with corresponding mutations have been listed in S1 Text. After purification, GST-domain I (GST-DI) and His-SUMO-domain I (His-SUMO-DI) were dialyzed against dialysis buffer (50 mM Tris-HCl, pH 7.5, 100 mM NaCl, 5 mM MgCl2, 10% glycerol, 0.5% NP-40). A GST pulldown assay was performed as described previously [70]. Briefly, 10 μg of GST or GST-fusion proteins were mixed with 5 μg of His-SUMO-DI in binding buffer (20mM Tris-HCl, pH 7.2, 0.5 M NaCl, 200KCl, and 1% NP-40) for 3 h at 4°C on a rotating platform. Then the mixture was added to glutathione beads and incubated overnight at 4°C. After washes using binding buffer, bound material was eluted with 50 μL of SDS sample buffer and heated for 10 min at 95°C. After centrifugation, these samples were analysed by Western blot using anti-GST and anti-His antibodies.
His-SUMO-DI proteins were cleaved with SUMO protease to produce native domain I. Following purification as in S1 Text, domain I was incubated with in vitro transcribed [α-32P] radiolabelled RNAs as described previously [32]. Then aliquots of each binding reaction were applied to a pre-assembled slot blot apparatus and filtered through firstly a nitrocellulose membrane (Schleicher & Schuell) to capture soluble protein-RNA complexes, and secondly a Hybond-N nylon membrane (Amersham Biosciences) to bind free RNA. After washing and air drying of both membranes, quantification of radioactivity was performed by phosphoimaging using an FLA 5000 Imaging system (Fuji), and ImageJ software. These data were fitted to the hyperbolic equation R=Rmax×P/(Kd+P). R is the percentage of bound RNA, Rmax is the maximal percentage of RNA competent for binding, P is the concentration of Domain I, and Kd is the dissociation constant [32].
Co-immunoprecipitation experiments were performed in Huh7.5 cells 72 h.p.e. with mJFH-1 virus RNA using polyclonal anti-Core or monoclonal anti-NS5A antibodies and Dynabeads™ Protein G (Thermo Fisher Scientific), following the manufacturers protocol. Immunoprecipitated proteins were subjected to immunoblotting and co-immunoprecipitated RNA was extracted by TRIzol reagent and then quantified by qRT-PCR.
Statistical analysis was performed using unpaired two-tailed Student’s t tests, unequal variance to determine statistically significant differences from the results for the wild type (n≥3). Data in histograms are displayed as the means ± S.E.
|
10.1371/journal.ppat.1005692 | Multi-epitope Models Explain How Pre-existing Antibodies Affect the Generation of Broadly Protective Responses to Influenza | The development of next-generation influenza vaccines that elicit strain-transcendent immunity against both seasonal and pandemic viruses is a key public health goal. Targeting the evolutionarily conserved epitopes on the stem of influenza’s major surface molecule, hemagglutinin, is an appealing prospect, and novel vaccine formulations show promising results in animal model systems. However, studies in humans indicate that natural infection and vaccination result in limited boosting of antibodies to the stem of HA, and the level of stem-specific antibody elicited is insufficient to provide broad strain-transcendent immunity. Here, we use mathematical models of the humoral immune response to explore how pre-existing immunity affects the ability of vaccines to boost antibodies to the head and stem of HA in humans, and, in particular, how it leads to the apparent lack of boosting of broadly cross-reactive antibodies to the stem epitopes. We consider hypotheses where binding of antibody to an epitope: (i) results in more rapid clearance of the antigen; (ii) leads to the formation of antigen-antibody complexes which inhibit B cell activation through Fcγ receptor-mediated mechanism; and (iii) masks the epitope and prevents the stimulation and proliferation of specific B cells. We find that only epitope masking but not the former two mechanisms to be key in recapitulating patterns in data. We discuss the ramifications of our findings for the development of vaccines against both seasonal and pandemic influenza.
| The current influenza vaccine requires frequent updating in order to protect against small changes in the virus from one year to the next as well as larger changes associated with the emergence of new influenza strains from zoonotic reservoirs that cause pandemics. There is a considerable interest in developing “universal” vaccines that will boost immune responses to the conserved regions of the virus, in particular, to the stem region of the major virus surface molecule hemagglutinin (HA). However, recent data reveals that vaccination results in very limited boosting of antibodies to the stem of HA. We use mathematical models to explore different hypotheses that may explain why vaccination does not boost antibodies to the conserved parts of the virus. By confronting our models with the data from the human vaccination trials we found that the key mechanism preventing effective boosting of the responses to the stem of HA is masking of the stem by pre-existing antibodies developed during previous infections and vaccinations. We discuss how this masking effect could be overcome in a “universal” influenza vaccine.
| Both seasonal and pandemic influenza pose significant public health concerns. Seasonal influenza in the U.S. is estimated to lead to an economic burden of $87.1 billion [1], and pandemic influenza poses a grave threat to public health, as witnessed during the 1918–1919 Spanish influenza outbreak [2].
We are currently able to generate vaccines against seasonal influenza based on knowledge of its global patterns of spread and mechanisms of evolution, such as antigenic drift, that lead to gradual annual changes in the surface proteins of the virus. However, given current immunization technologies, a new vaccine must be formulated each year; this endeavor is costly, and estimates of vaccine effectiveness vary widely and differ depending on whether they are focused on symptoms, infection or transmission [3–7]. Moreover, current vaccine technologies are not protective against pandemic influenza strains, to which people have little or no pre-existing humoral immunity. Pandemic influenza generally occurs due to larger antigenic changes (shifts), and when these novel strains enter the human population they typically cause severe disease [2, 8].
The humoral immune response is most strongly stimulated by hemagglutinin (HA), the major surface molecule, which has a distinct head and stem structure [9–12]. Current influenza vaccines target the head of hemagglutinin, which has multiple epitopes that vary from year to year. In contrast, the stem is highly conserved and remains largely the same over time. In fact, there are only a couple of different stem types, even across influenza subtypes. The stem is therefore a desirable target for immunization because a vaccine that could elicit antibodies that bind to the stem epitopes would be useful across years and even, likely, for novel pandemic strains that may arise in the future.
Recent experimental studies in mice and ferrets show that it is possible to generate high levels of antibody to the stem of HA using novel vaccines, and these antibodies can provide strain-transcendent immunity in animal model systems [13–19]. One of the important differences between these animal model systems and humans is that in contrast to naive mice and ferrets, humans typically have prior immunity generated by exposure to multiple strains that have circulated in the past. Consequently, to move these vaccines to humans, it would be helpful to develop a quantitative understanding of how pre-existing immunity affects the ability of vaccines to boost antibody responses in general and the response to the stem of HA in particular.
In this paper we develop mathematical models to help us explore different hypotheses for how pre-existing antibody affects boosting following immunization. We define the magnitude of the boost as the fold increase in antibody following immunization. The models are confronted with reanalysis of recent data shown in Fig 1 measuring how immunization with a vaccine containing HA from novel strains of influenza boosts antibodies specific to the stem and head of HA [20]. Fig 1 shows that: prior to immunization there were on average higher levels of antibodies to the stem than the head of HA (and no individuals with very low titers to the stem); and immunization caused antibodies against stem epitopes to be boosted less than those against head epitopes (Fig 1A and 1B, t − test p − value < 0.0001 and Fig A Panels A,B in S1 Text). In fact, if we look across both epitopes, increasing pre-exposure titers led to lower boosting of responses to both epitopes (Fig 1C). Similar data was obtained following immunization with inactivated H1N1 and is shown in Fig A in S1 Text. In the H1N1 study there was more overlap between the prevaccination antibody titers to the head and stem of HA. The heterogeneity in immune responses of individuals in the population makes it hard to discern the rules which govern how pre-existing immunity affects the boosting of antibody responses following immunization from these data.
We use the models together with statistical analysis to explore three hypotheses that may explain why pre-existing antibodies reduce the boosting of the humoral immune response following immunization; in doing so, we parsimoniously account for the differences in the immune responses to head and stem epitopes. The antigen clearance model (ACM) proposes that pre-existing antibodies, which bind to epitopes on antigens, cause rapid clearance of the antigen. The reduced antigen load results in less expansion of B cells and less boosting of antibody. The Fc receptor-mediated inhibition model (FIM) proposes that pre-existing antibodies bind antigen and these antigen-antibody complexes inhibit the activation of specific B cells. The proposed mechanism involves antibody forming a complex with antigen and antigen-specific B cells recognizing these immune complexes presented via complement and Fc-receptors on follicular dendritic cells in the germinal center. The co-crosslinking of the B cell receptor (bound to the antigenic epitope) and the FcγRIIB (bound to the Fc portion of the antibody in the immune complex) leads to inhibition of B cell activation [21]. The epitope masking model (EMM) proposes, instead, that pre-existing antibodies that bind to epitopes on the current strain of HA mask these epitopes, thus inhibiting the binding and proliferation of B cells specific for the same and nearby epitopes but not B cells specific for distant epitopes. This inhibition is because the stimulation of epitope-specific B cells requires their binding to the epitope and physical constraints associated with the size of antibodies prevent B cell from binding to epitopes with attached antibody. Further expansion of specific B cells and production of antibody to epitopes that are masked by pre-existing antibody is thereby downregulated despite the continued presence of antigen.
We compare the models’ predictions regarding antibody titers against hemagglutinin’s stem and head with the aforementioned data. We find that all three models can account for features of the data shown in Fig 1. We then show that reanalysis of the data in a manner that takes into account linkage between the responses to head and stem epitopes within individuals allows us to discriminate between the models. We find that only the EMM is able to recapitulate key patterns regarding the relative boosting of responses to the head and stem epitopes within individuals. We show that this conclusion is robust to a variety of alternative model assumptions, including model expansion to multiple epitopes on the head of HA. Finally, we discuss the implications of our findings for the development of a strain-transcendent influenza vaccine.
The modeled interactions between antigen and the immune system are depicted together with corresponding equations in Fig 2A and explained in detail in the Methods section. The objective is to use models to explore how prevaccination antibodies affect the boosting of antibody responses following immunization. The key features of the data are the measurements of antibodies to the head and stem of HA [20]. Accordingly, we use a minimal model that focuses on epitopes on the head and stem of HA and their interactions with B cells and antibodies specific for these epitopes. At this stage we have a simple B cell clonal expansion model for the generation of immune responses. Complex interactions such as differentiation of B cells and interactions with other cells such as follicular dendritic cells and T cells in germinal centers that underlie the process of clonal expansion are not explicitly included. This is because the experimental data does not include measurements of these quantities. Consequently, we use simpler models and focus on generating robust qualitative predictions. We test these predictions by confronting them with existing experimental data. In these circumstances the results of simpler models are typically more robust than those of complex models [22].
In the basic model, antigen stimulates clonal expansion of specific B cells, which produce antibody. Antibody production continues until all the antigen has decayed. The basic model predicts that the fold increase in antibody titers is independent of initial titers; therefore, under this model, the boost in antibody levels resulting from a second immunization is the same as that from the initial immunization. This is shown by simulations in Fig 2B.
The ACM, FIM and the EMM all predict that pre-existing antibody reduces the fold increase in antibody production following immunization, a pattern that is consistent with data ([20] and Fig 1). Thus secondary immunization with the same antigen leads to lower boosting of antibody in comparison to the primary immunization (Fig 2C and 2D). However, the underlying mechanisms of the ACM, FIM and EMM are different. In the ACM, antigen bound to antibody (Hb) is removed from the system faster than free antigen (i.e., db > df, green text in Fig 2A). Because bound antigen is removed more quickly, it has less time to stimulate B cells, resulting in less boosting (Fig 2C). In the FIM, antigen-antibody complexes (Hb) reduce the activation of B cells as described by the parameter α (see additional term in blue color in equation for B cells in Fig 2A). 1/α is the concentration of antigen-antibody complexes that reduces the growth rate of B cells by a factor of 2. If α = 0 there is no Fc-mediated inhibition. In the EMM, antigen is cleared at the same rate whether bound or free (i.e., db = df); however, bound antigen is unable to stimulate B cell proliferation because the sites to which the B cells would bind are sterically blocked (i.e., orange text and δ = 0 removes the two terms in corresponding equation for B cells in Fig 2A). This also results in less boosting upon re-exposure (Fig 2E).
The data in Fig 1C showing that increasing the amount of pre-existing antibody reduces the extent of boosting is consistent, thus far, with all three hypotheses (i.e. the ACM, FIM and EMM). We now tackle the problem of discriminating between the ACM, FIM and EMM by identifying scenarios where the models give rise to different predictions. In what follows we simulate the responses to immunization with HA using a multi-epitope framework with three epitopes, but we note that we get the same results using a two-epitope framework.
The experimental data show boosting of antibodies to the head and stem of HA following immunization with H5N1 (Fig 1) and H1N1 (see Fig A in S1 Text) vaccines. In the previous sections we have considered the consequences of changing prevaccination immunity to one epitope (the stem epitope). Fig 5 shows that the models give different predictions when we vary the prevaccination immunity to both head and stem epitopes simultaneously. In Fig 5, we consider a number of individuals with different levels of pre-existing antibody to the head and stem of HA and plot the boosting of the responses to both head and stem; points connected by a line are from the same individual. In the basic model, boosting is independent of the level of pre-existing antibody; thus, the lines are all horizontal and at the same height. In the ACM, pre-existing antibody to either the head or stem of HA clears the entire antigen (i.e., both head and stem) and reduces the boosting to both head and stem equally. In the FIM, antigen-antibody complexes inhibit the activation and proliferation of B cells specific for both epitopes. Thus, in both the ACM and FIM the lines joining each individual are horizontal; however, in comparison with the basic model, the boost is reduced by pre-existing antibody to both head and stem. In contrast, in the EMM, antibodies to the head and stem of HA can be boosted by different amounts (Fig 5C). This is because pre-existing antibody to the stem of HA masks only the stem epitope, thus reducing the boosting of responses to the stem and not to the head (and vice versa for antibodies to the head). Consequently, the lines for all individuals fall along a diagonal with a negative slope.
To discriminate between the models we compare the predictions with data from vaccine trials in humans with HAs from H5N1 or H1N1 strains (Fig 1 and Fig A in S1 Text). The assays used in the study allowed for independent measurement of stem and head antibody titers in each individual [20, 25]. The data for boosting of antibodies to the head and stem of HA such as that shown in Fig 1 were reanalyzed to generate a plot similar to that in Fig 5. The reanalyzed data is plotted in Fig 6. Clearly, the pattern observed in the data is consistent with only the EMM. The EMM predicts that low pre-existing antibody titers to epitope on the stem allow for a strong boost to that epitope regardless of whether titers to the epitopes on the head are high or low (and vice versa). In line with this prediction, the degree of boosting for a given epitope shows a significant negative correlation with the initial titers for that epitope but no significant correlation with the initial values for the other epitope (see Figs C,D and Table A in S1 Text).
We finally consider an ensemble of models that includes all possible combinations of antigen clearance, Fc-mediated inhibition and epitope masking (Fig E in S1 Text). The results are complex and indicate potentially interesting interactions between the different mechanisms of antigen clearance, Fc-mediated inhibition and epitope masking. For example, combinations that include both the FIM and EMM models (i.e. FIM+EMM and ACM+FIM+EMM) can generate slopes for individuals that are positive in comparison with the zero or negative slopes seen when we have the ACM, FIM or EMM alone. All combinations (of two or all three models) predict that the slope for each head-stem pair (slopes of individuals) will be less steep than the slope of the best fit line to the data set as a whole. This is not consistent with the observations as can be seen in Table 2 where we see that the average for the slopes of individuals is not significantly different from the slope fitted to all data points (p − values = 0.40 and 0.31 for H5N1 and H1N1, respectively). In conclusion, the data suggest that epitope masking is the key factor in reducing the boosting of antibody and that antibody-mediated antigen clearance and Fc-mediated inhibition of B cell activation play no more than a minor role.
Epitope masking was found to be the dominant factor for explaining patterns in the data; we therefore use the EMM to predict the optimal vaccination strategy for boosting stem antibody titers. We find that increasing the dose of stem antigen used in the vaccine can help counteract the effects of epitope masking and allow for the generation of stronger antibody responses to the stem. There is a threshold level of antigen above which the response is quite strong; this level is equal to the amount of antibody present before the vaccine. Once the antigen dose in the vaccine exceeds the amount of antibody present, the free antigen successfully stimulates a boost to stem antibody titers. This is illustrated in Fig F in S1 Text where we plot how both the level of pre-existing immunity and antigen-dose affect the boosting of antibodies to a given epitope.
This paper uses simple mathematical models to explore how pre-existing antibody affects the generation of recall responses following immunization. We apply these models to the context of antibody responses to HA which is the main target of humoral immunity to influenza. The first step in this paper was to show that multiple models of immunodynamics, the ACM, FIM and EMM are consistent with previous observations describing how prior antibody limits the boosting of antibodies to HA upon re-exposure. In particular, we extend our previous study that described the epitope masking hypothesis [26] by showing that the clearance of antigen by antibodies (i.e., the ACM) and Fc-mediated inhibition of B cell activation (i.e., the FIM) could also explain these observations. This is because, in all three models, the presence of pre-existing antibody leads to lowering the expansion of specific B cells and less boosting.
In order to discriminate between the ACM, FIM and EMM we determine situations in which they give rise to different predictions. We find that reanalysis of the data in a manner that takes into account information previously ignored (i.e., the relative boosts to head and stem antibodies within an individual) allows us to discriminate between the two models. This discrimination is possible because the three models provide different predictions in the presence of multiple epitopes on the same antigen to which the host has differing levels of pre-existing immunity. Under the ACM, the whole antigen is affected simultaneously. Under the FIM, the antigen-antibody complexes reduce the stimulation of all antigen-specific B cells. This results in the boosting of antibodies to all the epitopes being similar in both the ACM and FIM. In the EMM, epitopes can be affected independently, and the boosting of antibodies to one epitope is not affected by antibodies to the other epitope. Comparison of these predictions with the data allows us to reject the ACM and FIM models—only the EMM is consistent with the data, and suggests that the masking of antigen epitopes by antibodies plays a key role in modulating recall responses to influenza.
We have considered how steric-interference and epitope-masking affect the generation and boosting of antibody responses. Steric interference in the binding of antibodies to epitopes on the head of HA may also have important consequences for the neutralization of live virus and this has been described in [23, 27]. The antibodies binding to non-neutralizing epitopes on the head of HA may sterically prevent neutralizing antibodies binding to nearby epitopes, and it has been suggested that HA in vaccines should be engineered to remove the non-neutralizing epitopes [27]. Our study differs from these studies in focusing on the boosting of responses rather than the neutralization of virus. Interestingly, our analysis suggests that there is little if any steric interference between the binding of antibodies to the epitopes on the head and stem of HA (see Table A in S1 Text) and that the key factor limiting the boosting of antibodies to conserved epitopes on the stem of influenza is pre-existing antibodies to these epitopes rather than antibodies to the head of HA.
The data in Fig 1A shows the response to HA from H5N1 which is not circulating in the human population, and we see that the prevaccination level of antibody to the head is much lower than that to the stem. The question is whether there are inherent differences in the head and stem that could account for the observations for the different levels of boosting to head and stem that we see in Fig 1B. We point out a few reasons against this being a major factor. First, our current models parsimoniously explain the observations. Second, if we look at the data in Fig A in S1 Text, which describes the boosting followed by vaccination with H1N1, we see that there is more overlap between the prevaccination titers to head and stem of HA and the responses to the head and stem are comparable. For example, three data points with low prevaccination titers to the stem reach post-vaccination titers similar to points having low prevaccination titer to the head of HA.
Both H5N1 and H1N1 studies show that prevaccination antibodies to the stem of HA have a minimal impact on the expansion of responses to the head of HA, and similarly prevaccination antibodies to the head do not significantly affect the boosting of antibodies to the stem of HA. This can be seen in Figs C,D and Table A in S1 Text. We note that there is some indication that increased levels of pre-existing antibodies to the head of HA might reduce the boosting of antibodies to the stem, though the p − value does not reach statistical significance (p − value = 0.137 in Fig D in S1 Text). Our modeling approach can be easily extended to consider interference between binding of antibodies to the head and stem of HA. As the relevant experimental data becomes available this should allow assessment of its role in limiting boosting of antibody to stem during scenarios of immunization with an antigenically drifted vs. shifted HA [15, 16, 28].
One could argue that if an antigen is cleaved into two separate parts then all three models (ACM, FIM and EMM) could be consistent with the independent boosting of responses to epitopes on different parts. This is unlikely to be the case for HA because the stem of HA does not maintain its native conformation in the absence of being linked to the head of HA [29, 30], so it would lead to more head than stem antigen and we would expect much higher responses to the head of HA under the same experimental settings. We have a limited dataset, but in Fig 6B we see three individuals with low preexisting antibody to the stem. The lines joining the boosting of head and stem in these individuals do not have lower slopes than the average line, indicating that there are comparable levels of stem and head epitopes.
We have used simple phenomenological models to generate qualitative predictions. This is because, in the absence of detailed information on the parameter values or sufficient data to estimate them, simpler models frequently generate more robust qualitative results than complex ones [22]. The conclusions presented here are robust to a number of variations in the model specification. First, they hold up in the context of different models of antigenic drift (i.e., two-epitope model and a three-epitope model with steric interference) and are robust to considering interactions between antibodies and B cells against multiple epitopes on the head of HA. Second, our results are not highly sensitive to parameter values within biologically reasonable ranges (see Table B in S1 Text for the choice of model parameters and Fig G in S1 Text).
The immune response to influenza vaccination is complex and not fully understood. The extensions of the models we have proposed can play an important role in furthering our understanding. An important next step is to include the differentiation of B cells during responses and reactions occurring in different locations such as the site of infection and the germinal centers of lymph nodes as well as CD4 T cell help and affinity maturation [31–33]. Doing so would require a much more complex model, and ideally would be done in conjunction with an animal model system which allows measurement of the relevant parameters and testing of various components of the model. Another extension involves going from vaccination to natural infection. This will involve incorporating resource (target cells) limitation and innate immunity [34–38] as well as T cells [39–41]. These models could be used to consider the effect of pre-existing antibodies on infection with drifted and shifted virus strains. Early studies suggested that the response of an individual would be dominated by the antibodies expanded by the first strain encountered; this was termed “original antigenic sin” [42, 43]. More recent studies have shown the evolution of the response is more complex and used the terms “antigenic seniority” and “backboosting” to describe how exposure to the current strain can lead to boosting of responses to strains that were encountered previously [44–46]. Multi-epitope models which explicitly consider the different head epitopes together with antigenic maps of changes in these epitopes may help elucidate this complex dynamics.
Another area that requires attention is the measurement of vaccine efficacy. There are many facets of vaccine efficacy; vaccines may protect against pathology (VEP), susceptibility to infection (VES) and/or infectivity and transmission (VEI) [3, 4, 47], and estimates of influenza vaccine efficacy vary widely (e.g., the median monovalent pandemic H1N1 vaccine effectiveness in five observational studies was 69% and the range was 60% to 93% [5]). A more detailed understanding of the immune dynamics in response to influenza will be key to disentangling these complexities. Extending the EMM from vaccination to natural infection will be necessary both to further test the model and to predict its impacts on influenza epidemiology.
This study sheds light on puzzles of influenza immunity and vaccination. It elucidates rules (e.g., epitope masking) and key parameters (e.g., antigen dose) associated with the humoral immune response to influenza vaccination in non-naive hosts and may help guide the transfer of next-generation, stem-specific influenza vaccines from animal models to humans. The results suggest that explanations for differences in the antibody response to head and stem epitopes need not invoke highly different mechanisms; rather, they may result from differing pre-exposure antibody titers. Our model which is based on the results of vaccination studies in humans suggests that generating high levels of antibodies to the stem of HA (required for broad strain-transcending immunity) will involve delivery of a sufficiently high dose of antigen which overcomes the effect of epitope-masking. It would be interesting to extend models such as the ones we describe in this paper to better understand the success of immunization strategies in mice and ferrets [13–19], such as those involving nanoparticle based vaccines, that generate high levels of antibodies to the stem of HA [14, 19]. In particular, to determine whether the success of nano-particle based immunization is due to the maintenance of high levels of antigen for an extended period of time or due to other reasons such as steric accessibility of the stem in the nano-particles.
In conclusion, the confrontation of the predictions of qualitative models with a reanalysis of experimental data leads us to conclude that epitope masking is an important factor in the immune response following re-exposure to influenza’s HA. There is a clear need for a new generation of more effective and cross-protective vaccines, and understanding the key mechanisms and parameters that drive the generation of humoral immunity, such as epitope masking, prevaccination antibody titers and antigen dose, is critical.
In the following sections we describe the models we consider in the paper.
The one-epitope model for an antibody response to an antigen includes four variables such as free antigen (Hf), bound antigen (Hb), B cells specific for the antigen (B), and the antibodies (A) secreted by B cells. In accord with clonal selection, B cells are stimulated and proliferate in a manner dependent on the concentration of antigen. We consider a number of scenarios. In the basic model, antigen stimulates clonal expansion of specific B cells, which produce antibody. Antibody production continues until all the antigen has decayed. In the antigen clearance model (ACM) antigen bound to antibody is removed faster than free antigen. In the FcγRIIB mediated inhibition model (FIM) antigen-antibody complexes inhibit the stimulation and proliferation of B cells. In the epitope masking model (EMM) antibody binding to an antigen masks the epitope preventing it from stimulating B cells. The equations below describe these four models (see also Fig 2A).
For the basic model db = df = 0.5, α = 0, and δ = 1. For the ACM df = 0.5, db = 3, α = 0, and δ = 1. For the FIM db = df = 0.5, α = 0.01 (>0 in general), and δ = 1. For the EMM db = df = 0.5, α = 0, and δ = 0. The model parameters were chosen to obtain the key features of a typical antibody response. We rescaled the initial values of antibodies and B cells to unity at the naive state (prior to the first vaccination), and set a = dA, so that at equilibrium in naive or memory states we have B ≈ A. For the recall responses the initial values of antibodies and B cells were set equal to the level of preexisting immunity shown on the corresponding figures. Parameter values are specified in Table 1.
We note that B cell stimulation (by antigen) and inhibition (through FcγRIIB) are saturating functions of antigen and antigen-antibody complexes, respectively. The antigen density for half-maximal stimulation of B cells is equal to ϕ and the density of antigen-antibody complexes at which inhibition of B cells activation is half-maximal is equal to (1/α). We choose the term 1/α so that α = 0 corresponds to the absence of Fc-mediated inhibition.
We extend one-epitope model to two-epitope model by considering an antigen with one epitope on the head of HA (X) and one epitope on the stem of HA (S). We let BX and AX represent B cells and antibodies specific for epitope X (and similarly BS and AS for epitope S). The free antigen is HXS, and there are three additional states for antigen: HOS, HXO and HOO, representing antigen with antibodies bound to X, S or both epitopes, respectively (see schematic in Fig 3A). Parameters are the same as for the one-epitope model.
This model considers an antigen with two epitopes (X and Y) on the head of HA and one epitope (S) on the stem of HA. Steric interference for the antibodies binding to the two epitopes on the head of HA is introduced into model with parameter β. Parameter β = 0 corresponds to the case of no steric interference and β = 1 corresponds to the case when antibody bound to one epitope on the head completely blocks binding of antibodies to the other epitope on the head. The scheme showing the transitions between the different states of an antigen with epitopes X, Y and S is shown in Fig B in S1 Text. The corresponding model equations are shown below. Parameters are the same as for the one-epitope model.
|
10.1371/journal.ppat.1007358 | IFN-γ immune priming of macrophages in vivo induces prolonged STAT1 binding and protection against Cryptococcus neoformans | Development of vaccines against opportunistic infections is difficult as patients most at risk of developing disease are deficient in aspects of the adaptive immune system. Here, we utilized an experimental immunization strategy to induce innate memory in macrophages in vivo. Unlike current trained immunity models, we present an innate memory-like phenotype in macrophages that is maintained for at least 70 days post-immunization and results in complete protection against secondary challenge in the absence of adaptive immune cells. RNA-seq analysis of in vivo IFN-γ primed macrophages revealed a rapid up-regulation of IFN-γ and STAT1 signaling pathways following secondary challenge. The enhanced cytokine recall responses appeared to be pathogen-specific, dependent on changes in histone methylation and acetylation, and correlated with increased STAT1 binding to promoter regions of genes associated with protective anti-fungal immunity. Thus, we demonstrate an alternative mechanism to induce macrophage innate memory in vivo that facilitates pathogen-specific vaccine-mediated immune responses.
| Fungal infections are a significant global health problem that can affect anyone, however, individuals with a weakened immune system are most at risk. Cryptococcus neoformans infections can progress to meningitis in immune compromised individuals accounting for nearly 220,000 new cases annually, resulting in 181,000 deaths. Vaccine strategies tend to target CD4+ T cells for the generation of protective memory responses. However, immune compromised individuals have decreased numbers of these adaptive cells, providing a challenge for anti-fungal vaccine design. Here, we define a cellular mechanism by which macrophages, an innate cell population, generate protective immune responses against C. neoformans following initial exposure to a C. neoformans strain that secretes IFN-γ. We determined that the macrophages primed in vivo have heightened proinflammatory cytokine responses upon secondary exposure to C. neoformans in a manner that is mTOR-independent, yet dependent on histone modification dynamics. We show that IFN-γ primed macrophages can maintain STAT1 binding to the promoter regions of key proinflammatory genes long after the initial exposure. Remarkably, our studies show long-lived, cryptococcal-specific protective immunity in vivo. The results presented herein demonstrate that innate cell populations, namely macrophages, can be utilized as vaccine targets to protect against cryptococcal infections in immune compromised populations.
| Vaccines in clinical use today were designed from the onset to generate antigen-specific memory T and/or B cell (antibody) immune responses (reviewed in [1–3]). However, vaccines designed to elicit T cell and/or antibody-mediated immune responses are predicted to be ineffective in patients rendered immune compromised due to diseases such as HIV/AIDS or immunosuppressive therapies to prevent solid organ transplant rejection or ameliorate various autoimmune diseases. Novel approaches in vaccine design are needed to induce protective immunity in the increasing population of immunocompromised individuals.
The ability of innate immune cells to possess memory like-responses upon re-exposure to an antigen has been documented in plants, insects, and more recently, mammals. Memory-like responses by innate immune cells have been termed innate memory or “trained” immunity (reviewed in [4–7]). Trained immunity occurs independent of B and T cell adaptive responses, increases resistance of the host to reinfection, and involves cells including monocytes, macrophages, and NK cells [4]. Studies have shown that following the initial antigen stimulation, innate cells are capable of enhanced pro-inflammatory cytokine recall responses following secondary exposure to intracellular pathogens at 7 days post training [8, 9]. Thus, this concept suggests that immunization strategies that drive memory-like innate immune responses may provide a novel and effective mechanism for inducing vaccine-mediated protection in immunocompromised patients.
Our laboratory has utilized a fungal vaccine model that delivers interferon-γ (IFN-γ) in vivo to demonstrate the induction of protective immunity against disease in vaccinated B cell-deficient mice and CD4+/CD8+ T cell-depleted mice [10, 11]. These studies reveal that protective immunity can be achieved in hosts devoid of immune cells traditionally considered necessary for adaptive immunity and provide proof-of-concept that protection can be achieved in immunocompromised patients. However, the effector cell population and mechanism responsible for protection is unknown.
In the current studies, we sought to determine the mechanism(s) underlying innate memory using our protective fungal vaccine model. We observed that protectively immunized B cell KO mice that were subsequently depleted of T cells, neutrophils, and natural killer (NK) cells were protected against challenge with the opportunistic fungal pathogen Cryptococcus neoformans; a significant cause of morbidity and mortality in HIV+/AIDS patients world-wide [12–14]. Macrophages from protectively immunized mice displayed enhanced antigen-specific cytokine recall responses using a mechanism distinct from the established trained immunity paradigm. The transcriptome of pulmonary macrophages isolated from the protectively immunized mice early post-challenge demonstrated increased expression of genes associated with protective anti-fungal immune responses; namely, genes associated with the signal transducer and activator of transcription 1 (STAT1) signaling pathway. Enhanced binding of STAT1 to the promotor regions of known IFN-γ-induced genes was observed in immunized, unchallenged pulmonary macrophages, correlating with these macrophages’ ability to quickly respond to secondary challenge. Thus, IFN-γ priming of macrophages in vivo resulted in the establishment of antigen-specific innate memory-like responses through 70 days post-immunization and provided complete protection against secondary challenge in the absence of adaptive immune cells. Altogether, our studies demonstrate the feasibility of vaccine strategies designed to enhance innate immune responses against specific pathogens to provide protection against diseases that target immunocompromised individuals.
Previous studies showed that mice given an experimental pulmonary infection with a fungal pathogen, C. neoformans, engineered to produce interferon-γ (IFN-γ), denoted H99γ, were protected against subsequent challenge with the fully pathogenic, non-IFN-γ producing C. neoformans strain H99 [15]. Protection was also observed in H99γ immunized mice deficient in B cells or depleted of CD4+ and/or CD8+ T cells and challenged with wild-type (WT) yeast [10, 11, 15]. To elucidate the effector cell population required for protection, B cell KO mice were protectively or non-protectively immunized with C. neoformans strain H99γ or heat killed C. neoformans strain H99γ (HKH99γ), respectively. The mice were rested for 70 days, depleted of both CD4+ and CD8+ T cells or given isotype control antibodies and then challenged with WT C. neoformans strain H99 (Fig 1A). Cell depletions were maintained throughout the observation period and were verified by flow cytometry (S1 Fig). B cell KO mice and B cell KO mice depleted of both T cell subsets showed a 90% (Fig 1A; p = 0.3173) and 80% (Fig 1B; p = 0.3613) survival rate, respectively. These findings led us to hypothesize that the effector cell population responsible for protection in B and T cell-deficient mice is a member of the innate immune system. To investigate this hypothesis, we depleted protectively immunized B cell KO mice of both subsets of T cells as well as natural killer (NK) cells and neutrophils and subsequently challenged the mice with WT cryptococci (Fig 1B; S1 Fig). Remarkably, the mice rendered deficient in adaptive immune cells in addition to these two innate cell populations were 100% protected against challenge (p = 1.0), demonstrating that neither T cells, B cells, neutrophils nor NK cells were necessary for protective immunity.
Previous studies by our lab and others have demonstrated that classically activated macrophages (M1) are important for protective immune responses against C. neoformans as they produce pro-inflammatory cytokines and can kill cryptococci via nitric oxide production [16–22]. Recent studies investigating innate memory-like responses have shown that trained monocytes/macrophages have elevated cytokine recall responses following secondary exposure to an antigen [8]. Therefore, we sought to investigate whether macrophages from protectively immunized mice were trained to respond quickly upon secondary exposure to C. neoformans. Macrophages were isolated from the lungs of mice that were immunized with C. neoformans strain H99γ or HKH99γ and rested for 70 days as well as naïve mice. The macrophages were cultured ex vivo for 24 hours with a C. neoformans strain H99 calcineurin subunit α deletion mutant (cna1Δ) which is unable to proliferate at 37°C and, thus, will not outgrow the macrophages in culture. The supernatants were then analyzed for the production of cytokines associated with pro-inflammatory and protective anti-cryptococcal responses. Interestingly, pulmonary macrophages from protectively immunized mice stimulated with cryptococci produced significantly more IL-2 (p = 0.0013) and IFN-γ (p = 0.0191), cytokines associated with protective responses (Fig 1C and 1D). No difference in TNF-α production was detected between stimulated versus unstimulated pulmonary macrophages from protectively immunized mice (Fig 1E).
As the spleen is a reservoir for immune cells including macrophages, and macrophages are known to traffic back to the spleen after the resolution of infection, we also tested splenic macrophages from immunized and naïve mice for cytokine recall abilities. A significant increase in IL-2 (p = 0.0035), IFN-γ (p = 0.0134) and TNF-α (p = 0.0289) production was detected by stimulated macrophages from protectively immunized mice compared to non-protectively immunized mice (Fig 1F–1H). No significant increase in cytokine production was detected in macrophages from naïve or HKH99γ immunized mice from both pulmonary and splenic macrophages (Fig 1C–1H). To ensure that the cytokines detected in culture were produced by macrophages, we analyzed the intracellular cytokine production of the macrophages by imaging flow cytometry. After 6 hours in culture, macrophages were confirmed positive for the macrophage markers CD11b and CD64. We did not further access F4/80 cell surface expression as anti-F4/80 antibodies still bound to the cells following positive selection are likely to sterically hinder binding of anti-F4/80 antibodies used for flow cytometry. The flow cytometry data confirmed that the CD11b+CD64+ macrophages did indeed produce the cytokines IL-2, IFN-γ, and TNF-α (S2 Fig).
Previous data from our lab indicated that macrophages from protectively immunized mice are more fungistatic than those from non-protectively immunized mice [19]. In the current study, we sought to determine if there were differences in phagocytic capabilities of the macrophages. Splenic macrophages isolated from HKH99γ immunized or H99γ immunized mice were cultured with an mCherry expressing strain of C. neoformans (KN99mCH) for 6 hours and subsequently analyzed for macrophage association with or internalization of the cryptococci as described previously [23]. We observed a shift in increased association of macrophages from protectively immunized mice with cryptococcal cells, though the difference was not significant (Fig 2). However, significantly more macrophages from H99γ immunized mice internalized cryptococci compared to macrophages derived from HKH99γ immunized mice. Altogether, these data suggest that immunization with the IFN-γ producing strain primes the macrophages to respond to a secondary exposure to the yeast both by increased internalization of the pathogen and increased proinflammatory cytokine production, which likely aids in protection during in vivo challenge.
Cytokine recall responses of pulmonary and splenic macrophages from protectively immunized mice to C. neoformans appeared to follow a similar trend. Next, we sought to assess the specificity of the cytokine recall response to other members of the Cryptococcus species complex. Previous studies have shown the importance of T cell responses to protect against disparate serotypes in vivo [24], however, evaluation of the macrophage contribution to protection is unknown. Considering that significantly more macrophages can be obtained from splenic tissues than lung, we elected to use splenic macrophages to significantly reduce the number of mice needed to perform subsequent assays. We cultured macrophages from HKH99γ immunized or H99γ immunized mice for 24 hours with heat killed C. neoformans strain H99 (serotype A), C. deuterogattii strain R265 (serotype B), C. bacillisporus strain WSA87 (serotype C), C. deneoformans strain 52D (serotype D) or media alone and measured the cytokines present in the supernatants. Macrophages from protectively immunized mice produced significantly more IL-2 when stimulated with serotypes A, B, and D compared to stimulated macrophages from naïve and non-protectively immunized mice (Fig 3A). The macrophages from protectively immunized mice also produced significantly more IL-2 when stimulated with serotypes A, B, and D and more IFN-γ and TNF-α when stimulated with serotype A compared to unstimulated cells (Fig 3A–3C). For IFN-γ, there was a trend towards increased cytokine production in macrophages from H99γ immunized mice stimulated with all four serotypes, however it was not statistically significantly increased compared to either unstimulated cells or stimulated cells from naïve or HKH99γ immunized mice. Nevertheless, the lack of statistical significance observed is likely due to the lack of IFN-γ production by macrophages from naïve or non-protectively immunized mice following stimulation with the various serotypes, which prevented the performance of statistical analysis. Interestingly, the levels of TNF-α was relatively low, with the exception of macrophages from protectively immunized mice stimulated with serotype D. Overall, there was no significant change in TNF-α secretion measured across all conditions except for macrophages from H99γ immunized mice stimulated with serotypes A and D compared to unstimulated cells and compared to serotype D stimulated cells from naïve and HKH99γ immunized mice (Fig 3C). Previous studies show that immunization of mice with heat killed C. neoformans strains (including HKH99γ) or live C. deneoformans 52D does not result in protective immunity [11, 15, 24]. However, these data suggest that immunization with H99γ stimulates macrophages to provoke a memory-like response following stimulation with other Cryptococcus serotypes.
Our next question was whether or not immunization with C. neoformans H99γ can induce memory-like responses against non-cryptococcal antigens. To answer this question, we cultured splenic macrophages from immunized and naïve mice for 24 hours with LPS, heat killed Candida albicans, heat killed Staphylococcus aureus or C. neoformans cna1Δ to access the specificity of the cytokine recall response. Cytokine levels in cultures containing macrophages from protectively immunized mice appeared to only produce significantly elevated levels of several cytokines upon stimulation with C. neoformans. Macrophages from protectively immunized mice exposed to C. neoformans secreted IL-2 at significantly elevated levels compared to macrophages cultured in media alone or macrophages from naïve and non-protectively immunized mice (Fig 4A). Comparable levels of IFN-γ were secreted by macrophages from all immunization groups when stimulated with S. aureus, however, only the macrophages from the protectively immunized mice produced significant levels of IFN-γ in response to C. neoformans (Fig 4B). Likewise, TNF-α was detected in supernatants of all immunization groups during co-culture with S. aureus and C. albicans at comparable levels suggesting no trained response to these microbes. However, TNF-α production was increased by 4-fold in protectively immunized mice compared to all other groups following cryptococcal stimulation, although not statistically significant (Fig 4C). Granulocyte macrophage-colony stimulating factor (GM-CSF) and Th2-associated cytokines IL-4 and IL-5 were only significantly increased in macrophages from the protectively immunized mice co-cultured with Cryptococcus, indicating that these responses were also Cryptococcus-specific (Fig 4D–4F). We do note that the IL-5 levels observed were relatively low (<8 pg/ml), and likely not biologically relevant. Similarly, the levels of IL-12p70 were low and did not appear to show bias to any immunization group or stimuli (Fig 4G). Interestingly, macrophages from all groups produced elevated levels of the immune regulatory cytokine IL-10 when stimulated with C. albicans (123–282 pg/ml) or S. aureus (57–67 pg/ml; Fig 2H). However, IL-10 was not produced in response to C. neoformans stimulation in any immunization group (Fig 4H). Overall, these data suggest that macrophages from protectively immunized mice have a significant C. neoformans-specific cytokine recall response and are in contrast with previous reports demonstrating a lack of specificity in the macrophages/monocytes trained with C. albicans or β-glucan [8]. Thus, an alternative mechanism may be responsible for the induction of enhanced cytokine recall responses by macrophages from H99γ immunized mice.
Previous studies have demonstrated that inhibition of mTOR results in a loss of β-glucan induced trained immunity in monocytes [9]. We, therefore, determined the outcome of mTOR inhibition on cytokine recall responses by macrophages isolated from protectively immunized mice following exposure with C. neoformans. We observed similar levels of total and phosphorylated mTOR (p-mTOR) within macrophages from protectively immunized mice following exposure to Cryptococcus in vitro compared to p-mTOR levels in unstimulated macrophages (Fig 5A). Treatment of macrophages from protectively immunized mice with torin, an ATP-competitive inhibitor of mTOR kinase activity, concurrent with C. neoformans stimulation resulted in a decrease in p-mTOR below baseline levels observed in unstimulated macrophages. However, torin treatment did not suppress production of IL-2 or TNF-α by macrophages from protectively immunized mice (Fig 5B and 5D). While not statistically significant, lower levels of IFN-γ production were observed in macrophages from protectively immunized mice in the presence of torin which may have a biological impact on subsequent macrophage recall responses (Fig 5C). These data suggest that the cytokine recall responses of IFN-γ primed macrophages is independent of the mTOR pathway and is induced by a mechanism distinct from that observed for β-glucan or C. albicans induced trained immunity.
We performed whole transcriptome analysis of macrophages isolated from protectively compared to non-protectively immunized mice at days 1 and 3 post-challenge with WT C. neoformans yeast as well as macrophages from immunized but unchallenged mice. Ingenuity Pathway Analysis (IPA) of RNA-Seq results showed a rapid up-regulation of canonical pathways for interferon and IL-17 signaling in macrophages from protectively immunized mice compared to non-protectively immunized mice at day one post-challenge (Table 1). Notably, transcripts for STAT1, SOCS1, IFIT2, NOS2, IFNG, CXCL10, and IL6 are among the genes up-regulated in the macrophages from protectively immunized mice at day-one post-challenge. Also, network analysis showed an up-regulation of the STAT1 network in macrophages from the protectively immunized mice (Fig 6A and 6B), confirming our previous findings [19]. Gene ontology (GO) analysis of RNA-Seq results from macrophages isolated at day 1 post-challenge resulted in GO terms associated with pro-inflammatory and anti-microbial pathways (Fig 6C; S1 Table). IPA of RNA-Seq results from macrophages isolated at day 3 post-challenge showed a rapid up-regulation of canonical pathways associated with T and B cell signaling, communication between the innate and adaptive immune system, and overall cytokine responses (Table 1). IPA also revealed up-regulation of networks associated with protective immunity and M1 macrophage activation as transcripts for genes including SOCS1, STAT4, IL12B1, and IFNG were up-regulated in the primed macrophages (Fig 6D and 6E). GO analysis of RNA-Seq results from macrophages isolated at day 3 post-challenge revealed up-regulation of pathways associated with a protective immune response in macrophages from protectively immunized mice (Fig 6F; S2 Table). Analysis of genes from immunized, unchallenged mice did not reveal significant changes in immune related genes with the exception of VCAM1 (2.18 fold increase) and CXCL9 (3.01 fold increase) in protectively immunized compared to non-protectively immunized mice. Altogether, these data demonstrate that immunization with C. neoformans strain H99γ primes the anti-microbial response of macrophages in vivo to secondary exposure to the yeast.
Studies show that β-glucan training of monocytes induces histone modifications related to genes associated with protective immunity against various microbial pathogens [4, 7, 8, 25]. To investigate how IFN-γ priming results in the anti-cryptococcal activity of the macrophages, we cultured splenic macrophages from protectively and non-protectively immunized mice or naïve mice with 5′-deoxy-5′-methylthioadenosine (MTA), a histone methyltransferase (HMT) inhibitor, pargyline, a histone demethylase (HDM) inhibitor, givinostat, a histone acetyltransferase (HDAC) inhibitor, or epigallocatechin gallate (EGCG), a histone acetyltransferase (HAT) inhibitor. As shown in Fig 1, stimulation of the macrophages with cna1Δ induced elevated cytokine production only in macrophages from protectively immunized mice. Interestingly, co-culture with the HMT inhibitor MTA and HDAC inhibitor givinostat significantly decreased IL-2 and IFN-γ production by macrophages from protectively immunized mice (Fig 7A and 7B). In addition, EGCG reduced IFN-γ production by macrophages from protectively immunized mice (Fig 7B). TNF-α production showed a similar trend to IFN-γ, however the values were not statistically significant (Fig 7C). Overall, these data suggest that the increased cytokine recall potential of macrophages from protectively immunized mice is due to changes in histone modifications elicited by immunization with C. neoformans strain H99γ.
We subsequently focused on modifications of methylation patterns on histones in macrophages. We analyzed cytokine recall responses by splenic macrophages from naïve, non-protectively immunized and protectively immunized mice when cultured with inhibitors for enzymes that induce specific methylation patterns on histones. The EZH2 inhibitor GSK343 prevents tri-methylation of histone 3 lysine 27 (H3K27me3) which is associated with repression of transcription. UNC0638, an inhibitor of G9a, prevents the repressive modification H3K9me3. Finally, we used an inhibitor of the MLL complex, MI-2, thus preventing addition of H3K4me3 which is associated with active transcription. As expected, inhibition of H3K4me3 resulted in a decrease in IL-2 (p < 0.05), IFN-γ (p < 0.01) and TNF-α (p < 0.01) cytokine production by in vivo IFN-γ primed macrophages (Fig 7D–7F). In addition, inhibition of the repressive mark H3K27me3 also resulted in decreased IL-2 (p < 0.05), IFN-γ (p < 0.01) and TNF-α (p < 0.01) cytokine production. This may indicate an indirect role for H3K27me3 that aids in cytokine production by the trained macrophages. Inhibition of H3K9me3 resulted in a decrease in IFN-γ, although not significant (p > 0.05), and TNF-α (p < 0.05) cytokine production, but not IL-2 production by IFN-γ trained macrophages.
IFN-γ immune primed macrophages rapidly up-regulate the STAT1 pathway following challenge with C. neoformans in vivo. Consequently, we examined the effects of the methylation inhibitors on STAT1 activation. Stimulation of splenic macrophages from protectively immunized mice with C. neoformans increases total STAT1 and phosphorylated STAT1 (pSTAT1) compared to unstimulated macrophages (Fig 7G). Interestingly, co-culture of the macrophages with C. neoformans and GSK343 or MI-2 resulted in decreased expression of total STAT1 protein and pSTAT1 (Fig 7G), suggesting a role for H3K27me3 and H3K4me3 in STAT1 pathway activation. Inhibition of H3K9me3, however, had no effect on total STAT1 or pSTAT1 expression. Altogether, these data suggest that changes in the methylation patterns contribute to the cytokine recall responses and activation of the STAT1 pathway in macrophages from protectively immunized mice.
Data from the previous assays show the importance of STAT1 signaling to the enhanced cytokine production observed in macrophages from protectively immunized mice. We next sought to determine if the immunization with C. neoformans strain H99γ led to an alteration in STAT1 binding at promotor regions of certain known IFN-γ-induced genes. For these experiments, mice were immunized with either HKH99γ or H99γ and allowed to rest for 70 days. Macrophages were then isolated from the lungs of the immunized mice and chromatin immunoprecipitation (ChIP) was performed using an antibody against STAT1. The resulting DNA was amplified by real-time ChIP qPCR using a custom ChIP array. Target genes were chosen based on the results of the RNA-seq analysis on day 1 post-challenge, selecting genes that were significantly up-regulated in macrophages from H99γ immunized mice compared to macrophages from HKH99γ immunized mice. The data revealed an overall trend for increased percent IP fold enrichment of STAT1 binding to sites 1kb upstream of the transcription start site for the selected genes in macrophages from protectively immunized mice compared to non-protectively immunized mice (Table 2). We observed significantly increased binding of STAT1 to promotor regions of guanylate binding protein 6 (GBP6), CXCL10, and CIITA. We also detected increased binding of STAT1 to promotor regions of IRF-1, CXCL11, and IFIT2 with p-values approaching statistical significance. Again, transcripts for each of these genes are significantly elevated post-challenge suggesting that immunization with H99γ induced lasting STAT1 binding at promotor regions of select IFN-induced genes within macrophages of protectively immunized mice thus priming them to readily respond to a second Cryptococcus challenge.
In the present study, we demonstrate that macrophages from mice immunized with an IFN-γ producing strain of C. neoformans are primed to rapidly respond to a subsequent exposure to this pathogen. The effect appears to be STAT1-dependent, as demonstrated by expedited transcription of genes that are directly and indirectly associated with the STAT1 pathway as well as increased binding of STAT1 at promotor regions of select genes. Activation of STAT1 in macrophages results in their polarization toward an M1 phenotype which has been widely demonstrated as anti-cryptococcal and necessary for protective immunity against C. neoformans [16, 17, 19–22, 26]. To date, myelofibrosis treatment of two patients with Ruxolitinib, which inhibits JAK1,2 and likely STAT1, is hypothesized to have contributed to the development of cryptococcosis in these individuals [27, 28].
Macrophages primed with IFN-γ respond quickly to a secondary signal, resulting in tumoricidal and microbicidal activities of these cells [29–37]. It seems likely that by immunizing mice with the IFN-γ producing strain of C. neoformans, the macrophages are primed in vivo to respond to a subsequent exposure to cryptococci. Remarkably, these macrophages remain trained at least as long as 70 days post-immunization. At this point, we are unable to ascertain if the macrophages in the lungs at day 70 post-immunization are the same macrophages present during immunization. Alveolar macrophages are known to self-renew [38], therefore some macrophage turnover is expected. Future studies to address the question of macrophage turnover and long-term training are required.
Trained immunity refers to a non-specific response of innate immune cells that is independent of T and B cells and increases resistance against reinfection [4–7]. Studies show that β-glucan, a major component of fungal cell walls, trains monocytes to respond with enhanced production of IL-6 and TNF-α not only when stimulated with C. albicans, but also following stimulation with LPS, Mycobacterium tuberculosis, and the TLR-2 agonist Pam3Cys [8] which differs from our model which suggests the C. neoformans/IFN-γ priming is specific for C. neoformans. A recent study investigating β-glucan training concluded that the trained monocytes displayed enhanced survival in vitro via partial inhibition of apoptosis and that the increased survival of these trained cells explained the elevated levels of pro-inflammatory cytokines elicited by LPS challenge compared to untrained cells [39]. This study also showed that systemic administration of β-glucan to mice rendered the animals more responsive to LPS challenge after 4 days, however the effect was short lived as the enhanced cytokine responses were lost by day 20 post-β-glucan immunization [39]. These results differ with those in our model in which animals are still protected and macrophages exhibit heightened cytokine responses as far as 70 days post-IFN-γ priming in vivo.
Immunization with the IFN-γ producing C. neoformans strain requires the STAT1 pathway to induce protection [16, 17], thus, the mechanisms for eliciting IFN-γ innate memory and β-glucan training are likely different. Training with β-glucan signals through the C-type lectin receptor Dectin-1 [8] which specifically recognizes β-glucan [40]. However, C. neoformans possesses a polysaccharide capsule that efficiently masks detection of several cell wall components, including β-glucan. Thus, it is unlikely that β-glucan in the C. neoformans cell wall would trigger innate memory in the macrophages. In addition, we demonstrate that the memory-like response of macrophages from protectively immunized mice does not require mTOR activity during the anamnestic response; a signaling pathway required for β-glucan induced training [9]. In fact, cytokine production increases when mTOR activation is inhibited. Recent studies have shown that IFN-γ priming can inhibit mTORC1 (of which mTOR is the catalytic subunit) in human macrophages, increasing pro-inflammatory cytokine production by translational suppression of repressors of inflammation [41]. A recent study using M. bovis Bacillus Calmette-Guerin (BCG) to induce memory-like innate immune priming in mice revealed that the BCG bacteria can infiltrate the bone marrow and prime the hematopoietic stem cells (HSCs) to generate monocytes/macrophages that are more antimycobacterial than controls [42]. This priming appears to be dependent on IFN-γ signaling and STAT transcription factors [42], while conversely, training of HSCs with β-glucan occurs via IL-1β and activation of the GM-CSF/CD131 axis [43]. These data indicate that there are disparate mechanisms that can lead to innate training in mice. In addition, defects in trained immunity have been documented in patients with STAT1-mediated chronic mucocutaneous candidiasis (CMC) and following blockade of IFN-γ in PBMCs of healthy donors primed with C. albicans [44]. Thus, priming macrophages via activation of the STAT1 pathway does appear to be important for memory-like innate priming with a correlation to training in context with microbial infection rather than with microbial by-products like β-glucan. The data presented herein are the first to show that in vivo IFN-γ priming of macrophages can extend the memory-like phenotype for weeks, providing protection against subsequent infection.
Previous studies in our lab have shown that mice with macrophages deficient in STAT1 are unable to mount a protective immune response against C. neoformans strain H99γ [16]. In the current studies, we demonstrate by whole transcriptome analysis that macrophages from protectively immunized, yet unchallenged, mice have few transcript variances between groups. These data are in line with other RNA sequencing studies that do not show transcriptional differences in trained monocytes/macrophages in their resting state [45]. Remarkably, we detected robust activation of the STAT1 pathway and interferon-related canonical pathways in protectively immunized mice as early as day one post-challenge with WT cryptococci. Previous studies indicated that at day one post-challenge with C. neoformans, pulmonary macrophages from protectively immunized mice have increased gene expression for STAT1 and downstream members of the STAT1 pathway, as well as phosphorylation of STAT1, confirming the RNA-seq analysis [19]. By day three post-challenge, pathways associated with innate/adaptive immune communication are up-regulated. Since this immune response in macrophages trained by C. neoformans/IFN-γ happens so quickly, it suggests that the innate arm of the immune system is responsible for protection. Furthermore, these studies show that protection occurs in the absence of B and T cells, as well as NK cells and neutrophils, demonstrating the critical role that macrophages play in vaccine-mediated immune responses. It must be noted that cell types not depleted, such as dendritic cells, may be playing a role in the protective immune response in the multi-cell type depleted mice. Further investigation into DC memory-like responses is ongoing.
In the C. neoformans/IFN-γ training model, macrophages and other immune cells in mice immunized with C. neoformans strain H99γ are simultaneously exposed to cryptococcal antigen and IFN-γ. IFN-γ priming of human monocytes in vitro can induce sustained occupancy of transcription factors STAT1, IRF-1, and associated histone acetylation at promoters and enhancers at the TNF, IL6, and IL12B loci [46]. In these studies, priming did not induce transcription but created a poised environment that enhanced gene transcription during secondary stimulation with LPS [46], signifying that IFN-γ/STAT1 activation is important for mediating anti-microbial responses to a variety of stimuli including TLR agonists [47]. These data are similar to what we have observed in pulmonary macrophages from protectively immunized mice in that STAT1 is bound to the promotor regions of certain IFN-γ induced genes, including CXCL10, CIITA, and GBP6 at higher rates than in macrophages from non-protectively immunized mice. Transcription of these genes does not appear to be active as there is no difference in cytokine production in the macrophages when left unstimulated ex vivo.
Investigation into specific histone modifications in methylation patterns revealed that H3K27me3, H3K4me3 and, to a lesser extent, H3K9me3 are important for cytokine recall responses of macrophages following stimulation with C. neoformans. We have thus far observed that inhibition of EZH2, which tri-methylates H3K27, resulted in decreased cytokine production and inhibition of STAT1 phosphorylation. Tri-methylation of H3K4 is often associated with active transcription of genes in macrophages [48–50]. M1 macrophages up-regulate the histone methyltransferase MLL complex which adds the H3K4me3 mark [51, 52]. H3K4me3 is found at the promoter site for pro-inflammatory chemokine and M1 macrophage marker CXCL10, correlating with MLL activation and activity [52]. Furthermore, NF-κB can recruit the MLL complex to add H3K4me3 and activate transcription of the Nos2 and IL6 genes in Listeria monocytogenes infected macrophages in synergy with STAT1 [51]. In the current studies, inhibition of the MLL complex in IFN-γ trained macrophages with MI-2 resulted in decreased pro-inflammatory cytokine production, further confirming that H3K4me3 is important for promotion of M1 macrophage activity and anti-cryptococcal responses. Overall, these data emphasize that dynamic changes in histone modifications must occur upon re-stimulation with cryptococci in order to induce pro-inflammatory cytokines that are important for protection. Future studies will determine the genes targeted by specific histone modifications that occur after IFN-γ training that likely aid in the rapid response of macrophages during in vivo challenge with WT C. neoformans.
IFN-γ training of macrophages is a realistic novel therapeutic approach that selectively targets IFN-γ associated genes and leaves residual TLR and CLR functions intact for host-defense mechanisms [46]. Here, we demonstrate for the first time that macrophages can be primed in vivo with C. neoformans/IFN-γ to expediently activate the STAT1 pathway, maintain the innate memory phenotype for at least 70 days, and provide complete protection against secondary challenge. It is likely that the concurrent and sustained IFN-γ production by the genetically modified C. neoformans strain primed the macrophages to respond to Cryptococcus in a more specific manner compared to that observed during typical exposures. Other vaccine strategies including the use of genetically modified Cryptococcus strains, crude cryptococcal extracts or recombinant proteins such as mannoproteins or heat shock proteins to induce cell-mediated immune responses induced varying levels of protection against challenge [53–59]. A novel strategy using β-glucan particles as an adjuvant delivery platform packed with alkaline extracts from cryptococcal capsule mutant strains has recently shown efficacy against C. neoformans and C. gattii in a CD4+ T cell-dependent manner [60]. Antibody mediated immunity (AMI) contributes to protective responses; however, there is no conclusive evidence that protection can be achieved via AMI alone. Considering that the majority of individuals who succumb to cryptococcosis are T-cell deficient, innovations in the field that target host innate immune cells are potentially the best poised to protect against this fungal pathogen. Taken together, these results show that vaccines and/or immune therapies can be designed to induce innate cell memory-like immune responses that aid in protective responses among immunocompromised individuals who are most at risk of developing cryptococcosis and other life-threatening diseases.
All animal experiments were conducted following NIH guidelines for housing and care of laboratory animals and in accordance with protocols approved by the Institutional Animal Care and Use Committee (protocol number MU021) of the University of Texas at San Antonio. A scoring-system for assessment of animal distress was established before infection experiments were started. Based on these guidelines, general condition and behavior of the animals was controlled by well-educated and trained staff. Depending on the progress of the disease, animals were monitored twice daily during the “day-phase” (7:00 am to 7:00 pm). In order not to disturb the circadian rhythm of the animals, there was no monitoring after 7:00 pm. Humane endpoint by CO2 asphyxiation followed by cervical dislocation was conducted if death of the animals during the following hours was to be expected.
Female BALB/c (H-2d) mice (National Cancer Institute/Charles River Laboratories) and female B cell KO mice C.Cg-Cd19tm1(cre)CgnIghb/J and control BALB/cByJ mice (The Jackson Laboratory, Bar Harbor, ME) were used throughout these studies and housed at The University of Texas at San Antonio Small Animal Laboratory Vivarium and handled according to guidelines approved by the Institutional Animal Care and Use Committee.
C. neoformans strain H99, C. neoformans strain H99γ (derived from H99 serotype A, mating type α) [15], C. neoformans calcineurin mutant (cna1Δ, derived from H99, mating type α, and C. deuterogattii strain R265 (each kind gifts from Dr. Joseph Heitman), mCherry expressing C. neoformans strain (KN99mCH, a kind gift from Dr. Jennifer Lodge), C. bacillisporus strain WSA87, and C. deneoformans strain 52D (each kind gifts from Dr. Brian Wickes) were recovered from a 15% glycerol stock stored at -80°C prior to use in the experiments described in this study. The strains were maintained on yeast extract/peptone/dextrose (YPD) medium agar plates (Becton Dickinson, Sparks, MD). Yeast cells were grown for 16–18 h at 30°C with shaking in liquid YPD broth, collected by centrifugation, washed three times with sterile phosphate buffered saline (PBS), and viable yeasts were quantified using trypan blue dye exclusion on a hemacytometer. As a control for immunization studies, C. neoformans strain H99γ was grown, washed, and counted as stated above, then heat killed by boiling at 65°C for 1 hour and killing verified by plating on YPD agar (HKH99γ). For cytokine recall assays, Staphylococcus aureus strain UAMS-1 and Candida albicans SC5314 were grown in Luria-Bertani medium (Fisher Scientific, Houston, TX) at 37°C or YPD broth at 30°C, respectively, for 18 hours. The cultures were then washed, counted, and heat killed as described above.
Mice were anesthetized with 2% isoflurane using a rodent anesthesia device (Eagle Eye Anesthesia, Jacksonville, FL) then given an intranasal inoculation with 1 X 104 CFU of C. neoformans strain H99γ or heat killed C. neoformans strain H99γ in 50 μl of sterile PBS and allowed 70 days to resolve the infection. Subsequently, the immunized mice received a second experimental pulmonary inoculation with 1 × 104 CFU of wild-type C. neoformans strain H99 in 50 μl of sterile PBS. The inocula used for nasal inhalation were verified by quantitative culture on YPD agar. Mice were euthanized on predetermined days by CO2 inhalation followed by cervical dislocation, and lung tissues were excised using aseptic technique. Alternatively, mice intended for survival analysis were monitored by inspection twice daily and euthanized if they appeared to be in pain or moribund.
Mice were depleted of CD4+ and/or CD8+ T cell subsets via intraperitoneal (ip) administration of 200 μg anti-CD4 (GK1.5, rat IgG2b) and 200 μg anti-CD8α (2.43, rat IgG2b) antibodies (National Cell Culture Center) in 200 μl PBS as previously described beginning at 48 hours prior to challenge and administered weekly [11]. For neutrophil depletions each mouse received ip injections of 200 ug anti-Ly6G (1A8)(BioXcell) in 100 μl every other day. For NK cell depletion, each mouse received ip injections of 0.5 ug anti-asialo-GM (Wako USA) every 3 days. Neutrophil and NK cell depletions were started 24 hours prior to challenge. Control IgG2a and IgG2b isotype control antibodies were used (eBioscience Inc., San Diego, CA). All depletions were maintained through the entirety of the study at concentrations and schedules chosen following studies testing different dosages and schedules to determine the optimum dosage and schedule for each antibody. The efficiency of all cell depletions in the lungs and spleens was assessed by flow cytometric analysis using antibodies that adhere to epitopes distinct from those adhered to by the depletion antibodies (S1 Fig). Antibodies depleted approximately 95% of neutrophils, 87% of NK cells, 98% of CD4+ T cells, and 98% of CD8+ T cells.
Lungs were excised and digested enzymatically at 37°C for 30 min in 10 ml digestion buffer (RPMI 1640 and 1 mg/ml collagenase type IV (Sigma-Aldrich, St. Louis, MO) with intermittent (every 10 min) stomacher homogenizations. The digested tissues were then successively filtered through sterile 70- and 40-μm nylon filters (BD Biosciences, San Diego, CA) to enrich for leukocytes, and then cells were washed with sterile HBSS. Erythrocytes were lysed by incubation in NH4Cl buffer (0.859% NH4Cl, 0.1% KHCO3, 0.0372% Na2EDTA [pH 7.4]; Sigma-Aldrich) for 3 min on ice followed by a 2-fold excess of PBS. Standard methodology was employed for the direct immunofluorescence of pulmonary leukocytes as previously described [11]. Cells were incubated with CD16/CD32 (Fc Block; BD Biosciences) and fluorophore-conjugated antibodies were used to stain for cell surface staining (S3 Table). The samples were acquired on a BD FACSArray flow cytometer (BD Biosciences) and data analyzed using FlowJo Software (Ashland, OR).
A single cell suspension of pulmonary leukocytes was acquired as described above. The resulting leukocyte population was then depleted of CD3+ cells using an α-CD3 microbead kit (Miltenyi Biotec, Auburn, CA), then enriched for macrophages using biotinylated α-F4/80 and subsequent binding of anti-biotin magnetic beads (Miltenyi Biotec) according to the manufacturer’s recommendations. Splenic macrophages were isolated as described for the pulmonary macrophages, omitting the collagenase digestion. Purity was verified using Labeling Check Reagent-APC (Miltenyi Biotec) and F4/80-PE on a BD FACSArray Flow Cytometer (BD Biosciences) as described above and purity of 85–95% was routinely achieved with less than 3% T cells detected.
Reagents used were as follows: LPS (Escherichia coli O111:B4, Sigma Aldrich); GSK 343, MI-2 (hydrochloride), UNC0638 and Torin 1 (Cayman Chemical, Ann Arbor, MI); MTA (5’-Deoxy-5’-(methylthio)adenosine) and pargyline (Sigma Aldrich); EGCG ((-)-Epigallocatechin gallate) and givinostat (ITF2357) (Selleck Chemicals, Houston, TX).
Splenic macrophages (5 × 105/well) derived from mice immunized with C. neoformans strain H99γ or HKH99γ on day 70 post-immunization or naive female BALB/c mice were cultured in RPMI complete media with or without C. neoformans cna1Δ, heat killed C. neoformans H99, heat killed C. deuterogattii R265, C. bacillisporus WSA87, or heat killed C. deneoformans 52D (at a 1:1 ratio) at 37°C and 5% CO2 in 96 well round-bottom culture plates (Becton Dickinson Labware, Franklin Lakes, NJ). The culture supernatants were collected after 24 h, protease and phosphatase inhibitor cocktail (Thermo Fisher, Rockford, IL) was added and samples were stored at -20°C until analysis. Cytokine production was analyzed using the Bio-Plex protein array system (Luminex-based technology; Bio-Rad Laboratories, Hercules, CA). Alternatively, splenic macrophages were cultured in media alone, media +/- dimethyl sulfoxide (DMSO; Fisher Scientific), dimethylformamide (DMFM; Fisher Scientific), and various histone modification inhibitors or in media containing C. neoformans cna1Δ, 0.3 ug/ml LPS, heat killed S. aureus (5 × 105/well) or heat killed C. albicans (5 × 105/well) +/- histone modification inhibitors for 24 h and cytokine levels in supernatants were analyzed by Bio-Plex Protm Mouse Cytokine Th1/Th2 8-plex Assay (Bio-Rad) or IL-2 ELISA (Affymetrix). Culture of macrophages with the carriers DMSO and DMFM resulted in cytokine production similar to media alone.
For Western blotting of macrophages were stimulated as described above. After stimulation and collection of supernatants, cells were lysed in 35 μl of lysis buffer (Millipore) and biological replicates pooled. Equal amounts of protein were subjected to SDS-PAGE electrophoresis. Primary antibodies [1:500 and 1:50 000 (β-actin)] in 5% (w/v) BSA/TBST (5% bovine serum albumin/TBST) were incubated overnight at 4°C. HRP-conjugated anti-rabbit antibody at a dilution of 1:5000 in 5% (w/v) BSA/TBST was used for 1 hour at room temperature. The following antibodies were used: β-actin antibody, mTOR antibody, phospho-mTOR antibody (Ser2448), STAT1, and phosphor-STAT1 (Tyr701) antibody (Cell Signaling, Leiden, Netherlands; S3 Table). At least 3 different individual experiments were repeated for each Western blot experiment.
Total RNA was isolated from purified pulmonary F4/80+ cells using TRIzol reagent (Invitrogen, Carlsbad, CA) and then DNase (Qiagen, Germantown, MD) treated to remove possible traces of contaminating DNA according to the manufacturer’s instructions. Total RNA was subsequently recovered using the Qiagen RNeasy kit. RNA integrity and concentration was assessed via Bioanalyzer using the Agilent RNA 6000 Nano Kit according to the manufacturer’s recommendations (Agilent, Santa Clara, CA). Minimum acceptable RNA integrity number (RIN) was set at 7 for use in RNA-sequencing studies. Library preparation and sequencing were performed by UT Southwestern Medical Center Genomics and Microarray Core Facility, Dallas, TX. Resulting RNA sequences were deposited in the Sequence Read Archive, PRJNA420072.
Data normalization and differential gene expression was determined using a method that incorporates an internal-standard based approach of normalization and an associative t-test to minimize false positive determinations as previously described [61, 62] and performed by UT Southwestern Medical Center Genomics and Microarray Core Facility, Dallas TX. Genes exhibiting normalized expression values 20 times the standard deviation of the statistically defined background were considered expressed. Genes differentially expressed ≥2 fold passed the standard t-test significance level of p<0.05 and passed an associative t-test threshold to eliminate false positive determinations.
Functional pathway and network analyses of differentially expressed genes were performed using Ingenuity Pathway Analysis (IPA) (Qiagen, Redwood City, CA). The Ingenuity Knowledge Base, a repository of biological interactions, was used as a reference set. The functional analysis module in IPA was used to identify over-represented molecular and cellular functions of differentially expressed genes. The probability that each biological function assigned to the data set was due to chance alone was estimated, and a false discovery rate (FDR) <0.05 was used to correct for multiple comparisons. Over-represented canonical signaling and metabolic pathways in the input data were determined based on two parameters: (1) The ratio of the number of molecules from the focus gene set that map to a given pathway divided by the total number of molecules that map to the canonical pathway, and (2) a P-value calculated by Fisher’s exact test that determines the probability that the association between the focus loci and the canonical pathway is explained by chance alone. Network analysis used focus genes as “seeds” to infer de novo interaction networks. Direct interactions between focus loci and other molecules were inferred based on experimentally observed relationships supported by at least one reference from the literature. Additional molecules from the Ingenuity Knowledge Base were added to the network to fill or join smaller networks. The network score was based on the hypergeometric distribution and calculated with the right-tailed Fisher’s exact test. A higher score indicates a lower probability of finding the observed number of focus molecules in a given network by chance.
Gene ontology analysis was performed using the DAVID functional analysis tool [63]. The Bonferroni, Benjamini, and FDR (false discovery rate) were used for multiple test correction.
Pulmonary F4/80+ cells (5 x 106) isolated from immunized mice were crosslinked, lysed, and sonicated with the Bioruptor ultrasonicator (Diagenode, Denville, NJ). 10% of chromatin was reserved as input control. The remaining chromatin was immunoprecipitated with antibodies against Normal IgG or STAT1 (Cell Signaling Technology, Danvers, MA) bound to Protein A/G Magnetic Beads (Pierce Biotechnology, Rockford, IL). Immunoprecipitated DNA and Input DNA was reverse crosslinked, eluted, and purified with the IPure kit v2 (Diagenode) according to manufacturer’s instructions.
Immunoprecipitated DNA was used as a template for real time ChIP qPCR analysis using EpiTect ChIP Custom qPCR Arrays (Qiagen). Genes were chosen based on RNA-seq day one post-challenge IPA analysis. Primer sequences for regions 1kb upstream of the transcription start site of IRF-1, iNOS, GBP2, GBP5, GBP6, CXCL9, CXCL10, CXCL11, SOCS-1, CIITA, IFIT2, and IFI47 were used in the array. A master mix consisting of 2ul DNA per 25ul reaction was mixed with RT2 SYBR Green qPCR master mix (Qiagen) according to manufacturer’s recommendations and Real Time ChIP qPCR arrays were performed using 7300 real-time PCR System (Applied Biosystems, Foster City, CA). Percent IP was calculated based on comparative threshold values using the SuperArray ChIP-qPCR Data Analysis Template supplied by Qiagen according to manufacturer’s instructions.
Survival data was analyzed using the log-rank test (GraphPad Software). The unpaired Student’s t test was used to analyze comparisons between two groups to detect statistically significant differences. For multiple comparisons, a one-way ANOVA or two-way ANOVA with the Tukey’s multiple comparison test was performed. Significant differences were defined as *p < 0.05, **p < 0.01 or ***p < 0.001.
|
10.1371/journal.pntd.0005715 | Post-dengue acute disseminated encephalomyelitis: A case report and meta-analysis | Dengue is one of the most common infectious diseases. The aim of this study was to systematically review acute disseminated encephalomyelitis (ADEM) and to represent a new case.
We searched for articles in nine databases for case reports, series or previous reviews reporting ADEM cases in human. We used Fisher’s exact and Mann-Whitney U tests. Classification trees were used to find the predictors of the disease outcomes. We combined findings using fixed- and random-effects models. A 13-year-old girl was admitted to the hospital due to fever. She has a urinary retention. The neurological examinations revealed that she became lethargic and quadriplegic. She had upper limbs weakness and lower limbs complete paraplegia. Her status gradually improved after the treatment. She was nearly intact with the proximal part of her legs had a mild weakness in discharge. The prevalence of ADEM among dengue patients was 0.4% [95% confidence intervals (95% CI) 0.1–2.5%], all neurological disorders among dengue was 2.6% [95% CI 1.8–3.8%], and ADEM among neurological disorders was 6.8% [95% CI 3.4–13%]. The most frequent manifestation of ADEM was altered sensorium/consciousness (58%), seizures and urination problems (35%), vision problems (31%), slurred speech (23%), walk problems (15%) then ataxia (12%). There was a significant difference between cases having complete recovery or bad outcomes in the onset day of neurological manifestations being earlier and in temperature being higher in cases having bad outcomes (p-value < 0.05). This was confirmed by classification trees which included these two variables.
The prevalence of ADEM among dengue and other dengue-related neurological disorders is not too rare. The high fever of ADEM cases at admission and earlier onset day of neurological manifestations are associated with the bad outcomes.
| We presented a 13-year-old girl of ADEM following dengue infection. She was totally alert and had a low grade of fever with no focal neurologic deficits, on admission. We revealed that the prevalence of either ADEM or all neurological disorders among dengue patients was not too rare. Moreover, we found that the most common manifestation of ADEM was altered sensorium/consciousness followed by seizures and urination problems then vision problems. These manifestations should be considered in the diagnosis and management of dengue-infected patients. Also, this requires shedding the light on the total global cases of ADEM from the annual incidence of dengue. The onset of ADEM can be early or late after dengue infection. Hence, clinicians should pay attention that it can be early or late that patients can forget about their fevers. Moreover, the onset day of neurological manifestations and patients’ temperature were significantly associated with the disease outcome.
| Dengue, a worldwide prevalent mosquito-borne infectious disease, is a flavivirus spread by several species of Aedes type mosquitos, mainly Aedes aegypti [1]. Dengue has become a dangerous burden and is widely spread in more than 110 countries [2, 3]. The incidence of dengue has increased to reach 30-fold throughout the past 50 years [4]. Annually, between 50 and 528 million people have the infection and about 10,000 to 20,000 deaths [5–8]. Dengue has a wide variety of manifestations, from fever to dengue shock syndrome and/or multiple organs failures [1, 9]. There are a series of biological predictors such as immune cytokines [10–12], circulating DNA [13], microalbuminuria [14], nonstructural protein 1 [15–17], IgM, IgG [18], IgA [19] and endothelial cell damage, as well as dysfunction predictors, have been evaluated [20]. However, no efficient marker for the prediction of severe dengue infection has been discovered [20–22]. Although neurological problems of dengue virus (DENV) have also been reported, the incidence of this group is uncommon between 0.5 and 6.2% [23]. A previous systematic review has investigated the factors associated with DENV and revealed that these factors included the neurological signs [24]. DENV associated neurological problems can be divided into DENV direct invasion and para- or post-infectious disease [3]. These neurological DENV include encephalopathy, encephalitis, immune-mediated syndromes as acute disseminated encephalomyelitis (ADEM) and Guillain-Barré syndrome (GBS), neuromuscular complications as hypokalemic paralysis and dengue-associated stroke [25–30]. ADEM in dengue is very rare and it may occur during the acute phase or post-infectious phase of dengue. It is known to involve an immune-mediated mechanism in which the cytokine overproduction is triggered by DENV [25]. There was another theory which is the immune-mediated attack by autoantibodies and/or T-cells to central nervous system myelin structure. This leads to acute demyelination of the white matter of the brain, spinal cord or both. Thus, it results in an altered mental status and focal neurologic findings in ADEM patient such as paralysis [3]. Although ADEM causes a significant impact on dengue patients, data about this complication is still lacking. Understanding of this complication provides a potential insight into the clinical picture of DENV infection. Thus, this study aimed to conduct an extensive systematic review and meta-analysis of the literature on the ADEM manifestations in dengue with a new case report.
All the methods were performed in accordance with the relevant approved guidelines, regulations and declaration of Helsinki. The experimental protocols were approved by Children’s Hospital No.2 in Ho Chi Minh City in Vietnam. Written informed consent was obtained from the parents to have their girl’s details and accompanying images published and approved by the aforementioned hospital. Moreover, all patient’s data was analyzed anonymously.
This systematic review was performed according to the Preferred Reporting Items for Systematic Review and Meta-analyses statement (PRISMA) (S1 Table) [31]. We had developed and registered a protocol of methods (CRD42016047583). From inception to the 12th of September 2016, we searched for suitable studies in nine databases including; PubMed, Google Scholar, Institute of Science Index (Web of Science), Scopus, Popline, World Health Organization Global Health Library, Virtual Health Library, New York Academy of Medicine Grey Literature Report, System for Information on Grey Literature in Europe and cross-references from the included articles and previous reviews. The search strategy used was (ADEM or encephalomyelitis) and dengue. Three independent reviewers initially scanned primary titles and abstracts (when available) to select potential full-text articles for further scrutiny according to the inclusion and exclusion criteria. The inclusion criteria were as following; case reports, case series, previous literature reviews or systematic reviews discussing post-infectious immune-mediated ADEM of dengue infection in human. Exclusion criteria were as following; other study designs rather than case reports, case series, previous literature reviews or systematic reviews, other complications rather than ADEM, overlapped data sets, data which could not be extracted, duplicated studies and unreliable or incomplete data, no full-text available, abstract-only articles (conference, letters, commentaries), or thesis, books, review editorial or author response. When the title and abstract were not rejected by any reviewer, the full-text of the article was obtained and carefully reviewed for inclusion by the three reviewers. Inclusion or exclusion of each study was determined by discussion and consensus between the three reviewers. When the disagreement occurred, a consensus decision was made following discussion with a senior reviewer.
Data were extracted by three authors and were checked by at least another author. The disagreement was resolved via discussion and a consensus reached between the three authors. The data extraction form in an Excel file was developed by two authors based on a pilot review and extraction. The data extracted included the first author, year of publication, year of patient recruitment, study design, country of origin and characteristics of the population (infant, children, adult), gender, age at examination of included individuals, the manifestations, the blood and CSF analyses, medications used, visual and neurological examinations, renal and liver function tests and outcome of each patient. If there were more than one value from the examination, the nadir value, e.g. the lowest platelets, the highest packed cell volume (PCV), was extracted. Papers published by the same research group and studying the same factor were checked for potentially duplicated data based on the year of patient recruitment and hospital where the patients were recruited. When duplications were noted, the largest data set was used for our study.
Fisher’s exact and Mann-Whitney U tests were used for categorical and continuous variables, respectively. The values with different units were converted into one common unit to make the values comparable. The classification tree models were used to find the independent predictors that best predict bad outcomes (partial recovery or death) versus complete recovery as well as complete recovery versus partial recovery [32]. In particularly, 20 variables including age, sex, clinical examination’s variables and the steroid treatment and its administration route were included to build classification tree models. We selected the maximum depth of the tree to be five to construct a tree of reasonable complexity. If the tree is too complex, it is difficult to apply. Likewise, we chose the minimum number of observation at each leaf node is equal to five to prevent the tree from sub-dividing into overly specific nodes that contain little supporting data. The performance measures of the tree were accuracy (1 –misclassification error) and its 95% confidence intervals (95% CI), sensitivity, specificity, positive predictive (PPV) and negative predictive values (NPV). The statistical significance was considered when the p-value was < 0.05. Data were analyzed using SPSS version 23.0, and R software version 3.3.2.
Meta-analyses were performed using Comprehensive Meta-analysis (CMA) software version 3 (Biostat, NJ, USA) when there was more than one study. Dichotomous variables were analyzed to compute pooled event rate (ER). A fixed-effect model [33] was used when there is no evidence of a heterogeneity between studies, otherwise, a random-effects model was chosen [34]. Heterogeneity between studies was evaluated using the Q statistic and I2 test which describes the percentage of variability in the effect estimates that is because of heterogeneity beyond sampling error [34, 35]. To evaluate the presence of publication bias, we performed Begg’s funnel plot [36] and Egger’s regression test [37, 38] when there were five or more studies in the analysis. The publication bias was considered significant when the p-value was < 0.1. If the publication bias was found, the trim and fill method of Duvall and Tweedie was performed by adding studies that appeared to be missing [39, 40] to enhance the symmetry [41]. The adjusted pooled effect size and its 95% CI were computed after the addition of potential missing studies.
A 13-year-old girl was admitted to Children’s Hospital No.2 in Ho Chi Minh City in Vietnam on 31st August 2016 because of fever for seven days. On admission, she was totally alert and had a low grade of fever. The examination found no focal neurologic deficits. Her total blood count showed that white blood cell (WBC) and platelet counts were 11,000/μL and 182,000/μL, respectively. C-reactive protein level was 3 mg/L. On the 8th day (the day after admission) of her disease course, she recovered from fever but first began to complain of no passage of urine. She was found to lose the sensation of urinating and have urinary retention. She then needed insertion of the indwelling urinary catheter. On the 9th day (the 2nd day of admission), the serologic test revealed that serum dengue IgM was positive. Serum PCR dengue was negative. Laboratory tests of liver and renal functions and electrolytes did not show any particular abnormality. On the same day, she could not move her legs, began to lose her consciousness and showed signs of confusion. The neurological examination then found that she became lethargic and quadriplegic with no abnormal sign of cranial nerves. She had normal ocular fundus, opened her eyes with painful stimulation, answered her name and then drifted back to sleep. Muscle strength and tendon reflexes of the upper extremities were 2/5 and 2+, respectively, with the weakness of upper limbs (UL). The sensory functions were nearly intact. The muscle strength and tendon reflexes of the lower extremities were 0/5 and 2+, respectively, with complete paraplegia of lower limbs (LL). Bilateral (B/L) Hoffmann and Babinski tests were positive. Cerebrospinal fluid (CSF) collected on 10th day showed pleocytosis (61 cells/mL); elevation of protein, 1.68 g/L; glucose, 0.43 g/L; chloride, 138 mmol/L; lactate 4.23, mmol/L. Her CSF was positive for ELISA dengue IgM but negative for ELISA Japanese encephalitis virus IgM. PCR Zika virus in blood and in urine was negative. Magnetic resonance imaging (MRI) of the brain and spinal cord did not show any particular abnormality (Fig 1 and S1–S3 Figs). An electromyogram showed that motor and sensory functions were normal on both UL and LL (S4 and S5 Figs).
From her clinical course and laboratory tests, she was diagnosed as ADEM following dengue infection without warning signs. For treatment of ADEM, high dose of methylprednisolone (30mg/kg/day) for five days was given, beginning on the 3rd of September. The oral low dose of prednisone (1 mg/kg/day) was then used for four weeks. Her alertness improved gradually. On the fourth day of methylprednisolone course, she opened her eyes responding to voice, oriented and answered word by word correctly. Although urinary retention remained, her muscle strength of upper and lower extremities increased to 4/5 and 2/5, respectively. After two weeks of oral prednisolone, limbs weakness was significantly improved and after four weeks, sphincter function was back to normal. She was nearly intact with the proximal part of her legs had a mild weakness when she was discharged from the hospital after four weeks of admission (five weeks since fever onset).
From nine databases, we identified 690 potentially relevant publications. After excluding duplicates and screening titles and abstracts, we retrieved 34 articles for full-text review. Of these, 25 articles met our inclusion criteria. Four additional articles, from manual search, were identified. Finally, 29 articles were in this systematic review including; 15 case reports, 10 case series, 1 literature review with 1 case, and 3 literature reviews (Fig 2).
The total sample size was 1,163 dengue patients including 165 patients with neurological complications. Of those 165 patients, there were 29 ADEM cases including three cases of rare types from ADEM; one case with neuromyelitis optica [42], two cases with meningoencephalitis [43, 44], and our new case. The characteristics of the 29 cases are shown in Table 1. Among 29 ADEM patients, there were three cases were described in a group of neurological manifestations with limited individual data [43, 45]. There were more males than females (18 and 8, respectively), the median age was 20 (range 9 days to 65 years). Most of the dengue cases were diagnosed based on IgM followed by IgG.
The three included literature reviews, in general, discussed the pathogenesis of DENV and its accompanying neurological complications, their pathogenesis, and their incidence. The first review discussed the neuropathogenesis of DENV illness, its neurological complications, the diagnosis, and treatment of these diseases. It discussed also the epidemiology of DENV and the increasing prevalence and incidence of the disease and its extension to new countries. The neuropathogenesis of DENV includes three ways; the metabolic disturbance causing encephalopathy, direct central nervous system invasion (especially, by DENV-2 and -3) causing mainly encephalitis, and autoimmune reaction mechanism. The neurological complications discussed included encephalitis and meningitis, being the most common complication and caused by direct invasion, ADEM, and its rare type neuromyelitis optica, by an immune-mediated process, myelitis, either by an immune-mediated mechanism or by direct invasion, GBS, and mononeuropathies, by autoimmune mechanisms, and myositis [46].
While the second one discussed the various neurological complications, their diagnosis, and the treatment. The neurological complications included dengue encephalopathy, describing it as the most commonly reported neurological disorder associated with DENV and stating that in a retrospective study in Indonesia, 6% (152) of patients with DHF were admitted with encephalopathy. Encephalitis was described in five studies. Post-dengue immune-mediated diseases were discussed and included acute transverse myelitis, GBS, ADEM, and its rare type neuromyelitis optica. Also, cerebrovascular complications (with unknown incidence) and dengue muscle dysfunction (ranging from 66 to 100% in different studies) and neuro-ophthalmic complications (about 10 to 40% in different studies) were described [47].
Finally, the last one discussed the epidemiology, transmission of DENV, its clinical manifestations, its neurological complications and their pathogenesis, the diagnosis, and management of the disease. The pathogenesis of the neurological complications included immune-mediated reactions, metabolic disturbance, and direct invasion. This study described encephalopathy as the most common neurological manifestation of DENV infection. Other neurological complications were encephalitis, myelitis, GBS, myositis, and hypokalemic paralysis. However, ADEM was described as a rare complication [48].
The prevalence of neurological disorders (n = 27) among dengue patients (n = 1,024) in two studies was 2.6% [1.8–3.8%] without an evidence of heterogeneity (Fig 3A).
Pooling two studies enrolling all available dengue patients (n = 1,024) revealed that the prevalence of ADEM (n = 3) among dengue patients was 0.4% [0.1–2.5%] with a moderate heterogeneity (Fig 3B).
In seven studies recruiting dengue patients (n = 144) with neurological disorders (n = 8), the prevalence of ADEM was 6.8% [3.4–13%] without an evidence of heterogeneity nor publication bias, Egger’s test (p-value = 0.8) (Fig 3C).
We could only analyze manifestations in 26 cases due to the lack of information in three cases [43, 45].
The onset day of neurological manifestations after initial dengue symptoms ranged from day 3 to day 19 (median = 7). Most of the cases had a fever on the ADEM onset (25 cases) and 4 cases were not described (ND). The reasons for ADEM admission were fever, vomiting, urination problems, arthralgia, seizures, walk problems, chills, altered sensorium, asthenia, hemiparesis, vision problems, paraparesis, paresthesia and hyperreflexia, abdominal pain, thrombocytopenia, lethargy, poor feeding and seizures, weakness, headache, rigors, and/or myalgia.
The most frequent manifestations and signs related to dengue were fever (22/26, 85%), thrombocytopenia and vomiting (13/26, 50%), headache (11/26, 42%), erythema/rash (9/26, 35%), myalgia (8/26, 31%), arthralgia (6/26, 23%), chills (5/26, 19%), leukocytopenia and restless (4/26, 15%) then retro-orbital pain, rigors and lethargy (3/26, 12%). While the most frequent manifestations and signs related to ADEM were altered sensorium/consciousness (15/26, 58%), seizures and urination problems (9/26, 35%), vision problems (8/26, 31%), slurred speech (6/26, 23%), walk problems (4/26, 15%) then ataxia (3/26, 12%) (Fig 4 and S2 Table).
Liver function tests were normal in 5 cases, abnormal then normal in 1 case, abnormal in 9 cases and ND in 14 cases. The renal function tests were normal in 3 cases, 1 case had urea: 46 mg/dl and creatinine 1.1 mg/dl, 1 case had acute kidney injury and metabolic acidosis and ND in 23 cases. Urinalysis was normal in 3 cases and ND in 26 cases. Chest-X rays were normal in 4 cases, suggestive of acute respiratory distress syndrome (ARDS) in 1 case and B/L fluffy shadows as well as ARDS in another case and ND in 23 cases. The cardiovascular system examination was normal in two cases and ND in 27 cases. The pulses per min were normal in most cases, median (range) was 110 per minute (92–140), stable in 2 cases but ND in 9 cases. The abdominal findings were normal in 1 case, distended urinary bladder, splenomegaly 2 cm and hepatomegaly 3 cm in another case, markedly distended and palpable urinary bladder and spleen tip palpable, 1 case had pain, mild hepatomegaly in another case and ND in 24 cases. The arterial blood pressure (ABP) was normal in most cases, median (range) of systolic blood pressure was 100 mmHg (90–130) while of diastolic blood pressure was 60 mmHg (60–80) and ND in 19 cases. The tourniquet test was positive in 3 cases and ND in 26 cases. The respiratory rate was 28/min in 1 case, 30/min in 1 case and stable in another case and ND in 26 cases (S3 Table).
The results from CSF analysis showed that most of the cases have elevated protein levels (13 cases), normal glucose (11 cases), pleocytosis (6 cases), positive for ELISA dengue IgM in 2 cases and ND in 12 cases. The results from MRI of brain and spinal cord showed that most of the cases have abnormalities such as T2 lesions, demyelination 2 cases and B/L hemorrhagic demyelination in 1 case, cervical/proximal dorsal cord edema deep white matter, cortical and Pontine swellings in 1 case, cervical/proximal dorsal cord edema in 1 case, cervicodorsal cord swelling in another case and no description for MRI of spinal cord in 20 cases while in 7 cases only for MRI of brain (S4 Table).
The power grade ranged from 1/5 to 5/5. Moreover, deep tendon reflexes (DTRs) were normal (3 cases), increased/hyperreflexia (3 cases) or brisk (3 cases), absent/hyporeflexia (2 cases) and ND (18 cases). Plantar responses were B/L extensor (9 cases), B/L Babinski sign (1 case), flexor (1 case), B/L Hoffmann and Babinski signs (1 case), left Babinski sign and absent response in right foot (1 case) nonresponsive B/L (1 case) and ND in 15 cases. Cranial nerves were normal (6 cases), unable to be examined (3 cases) facial nerve palsy (1 case), B/L ptosis (1 case) and ND (18 cases). Furthermore, the ULs and LLs were abnormal in most cases, normal in 1 case only and ND in 11 cases for LLs and cases for ULs 13 cases. The motor system was ND in 25 cases and has abnormalities (hemiparesis or quadriparesis) in the remained ones (S5 Table).
The blood analysis showed an increased WBC in most cases, median (range) of WBC = 45 × 108/L (0.011–31×109), while of platelet was 60 × 109 (1.1 × 109–328 × 109), of hemoglobin (Hb) was 107000 mg/L (112–1620000), of Glasgow coma scale (GCS) was 7 (6–9), of PCV was 38.1% (30.1–48.4), of alanine transaminase (ALT) was 214 U/L (36.3–123000), of aspartate transaminase (AST) was 185 U/L (44–199000), of albumin was 3.05 g/dL (2.3–4.2), of creatine was 1.2 mg/dL (0.6–4.3), of glucose was 103.8 mg/dL (69–138), of urea was 46 mg/dL (21.9–102) (S6 Table).
The optic nerves were unable to be examined (3 cases), normal (3 cases), B/L ptosis (1 case), optic neuritis (1 case), B/L involvement (1 case), moderate B/L optic atrophy (1 case) and ND (19 cases). The fundus examination was normal in 9 cases, showed pallor of optic discs in 1 case, B/L papilledema which was more severe in the right eye in another case and ND in 18 cases. Pupils have normal size and reaction to light in 5 cases, sluggish reaction to light in 1 case, mid-dilated and equal in size with sluggish reaction to light in 1 case, mid-dilated, symmetrical with sluggish reaction to light in 1 case, dilated in 1 case, B/L mid-dilated, symmetrical and sluggishly reacting to light in another case and ND in 19 cases. The visual acuity was deteriorated in left eye then in right eye then in both eyes gradually deteriorated in 1 case, another case had a severe visual impairment in right eye (only light perception) and a slight visual disturbance in left eye (VA = 20/25), 1 case with a severe right visual impairment and ND in 26 cases (S7 Table).
The follow-up period ranged from 28 days to 5 years (S2 Table). Most specific treatments used for ADEM were oral or intravenous (IV) corticosteroids including methylprednisolone (11 cases), prednisolone (7 cases) and dexamethasone (5 cases) or human immunoglobulin (1 case). Other treatments used were anticonvulsant medications such as phenytoin (2 cases) and phenobarbitone (1 case), oral or intravenous antipyretics and anticonvulsant such as dipyrone, paracetamol, pulse therapy, dopamine, noradrenaline, and lorazepam (1 case for each treatment). The outcomes in these cases were either death (3 cases), partial recovery (7 cases), complete recovery (16 cases) or ND (3 cases). The cases with partial recovery were either; had mild B/L visual disturbance, dysuria, and dyschezia remained [49], was able to walk with a minimal support [50], wanted to carry further treatment in the hospital [51], a slight residual cerebellar ataxia [52], the frontal symptoms persisted [53], mild ataxia and dysarthria [54]. The three cases died due to; myalgia, jaundice, conjunctival hemorrhage, hematuria, oliguria, shortness of breath, became stuporous, acute respiratory distress syndrome (ARDS), acute kidney injury and metabolic acidosis [55], intracranial tension [56] or B/L hemorrhagic demyelination [57] (Table 2).
Our results showed that the body temperature levels in the complete recovery group were significantly lower than those of the partial recovery (p-value = 0.026) and bad outcomes groups (p = 0.03). While there was a significant difference between cases having complete recovery or bad outcomes on the onset day of neurological manifestations which was found started earlier in cases having partial recovery (p = 0.03) and bad outcomes (p = 0.006) as compared to patients with complete recovery. Other factors including gender, steroid treatment, and its administration route, age, hemoglobin, platelet, WBC, GCS, PCV, ALT, AST, pulse, systolic blood pressure, diastolic blood pressure, urea, glucose, creatinine or albumin were not associated with the ADEM outcomes (Table 3).
We then selected all of the aforementioned 20 variables to build classification tree models for bad outcomes (partial recovery or died) versus complete recovery and for partial recovery versus complete recovery. Interestingly, that both classification trees including the onset day of neurological manifestations (the cut-off point at 9.5 days), and temperature (the cut-off point at 101.2°F) are the best models (Fig 5). Both trees enhance the results from uni-variable analysis indicating that the earlier onset day of the neurological manifestations (< 9.5 days) and higher fever when presenting ADEM (≥ 101.2°F) were associated with the bad outcomes and partial recovery. The performance of tree that classified bad outcomes versus complete recovery is at the accuracy of 84.6% [65.1–95.6%] with the sensitivity of 80%, specificity of 87.5%, PPV of 80% and NPV of 87.5%. Whereas, the performance of tree that classified complete recovery versus partial recovery is at the accuracy of 82.6% [61.2–95.1%] with the sensitivity of 87.5%, specificity of 71.4%, PPV of 87.5% and NPV of 71.4% (Fig 5).
Our meta-analysis revealed that the prevalence of patients with neurological disorders within dengue patients is not rare (2.6%). Though ADEM was reportedly stated as a rare condition [57–59], the incidence could be higher because of the high global burden of dengue infection. In a previous study, it revealed that dengue was present in 4–47% of patients with encephalitis in the endemic regions [60]. It is well-known that encephalopathy is the most common neurological disorder accompanying DENV infection [47, 48]. Thus, this should push health cares to estimate the total global cases of ADEM from the annual incidence of dengue because the number of post-dengue ADEM is underestimated probably due to the neglect of the clinicians and patients, hence our current proposal is to screen all patients with neurological manifestations against dengue and most other flaviviruses such as Zika to investigate the number of ADEM cases in a multi-center study in Vietnam and Philippines. Moreover, we suggest adding such neurological complications in the dengue WHO guidelines, so they get no neglect.
The onset of ADEM ranged from day 3 to day 19 from dengue infection. The clinicians should be aware that patients can present with early or late onset of ADEM symptoms and patients may not always mention a recent history of fever. Moreover, there was a significant difference between cases having complete recovery or bad outcomes only in two factors which were the onset day of the neurological manifestations being earlier and the temperature being higher in cases having bad outcomes or partial recovery through our uni-variable analysis. This finding is supported by classification tree models including the onset day of the neurological manifestations and temperature. Both trees indicate that the earlier onset day of the neurological manifestations (< 9.5 days) and higher fever when presenting ADEM (≥ 101.2°F) were associated with the bad outcomes. These findings require an attention from physicians regarding the temperature of the dengue cases to be managed well once elevated.
The most frequent manifestations related to dengue infection arranged from the most frequent to the least frequent were; fever, thrombocytopenia, and vomiting, headache, erythema/rash, myalgia, arthralgia, chills, leukocytopenia, and restlessness then retro-orbital pain, rigors, and lethargy. Noteworthy, vomiting, rash, and leukocytopenia are classified as dengue without warning signs in the WHO 2009 guidelines while thrombocytopenia, restlessness, and lethargy are classified as dengue warning signs [1]. While the main manifestations related to ADEM arranged from the most frequent to the least frequent were; altered sensorium/consciousness, seizures and urination problems, vision problems, slurred speech, walk problems, then ataxia. Similarly, altered consciousness is classified within the severe dengue signs, in the severe CNS involvement, in the WHO 2009 guidelines. However, other ADEM manifestations are not mentioned in it [1], maybe because ADEM sometimes appears late after dengue. Hence, we suggest adding them to the guidelines of severe organs involvement stage.
The results from MRI of the brain and spinal cord showed that most of the cases have abnormalities such as T2 lesions. In contrast to our results which showed no abnormalities. The attention for the MRI of either the brain or spinal cord findings should be paid more due to its unlimited importance in diagnosis and treatment of ADEM cases. Most of the outcomes in these cases were relatively good because most of them showed either partial recovery or complete recovery. There was no significant difference between cases with bad outcomes or complete recovery in the treatment used. Unlike a previous literature review which suggested that steroids are promising in the treatment of ADEM during its active phase [61].
Till now, there is no study described the mechanism of post-dengue ADEM. The neurological complications of dengue infection have been considered to be due to systemic complications of dengue and not related to its neurotropic nature [48, 61–65]. After the demonstration of neural tropism of dengue virus, the neurological manifestations of dengue infection are categorized as (1) related to direct neurotropic effects of the virus (myelitis, meningitis, myositis, rhabdomyolysis, and encephalitis), (2) related to systemic or metabolic complications of dengue (encephalopathy, stroke) and (3) post-infectious immune-mediated complications (GBS, transverse myelitis, ADEM). ADEM usually occurs following a viral infection but may appear spontaneously, after bacterial, parasitic infection or vaccination. Most cases follow a nonspecific upper respiratory tract infection. Although it occurs in all ages, most reported cases are in infants and adolescents [66]. The post-infectious ADEM usually begins late in the course of viral infections including measles, chickenpox, mumps, rubella, influenza, EBV and nonspecific respiratory infections. The pathophysiology involves a transient auto-immune response directed at myelin or other self-antigens, possibly by a non-specific activation of auto-reactive T-cell clones or by molecular mimicry [63, 67, 68]. As with other viruses, the pathogenesis underlying dengue-associated ADEM may result from an immune system mediated-process [25].
A limitation of this study was the small number of included studies (a total of 29 ADEM cases) and reported cases, with some missing values, included in the uni-variable analysis, meta-analysis, and the classification tree models. Moreover, our results should be interpreted with caution because most cases depended on IgM ELISA which has a probable diagnosis [1] but with a high specificity [1, 69–74]. In conclusion, our analysis of the case report and other included cases revealed that the onset day of neurological manifestations and temperature in the ADEM patients were associated with the disease outcome and can predict it. Moreover, we found that the most frequent dengue manifestations were fever, thrombocytopenia, vomiting, and headache while the most frequent ADEM manifestations were altered sensorium/consciousness, seizures urination problems, and vision problems. The serious manifestations after dengue infection continue to be reported. These manifestations should be considered in the diagnosis and management of patients with dengue infection. The prevalence of ADEM among dengue and other dengue-related neurological disorders is not too rare. Since the incidence of ADEM is not known well, future larger studies are necessary to accurately investigate ADEM.
|
10.1371/journal.pntd.0004792 | Over Six Thousand Trypanosoma cruzi Strains Classified into Discrete Typing Units (DTUs): Attempt at an Inventory | Trypanosoma cruzi, the causative agent of Chagas disease, presents wide genetic diversity. Currently, six discrete typing units (DTUs), named TcI to TcVI, and a seventh one called TcBat are used for strain typing. Beyond the debate concerning this classification, this systematic review has attempted to provide an inventory by compiling the results of 137 articles that have used it. A total of 6,343 DTU identifications were analyzed according to the geographical and host origins. Ninety-one percent of the data available is linked to South America. This sample, although not free of potential bias, nevertheless provides today’s picture of T. cruzi genetic diversity that is closest to reality. DTUs were genotyped from 158 species, including 42 vector species. Remarkably, TcI predominated in the overall sample (around 60%), in both sylvatic and domestic cycles. This DTU known to present a high genetic diversity, is very widely distributed geographically, compatible with a long-term evolution. The marsupial is thought to be its most ancestral host and the Gran Chaco region the place of its putative origin. TcII was rarely sampled (9.6%), absent, or extremely rare in North and Central America, and more frequently identified in domestic cycles than in sylvatic cycles. It has a low genetic diversity and has probably found refuge in some mammal species. It is thought to originate in the south-Amazon area. TcIII and TcIV were also rarely sampled. They showed substantial genetic diversity and are thought to be composed of possible polyphyletic subgroups. Even if they are mostly associated with sylvatic transmission cycles, a total of 150 human infections with these DTUs have been reported. TcV and TcVI are clearly associated with domestic transmission cycles. Less than 10% of these DTUs were identified together in sylvatic hosts. They are thought to originate in the Gran Chaco region, where they are predominant and where putative parents exist (TcII and TcIII). Trends in host-DTU specificities exist, but generally it seems that the complexity of the cycles and the participation of numerous vectors and mammal hosts in a shared area, maintains DTU diversity.
| Trypanosoma cruzi, the causative agent of Chagas disease, has been classified into six genetic groups (discrete typing units, DTUs) named TcI-TcVI and a seventh one called TcBat. Currently, several genetic molecular markers are used to classify the strains after their isolation in culture or directly from biological samples. The current inventory compiling the published works aiming to identify the DTUs of T. cruzi strains accumulated a total of 6,343 identifications. Although this inventory is not free of sampling bias, like all samples, it is the largest sampling to date and hence likely represents the closest picture of the current diversity of T. cruzi strains (i) circulating throughout the endemic area from the southern United States to Argentina and (ii) circulating in vectors as well as in wild and domestic mammals, and humans. Data analysis helps identify trends and provides a basis for further comparisons of new data, in a context where human factors (migration, vector control, urbanization, deforestation, agricultural expansion, resource exploitation) influence the epidemiological patterns of Chagas disease.
| Trypanosoma cruzi is a pathogenic microorganism, the causative agent of Chagas disease, characterized by high genetic and phenotypic intraspecific diversity. Population genetics suggests that clonality is an important mode of propagation of the natural populations of T. cruzi [1], although, likely sexual reproduction [2, 3] and recombination events occur to some extent and are important mechanisms that generate genetic diversity within the taxon, as discussed in a recent review [4].
The consensual nomenclature recognizes six discrete typing units (DTUs) named TcI to TcVI and a recently proposed seventh, Tcbat [5–7]. This classification is widely used as a reference in epidemiological studies. However, there is not consensus on the best method to identify the different DTUs. Similarly, the evolutionary relationships between the DTUs and therefore the evolutionary history of T. cruzi continue to be researched [8]. Several mechanisms of evolution have been recognized such as clonality, hybridization, and conventional and nonconventional genetic exchanges. In addition, several studies have demonstrated the extraordinary plasticity of the T. cruzi genome. The evolutive relationships among these DTUs has not been fully elucidated, but two of them (TcV and TcVI) clearly have a hybrid origin with TcII and TcIII as putative parents [9] according to the authors, TcIII and TcIV could also originate from a hybrid between TcI and TcII [10, 11] but some claim that is not the case [12, 13]. TcI and TcII remain two pure lines that are evolving separately from a common ancestor dating from approximately 1–3 million years ago [11, 13].
The main properties of the different DTUs have been reported previously [3, 5, 14, 15]. Briefly, (i) TcI has a wide distribution, from the southern United States to northern Argentina and Chile; this DTU is the most frequently sampled in sylvatic cycles, but it is also frequent in domestic cycles and it is the dominant DTU responsible for the transmission of Chagas disease in endemic countries located north of the Amazon basin; (ii) studies show that TcII, V and VI are more likely to be associated with domestic cycles and patients with chronic Chagas disease in the Southern Cone countries and Bolivia; (iii) TcIII and IV are mainly sampled in rainforest sylvatic cycles; (iv) Tcbat previously identified in bats, has recently been found in humans [7, 16–18]. It is well known that various DTUs can coexist in the same vector and in a single host [19–21].
The different DTUs present substantial genetic diversity. Various reports have shown that the parasite’s genetic diversity has a profound impact on its epidemiological, biological and medical characteristics [22]. Consequently, it is indispensable to characterize the genotypes that are circulating in space and in hosts. Moreover, the tracking of the different genotypes is of great interest in eco-epidemiology, providing a better understanding of epidemiological systems.
After the biogeographic overview of T. cruzi DTUs by Miles and his colleagues [23], no other exhaustive review has been done, while very numerous new genotyping studies using new genetic markers and additional parasite strains have been conducted. Although we are conducting studies on the limits of DTUs classifications of T. cruzi strains and their actual existence as genetically separated units, it seemed important to take all existing data that refer to the current classification and to examine the geographic properties and host specificities of the different DTUs.
Data were obtained from a total of 137 articles (including our own published results) selected after searching PubMed (http://www.ncbi.nlm.nih.gov/pubmed) with “DTU”, “genetic characterization”, “lineage”, “genotype”, “isozyme”, “isoenzyme”, and “Trypanosoma cruzi” as key words. This research, as exhaustive as possible, was updated to April 27, 2016. Research has also been conducted by authors having worked on the genetic characterization of T. cruzi strains. For our published data, additional data, not present in the publications, was included in the current inventory because this information was available from our own records. For example, the names and data concerning the strain origins analyzed in Barnabé et al. [24] were added here. The publications included in the inventory used genetic markers that allowed DTU typing according to the consensual nomenclature in 6–7 DTUs [5, 6]. Moreover, in some cases correspondences between typing methods with different markers were used for the data interpretation [6, 25]. The data are shown in an Excel spreadsheet (S1 Table) where each line corresponds to a single determination from an isolate, a strain, a laboratory clone, mammal blood or tissue samples, and different vector digestive tract samples (“sample type” column in S1 Table). Several lines were recorded when different DTUs were detected in a strain and its laboratory clones. When more than 1 DTU was detected in one vector or mammal host (mixed infection), several lines corresponding to each DTU were recorded in the file. A total of 6,343 determinations were compiled. Each of them has a code corresponding to the strain/sample name reported in the publications, except for the records not identified with a name but only counted in publications, which we have labeled “anonymous”. In a few publications, undistinguished DTUs were reported for part of the identifications; consequently, additional categories were created for them: TcI/TcII (three cases), TcII/TcV/TcVI (26 cases), TcII/TcVI (two cases), TcIII/TcIV (31 cases), and TcV/VI (47 cases). These undistinguished DTUs accounted for 1.7% of the total inventory. The geographical origin was informed by the country name (no missing data), the upper continental subdivision of North, Central, and South America, the upper administrative divisions such as state, province, department or region, and the lower administrative divisions such as municipality, province, or community according to the information existing in the publications. The collection dates of the strain or biological samples were not always documented (52.5% of missing year data). Host origin was generally informed by the species (31 missing data), and columns were added indicating the order, genus and tribe for the triatomines. Also, the cycles to which the different hosts belonged were classified as “domestic” when the hosts were living and/or were captured in the intra- and peridomicile areas, and “sylvatic” when the hosts were captured in the field outside domestic areas. When the location of the capture site was missing, the wild mammals where classified in the sylvatic cycle except for the synanthropic species such as opossums and rodents for which the information was considered as unknown (uk). The information on the methods used for the characterization of the DTUs is also included in S1 Table. The first column indicates if the DTU was characterized at nuclear or mitochondrial level or both, the second one indicates the method(s) used, and the third one on the markers, the names of the genes, or the number of loci for MLMT (multilocus microsatellite typing) and MLEE (multilocus enzyme electrophoresis).
The 6,343 samples of T. cruzi DTUs compiled in this review were identified in vectors and mammalian hosts from 19 different countries, covering an area from the southern United States to Argentina (S2 Table). No data is available from Belize in Central America, and Uruguay and Guyana in South America. The vast majority of data relate to South America (90.7%). The DTUs were identified in 86 genera (32 missing cases), 158 different species of which 42 are vectors belonging to 7 genera (Dipetalogaster, Eratyrus, Meccus, Mepraia, Panstrongylus, Rhodnius, and Triatoma). Approximately of the identifications in South America 49.3% were from vector species; however, in North and Central America most of the identifications were from vectors (69.3% and 65.8% respectively). The mammal species belong to nine orders of which the most represented is the Primate order (61.5%), because 59.4% of the identifications in mammals were made in samples from humans (n = 1902). One-third of the DTU identifications (31.0%) corresponded to parasites from hosts (vectors and mammals) captured in sylvatic ecotopes, 57.6% from intra- and peridomestic hosts, and the others were undetermined (n = 719, 11.3%) because in several studies the origin of the vectors was not specified.
In 1.7% of the samples, the DTU (n = 109) was reported as a group of DTUs: (i) in one dog, 15 coati from Brazil, and ten triatomines from Argentina, TcII, TcV, and TcVI were not distinguished; (ii) TcII or TcVI was reported in two T. infestans from Paraguay; (iii) 47 infections with TcV or TcVI in dogs, humans, T. infestans from Chile and Bolivia and P. megistus in Brazil were reported; (iv) in 31 vectors and mammal hosts from Brazil and Mexico TcIII/TcIV were not discriminated; and (v) in three cases TcI and TcII were not discriminated in T. pallidipennis. In the 6234 other records, TcI was found in approximatively 60.0% of the overall identifications; TcII, TcV and TcVI were identified in around 10% each; and TcIII, TcIV and Tcbat were rarer with percentages ≤ 3.6%. Fig 1 presents the proportions of DTUs observed, excluding from the calculation the ambiguous DTU determinations over the entire endemic area, and in North, Central and South America (see below).
According to the current available records, the DTU distribution was different between North, Central, and South America (Fig 1). In Central America only two DTUs (TcI and TcIV) were identified while all DTUs were detected in South America. In North America the latest studies have identified TcII, TcV and TcIII in addition to TcI and TcIV, which remain the major strains, in Central America. In South America the DTU distribution was highly variable depending on the country, and the current trend is a predominance of TcI north of the Amazon and the presence of all DTUs south of the Amazon with abundance of TcV and TcVI (Fig 2).
Tcbat is a recently proposed DTU that is genetically more closely related to TcI than to any other DTU. Therefore this DTU is probably underestimated because it is not recognized by the markers used in many publications, and consequently it may have been erroneously equated with TcI. This DTU was identified in 59 bats belonging to 12 different species in Brazil, Colombia, and Ecuador [16, 17, 26, 27], in one specimen of T. sordida from the State of Mato Grosso do Sul State in Brazil [28], and in a Colombian patient infected with a mixture of TcI and TcBat [18].
As mentioned above, TcI was the most frequently identified DTU in the overall sample, with a lower percentage in South America (58.2%) than in North America (79.5%) and Central America (93.3%). It was identified in all the countries included in the study. In South America, the low frequencies of TcI in Argentina (19.9% of 589 determinations) and Paraguay (2.8% of 181) contrasted with the proportions of this DTU in the other South American countries (at least > 47.0%) (Fig 2).
TcII was much more rarely identified (9.6% of overall DTUs identified). It was not identified in Central America out of 120 identifications, and only 13 identifications were reported from North America out of 459 (2.8%). Eight of these 13 TcII were found in Mexico, four in T. dimidiata captured in domestic cycles in the state of Veracruz [29] and four in Meccus pallidipennis collected in Michoacan [30]. The five other identifications were from mice and rats captured in the immediate surroundings of the dwelling of the first described autochthonous case of T. cruzi transmission in Louisiana, near New Orleans [31, 32]. In South America, TcII presents a higher proportion, reaching 10.4% and was reported in Colombia, Surinam, Peru, Bolivia, Brazil, Argentina, Paraguay and Chile.
TcIII and TcIV, which are thought to result from ancestral hybridization between TcI and TcII, reached 3.4% and 3.6% of the identifications, respectively. In North America, both of these DTUs were reported in Mexico in several publications [29, 30, 33, 34], but for the moment only TcIV has been identified in the US [24, 31, 35, 36]. In Central America, only TcIV has been identified in Guatemala in humans and vectors [37, 38]. In other Central American countries, neither TcIII nor TcIV has been reported. In South America, TcIII could be more cosmopolitan (Argentina, Bolivia, Brazil, Chile, Colombia, Paraguay, Peru and Venezuela) than TcIV, which has not yet been identified yet in Argentina, Chile and Paraguay.
The last two DTUs, TcV and TcVI, were the recent hybrids, derived from hybridizations between TcII and TcIII. These DTUs showed the most differential geographical distribution. Indeed, TcV was identified in North America in exceptional cases in Mexico (Veracruz) in T. dimidiata as well as above-mentioned TcII [29]. TcV and TcVI have never been identified in US in 148 determinations, nor in Central America in 120 cases. In contrast, in South America, these DTUs together have frequently been identified in several countries, Argentina (76.9%), Bolivia (44.6%), Chile (28.6%) and Paraguay (55.2%)—but very rarely in others such as in Colombia (1.1%) [24, 39–41], Ecuador (3.3%) [42], and Brazil (1.5%) [24]. In Peru they were identified in 13.0% [24, 43, 44]. Moreover, when the two DTUs coexist, different proportions can be observed in the different countries. The most remarkable case was the identification of TcV and TcVI in Bolivia with 43.1% and 1.0% respectively, while in Argentina TcVI was more common (50.0%) and TcV less frequently detected (26.5%).
For many years, the characterization of T. cruzi strains was mostly conducted with specific goals in limited geographical areas and consequently with a limited number of strains. The current compilation, based on the consensus nomenclature of six DTUs, reached an accumulated number of 6,343 identifications. However, T. cruzi genotyping is associated with many biases and trapping methods, and several caveats must be considered, such as (i) unequal distributions of the research groups in the eco-epidemiology of T. cruzi in different countries, resulting in nonhomogeneous information; (ii) selection of some DTUs during the culture step; (iii) differential parasitemia levels in hosts, facilitating the isolation by hemoculture or xenodiagnosis, or facilitating the direct detection of some DTUs over others; (iv) markers’ differing ability to detect the different DTUs; (v) overrepresentation of humans in the overall sample; (vi) scarcity of mammals that are difficult to trap; (vi) difficulties discriminating closely related DTUs; and (vii) use of a nonstandardized set of reference strains. Despite of this nonexhaustive list of biases, the data reported herein constitute the most complete picture of the DTU distribution in the endemic area of Chagas disease.
The purpose of this review is not to discuss the current nomenclature of T. cruzi in six DTUs. Indeed, there is an increasing number of new genetic analyses of T. cruzi strains, especially from sylvatic cycles, which show that it is increasingly difficult to obtain a relevant genetic structure that divides into six statistically supported clusters with the most in vogue genetic markers, microsatellites and nuclear sequence polymorphisms [68–70]. Moreover, at the mitochondrial level, we recently assessed that three robust clusters that we named mtTcI, mtTcII and mtTcIII actually exist [8]. The mtTcI cluster includes only strains belonging to the TcI DTU, the mtTcII includes only those belonging to the TcII DTU and mtTcIII includes strains belonging to several DTUs: TcIII and TcIV (ancient hybrids of TcI/TcII), TcV and TcVI (recent hybrids TcII/TcIII) and even TcI (a result of mitochondrial introgression for some strains labeled TcI with nuclear markers). These last few years, a number of studies aiming to characterize T. cruzi strains have used the nomenclature of six DTUs, so we proposed to examine the eco-epidemiological features of these DTUs and highlight new knowledge that may challenge the current paradigm.
Based on the available typing data, the first outstanding result is the predominance of TcI strains. This DTU, genetically diversified, is found throughout the geographic distribution of T. cruzi and in all cycles where it is always dominant. There are probably no ecological systems (i.e. geographical areas where the parasite evolves between mammalian hosts and vectors specific species) where TcI is absent. However, it appears that TcI strains do not develop well in some mammal species such as those within the order Cingulata since this order is rarely infected with TcI (Table 3). The ecological systems are usually complex networks of relationships involving many species of mammals and vectors, and strain diversity may be maintained because of differential interactions between the parasite’s hosts and genotypes. TcI is an old DTU that has evolved since 3–16 MYA as previously proposed [71], and its very high genetic diversity is consistent with a long-term evolution. Moreover, recombination between TcI strains appears to be more frequent than previously thought [2, 3, 72]. The recombination events (i.e. sex) generally increase the variability of the organisms and thus increase their resilience, allowing new areas to be conquered and especially new hosts that have probably played a key role in the large dispersion and adaptation of TcI. Another question is the geographical origin of TcI. A North-South clustering was recognized, even if some incongruence remains to be explained [73–75]. In an analysis of TcI, the Gran Chaco region was proposed as an origin, while human TcI may have a North/Central American origin [75–77]. It should be noted that if the current trend is to propose sub groups within TcI, the presence of subunits, evolving separately, must be previously evidenced which is not yet the case. Also, it has been proposed that marsupial species of the family Didelphidae family are the ancestral hosts of TcI [78] given that, among others, TcI predominates in these animals. Based on our recent analysis of COII and CytB gene sequences previously deposited in GenBank [8], we evaluated the haplotype and nucleotide diversities of TcI within the order Didelphimorphia, and we observed that these indices were comparable to those obtained for all the other orders of wild mammals combined. This assesses the larger genetic diversity in marsupials than in other animals, supporting a longer association. The remarkable expansion of TcI, which invaded most of environments, does not allow its origin to be determined from the picture of its geographical distribution alone.
TcII is a DTU as old as TcI, but it has been sampled much more rarely. The strains belonging to this DTU carry mitochondrial genes (mtTcII mitochondrial cluster) whose sequences show substantial genetic divergence from TcI. Moreover, this DTU presents a much lower genetic diversity than TcI. For example, the haplotype diversity of COII and CytB genes are 0.39 and 0.48, while for mtTcI they are 0.81 and 0.58 respectively [8]. A similar level of differences is also observed for nucleotide diversity. The available data on the geographical distribution of TcII suggest that it is absent or extremely rare in some ecosystems (Central and North America). It seems that TcII strains would not have had the same expansion capacity as TcI among the wild cycles, and they probably found refuge mostly in certain wild mammals. TcII is already reported in different wild mammals of the Chiroptera, Cingulate, Didelphimorphia and Primate orders. However, its strong association with primates in the Atlantic Coastal Rainforest in Brazil should be noted [79]. In humans, it is relatively abundant, accounting for 20% of human strains, but it is highly abundant in Brazil (66% of human strains identifications) and rare in most other countries except Colombia (15%) and Chile (30%). For now, its geographical distribution is more consistent with a South American origin, and further south than north of the Amazon basin where this DTU is more abundant.
TcIII and TcIV are DTUs that do not seem to be present throughout the entire endemic area. First, it is important to note that the genetic data do not clearly define these two groups separately. The genetic diversity of TcIII-TcIV is very large and the monophyly of each DTU is not really highlighted. Several studies showed that these strains are the result of ancient hybridization(s) between TcI and TcII strains, which suffer over time from genetic rearrangements, decreasing their level of heterozygosity at the expense of mosaic mitochondrial and nuclear genes [80]. Recombination events have probably occurred several times and this would have given a mtTcIII group composed of polyphyletic subgroups of strains. Therefore, the wild strains from the US, attributed to TcIV, seem to be a monophyletic subgroup differing from the others long ago [81], but whose closest ancestors have probably disappeared. There is little doubt that TcIII and TcIV DTUs have a sylvan origin, but these strains infect humans more than occasionally: the current database shows that TcIV is reported in 84 human cases in six countries (Brazil, Colombia, Ecuador, Guatemala, Peru and Venezuela), and 11 canine cases. Similarly, TcIII is reported in 26 human cases in Brazil and Paraguay.
The two TcV and TcVI DTUs include strains derived from the hybridization of TcII and TcIII strains [9]. They are usually considered hybrids and they are heterozygous at several loci and SNPs (single nucleotide polymorphisms). In our database, a total of 21.3% of the determinations belong to these DTUs. Some of these strains have spread across large geographic areas through the clonal propagation mode [82]. Both DTUs are clearly associated with domestic cycles since only 10.5% of them are identified in hosts from wild cycles. They are identified in some Didelphimorphia and in different species of rodents but only in the Southern Cone countries and Bolivia. Previously, the Gran Chaco region was proposed as the original location of these DTUs, where they are very abundant and where the putative parents are also present [15], and this hypothesis fits well with the current observed distribution of these DTUs.
The universe of Hemiptera vectors of T. cruzi or potential vectors is huge since currently over 141–147 triatomine species are recognized, about 130 occur in the Americas, and it appears that all of these are able to transmit the parasite. Most of these species are involved in wild cycles with at least 100 species of mammals playing a role of host and/or reservoir. In the current data, only 37 species of vectors are included and for the majority of them, very few DTU determinations were made, even though these vectors are generally widely distributed. Similarly, the knowledge of the parasite genetic variants that infect mammals, except for humans, and to a lesser extent for Didelphidae, is very limited. In various regions, in a context of high anthropization and climate changes, it is urgent to study the impact of these environmental modifications on potential vectors and their hosts.
Several studies of experimental infections of vectors with different strains of T. cruzi showed differences in susceptibility [83] and even suggested that the strains are pathogenic and induce more or less deleterious effects in bugs [84]. Few studies relate comparisons of DTUs in experimental infections in a single triatomine species. For T. infestans in which this was done, significant developmental differences in the vector were observed depending on the DTU it was infected with, and after experimental double infections: in 50% of cases, only one of the two DTUS was detected after a few days of infection [85, 86]. As a field observation, we can report the case of Triatoma sordida, a primary vector in the northeast of the city of Santa Cruz, in Bolivia, in which TcI was predominantly detected while in mammals of the same area, TcV was a major strain [59]. In wild mammal hosts, experimental infections of two important reservoirs in the US (placental and marsupial) showed DTU-mammal association [87].
Examples could be multiplied but we can already conclude that the vectors and even the wild mammal hosts can influence the distribution of DTUs. Whatever the host, there is a balance between parasite genotypes and hosts which probably depends on environmental conditions such as outside temperature for vectors or immune and nutritional status for mammals. The diversity of hosts, and environmental conditions certainly explain the maintenance of parasitic diversity and the emergence of new variants by natural selection. Therefore the distribution of DTUs reported here, although very informative, is only a temporary picture that will inevitably evolve over time, above all if drastic environmental changes occur such as deforestation, intensive farming, urbanization, and unexpected climatic upheavals.
|
10.1371/journal.pntd.0000630 | Nodular Worm Infection in Wild Chimpanzees in Western Uganda: A Risk for Human Health? | This study focused on Oeosophagostomum sp., and more especially on O. bifurcum, as a parasite that can be lethal to humans and is widespread among humans and monkeys in endemic regions, but has not yet been documented in apes. Its epidemiology and the role played by non-human primates in its transmission are still poorly understood. O. stephanostomum was the only species diagnosed so far in chimpanzees. Until recently, O. bifurcum was assumed to have a high zoonotic potential, but recent findings tend to demonstrate that O. bifurcum of non-human primates and humans might be genetically distinct. As the closest relative to human beings, and a species living in spatial proximity to humans in the field site studied, Pan troglodytes is thus an interesting host to investigate. Recently, a role for chimpanzees in the emergence of HIV and malaria in humans has been documented. In the framework of our long-term health monitoring of wild chimpanzees from Kibale National Park in Western Uganda, we analysed 311 samples of faeces. Coproscopy revealed that high-ranking males are more infected than other individuals. These chimpanzees are also the more frequent crop-raiders. Results from PCR assays conducted on larvae and dried faeces also revealed that O. stephanostomum as well as O. bifurcum are infecting chimpanzees, both species co-existing in the same individuals. Because contacts between humans and great apes are increasing with ecotourism and forest fragmentation in areas of high population density, this paper emphasizes that the presence of potential zoonotic parasites should be viewed as a major concern for public health. Investigations of the parasite status of people living around the park or working inside as well as sympatric non-human primates should be planned, and further research might reveal this as a promising aspect of efforts to reinforce measures against crop-raiding.
| The disease caused by the nodular worm Oesophagostomum bifurcum can be lethal in humans and is thus of major human health significance in certain African regions. There are still gaps in the understanding of the epidemiology of the disease, including the role of non-human primates as reservoirs of the infection. We recently conducted a survey in a community of wild chimpanzees (Pan troglodytes schweinfurthii) in Kibale National Park, Western Uganda. O. stephanostomum is so far the only species previously found in chimpanzees. A total of 311 stool samples was examined and revealed that high-ranking males are more infected than other individuals. These chimpanzees are also the more frequent crop-raiders. Moreover, we reported for the first time molecular evidence for O. bifurcum in addition to O. stephanostomum in chimpanzees. Our results raise public health concerns for a neglected infection in regions where spatial proximity between great apes and humans are increasing because of forest fragmentation.
| Nodular worms (Oesophagostomum spp.) are commonly found as nematode parasites of pigs, ruminants and primates, including humans. In endemic foci in Africa, especially in Ghana and Togo, a high prevalence of Oesophagostomum bifurcum infection has been reported in human populations, one million being estimated at risk [1],[2]. Patients are mostly children aged <10 years [2]. Clinical disease, due to encysted larvae, known as oesophagostomosis, sometimes leads to death [1]–[3]. The distinction between hookworm and nodular worms eggs is not possible [2] and the definitive diagnosis of oesophagostomosis in humans involved exploratory surgery or ultrasound examination. Transmission occurs through the ingestion of the infective third-stage larvae (L3) but the factors explaining such a high regional prevalence remain unknown. Eight species of Oesophagostomum have been recognized so far to occur in non-human primates [4]. Among them, O. bifurcum, O. stephanostomum and O. aculeatum are also reported in humans [3]. Human cases have been attributed to a zoonotic origin, non-human primates being proposed as a potential reservoir [3]. However experimental infection of rhesus monkey (Macaca mulata) showed that O. bifurcum obtained from humans did not effectively infect monkeys [5]. In addition, significant variations exist in lengths of adult worms isolated from humans and non-human primates [4]. The geographic distribution in humans and some non-human primates is not overlapping [6],[7] and recent molecular findings demonstrated a genetic host-affiliated sub-structuring within O. bifurcum [6],[7]. Among great apes, especially chimpanzees, bonobos and gorillas, prevalence of strongyle eggs in stools is often high and O. stephanostomum was the only species of Oesophagostomum identified so far [8]–[10]. However, little is known about the intensity of infection in terms of parasite load and clinical signs in great apes. It has been reported that wild apes develop clinical signs of oesophagostomosis as soon as in captivity [11] while the presence of parasites remains asymptomatic in wild animals. Recently fatal cases have been described in African apes from sanctuaries [12] and collected parasites were diagnosed as O. stephanostomum. Nevertheless, because of the phylogenetic and spatial proximity between humans and chimpanzees, potential transmission is not excluded especially in Uganda where human oesophagostomosis has been reported [4]. Around Kibale, population density is high (up to 512 ind/km2) [13] and chimpanzees regularly crop-raid. Additionally recent findings confirmed that human-related diseases should be considered as a high threat for endangered apes [14]–[18]. As a consequence, it has been emphasized that investigations on potential cross-transmission should be reinforced. We report hereafter the results of our recent finding about nodular worm infection in wild chimpanzees (Pan troglodytes schweinfurthii) in the framework of a long-term health monitoring of the community of Kanyawara in Kibale National Park (Uganda).
The studied chimpanzees (Pan troglodytes schweinfurthhii) belonged to one community in Kibale National Park (766 km2, 0°13′–0°41′N, 30°19′–30°32′E), located in Kanyawara area. This community counted 52 chimpanzees in 2006. Ages presented are those estimated in 2006. Their home range is close to the boundary of the Park and Kanyawara chimpanzees are sometimes entering plantations for crop-raiding. Stool samples were collected from identified individuals within the minutes following defecation.
We performed analyses on two series of fecal samples (Table 1). From December 2005 to March 2006, a total of 295 fecal samples was collected from 33 chimpanzees, 17 females (13 adult females and four immature females) and 16 males (9 adult chimpanzees including five dominant individuals, four subordonate individuals and seven immature males) (set 1); coproscopy, coproculture and molecular analysis were performed on the total or parts of this set. In October 2008, 16 samples were collected from 10 identified chimpanzees, 5 females and 5 males. These samples were dried for further molecular analysis (set 2). Indeed, since coprocultures in field conditions and diagnosis of third-stage larvae (L3) are laborious and require skilled personnel for identification, we wished to test a molecular method using dried feces.
For each sample of set 1 (n = 295), two grams of fresh stool were preserved in 18 mL of 10% formalin saline solution, then smears made with 50 µL of the suspension were microscopically examined. MacMaster flotation was performed at the field station on fresh stools within the day of collection. MacMaster cells were filled with one mL of filtrat of two grams of fresh stools diluted in 30 mL of magnesium sulfate. However, as electricity was not available every day, only 100 samples could be examined.
With both methods, strongyloid eggs were identified according to their size, color, shape and morula aspect (16–32 cells) and they were counted. Egg per gram (epg) counts were corrected according to the fecal consistency (ie ×2 for soft stools and ×3 for diarrheic stools) [12]. Arithmetic corrected mean was calculated including infected and non infected samples (mean abundance). Larvae of Probstmayria sp. and larvae of unidentified species as well as ciliates were also counted during coproscopy (data not shown).
To confirm identification of strongyloid eggs in set 1, coprocultures were performed with 16 stool samples from 13 individuals (5 males, 8 females). After 10 to 15 days of incubation, larvae were collected by Baermann technique and preserved in 95% ethanol. Third-stage larvae (L3) of Oesophagostomum spp. obtained were microscopically diagnosed (long filamentous tail of the sheath, triangular intestinal cells, and length of the larvae [4]).
Molecular characterization was performed on samples from sets 1 and 2. With the mixture of larvae obtained from each above coproculture (n = 16 samples, set 1), DNA was prepared using Nucleo-spin Tissue (Macherey-Nagel) and ITS2 region was amplified using the primers NC1 and NC2 as described previously [19]. ITS2 sequence of O. stephanostomum is characterized by 2 digestion sites for NLaIII while ITS2 sequence of O. bifurcum is characterized by a unique digestion site. RFLP were analyzed after digestion of the ITS2 sequence. Sequencing was performed on ITS2 sequences and compared to published data (GenBank accession numbers: AF136576 for O. stephanostomum; AF136575 and Y11733 for O. bifurcum). Another PCR test was performed from DNA obtained from the cultured samples to compare the two methods. We used a semi-nested PCR followed by direct sequencing as described before [20].
From set 2, 16 samples of 4 g fresh feces were stored dried on 20g of silicagel bead. Before DNA extraction, vegetal debris was removed in order to avoid PCR inhibition. DNA was extracted from 200mg of dried feces without culture by using the QIAMP DNA Stool Kit (Qiagen, Chatsworth, CA) according to instructions with the following modifications: in step 3, the suspension of 200 mg with buffer ASL was incubated overnight at 70°C and in step 12, the solution was incubated one hour with proteinase K at 70°C. Direct sequencing after PCR using NC1 and NC2 primers was performed.
Table 1 presents results obtained from the two sets of collection with the different methods of analysis.
Strongyloid eggs were detected in 12% of the 295 feces examined with direct smears, that is 60% of the chimpanzees (n = 33). The arithmetic mean corrected parasite load of strongyloid eggs was 52±12 epg. The diarrheic samples had a significantly higher oesophagostomine egg counts (225 epg, n = 17) than the firm feces (19 epg, n = 217) (Kruskal-Wallis test: P<0.01). No other factor was significantly affecting egg counts by direct examination although egg counts tend to be affected by hierarchical status in males (dominants: 42 epg, n = 33, subordinates: epg: 14, n = 42).
Strongyloid eggs were detected at least once from all the chimpanzees (n = 29) with MacMaster method. The proportion of positive samples for strongyloid eggs with Mac Master flotation was 91%. The arithmetic mean corrected parasite load of strongyloids was 140±58 epg. Values of corrected epg were significantly different according to hierarchical status in males (dominants: 232 epg, n = 10, subordonates: 88 epg, n = 13; Mann-Whitney test; P value = 0.005) and fecal consistency (firm feces: 90 epg, n = 73; soft feces: 239 epg, n = 20; diarrheic feces: 414 epg, n = 7; Kruskal-Wallis test: P value = 0.021). No difference according to the sex, the age or the sampling period of the day was found.
L3 characteristic of Oesophagostomum were found after coproculture and microscopic examination in 12/16 samples from 10/13 chimpanzees (3 males; 10 females) of the set 1.
PCR-RFLP conducted on larvae from the 16 coprocultured samples (set 1) from wild chimpanzees identified O. stephanostomum and O. bifurcum. ITS2 sequence of O. stephanostomum, characterized by 2 digestion sites for NLaIII, was identified from fecal samples from 2 chimpanzees (MS, male and AL, female). ITS2 sequence of O. bifurcum, characterized by a unique digestion site, was identified from one fecal sample from one chimpanzee (AJ, male) (Fig. 1). Sequencing performed on these samples confirmed the presence of the two species. All but one samples revealed DNA sequences showing 99% of homology with Panagrolaimus sp. (AY878405 from Genbank) and one sample revealed 82–88% homology with Necator sp. (AF217891 from Genbank) nematodes. With semi-nested PCR and direct sequencing, O. stephanostomum was identified in one of the two chimpanzees positive with PCR-RFLP (MS, male) and O. bifurcum in five chimpanzees (Fig. 2). In the second set of fecal samples, which were stored dried, Oesophagostomum DNA was found in 14 of the 16 fecal samples. All the 10 chimpanzees sampled in set 2 were positive.
All the sequences obtained from extraction of larvae DNA and dried feces were identical between them for each of the two Oesophagostomum species. The sequences corresponding to O. stephanostomum were 100% identical to the ITS2 reference sequence of O. stephanostomum collected from a chimpanzee in Tanzania (GenBank accession number AF136576) (BLASTn). In the sequences we obtained there was no mixed sequence signals in positions 116, 176, 197 in which the reference sequence shows polymorphism. (fig. 2). Among the three sites where polymorphism is observed in ITS-2 sequences for human and monkeys for O. bifurcum (positions 56, 112, 162), there is no nucleotides equivoque in our sequences. Our sequences were identical to the ITS2 sequence of O. bifurcum collected from a human (100% of identity with GenBank accession number Y11733) and different in position 112 from the one documented in Cercopithecus mona (GenBank accession number AF136575). The species O. stephanostomum was found in 11 chimpanzees. The species O. bifurcum was found in 5 chimpanzees. One chimpanzee was co-infected by two Oesophagostomum species in the first period (AL, female) and four chimpanzees (2 females AJ and WL and two males AL and ST) were infected by the two species if we considered both periods (December 2005 to March 2006 and October 2008). Whatever the method used to identify specimens of the genus Oesophagostomum (coproculture or any molecular characterization), a total number of 26 of 32 samples (81%) were positive corresponding to 15 chimpanzees of the 18 sampled (83%).
In this study we used several methods to survey parasite status of wild chimpanzees.
We compared nodular worm eggs counts between individuals of different classes of age, sex and dominance rank. We demonstrated for the first time that wild chimpanzees could be infected by O. bifurcum. The presence of two Oesophagostomum species (O. stephanostomum and O. bifurcum) was reported in the same chimpanzees community. Our results, based on RFLP-PCR and semi-nested PCR-direct sequencing and PCR from dried stools, extend our understanding of the epidemiology of O. bifurcum, confirm accuracy of alternative method (DNA extraction from dried stools) to coproculture and raise public health awareness for a neglected disease.
However while the substantial increase of accuracy of PCR compared to coproscopy has been previously shown [21], technical difficulties and limitations of stool analyses and culture due to field conditions when studying wild chimpanzees have to be considered. They are overcome by using PCR directly on dried stools. Additionally, the sensitivity of molecular analysis was higher when applied directly on dried samples than samples obtained from coproculture. With both methods of coproscopy, we determined that high-ranking males in Kanyawara chimpanzee community had higher parasite burdens during the study period. Our results also provided evidence that these free-ranging chimpanzees are infected by two Oesophagostomum species (O. stephanostomum and O. bifurcum). The species O. bifurcum is responsible for human and monkeys infections and had never been characterized in wild great apes as chimpanzees so far. The species O. stephanostomum is detected in great apes and this species was recently associated with nodular lesions in chimpanzees and a gorilla from sanctuaries [12]. Behavioral patterns of Pan troglodytes may explain that males are more infected than females by strongyloid parasites: male chimpanzees are staying all their life in their native community while females migrate. Males develop close relationships, indulging in very long grooming sessions where individuals are staying in close proximity. Our results are consistent with previous studies showing that both testosterone and cortisol were positively associated with gastrointestinal parasite infections in Kibale chimpanzees [22] suggesting that stress of high-ranking males may alter an efficient immune response. Additionally males are visiting plantations in the edge of the forest more frequently than females, encountering conditions favoring parasite transmission from humans and non-human primates: people being very close to the forest are usually not using latrines and monkeys in the edge such as red colobus are more infected than those from the interior, egg counts for Oesophagostomum being 10 times higher [23].
In spite of the severe health problem caused by oesophagostomosis to humans, epidemiology and transmission of the disease are still poorly understood [24]. While colobus monkeys were not infected in surveys conducted in Ghana [20],[21], in Kibale NP, primates including the arboreal red colobus (Piliocolobus tephrosceles) and black and white colobus (Colobus guereza) and more terrestrial species such as olive baboons (Papio anubis) were proved to be infected by Oesophagostomum sp. [23],[25],[26]. Diagnosis of the parasites species was not conducted in monkeys but previous findings suggesting no risk of infection for arboreal colobus monkeys [24] was not supported at the genus level in this area. The role of chimpanzees and other primates in the cycle needs thus to be further explored. Oesophagostomum bifurcum nematodes from chimpanzees may be genetically distinct from O. bifurcum nematodes from other primate species including humans as previously demonstrated [6],[7]. However, chimpanzees are more closely related to humans than non-human primates species investigated so far (colobus, baboons, patas and Mona monkeys) and investigating the genetic variability of O. bifurcum between chimpanzees and other primates would be interesting. Moreover, the home ranges of chimpanzees from Kibale NP include areas where human beings are present. Chimpanzees are visiting plantations bordering their forest home range and males, especially high-ranking males, which have higher infection level, more frequently. Humans are working or entering inside the park (researchers, research assistants, other employees from the park, poachers…). Chimpanzees are also exploiting resources also used by other non-human primates. For these reasons, even if the origin of infection is unknown, the zoonosis risk can not be excluded. Outbreaks of oesophagostomosis in human population have not been documented in the study area. However, an investigation of the parasite status in humans living in the villages surrounding the park should be planed. The presence of potentially zoonotic parasites in chimpanzees in a context where proximity between human and apes is increasing (ecotourism, crop-raiding, research…) should be viewed as a point of concern for the future of public health in this region and elsewhere. For economic and health reasons, prevention of crop-raiding programs should be reinforced.
|
10.1371/journal.pcbi.1004107 | Protein Complexes in Bacteria | Large-scale analyses of protein complexes have recently become available for Escherichia coli and Mycoplasma pneumoniae, yielding 443 and 116 heteromultimeric soluble protein complexes, respectively. We have coupled the results of these mass spectrometry-characterized protein complexes with the 285 “gold standard” protein complexes identified by EcoCyc. A comparison with databases of gene orthology, conservation, and essentiality identified proteins conserved or lost in complexes of other species. For instance, of 285 “gold standard” protein complexes in E. coli, less than 10% are fully conserved among a set of 7 distantly-related bacterial “model” species. Complex conservation follows one of three models: well-conserved complexes, complexes with a conserved core, and complexes with partial conservation but no conserved core. Expanding the comparison to 894 distinct bacterial genomes illustrates fractional conservation and the limits of co-conservation among components of protein complexes: just 14 out of 285 model protein complexes are perfectly conserved across 95% of the genomes used, yet we predict more than 180 may be partially conserved across at least half of the genomes. No clear relationship between gene essentiality and protein complex conservation is observed, as even poorly conserved complexes contain a significant number of essential proteins. Finally, we identify 183 complexes containing well-conserved components and uncharacterized proteins which will be interesting targets for future experimental studies.
| Though more than 20,000 binary protein-protein interactions have been published for a few well-studied bacterial species, the results rarely capture the full extent to which proteins take part in complexes. Here, we use experimentally-observed protein complexes from E. coli or Mycoplasma pneumoniae, as well as gene orthology, to predict protein complexes across many species of bacteria. Surprisingly, the majority of protein complexes is not conserved, demonstrating an unexpected evolutionary flexibility. We also observe broader trends within protein complex conservation, especially in genome-reduced species with minimal sets of protein complexes.
| Abundant genome sequencing revealed an astounding diversity among bacterial genomes. Even species that inhabit the same environment may only share a fraction of their genes. This raises the question how these organisms have adapted to their environments using only a limited number of genes. Here, we investigate the protein complements across bacterial genomes, how proteins are combined into protein complexes across species, and whether these complexes have been conserved across diverse branches on the prokaryotic tree of life.
Other studies have compared the interaction networks of S. cerevisiae, S. pombe and E. coli made possible by systematic screens of genetic interactions and have found notable differences in their structure and content [1,2]. Studies comparing baker’s yeast and fission yeast found that essentiality also varies between species [3]. This might be explained by functional redundancy and the importance of mechanism over structure. The extent of the differences might be unexpected but make sense when seen in the light of evolutionary flexibility [1].
Numerous studies of protein-protein interactions have revealed the organization of proteomes into networks of interactions as well as protein complexes. Systematic surveys of protein complexes exist for only a few bacterial species, namely E. coli [4,5] and Mycoplasma pneumoniae [6]. The list of binary protein-protein interactomes is clearly larger but has not been considered in this study. Based on this limited dataset, we investigated whether the complexes found in a few model organisms are sufficient to reconstruct homologous protein complexes in other species. This is a particular challenge in prokaryotes as the genomes of most species are highly divergent from the few model species used here. However, E. coli and Mycoplasma provide two important paradigms: E. coli is a generalist that can live under a variety of conditions while Mycoplasma is a specialized parasite that requires host cells to grow. With ~4,300 and ~700 genes, respectively, they represent medium-sized as well as minimal genomes and thus medium and minimal diversity of protein complexes.
Few studies have investigated the evolution and diversity of protein complexes across a wide range of taxa. This is not surprising given that large-scale experimental data has only become available in recent years. In combination with a large number of completed genome sequences we can use this data to evaluate the extent to which protein complexes are likely to be conserved across microbial species. Furthermore, we can evaluate the biological role of proteins and complexes of unknown function across many species.
Existing studies comparing sets of interactome data, including pure bioinformatics approaches [7] have generally limited their comparisons to a few well-characterized protein-protein interaction networks, such as comparisons of S. cerevisiae, S. pombe and E. coli [2,7,8]. Methodological frameworks for predicting co-evolution on the basis of gene presence/absence [1,9] may also be employed to predict novel interactions in other species. In this study, we focus on eight distinct bacterial species, seven of which have been the subject of essentiality screens and two of which have comprehensive protein complex surveys available. We then expand the focus to a set of 894 bacterial genomes.
In order to compare genomes and protein complexes across species, we couple the results of mass spectrometry-characterized protein complexes [5,6] with databases of gene orthology [10] and essentiality [11] to characterize interaction conservation within protein complexes. Furthermore, we use the perspective of genome reduction to evaluate patterns across levels of protein conservation. Comparing sets of protein complexes from divergent bacterial species (in this case, E. coli and M. pneumoniae) alleviates some of the bias inherent in using a single species as a universal model. Rather, observing which protein complexes and their components are present in two otherwise distinct species allows us to draw conclusions about how critical these components are to microbial life.
In general, small microbial genomes are enriched for proteins which are conserved across bacteria (Fig. 1; Table 1). This trend is most noticeable when paralogy is eliminated, either by removing all but one in a paralogous group (PG) or by the natural effects of genome reduction, as is seen in many of the smallest bacterial genomes. In both cases, average protein conservation decreases as genome size increases. Genes of larger genomes, such as that of Pseudomonas aeruginosa, may be conserved across 20 to 30 percent of all other bacteria, on average. The most minimal genomes, including those of Mycoplasma species, may share their orthologous groups (OGs) with 60 to 80 percent of other bacterial species, on average. These results are reasonable and expected: reduced genomes, by definition, have lost sequence space but have not lost the loci most crucial to bacterial life itself. Furthermore, though most genomes show an increasing fraction of paralogs being conserved as their size increases (Fig. 1), many of the most reduced genomes actually show greater average conservation when potential paralogs are removed. The paralogous protein-coding loci in reduced genomes may be enriched for accessory genes rather than broadly-conserved core genes.
The presence of multiple members within a single orthologous group has an effect on average gene conservation. Here, we display this effect as the difference between average locus conservation and average OG conservation. (Fig. 1, orange vs. blue dots). The difference between the values is an approximation of the level of paralogy across each genome; larger genomes appear to contain more paralogs than smaller genomes, especially as genome size falls below 1 Mb. The effect on average gene conservation is expected, as using orthology-based comparisons compresses paralogs into single OGs. Within our data set, 20 genomes (within 10 unique genera) under 3 Mb had greater average conservation among OGs than among individually-considered loci. The smallest genome in the set, that of the cicada endosymbiont Hodgkinia cicadicola [12] demonstrates no difference at all in average conservation between OGs and individual loci. All genomes greater than 3 Mb had higher average conservation among individually-considered loci than among OGs.
In this study, we used the literature-curated set of EcoCyc E. coli protein complexes and the protein complexes isolated by Hu et al. [5] as a set of experimentally-determined complexes for E. coli (Fig. 2A). The set of experimentally-determined Mycoplasma pneumoniae complexes identified by Kühner et al. [6] was also included in the comparison as a distantly-related, minimal set. Though these datasets differ in content and approach, both E. coli data sets contain about 300 complexes. Most complexes in the EcoCyc set contain from 2 to 4 unique proteins while the Hu set contains a comparatively higher number of complexes (more than 30) containing 5 or more unique protein components (i.e, unique proteins mapping to different orthologous groups). Note that some of the Hu et al. complexes appear to represent subsets of full complexes (i.e., the full ribosome constitutes a single complex in EcoCyc but is represented by several complexes in Hu et al.). Also, the EcoCyc set is partially redundant (i.e., each RNA polymerase holoenzyme is represented as a different protein complex, as are the F1 and F0 subregions of ATP synthase).
The size of the complexes within the data set produced by Kühner et al. appears to differ in distribution from those characterized by Hu et al. (Fig. 2A). Specifically, most M. pneumoniae complexes with two or more unique members contain just those two unique proteins. The cross-species discrepancy may also result from methodology, though Kühner et al. suggest it is representative of authentic biological differences between the two species. M. pneumoniae contains fewer unique proteins than E. coli does and this difference limits the number of unique proteins seen in any single complex.
The exact protein complexes defined by each data set differ. Pairwise comparison of presence or absence of proteins in each complex is improved by mapping components to orthologous groups but few complexes appear to be present in an identical form across all three data sets. Fig. 2B provides four examples of the types of complex matches seen across the data sets. For instance, the DNA polymerase III holoenzyme (EcoCyc: CPLX0–3803) contains 9 unique proteins as per EcoCyc but its closest match in the Hu set contains 7, including two proteins not found in any EcoCyc complex. The “missing” proteins from the EcoCyc complex are found in other Hu complexes. The Hsp70 chaperone complex (EcoCyc: HSP70-CPLX) provides another example: The M. pneumoniae complexes provide a better match for the EcoCyc complex than the Hu set does. Topoisomerase IV (EcoCyc: CPLX0–2424) has a good match in all three data sets though the representative Hu complex contains an additional protein. Lastly, RecBCD serves as an example of a good E. coli-specific match with no components present among the M. pneumoniae complexes.
In the aggregate, most EcoCyc complexes do not have reliable matches in the other experimental sets (Fig. 2C). Using all 285 EcoCyc complexes as a guide, their best matches in the other sets are classified as “good” if they contain at least half of the same unique proteins (as members of orthologous groups) or “poor” if they contain a match of less than half of the EcoCyc complex’s components. No complex of a size greater than 4 unique proteins has a good match in both the Hu et al. and Kühner et al. complex sets. 28 complexes (9.8%) of the complexes of size 4 or less have good matches in both sets. The majority of the complexes in this size class (153 out of 246) contained at least one matching component in the Hu E. coli complexes but no match among the Kühner et al. M. pneumoniae complexes.
The set of M. pneumoniae complexes serves as a rough model for the complexes most commonly found across bacterial species (see S9 Table and S10 Table) for the predicted conservation of each complex). It is an imperfect model: out of 116 complexes, only 28 are fully conserved (that is, each of their components are present as orthologs) in the 7 other model species in this study. 39 M. pneumoniae complexes appear to share at least 2/3 of their components with all the other species, though 75 complexes share at least half. Just one complex contains components entirely specific to M. pneumoniae (complex 87, composed of uncharacterized proteins Mpn036 and Mpn676, respective UniProt entries P75078 and P75116).
The variability between the EcoCyc and Hu datasets has a direct impact on the usefulness of these complexomes as models for other bacterial species. In any case, the EcoCyc and Hu complex sets provide the most comprehensive complex set currently available for E. coli. The intersection of the two sets (Fig. 3A) is indeed limited: just 576 unique orthologous groups are shared between the sets and just 132 complexes appear to be “good” matches between the sets. Using these 132 complexes as a model for those in P. aeruginosa shows that up to 120 of the complexes may be conserved based on orthologous components present in the P. aeruginosa genome. If the yet-uncharacterized P. aeruginosa complexome contains roughly the same number of complexes as those for E. coli then this prediction method misses more than half (that is, around 150) of the potential complexes unless we also use the unique complexes of each set. We used these results as evidence that the data sets should be used as independent models rather than as an intersecting set: losing more than half of the potential model complexes simply due to inconsistencies across data sets may be too limiting for a broad cross-species comparision.
Fig. 3B displays distributions of protein complex conservation across four bacterial species other than E. coli. (M. pneumoniae complexes were not used in this comparison.) These plots provide the median and interquartile range of protein complex conservation fractions in each species, using either EcoCyc or Hu et al. complexes as a model of the complex set. A comprehensive set of protein complexes has not been identified for any of these species as of yet. Following the results shown in Fig. 1, however, we may predict that most bacterial protein complex component sets should share at least half of their OGs with all other bacterial genomes, on average. Basic biology also plays a role here: we expect a subset of crucial protein complexes like polymerases to be well-conserved across all species. The set of all EcoCyc complexes, appears to be highly-conserved in P. aeruginosa (the entire interquartile range lies between full and 75% complex conservation, showing the average EcoCyc complex is well-represented in P. aeruginosa) but shows a greater range of conservation across the three other species. The Hu complexes show lower complex conservation median values than EcoCyc for all but H. pylori and lower variability for all but P. aeruginosa. Here, the median values are not as useful as the conservation ranges: the distance between the highest and lowest values includes every possibility from 0 to 100% conservation using either model of E. coli complexes. We see that the two species most closely related to E. coli in this set—P. aeruginosa and C. crescentus—produce different median values and interquartile ranges between the sets across all protein complexes. Components of complexes in the two E. coli sets, used as models, are clearly conserved differently across bacterial species. A higher-resolution comparison is necessary to determine which complexes are highly-conserved.
Although the size distribution is different in E. coli and Mycoplasma, we hypothesized that homologous complexes should be very similar, both in size and composition. However, this is not true: few complexes share even half of their components across the data sets (Fig. 2C). The majority of complexes shows less than 50% overlap between EcoCyc and Hu, but also between Hu and Mycoplasma. This suggests that there are both technical (E. coli vs E. coli) but also biological reasons (E. coli vs. Mycoplasma) for these differences.
To get a more global yet more detailed picture of protein complex conservation, we compared conservation across 8 bacterial species, including the two species for which full protein complex sets exist. The EcoCyc complex set was used as a standard to which all other species were compared. Fig. 4 provides three examples of the ways protein complexes may or may not be conserved across species. Conservation of protein complexes may be roughly grouped into three categories: well-conserved complexes, complexes with a core set of proteins conserved, and those in which no core set appears to be consistently conserved. As conservation and essentiality may be related to paralogy, we also compared the components of these complexes on the presence or absence of paralogs.
It is commonly assumed that highly conserved proteins must be important and thus should be essential in many cases. Interestingly, this is often not true (Fig. 4). For example, the well-conserved succinate dehydrogenase components are essential in only 3 of the species shown. The four components of this complex (as defined by the default structure in E. coli) are present only in Pseudomonas aeruginosa and Caulobacter crescentus. Helicobacter pylori and B. subtilis encode 3 out of 4 components and the other 3 species appear to have lost the entire complex. Similarly, the Bam outer membrane protein assembly complex (EcoCyc: CPLX0–3933) shows partial essentiality across the complex in 4 species though its components are well conserved in only 3 species. This complex has a similarly patchy pattern of conservation, with any number from zero to all 5 components conserved. In the case of H. pylori Bam complex, what initially seems like a lack of conservation may be the result of component replacement by functionally similar proteins [14]. By contrast, F1 ATP synthase is conserved in all species examined. These examples show that most complexes are less well conserved than their often important functions indicate (as measured by the presence of essential proteins in these complexes).
Fig. 5A displays all EcoCyc E. coli complexes with at least one component present in M. pneumoniae. In this case, fraction of essentiality (the number of protein components found to be essential out of all protein components present) is shown. Fig. 5B displays conservation fractions of all E. coli complexes with at least one protein conserved in M. pneumoniae, though not necessarily present in a M. pneumoniae complex. A complete survey of all EcoCyc complexes across these species in terms of conservation and essentiality is provided in S1 Fig. and S2 Fig., respectively. Conservation fraction was established as the fraction of unique proteins in a defined complex present in the target species. Notably, proteins of only 21 complete EcoCyc complexes are fully conserved across all 8 species, or just 15 complexes when subunits and alternate forms (i.e., RNA polymerase with different sigma factors) are removed. An additional 19 complexes are fully conserved across all species but the two Mycoplasma species. The remaining complexes vary extensively in their degree and extent of conservation. A number of complexes are well conserved across E. coli, P. aeruginosa, C., crescentus, H. pylori, and B. subtilis but not S. sanguinis or the Mycoplasma (e.g. succinate dehydrogenase, EcoCyc: SUCC-DEHASE). Overall, of the 176 EcoCyc complexes of 3 or more unique proteins (S1 Fig.), 128 appear to have lost at least one unique protein component in one or more species. This demonstrates that protein complexes are far more flexible in evolutionary terms than previously assumed.
Protein complex function varies in a similar way as conservation (Fig. 5B). As expected, many of the most highly conserved complexes are directly involved in DNA replication, transcription, or translation. Many protein complexes of varying conservation fractions are transport complexes—as bacterial membrane structures vary across species, some degree of transporter component evolution is also expected. At least six distinct complexes involved in DNA modification or repair demonstrate less than perfect conservation.
E. coli complexes serve as a “gold standard” for protein complexes across bacteria only in cases where most or all of the components of a complex are broadly conserved. This property is true of just a small fraction of complexes. Fig. 5D displays conservation fractions for all 285 E. coli complexes in the EcoCyc set, clustered by similarity of their conservation patterns across the 7 other species used in this study. Just 21 complexes appear to be fully conserved (that is, orthologs of each of their components are present) in all other species. This is a broad taxonomic range, so a more relaxed cutoff may be appropriate to predict a complex is conserved; even so, only 28 complexes contain at least 2/3 of the E. coli components across all species. Lowering the cutoff to at least half of the E. coli components still yields only 34 complexes. The lack of broad conservation is not, however, a matter of full complex presence or absence across species. Rather, many complex components appear to be conserved independently from other members of their complex. Similarly patchy conservation can be seen for essentiality (Fig. 5C), as the most broadly well-conserved complexes (far left) generally retain essentiality across species but less consistently-conserved complexes do not, though they may retain essentiality while appearing to lose complex components.
E. coli is frequently used as a model organism for bacteria in general. Using the literature-curated set of protein complexes from EcoCyc, we sought to determine how well this protein complexome serves as a model for complexes in other bacterial species. A comparison of the fractional conservation of each EcoCyc complex across 894 different bacterial genomes was the result (Fig. 6; see S5 Fig. for an expanded version). The genomes in this comparison were arranged as per NCBI taxonomy definitions, revealing patterns in complex conservation closely corresponding to numerous taxonomic boundaries. Hierarchical clustering of each E. coli model complex (specifically, UPGMA) on the basis of its fractional conservation across all other species reveals groups of complexes with similar patterns of predicted conservation.
The species with the most overall conservation of the E. coli complexes are, unsurprisingly, those most closely related to E. coli. Roughly a third of the complex set is conserved across all species with the minimal Rickettsia and Mycoplasma genomes, among others, serving as notable exceptions. The middle third shows the most difference in conservation between the Proteobacteria and all other species. The Lactobacillales show the most difference in conservation among these complexes, to the degree that they resemble Cyanobacteria more closely among this subset. The last third (far left of Fig. 6) of the complexes demonstrate the most variable conservation across all species. Many of these complexes are missing or partially conserved among the Proteobacteria yet are fully present in many Firmicutes species and even in extremophiles like Thermus or Thermotoga species. Overall, out of 285 EcoCyc complexes, 12 (~4%) have at least one component present in all 894 bacterial genomes in the set. None are perfectly conserved across all genomes but 14 complexes appear to be conserved across at least 95% of the genomes. If potential complex conservation is generously defined as conservation of at least half of the complex components, 3 EcoCyc complexes are potentially conserved across all 894 genomes, 25 are potentially conserved across 95% of the genomes, and 186 are potentially conserved across at least half of the genomes. Variance across the full set of complex conservation fractions is 0.189. Because conservation of these complexes follows the existing taxonomy well, some generally well-conserved complexes like RNA polymerases may be missing from entire genera.
The experimentally-determined protein complexes identified by Hu et al. were also used as a model of the E. coli complexome (S6 Fig.). Most complexes appear to have partial conservation across nearly all species using this model. Distinctions are still seen among the minimal genomes of the Rickettsiales as well as the Mycoplasma and the genomes of related species. Out of 310 Hu et al. complexes, 16 (~5%) have at least one component present in all 894 bacterial genomes in the set. As with the EcoCyc complexes, none are perfectly conserved across all genomes but a single complex (complex 271) appears to be conserved across at least 95% of the genomes. Using the same 50% cutoff for potential complex conservation as used above, no Hu complexes appear to be conserved in all 894 genomes, 10 are potentially conserved across 95% of the genomes, and 182 are potentially conserved across at least half of the genomes. Though these Hu et al. complex values appear similar to those for the Ecocyc complexes, variance across the full set of Hu complex conservation fractions is 0.097, indicating less variability among the values than that seen for the EcoCyc complexes. This lesser variance can also be seen in the surprising consistency across taxonomic lines (S6 Fig.).
Both the literature-curated EcoCyc model and the Hu et al.-based experimental model were evaluated by comparision to a randomized version of their respective components. For the literature-curated model, Pearson correlation was 0.185, while for the experimental model, Pearson correlation was 0.293. The higher correlation value for the experimental model indicates it is closer to a random distribution of complex correlation fractions across the species set. We do not expect complexes to be conserved in a random pattern so this may indicate the Hu et al. complex set is less useful than the EcoCyc complex set for prediction across this wide range of genomes.
Mycoplasma species have highly reduced genomes and it is generally assumed that they have retained mostly essential proteins. In fact, the fraction of conserved essential proteins is much higher when comparing Mycoplasma pneumoniae to E. coli than vice-versa (Fig. 7). In these comparisons, all complex components are searched for in full genomes and essentiality is assigned based on the target species. Among the full set of Hu et al. E. coli complexes, complexes have an average conservation fraction of 0.198±0.230 and an average essentiality fraction of 0.122±0.196 in M. pneumoniae. High variability in conservation among complexes is expected as complex components, like single proteins, are subject to a broad variety of evolutionary pressures. Among the 53% of complexes with at least one component present in M. pneumoniae, the average fractions increase to 0.375±0.184 and 0.231±0.218, respectively. Among the full set of Kühner et al. M. pneumoniae complexes, complexes have an average conservation fraction of 0.716±0.292 and an average essentiality fraction of 0.32±0.332 in E. coli. Among the 95% of complexes with at least one component present in E. coli, the average fractions increase to 0.755±0.245 and 0.337±0.332, respectively. Overall, Mycoplasma protein complex components are more likely to be present and essential in E. coli than E. coli protein complex components are in Mycoplasma.
One possible explanation for the lower fraction of conserved essential proteins in E. coli is the presence of paralogs that renders duplicate genes non-essential, given the presence of an additional copy with a redundant function. We performed comparisons of the fraction of conservation of each complex and its sum of paralogy (that is, the total number of all copies of all genes coding for the complex components in the target species). As the number of paralogs for each gene was broadly defined using orthologous groups, these numbers are considered maximum possible values rather than specific counts of known paralogous regions. We observed an inverse trend between E. coli complexes vs. M. pneumoniae (S3A Fig.) and vice versa (S3B Fig.): the more paralogs they have in E. coli the less conserved these proteins were in Mycoplasma and vice versa. More specifically, E. coli complexes with a conservation fraction greater than 0.6 in M. pneumoniae all had total paralogy sums lower than 40 though more poorly-conserved complexes had paralogy sums between 2 and about 100. M. pneumoniae complexes with a conservation fraction greater than 0.6 in E. coli had a range of sums of paralogy between 2 and nearly 80. The more poorly-conserved complexes all had paralogy sums of 60 or less. Calculated Pearson anti-correlation for E. coli complexes vs. M. pneumoniae (S3A Fig.) was -0.04 and Pearson correlation for M. pneumoniae complexes as a model for E. coli (S3B Fig.) was 0.05, indicating limited to no overall correlation in either full comparison. As is the case with conservation of complexes across all species (Fig. 6), correlation may be case-specific.
The fraction of essential components in protein complexes is non-random and may be greater than expected, depending upon the complexes compared (Fig. 8). When compared to random assortment, Hu et al. E. coli complexes have more essential proteins than expected by chance (Fig. 8A). A Spearman anti-correlation of -0.25 was found. E. coli complexes from EcoCyc (Fig. 8B) demonstrate similar trends, with a Spearman anti-correlation of -0.22. M. pneumoniae complexes from Kühner et al. (Fig. 8C) show a trend of declining essentiality compared to randomized essentiality fractions of 0.6–08. A Spearman anti-correlation of -0.03 was found for this M. pneumoniae complex set. Both E. coli anti-correlations show a weak relationship.
Paralogy was also examined as a function of essentiality (S4 Fig.). Here, average paralogy values were determined for each complex to minimize the impact of complex size, especially as only one or two components of a complex may be essential. E. coli complexes from Hu et al. (S4A Fig.) decrease in average paralogy as their fraction of essentiality decreases. In total, compared to random assortment, far more E. coli complexes than expected appear to have essentiality fractions of 0.4 or more. E. coli complexes from EcoCyc (S4B Fig.) demonstrate similar trends. M. pneumoniae complexes from Kühner et al. (S4C Fig.), however, do not appear to retain the same relationship between essentiality and average paralogy. Additionally, more M. pneumoniae complexes than expected were found to have essentiality fractions of 0.2 or less while fewer than expected had essentiality fractions greater than 0.6. Spearman anti-correlation was not statistically significant at -0.03. Overall, essentiality and average paralogy appear to be related for E. coli but not for M. pneumoniae complexes, probably because M. pneumoniae contains relatively few paralogs. Example complexes from each of these sets and the OGs shared between them are provided in S4D Fig.; in each example, at the majority of the complex components are essential but their representative genomes contain few paralogs coding for redundant complex components.
Protein complexes are attractive targets for functional analysis, given that proteins are embedded in a functional context. This is especially true for proteins of unknown function that are part of a complex (Fig. 9A, B). Here, conservation is defined as greater than 0.5 conservation fraction and essential complexes are those with at least one essential component in the target species. Among the highly conserved components, many are essential in 4 or more of the 8 species. Using more than one species reduces the effect of noise and inconsistency across essentiality screens. Starting with 39 EcoCyc-defined complexes that contain unknown proteins, at least 15 appear to be conserved in all other species in this study but the Mycoplasma. Fig. 9C displays example complexes for the Hu (E. coli) and Kühner (Mycoplasma pneumoniae) complexes, respectively. Unlike in parts A and B, the complexes shown are experimental results rather than literature-defined complexes. Each complex contains at least one component of unknown or unclear function, whether in the context of the protein complex or broader cellular function. For instance, complex 66 from Hu et al. (Fig. 9C) consists of 6 unique proteins of which 3 are of unknown function (or remain without annotation). Of the 6 proteins, 3 are highly conserved and 1 of those three is frequently essential. The E. coli protein MraZ, present in Hu complex 149, is shown here as a protein of unknown function but was recently found to be a transcriptional regulator involved in multiple pathways [15]. More than 149 Hu et al. E. coli complexes and 34 Kühner et al. Mycoplasma pneumoniae (183 in total) complexes contain at least one component of unknown function. Of these, 109 Hu et al. E. coli complexes and 19 Kühner et al. M. pneumoniae complexes contain components highly conserved as essential proteins. The full list of experimental complexes with unknown components is available in S11 Table.
The substantial variation among protein complexes across species supports the notion that these complexes are much more malleable than previously thought. A possible explanation of this is that the function of a complex is more important than its content. Complexes can serve the same role yet contain different proteins and when one function is lost, others can fill in the gap. Other studies have found that functional redundancy can lead to variation and that there is little overlap in terms of protein interaction among species [2,3]. While mutational change in a protein complex may have catastrophic potential, complexes are not immutable. In fact, several complexes that are essential in some species have varying composition in other species. For instance, 5 out of 9 components of the E. coli Sec translocation complex (EcoCyc: SEC-SECRETION-CPLX) are well-conserved across species from P. aeruginosa to M. genitalium. One of these components, SecA, has been found to be essential in all species focused on in this work with the exception of S. sanguinis; orthologs of this protein are present in all 894 bacterial genomes examined. The remaining 4 E. coli components are more variable in conservation across species. For instance, YajC is present in 727 out of the same 894 genomes. Strong selection pressure seems to avoid mutations that render the entire complex ineffectual. This may explain why we have observed a higher level of conservation for protein complex components than for proteins in general (Fig. 1).
The essential “core” components of protein complexes may be conserved across taxonomic levels while “accessory” components may not [1]. Given their multiple interactions, proteins within protein complexes should not only be more highly conserved than “un-complexed” ones, but should retain their essential roles if their fellow complex members are present [16,17].
Components of protein complexes are, on average, more likely to be present in other bacterial species than proteins not in complexes [1]. This is a result of high conservation among sets of large, essential complexes. 128 out of 285 literature-verified E. coli protein complexes are fully present in B. subtilis, 30 of which are also completely present in M. genitalium. For instance, all components of the ATP synthase complex (EcoCyc: ATPSYN-CPLX) are present in all species examined, though they do vary in essentiality. B. subtilis essentiality screens found no essential genes in ATP synthase, while those for M. pneumoniae found all but one component to be essential. Other complexes—predominantly those with transmembrane domains and/or transporter functions—are more variable in both conservation and essentiality, though they provide examples of how dispensable accessory proteins may be.
Some protein complexes with essential functions in E. coli may not be present in other species. The lipopolysaccharide transport complex (EcoCyc: CPLX0–7992) serves as an excellent example: all seven of the Lpt proteins in this complex have been found to be essential in E. coli though their conservation is limited to other Gram-negative species including C. crescentus and P. aeruginosa. We found that most transmembrane protein complexes follow this pattern. Interestingly, species with partial complex component conservation vs. E. coli may highlight situations in which core elements of a complex are conserved but have been modified to carry out other functions or adapted to special physiological circumstances. For example, 3 out of 4 of the succinate dehydrogenase complex (EcoCyc: SUCC-DEHASE) components in E. coli are also present in B. subtilis but not at all in S. sanguinis. This is an especially interesting example as two of the components, SdhC and SdhD, are inner membrane proteins, though only SdhC is present in the three-component B. subtilis succinate dehydrogenase. We conclude that membrane proteins and their complexes are particularly malleable, given their role in signaling and transport which reflects adaptations to specific environments and the nutrients present in them.
Smaller and more reduced bacterial genomes (that is, relative to E. coli) appear to code for a greater fraction of highly-conserved protein complexes. This conservation is evident in comparisons of the Mycoplasma pneumoniae protein complexes. In an examination of these protein complex components across more than 800 bacterial genomes, we found that species such as M. pneumoniae offers a better model of the protein complexes most critical to bacterial life. Protein complexes observed in M. pneumoniae may not only have retained a core set of functions but also utilized a higher degree of multifunctionality among its metabolic enzymes [18,19].
Surprisingly, many essential proteins are poorly conserved and essentiality itself is often not conserved across species (Figs. 4 through 6). This suggests that many functions can be replaced by non-homologous displacement [20] and that genomes are more malleable in evolutionary terms than previously expected. Clearly, this evolutionary flexibility has contributed much to the success of microbes to populate all possible environments on the planet. Variability in complex conservation highlights a limitation with this study: we are unavoidably limited by the availability of sequenced bacterial genomes. Newly-characterized genomes may reveal additional variation or consistency among protein complexes even if they are highly reduced in other respects.
As with their protein components, individual complexes reveal underlying evolutionary processes (Fig. 6 and S5 Fig.). The most highly-conserved complexes are those with functions critical to microbial life, including transcription, translation, and transcript degradation. Though different RNA polymerase (RNAP) holoenzymes (that is, RNA polymerases with different sigma factors) were considered as distinct complexes in this study, all bacterial species unsurprisingly retained at least one type of RNAP. The ribosome (EcoCyc: CPLX0–3964) is also well-conserved though its size and high level of conservation may obscure cross-species differences.
Variable conservation of some complexes is visible even among the Escherichia genomes. CPLX0–7909 (the RnlA-RnlB toxin-antitoxin complex) only appears to be present in K-12 E. coli but also in single species of Shewanella and Photobacterium. This toxin-antitoxin system has a role in bacteriophage resistance in E. coli [21] but it is unclear if this function may be retained in distantly related bacteria. CPLX0–2001 (the ferric dicitrate transport system) provides an example of more gradual change. This complex spans the membrane, suggesting its conservation should be membrane-dependent. This appears to be the case as it is well conserved across most Proteobacteria (except the Rickettsiales and Buchnera species) yet is poorly-conserved across most of the species traditionally considered Gram positive. A subset of complexes, including CPLX0–1163 (HslVU protease) and ABC-56-CPLX (aliphatic sulfonate ABC transporter), fit a strict co-conservation model: these complexes are almost always present in their full form rather than as a fraction of the E. coli model complex. These complexes are exceptions rather than the rule. Using E. coli as a model, few complexes are conserved perfectly across a wide range of species; in fact, most complexes are fractionally conserved.
All data management was performed using in-house Python scripts (SPICEDNOG; available at http://github.com/caufieldjh/spicednog). Statistical analysis and clustering was performed using R package vegan [22].
The full set of protein complexes from Escherichia coli K-12 W3110 as defined by Hu et al. [5] was assigned membership in orthologous groups (OGs) from version 3 of the eggNOG database [10] such that each protein in a complex was assigned to a single OG. The remaining loci were referred to using their original locus identifiers (in this case, their b-codes) and were retained for all further analysis. The process was repeated for all protein complexes isolated by Kühner et al. [6] from Mycoplasma pneumoniae M129 and for E. coli protein complexes defined by the EcoCyc database [13]. A representative set of six other species (Bacillus subtilis 168, Caulobacter crescentus, Helicobacter pylori 26695, Mycoplasma genitalium G37, Pseudomonas aeruginosa UCBPP-PA14, and Streptococcus sanguinis SK36) for which whole-genome gene essentiality data was selected for in-depth analysis. This species set is referred to as the focused set. Lists of all protein-coding loci for each species were obtained using the respective full proteome sets from UniProt (see S4 Table for taxonomy IDs corresponding to all genomes used). Essentiality data was collected from the Database of Essential Genes [11]. Protein structures were obtained from the Protein Data Bank (www.rcsb.org, [23]) and are referenced where used.
A set of 894 species, referred to as the large set, was also prepared using every bacterial species present in eggNOG v.3 and in the NCBI Taxonomy database [24]. The trees shown in Figs. 4, 5, and 6 are cladograms intended to show the general relationship between species within context of consensus taxonomy.
Each locus in each genome was assigned to a single orthologous group (OG) as in eggNOG v.3 [10], such that all loci were assigned to a COG, a NOG, or a bactNOG, depending upon the most widely-conserved group assignment available (see Powell et al. [10] for details regarding OG levels). Next, the presence of each locus was determined across the entire set of bacterial species; a locus seen in half of all bacterial species would be assigned a conservation value of 0.5. This presence was averaged across all loci to generate a value for average locus conservation for each genome. This value was adjusted based on locus coverage in eggNOG (i.e., if only 70 percent of the loci in a genome mapped to eggNOG OGs, the average value was reduced by 30 percent.) An identical set of comparisons were performed for all loci with predicted paralogs (that is, loci with the same OG assignment) removed prior to comparison. Subsets of selected species were also prepared such that they included only loci with the same orthologous groups as those seen in the Hu et al., EcoCyc, or Kühner et al. protein complex sets. Genome sizes were retrieved from NCBI GenBank and KEGG GENOME [25].
The observed distributions of essential genes among those coding for protein complex components were obtained using protein complex sets [5,6,13], eggNOG (v. 3) [10], and the Database of Essential Genes [11] as defined above. For a single protein complex, an essentiality fraction was defined as the fraction of all genes in a complex found to be essential, out of the set of all unique protein-coding genes in the complex. The conservation scores were used to judge participation of a complex within a dataset, establishing a maximum for each species and dataset combination. Second, essentiality fraction was found by linking each essential protein to an OG. In instances where multiple proteins shared OGs but not essentiality, essentiality was considered as the primary case and the OG was counted as essential.
A random model was created for the purpose of comparing the data set to background noise. The random model, meant to represent a collection of randomly sized complexes, was populated by proteins that have been randomly assigned essential status. The complex sizes were randomly assigned a value from three to ten. Each complex was then assigned protein values of either essential or non-essential status. The probability of being essential was determined by the overall percent of essential genes within the organism, while the random model size is equal to the maximum of the species and dataset being compared. This random model was then put through the same binning process as the observed data. The mean of each bin was obtained after 10,000 replications. This results in s bins of a size that is no longer equal to the actual data set but demonstrates an appropriate background noise level for comparison purposes. The log2(Observed/Expected) values are plotted in Fig. 8 to show any significant difference between observed essentiality and expected.
The general scheme for data analysis was as follows: (1) A list of all orthologous groups (OG) was produced for each of 894 bacterial species found in the large set defined above. (2) Presence or absence of each OG was determined for all species. (3) Repeated OGs were removed from each list and step 2 was repeated. (4) The list from step 1 was used to map OGs to the components of three sets of protein complexes. The complexes were compared to search for cross-data set complex matches. Gene essentiality was also mapped to each OG in a species-dependent basis. (5) A list of 8 taxonomically-divergent species was selected and used to define fractional conservation and fractional essentiality of each protein complex.
OGs were used as the basis of comparison for similarity between data sets. Complex size was defined as the number of unique proteins isolated from a complex; i.e. a complex may contain 3 unique OGs but 4 distinct protein components, yielding a complex size of 4. For each complex, the presence of each OG within the complex was assayed in the full proteome sets of the seven other representative species. The resulting binary presence/absence values were combined to produce a value for the percent complex conservation. This value intentionally disregards any gene context similarity (that is, an OG may be present in two genomes even if neighboring genes differ between the genomes) and simply expresses the fraction of complex components which a specific genome may code for. When a target proteome did contain a specified complex component, the number of paralogs of the component-coding gene was determined as the number of proteins in the list mapping to the same OG. While further verification, may be necessary to define any of these protein-coding genes as true paralogs, we simply used the OGs (including paralogs) as determined by eggNOG.
All protein complex components were also assigned binary essentiality values using published assays specific to the species listed above. These values were used to define the essentiality fraction of each potentially conserved complex, i.e. an E. coli complex for which 80% of the components appear to be conserved in M. pneumoniae but only 60% of the components may be essential in the latter species.
A broader comparison was prepared using the list of 894 species as defined above. Genome sizes for each species were retrieved from the KEGG GENOME Database (http://www.genome.jp/kegg/genome.html, [25]). For each species, the total number of OG-mapped protein-coding loci was divided by the total number of loci to produce a value for percentage mapped. Using the list of all OGs in the species, each OG was compared with all other species to determine its conservation across Bacteria. Adjusted average locus conservation for a particular genome, CAAL(g), was calculated as:
CAAL(g)=m(∑CL(g)L(g))N
where CL is the number of genomes in which the locus is present, L(g) is the number of loci in the genome, N is the total number of genomes, and m is the percentage of loci mapped by eggNOG v.3. Values are adjusted using the fraction of loci actually mapped so unmapped loci lower the effective conservation.
An identical list of values, but with repeated OGs reduced to a single occurrence, was averaged to produce average OG conservation. This modification removes the effect of counting loci more than once when they share OGs, as may happen when two or more loci are paralogous. Adjusted average OG conservation for a particular genome, CAAO(g), was calculated as:
CAAO(g)=m(∑CL(g)O(g))N
where CL is the number of genomes in which the locus is present, O(g) is the number of unique OGs in the genome, N is the total number of genomes, and m is the percentage of loci mapped by eggNOG v.3.
Species/strains were sorted by genome size and compared to the average conservation values. For the set of all bacterial genomes, N = 943, though Fig. 1 presents the results after removing 45 genomes of very similar size and sequence. For a subset of species, the Average Locus and Average OG Conservation values were calculated using only OGs found in published protein complex data sets.
Mapping of fractional complex conservation across species was performed as follows for both the focused set (8 species) and the large set. A cladogram of all species in the set was prepared using the Interactive Tree of Life (iTOL, [26]) project as per NCBI taxonomy. All protein components were mapped to eggNOG v.3 OGs and complex size was determined as defined above. Conservation fraction of each complex in each species was defined as the number of complex component OGs shared between the model (an E. coli complex) and the target genome over the size of the model complex. Heatmaps were prepared using the R heatmap.2 function in the gplots package. Randomized models of the large set heatmaps (Fig. 5, S5 and S6 Figs.) retaining the same species order but with a randomized distribution of conservation fractions were prepared using the R function randomizeMatrix (in the picante package [27]) and the ‘richness’ null model to respect overall conservation levels.
|
10.1371/journal.pcbi.1004748 | Kernel Architecture of the Genetic Circuitry of the Arabidopsis Circadian System | A wide range of organisms features molecular machines, circadian clocks, which generate endogenous oscillations with ~24 h periodicity and thereby synchronize biological processes to diurnal environmental fluctuations. Recently, it has become clear that plants harbor more complex gene regulatory circuits within the core circadian clocks than other organisms, inspiring a fundamental question: are all these regulatory interactions between clock genes equally crucial for the establishment and maintenance of circadian rhythms? Our mechanistic simulation for Arabidopsis thaliana demonstrates that at least half of the total regulatory interactions must be present to express the circadian molecular profiles observed in wild-type plants. A set of those essential interactions is called herein a kernel of the circadian system. The kernel structure unbiasedly reveals four interlocked negative feedback loops contributing to circadian rhythms, and three feedback loops among them drive the autonomous oscillation itself. Strikingly, the kernel structure, as well as the whole clock circuitry, is overwhelmingly composed of inhibitory, rather than activating, interactions between genes. We found that this tendency underlies plant circadian molecular profiles which often exhibit sharply-shaped, cuspidate waveforms. Through the generation of these cuspidate profiles, inhibitory interactions may facilitate the global coordination of temporally-distant clock events that are markedly peaked at very specific times of day. Our systematic approach resulting in experimentally-testable predictions provides insights into a design principle of biological clockwork, with implications for synthetic biology.
| Sleep/wake cycles in animals exemplify daily biological rhythms driven by internal molecular clocks, circadian clocks, which are important for plant life as well. The plant circadian clock is highly complex, eluding our understanding of its design principle. Based on the computational simulation of Arabidopsis thaliana, we successfully identified a kernel of the plant circadian system, the critical genetic circuitry for clock function. The kernel integrates four major negative feedback loops that process molecular circadian oscillations. Surprisingly, the plant clock circuitry was found to be overwhelmingly composed of inhibitory, rather than activating, interactions among genes. This fact underlies plant circadian molecular profiles to often exhibit sharply-shaped, cuspidate waveforms, which indicate clock events that are markedly peaked at very specific times of day. Our work presents experimentally-testable predictions, with implications for synthetic biology.
| A variety of living organisms on Earth features built-in molecular clock machineries that control the organism’s daily activities [1]. These internal time-keepers, circadian clocks, generate endogenous oscillations of gene expression with ~24 h periodicity, enabling the anticipation of diurnal environmental variations and the coordination of biological processes to the optimal times of day. Examples of such biological processes include sleep/wake cycles in animals, emergence from the pupal case in fruit flies, spore formation in fungi, and leaf movements in plants [2–4]. Disruption of circadian rhythmicity is associated with a wide range of pathophysiological conditions, indicating the importance of clock functions in homeostasis [5–8]
Compared to other organisms, such as fungi, insects, and mammals whose circadian systems have been well studied, a molecular understanding of the plant circadian system is still elusive. Numerous molecular and genetic approaches using Arabidopsis thaliana have facilitated the discovery of more than 20 plant clock genes as well as their regulatory interactions [1, 9, 10]. The emerging picture from this effort suggests that the core regulatory circuit of the plant circadian system is more complex than in other organisms [9, 11–13]. The apparent complexity of the plant clock machinery raises a fundamental question: are all the regulatory interactions between clock genes equally necessary for the establishment and maintenance of plant circadian rhythms? In other words, can we distinguish more important from less important regulatory interactions for normal clock functioning? Answering this question involves an attempt to prioritize our focus amongst numerous regulatory interactions, in order to simplify a global view of, and thereby elicit an essential principle of, the plant clock organization. Despite the fundamental importance of this issue, a satisfactorily systematic approach has not been taken yet; thus, this topic is the focus of our study. In the case of other biological processes, finding essential subnetworks out of the whole has been of wide interest for both scientific and engineering purposes [14–18].
Properly designed experiments may be one way to address this issue, but often require laborious and costly efforts. Complementary to experiments, mathematical models help biological findings by predicting the effects of genetic and non-genetic perturbations, where experimental access could be limited or unavailable. Utility of mathematical models has been well documented in earlier studies of circadian rhythms [19–22]. An initial mathematical model of the plant circadian system was constructed based only on three genes, LATE ELONGATED HYPOCOTYL (LHY), CIRCADIAN CLOCK ASSOCIATED 1 (CCA1), and TIMING OF CAB EXPRESSION 1 (TOC1) [22]. This model has evolved to include five times more components to date [23, 24]. Additionally, models that incorporate the downstream targets of the core circadian system are starting to gain attention [25]. These models have certainly served a significant role in enhancing our understanding of the plant circadian clock. Nevertheless, to the best of our knowledge, none of these studies has fully attempted to specify the functionally essential interactions between clock genes in a systematic and comprehensive way.
Central to our approach to the plant circadian system is the concept of a kernel. We define a kernel as a collection of minimal functional sets, each comprising all molecular components (genes and gene products) in the system and only a part of their regulatory interactions, which must be present to generate the temporal trajectory of molecular concentrations close to wild type (WT). In this definition, we refer to a collection of minimal sets to cover cases with multiple minimal sets. Based on an Arabidopsis clock model constructed in this study, our analysis shows that the kernel structure combines four negative feedback loops whose interplay effectively accounts for circadian rhythmicity in Arabidopsis. Strikingly, the kernel structure, as well as the whole clock circuitry, was found to be overwhelmingly composed of inhibitory interactions between genes. We subsequently present a mechanistic reason for the prevalence of such inhibitory interactions in the plant clock. These results provide a systematic and unique view of the plant circadian oscillators, with experimentally testable predictions to enhance our understanding of biological time.
We began by constructing a mathematical model of the core circadian oscillator in plant Arabidopsis thaliana. For this model construction, we applied system identification techniques to publicly available time course data of mRNA and protein expression (Materials and Methods). The resulting model consists of 24 ordinary differential equations (ODEs), describing a rate of a concentration change of each mRNA, protein, or protein complex (S1 Text). Experimentally-verified molecular interactions were primarily incorporated in the model, which then contains a total of 40 transcriptional and post-translational interactions between components, along with light-dependent regulations. Fig 1A shows a global architecture of the core gene circuit considered in our model.
In comparison with previous models [23, 24, 26], the new model is mainly based on the model (P2013) by Pokhilko et al. [23], but we filtered out hypothetical or outdated molecular interactions and adopted some recent findings [24]. Compared to our earlier work [26], which uses a discrete-time model for control design purposes, here we have constructed a continuous-time model, with revised interactions compatible with recent knowledge. Full details of the model comparisons are presented in S1 Text. Overall, we stress that our current model does not intend to outperform other existing models in its accuracy through the inclusion of all up-to-date information. Rather, the priority was to construct a model which is compact, yet biologically relevant, in accordance with recent experimental knowledge. We expect that this model is suitable enough for our main purpose of kernel identification, without further sophistication of the model structure.
Because we are ultimately moving forward to identify the kernel structure responsible for circadian rhythms in WT plants, time series data of mRNA and protein expression from WT, not from mutants, were used during model construction. Mutant data were used only to validate the constructed model, as will be described later. Specifically, we estimated the parameters of the model by fitting the simulation results to WT mRNA and protein expression profiles over time, under five different light conditions: equal length light-dark cycle, i.e., 12 hours of light and 12 hours of dark (12L:12D), 16 hours of light and 8 hours of dark (long day), 8 hours of light and 16 hours of dark (short day), constant light (LL), and constant dark (DD). These expression profiles were obtained from publicly available experimental literature and databases (S1 Table). Because the absolute levels of mRNAs and proteins were difficult to ascertain from their sources, we normalized the expression levels into dimensionless values (≤1) with arbitrary scales. As a proxy for the LHY/CCA1 information, we adopted the LHY expression data, because they were often better in the quality than CCA1’s. Constraining the model output to fit all these datasets gave rise to a total of 97 estimated parameters of the model equations, along with 51 coefficients that scale each light condition’s mRNA and protein levels relative to the levels under 12L:12D cycles (see S1 Text). Our model does not separate nuclear from cytosolic proteins [27, 28], due to incomplete availability of the relevant expression data and to avoid increasing model complexity.
What is the resulting performance of our model (MF2015)? We found that MF2015 captures well the overall temporal patterns of gene expression from WT (Fig 1B–1G; for comparison with P2013, see S1 Text). Also, the free running rhythms in WT are in good agreement with experimental values [29, 30]: 25.2 h (model) and 24.6 h (experiment) in LL, and 25.8 h (model) and 25.9 h (experiment) in DD. However, these results cannot validate MF2015, because we estimated the model parameters from the WT data. To directly test the predictive power of the model against an independent dataset, we computed the altered rhythmicity under different genetic perturbations. The simulated mutants are 76.2% accurate when the clock periods are quantitatively compared to experimental values (see S1 Text). Qualitative agreement (lengthened period, shortened period, or arrhythmia) is observed for 85.7% of the simulation outcomes and experimental results (S1 Text). Moreover, the simulation predicts the substantial elevation (reduction) of ZEITLUPE (ZTL) protein levels in LL (DD), matching the experimental finding [31]. This result is the first accurate reproduction of ZTL performance through computational modeling (S6 Fig). Taken together, MF2015 is greatly supported by an array of experimental evidence in terms of its predictability. Note that P2013 yields the simulated mutant periods in 42.9% quantitative agreement with experimental values.
In general, the simulation outcomes were robust to a wide range of kinetic parameter variations and transient molecular concentration changes (S1 Text). A few exceptions that convey the system’s sensitive response involve the variations of parameters in PSEUDO RESPONSE REGULATOR 5 (PRR5) mRNA degradation, EARLY FLOWERING 3 (ELF3) inhibition by LHY/CCA1, and light-responsive protein production. Whether they represent genuine biological factors or model incompleteness is unknown. Meanwhile, the overall robustness to parameter variations indicates the presence of multiple parameter sets for the model. Interestingly, alternative parameters that we examined did not make much of an improvement in the predictability of mutant period lengths (S1 Text). Moreover, such alternative parameters of the model are unlikely to change the main results of our study, as kernel identification and analysis involve parameter re-optimization processes.
Our modeling of the core circadian system (MF2015) encouraged us to address difficult mechanistic questions. Among all 40 molecular interactions and light regulations in the system, which interactions (and light regulations) are minimally necessary to shape the circadian mRNA and protein expression profiles observed in WT across different light conditions? We refer to this collection of minimal sets as the kernel of the circadian system. In the next paragraph, both molecular interactions and light regulations are referred to simply as interactions.
Sheer screening of interaction sets, where removal severely distorts clock rhythmicity, would not be sufficient to identify a kernel structure. If this distortion is repaired by a readjustment of kinetic parameters, the removed interactions are not likely to be essential in their network-topological properties; rather, their knockout effect is simply dependent on specific parameters. Therefore, the knockout effect in distorting clock rhythms should be double-checked with re-optimized parameters. If the knockout effect remains severe even after parameter re-optimization, the removed interactions can now be said to be essential in their network topological properties. Ideally, our kernel discovery procedure would be to search through all possible combinations of interactions, and examine the effects when the interactions in each combination are removed, followed by parameter re-optimization to best fit the WT expression profile of every clock component across different light conditions. This strategy, although ideal, is extremely computationally demanding and therefore impractical. Instead, we devised a heuristic approach that consists of the following steps (Materials and Methods, and S1 Text): first, we measure the knockout effect of each interaction on the WT expression patterns under the five different light conditions. Then, we prune those interactions from weak to strong knockout effects until discovering any single clock component that fails to produce rhythms similar to WT. Next, among the remaining interactions, we choose those with knockout effects below a certain threshold. Each chosen interaction is deleted, and parameter re-optimization follows to fit the WT expression data. If parameter re-optimization recovers the WT rhythms for every clock component, this interaction is completely removed from the system. The implementation of these steps, complemented by an additional step to allow multiple solutions, leaves a fraction of the interactions, which yet connect all the molecular components in the system. This interaction set corresponds to our estimated kernel structure. For a detailed description of the kernel identification, see S1 Text.
Using MF2015, we found that the kernel of the plant circadian system consists of 22 transcriptional and post-translational interactions and light regulations, which seamlessly involve all molecular clock components in the system. In other words, at least half of the 40 interactions/regulations in the whole system are required to form the WT rhythms across the five different light conditions. Notably, the kernel structure harbors four negative feedback loops, termed loops I to IV (Fig 2; compare with Fig 1A). In the kernel, the only negative feedback other than these four loops is the autoinhibition of the EVENING COMPLEX (EC) genes through the EC, and this effect remains localized to the EC formation and thus not our focus here. Loops I to IV host at least one of the PSEUDO RESPONSE REGULATOR (PRR) genes each, and are interlocked by having LHY/CCA1 in common: loop I includes LHY/CCA1, PRR5, and TOC1 (Fig 2A). Loop II has LHY/CCA1, PRR7, and TOC1 (Fig 2B). Loop III involves LHY/CCA1, PRR7, and the EC, along with the EC subcomponents (Fig 2C). Lastly, loop IV includes LHY/CCA1, and PRR9 regulated by light (Fig 2D). Accordingly, TOC1 interconnects loops I and II, while PRR7 interconnects loops II and III. Each of loops I, II, and III includes a cyclic structure of triple inhibitions, known as a repressilator (Fig 2A–2C) [32]. A repressilator structure can exhibit sustained oscillation under proper conditions. Of note, loop I has one more interaction added to this repressilator structure, i.e., the inhibition of PRR5 by LHY/CCA1. The direction of this inhibitory interaction is exactly opposite to the repressilator’s overall cyclic direction, and thus is supposed to be antagonistic to the oscillatory capability of the loop (see below). Among the four loops, loop IV in Fig 2D is the simplest one, having only a pair of single positive and negative connections between two morning-expressed components, coupled with light.
To our knowledge, loops I and II have not been previously described, whereas loop III recapitulates a repressilator structure previously reported [33]. Loop IV has been previously termed the morning loop [9, 34, 35]. Therefore, our unbiased and systematic approach to kernel identification does not only recover previously characterized gene circuits (loops III and IV), but also suggests new circuits (loops I and II) that may be crucial for Arabidopsis clock function.
Owing to the above kernel identification, the complex plant clock circuitry has been greatly simplified, converging on the four negative feedback loops that structure the kernel. We next considered an in-depth mechanistic analysis of the individual feedback loops as well as their interrelations.
An immediate question is, among the four negative feedback loops, which of the loops critically support the generation of autonomous molecular oscillations observed in WT. By definition, every element in the kernel must play a significant role in shaping the oscillatory profiles. However, it does not mean that their contributions to the creation of the autonomous oscillation are necessarily equivalent to each other. Moreover, the current kernel structure is a full repertoire of interactions necessary for all five different light conditions mentioned above. Clearly, only separate simulations of constant, free running conditions will answer this question for the endogenous, autonomous oscillation.
To test the capability of individual loops to generate autonomous oscillations close to WT, we simulated LL using a computational model of each isolated loop, with kinetic parameters re-optimized for the WT expression data in LL (S1 Text). Given the WT expression profiles, this parameter re-optimization was expected to reveal the maximum oscillatory capacity of each loop structure regardless of its specific MF2015 parameters. It infers a natural bound of the loop’s contribution to the WT endogenous oscillations − a natural bound imposed by the loop’s structure itself rather than by specific parameters. From this simulation, we found that loops I, II, and III in LL were clearly able to generate sustained oscillations similar to WT (Fig 3A and 3B), whereas loop IV failed (Fig 3C). In fact, if equipped with other parameters, oscillations can be maintained even by loop IV, but at the expense of its specific oscillatory patterns, in far deviation from the experimental profiles. Once loop IV undergoes a parameter adjustment to fit the experimental profiles, it loses sustained oscillation.
The endogenous oscillatory capability of individual loops I to III raises an intriguing possibility: can the plant circadian rhythm be robust to the breakage of some loop(s), if buffered by the other loop(s’) activity? To explicitly address this question, we inactivated loop I in MF2015 by blocking the inhibition of LHY/CCA1 by PRR5. Likewise, we inactivated both loops II and III simultaneously, by blocking the inhibition of LHY/CCA1 by PRR7. The MF2015 simulation of LL demonstrates that either of these two “mutations” largely restores the circadian gene expression profiles observed in WT, if accompanied by parameter re-optimization (Fig 3D). As can be predicted, the simultaneous blockage of both PRR5 and PRR7’s inhibitory actions on LHY/CCA1 in MF2015 inactivated all three oscillatory loops I to III, and thus abolished the circadian rhythmicity itself of gene expression, even when accompanied by parameter re-optimization. This prediction is well supported by an experimental report that the Δprr5/prr7 double mutant in constant conditions exhibits almost arrhythmic mRNA levels of clock-controlled genes, although each single mutant retains free running rhythmicity [36]. Moreover, the above simulation forecasts that only the removal of the two inhibitory interactions, rather than the entire double gene deletion, is necessary to cause severely abnormal clock gene expression. In sum, we find that under certain circumstances loop I can buffer the loss of loops II and III, and vice versa. Similarly, we computationally blocked PRR7 inhibition by TOC1, and that by the EC, to inactivate loop II and loop III, respectively. Again, simulated mutant outcomes suggest that loop II and loop III can buffer the loss of each other. Taken together, these results indicate complementary relationships between loops I, II, and III in the management of endogenous circadian oscillations.
While loops I to III exhibit the fundamental capacity to generate endogenous oscillations similar to WT, loop IV lacks such capability. We therefore conjectured that, among all the four loops, loop IV is unlikely to exert the strongest regulation on the clock gene expression, if these genes are regulated by the other loops as well. Indeed, the LHY/CCA1 inhibition by PRR9 (in loop IV) was consistently weaker than either the LHY/CCA1 inhibition by PRR7 or that by PRR5 (in loops I to III), throughout our simulation with various re-optimized parameters (S1 Text). Previous experimental data from LL have shown that a Δprr9 knockout has a smaller effect on LHY and CCA1 expression than a Δprr7 knockout [29]. Fig 3E shows that LHY mRNA levels, on average, increased by 60.5% and 16.7% in the Δprr7 and Δprr9 mutants, respectively, consistent with our computational prediction; a similar trend was also observed for CCA1 mRNA [29]. Despite the loop IV’s relatively weak role in free running rhythmicity, it should be noted that, in our current kernel structure, loop IV is the only negative feedback loop which senses external light stimulus (Fig 2D) and thereby contributes to the entrainment of the kernel dynamics to light. We cannot entirely exclude the possibility that more loops may come into play in light sensing of the kernel as our model becomes updated.
The efficacy of our simple kernel structure to interpret the clock dynamics is further exemplified by loop I. In addition to the basic repressilator structure, loop I holds a unique topological feature of reciprocal inhibitory interactions between LHY/CCA1 and PRR5 (Fig 2A). In particular, the inhibition of PRR5 by LHY/CCA1 is placed in opposition to the repressilator’s overall cyclic direction, and thus may retard the loop’s inherent oscillation. In fact, this retardation effect was found to affect the oscillation of the whole clock circuitry, because of the structural interconnection between loop I and the whole. For example, the simulation of MF2015 in LL demonstrates that a 20% increase in PRR5 inhibition by LHY/CCA1 slows down the circadian rhythm, resulting in a 3.3 h lengthened period, whereas a 20% decrease in this inhibition shortens a period by 2.9 h (Fig 3F). This experimentally-testable idea might be hard to conceive without the simplicity of the loop-I structure.
In the kernel, LHY/CCA1 interlocks all loops I to IV, indicating its central role in the circadian oscillator. The adverse effect of the Δlhy/cca1 double knockout on model performance is supported by experimental evidence [37, 38]. From the entire kernel structure in Fig 1A, compared with loops I to III, one can notice the presence of TOC1 inhibition by the EC. This inhibition is the only regulatory interaction with its regulated target (TOC1) in the loops, while the interaction itself is not a part of major negative feedback loops in the kernel. This fact prompted us to investigate whether TOC1 inhibition by the EC should be retained in our kernel. The simulated removal of this inhibition from the kernel apparently distorted, e.g., the LHY mRNA and TOC1 protein profiles, even when accompanied by parameter re-optimization (S7 Fig). Therefore, we keep in the present kernel structure TOC1 inhibition by the EC.
In conclusion, our model is supported by current experimental data and indicates that the plant circadian oscillator is an orchestrated interaction of mainly four negative feedback loops in the kernel. In the face of the larger complexity of the full circuitry, our simplified loop structures may offer an efficient way to understand the plant clock mechanisms, as well as predict circadian dynamics that has not yet been characterized.
Among the four major negative feedback loops in the kernel, loops I to III have the repressilator-like structures that are entirely composed of inhibitory interactions. Only loop IV includes an activating interaction. Regarding the central role of these feedback loops in circadian rhythms, why does the plant circadian system favor such inhibitor-enriched loops for its function? Indeed, recent molecular studies of the plant circadian system have indicated that inhibitory relationships outnumber activating regulations among all clock genes [39]. The full circuitry considered in MF2015 is dominated by inhibitory interactions, and this feature becomes even more prominent in its kernel structure, harboring only one activating interaction (Fig 1A). The dominance of such inhibitory interactions distinguishes the plant clock from other circadian systems, including those of mammals and fungi, which have comparable numbers of inhibitory and activating interactions [11–13].
This issue can begin to be addressed by considering that the kernel structure is designed for the production of temporal gene expression patterns close to WT (Fig 4A). Therefore, we presumed that many inhibitory regulations, at least in the kernel, may generate specific waveforms of the WT expression profiles. We do observe, in fact, that a number of Arabidopsis clock genes often exhibit particular waveforms of mRNA and protein expression (Figs 1B–1G and S1–S5). This waveform is characterized by an asymmetry between the acrophase and bathyphase, as schematized in Fig 4B: the acrophase shows a relatively sharpened peak, whereas the bathyphase can be approximated as flat. Regarding the overall acuteness around a particular peak phase, we here describe this pattern as cuspidate. For comparison, a common sinusoidal wave is not cuspidate, having a symmetrically rounded shape to the acrophase and bathyphase.
To examine the possible relevance of inhibitory regulation in cuspidate waveforms, we created a mathematical system consisting of a single transcription factor, either an inhibitor or activator, and its own target gene (Fig 4A and Materials and Methods). We formulated the model equations similar to MF2015. On the assumption that the target gene shows a near cuspidate ~24h-period expression pattern of proteins (Fig 4B), we conversely asked what specific abundance profile the transcription factor (inhibitor or activator) should have for the production of that target gene profile. Our simulation results highlight a clear difference between inhibitor and activator cases, when the target gene exhibits a cuspidate pattern (S1 Text). The inhibitor or activator tends to have a large or small phase difference, respectively, of ~8 to 12 hours or ≲4 hours with the target gene in their protein profiles, as shown in Figs 4C–4E and S8. In other words, an inhibitor (activator) and its target have a roughly antiphase-like (inphase-like) relationship. Otherwise, the target gene’s protein expression waveform will not be cuspidate but will exhibit a more smoothened profile (S9 Fig). These facts were initially observed in our simulation with simplified, yet realistic, protein expression profiles, such as that in Fig 4B. Even without such simplification, adopting empirical protein expression patterns for our simulation consistently supported the above results (S10 Fig). We also note that the cuspidate waveforms in the plant clock do not simply result from the sampling intervals of experimental data, as different interpolation methods for these data points (and the absence of such interpolation itself) gave similar profiles.
Provided that a cuspidate profile confers accurate timing of biological events around the peak phase, what is the implication of our simulation results involving the cuspidate waveform and inhibitory or activating regulation? Inhibition-induced large phase differences between the genes correspond to the global coordination of multiple clock events, distant from each other in their peak times. Conversely, activation-induced small phase differences between the genes may coordinate only the clock events nearby in time. It is possible that activating regulation might also induce larger phase differences between the genes, but would not generate cuspidate profiles in this case (S9 Fig). This fact explains why the kernel does not keep the activating regulations by REVEILLE 8 (RVE8), whose target genes have large phase differences with RVE8, yet exhibit cuspidate profiles (hence, those profiles are presumably more attributed to other regulators of these target genes). To summarize, inhibitory interactions in the plant clock seem to support the temporal coordination of distant clock events peaked at very specific times. However, it should be stressed that inhibitory interactions do not necessarily result in cuspidate waveforms in all cases. Rather, obtaining such waveform profiles requires inhibitory interactions when involving genes with large phase differences in their peak expression. Employing the terms in propositional logic, the presence of both cuspidate waveforms and large phase differences is close to a sufficient condition to implicate inhibitory regulation as their cause, but is not the necessary condition. We also note that our current definition of a cuspidate waveform is largely qualitative, based on a particular type of asymmetry between the acrophase and bathyphase. Mathematically more rigorous characterization, along with the inclusion of other possible waveforms in our framework, deserves investigation.
Within the MF2015 kernel structure, a cuspidate-waveform gene which has multiple inhibitors tends to have larger phase differences with its strongest inhibitor, consistent with our framework. For example, the transcription of a cuspidate-waveform gene, PRR5, is repressed by both LHY/CCA1 and TOC1. There is a large phase difference between PRR5 and LHY/CCA1 proteins, ~8 h compared to the ~4 h difference between PRR5 and TOC1 proteins in a 12L:12D cycle. Supportively, MF2015 suggests that LHY/CCA1 inhibits PRR5 expression ~17 times more than TOC1 (S1 Text). This fact indicates that the primary role of the PRR5 inhibition by LHY/CCA1 is to ensure the PRR5’s cuspidate waveform. Lowering the relative contribution of this inhibition (i.e., alleviating the repression by LHY/CCA1 while strengthening that by TOC1) reduces the peak-to-trough change in the PRR5 expression over time (performed under 12L:12D cycles to control for the periods of different expression profiles; see S11 Fig). Our analysis accounts well for why PRR5 inhibition by LHY/CCA1 is present in the clock, although it is antagonistic to the system’s overall oscillatory capability as noted previously in relation to loop I. However, we recognize that it may be hard to treat separately multiple transcription factors regulating the same gene when considering their regulatory effects. Even in this case, we suggest that the combined activity profile of those transcription factors, which can be mapped into a mathematically equivalent single transcription factor’s profile, should follow our aforementioned condition when the target gene displays a cuspidate waveform.
Generally, it is known that dynamical systems with activating interactions alone do not easily generate oscillations; inhibitory interactions are also necessary. Specifically, an odd number of inhibitions need to be arranged along a feedback loop, if the loop is not too long [40–42]. In addition to this basal level of inhibitory interactions required, an abundance of cuspidate-waveform genes in the plant oscillator tips the balance in favor of a greater number of inhibitory interactions, resulting in their dominance, according to our hypothesis [cuspidate protein profiles include LHY, PRR5, TOC1, EARLY FLOWERING 4 (ELF4), LUX ARRHYTHMO (LUX), and GIGANTEA (GI) profiles in S1–S5 Figs, and comprise at least half of the available protein profiles. Among the corresponding genes, light-responsive genes are only LHY and GI (Fig 1A), which yet maintain cuspidate expression patterns in LL and DD (S4 and S5 Figs). It indicates that these patterns are largely independent of light stimulation]. For example, in loop I of the kernel, we note that both morning (LHY/CCA1) and evening (TOC1) genes show cuspidate profiles with a large phase difference between them, and are thus likely to require their own inhibitors. The simplest solution would be to have the two genes repressed by each other, but this solution, with an even number of inhibitions, would not generate oscillations. Hence, one more inhibitor, PRR5, is necessary and the subsequent introduction of the double negative connection from TOC1 to LHY/CCA1 through PRR5, combined with the TOC1 inhibition by LHY/CCA1, completes the repressilator structure. In addition, PRR5 should maintain a large phase difference with LHY/CCA1, because of the LHY/CCA1’s cuspidate profile. Consequently, PRR5 should show a small phase difference with TOC1. Because of this small phase difference, the inhibition of PRR5 by TOC1 cannot alone produce the empirically-observed cuspidate PRR5 profile. Therefore, PRR5 requires an additional inhibitor with a large phase difference, LHY/CCA1. The resulting inhibition of PRR5 by LHY/CCA1 now completes the full loop-I circuit. Through this analysis of loop I, the underlying mechanism of oscillatory dynamics with cuspidate waveforms was found to explain not only the prevalence of inhibitory interactions, but also the very specific, fine-resolution structure of loop I, revealing the loop’s organizing principle.
Motivated by the intriguing connection between the shape of the waveforms and inhibitory regulation in plants, we asked if such relationships are observed in other circadian systems. Notably, a prevalence of inhibitory interactions per se is not conserved in other organisms: the core circadian systems of other organisms are usually simpler than those of plants, and involve feedback loops with comparable numbers of positive and negative interactions [11–13]. Those interactions are not necessarily transcriptional, and thus, caution should be taken when they are analyzed in our waveform-shape framework, which has been derived from the mathematical models of transcriptional regulation. Despite this caveat, in a preliminary analysis below, we applied our framework to both transcriptional and non-transcriptional interactions, considering their possible mathematical similarity at the coarse-grained level.
In the core circadian clock of the fungus Neurospora crassa, WHITE COLLAR-1, 2 (WC-1 and WC-2) proteins form a WHITE COLLAR COMPLEX (WCC) that activates the expression of frequency (frq) gene. The expressed FRQ protein subsequently blocks the WCC activity by the clearance of WC-1 [12]. In this negative feedback loop, WC-1 is suppressed by FRQ, which is upregulated by WC-1. From the experimental data [43], we observed that WC-1 exhibits a cuspidate profile, when having a large phase difference (~11 hours in DD) with FRQ. At the same time, FRQ shows a smooth sinusoidal profile. Despite multiple complicating factors in a rigorous analysis of species other than plants, this preliminary result from the Neurospora data is supportive of a relation between waveform-shape, phase differences, and interaction types (activation or inhibition), which is suggested by our waveform-specifying framework.
In this study, we explored the underlying mechanism of the plant circadian system through a systematic in silico analysis of the clock gene circuitry, revealing its kernel architecture to be an interaction between four negative feedback loops dominated by inhibitory regulations (Figs 1A and 2). The kernel encompasses about half of the currently known interactions in the system, and they must be present to generate molecular rhythms close to WT. The other interactions not belonging to the kernel may play a role to improve the system’s robustness to diverse disturbances (S1 Text), or may be required to form WT rhythms but under light conditions that have not been considered here due to limited data availability. A follow-up analysis is warranted for a more holistic understanding of plant circadian dynamics. Overall, our study illustrates the remarkable utility of mechanistic simulations, which can complement experimental approaches, in deciphering important biological processes [44–46] such as circadian rhythms.
We suggested that a preponderance of inhibitory interactions at the core of the plant clock reflects abundant cuspidate profiles of clock genes, and facilitates the global coordination of temporally-distant clock events which are sharply peaked at very specific times. We envisage that this type of cuspidate waveforms helps confer high-resolution timing to many subsequent downstream tasks in plant physiology and development [35, 47]. Whether a certain class of waveforms other than cuspidate shapes will also benefit from inhibitory interactions will be an interesting issue to address.
Besides the effect on waveforms, alternative hypotheses might be possible to explain the prevalence of the inhibitory interactions, e.g., in the context of stochasticity in molecular events, or the system’s response time [48–50]. Yet, we are not aware of any explicit link or evidence to connect those mechanisms to dominant inhibitory interactions in the plant clock. Nevertheless, the possible relevance of those mechanisms deserves active investigation, towards a comprehensive picture of the plant circadian system viewed from various angles.
The four negative feedback loops within the kernel present an array of interesting predictions, which are experimentally testable. The Δprr5/prr7 double mutation severely impairs the free running rhythmicity of clock-controlled gene expression [36]. According to our prior discussion of the loops-I-to-III inactivation, only the removal of both PRR5 and PRR7’s inhibitory actions on LHY and CCA1, rather than entire deletions of PRR5 and PRR7, should suffice to phenocopy the double mutant, or at least, to considerably alter clock gene expression patterns. Additionally, from the reciprocal inhibitions within the unique loop-I structure, we suggested that an increase of the PRR5 inhibition by LHY/CCA1 would lengthen the free running period and that the opposite perturbation would shorten the period (S1 Text). Furthermore, in the context of inhibitory interactions and cuspidate waveforms, we proposed that decreasing the PRR5 inhibition by LHY/CCA1 under 12L:12D cycles, balanced by strengthening the PRR5 inhibition by TOC1, would reduce the peak-to-trough change in the PRR5 expression profile (S1 Text). Experimental validation of all these predictions would require manipulation of specific interactions between genes, rather than the alteration or deletion of the functionality of the entire gene itself. This could be achieved, for example, by modifying key cis-regulatory elements at the relevant promoter sites. Any discrepancy between experimental and computational results might be useful for our model improvement. Further consideration of protein segregation into different cellular compartments [27, 28], stochastic fluctuation in mRNA and protein concentrations [49, 51, 52], stimulus by temperature changes and endogenous sugar supply [53, 54], and tissue-specific clock regulation [55] offers additional avenues towards more complete mathematical models. Various methods to infer biological networks would also contribute to this direction [56–59]. Finally, our systematic approach advances the goal for a fundamental design principle of biological clockwork [53, 60–62], as well as for an optimal circuitry design in synthetic biology [32, 63, 64].
We constructed our mathematical model (MF2015) of the core circadian clock in Arabidopsis by applying system identification techniques [65]. Transcriptional, post-translational, and light regulations of molecular components were considered for model construction, primarily based on experimentally verified knowledge. The model consists of 24 ODEs employing Michaelis-Menten kinetics. Each ODE describes the concentration rate change of the corresponding mRNA, protein, or protein complex: typically, for mRNAs, ċm(t) = f1[{cTF(t)}, {h}, {θ}]–g1[cm(t), {θ}], and for proteins, ċp(t) = f2[cm(t), {θ}]–g2[cp(t), {θ}]. Here, cm (cp) denotes mRNA (protein) concentration, cTF denotes the transcription factor concentration, t is time, the function f1 (f2) describes transcriptional (translational) mechanisms, the function g1 (g2) describes mRNA (protein) degradation, θ’s are model parameters, h’s are the Hill coefficients, and {…} includes single or multiple elements. If experimental evidence indicates that transcription factors form a dimer, we set the Hill coefficient to be 2, otherwise, it is set to 1 [33, 66]. Transcriptional regulation in f1 is modeled by θ1(cTF)h/[θ2h+(cTF)h] for activation or θ1/[θ2h+(cTF)h] for inhibition. The regulatory effect of multiple activators (inhibitors) is modeled by the summation (product) of individual regulatory effects, with some exceptions such as PRR proteins (S1 Text) [29, 67]. We model the binding of ZTL and GI proteins by adapting the alternative Michaelis-Menten relation in [68]. For the model parameter estimation, we collected experimental time course data of mRNA and protein levels in WT Arabidopsis from publicly available sources listed in S1 Table. Because the absolute mRNA and protein levels were difficult to ascertain from their sources, we normalized the mRNA and protein levels into dimensionless values (≤1) with arbitrary scales (S1 Text). We compared the simulation results with experimental data and applied the prediction error method with a quadratic criterion [65] to estimate the parameters; minimization of a mean squared error between the simulated and experimental data gave rise to the estimated parameters. Before the minimization, the initial parameters were chosen using a linear least square method described in [26]. The minimization was performed using the MATLAB function fminsearch. In cases where constraints need to be imposed on the parameters to avoid over-fitting or biologically unrealistic solutions, the MATLAB function fmincon was used. Full details of the model construction, equations, and parameters are presented in S1 Text.
In this study, a kernel is defined as a collection of spanning subgraphs that satisfy the following condition: each spanning subgraph contains all molecular components in the system and a minimal subset of their regulatory interactions (including light regulation), which are necessary to generate the temporal trajectory of molecular concentrations close to those of WT. Identification of the exact kernel demands very extensive computational resources; therefore, we used a heuristic approach to estimate the kernel structure. In this procedure, both molecular interactions and light regulations are referred to simply as interactions. First, we simulated the knockout effect of each interaction on WT expression patterns under five different light conditions. The knockout effect was quantified for each molecular component and light condition, by a root mean square error (RMSE) between the simulated mutant and WT expression profiles of the component in that light condition (S1 Text). After deletion of a given interaction, we identified the largest value (RMSEmax) among RMSEs for all components and light conditions except for GI and ZTL proteins in LL (RMSEGI,LL and RMSEZTL,LL). Based on our manual inspection, the model outputs appear to remain robust if they simultaneously satisfy RMSEmax ≤ 0.2, RMSEGI,LL ≤ 0.5, and RMSEZTL,LL ≤ 0.5 (because GI and ZTL levels are substantially elevated in LL, they allow relatively large RMSEs). From MF2015, we pruned all interactions with small knockout effects (RMSEmax ≤ 0.2, RMSEGI,LL ≤ 0.5, and RMSEZTL,LL ≤ 0.5). The simulated profiles with the only remaining interactions after the pruning still showed RMSEmax ≤ 0.2, RMSEGI,LL ≤ 0.5, and RMSEZTL,LL ≤ 0.5. Among these remaining interactions, we focused on the interactions that satisfy RMSEmax ≤ 0.3, RMSEGI,LL ≤ 0.8, and RMSEZTL,LL ≤ 0.8. We found that some of these interactions can be additionally removed from the system because the simultaneous deletion of those interactions eventually resulted in RMSEmax ≤ 0.2, RMSEGI,LL ≤ 0.5, and RMSEZTL,LL ≤ 0.5, when parameter re-optimization was performed (S1 Text). We did not attempt to delete interactions with larger RMSEmax, RMSEGI,LL, or RMSEZTL,LL (RMSEmax > 0.3, RMSEGI,LL > 0.8, or RMSEZTL,LL > 0.8) with the original parameters, because these RMSEs were not usually reduced to RMSEmax ≤ 0.2, RMSEGI,LL ≤ 0.5, and RMSEZTL,LL ≤ 0.5 after parameter re-optimization. The exception to these procedures was the PRR7 inhibition by TOC1. This interaction was removed initially because of small RMSEs caused by the deletion. In fact, the small RMSEs resulted from the PRR7 inhibition by the EC, which buffered the loss of the inhibition by TOC1. The PRR7 inhibition by TOC1 and that by the EC are almost equivalent to each other, because of the same target gene (PRR7) and regulation type (inhibition), and similar TOC1 and EC profiles in MF2015. Indeed, the removal of the PRR7 inhibition by the EC from MF2015 was compensated for by the inhibition by TOC1 when accompanied by parameter re-optimization. Because of the equivalence of these two inhibitory interactions, we reinstated the PRR7 inhibition by TOC1 in the kernel structure. No other interaction was reinstated due to a lack of such equivalence. The simulation of the resulting kernel structure with re-optimized parameters produces the WT expression profiles that capture the overall experimental and MF2015-simulated profiles (S1–S5 Figs). Further details of the kernel identification are presented in S1 Text.
To investigate how transcriptional regulation affects the formation of cuspidate profiles, we considered a mathematical system containing a single transcription factor (either an inhibitor or activator) and its own target gene (Fig 4A). The ODEs for this system are given by x˙m(t)=g[xTF(t),h,{α}]−λmxm(t) and x˙p(t)=xm(t)−λpxp(t), where xTF denotes the transcription factor concentration, xm (xp) denotes the target gene’s mRNA (protein) concentration, t is time, g = xTFh/(α1+α2xTFh) + α3 if the transcription factor is an activator, g = 1/(α1+α2xTFh) if the transcription factor is an inhibitor, h is the Hill coefficient, and α’s and λ’s are constants. In the equation for x˙p(t), without loss of generality, we omitted the coefficient for a protein synthesis rate per mRNA in front of xm(t). Therefore, technically, xm(t) should be interpreted as the protein synthesis rate, rather than as the mRNA concentration itself. Although the equation for x˙m(t) was formulated for the case of a single transcription factor, it generally works for multiple transcription factors as well, because the combined activity profile of these transcription factors (represented by g) can be mapped into a mathematically equivalent single transcription factor’s profile. To generate xp(t) having a cuspidate waveform schematized in Fig 4B, we considered various forms of xTF(t) and activating and inhibitory regulations. Given the form of xTF(t), we computed xp(t) with the parameters that best fit xp(t) into a cuspidate profile in Fig 4B. The resulting xp(t) was compared to Fig 4B, and their similarity was evaluated. Further details are presented in S1 Text.
|
10.1371/journal.ppat.1004750 | Japanese Encephalitis Virus Nonstructural Protein NS5 Interacts with Mitochondrial Trifunctional Protein and Impairs Fatty Acid β-Oxidation | Infection with Japanese encephalitis virus (JEV) can induce the expression of pro-inflammatory cytokines and cause acute encephalitis in humans. β-oxidation breaks down fatty acids for ATP production in mitochondria, and impaired β-oxidation can induce pro-inflammatory cytokine expression. To address the role of fatty-acid β-oxidation in JEV infection, we measured the oxygen consumption rate of mock- and JEV-infected cells cultured with or without long chain fatty acid (LCFA) palmitate. Cells with JEV infection showed impaired LCFA β-oxidation and increased interleukin 6 (IL-6) and tumor necrosis factor α (TNF-α) expression. JEV nonstructural protein 5 (NS5) interacted with hydroxyacyl-CoA dehydrogenase α and β subunits, two components of the mitochondrial trifunctional protein (MTP) involved in LCFA β-oxidation, and NS5 proteins were detected in mitochondria and co-localized with MTP. LCFA β-oxidation was impaired and higher cytokines were induced in cells overexpressing NS5 protein as compared with control cells. Deletion and mutation studies showed that the N-terminus of NS5 was involved in the MTP association, and a single point mutation of NS5 residue 19 from methionine to alanine (NS5-M19A) reduced its binding ability with MTP. The recombinant JEV with NS5-M19A mutation (JEV-NS5-M19A) was less able to block LCFA β-oxidation and induced lower levels of IL-6 and TNF-α than wild-type JEV. Moreover, mice challenged with JEV-NS5-M19A showed less neurovirulence and neuroinvasiveness. We identified a novel function of JEV NS5 in viral pathogenesis by impairing LCFA β-oxidation and inducing cytokine expression by association with MTP.
| Lipids are involved in various steps of viral infection, and viruses may alter lipid metabolism to facilitate efficient viral replication. To address whether long-chain fatty acid (LCFA) metabolism is affected by Japanese encephalitis virus (JEV) infection, the leading cause of viral encephalitis in Asia, we compared the oxygen consumption rate of mock- and JEV-infected cells cultured with or without LCFA. LCFA utilization was impaired in JEV-infected cells, and higher pro-inflammatory cytokine expression was induced when LCFA was the major energy source. JEV nonstructural protein 5 (NS5) interacted with mitochondrial trifunctional protein, an enzyme complex involved in LCFA β-oxidation, and the interaction impaired LCFA β-oxidation, enhanced cytokine production, and contributed to JEV pathogenesis. The M19 residue of NS5 is involved in its interaction with MTP and the recombinant JEV with NS5-M19A mutation was less able to block LCFA β-oxidation, induced lower levels of cytokine production and showed less neurovirulence and neuroinvasiveness than wild-type JEV. Thus, impaired LCFA β-oxidation and enhanced cytokine production induced by JEV NS5 may provide new insight into JEV virulence.
| Japanese encephalitis virus (JEV) is a member of the Flaviviridae family of many important human pathogens such as yellow fever virus, dengue virus (DENV), West Nile virus (WNV), tick-borne encephalitis virus (TBEV) and hepatitis C virus (HCV) [1]. JEV is the leading cause of viral encephalitis in Asia, with more than 50,000 cases and 10,000 deaths annually [2,3]. The genome of JEV is a positive-sense RNA encoding a polyprotein that is proteolytically processed into three structural proteins (core [C], precursor of membrane [prM], and envelope protein [E]) and seven nonstructural proteins (NS1, NS2A, NS2B, NS3, NS4A, NS4B, and NS5). NS5 is the largest flaviviral protein, with enzymatic activities of methyltransferase (MTase) and RNA-dependent RNA polymerase (RdRP) required for viral replication [4–6]. Interferon (IFN) antagonistic roles have also been demonstrated for flaviviral NS5; for example, the NS5 proteins of JEV, WNV and TBEV can block IFN-triggered JAK-STAT signaling [7–9] and DENV NS5 can cause degradation of STAT2 protein [10].
Lipids are involved in various steps of viral infection, such as viral entry, RNA replication, virion assembly and energy supply, and viruses are known to modulate cellular lipid metabolism [11,12]. Fatty acids synthesized from acetyl-CoA by lipogenesis may serve as precursors to produce lipid components or be broken down for ATP production via β-oxidation [13]. Long-chain fatty acids (LCFAs) are transported into mitochondria with the help of carnitine, then β-oxidation splits LCFA into acetyl-CoA via a four-step reaction [14]. Three of the 4 enzymatic activities of β-oxidation are catalyzed by a protein complex called mitochondrial trifunctional protein (MTP), consisting of hydroxyacyl-CoA dehydrogenase subunit A and subunit B (HADHα and HADHβ) [14,15]. The hallmark of MTP deficiency is accumulation of long-chain 3-hydroxy fatty acids [16–18], which, trapped inside the mitochondrial matrix, induce reactive oxygen species (ROS) production and pro-inflammatory cytokine expression [19,20]. Blockage of LCFA β-oxidation may also increase glucose consumption and result in hypoglycemia [21,22], which is deleterious to the central nervous system (CNS) [23]. Thus, patients with MTP deficiency might have serious complications, with damaged organs and long-term irreversible neuropathic complications with progressive encephalopathy.
Positive-sense RNA viruses induce intracellular membrane rearrangements to create favorable sites for viral replication. To synthesize and reorganize the intracellular membranes, HCV increases de novo synthesis and uptake of fatty acids, and also inhibits β-oxidation [24,25]. These lipid modulations may lead to abnormal accumulation of fat deposits in the liver (steatosis) [26,27] and are associated with chronic inflammatory response, features commonly seen in HCV patients [28]. Impaired LCFA β-oxidation has been implicated in influenza-associated neuronal disease, because patients with fatal and handicapped influenza-associated encephalopathy showed increased serum acylcarnitine ratio of C16:0+C18:0 to C2 [29]. Acute Japanese encephalitis (JE), characterized by inflammatory mediators in the brain, can develop in humans stung by a JEV-infected mosquito. The neurological dysfunction caused by activated immune cells is via induction of pro-inflammatory cytokines and ROS production, which leads to increased permeability of the blood brain barrier [30–32]. The levels of tumor necrosis factor α (TNF-α) and interleukin 6 (IL-6) were elevated in serum and cerebrospinal fluid of JE patients, with their induction associated with fatal outcome of JE [33,34].
Despite the link between lipid metabolism and inflammation in viral diseases, the role of fatty acid metabolism in JEV infection is largely unknown. In this study, we addressed whether LCFA β-oxidation is modulated by JEV infection and its potential involvement in JEV pathogenesis. We further studied the molecular mechanism of how JEV hinders cellular β-oxidation and found that JEV NS5 interacts with HADHα and HADHβ, two subunits of the enzyme complex MTP involved in LCFA β-oxidation. Furthermore, the recombinant JEV carrying a mutated NS5, with less binding ability with MTP, was less able to block LCFA β-oxidation, triggered reduced cytokine production, and featured less virulence. We discuss the novel function of JEV NS5 in modulating LCFA β-oxidation and cytokine induction.
Palmitic acid is the most common fatty acid in animals, so we and others have been using sodium palmitate conjugated bovine serum albumin (PA-BSA) to study fatty acid β-oxidation [35]. During fasting, fatty acid oxidation becomes the major energy source [36,37] and oxygen consumption will be mainly resulted from β-oxidation in cells cultured with PA-BSA under starvation condition (without serum). To address whether JEV infection modulates cellular LCFA β-oxidation, we measured the oxygen consumption rate (OCR) in cells cultured with a noncytotoxic dose of PA-BSA (S1 Fig) or BSA control by using a metabolic XF24 analyzer [35]. In JEV-infected human A549 cells cultured with BSA, OCR values continued to increase from 6 to 24 h post-infection (hpi; Fig. 1A, JEV + BSA). However, in JEV-infected cells cultured with PA-BSA, the OCR values increased in the beginning, then decreased from about 11 hpi until the end of the recording (Fig. 1A, JEV + PA-BSA). Mock-infected cells did not show the distinct OCR patterns with BSA and PA-BSA treatments (Fig. 1A). Changes in OCR values represented by area under the curve (AUC) similarly showed that the AUC OCR was lower in JEV-infected A549 cells cultured with PA-BSA than BSA (Fig. 1B). This phenomenon was not limited to a single cell type and also occurred in JEV-infected human neuroblastoma HTB-11 cells cultured with PA-BSA or BSA (S2A Fig). Thus, the reduced OCR in JEV-infected cells cultured with palmitate indicates that JEV cannot utilize LCFA efficiently, probably because of blocked β-oxidation.
We then assessed the effect of impaired β-oxidation on JEV replication and cytokine induction. The levels of viral NS3 protein expression and viral progeny production were lower in cells cultured with PA-BSA than BSA (Figs. 1C and S2B) and this reduction could be rescued by serum supplement (Figs. 1D and S2C). The induction of IL-6 and TNF-α was higher in JEV-infected cells cultured with PA-BSA (Figs. 1E, 1F, S2D, S2E), even though viral replication was reduced under this condition. Interleukin 10 (IL-10), but not IL-4 and IL-13, was also induced in JEV-infected cells cultured with PA-BSA (S3 Fig). Furthermore, this cytokine induction depended on ROS generation, because of stronger ROS signals seen in JEV-infected cells cultured with PA-BSA than BSA (S4A Fig) and treatment with N-acetylcysteine (NAC), a free radical scavenger, reduced the levels of cytokine induction (Fig. 1E and 1F). Moreover, nuclear translocation of NFκB, an indicator of NFκB activation, was more prominent in JEV-infected cells cultured with PA-BSA than that with BSA (91% vs. 62%) (S4B Fig). Our data thus suggest that JEV-impaired LCFA β-oxidation can lead to ROS generation, NFκB activation and cytokine induction.
Three of the four enzymatic activities of LCFA β-oxidation are catalyzed by protein complex MTP [14], so we investigated whether JEV modulates LCFA β-oxidation by changing the expression and/or localization of MTP. The protein expression levels of the two subunits of MTP, HADHα and HADH, were similar between JEV- and mock-infected cells (Fig. 2A, lanes 1–2). Furthermore, by using a mitochondria isolation kit (S5A Fig), these two MTP proteins were detected in the heavy membrane fraction (H) containing mitochondria of mock- and JEV-infected cells (Fig. 3A, lanes 2 and 4). We then explored whether certain JEV proteins might interact with MTP by using immunoprecipitation (IP)–Western analysis. The plasmids expressing individual Flag-tagged JEV proteins were co-transfected with that of V5-tagged HADHα. In cells expressing HADHα-V5 plus NS5-Flag but not other viral proteins, anti-Flag affinity gel also brought down V5-tagged HADHα (Fig. 2B). Furthermore, endogenous HADHα and HADHβ were identified as NS5-interacting proteins by LC-MS/MS proteomic analysis of cellular proteins co-immunoprecipitated with Flag-tagged NS5 (Figs. 2C and S6). The interaction of NS5-Flag with HADHα-V5-His and HADHβ-HA was demonstrated by IP—Western analysis (Fig. 2D). The virus-expressed NS5 also interacted with HADHα and HADHβ, as demonstrated by IP—Western analysis of cellular lysates with JEV infection plus HADHα-V5-His or HADHβ-V5-His transfection (Fig. 2E).
JEV NS5 protein expressed by plasmid transfection or viral infection was detected in cytosolic (C) and mitochondria-containing heavy membrane (H) fractions (Fig. 3A, 3B and 3C) by using 3 different isolation protocols outlined in S5 Fig. To better understand the subcellular localization of JEV NS5, we performed Proteinase K resistance assay on the crude mitochondria isolated from HEK293 cells with JEV infection or JEV NS5-Flag overexpression. As shown in Fig. 3D, Proteinase K digested the mitochondrial outer membrane protein TOM70, whereas the proteins in intermembrane-space (Cytochrome c) and inner-membrane (HADHα and HADHβ) were protected. Importantly, some of the NS5 proteins were resistant to Proteinase K cleavage, suggesting enclosure of NS5 by membrane structure. JEV NS3 and E proteins were detected in the cytosolic and membrane-containing fractions (S5 Fig) as previously reported [38,39]. Furthermore, E but not NS3 protein was resistant to Proteinase K-mediated cleavage (S5C Fig), in accordance with the known locations for NS3 and E in cytosol and inside the endoplasmic reticulum (ER), respectively. Different fractionation patterns were noted between mock- and JEV-infected cells; for example, the ER protein calreticulin was mainly detected in the cytosolic/light microsomal membrane fractions of mock cells, but its location slightly shifted to the heavy membrane fraction (S5B Fig), probably due to the intracellular membrane rearrangements known to be caused by many positive-sense RNA viruses, including JEV [11]. Furthermore, co-localization of NS5 with mitochondria and with HADHα or HADHβ was detected by confocal microscopy (Fig. 3E and 3F). Similar to with JEV infection, NS5 protein expression did not change the expression level (Fig. 2A, lane 3) or cellular location (Fig. 3A, lanes 5–6) of HADHα and HADH. Thus, JEV NS5 can locate in a rearranged heavy membrane structure containing mitochondrial proteins and interact with subunits HADHα and HADHβ of the LCFA β-oxidation enzyme complex MTP.
To test whether NS5 is involved in impaired LCFA β-oxidation, we measured the OCR of A549 cells with or without NS5 overexpression cultured with PA-BSA or BSA. The AUC OCR was significantly lower in NS5-overexpressing cells cultured with PA-BSA than BSA, whereas AUC OCR was higher in vector control cells cultured with PA-BSA than BSA (Fig. 4A). Levels of IL-6 and TNF-α were higher in NS5-overexpressing cells cultured with PA-BSA than BSA (Fig. 4B and 4C). Moreover, this cytokine induction phenomenon was JEV NS5-specific, since cells expressing other viral proteins such as JEV NS1, NS2A and DENV NS2B3 did not show TNF-α induction with PA-BSA treatment (S7 Fig).
To identify the region of NS5 interacting with HADHα or HADH, we co-expressed Flag-tagged full-length or a series of truncated NS5 used previously [7] (Fig. 5A), with V5-tagged HADHα or HA-tagged HADHβ. Anti-Flag affinity gel co-immunoprecipitated HADHα and HADHβ with the NS5 proteins containing N-terminal 1–270 residues but not with the N-terminal—deleted NS5 (167–905) (Fig. 5B and 5C). To identify the crucial amino acids of NS5 (1–270) participating in this interaction, we created NS5 mutants by random mutagenesis and screened for their ability to bind with HADHα or HADHβ by IP—Western analysis. The NS5 mutant with residue 19 changed from methionine to alanine (M19A) showed reduced binding with endogenous HADHα and HADH (Fig. 6A), despite the cellular distribution of NS5-WT and NS5-M19A was similar (Fig. 6B). To verify whether the MTase activity located at the N-terminus of NS5 is involved in the interaction with MTP, we site-specifically mutated the enzyme catalytic tetrad KDKE motif [40] by creating the K61A, D146A, K182A and E218A mutants of JEV NS5. NS5-K61A and-D146A, but not-K182A and-E218A mutants, showed reduced binding with MTP (Fig. 6A), which suggests that the MTase enzyme activity per se is not essential for this protein—protein interaction.
To test whether NS5 association with MTP is involved in impaired LCFA β-oxidation, we compared the OCR of A549 cells with wild-type NS5 (NS5-WT) or NS5-M19A overexpression. Higher AUC OCR was noted in PA-BSA-treated NS5-M19A cells (S8A Fig), indicating that NS5-M19A was less able to block LCFA β-oxidation than NS5-WT. Consistently, lower IL-6 and TNF-α induction was seen in cells expressing NS5–M19A (S8B and S8C Fig). Furthermore, IL-6 protein level was higher in NS5-overexpressing cells when compared to GFP control and NS5-M19A mutant (S8D Fig).
We then created recombinant JEV with NS5 mutation by using a JEV infectious clone [41]. Since NS5-K61A mutation hampers JEV replication [42], we selected NS5-M19A and-D146A for recombinant JEV generation. JEV with NS5-M19A, but not-D146A mutation was recovered, likely because the D146A mutation will abolish its MTase activity and lose viral replication ability as reported for WNV [43]. JEV-NS5-M19A was infectious and produced similar plaque morphology as with wild-type JEV (JEV-WT) in BHK-21 cells (Fig. 6C). The viral NS3 protein expression and viral progeny production of JEV-WT and JEV-NS5-M19A were similar in A549 cells (Fig. 6D). However, the binding of NS5-M19A to HADHα-V5-His and HADHβ-HA was lower than that of NS5-WT in the context of virus infection (Fig. 6E), despite their cellular localization was similar (Fig. 6F).
The OCR and AUC OCR were higher in PA-BSA—treated A549 cells infected with JEV-NS5-M19A than JEV-WT, while with similar values with BSA treatment (Fig. 7A and 7B). Thus, JEV-NS5-M19A was less able to block LCFA β-oxidation and induced lower levels of IL-6 and TNF-α than JEV-WT (Fig. 7C and 7D). Furthermore, as compared with JEV-WT infection, even with serum-containing medium, JEV-NS5-M19A infection triggered significantly lower levels of IL-6 and TNF-α while producing slightly less viral RNA (Fig. 7E, 7F and 7G). Thus, M19 of NS5 is involved in its interaction with MTP and affects the ability of JEV to impair LCFA β-oxidation and induce cytokine expression.
To investigate the impact of LCFA β-oxidation on JEV infection in vivo, we compared the neurovirulence of JEV-WT and JEV-NS5-M19A in mice with intracerebral (i.c.) virus injection. All mice died with injection of 20 or 2 plaque-forming units (PFU) of JEV-WT or JEV-NS5-M19A, whereas 80% and 60% of mice survived from challenge with 0.2 PFU of JEV-NS5-M19A and JEV-WT, respectively (Fig. 8A). The 50% lethal dosage (LD50) was calculated as 4 x 10-1 and 9.48 x 10-1 PFU for JEV-WT and JEV-NS5-M19A, respectively, for a 2.37-fold increase for JEV-NS5-M19A. The levels of viral RNA replication were similar in the brains of mice infected with JEV-WT or JEV-NS5–M19A (Fig. 8B). However, IL-6 and TNF-α induction was higher in brains of mice infected with JEV-WT than JEV-NS5–M19A (Fig. 8C and 8D).
The difference between these two viruses was more obvious on challenge with an intraperitoneal (i.p.) injection plus i.c. puncture with PBS (i.p. plus i.c. route). The LD50 for JEV-WT and JEV-NS5-M19A was 2 x 103 and 1.38 x 104 PFU, respectively, for a 6.92-fold increase for the NS5-M19A mutated JEV (Fig. 9A). The levels of JEV titers and viral RNA were higher in mouse brains inoculated with JEV-WT than JEV-NS5–M19A (Fig. 9B and 9C). Furthermore, IL-6 and TNF-α gene induction was higher in mouse brains challenged with JEV-WT than JEV-NS5–M19A (Fig. 9D and 9E) and IL-6 protein could be detected in the sera of mice with JEV-WT infection (Fig. 9F). Thus, JEV NS5 can bind with MTP and hinder its ability to catalyze LCFA β-oxidation, which then induces cytokine production and contributes to viral pathogenesis.
Flaviviral NS5 contains 2 enzymatic domains: RdRP on its C-terminus required for viral RNA replication and MTase on its N-terminus needed for viral RNA stability and efficient translation [4,5]. The 2'-O methylation on the viral RNA 5' cap catalyzed by NS5 MTase contributes to escape from the IFIT-mediated host antiviral response for WNV and JEV [44,45]. Several flaviviral NS5 proteins such as those from JEV, WNV, TBEV and DENV have been identified as IFN signaling antagonists [7–10] by targeting STAT2 and not-yet-identified mechanisms. Here, we discover a new function of flaviviral NS5: JEV NS5 interacts with MTP, an enzyme complex involved in LCFA β-oxidation and interferes with the catabolism of LCFA. The accumulated LCFA triggers oxidative stress, activates NFκB, induces pro-inflammatory cytokine production and contributes to JEV pathogenesis. Thus, besides being the enzyme involved in virus replication, flaviviral NS5 also functions as an immune modulator by affecting the host immune system such as IFN signaling and cytokine production. Furthermore, these two immunomodulation functions of NS5 may not be mediated by the same molecular mechanism, since JEV-WT and JEV-NS5-M19A show different degree of LCFA β-oxidation impairment, but both can trigger IFN production and block IFN signaling (S9 Fig).
Different cellular distribution has been reported for flaviviral NS5 proteins. For example, the NS5 proteins of DENV-2 and DENV-3 [46,47] mainly locate in the nuclei, but those of JEV [7,48,49], WNV [8,46,50], DENV-1 and DENV-4 [47] are in the cytoplasm. By using fractionation and confocal microscopy assays, JEV NS5 was detected in the cytosolic fraction and membrane-containing fractions including mitochondria (Figs. 3 and S5). Although no conventional mitochondria targeting sequence (MTS) was predicted, JEV NS5 shows ~1/3 in probability of translocation to mitochondria (Mitoprot score: 0.3351) by Mitoprot software (http://ihg.gsf.de/ihg/mitoprot.html) [51,52]. We suspect that the nonconventional mitochondria import pathways such as that used by microtubule-associated protein 4 (MAP4) [53] and human apurinic/apyrimidinic endonuclease [54] might be adapted by JEV NS5. Another possibility is that JEV NS5 may enter mitochondria with the help of other cellular proteins such as Hdj2, which is known to regulate mitochondrial protein import [55] and has been reported to interact with JEV NS5 [56]. Furthermore, the NS5 proteins in the crude mitochondrial fractions migrated slightly slower than the ones in cytosolic fractions (Fig. 3A). Thus, JEV NS5 might gain access to mitochondria through certain protein modifications such as the phosphorylation-mediated mitochondrial translocation of cytosolic proteins [57] reported for MAP4 [53] and Parkin [58]. However, mitochondrial translocation of NS5 does not guarantee its interaction with MTP, since the N-terminal—deleted NS5 (167–905) and M19A-mutated NS5 show reduced binding with MTP (Figs. 5 and 6) but could still be detected in mitochondria (Figs. 6 and S10A).
The recombinant JEV-NS5-M19A mutant and JEV-WT replicated to similar levels, but JEV-NS5-M19A was less able to block LCFA β-oxidation and triggered lower cytokine levels than the wild type in cultured cells (Figs. 6 and 7). Furthermore, JEV-NS5-M19A exhibited attenuated neurovirulence and neuroinvasiveness as compared with JEV-WT in challenged mice (Figs. 8 and 9). According to the crystal structure of JEV NS5 [59], M19 residue is located on a linker between two helix structures (S10B Fig). Since linker peptide mutants may affect protein folding and lead to conformational changes [60,61], we suggest that structural integrity of the linker with residue M19 on NS5 may be essential for maintaining its functional interaction with MTP.
Similar to our findings with JEV, infection with HCV and human cytomegalovirus (HCMV) impairs fatty acid β-oxidation [25,62], whereas DENV infection increases fatty acid β-oxidation [63]. We also noted that DENV-2 infection was less able to block β-oxidation than JEV infection (S11A Fig). DENV-2 NS5 mainly located in cell nuclei and did not interact with MTP (S11B and S11C Fig). Furthermore, inhibition of β-oxidation by etomoxir reduced DENV replication [63] but has no effect on JEV replication (S12 Fig). Thus, NS5 proteins of JEV but not that of DENV-2 interact with cellular MTP and these two viruses interplay with cellular fatty acid β-oxidation in different ways.
Viruses use various mechanisms, such as by affecting gene expression and protein—protein interaction, to regulate fatty acid β-oxidation. Genes including HADHα and peroxisome proliferator-activated receptor α, a transcription factor required for the expression of genes involved in fatty acids metabolism, are downregulated in patients with HCV cirrhosis and hepatocellular carcinoma [26]. HCV core protein can induce various alterations in lipid metabolism by increasing the expression of genes involved in lipogenesis and decreasing that of genes involved in β-oxidation and secretion of fatty acids [25,64]. For HCMV infection, a cellular IFN-induced protein named Viperin is translocated into mitochondria to interact with MTP and inhibit fatty acid β-oxidation [62,65]. Since JEV infection causes protein degradation of Viperin [66], Viperin redistributed to mitochondria may not be adapted by JEV to block MTP. AMP-activated kinase (AMPK) plays a role in cellular energy homeostasis; activation of AMPK can inhibit fatty acid synthesis and restrict infection of several RNA viruses [67]. We also addressed whether AMPK is involved in JEV infection by treating cells with an AMPK activator A769662, which showed no effect on JEV replication in cells with or without palmitate pretreatment (S13B and S13C Fig). Thus, different from Kunjin virus [67], AMPK may not be involved in JEV replication and may not contribute to impaired β-oxidation in JEV-infected cells.
Under the well-fed condition, glucose is the major substrate for ATP production, but when glucose level is low or with excess fatty acid content, fatty acids will become the alternative source for energy production [36,37]. JEV infection consumes ATP [68], so ATP levels were lower in JEV-infected cells as compared to mock (S14 Fig). Because of LCFA β-oxidation impairment by JEV, further reduction of ATP was seen in cells cultured with PA-BSA (S14 Fig). Normally, glucose is mainly broken down by oxidative phosphorylation in mitochondria, but under hypoxia and stress conditions such as virus infection [69], glycolysis occurring in cytoplasm will dominate [70,71]. For example, HCV and HCMV infection induces glycolysis [72,73] and the activity of some glycolysis enzymes is increased during JEV infection [68,74]. Glycolysis produces lactate and causes acidification of the extracellular space, called lactic acidosis [75], seen in patients with LCFA β-oxidation deficiency [76,77] and Japanese encephalitis (JE) [78]. Thus, the high lactate secretion in patients with JE might be a metabolic symptom due to impaired LCFA β-oxidation during JEV infection.
Fatty acids can generate intracellular ROS via several mechanisms [79] and fatty acid metabolism has been implicated in viral pathogenesis. For example, the expression of pro-inflammatory cytokines IL-6 and TNF-α was higher in hepatitis B virus X-protein—transgenic mice fed a high-fat rather than normal diet [80]. Furthermore, a non-neurotropic influenza A virus replicated to increased levels in mice lacking carnitine transporter OCTN2, a gene required for LCFA β-oxidation, and resulted in increased brain vascular permeability and encephalopathy [29]. Thus, disordered mitochondrial β-oxidation increases the risk of brain damage caused by influenza A virus infection. JEV is a neurotropic virus [5] that causes encephalitis by attracting immune cells across the blood brain barrier to induce the inflammatory response and brain pathology [30–32]. When fatty acid β-oxidation is impaired, the accumulation of LCFA elicits protein oxidative damage and decreases antioxidants in the cerebral cortex [18]. Thus, impaired LCFA β-oxidation may facilitate membrane proliferation and rearrangement in JEV-infected cells, but then likely contributes to JE-associated brain damage. Our finding that JEV NS5 associates with MTP and can inhibit fatty acid β-oxidation may shed new light on JEV-triggered pathogenesis and provide a novel target for future drug development.
Human lung epithelial carcinoma A549 cells (ATCC, CCL-185) were cultured in F-12 medium (Invitrogen, Grand Island, New York, USA) containing 10% fetal bovine serum (FBS). Human neuroblastoma SK-N-SH cells (ATCC, HTB-11) were grown in Minimum Essential Media (Invitrogen) containing 10% FBS. Human embryonic kidney 293T cells (HEK293T; ATCC, CRL-11268) were cultured in Dulbecco’s modified Eagle’s medium (Invitrogen) containing 10% FBS. Baby hamster kidney BHK-21 cells (ATCC, CCL-10) were grown in RPMI 1640 medium (Invitrogen) containing 5% FBS. JEV strain RP-9 [81] (GenBank accession no. AF014161) was propagated in the C6/36 mosquito cell line grown in RPMI 1640 medium supplemented with 5% FBS.
The JEV-NS5-M19A mutant was generated by single-primer mutagenesis [82] with the primer 5΄-GAAGGAAAAACTAAATGCCGCGAGCAGAGAAGAGTTTTTTAAATACCG-3΄ (mutated sequence underlined) with a JEV infectious clone as described [41]. For viral infection, cells were adsorbed with virus at the indicated multiplicity of infection (MOI) for 2 h at 37°C, then unbound virus was removed by a gentle wash with HBSS (Invitrogen). At the indicated times post-infection, culture supernatants were sequentially diluted for plaque-forming assays on BHK-21 cells as described [41]. To establish JEV NS5-overexpressing cells with EYFP-tagged mitochondria, pTY-EF-NS5-Flag cells [7] were transduced with EYFP-Mito-expressing lentivirus for 24 h, then selected with 5 μg/ml puromycin for 72 h.
N-acetylcysteine (NAC) (A7250), Proteinase K (P2308), recombinant human IFN-αA/D (I4401), and Etomoxir (E1905) were from Sigma (St. Louis, MO, USA). A769662 (sc-203790) was from Santa Cruz Biotechnology (Dallas, Texas, USA).
Cytotoxicity was assessed by use of the Cytotoxicity Detection Kit (LDH) (Roche, Basel, Switzerland). Cell viability was determined by using AlamarBlue (Invitrogen) cell viability assay and trypan blue exclusion assay (Gibco, Grand Island, NY, USA). Briefly, A549 cells were incubated with the indicated concentration (0–400 μM) of PA-BSA for 24 h. Cell-free supernatants were collected and used in LDH assay as instructed by the manufacturer. The viable cells stained with AlamarBlue were determined by measurement of spectrophotometric absorbance with a microplate reader. The cells were mixed with an equal volume of trypan blue then survival cell numbers were determined using an automated cell counter (Countess; Invitrogen).
Oxygen consumption rate (OCR) in A549 and HTB11 cells was measured in serum-free F-12 medium containing 0.5 mM L-carnitine (C0158; Sigma), an essential addition to transport palmitate into mitochondria. Sodium palmitate (P9767; Sigma) was conjugated with fatty acid free bovine serum albumin (BSA) (A7030; Sigma) (PA-BSA) at a 6:1 molar ratio by a protocol from Seahorse Bioscience (North Billerica, MA, USA). Briefly, sodium palmitate was solubilized in 150 mM NaCl by heating up to 70°C. BSA was dissolved in 150 mM NaCl and warmed up to 37°C with continuous stirring. Solubilized palmitate was added to BSA at 37°C with continuous stirring. Then, the conjugated palmitate—BSA (PA-BSA) was aliquoted and stored at -20°C for assessing β-oxidation of long-chain fatty acid [35]. After the injection of PA-BSA or BSA, OCR values were real-time recorded every 8 min from 6 to 24 h post infection with use of an XF24 analyzer (Seahorse Bioscience) and the area under the curve (AUC) OCR was calculated.
Cells were lysed with RIPA buffer (10 mM Tris, pH 7.5, 5 mM EDTA, 150 mM NaCl, 0.1% SDS, 1% TritonX-100, 1% sodium deoxycholate) containing a cocktail of protease inhibitors (Roche). Equivalent amounts of proteins determined by the DC Protein Assay Kit (Bio-Rad, Hercules, CA, USA) were separated by SDS-PAGE and transferred to a nitrocellulose membrane (Hybond-C Super; Amersham, Buckinghamshire, UK). Nonspecific antibody binding sites were blocked with skim milk in phosphate-buffered saline (PBS) with 0.1% Tween 20 (PBST), then reacted with primary antibodies for HADHα (sc-82185), HADH (sc-55661 and sc-271496) and α Tubulin (sc-5546) from Santa Cruz Biotechnology; actin (NB600-501; Novus Biologicals, Littleton, CO, USA); GAPDH (GTX100118) and TOM70 (GTX85353) from GeneTex (Irvine, CA, USA); Cytochrome C (#556433; BD, Franklin Lakes, New Jersey, USA); Calreticulin (#2891), phospho-STAT1 (Tyr701) (#9171), STAT1 (#9172), phospho-AMPKα (Thr172) (#2535), and AMPKα (#2532) from Cell Signaling Technology(Danvers, MA, USA); GFP (#11814460001; Roche); V5-tag (V8012) and Flag-tag (F7425) from Sigma; or HA-tag (MMS-101R; Covance, Princeton, New Jersey, USA), and then incubated with an appropriate horseradish peroxidase-conjugated secondary antibody (Amersham). Signals were detected by enhanced chemiluminescence (Amersham).
Total cellular RNA was prepared with use of an RNeasy Mini Kit (Qiagen, Hilden, Germany) and cDNA was reverse-transcribed from 1 μg total RNA by use of the SuperScript III First-Strand Synthesis System (Invitrogen). qPCR involved use of TaqMan Universal PCR Master Mix (Invitrogen) with commercial probes for IL-6 (Hs00985639 and Mm00446190), TNF-α (Hs01113624 and Mm00443260), IL-10 (Hs00961622), IL-4 (Hs00174122), IL-13 (Hs00174379), IFN-β (Hs01077958) and GAPDH (Hs02758991 and Mm99999915) (Applied Biosystems, Foster City, CA, USA) as well as primers for JEV viral RNA (5΄-AGAACGGAAGATAACCATGACTAAA-3΄and 5΄-CCGCGTTTCAGCATATTGAT-3΄). The relative expression of genes was assessed by the comparative threshold cycle method and normalized to that of GAPDH.
Cells were stained with 50 μM 2’,7’-dichlorofluorescin diacetate (DCFH-DA) (OxiSelec Intracellular ROS Assay Kit; Cell Biolabs, San Diego, CA, USA) for 30 min and examined under an inverted fluorescent microscope.
Cells were fixed with 4% formaldehyde in PBS for 20 min at room temperature, then washed twice with PBS. Cells were permeabilized in PBS containing 0.2 or 0.5% Triton X-100 for 5 min and blocked with skim milk in PBS or 3% BSA in Tris-buffered saline (TBS), then incubated with primary antibodies for NFκB p65 (sc-372; Santa Cruz), HADHα (sc-374497; Santa Cruz), Flag-tag (F7425; Sigma), or HA-tag (MMS-101R; Covance) diluted in TBS with 2% BSA overnight at room temperature before being washed with TBS, then with appropriate Alexa Fluor-conjugated secondary antibodies (Alexa Fluor 647 goat anti-mouse [A21236] or Alexa Fluor 568 goat anti-rabbit [A11036] from Invitrogen) for 1 h at room temperature. Cells were photographed under a fluorescence microscope or a Zeiss LSM510 Meta Confocal Microscope with a 100X objective. Co-localization was visualized by use of the ZEN 2011 (Zeiss, Oberkochen, Germany) co-localization module.
The Qproteome Mitochondria Isolation Kit (Qiagen) was used to isolate crude mitochondria from HEK293T cells according to the manufacturer’s instruction as outlined in S5A Fig. Two other biochemical approaches of cellular fractionation were also performed as previously described [83,84]. As outlined in S5B Fig, cells were washed once with cold PBS, scraped off from culture plate, and lysed in homogenization buffer [20 mM HEPES (pH 7.5), 70 mM sucrose and 220 mM mannitol] by 30 strokes in a Dounce homogenizer. The homogenate was centrifuged at 800 g for 5 min to precipitate the nuclei, and the resulting supernatant was further centrifuged at 10,000 g for 10 min (4°C) to precipitate the crude mitochondrial fraction. The resulting supernatant was further centrifuged at 100,000 g for 30 min (4°C) to precipitate light membrane organelles, and the final supernatant was used as the cytosolic fraction. Another biochemical method [84] was outlined in S5C Fig. Briefly, JEV-infected HEK293T cells were washed once with cold PBS, scraped off from culture plate, and lysed in mitochondria buffer [10 mM Tris/MOPS (pH 7.4), 0.1 mM EGTA/Tris (pH 7.4) and 250 mM sucrose] by 20 strokes in a Dounce homogenizer. Part of the homogenate was centrifuged at 16,200 g for 30 min, and the resulting supernatant was used as cytosolic fraction. The rest homogenate was centrifuged at 600 g for 5 min to precipitate the nuclei and unbroken cells, and the resulting supernatant was further centrifuged at 7,000 g for 10 min (4°C). Then the resulting pellet was resuspended and centrifuged again at 10,000 g for 10 min (4°C) to precipitate the crude mitochondrial fraction. Isolated mitochondrial fractions were lysed and examined by Western blot analysis.
The Proteinase K resistance assay was performed as previously described [84]. Briefly, the crude mitochondria pellet was washed once with mitochondria buffer. Then, the pellet resuspended in mitochondria buffer was treated with Proteinase K on ice for 30 min. After adding 2 mM phenylmethylsulfonyl fluoride (PMSF) to quench the protease reaction, samples were centrifuged at 15,000 g for 10 min (4°C). The resulting pellet was washed with mitochondria buffer plus 1 mM PMSF and centrifuged again. The reactants were subjected to Western blot analysis with the indicated antibodies.
HEK293T cells were transfected with the indicated plasmids by use of Lipofectamine 2000 (Invitrogen). After 24 h, cells were lysed with IP lysis buffer (50 mM Tris-HCl, 150 mM NaCl, 1 mM EDTA, 1% Triton X-100, pH 7.4) containing a cocktail of protease inhibitors (Roche). Cell lysates were immunoprecipitated with anti-Flag M2 (A2220),-HA (E6779),-V5 affinity gel (A7345) or Nickel beads (P6611; Sigma) overnight at 4°C. Proteins were eluted by sample buffer and examined by Western blot analysis with the indicated antibodies.
Cell lysates of A549, GFP-A549 and NS5-Flag-A549 cells were immunoprecipitated with control IgG or anti-Flag affinity gel. Proteins in the immune complexes were separated by SDS-PAGE and visualized by staining with SYPRO Ruby (Invitrogen). The extra protein bands, which bound to NS5-Flag but not the control, were excised for in-gel trypsin digestion and analyzed by LC-MS/MS.
We created NS5 mutants by random mutagenesis with mutagenic dNTP analogs with the JBS dNTP-Mutagenesis Kit (PP-101; Jena Bioscience, Jena, Germany). The technique involves incorporation of the dNTP analogs 8-oxo-dGTP and dPTP, which induce base mispairing upon DNA amplification.
Animal studies were conducted according to the guidelines outlined by Council of Agriculture, Executive Yuan, Republic of China. The animal protocol was approved by the Academia Sinica Institutional Animal Care and Utilization Committee (Protocol ID 11-11-245). All surgery was performed under sodium pentobarbital anesthesia and every effort was made to minimize suffering.
Groups of 4-week-old C57BL/6 mice were challenged intracerebrally (i.c.) with 30 μl JEV-WT or JEV-NS5-M19A for neurovirulence testing or intraperitoneally (i.p.) with 500 μl JEV-WT or JEV-NS5-M19A and i.c. injected with 30 μl PBS (i.p. plus i.c.) to damage the blood-brain-barrier as previously described [85] for neuroinvasiveness testing [86].
A human IL-6 ELISA kit (BMS213INST; eBioscience, San Diego, CA, USA) was used to detect IL-6 secretion in PA-BSA treated A549 cells. A mouse IL-6 ELISA kit (EM2IL6; Thermo Fisher Scientific, Waltham, MA, USA) was used to detect IL-6 secretion in mouse sera samples.
Cells in 96-well plates were mixed and incubated with an equal volume of CellTiter-Glo Reagent (G7572; Promega, Fitchburg, WI, USA) for 12 min, and then ATP levels were determined by measurement of luminescent signal.
Data are shown as mean±SD. The data for the AUC OCR was compared by ANOVA and post-hoc Tukey test with use of Prism 4 (GraphPad; La Jolla, CA, USA). The two-tailed Student t test was used for comparisons between 2 groups. The p values of survival curves were analyzed by the log-rank test using SigmaPlot 10 (Systat Software; San Jose, CA, USA). P < 0.05 was considered statistically significant.
|
10.1371/journal.pgen.1004752 | Gene Pathways That Delay Caenorhabditis elegans Reproductive Senescence | Reproductive senescence is a hallmark of aging. The molecular mechanisms regulating reproductive senescence and its association with the aging of somatic cells remain poorly understood. From a full genome RNA interference (RNAi) screen, we identified 32 Caenorhabditis elegans gene inactivations that delay reproductive senescence and extend reproductive lifespan. We found that many of these gene inactivations interact with insulin/IGF-1 and/or TGF-β endocrine signaling pathways to regulate reproductive senescence, except nhx-2 and sgk-1 that modulate sodium reabsorption. Of these 32 gene inactivations, we also found that 19 increase reproductive lifespan through their effects on oocyte activities, 8 of them coordinate oocyte and sperm functions to extend reproductive lifespan, and 5 of them can induce sperm humoral response to promote reproductive longevity. Furthermore, we examined the effects of these reproductive aging regulators on somatic aging. We found that 5 of these gene inactivations prolong organismal lifespan, and 20 of them increase healthy life expectancy of an organism without altering total life span. These studies provide a systemic view on the genetic regulation of reproductive senescence and its intersection with organism longevity. The majority of these newly identified genes are conserved, and may provide new insights into age-associated reproductive senescence during human aging.
| Female reproductive senescence is one hallmark of human aging, and as the germline ages, there is increased incidence of chromosome non-dysjunction and DNA damage. Delayed childbearing is a general feature in modern society, resulting in high risk of infertility, miscarriage and birth defects. Thus, understanding the molecular mechanisms regulating reproductive senescence and its association with somatic senescence is increasingly relevant to human health, and will shed light on the prolongation of reproductive longevity and the improvement of post-reproductive health. Here we conducted a genomic screen in Caenorhabditis elegans, searching for genetic regulators of reproductive senescence. We identified 32 gene inactivations that extend reproductive lifespan. Functional characterization of these genes has revealed their interactions with insulin/IGF-1 and TGF-β signaling pathways, their effects in different genders on the regulation of reproductive longevity, and their implications in the control of healthy life expectancy. Many of these genes are conserved between worms and humans. Our studies thus provide new insights into the molecular control of reproductive aging and the mechanistic link between reproductive senescence and organism longevity.
| Age-associated reproductive decline is one of the first aging phenotypes to manifest in humans. As women age, they experience both a decline in fertility and an increased risk of miscarriage and birth defects [1], [2], and completely cease reproduction (reproductive senesce) after menopause. The mean age of natural menopause is 50–51 years in the western world, with a variation from 40–61 years [3]. Studies of menopausal onset demonstrated an association between mothers and daughters and also between sister pairs, suggesting the involvement of genetic factors in regulating reproductive senescence [4], [5], [6]. The genetic contributions have been estimated to range from 30 to 85% [4], [5], [6]. However, the exact genes and signaling pathways that regulate the onset and progression of reproductive senescence remain largely unknown.
C. elegans has been proven an effective model for studying reproduction and aging processes using various genetic approaches, genome-scale screens and live microscopy imaging. Although evolutionarily distant, C. elegans and humans share many conserved molecular pathways. C. elegans also undergoes reproductive senescence and ceases reproducing progeny after one third of their lifespan [7]. Like the increase in chromosome nondisjunction as human females age, aging C. elegans exhibits increased chromosome nondisjunction during aging [8], suggesting conserved consequences of germline aging. Several endocrine factors were shown to influence the onset of reproductive senescence in C. elegans, including mutations in the daf-2/insulin/IGF-1 receptor gene, and the sma-2/TGF-β receptor linked Smad transcription factor gene [7], [8], [9]. Inactivating these genes significantly increases reproductive lifespan, and improves the quality of oocytes in later age [7], [8], [9]. Importantly, these genetic factors are also implicated in the regulation of mammalian reproductive lifespan and oocyte quality maintenance [10], [11].
Reproductive aging closely interacts with the decline of somatic health. The positive association between late reproduction and longevity has been observed in various invertebrate and vertebrate animals models and in human population studies. In Drosophila melanogaster, long-lived strains can be generated after more than 25 generations of artificial selection of late progeny produced by aged females, which show extended mean lifespan exceeding the maximum lifespan of the original parental strains [12], [13], [14], [15]. In wild chimpanzees, age-associated fertility declines are well correlated with decrease in survival probability [16]. In women, the age at natural menopause is inversely correlated with the rate of all-cause mortality during aging, and females who reproduced their last offspring at advanced age (>50 years) are far more likely to survive to 100 years old [17], [18]. Despite these association studies, the molecular mechanisms underlying the intersection between reproductive aging and somatic aging remain unclear.
Here, we report a comprehensive functional genomic screen to search for genes regulating reproductive senescence in C. elegans, and identified 32 gene inactivations that extend reproductive lifespan. Genetic epistasis analysis revealed interactions of some of these genes with insulin/IGF-1 and TGF-β endocrine signaling pathways. We also characterized their requirement and sufficiency in either hermaphrodites or males to promote reproductive longevity. Further demographic analysis showed that 22 of these gene inactivations reduce the mortality rate at young adulthood with or without increasing total life expectancy. Together these new reproductive aging regulators provide an entry point to further characterize the molecular control of reproductive aging, and reveal important mechanistic insights into the intersection between reproductive senescence and organism longevity.
Wild type C. elegans hermaphrodites produce progeny for 7 days at 20°C or for 3 days at 26°C, before they undergo reproductive senescence and cease reproduction. To identify genetic regulators of reproductive aging, we designed a genome-wide RNAi screen to identify those gene inactivations that postpone reproductive senescence in C. elegans hermaphrodites. To facilitate high-throughput screening, we used a temperature-sensitive embryonic lethal mutation in a collagen gene, sqt-3(e2117), to destroy all progeny produced during the normal phase of C. elegans reproduction so that we could observe only those viable progeny produced after normal reproductive senescence. In the RNAi screen (Figure S1A), C. elegans genes were inactivated one gene per well by feeding an Escherichia coli strain expressing an approximately one kilobase C. elegans segment of double stranded RNA to the sqt-3(e2117) mutant animals starting at the first larval stage (L1). The animals were raised at 15°C during larval development, and shifted to the 26°C non-permissive temperature at the fourth larval stage (L4). At this non-permissive temperature, the eggs that are laid during the normal reproductive period are killed by the temperature-sensitive sqt-3 aberrant collagen mutation. After the normally 3-day reproductive period at 26°C, the adults were shifted back to 15°C on the fourth day, and screened for the presence of live progeny in the well three days later. As expected, we detected no live progeny from the control animals that were fed with bacteria expressing no dsRNA or “non-scoring” dsRNAs (Figure S1B). Reduction of daf-2 activity by mutation delays the onset of reproductive senescence and extends healthy reproductive lifespan [7]. As a positive control, we scored about 15 live progeny generated per well by the normally reproductively senescent adults fed with daf-2 RNAi bacteria (Figure S1B). This proves the principle of the genomic screening strategy to systemically explore the regulatory mechanisms of reproductive senescence. By screening 18,413 such dsRNAs, we could survey 94% of C. elegans genome for gene inactivations that like daf-2 gene inactivation delay reproductive senescence.
Out of 18,413 gene inactivations screened, 58 candidates were identified that extend reproductive longevity of sqt-3 mutant animals (Figure S1C and Table S1). These gene inactivations were further examined in wild type and in an RNAi hypersensitive strain, nre-1(hd20)lin-15b(hd126) where RNAi potency is enhanced, especially in neurons [19]. We found that 32 gene inactivations prolong reproductive lifespan by more than 25% in the nre-1(hd20)lin-15b(hd126) background (Figure 1A and Table S2), and increase late reproduction of progeny after the first 3 days of adulthood (Figure S2). Twenty-six of the gene inactivations extend reproductive lifespan nearly equivalently in the enhanced neuronal RNAi background and in wild type; the 6 genes that show phenotypes only in the enhanced neuronal RNAi strain background may act in neurons because wild type C. elegans generally does not silence neuronal gene functions by RNAi (Figure 1B and Table S3).
One manner by which gene inactivations can appear to increase reproductive longevity is if that gene inactivation slows down the developmental process. That is, if development of the animal or its germline during larval stages is progressing at a decreased rate, the onset of reproduction as well as its cessation may be delayed. On the other hand, if time to initial reproduction is normal, but the cessation of reproduction is delayed, the gene is more likely to regulate reproductive senescence. To distinguish these two possibilities, we examined the time needed to develop from the L1 stage to the onset of reproduction in adulthood for each of the gene inactivations. Nearly all of the reproductive senescence candidate gene inactivations have a normal developmental rate (Figure S3), suggesting that metabolic rate or other gross features of developmental control are not affected by these gene inactivations. Three gene inactivations prolong the developmental time to adulthood, but to a much lesser extent than their effects on the reproductive lifespan extension (Figure S3). Therefore, most of the identified genes indeed affect the program of reproductive senescence, and their effects on reproductive longevity are not simply the result of a developmental delay.
Functional annotation analysis classified these newly identified genes into different categories, such as signaling transduction, gene expression/translation control, metabolic maintenance, ion transport and innate immunity defense (Figure 2A). To further understand the regulatory mechanisms of these genes, we examined their interactions with the known reproductive longevity pathways regulated by insulin/IGF-1 and TGF-β signaling [7], [8], [9]. Mutations in the insulin/IGF-1 receptor, daf-2(e1370), cause increased reproductive lifespan, which is fully suppressed by a mutation in the negatively regulated FoxO transcription factor, daf-16(mgDf47) [7]. To analyze the functions of the identified genes in insulin/IGF-1 signaling, we inactivated the newly identified reproductive senescence genes in the sqt-3(e2117) strains carrying either daf-2(e1370) or daf-16(mgDf47) mutation, performed the similar temperature switching procedure as shown above, and examined whether reproductive senescence could be delayed. As shown in Figure 2B, ten gene inactivations delay reproductive senescence in the background with either daf-2 or daf-16 mutation, suggesting that these genes regulate reproductive senescence independently from insulin/IGF-1 signaling. Among the other 22 genes, nine of them act independently of daf-16, but their inactivations have no additive effects with the daf-2 mutation; the effects of four genes are dependent on daf-16, but are additive to daf-2; and the other nine gene inactivations fail to delay reproductive senescence in either mutant background. These 22 genes are likely acting in insulin/IGF-1 signaling, but at different positions in the pathway.
TGF-β signaling also regulates C. elegans reproductive aging; mutations in sma-2(e502), which encodes the TGF-β-receptor-regulated Smad transcription factor, increase reproductive lifespan more than two-fold [8], [9]. We found that five of the identified gene inactivations interacting with insulin/IGF-1 signaling can further extend the reproductive longevity of the sma-2 mutant, as is also true for the daf-2 gene inactivation (Figure 2B and 2C, highlighted with pink). On the other hand, eight other gene inactivations delay reproductive senescence of the daf-2 and daf-16 mutants, but fail to do so in the sma-2 mutant (Figure 2B and 2C, highlighted with blue). These data suggest that insulin/IGF-1 and TGF-β signaling modulate the process of reproductive aging in parallel, which is consistent with previous findings [8]. However, our results also show that these two pathways converge on a large group of common downstream targets, considering 17 of the candidate genes interact with both pathways (Figure 2B and 2C, highlighted with purple).
The two most potent effectors, nhx-2 and sgk-1, however act independently of either pathway (Figure 2B and 2C, highlighted with green). In addition to insulin/IGF-1 and TGF-β signaling, caloric restriction using eat-2 mutants also prolongs reproductive lifespan in C. elegans [7]. We found that RNAi inactivation of either nhx-2 or sgk-1 further enhances reproductive lifespan extension in the eat-2(ad465) mutant (Figure 2D). Because the eat-2 mutation causes caloric restriction via a disability in the actual ingestion of bacteria, it is unclear if these enhancing gene inactivations act via a regulatory mechanism on the caloric restriction pathway or in parallel to that pathway. An sgk-1(mg455) null mutation and nhx-2 RNAi inactivation exert additive effects on reproductive lifespan extension (Figure 2E), suggesting their parallel functions in regulating reproductive aging. The mammalian homologs of these two genes are the sodium/hydrogen exchanger and the serum-and glucocorticoid-inducible kinase respectively, which both regulate sodium reabsorption [20], [21]. Our results suggest that sodium homeostasis may play a crucial role in the regulation of reproductive senescence.
Reproductive activities are under the control of complex interaction between germline and somatic tissues. To test whether the candidate genes function in the germline or in somatic tissues to regulate reproductive longevity, we inactivated the 32 reproductive senescence genes in rrf-1(pk1417) mutant animals, for which RNAi is predominantly effective only in the germline with a minor effect in the intestine [22], [23]. We found that only 10 gene inactivations prolong the reproductive lifespan of the rrf-1 mutant animals (Figure 3 and Table sS4). Among them, the effect of nhx-2/sodium/hydrogen exchanger, moma-1/apolipoprotein O homologue or daf-2/insulin-like receptor in the rrf-1 background is much weaker than that in wild type (Figure 3), suggesting that their functions in somatic tissues are also involved in the regulation of reproductive longevity. Moreover, the other 22 gene inactivations failed to increase reproductive lifespan in the rrf-1 mutant; those genes likely regulate reproductive longevity systemically via their effects in somatic tissues. Together these studies indicate the significance of somatic functions in the regulation of reproductive senescence.
Self-fertilizing C. elegans hermaphrodites generate a limited number of sperm by virtue of a pulse of spermatogenesis before a longer run of oogenesis, which constrains their reproductive capacities. However when mated to males which produce far more sperm over a longer period, hermaphrodites use male sperm preferentially, produce double the number of progeny, and reproduce for a much longer time period [7]. To test whether the newly identified genes exert effects on the reproductive longevity of mated animals, we first inactivated these genes in hermaphrodites by RNAi and then crossed those animals with males feeding with control bacteria expressing no dsRNAs. Under this condition, sperm from the males are not affected by RNAi inactivation. We found that 19 genes when inactivated only in mated hermaphrodites are sufficient to promote reproductive longevity (Figure 4A and Table S5), suggesting that these gene inactivations predominantly delay age-associated oocyte senescence.
For the other 13 candidate genes, their inactivations in hermaphrodites alone are not sufficient to enhance reproductive longevity after mating. However, when they are also inactivated in males, 11 of them increase the reproductive lifespan of the mated hermaphrodites (Figure 4B and Table S5), suggesting the requirement of their functions in males to modulate the process of reproductive senescence. To test whether these 11 genes affect sperm production, we examined the effects of their inactivations on brood size in self-fertilizing hermaphrodites where the total number of progeny is determined by sperm quantity. We found that none of these gene inactivations increase total brood size (Figure S4). Therefore, the effect of these genes on reproductive lifespan extension is not simply a result of increased sperm quantity.
We also found five genes whose inactivation only in males are sufficient to prolong the reproductive lifespan of hermaphrodites after mating, including daf-2, F54E2.1, Y46G5A.20, srz-1 and F20B10.3 (Figure 4C). Inactivation of srz-1 and F20B10.3 in males alone prolongs reproductive lifespan by 14% and 13%, respectively (Figure 4C), which are comparable to the 15% and 13% extension caused by their inactivation in both hermaphrodites and males (Figure 4B). On the other hand, male-only inactivation of Y46G5A.20 results in 18% reproductive lifespan extension (Figure 4C), which is significantly reduced compared to the 42% extension when Y46G5A.20 is inactivated in both hermaphrodites and males (Figure 4B). For daf-2 and F54E2.1, their inactivation in either hermaphrodites or males alone is sufficient to prolong reproductive lifespan (Figure 4A and 4C). Interestingly, when mated with the daf-16(mgDf50) mutant hermaphrodites, the male-only daf-2 inactivation failed to prolong reproductive lifespan (Figure 4D). Together, these studies suggest that the effects of the gene inactivations on reproductive longevity occur not only in self-fertilizing hermaphrodites, but also in mated animals. Furthermore, sperm and seminal fluid transferred by mating may induce a humoral response that regulates the process of reproductive senescence in hermaphrodites systemically, and insulin/IGF-1 signaling is involved in this humoral response.
To characterize the interaction between reproductive aging and somatic aging, we inactivated the 32 genes in the nre-1(hd20)lin-15b(hd126) RNAi hypersensitive strain, and examined their effects on organismal lifespan and on age-specific patterns of mortality. In the fitted mortality rate curve under the Gompertz-Makeham model (Figure 5A), the slope of the curve defines the demographic rate of aging (RoA) showing the rate of increase in mortality with age; whereas the intercept with the y-axis defines initial mortality rate (IMR) or “frailty” that represents the mortality rate at the defined time zero, in this case young adulthood and is related to the baseline mortality [24].
We found that the majority (25) of the identified reproductive senescence genes affect somatic mortality trajectories (Figure 5B–D). Based on their effects on different aging-related parameters, these genes can be classified into three groups. First, gene inactivations of three genes, including daf-2, R07H5.9 and C05E11.6 decrease RoA without affecting IMR (frailty) (Figure 5B–D, Figure 6A–C). As a result, the median and maximum lifespan (defined here as age at 1% estimated survival) are both significantly increased upon their inactivations (Figure 6A–C). Secondly, gene inactivation of the two sodium homeostasis regulatory genes nhx-2 and sgk-1, reduce IMR but not RoA (Figure 5B–D, Figure 6D and 6E). Their inactivations also lead to median and maximum lifespan extension (Figure 6D and 6E), which is consistent with previous findings [25], [26]. The third group includes 20 gene inactivations that decrease IMR, but increase RoA (Figure 5B–D). As a result, these genes have little effect on median lifespan and cause no change in maximum lifespan (Figure 5B–D, one example shown in Figure 6F). Together, these results suggest that most of the gene inactivations that delay reproductive senescence decrease the baseline probability of death and consequently postpone the age-associated rise in mortality.
Our genome scale screen for gene inactivations that delay reproductive senescence identified 32 genes that normally function to mediate reproductive senescence. Characterization of synergies and epistasis of the reproductive lifespan extension induced by these gene inactivations in strains also carrying mutations in insulin/IGF-1 and TGF-β signaling pathways places them into distinct classes. Insulin/IGF-1 and TGF-β signaling are two independent mechanisms to modulate reproductive senescence in C. elegans, which are also implicated in regulating mammalian reproductive span and oocyte quality maintenance [10], [11]. Most of our newly identified genes interact with either or both of these two signaling pathways.
sgk-1 and nhx-2 are the two most potent effectors identified from the screen and regulate reproductive senescence through a mechanism independent of insulin/IGF-1 signaling, TGF-β signaling or caloric restriction. The mammalian homologues of sgk-1, the serum and glucocorticoid activated kinase and nhx-2, a sodium/hydrogen exchanger both modulate sodium reabsorption [20], [21]. Thus our data suggest an involvement of sodium homeostasis in the regulation of reproductive senescence in C. elegans. Up-regulation of mammalian SGK-1 has been implicated in reproductive failure in women [27]. Therefore, although worms and humans have very distinct reproductive strategies, they may share some common genetic regulatory factors. Twenty-one of the newly identified C. elegans genes from the screens are well conserved in human (Table S6), and may provide new insights into the age-associated reproductive senescence in human.
Reproductive success requires proper functions of both oocyte and sperm, and their coordination. In several species including C. elegans, sperm can signal to oocyte and to the somatic gonad, and sperm-derived signals are responsible for promoting both oocyte maturation and ovulation [28]. Males can also influence hermaphrodite lifespan in C. elegans [29], [30], through secreting diffusible compounds [29]. We found 19 gene inactivations in hermaphrodites only that can promote reproductive longevity. In these cases, mated hermaphrodites use the sperm from the control males not undergoing gene inactivation. Thus the extension of reproductive lifespan under these conditions is caused by the gene inactivations that regulate oocyte activities directly or indirectly. There are 11 genes that require inactivation in both hermaphrodites and males to prolong reproductive lifespan, suggesting the significance of sperm and/or seminal fluid in regulating reproductive aging. Among those 11 genes, only ilys-3 and F25H8.1 are sufficient to prolong reproductive lifespan when inactivated in the rrf-1 mutants, which may exert cell-autonomous effects on the activity of sperm and/or seminal fluid. However, the other nine genes are likely to function in somatic tissues to affect sperm and seminal fluid activity cell-nonautonomously. On the other hand, we also found that five gene inactivations in the male alone, including daf-2, F54E2.1, Y46G5A.20, srz-1 and F20B10.3, are sufficient to prolong the reproductive lifespan of hermaphrodites carrying non-gene inactivated oocyte. These data suggest that sperm produce humoral signals to actively modulate the process of reproductive senescence, and these five genes are likely involved in the sperm humoral signaling. These humoral signals are likely to be secreted steroids and/or peptides; while it is also possible that the sperm or seminal fluid from the males undergoing RNAi treatment may bring gene-inactivating siRNAs into the hermaphrodite. We expect that future characterization of the regulatory mechanisms of those genes will provide insights into the humoral communication between males and females.
Reproductive success is the currency of evolution. To achieve this goal, somatic physiology must be well maintained to support reproductive activities. Thus it is expected that the onset of reproductive senescence would affect somatic maintenance and organism aging. Our data support this, and further reveal that delaying reproductive senescence may have predominant effects on healthy lifespan rather than total lifespan. In our demographic analysis, 3 gene inactivations reduce the mortality rate increase with age (RoA), while 22 gene inactivations decrease the baseline mortality (IMR). IMR measures the initial vulnerability to physiological damage, and its decrease is related to a delay in the onset of chronological aging, or an increase in the initial quality of the organism [24], [31]. Thus, these 22 gene inactivations likely increase healthy life expectancy without necessarily slowing the aging process. Further characterization of these genes may help us find new ways to improve healthy life expectancy, and understand the crucial link between late reproduction and organism longevity during evolution.
N2 Bristol was used as the wild-type strain. Other C. elegans mutant alleles: sqt-3(e2117)V, nre-1(hd20)lin-15b(hd126)X, daf-2(e1370)III, daf-16(mgDf47)I, sma-2(e502)III, and rrf-1(pk1417)I, eat-2(ad465)II, daf-16(mgDf50)I.
We carried out a large-scale RNAi screen using the sqt-3(e2117) strain. RNAi bacteria were cultured 12 h in LB with 50 µg/ml ampicillin and seeded onto 24-well RNAi agar plates containing 5 mM isopropylthiogalactoside (IPTG). The plates were allowed to dry in a laminar flow hood and incubated at room temperature overnight to induce dsRNA expression. Approximately 20∼30 synchronized L1 larvae were placed onto agar plate wells where each E. coli strain expressing a distinct C. elegans dsRNA corresponding to one gene, and the worms were allowed to develop at 15°C till L4 and shifted to 26°C. After 4 days, the worms were shifted back to 15°C and scored for progeny production after 2∼3 more days of incubation. Worms feeding on bacteria carrying the empty vector (L4440) were used as control. We routinely screened ∼2000–3000 RNAi clones in one experiment. RNAi wells in which live progeny were observed were scored as positives. These “positive” RNAi clones were retested at least three more times using similar high-throughput screening strategy described above. RNAi clones that were scored as positive in all the rescreening tests were deemed “primary positives”. Primary positives were retested three times in RNAi self-fertilizing reproductive lifespan assays (see below) using the nre-1(hd20)lin-15b(hd126) or wild type N2 strain. The RNAi clones that were scored positive three times in the reproductive lifespan assays define the final list of enhanced reproductive longevity gene inactivations.
daf-2(e1370);sqt-3(e2117), daf-16(mgDf47));sqt-3(e2117) and sma-2(e502);sqt-3(e2117) double mutants were generated. Approximately 20∼30 synchronized L1 larvae were placed onto agar plate wells where each E. coli strain expressing a distinct C. elegans dsRNA corresponding to one gene, and the worms were allowed to develop at 15°C till L4 and shifted to 26°C. For daf-2(e1370);sqt-3(e2117), the worms were shifted back to 15°C after 9 days and scored for progeny production after 2∼3 more days of incubation. For daf-16(mgDf47));sqt-3(e2117) and sma-2(e502);sqt-3(e2117), the worms were kept at 26°C for 4 days and 9 days, respectively. Worms feeding on bacteria carrying the empty vector (L4440) were used as control, in which no progeny was detected after temperature shifting back to 15°C.
RNAi bacteria were prepared as described above in 60 mm RNAi agar plates. Approximately 20∼30 synchronized L1 larvae were placed onto the RNAi containing plates and allowed to develop at 20°C until the L4 stage. For each genotype and RNAi treatment, three plates were prepared. At the L4 stage, 10 larvae from each plate were hand- picked and transferred to a fresh RNAi agar plate containing the same RNAi bacteria. The animals then were kept at 20°C and were transferred to fresh RNAi plates every day. After transferring, the old plates were kept at 20°C and scored for the number of progeny two days later. Reproductive lifespan is defined as the first day of reproduction (reproductive lifespan = 1) to when no progeny were scored. The reproductive lifespan of RNAi-treatment groups are compared with that of the control groups (bacteria carrying the empty vector) using a student T-test.
RNAi bacteria were prepared as described above in 35 mm RNAi agar plates. Synchronized L1 larvae (nre-1(hd20)lin-15b(hd126) or daf-16(mgDf50)) were placed onto the RNAi containing plates and allowed to develop at 20°C. Next, L4 hermaphrodites were mated to young males (nre-1(hd20)lin-15b(hd126) at a 1∶2 ratio for 48 hours before being transferred to individual RNAi containing plates. Successful mating was determined by the production of male progeny each day. The individual animal was transferred to fresh plates daily until no progeny scored for at least two days. For each individual, the last day of live progeny production was recorded as the day of reproduction cessation. For each experiment, at least 10 individual hermaphrodites were included. Statistical analyses were performed using SPSS software (http://www-01.ibm.com/software/analytics/spss/). Each RNAi-treatment population is compared with that of the population treated with control RNAi (bacteria carrying the empty vector) using a log rank test.
RNAi bacteria were prepared as described above in 60 mm RNAi agar plates. Approximately 30∼40 synchronized L1 larvae (nre-1(hd20)lin-15b(hd126)) were placed onto RNAi-containing agar plates, allowed to develop at 20°C. The animals then were kept at 20°C and transferred to fresh RNAi plates every two days or every seven days during or past the reproductive period respectively. For each RNAi treatment, 3 plates (around 100 worms) were prepared and scored every day by gentle prodding with a platinum wire to test for viability. Lifespan is defined as the first day of adulthood (adult lifespan = 1) to when they were scored as dead. The same experiment was performed three times independently.
We generated lifetables from Nx, dx, and cx, the number entering, dying in, or censored in each age interval, respectively. Probability of death (qx) was estimated as qx = dx/Nx, and force of mortality as mx = −ln(1−qx). The Gompertz-Makeham model for age-specific mortality M(x) = M0⋅exp(G⋅x)+M∞, with initial mortality M0, rate of aging G, age x, and age-independent mortality M∞ can be shown to give the survival proportion S(x) = exp[(M0/G)(1−exp(G⋅x)−M∞⋅x]. We performed non-linear-least-squares fits to S(x) using a Trust-Region Reflective Newton algorithm implemented by the MATLAB fit() function (The MathWorks, Natick, MA). Kaplan-Meier estimates of the cumulative distribution (survival) function were computing using the MATLAB (The Mathworks, Natick, MA) function ecdf(). We also performed a log-rank test to compare each RNAi gene inactivation to it's associated vector control within the same experiment, using a Bonferroni step-down multiple testing p-value adjustment.
10 synchronized L4 hermaphrodite larvae (nre-1(hd20)lin-15b(hd126)) were transferred to fresh RNAi plates every day and the number of progeny was counted daily until reproduction cessation. The total number of progeny per hermaphrodite was calculated. The experiments were performed three times at 20°C. The RNAi-treatment groups are compared with the control groups (bacteria carrying the empty vector) using a student T-test.
10 synchronized L1 larvae (nre-1(hd20)lin-15b(hd126)) were allowed to develop at 20°C. After becoming L4, the animals were examined every half hour to note the transition to adulthood (oocytes in the germline). If any, the adult individual would be removed from the population. The time from L1 to adulthood when the animal was removed was recorded for each individual. The RNAi-treatment groups are compared with the control groups (bacteria carrying the empty vector) using a student T-test.
|
10.1371/journal.pbio.1000103 | Direct Observation of ATP-Induced Conformational Changes in Single P2X4 Receptors | The ATP-gated P2X4 receptor is a cation channel, which is important in various pathophysiological events. The architecture of the P2X4 receptor in the activated state and how to change its structure in response to ATP binding are not fully understood. Here, we analyze the architecture and ATP-induced structural changes in P2X4 receptors using fast-scanning atomic force microscopy (AFM). AFM images of the membrane-dissociated and membrane-inserted forms of P2X4 receptors and a functional analysis revealed that P2X4 receptors have an upward orientation on mica but lean to one side. Time-lapse imaging of the ATP-induced structural changes in P2X4 receptors revealed two different forms of activated structures under 0 Ca2+ conditions, namely a trimer structure and a pore dilation-like tripartite structure. A dye uptake measurement demonstrated that ATP-activated P2X4 receptors display pore dilation in the absence of Ca2+. With Ca2+, the P2X4 receptors exhibited only a disengaged trimer and no dye uptake was observed. Thus our data provide a new insight into ATP-induced structural changes in P2X4 receptors that correlate with pore dynamics.
| ATP is not only a source of intracellular energy but can act as an intercellular signal by binding membrane receptors. Purinergic receptors, which bind with nucleotides including ATP are known as P2 receptors and are divided into two types: ion channel-type P2X receptors and metabotropic-type P2Y receptors. P2X receptors are thought to undergo conformational changes in response to ATP binding, leading to the opening of transmembrane channels, through which cations enter the cells. A growing body of evidence shows that P2X receptors control various physiological and pathophysiological cellular responses. However, the receptor structure and the conformational changes it experiences upon stimulation remained to be clarified. Here, we employed an atomic force microscope (AFM) to observe P2X receptor behavior at the single channel level. We chose to analyze the P2X4 receptor, because it is known to increase the transmembrane pore size (i.e., pore dilation) in the absence of extracellular calcium. Activated P2X4 receptor exhibited a trimeric topology with a pore-like structure in the center. When calcium was present the receptor exhibited a trimer without a pore structure at its center. These structural changes corresponded well with the changes of ion permeability of P2X4 receptor.
| P2X receptors (P2XRs) are cell-surface ATP-gated cation channels, and seven subtypes (P2X1–7) are known [1]. One functional P2XR channel is composed of three subunits. Each P2XR subunit is predicted to have a large extracellular domain (ECD), two transmembrane-spanning domains (TMD), and N and C termini intracellular domains (ICD) [1]. It has been suggested that the second half of the ECD (residues 170–330) has sequence and secondary structure similarities to the catalytic site of class II aminoacyl-tRNA synthetase [2]. A six-stranded antiparallel β-pleated sheet structure is believed to exist in the ECD of P2XRs. 3-D homology modeling in P2X4Rs suggests that this region coordinates ATP binding and the allosteric coupling of the conformational changes in the ATP binding domain with corresponding changes at the transmembrane channel gate through a linker region (the α-helix between the β6 strand and TM2 region) [3]. In addition to the allosteric coupling of the ATP-binding sites at ECDs and the channel gate at TMD, P2XRs have different permeability states that were originally discovered by Cockcroft and Gomperts [4]. With P2X4Rs, extracellular Ca2+ levels greatly affect the permeability dynamics [5]. In the presence of Ca2+, P2X4R only opens a small cation-permeable channel pore but in the absence of extracellular Ca2+ it forms a larger pore that allows larger molecules including N-methyl-D-glucamine (NMDG)+, propidium iodide, and ethidium bromide (EtBr) to pass. Although there is a functional relationship between ECD and TMD, the ATP-induced structural changes in ECD are poorly understood. Recent extensive studies by Khakh's group have clearly demonstrated the allosteric coupling of ICDs and the ion channel permeability of P2XRs [6,7]. These results strongly support the hypothesis of the allosteric coupling of channel pores in TMD and other domains including ECDs.
In recent structural studies of P2XRs two approaches have been used: electron microscopy (EM) and atomic force microscopy (AFM). In EM, single particle averaging analysis and the Ni-NTA gold labeling of human P2X4Rs have clearly demonstrated the distance between the C-terminal tails, the molecular volume, and the 3-D structure [8]. In AFM research, an antibody tagging study has revealed the trimer structure of P2XRs [9,10]. AFM has the important advantage of allowing proteins to be observed under liquid conditions, and this makes it possible to activate P2XRs by ATP during AFM studies. In an AFM study combined with ATP treatment, P2XRs exhibited a pore-like structure [11]. In addition to drug treatment, AFM can be used for imaging both lipid bilayers [12] and proteins inserted in lipid membranes [13]. Extensive AFM studies by Engel and Müller's groups have obtained high-resolution topographs of many proteins including aquaporin [14], connexin [15], F-ATP synthase [16], and tubulin [17]. A recent study by Cisneros clearly demonstrated the topography of orientation regulated and covalently assembled homotrimer OpmF proteins [18]. In their report, the authors employed the single particle correlation averaging method to obtain 3-fold symmetrized images of OmpF trimer that are identical to the topographs of 2-D crystals of OmpF. Because many P2XR channels are also homotrimers, this approach can be used for the high-resolution imaging of P2XRs. Although the use of AFM provides significant advantages the imaging speed is usually very slow (several tens of seconds). Many ion channel reactions occur in less than a second, so fast scanning is essential for observing the P2XR reaction with AFM. To address this issue, we employed a recently developed fast-scanning AFM [19] that allows us to observe biological molecules including nucleic acids [20], lipids [12], and proteins [21,22] at high temporal resolution. Fast-scanning AFM in combination with single particle averaging is considered a powerful tool for analyzing single P2X4R channels with high spatial and temporal resolution.
The expression of rat P2X4R protein in human 1321N1 astrocytoma cells was estimated by western blotting. P2X4R was detected only in P2X4R gene-overexpressed cells (Figure 1A). In silver-stained native PAGE, only one band corresponding to a trimer (about 150 kDa) (Figure 1B) was observed. The same protein analyzed by SDS-PAGE and silver-staining exhibited a band corresponding to a monomer (about 50 kDa). For the AFM analysis of P2X4Rs, we used freshly cleaved mica as a substrate because it has an atomically flat surface and is usually used for protein observation with AFM. All the AFM images were presented as gray-scale height images. In many cases, the P2X4R particles were only loosely attached to the uncoated mica and so they moved during the AFM observation. To obtain a stronger attachment for electrostatic interactions, we coated the mica with positively charged poly-D-lysine (PDL) (1 mg/ml, 30 min at room temperature [RT]) and set the pH of the imaging buffer (AFM imaging buffer A) at 8.0 because the isoelectric point of P2X4R is pH 7.41. All the P2X4Rs on the PDL-coated mica were observed in AFM imaging buffer A. Under this condition, the P2X4Rs were attached stably to the substrate (Figure 2A). The P2X4R control particles were relatively homogenous and nearly all circular, ellipsoid, or triangular with obtuse angles (Figure 2B, upper panels). PDL-polymers were also observed (Figure 2B[ii], arrows). In this study, we defined the dimensions of the P2X4Rs as their diameter and height on the basis of our criteria (see also Materials and Methods and Figure S1). The nonstimulated P2X4Rs had a diameter of 12.6 ± 0.2 nm (mean ± standard error of the mean [SEM]) (n = 200) and a height of 2.3 ± 0.1 nm. To observe activated P2X4Rs, we added ATP before the AFM observation. ATP did not induce any significant changes at 100 μM (unpublished data), but the P2X4Rs changed greatly at 1 mM (Figure 2B, lower panels). Under this condition, at least several minutes of ATP treatment was required before the P2X4Rs underwent structural changes. After the structural changes caused by 1 mM ATP, most of the P2X4Rs appeared to be trimers (84.9 ± 5.0%, n = 393) (Figure 2C). The ATP-treated P2X4Rs had a diameter of 14.2 ± 0.2 nm (n = 205) and a height of 3.0 ± 0.1 nm. The diameter of one lobe in a P2X4R trimer was 5.9 ± 0.2 nm (n = 40). To obtain clear topographs of P2X4Rs, we averaged single P2X4R images by using the same approach employed by Cisneros et al. [18] and on the basis of our criteria (Figure S1). The nonsymmetrized averaging of ATP-treated P2X4Rs revealed a tripartite morphology (Figure 2D[i], right) that was enhanced by 3-fold rotational symmetrization (Figure 2D[ii], right). Nonstimulated P2X4Rs were circular or triangular with obtuse angles after averaging (Figure 2D, left panels). For averaging, we used the particles shown in Figure 2B(iii) (n = 60). Then, we checked whether these trimers were one unit of P2X4R trimers or simply three adjacent particles. If each lobe was an individual P2X4R trimer that was incidentally assembled into a trimer, the distance between lobes would not be significantly different from the distance between trimers. The distance between the lobes in a P2X4R trimer and the distance between two adjacent trimers were 8.7 ± 0.1 nm (n = 100, between lobes) and 35.5 ± 2.7 nm (n = 115, between trimers), respectively (Figure 2E). Sometimes, P2X4R particles on PDL-coated mica shifted position within the same scan area. In this situation, single lobes in a P2X4R trimer (15 min after 1 mM ATP treatment) did not move individually but moved along with the other two lobes (Figure 2F[i]). Enlarged images of single P2X4R trimer in a rectangle at 5 s are shown on the left in Figure 2F(ii). The nonsymmetrized and symmetrized averaging of ten particles in the same scan area at 0 s is shown in the center and on the right, respectively, in Figure 2F(ii).
To observe the ATP-induced continuous structural changes in P2X4Rs, we performed imaging using fast-scanning AFM with a scan rate of two frames per second. P2X4Rs were observed in AFM imaging buffer B. Under our conditions, faster scan rates than this degraded the signals and increased noise so that we were unable to obtain sufficient resolution. It is known that a mica surface is negatively charged [23], and so we used uncoated mica rinsed with a high concentration of KCl (1 M, 30 min at RT) to reduce electrostatic interactions between the mica surface and the ATP or P2X4Rs. Under this condition, many P2X4Rs shifted position during AFM imaging. To obtain a clear topology of P2X4R, ten P2X4R particles were averaged at the same time point. The resulting 3-fold symmetrized images of P2X4Rs clearly exhibited the structures at each time point. Before the uncaging (−2.5 to ∼0.0 s) of caged ATP (200 μM), P2X4R exhibited a circular structure (Figure 3, see also Video S1). At 0.5 s after uncaging, the P2X4R structure changed greatly and a clear trimeric structure was observed. After this change, the distances between individual lobes gradually increased (≈5 s). The conformational change in the nonsymmetrized P2X4R is also shown in Figure S2. The same reaction was reproduced in three independent experiments. Another result of the ATP-induced structural changes in P2X4R is shown in Figure S3. Some P2X4Rs were stable at one location during AFM imaging. Several examples of ATP-induced structural changes in a single P2X4R are shown in Figure S4. At a single particle level, although the P2X4R topologies were relatively blurry, individual subunits became visible after uncaging and appeared to move away from each other. When the ATP was washed off, the pore dilation-like structure returned to a circular structure (unpublished data).
To estimate the orientation of observed structures, P2X4Rs were reconstituted in a lipid bilayer. Figure 4A(i) is a diagram showing the predicted structure of a P2X4R subunit. A six-stranded anti-parallel β-plated sheet structure is reported to exist in the second half of the ECD in P2X4R subunits [2,3]. The entire structure of trimeric P2X4R is predicted on the basis of this homology modeling data, as shown in Figure 4A(ii). In AFM, this β-plated sheet structure should be observed as one large domain. Figure 4B shows our working hypothesis, which is that when P2X4Rs are reconstituted in a lipid bilayer and if they are inserted in an upward orientation, they should respond to ATP thus resulting in structural changes and increased Ca2+ permeability. When P2X4Rs were inserted in the lipid bilayer that formed on mica, the AFM images of P2X4Rs in membranes were similar to the P2X4Rs that were dissociated from the membrane. The P2X4Rs had circular structures in the control and trimeric structures after binding with ATP (200 μM, 1 min) (Figure 4C and 4D). P2X4Rs reconstituted in a lipid bilayer did not require as high a concentration of ATP as those on PDL-coated mica. Under this condition, the structures of most of the P2X4Rs (83.3 ± 5.4%, n = 70) changed into a tripartite form. The P2X4Rs in the membranes were 11.4 ± 0.3 nm in diameter and 5.8 ± 0.1 nm high (including the height of the membrane) in the control (n = 50) and 13.3 ± 0.3 nm in diameter and 6.1 ± 0.1 nm high after ATP addition (n = 100). The calculated height of the membrane was 4 nm. The AFM imaging of membrane-inserted P2X4Rs was performed in imaging buffer B.
For calcium imaging, the P2X4Rs were reconstituted in a lipid bilayer that was suspended over a 500 μm hole. The green fluorescence intensity of fluo-3 (50 μM, in hole) was significantly increased after ATP (100 μM) addition (Figure 4E). This green florescence was only detected in the hole (Fig. 4E[i]). The intensity of the green fluorescence increased rapidly for a few seconds after ATP addition and then increased gradually (Fig. 4E[ii], see also Video S2). The averaged trace was obtained from five individual experiments. An EtBr-dye uptake measurement was performed at the same time as the Ca2+ imaging. Here, no increase was observed in red fluorescence after ATP addition (Fig. 4E[iii]). The Ca2+ imaging was performed in Ca2+ imaging buffer.
In the time-lapse imaging of ATP-induced structural changes in P2X4R, we observed a characteristic pore dilation-like structure (Figure 3, ≈5.0 s). This pore dilation-like structure was also observed in membrane-reconstituted P2X4Rs (Figure 4D). Before the appearance of this structure, the P2X4Rs on the uncoated mica exhibited nondilated trimer structures (Figure 5A, center). We observed two P2X4R structures similar to these two different forms on PDL-coated mica (Figure 5B). 15 min after ATP (1 mM) addition, the P2X4Rs exhibited a nondilated trimer structure (Figure 5B, center) but they exhibited a pore dilation-like structure 30 min after ATP addition (Figure 5B, right). Then we estimated the dye uptake function of P2X4Rs using the same Ca2+ imaging system. EtBr-uptake imaging buffer containing no Ca2+ was used for this study. Here, ATP (100 μM) addition increased the red fluorescence intensity in the hole (Figure 5C, upper panels, see also Video S3). Under our conditions, the red fluorescence intensity started increasing within seconds of the ATP addition and then increased gradually (≈300 s) (Figure 5C, lower panel). When we measured dye uptake with 2 mM Ca2+ in an external solution, we observed no increase in red fluorescence intensity (Figure 4E[iii]). To confirm whether the effect of Ca2+ on dye uptake is related to the pore dilation-like structural changes, we compared the structures of P2X4Rs in the presence and absence of Ca2+. In this study, we used the same mica as we used for the time-lapse imaging, and we used AFM imaging buffer B or C for each condition. With 0 Ca2+, an averaged P2X4R image was obtained from 18 particles (Figure S5A[i]). The particles were selected from frames at least 5 s after uncaging. In this case, a pore dilation-like image was again obtained (Figure 5D[i]). In the presence of 2 mM Ca2+, no pore dilation-like averaged image was obtained but a nondilated trimer was observed (Figure 5D[ii]). An averaged P2X4R image was obtained from 18 particles (Figure S5A[ii]) at least 5 s after uncaging. The averaged images are obtained after 3-fold symmetrized averaging. Under both conditions, the majority of the P2X4Rs responded to ATP (0 Ca2+: 67.0 ± 2.8 %, n = 257; 2 mM Ca2+: 62.8 ± 2.7 %, n = 324). Nonsymmetrized averaging images of P2X4R under each condition are shown in Figure S5B.
Models of the ATP-induced structural changes of P2X4R based on our results are shown in Figure 6. In the control, three ECDs of P2X4Rs were close to each other and AFM revealed no individual subunits. Under this condition, neither Ca2+ nor EtBr can permeate the TMD pore. In the absence of Ca2+, the ECDs are disengaged and a tripartite topology was observed immediately after ATP binding (Figure 6A, center). Prolonged ATP treatment induces further disengagement of the three ECDs (Figure 6A, right). These two structures appear to correspond to the Ca2+ permeable and EtBr permeable states (Figure 6A, below). Under a 2-mM Ca2+ condition, P2X4R has a nondilated trimer structure regardless of the ATP exposure time (Figure 6B). In this situation, the TMD pores allow Ca2+ to permeate but not EtBr however it is unclear whether or not P2X4R is desensitized during ATP exposure.
Our main findings in this study are that (i) it is possible to achieve time-lapse imaging of the dynamic structural changes of P2X4Rs evoked by ATP; (ii) the three subunits are close to each other and it is impossible to observe individual subunits in the control but they disengage and move away from each other after ATP binding; and (iii) the two types of structural changes observed in AFM appear to correspond to two functional states, namely the Ca2+ permeable state and the dye permeable state.
Recent structural studies with direct imaging methods including EM and AFM or with other methods including fluorescence resonance energy transfer (FRET)-based analysis have provided strong motivation for structural studies of P2XRs. These reports clearly demonstrated trimeric stoichiometry using antibody-tagging [9,10] or Ni-NTA gold labeling on the His-tag of the C termini in P2XRs [8], and the shape, architecture, and size of P2X4Rs in a nonstimulated state and the distance between the C termini of P2XRs [8]. We needed to determine the way in which P2XRs change their entire structure in response to ATP binding. To address this issue we analyzed homotrimeric P2X4Rs. To this end, we overexpressed P2X4R gene in human 1321N1 astrocytoma cells. Because this cell does not express P2XRs [24], purified P2X4Rs from the membrane fraction of this cell are considered to form homotrimers. In fact, the purified P2X4R presented as a single band corresponding to a trimer (about 150 kDa) in native-PAGE but as a monomer (about 50 kDa) in SDS-PAGE, and purified P2X4R was functional as estimated in terms of Ca2+ permeability. We then observed P2X4Rs on mica but they did not attach to it stably. P2X4Rs on PDL-coated mica exhibited stable attachment but a high concentration of ATP was required to induce structural changes. The ATP has negative charges that may induce the strong attraction of ATP to the positively charged PDL. In fact, P2X4Rs reconstituted in a lipid bilayer or on mica without PDL coating responded to lower ATP concentrations. In addition to the high ATP concentration, a long period of ATP exposure was required when P2X4Rs were adsorbed on PDL-coated mica. This may be due to the strong attachment of P2X4Rs to mica. In a recent report, the N-terminal tagging of fluorescent proteins on P2X2Rs dramatically increased the ATP EC50 value, but this did not occur with small tetracystein (4C) tags labeled with fluorescein arsenical hairpin [7], implying that the limited spatial flexibility in the N-terminal domain of P2XRs may reduce the response to ATP. Koshimizu et al. have reported that the cytoplasmic intersubunit interaction prior to ATP binding in P2X2R contributes to the subsequent channel activity and conformational changes [25]. The strong attachment of P2X4R to mica may also affect the intersubunit interaction via the ICDs, which perhaps causes the reduced response of P2X4R to ATP. Under our conditions, the strong attachment of P2X4R may change the structural flexibility and/or the intersubunit interaction that reduces the responsivity to ATP. The reduced attachment of P2X4Rs to mica without PDL dramatically increased the velocity of the ATP response, and thus supported our hypothesis.
Despite the low ATP reactivity of P2X4Rs on PDL, we observed clear structural differences between the control and the ATP-treated condition. We believe that the three lobes observed after ATP addition were three individual subunits of one P2X4R trimer. First, the distance between the lobes was significantly smaller than that between trimers. If each lobe was an individual P2X4R trimer that was incidentally assembled into a trimer, the distance between lobes would not be significantly different from the distance between trimers. Second, during the AFM observation, some P2X4R trimers occasionally shifted position, and these trimers moved as trimers (i.e., the three lobes did not dissociate). Third, in time-lapse analysis, the circular structure changed into a trimer after ATP treatment both in the averaged particle images and in single particles. This result also suggests that the trimeric stoichiometry exists even in P2X4Rs before ATP binding. From this observation, we considered circular particles without individual subunits before ATP binding to be trimeric P2X4Rs because those subunits were closer together than the spatial resolution of our AFM system. If this is the case, the diameter of the P2X4R in the control should be double that of one lobe. In fact, the diameter of the P2X4R in the control (about 12.6 nm) was approximately double that of one lobe (5.9 × 2 nm). These three lobes were also observed when P2X4Rs were inserted into a lipid bilayer, suggesting that these lobes are the predicted six-stranded antiparallel β-pleated sheet structures in the ECDs of P2X4Rs. EM analysis of P2X4Rs revealed propeller-like domains in the ECDs [8] that were similar to the six-stranded antiparallel β-pleated sheet-like structure that we observed in the ECDs. In their report, the authors clearly demonstrated that the EM-based distance between the C termini of the P2X4Rs was 6.1 nm and the FRET-based distance between the C termini was 5.6 nm. The three propeller-like domains at the opposite end of the P2X4R to the gold-labeled C termini means the distances between these domains would be similar. As described above, when three lobes are assembled close together in the control, the distance between the centers of two lobes is twice the lobe radius (2.95 × 2 nm), which agrees well with the distance between P2X4R C termini estimated by FRET and EM [8]. As mentioned above, the AFM images of P2X4Rs in a lipid bilayer and on mica were comparable; this result strongly suggests the upward direction of the P2X4Rs on mica. However, the height of P2X4R on mica was less than the height of a lipid bilayer composed of phospholipids (about 4 nm) [12]. In nonsymmetrized averaging, one of the three lobes in the P2X4R trimer on mica was lower than the other two. The height of the P2X4Rs on mica was only slightly greater than that from the surface of a lipid bilayer to the top of the inserted P2X4Rs, indicating the possibility that the P2X4Rs do not stand vertically and TMD and/or ICD are bent during the AFM observation. From these observations, we concluded that P2X4Rs lean to one side on mica and TMD or ICD may be bent and concealed behind the ECDs. Similarly, the simple adsorption of P2X2Rs [11] on mica also results in these molecules having a top view-like structure in AFM images, thus supporting our conclusions.
In the time-lapse imaging, we observed two different structural changes: (i) from one circular structure to a trimeric structure (0.0 → 0.5 s after uncaging) and (ii) the subsequent moving away of each lobe (0.5 → 5.0 s). The second structural change reminds us of an important function of the P2X family, namely pore dilation. In an early study, Khakh et al. clearly demonstrated that P2X4R exhibits NMDG+ permeable pore dilation in the absence of extracellular Ca2+ [5]. In their work, the P2X4Rs exhibited sustained activity for several minutes, indicating that our pore dilation-like structure is not a desensitized P2X4R state. Our work represents direct evidence of the functional and structural relationship of pore dilation in P2X4R. Under a 0 Ca2+ condition, we observed both pore dilation-like structural changes in ECDs and EtBr uptake. This pore dilation-like change was reproducible under various conditions including on mica, on PDL-coated mica and in a lipid bilayer, strongly suggesting that this structural change is a fundamental reaction of P2X4R. At 2 mM Ca2+, we observed no EtBr uptake but there was a Ca2+ flow via P2X4R that also corresponded to the previous report [5]. Under this condition, the pore dilation-like structure of P2X4R was not observed but P2X4R trimers similar to the structure seen immediately after ATP binding were evident. The averaged trace of the green fluorescence intensity exhibited a near-plateau state after an initial increase. This result may indicate that the number of desensitized P2X4Rs increase during a long ATP exposure. From these observations, we considered that the structural changes in the ECDs of P2X4Rs are related to permeability dynamics. Recent reports on P2X7Rs suggested the possibility that their EtBr uptake is mediated by accessory Pannexin-1 (Panx1) channels [26]. In their report, the authors demonstrated that human 1321N1 cells express Panx1, so it is possible that there is functional coupling between overexpressed P2X4Rs and Panx1 in this cell. We concluded that EtBr can pass through P2X4R independent of Panx1 (at least in our study) for the following reasons. First, we used purified P2X4Rs and only a single band was observed in the native-PAGE/silver staining. As Panx1 (about 50 kDa) forms a hexameric channel [27], Panx1 contamination would be detected as another band (about 300 kDa). Second, Panx1 and connexins are known to have structural similarities [28] and connexins are observed as hexameric structures [15] in AFM. We observed no hexameric structures in our AFM study. Third, the issue of Panx1 coupling with P2X7R remains to be clearly settled because another group has demonstrated that P2X7R exhibits pore dilation independent of Panx1 [29]. P2X2R also exhibits the pore dilation independent of Panx1 [7]. These results indicate that Panx1 may not be a fundamental component of the pore dilation state of the P2XR family. Fourth, in contrast to connexin hemichannels, Panx1 is active at physiological extracellular Ca2+ concentrations [28]. In our simultaneous Ca2+/dye uptake measurement, EtBr uptake was not observed at 2 mM Ca2+. However, our data and these reports do not rule out the possibility of functional coupling between P2X4Rs and Panx1 in cells.
Thus, our present study provides direct evidence of structural changes in the ECDs of P2X4Rs that are involved in permeability dynamics. We have achieved the first direct, time-lapse imaging, to our knowledge, of ATP-induced structural changes of P2X4R using a new technique, namely fast-scanning AFM. Our approach provides new insights into the structure of P2XRs, and an extension of this approach to other P2X subtypes will help us to understand the structural and functional relationships of the P2XR family.
Reagents were obtained from the following sources. DMEM, EDTA, and FBS were purchased from Gibco. Aprotinin, bestatin hydrochloride, bromophenol blue, geneticin, glycine, leupeptin, NaCl, EGTA, penicillin, pepstatin A, PDL, protein A sepharose, SDS, streptomycin, sucrose, Tris-HCl, Triton X-100, HEPES, 3-[(3-Cholamidopropyl)dimethylammonio]-1-propanesulfonate, 4–2(aminoethyl) benzenesulfonyl fluoride hydrochloride (AEBSF), and L-α-phosphatidylcholine (PC) were obtained from Sigma-Aldrich. Geneticin was supplied by Invitrogen. The silver staining kit and MeOH were purchased from Wako Pure Chemicals. E-64 protease inhibitor was obtained from Calbiochem. Anti-P2X4 receptor antibody was supplied by Alomone Labs. Brain-derived phosphatidylserine (PS) was obtained from Avanti. Native mark (Invitrogen) microdialysis rods were purchased from Hampton Research. The spectrapor dialysis membrane was obtained from Spectrum Lab. n-octyl-β-D-glucopyranoside (βOG) was obtained from DOJINDO.
Human astrocytoma 1321N1 cells were maintained in DMEM, containing 5% (v/v) FBS, 100 μg/ml penicillin, and 100 μg/ml streptomycin (Sigma). For 1321N1 cells expressing P2X4R, 400 μg/ml G-418 (geneticin) was added. cDNA encoding rat P2X4R was subcloned into the pcDNA3.1 vector. Transfection was carried out with Superfect (QIAGEN) according to the manufacturer's protocol. 1321N1 cells successfully expressing P2X4R were confirmed by the ATP-induced increase in [Ca2+]i, and were isolated and proliferated.
P2X4R-expressing 1321N1 cells were cultured to confluence and then harvested by scraping. The cells were homogenized with a Teflon homogenizer in HEPES buffer containing 20 mM HEPES, (pH 7.4), 320 mM sucrose, 5 mM EDTA, 5 mM EGTA, and protease inhibitors (100 μM AEBSF, 80 nM aprotinin, 5 μM bestatin, 1.5 μM E-64 protease inhibitor, 2 μM leupeptin, and 1 μM pepstatin). Supernatants obtained by centrifuging the homogenate at 3,000g for 15 min at 4 °C were further spun at 38,400g for 15 min to obtain membrane pellets. The pellets were resuspended in buffer containing 20 mM HEPES, (pH 7.4), 1% CHAPS, 100 mM NaCl, 5 mM EDTA, 5 mM EGTA, and protease inhibitors. The sample was treated with anti-P2X4R antibody (10 μg) and incubated for 24 h at 4 °C with gentle agitation. Then protein A sepharose (1 mg) was added to the sample and incubated for 1 h at 4 °C. The sample was then centrifuged at 3,000 g for 5 min and the pellets were washed with buffer (20 mM HEPES, [pH 7.4], 100 mM NaCl, 5 mM EDTA, 5 mM EGTA, and protease inhibitors) three times and treated with 50 μl 0.1 M glycine-HCl (pH 2.7) to dissociate the P2X4R protein from the antibody. The supernatant was transferred to a new tube and added to 1/10 volume of 1M Tris-HCl (pH 8.5).
Purified protein was resolved in a native sample buffer (62.5 mM Tris-HCl, [pH 6.8], 15% glycerol, 1% deoxycholate, and 0.01% bromophenol blue) and was loaded onto 4%–13% acrylamide gradient gel. Native Mark was used as a marker for detecting the molecular weight of purified P2X4R. After native-PAGE, silver staining was undertaken following the manufacturer's protocol (Silver stain kit II, Wako). After electrophoresis, the gel was transferred into a container and fixed with a first fixation buffer (10% MeOH, 10% acetic acid and 40% H2O) for 10 min followed by a 10-min second fixation in a second fixation buffer (10% fixation buffer A and 90% H2O). Then the gel was incubated in an intensification buffer (5% intensification buffer, 47.5% MeOH, and 47.5% H2O) for 10 min and washed with H2O for 5 min. The gel was stained in a stain buffer (5% stain solution A, 5% stain solution B, and 90% H2O) for 15 min. After washing with H2O (3 min × three times), the gel was incubated in a developing buffer (5% developing solution and 95% H2O) until the protein bands became visible.
Cells and purified P2X4R protein were lysed with lysis buffer (containing 10 mM Tris, [pH 7.5], 150 mM NaCl, 1 mM EDTA, 1 mM EGTA, 1 % Triton X-100, 0.1% SDS, 1 mM sodium orthovanadate, 1% doxycholate, and 10 μg/ml each of aprotinin, bestatin, pepstatin A, leupeptin). For SDS-PAGE, the lysates were mixed with an equal amount of Laemli sample buffer (62.5 mM Tris/HCl, 20% glycerol, 2.5% SDS, 0.01% bromophenol blue, and 10% 2-merchapt EtOH) and then boiled at 95 °C for 5 min. Proteins were separated in 4%–13% acrylamide gradient gel and then visualized by silver staining. For western blotting analysis, electrophoresed proteins were transferred to the PVDF membrane and P2X4R protein was detected with anti-P2X4R antibody.
The AFM experiments were performed using an NVB500 high-speed AFM (Olympus Corporation). BL-AC7EGS-A2 cantilevers with a spring constant of 0.1 N/m (Olympus Corporation) were used in the tapping mode with an oscillation frequency of 800–1,000 kHz. PDL (0.1 mg/ml in H2O) was treated on mica for 30 min at RT. The sample was washed with imaging buffer A (25 mM Tris-HCl, [pH 8.0], 137 mM NaCl, 2.7 mM KCl), and then deposited on the mica and incubated for 30 min at RT. For the time-lapse imaging of the P2X4Rs, mica rinsed with a high-salt buffer (10 mM Tris-HCl, [pH 8.0], 1M KCl, 30 min at RT) without PDL coating and imaging buffer B (25 mM Tris-HCl, [pH 7.4], 137 mM NaCl, 2.7 mM KCl) were used. To observe P2X4Rs in the presence of Ca2+, imaging buffer C (25 mM Tris-HCl, [pH 7.4], 137 mM NaCl, 2.7 mM KCl, 2 mM CaCl2) was used. To activate the P2X4Rs, caged ATP (200 μM) was uncaged by UV illumination (BH2-RFL-T3, Olympus). ATP and caged ATP were dissolved in imaging buffers A and B, respectively. Images containing 192 × 144 pixels were obtained at a scan rate of 0.2 or 0.5 fps for static images and 2.0 fps for time-lapse imaging.
All AFM images were processed using Image J software (http://rsb.info.nih.gov/ij/). The P2X4R diameters were measured by using “segmented line selections.” The height and diameter were measured by using “Analyze-Plot profile” found on the menu bar. The 3-D images of P2X4R shown in Figure S1 were converted from 2-D AFM images with the Image J plug-in “interactive 3D surface plot” (http://rsbweb.nih.gov/ij/plugins/surface-plot-3d.html). The plug-in programs were downloaded from the Image J software homepage (http://rsb.info.nih.gov/ij/plugins/index.html). The P2X4R images were averaged with EMAN software [30] (http://blake.bcm.tmc.edu/eman/). Because the majority of the P2X4Rs exhibited a similar direction, we simply selected the P2X4R particles at random for averaging. All the P2X4R images used for EMAN processing were converted to TIFF files. The TIFF images were opened by boxer program and particles for averaging were selected. The selected images were processed with an averaging command in proc2d program. The resulting averaged image was saved in PNG file format. For 3-fold symmetrized averaging the P2X4Rs were rotated through three angles (0, 120, and 240°) with illustrator CS software and the file was converted to a TIFF file. The resulting three P2X4Rs were further averaged using EMAN software.
We first established criteria for determining the P2X4R particle center. Three types of P2X4Rs were observed in the control, namely those with triangular, circular, and ellipsoidal structures (Figure S1A). The center of the triangular P2X4R was defined as the center of a triangular circumcircle. The center of the circular P2X4R was defined as the center of an approximated circle. The center of the ellipsoidal P2X4R was defined as the intersection of the long and minor axes. The center of the trimeric P2X4R was defined as the center of a circle connecting the highest points of all subunits. Next, we established P2X4R size criteria. In the present study, we defined the P2X4R dimensions as diameter and height. The diameter of the triangular and circular P2X4Rs was defined as the diameter of the circles used for determining the particle center. The diameter of the ellipsoidal P2X4R was defined as the average value of the long and short axes. The diameter of the trimeric P2X4R was defined as the diameter of a circle that circumscribed the three lobes. The particle height in the control was simply defined as the distance between the top of the particle and the mica surface. The height of the trimeric P2X4R was defined as the average height of three lobes. The height of P2X4R in a lipid bilayer was defined as the total distance from the top of the particle to the membrane surface plus the height of the lipid bilayer (4 nm). Then, we established criteria for particle selection. The P2X4R diameters obtained from 600 particles including activated and nonactivated P2X4Rs exhibited a clear single distribution and the top 5% and bottom 5% of the particles were eliminated from the analysis. The remaining 90% of the particles indicated by the arrows in Figure S1C(i) were used for analysis. In addition to this, P2X4Rs that exhibited a large noise (Figure S1C[ii]) were also eliminated from the analysis. During AFM observation, the P2X4Rs did not always provide clear images. Although the P2X4Rs had a clear topology in some frames, it was not clear in others. When the ATP stimulated P2X4Rs were selected for averaging, P2X4R particles without subunit-like structures were eliminated from the averaging process.
We performed the averaging in accordance with an early study [18]. First, we selected individual P2X4R particles on the basis of our criteria and then averaged them using EMAN software (nonsymmetrized averaging). Under our conditions, most of the P2X4Rs exhibited similar directions, so we did not perform any additional processing before averaging. The resulting images were further rotated (0, 120, and 240°) and averaged again (3-fold symmetrized averaging). When the activated P2X4Rs were averaged, the P2X4Rs without the subunit-like structures observed in the control were eliminated.
Lipid mixtures (100 μl) for reconstitution were prepared from L-α-phosphatidylcholine/brain-derived phosphatidylserine (PC/PS = 1:1, 160 μM) with 160 mM n-octyl-β-D-glucopyranoside. Mixed micelles were added to 100 μl of 100 ng/ml P2X4R protein. Detergent was removed by dialysis using microdialysis rods and a Spectrapor dialysis membrane (molecular cut-off of 50,000) in a dialysis buffer (30 mM HEPES, 5 mM EDTA, 1 mM EGTA, 0.02% of NaN3). The P2X4Rs were dialyzed for 5 d and the buffer was changed every day.
Purified DNA was prepared from primary cultured rat cortex astrocytes using ISOGEN (Nippongene). Primary rat cortex astrocytes were cultured as described in detail in our previous work [31]. DNA isolation was performed in accordance with the manufacturer's protocol. Confluent cultured astrocytes in a 100-mm cell culture dish were washed three times with PBS and lysed with 1 ml of ISOGEN. After homogenization by pipetting, the cell lysate was transferred to a 1.5-ml tube. Then 0.2 ml of chloroform was added to the tube and the resulting mixture was incubated for 3 min at RT after vigorous shaking (15 s). The tube was centrifuged (12,000g) for 15 min at 4 °C and the inter/organic phases were transferred to a new tube. Next, ethanol (0.3 ml) was added to the tube and incubated for 3 min at RT. The tube was centrifuged (2,000g) for 5 min at 4 °C. The supernatant was discarded and 1.0 ml of 0.1 M sodium citrate (in 10% ethanol) was added to the tube. After 30 min incubation at RT, the tube was centrifuged (2,000g) for 5 min at 4 °C. The precipitate was mixed in 2 ml of 75% ethanol and incubated for 30 min at RT. The tube was then centrifuged (2,000g) for 5 min at 4 °C. The precipitate was dried and dissolved in H2O.
Calcium and dye uptake imaging of P2X4Rs was performed using a 500-μm hole cut in a plastic plate consisting of the bottom plate of a 60-mm cell culture dish. A Terumo syringe (25 gauge, 500 μm in diameter) was briefly heated with a gas burner and then pushed through the plastic plate. The resulting plastic burr around the hole was removed with a razor. Then 0.2 μl of imaging buffer containing 50 μM fluo-3 and 100 ng/μl DNA was placed in the hole. The top and bottom of the hole were covered by 1 μl of n-decane containing 2 mM PC/PS (1:1) and incubated for 5 min at RT. P2X4R-containing proteoliposome (0.5 μl) was supplied to the top surface of the hole and incubated for 10 min at RT. The bottom surface of the hole was covered with 2 μl of 10 mM Tris-HCl buffer (pH 7.4). Proteoliposome-containing buffer was carefully washed with 2 μl of 10 mM Tris-HCl buffer (pH 7.4) and then with 10 mM Tris-HCl buffer containing 20 μM EtBr with/without 2 mM CaCl2 (calcium imaging buffer or EtBr uptake imaging buffer). To stimulate the P2X4Rs, 1 μl of ATP (300 μM, the final concentration of ATP in the buffer was 100 μM) was added to the top of the hole. Calcium and dye uptake imaging was performed using a Zeiss LSM510 and ZEN2007 imaging system under a 5× objective. Throughout the functional analysis, fluo-3 was excited with the 488-nm line of an argon ion laser and the emitted light was collected using a 500–530-nm band-pass filter. EtBr was excited at 488 nm and the emission fluorescence was collected using 560–615-nm band-pass filters [32].
Average results are expressed as the mean ± SEM. Data were analyzed with the Student's t-test to determine the differences between groups. Significance was accepted when p < 0.05.
|
10.1371/journal.pgen.1007810 | Dynamic transcriptome profiles within spermatogonial and spermatocyte populations during postnatal testis maturation revealed by single-cell sequencing | Spermatogenesis is the process by which male gametes are formed from a self-renewing population of spermatogonial stem cells (SSCs) residing in the testis. SSCs represent less than 1% of the total testicular cell population in adults, but must achieve a stable balance between self-renewal and differentiation. Once differentiation has occurred, the newly formed and highly proliferative spermatogonia must then enter the meiotic program in which DNA content is doubled, then halved twice to create haploid gametes. While much is known about the critical cellular processes that take place during the specialized cell division that is meiosis, much less is known about how the spermatocytes in the “first-wave” in juveniles compare to those that contribute to long-term, “steady-state” spermatogenesis in adults. Given the strictly-defined developmental process of spermatogenesis, this study explored the transcriptional profiles of developmental cell stages during testis maturation. Using a combination of comprehensive germ cell sampling with high-resolution, single-cell-mRNA-sequencing, we have generated a reference dataset of germ cell gene expression. We show that discrete developmental stages of spermatogenesis possess significant differences in the transcriptional profiles from neonates compared to juveniles and adults. Importantly, these gene expression dynamics are also reflected at the protein level in their respective cell types. We also show differential utilization of many biological pathways with age in both spermatogonia and spermatocytes, demonstrating significantly different underlying gene regulatory programs in these cell types over the course of testis development and spermatogenic waves. This dataset represents the first unbiased sampling of spermatogonia and spermatocytes during testis maturation, at high-resolution, single-cell depth. Not only does this analysis reveal previously unknown transcriptional dynamics of a highly transitional cell population, it has also begun to reveal critical differences in biological pathway utilization in developing spermatogonia and spermatocytes, including response to DNA damage and double-strand breaks.
| Spermatogenesis is the process by which male gametes–mature spermatozoa–are produced in the testis. This process requires exquisite control over many developmental transitions, including the self-renewal of the germline stem cell population, commitment to meiosis, and ultimately, spermiogenesis. While much is known about molecular mechanisms regulating single transitions at single time points in the mouse, much less is understood about how the spermatogenic progenitor cells, spermatogonia, or the meiotic cells, spermatocytes, change during testis maturation. Our single-cell-mRNA-sequencing analysis is the first to profile both spermatogonia and spermatocytes from neonatal mice through adulthood, revealing novel gene expression dynamics and differential utilization of biological pathways. These discoveries help us to understand how the spermatogenic progenitors of this population modulate their activity to adapt to a changing testicular environment. Furthermore, they also begin to explain previously-observed differences—and deficiencies—in spermatocytes that are derived from the “first wave” of spermatogenesis in juvenile mice. Overall, this dataset is the first of its kind to comprehensively profile gene expression dynamics in male germ cell populations during testis development, enriching our understanding of the complex and highly-orchestrated process of spermatogenesis.
| Mammalian spermatogenesis requires proper establishment of the spermatogonial stem cell (SSC) pool, which resides within the seminiferous tubules of the testis and supports life-long germ cell development [1]. These progenitors give rise to all the differentiating germ cells, ranging from spermatogonia to spermatocytes to spermatids, and finally to mature spermatozoa. Despite the essential nature of this process, the genetic regulatory mechanisms underlying the many complex cellular transitions, and the maturation of this system during testis development, have yet to be fully described.
Gamete development in the mouse relies on a rare population of primordial germ cells, the bi-potential progenitors of all gametes, which are specified at embryonic day (E) 6.25 [2]. These cells migrate to and colonize the developing gonad, arriving at the genital ridge from E10.5 [3], and undergo abundant proliferation until E13.5. At this time, germ cells developing in an XX (female) gonad will enter the meiotic program as “oocytes”, while germ cells developing in an XY (male) gonad will become “prospermatogonia”, remaining relatively non-proliferative until shortly after birth [4,5]. Prospermatogonia are then able to adopt several fates [6]: in early postnatal life, a subset of these cells differentiate immediately into spermatogonia and continue to progress through spermatogenesis, to constitute the “first wave” of spermatogenesis. A second subset of prospermatogonia will undergo apoptosis, while the remaining prospermatogonia will become established within the testicular stem cell niche soon after birth, to become the self-renewing SSC population which will support “steady-state” spermatogenesis throughout life. This germline stem cell population makes up less than 1% of the cells of adult testes [7], and must balance self-renewal and differentiation to maintain a healthy male gamete supply. Thus, the first cohort of meiotically-active male germ cells enters the meiotic program without first entering a self-renewal SSC phase, clearly differentiating the first wave of spermatogenesis from the other subsequent waves.
SSCs are triggered to enter spermatogenesis coincident with a burst of retinoic acid (RA), which induces both the spermatogonial divisions and the entry into prophase I of meiosis [8–11]. Thus, in mice, male meiotic entry commences around postnatal day (PND) 10, in response to RA-induced expression of key genes, including ‘Stimulated by Retinoic Acid 8’ (Stra8) [8–11]. Spermatocytes execute many essential meiotic events including creation of double-strand breaks, synapsis of homologous chromosomes, and DNA repair and crossover formation, all of which are critical to proper segregation of homologs in the first meiotic division. Failure to properly execute any of these steps is known to result in potential chromosome mis-segregation, non-disjunction events, aneuploidy, and infertility (reviewed comprehensively in Baarends et al [12] and Gray & Cohen [13]).
While the developmental transitions which underlie germ cell differentiation and maturation have been broadly defined, the gene regulatory underpinnings of these transitions remain largely uncharacterized. Concurrent with our work presented herein, many groups have also investigated developmental transitions within the testis using single-cell sequencing, and have begun to shed some light upon genetic regulatory mechanisms of these processes [14–18]. Intriguingly, several new cell types have been identified, including previously unidentified somatic cells [14], and murine spermatogenesis has been extensively compared to human spermatogenesis [15], emphasizing the translational impact of these types of studies. A caveat of these studies, however, is their focus on single time points, or utilization of cell enrichment protocols that may bias the output. In this manuscript, we have performed the first single-cell sequencing developmental time series of the male mouse germline with comprehensive sampling, thereby capturing all germ cell types through the progression of postnatal testis maturation. The advent of single cell transcriptomics provides an invaluable tool for understanding gene expression dynamics at very high resolution in a large number of individual cells in parallel. Furthermore, single-cell sequencing reveals heterogeneity and potential plasticity within cell populations, which bulk mRNA sequencing is unable to accomplish, making it an ideal tool for profiling germ cell populations which rapidly progress through myriad developmental transitions.
We demonstrate that germ cells display novel gene regulatory signatures during testis development, while cells positive for single protein markers have the capacity to change dramatically with age, and therefore cells of a particular “identity” may differ significantly from postnatal to adult life. Intriguingly, we have also begun to identify differential expression of genes in critical biological pathways which may contribute to observed differences in the first-wave of spermatogenesis [19,20]. Dissecting the complex dynamics of these developmental transitions can provide critical information about the transcriptional landscape of both SSCs, spermatogonia, and spermatocytes, and the regulatory mechanisms that underlie the formation of a dynamic and functional complement of germ cells to support life-long spermatogenesis.
Mouse testes were collected at several postnatal time points, selected to represent distinct stages of germline development: postnatal day (PND) 6 (during SSC specification), PND14 (first appearance of pachytene spermatocytes during the first wave), PND18 (pachytene and diplotene spermatocytes from the first wave present), PND25 (spermatids present) and PND30 and adult (spermatozoa present) (Fig 1A) and subjected to single-cell RNAseq. The tissue was dissociated, and the resulting slurry subjected to 30% Percoll sedimentation to remove debris. The PND18 cell suspension was split and processed both with and without Percoll sedimentation as a technical control; due to similarities between libraries, the data from these libraries was thereafter combined (S1 Fig). Additionally, due to the proportionally high representation of sperm in the adult testis, it was necessary to increase representation of other germ cell types from these samples. To accomplish this goal, an adult testis suspension post-Percoll sedimentation was split in half and either positively magnetically-cell-sorted (MACS) for the cell surface marker THY1, in an attempt to enrich for spermatogonia [21], or negatively MACS-sorted for ACRV1, in an attempt to deplete testicular sperm [22]. While neither strategy can accomplish complete enrichment of spermatogonia or removal of spermatozoa, respectively, both adult libraries had a representative sample of all germ cell types (Fig 1B), and are therefore treated as adult replicates in these data. For each single-cell testis suspension, 4–5,000 cells per mouse were processed through the 10X Genomics Chromium System using standard protocols for single cell RNA sequencing. Libraries were sequenced to an average depth of 98M reads; on average, 91% of reads mapped to the reference genome. After standard data processing, we obtained gene expression profiles for approximately 1,100–2,500 cells per library (Figs 1B and 2A, S1 Table) with representation of between 2,500–5,000 genes per cell, and an average of nearly 32,000 mapped reads/cell post-normalization (also known as “UMI”; unique molecular identifier counts) (S2 Fig), all of which indicate the robustness of the sequencing method and comparability to UMI counts in other similar studies [14,15,17].
Primary cell clusters revealed by Seurat (S3 Fig) were identified and merged into superclusters (Fig 2B) based on characteristic marker gene expression (S2 Table). While somatic cells were evident in the clustering analysis, including Sertoli cells, smooth muscle, and epithelial and hematopoetic cells, they represent a minority (fewer than 25%) of the total cells profiled (S1 Table). In particular, Sertoli cells were primarily derived from the PND6 library (Fig 2) and exhibit expression of known immature-, mature-, and pan-Sertoli cell markers [23–28] (S4A Fig). This biased recovery is likely due to their increased representation at the neonatal time point, as well as processing steps which effectively removed these cells, which become considerably larger in older mice. This reduced retention of somatic cells in the processing of testes from older mice is consistent with other single-cell sequencing studies of the testis, and can only be mitigated by different dissociation or filtering methods and/or cell sorting [14]. Therefore, somatic cells were excluded from gene expression analysis during testis maturation, which focused on the spermatogonial and spermatocyte populations. Analysis of differentially expressed genes between cell clusters identified characteristic marker genes in each cell type, including Zbtb16 (Plzf), Sall4, Sohlh1, and Dmrt1 in spermatogonia; Meioc, Prdm3, Top2a, and Smc3 in early spermatocytes; Sycp1/2/3 and H2afx in spermatocytes; Acrv1, Izumo1, and Catsper3/4 in round spermatids; and Prm3, Izumo3, and Tssk6 in elongated spermatids (Figs 3, S5 & S6, S2 Table).
Critically, the germ cell type classifications are representative of the known timeline of the developing testis (Fig 1B), with only spermatogonia present at PND6, some early spermatocytes present at PND14, much greater representation of those spermatocytes at PND18, and appearance of more differentiated round and elongated spermatids from PND25 onwards. Interestingly, we observed the greatest enrichment of spermatids in the positively THY1-sorted adult sample, likely due to non-specific binding of the antibody to the developing acrosome. Despite this, the library contained strong representation of spermatogonia and spermatocytes and was therefore retained in the analysis. The negative ACRV1 sorting for the other adult sample retained representation of all germ cell types in the adult testis, including spermatogonia, which would otherwise have been poorly represented due to the much greater abundance of more differentiated cell types. Overall, both adult samples provide excellent representation for all germ cell types present in the adult testis and are therefore included in this sampling analysis.
Importantly, we also observe the expected down-regulation of X-linked genes during meiosis I. Numerous studies have demonstrated that Prophase I progression is associated with only partial synapsis of the X and Y chromosomes at the pseudo-autosomal region [29–31]. As a result, the sex chromosomes undergo progressive silencing throughout the unsynapsed chromatin, with complete silencing by pachynema [31–34]. As a positive validation of appropriate spermatocyte classification, a random sampling of X-linked genes were analyzed for their expression across cell types (S4B Fig). Despite several patterns of X-linked gene expression among cell clusters, the spermatocyte supercluster exhibits very little to no detection of all profiled X-linked genes, again confirming the robustness of our cell-type classification methods.
Cell-free RNA contamination from lysed cells is a well-known confounding feature in single-cell sequencing libraries, as highly-expressed transcripts from even a small number of lysed cells can become incorporated in the gel bead emulsions of single-cell microfluidics devices [35]. As a result of the incorporation of these transcripts into libraries of cells from which they did not originate, cells which do not endogenously express such transcripts can appear to have low levels of expression of these markers, potentially confounding data analysis due to technical artefacts. In this data set, genes highly expressed in elongated spermatids/sperm were detected at low levels in other cell types, including somatic cells, nearly exclusively in libraries derived from testes aged 25 days or older (S7 and S8 Figs), the only samples in which spermatids are present. Therefore, we believe the detection of these transcripts in both spermatogonia and spermatocytes of older mice is due to contamination from cell-free RNA derived from lysed spermatids. To mitigate the age-related biases this signal might pose in down-stream analysis, markers of the spermatid/sperm population (genes with a greater than 20:1 ratio of expression between spermatids and other germ cell types), such as Prm1/2, were removed from further analysis (S3 Table).
To better understand the developmental transitions that spermatogonia undergo as mice age, genes variably-expressed with age were identified by Model-based Analysis of Single Cell Transcriptomics (MAST) [36] (S4 Table) and visualized as a heatmap (Fig 4). As spermatogonia become proportionally rarer with age (and therefore later-aged individual libraries experience lower representation), spermatogonia from libraries PND18, PND25, and PND30 were merged with each other, as were spermatogonia from the two adult samples. While marker genes, indicated in the top row, remain quite constant over time, clear and novel differences can be observed in spermatogonial gene expression over time, particularly in spermatogonia derived from PND6 testes. Several genes, including those noted to the side of the heatmap, are observed to have robust differential expression over time, and may play a role in the establishment and growth of this spermatogonial pool which will support lifelong spermatogenesis.
In addition to MAST analysis for variable expression, Gene Set Enrichment Analysis (GSEA) Time Series analysis [37] of Reactome pathways revealed differentially-utilized pathways, which were then visualized using Enrichment Map in Cytoscape [38,39]. GSEA of variably-expressed genes in spermatogonia from mice of different ages reveals significant changes in many pathways, including increasing expression of genes related to RNA destabilization and protein degradation as well as WNT signaling, and decreasing expression of genes related to asparagine metabolism, various signaling pathways including TGFB, FGFR, and KIT, and transcriptional regulation (Figs 5 & S9A, S5 Table). In particular, many critical signaling receptors and ligands, including Kit and KitL [40–43], as well as Fgf8 and Fgfr1 [44,45], exhibit downregulation in spermatogonia derived from mice of increasing age, consistent with potentially altered paracrine signaling around the basement membrane of the seminiferous tubules during testis maturation.
It has been well established that meiotic regulation is distinct in the first wave of meiosis compared to that in subsequent waves [6,19,20]. Thus, we sought to characterize this phenomenon in terms of the transcriptome profile of spermatocytes at discrete developmental time points. Spermatocytes from the first meiotic wave compared to steady-state (adult) ages were also subjected to MAST analysis, as described above (Fig 6, S6 Table). For this analysis, spermatocytes were abundant enough from all libraries that each time point could be considered separately, except for PND6 in which spermatocytes are not yet present. Notably, spermatocytes from PND14, which are only just beginning Prophase I, demonstrate very distinct gene expression patterns from spermatocytes at later time points and are not representative of the full spectrum of meiotic cell types, as expected. Some genes, including those noted to the side of the heatmap, show robust differential expression with age, highlighting differences between spermatocytes derived from the first-wave (PND18) in contrast those which are derived from a self-renewing SSC population (adult). Therefore, these genes were chosen for further analysis and orthogonal validation.
GSEA time series analysis of Reactome pathway enrichment of variably expressed genes in spermatocytes also reveals intriguing differentially utilized pathways. From this analysis, we observe decreasing expression of genes related to translation and post-transcriptional regulation, and increasing expression of genes related to DNA replication, double strand break repair, and cell cycle regulation (Figs 7 & S9B, S7 Table). Most notable in the list of genes upregulated in spermatocytes of increasing age are those known to be essential to DNA repair, meiotic progression, and crossover formation including Brip1 [46], Brca1 and Brca2 [47–49], Rad51 [50], H2afx [51] and Atm [52]. Many of these pathways, particularly those related to double strand break repair (which initiates meiotic recombination), may be crucial for understanding the molecular mechanisms underlying fundamental differences in first-wave spermatocytes.
Candidate genes of interest (GOIs) identified in the single-cell sequencing data were investigated using immunofluorescence, which allowed us to validate changes in protein expression in the context of the native testis tissue. We used paraffin-embedded testis tissue sections at postnatal ages PND7, PND13, PND22, and 8 weeks of age to characterize the same range of postnatal testis development as was captured in the single-cell RNAseq data set. In an effort to reduce batch effects emerging from testis isolation on different days, and in light of the subtlety of some of the differential protein expression, we prioritized simultaneous processing over exact age-matching.
GOIs were selected by meeting several criteria including: representation across several biological pathways, significant differential expression in a given cell type between mice of different ages, and availability of a commercial immunofluorescence-verified and mouse-reactive antibody. For all validation of spermatogonial candidates, double-staining was performed with an antibody against ‘Promyelocytic Leukemia Zinc Finger Protein’ (PLZF; aka ZBTB16), a well-characterized marker of undifferentiated spermatogonia [53–55]. For all validation of spermatocyte candidates, double-staining was performed with an antibody against ‘Synaptonemal Complex Protein 3’ (SYCP3), to allow for visualization and staging of Prophase I-staged cells [13,56]. SYCP3 marks the nuclei of spermatocytes through leptotene, zygotene, pachytene, and diplotene stages of prophase I.
To profile a range of biological functions including metabolism, enzyme ‘Asparaginase-Like 1’ (ASRGL1; aka ALP1) was chosen for immunofluorescence analysis. ASRGL1 is known to catalyze the hydrolysis of L-asparagine [57] and to clear protein-damaging isoaspartyl-peptides [58], and while largely uncharacterized in the mouse, has been found to be highly expressed in the human cervix, fallopian tube, ovary, and testis [59]. Interestingly, ASRGL1 has been identified as a biomarker of endometrial cancer [59–62], as well as an antigen in rodent sperm [63]. In the single-cell sequencing dataset, Asrgl1 was observed to be highly expressed in spermatogonia from PND6 mice, with decreasing expression in this cell type in older mice (Fig 4, S4 Table). Interestingly, Asrgl1 was also shown to have dynamic expression in spermatocytes, with the lowest expression detected in PND14 spermatocytes and increasing expression in spermatocytes of older mice (Fig 6, S6 Table). These results at the mRNA level were corroborated by immunofluorescence data showing high expression of ASRGL1 protein in PND7 PLZF+ spermatogonia, with decreasing expression of the protein in PND22 and adult PLZF+ spermatogonia (Fig 8). Furthermore, PND13 first-wave pachytene spermatocytes expressed little ASRGL1 protein, with expression becoming abundant in pachytene spermatocytes from PND22 and adult testes (Fig 9).
Significant differences in RNA stability and processing genes were also observed in spermatogonia during postnatal testis development, with down-regulation of related pathways over time. The RNA binding protein ‘RNA Binding Motif Protein, X-linked-like 2’ (RBMXL2; aka HNRNPG-T) is a putative RNA regulator and splicing factor highly expressed in the mouse testis, specifically in germ cells [64], with critical functions in spermatogenesis [65]. Furthermore, disruptions in RBMXL2 expression and localization in human testes are associated with azoospermia in men [66]. In the single-cell data set, Rbmx2 mRNA was observed to be highly expressed in spermatogonia from the youngest mice, then decreasing with age (Fig 4, S4 Table), with expression persisting in spermatocytes at all ages. Immunofluorescence of RBMXL2 protein demonstrated high expression of the protein in all germ cells, including spermatogonia and spermatocytes. Close inspection, however, revealed relatively higher expression of RBMXL2 in PND7 PLZF+ spermatogonia compared to later time points, despite the relatively consistent expression of RBMXL2 protein in all other germ cell stages at all mouse ages (S10 Fig).
‘Double-sex and Mab-3 Related Transcription Factor B1’ (DMRTB1; aka DMRT6) is a transcriptional regulator known to coordinate the developmental transition from spermatogonial differentiation to meiotic entry [67]. As has been previously observed, Dmrtb1 mRNA was highly expressed in spermatogonia and early spermatocytes (Figs 4 & 6), which we confirmed at the protein level by immunofluorescence. At PND13, first-wave early leptotene spermatocytes, evidenced by spotty SYCP3, exhibited nuclear DMRTB1 staining, while pachytene spermatocytes from all mouse ages lost DMRTB1 expression. Interestingly, the nuclear staining in leptotene spermatocytes was only observed at the earliest time point, PND13, and not seen in early Prophase I spermatocytes of later spermatogenic waves (Fig 10).
Finally, DNA damage response proteins ‘RAD51 Recombinase’ (RAD51) and ‘Ataxia Telangiectasia Mutated’ (ATM) were profiled across spermatocytes from mice of increasing age, as these represent particularly interesting candidate proteins whose differential expression may be crucial to understanding aberrant recombination rates and chromosome segregation in the first wave. Intriguingly, both RAD51 and ATM showed subtly, but decidedly, decreased overall protein expression in first-wave pachytene spermatocytes, with increasing protein expression as mice age (Figs 11 and 12), as predicted from our scRNAseq data. Previous research has demonstrated significantly fewer RAD51 foci along chromosome cores of juvenile C57bl/6 spermatocytes compared to those at 12 weeks of age [19]. To test if this phenomenon is also observed in B6D2F1/J mice, RAD51 foci were quantified on zygotene chromosome cores in PND14, PND21, and adult spermatocytes. Unlike the differential foci counts observed in C57bl/6 mice, B6D2F1/J mice do not exhibit significant differences in RAD51 foci as a function of age (Fig 13A). Our RAD51 observations, therefore, demonstrate overall decreased protein expression in the nucleus of pachytene first-wave spermatocytes, with increasing expression by 3 weeks of age, but no alteration in localized RAD51 during zygonema (Fig 11). A similar expression dynamic was observed for ATM, which has robust cytoplasmic staining as well as diffuse nuclear staining in pachytene spermatocytes [68]. Our data demonstrate decreased expression of ATM in both cellular compartments in the first-wave spermatocytes, with increasing expression, particularly in the cytoplasm of these cells, by 3 weeks of age (Fig 12).
To further dissect DNA damage response in first-wave spermatocytes, γH2AX staining was performed on chromosome spreads from PND14, PND21 and adult spermatocytes (Fig 13B and 13C). Canonically, the γH2AX histone variant marks meiotic double-strand breaks (DSBs) [69] in leptonema and zygonema, becoming restricted to the unsynapsed “sex body”, or XY chromosomes, during pachynema [33]. Presence of γH2AX on pachytene autosomes indicates aberrant unrepaired DSBs. Quantification of γH2AX demonstrated a significantly greater average percentage of cells with persistent flares on pachytene autosomes compared to adult spermatocytes (Fig 13B). These age-dependent dynamics of critical DNA damage response regulators are likely to contribute to the health and viability of resulting spermatocytes and spermatozoa from these spermatogenic cycles, and may underlie some of the functional differences observed in the first wave of spermatogenesis [6,19,20].
Overall, these gene expression dynamics discovered from single-cell mRNA sequencing are reproducible at the level of protein expression in the context of the native tissue, and likely represent important transitions both in spermatogonia and spermatocytes during testis maturation. These data will be indispensable to investigate how gene expression dynamics help to regulate the many critical developmental events, including spermatogonial differentiation and meiotic progression, occurring in the developing mouse testis.
We have performed the first comprehensive sampling and screening of mouse germ cells from neonatal life through adulthood, to characterize transcriptional profiles at single-cell resolution. With the exception of the adult samples, which were sorted to minimize the over-riding sperm component, all libraries were generated from unsorted single-cell suspensions containing all testicular cells. Adult samples were minimally processed with a single-step magnetic cell sort to provide representation of all germ cell types in the adult testis. Previously, single-cell-sequencing studies on sorted cells have provided valuable information about specific, and marker-defined cell types [15,70]. The study presented here, however, focuses on profiling the germline during postnatal testis maturation, importantly capturing the first-wave of spermatogonia and spermatocytes which exhibit differences from later, steady-state spermatogenesis. Because of the progression of ages profiled, we have captured changes in gene expression at single-cell resolution to compare the developmental progression of spermatogenesis as mice age.
Germ cells subtypes in the testis are frequently defined by the presence or absence of particular protein markers, which can be visualized by reporter expression or immunofluorescence, or may be used for flow cytometry or other enrichment paradigms. Spermatogonia are often defined as cells which express a key complement of protein markers, such as PLZF or ‘GDNF family receptor alpha 1’ (GFRα1) [71–73]. While this is the best practice for visual identification of cells for which discrete markers have been elusive, such as spermatogonia, our analysis suggests that the biology of these cells during postnatal testis development is far more complicated than previously understood. Our analysis stresses that defining these cell populations on the basis of specific markers may be overly simplistic, despite being the current standard practice in the field. Primary spermatocytes similarly show age-specific patterns at the transcriptional level. Furthermore, cells possessing an SYCP3-positive synaptonemal complex indicative of pachynema also exhibit differences in immunolocalized proteins during testis maturation, indicating that they, too, exhibit distinct and variable expression patterns with increasing age. Critically, this analysis reveals that, while known markers may be useful for defining primary cell identity, there are many changes in spermatogenesis during mouse development that have been under-appreciated without the power of single-cell resolution of gene expression profiling.
Specifically, we show here that PLZF-defined spermatogonia, though retaining PLZF-positivity, are transcriptionally distinct at PND6 compared to later developmental time points. These transcriptional dynamics are also reflected by distinct differences at the protein level, with proteins such ASRGL1 being localized strongly in PLZF+ spermatogonia during the first weeks of life, but decreasing in expression in these cells around three weeks of age (Fig 8). Similarly, we show that Prophase I spermatocytes possess significant transcriptional differences in the first-wave compared to subsequent spermatogenic waves, with hundreds of differentially expressed genes across multiple regulatory pathways. Furthermore, direct inspection of pachytene spermatocytes from the first wave at PND13 to later spermatogenic waves reveals that, like spermatogonia, transcript dynamics are also reflected at the protein level. For example, proteins such as DMRTB1 are found only in the nuclei of first-wave leptotene spermatocytes as they transition from differentiated spermatogonia into the meiotic program, but not in spermatocytes from older mice (Fig 10). By contrast, other proteins such as ASRGL1 are in low abundance in early first-wave spermatocytes, but become more strongly immunolocalized in spermatocytes at increasing ages (Fig 9).
In addition to the transcriptional and translational dynamics in defined cell types, these data also reveal differential utilization of particular biological pathways over developmental time. Gene set enrichment analysis [37] utilizing the Reactome pathway database [38,39] has demonstrated that spermatogonia significantly change their transcriptional landscape as mice age, including downregulation of genes within essential meiotic-entry-associated SCF/KIT [43,74] and FGFR [75] pathways, including Kit and KitL (S9A Fig). While we cannot rule out the possibility that variable gene expression in spermatogonia is, in part, due to differential contributions of the spermatogonial stem cell population at different ages–with decreasing contribution as mice age–this dataset provides strong support for true variable gene expression in the spermatogonial pool. For instance, while spermatogonia derived from older mice exhibit downregulation of genes associated with FGFR signaling, including Fgf8 and Fgfr1, and could indicate decreased representation of an undifferentiated spermatogonial population [45], this is in opposition to observed coincident decreased expression of Kit and KitL which would support increased representation of an undifferentiated population [41,43,74] (S9A Fig). It is interesting to consider that these data may reflect overall changes to the paracrine signaling of the spermatogonial stem cell niche as well as the basement membrane as mice age. Overall, these data suggest spermatogonia may modulate their sensitivity for particular critical signaling pathways, which may affect their competency to commit to the meiotic program. Furthermore, pathways associated with mRNA stability and protein degradation are upregulated as the testis matures, suggesting that spermatogonia from older mice may change their capacity for post-transcriptional and post-translational regulation, possibly reflecting changing demands for growth and proliferation in older animals.
Similar to spermatogonia, spermatocytes also exhibit differential utilization of specific biological pathways with age, an observation that dovetails with the knowledge that spermatocytes derived from the first wave of spermatogenesis are functionally different to those spermatocytes derived from steady-state spermatogenesis. First-wave spermatocytes are known to exhibit several unique, and some detrimental, characteristics, including reduced recombination rate [19,20] and greater incidence of chromosome mis-segregation [20]. These features may help to explain some of reproductive deficits observed in spermatozoa derived from young fathers compared to older fathers [76,77]. Our data presented here demonstrate age-related upregulation of pathways associated with DNA replication and repair, double strand break repair, and response to DNA damage (S9B Fig), all of which may underlie the well-characterized differences between spermatocytes in the first wave compared to steady-state spermatocytes. Included in these sets of variably-expressed genes are known regulators of DNA damage response and cross-over formation including Rad51 [50], Brip1 [46], and H2afx [51], as well as Brca1 and Brca2 [47–49] and Atm [52,78], all of which increase in spermatocytes from older mice, effects which we have also shown at the protein level for both RAD51 and ATM (Figs 11 and 12). While the cause of this lower expression in the first wave is unknown, we have also shown that it is coincident with persistent γH2AX on autosomes during pachynema (Fig 13), pointing to systemic alterations in the processing of DNA damage during the initial waves of spermatogenesis. This interpretation is further supported by work from Lange et al, in which the authors find that reduced ATM expression results in increased DSBs [78]. These observations may suggest that steady-state spermatocytes acquire greater competency to cope with the DNA damage inherent to meiotic progression, and that spermatocytes in the first wave may not execute these pathways as successfully, resulting in the observed recombination differences and increased chromosome mis-segregation. Moreover, these early meiotic waves in boys are thought to be highly error-prone and thus these data may provide an explanation for the increased aneuploidy observed in the progeny of young fathers [76,77]. Notably, like spermatogonia, spermatocytes also experience alteration of pathways related to translation and mRNA stability, emphasizing the myriad ways in which gene expression is regulated in developing germ cells. Ultimately, this differential pathway utilization may help to explain not only the functional differences observed in spermatocytes and spermatozoa from juveniles, but may also improve our understanding of increased birth defects associated with young paternal age [76,77].
An important consideration in these studies is the changing architecture during the maturation of the murine testis, which spans the time points we have chosen for study. Between two to three weeks of age in the rodents [79], the seminiferous epithelium becomes segregated into basal and adluminal compartments due to the formation of the blood-testis barrier (BTB), involving the establishment of a variety of cell-cell junctions between Sertoli cells (reviewed comprehensively in Mruk & Chang [80]). The BTB allows for the physical separation between spermatogonia and pre-leptotene spermatocytes in the basal compartment from all more differentiated spermatocytes and spermatogenic cell types in the adluminal compartment. Functionally, the BTB allows for meiotic cells to be maintained and matured in an immune-privileged environment, disconnected from the circulatory system of the animal [81]. As a result of this maturation of the seminiferous epithelium and the cyclicity of spermatogenesis, germ cell groups can synchronously progress through spermatogenesis longitudinally through the tubule, resulting in the many germ cell stages of the seminiferous epithelial cycle [82]. Due to the formation of these structures taking place in the time frame between our PND6/14/18 libraries and our PND25/30/adult libraries, we cannot rule out the possibility that some of the differences we observe in first-wave spermatocytes are due to their maturation in a seminiferous tubule environment that has not yet established this architecture, or due to their different stages in the epithelial cycle. There is likely complex interplay between the maturation of the environment within the tubule, the cell-cell signaling taking place, and the resulting changes in biological pathway utilization in the first-wave spermatocytes compared to those in steady-state spermatogenesis.
An additional primary objective in undertaking this analysis was to potentially reveal new markers of the spermatogonial stem cell population. Despite many efforts to define this population by both cytoplasmic and nuclear markers, discrete markers of this population have remained elusive and controversial [83–86]. Our data are supportive of the high degree of heterogeneity of this cell population, not only within the population at a single age, but also across ages. Furthermore, candidate cell-type specific markers which have become accepted in the field, such as expression of ‘Inhibitor of DNA Binding 4’ (Id4) [87–89], show widespread detection in spermatogonia through spermatocytes, while markers such as Zbtb16 (Plzf) and Gfrα1 have much more restricted expression (S6 Fig). Importantly, expression patterns of these popular markers are not entirely self-consistent. It is therefore likely that the spermatogonial population is not discrete, but is indeed a continuous or plastic population [90,91]. These data represent a valuable resource to the study of the molecular mechanisms underlying SSC self-renewal and differentiation, though the biology may not be as simplistic as originally thought.
Taken together, these data represent the first comprehensive sampling and profiling of spermatogonia and spermatocytes during development of the mouse testis. These data emphasize the necessity of considering not only the protein markers for which individual cells are positive, but also the age of the cells being analyzed. These observations of dynamic gene expression in germ cell populations during postnatal testis development stress that germ cells of a particular age or identity possess distinct profiles and that consideration must be given to these dynamics when profiling germ cell populations. These data also represent an invaluable community resource for discovery of previously unknown gene expression dynamics and pathway contributions that may be critical for the many developmental transitions in the male germ cell population which are essential for successful spermatogenesis and fertility.
Work involving laboratory rodents was performed under an approved protocol, 2004–0063, from the Cornell Institutional Animal Care and Use Committee. Mice were maintained on standard conditions of light:dark cycling and temperature, with ad libitum access to rodent chow and water.
B6D2F1/J mice were generated by mating C57bl/6 female mice with DBA/2J male mice. All animal protocols were reviewed and approved by the Cornell University Institutional Animal Care and Use Committee and were performed in accordance with the National Institutes of Health Guide for the Care and Use of Laboratory Animals. Mice were maintained on standard light:dark cycles with laboratory mouse chow provided ad libidum.
Testes were collected from mice (n = 1 mouse, 2 testes for each time point) at postnatal (PND) days 6, 14, 18, 25, 30, and 8 weeks of age, and dissociated per standard protocols for germ cell enrichment [21]. Briefly, testes were held in 1X HBSS buffer before de-tunicating and moving tubules into 0.25% Trypsin. Tubules were further dissociated by trituration and addition of DNase to a final concentration of 7 mg/ml. Tubules were placed in a 37°C incubation for 5 minutes at a time, and then removed for further trituration. Incubations at 37°C were performed three times, until a cloudy suspension was achieved. Cells were passed through a 40 μM filter, spun down, and re-suspended in 10ml 1X Dulbecco’s PBS + 10% Knockout Serum Replacement (DPBS-S). This cell suspension was then layered on top of a 30% Percoll solution. Cells were then spun down again, and the resulting pellet was re-suspended in 1ml DPBS-S. As a technical control, cells from PND18 were split into two samples after the 40 μM filter, with one half of the cells processed with the Percoll gradient, and the other half directly re-suspended in its final buffer with no Percoll sedimentation, resulting in libraries “PND18” and “PND18pre”, respectively. Due to the similarities between these libraries (S1 Fig), the data from these libraries were thereafter combined and analyzed together as “PND18”.
For adult testes only, the resulting cell suspension was split in half and sorted with magnetic beads in two ways: (1) sperm-depletion was performed by incubating the cells for 30 minutes with 20 μl anti-ACRV1-PE (Novus Biologicals #NB500-455PE), washing with DPBS-S, incubating the cells for 30 minutes with 20 μl magnetic-bead-conjugated anti-PE (Miltenyi Biotec #130-048-801), and finally performing a negative magnetic selection. Cells were applied to a Miltenyi Biotec MACS LS column, and flow-through cells were collected, as sperm were to remain bound to the ferromagnetic column. (2) THY1+ spermatogonia were enriched by incubating the cells for 60 minutes with 20 μl magnetic-bead-conjugated anti-CD90.2 (THY1) (Miltenyi Biotec #130-102-464), and finally performing a positive magnetic selection. Cells were applied to the column, flow-through cells were discarded, and antibody-bound cells were eluted from the ferromagnetic column. These cells were then spun down and re-suspended in 1ml DPBS-S as above.
The resulting cells from all samples were submitted to the Cornell DNA Sequencing Core Facility for processing on the 10X Genomics Chromium System with a target of 4–5000 cells per sample. Sequencing libraries were generated using the 10X Genomics Chromium Single Cell 3′ RNAseq v2 kit, tested for quality control on an ABI DNA Fragment Analyzer, and run on a NextSeq platform with 150 base-pair reads. Libraries were sequenced to an average depth of 98M reads (range 77M-124M); on average, 91% of reads (range 89%-92%) mapped to the reference genome.
Data normalization, unsupervised cell clustering, and differential expression were carried out using the Seurat R package [92]. Batch effect and cell-cycle effect were removed by Combat [93] and Seurat together. Cells with less than 500 genes or 2000 UMIs or more than 15% of mitochondria genes were excluded from the analysis. Gene expression raw counts were normalized following a global-scaling normalization method with a scale factor of 10,000 and a log2 transformation, using the Seurat NormalizeData function. The top 4,000 highly variable genes were selected using the expression and dispersion (variance/mean) of genes. Combat removed batch effects. Seurat regressed the difference between the G2M and S phase, then followed by principal component analysis (PCA). The most significant principal components (1–30) were used for unsupervised clustering and t-Distributed Stochastic Neighbor Embedding (tSNE) analysis.
Cell types were manually identified by characteristic marker genes [18,94,95], and confirmed by SingleR (Single-cell Recognition) package. Differential expression analysis was performed based on the MAST (Model-based Analysis of Single Cell Transcriptomics) [36]. Gene Set Enrichment Time Series Analysis [37] used the differential expression based on each time point, after removing genes highly expressed in spermatids (S7 and S8 Figs). Pathways were visualized by EnrichmentMap [39] in Cytoscape [38].
Code availability: The scripts used for analysis and figure generation are available at https://github.com/nyuhuyang/scRNAseq-SSCs
Data availability: The single-cell RNAseq data have been deposited at GEO and are accessible through Series accession number GSE121904.
Testes were collected at PND7, PND13, PND22, and 8 weeks of age, cleaned of excess fat, and fixed in 0.1% formalin solution overnight before dehydration and embedding in paraffin. Note that for these experiments, while we used the same time frames for scSEQ samples representing the range of testis maturation from SSC specification through the first wave and into steady-state spermatogenesis, the precise days of collection were dependent on the simultaneous harvesting and processing of samples from all age groups to reduce batch effects and allow well-controlled comparisons among ages.
Fixed testes were sectioned at 5 μm onto glass slides by the Cornell Animal Health Diagnostic Center. To stain, sections were de-paraffinized by 3x, 5 minute washes in Histoclear followed by rehydration in 100% ethanol (2x, 5 minutes), 95% ethanol (2x, 5 minutes), 70% ethanol (1x, 5 minutes), water (1x, 5 minutes). Sections were then incubated in boiling antigen retrieval buffer (10 mM sodium citrate, 0.05% Tween-20, pH 6.0) for 20 minutes and left to cool. Sections were washed 3x, 5 minutes in 1X PBS + 0.1% Triton-X (PBST). Tissue sections were then incubated in blocking buffer [3% Goat Serum (Sigma), 1% Bovine Serum Albumin (Sigma), and 0.5% Triton-X (Fisher Scientific) in 1X PBS] and stained by incubation with primary antibodies against PLZF, SYCP3, RBMXL2, ASRGL1, DMRTB1, RAD51, ATM, and γH2AX (see S8 Table) overnight at 4°C. The following day, slides were washed 3x, 5 minutes in PBST and then incubated with secondary antibodies raised in goat against mouse (594 nm) and rabbit (488 nm) at 1:500 for 1 hour at 37°C. A secondary antibody-only control was included to assess background staining. Sections were further stained with DAPI to visualize nuclei, mounted and analyzed on an Epifluorescent Zeiss Axioplan microscope. For a given set of antibodies, images were exposed equivalently for all samples from different time points to generate images for relative comparison of intensity over time.
Spreading of meiotic chromosomes was performed using the “drying down method” as previously described [96]. Briefly, testes were simultaneously collected from mice aged 14 days, 21 days, or 8 weeks, de-tunicated, and incubated in hypotonic extraction buffer for one hour on ice. Tubules were then minced in a bubble of 0.03% sucrose and the cell suspension applied to 1% paraformaldehyde wetted slides. Cells were allowed to “spread” for two hours in humidity, followed by drying. Slides were subsequently stained with primary antibodies against either RAD51 or γH2AX in conjunction with primary antibody against SYCP3 (see S8 Table). Secondary antibodies used are the same as those used for testis section immunofluorescence. For analysis of RAD51 foci, SYCP3 was used to identify zygotene spermatocytes. Two animals per age and at least 30 cells per mouse were analyzed for number of RAD51 foci on the chromosomes cores per cell. Significance was determined by Kruskal-Wallis test. For analysis of persistent γH2AX, SYCP3 was used to identify pachytene spermatocytes. Two animals per age and at least 30 cells per mouse were analyzed for presence or absence of γH2AX on the autosomes. Significance was determined by a one-way Anova.
|
10.1371/journal.pcbi.1002982 | Sustained Firing of Model Central Auditory Neurons Yields a Discriminative Spectro-temporal Representation for Natural Sounds | The processing characteristics of neurons in the central auditory system are directly shaped by and reflect the statistics of natural acoustic environments, but the principles that govern the relationship between natural sound ensembles and observed responses in neurophysiological studies remain unclear. In particular, accumulating evidence suggests the presence of a code based on sustained neural firing rates, where central auditory neurons exhibit strong, persistent responses to their preferred stimuli. Such a strategy can indicate the presence of ongoing sounds, is involved in parsing complex auditory scenes, and may play a role in matching neural dynamics to varying time scales in acoustic signals. In this paper, we describe a computational framework for exploring the influence of a code based on sustained firing rates on the shape of the spectro-temporal receptive field (STRF), a linear kernel that maps a spectro-temporal acoustic stimulus to the instantaneous firing rate of a central auditory neuron. We demonstrate the emergence of richly structured STRFs that capture the structure of natural sounds over a wide range of timescales, and show how the emergent ensembles resemble those commonly reported in physiological studies. Furthermore, we compare ensembles that optimize a sustained firing code with one that optimizes a sparse code, another widely considered coding strategy, and suggest how the resulting population responses are not mutually exclusive. Finally, we demonstrate how the emergent ensembles contour the high-energy spectro-temporal modulations of natural sounds, forming a discriminative representation that captures the full range of modulation statistics that characterize natural sound ensembles. These findings have direct implications for our understanding of how sensory systems encode the informative components of natural stimuli and potentially facilitate multi-sensory integration.
| We explore a fundamental question with regard to the representation of sound in the auditory system, namely: what are the coding strategies that underlie observed neurophysiological responses in central auditory areas? There has been debate in recent years as to whether neural ensembles explicitly minimize their propensity to fire (the so-called sparse coding hypothesis) or whether neurons exhibit strong, sustained firing rates when processing their preferred stimuli. Using computational modeling, we directly confront issues raised in this debate, and our results suggest that not only does a sustained firing strategy yield a sparse representation of sound, but the principle yields emergent neural ensembles that capture the rich structural variations present in natural stimuli. In particular, spectro-temporal receptive fields (STRFs) have been widely used to characterize the processing mechanisms of central auditory neurons and have revealed much about the nature of sound processing in central auditory areas. In our paper, we demonstrate how neurons that maximize a sustained firing objective yield STRFs akin to those commonly measured in physiological studies, capturing a wide range of aspects of natural sounds over a variety of timescales, suggesting that such a coding strategy underlies observed neural responses.
| It is widely believed that sensory representations are optimized to process the stimuli to which they are exposed in natural environments [1]. Of particular interest is understanding the computational principles that underlie the generation of observed neural firing patterns. A popular hypothesis explored in recent years assumes that neural populations optimize a sparse code. This means that at any given time, only a small subset of a neural population fires to encode a given stimulus [2]. Such a representation is attractive for reasons of coding efficiency (see, e.g., [3]) and conservation of physiological resources [4]. The sparse coding hypothesis has enjoyed particular success in studies of vision (e.g., [5], [6]), and has also been supported more recently by both neurophysiological [7], [8] and computational studies [9]–[11] of the auditory system.
However, it has also been observed that some central auditory neurons, when driven by their preferred stimuli, exhibit sustained firing rates. Measuring from auditory thalamus and primary auditory cortex, Wang et al. observed that sustained responses were not simply phase-locked to the fast dynamics of the stimulus, suggesting that this rate-based code represented a meaningful, non-isomorphic transformation of the stimulus [12], [13]. Indeed, such a code is particularly important for audition since it directly addresses the issue of how to indicate the continued presence of a sound in a complex acoustic environment. Results from Petkov et al. have also illustrated how sustained responses play a role in auditory scene analysis, forming part of the neural basis for the perceptual restoration of foreground sounds against a cluttered background [14]. Moreover, Wang has argued that a rate-based representation is critical for matching fast temporal modulations present in natural sounds to slower rates found in higher cortical areas [15]. Slower dynamics in acoustic signals are believed to be the main carrier of information in speech and music [16]; are commensurate with temporal dynamics of stream formation and auditory grouping [17]; and may play an important role in multi-modal sensory integration [15]. Related computational studies in vision have suggested how this principle may underlie the shapes of simple and complex cell receptive fields in primary visual cortex [18], [19]. Importantly, a sustained firing rate, i.e., one that is persistent and therefore slowly changing over time, is related to slow feature analysis, a well-known method for extracting invariances from sensory signals [20] (see Discussion). To the best of our knowledge, however, there are no computational studies that explicitly consider the implications of a sustained firing-based code in central auditory areas.
At first glance, the two coding schemes are seemingly at odds: on the one hand a sparse code seeks to minimize the activity of a neural population whereas a sustained firing-based code requires that neural responses persist over time but still form an efficient representation of the stimulus. However, it appears that central auditory responses can strike a balance between the two strategies, with a large, transient population response at the onset of a sound, and a sparse subset of preferentially driven neurons exhibiting a strong, sustained response throughout the sound's duration [15], [21]. This picture suggests a mechanism for detecting and tracking target sounds in noisy acoustic environments and for generating a persistent signal that facilitates a stable perceptual representation. From a computational perspective, a better understanding of these mechanisms can inform models of auditory scene analysis as well as signal processing schemes for hearing prosthetics and automated sound processing systems.
A general computational approach for exploring the effects of particular coding strategies in sensory systems is based on optimizing a statistical objective criterion that quantifies the principle governing the transformation between stimulus and internal representation. Upon convergence, one then compares the emergent representation to known properties of the sensory system being studied [1]. Here, we apply this framework to explore how optimizing a sustained firing criterion influences the shapes of model auditory spectro-temporal receptive fields (STRFs) when processing natural sounds, and we compare the emergent ensembles to those obtained by optimizing a sparse coding objective. STRFs describe the linear mapping between a spectro-temporal stimulus and an instantaneous firing rate [22], and have proven useful not only for describing basic processing aspects of auditory neurons [23], [24], but also for shedding light on the nature of task-driven plasticity [25]. Figure 1 illustrates how a spectro-temporal stimulus is mapped to a set of instantaneous neural firing rates, whose ensemble response according to a desired coding strategy directly shapes the mapping.
In this paper, we show how this framework allows us to not only explore how the timescales of natural sounds are captured by and reflected in an emergent sensory representation, but reveal key similarities between choice of a sustained versus sparse code. Moreover, we demonstrate how a sustained firing-based code suggests a mechanism for an emergent discriminative representation for ensembles of natural stimuli.
We defined a sustained neural response as one where firing rate energy changes relatively slowly and is consequently highly correlated over time. In particular, we were interested in the characteristics of ensembles of model STRFs that promoted sustained responses over a specified time interval . Denoting the response of the neuron as , where is the STRF and is a spectro-temporal stimulus, we quantified this principle using the following objective function:(1)where denotes time average. Observe that represents the sum of correlations between signal energies of the neuron over a time interval defined by across an ensemble of neurons. If a neuron yielded a sustained response, then each of the would vary smoothly over the specified interval and we expect to be large. Moreover, choice of the correlation interval allowed us to directly explore the effect of different timescales on the ensembles that maximized Eq. 1. Finally, the weights were chosen to be linearly decaying for to , reflecting the intuition that recent activity of a neuron likely has more influence on the current output than the past. Note that these weights could be adapted to specifically model, for example, positive- or negative-monotonic sustained responses observed in physiological studies [13]. Full details of the optimization procedure can be found in Methods.
Alternatively, we explored an objective function that promoted sparsity. A natural way to induce sparsity in a population code is by enforcing a population response whose firing rate distribution is highly peaked near zero (representing frequent weak responses), but has long tails (representing infrequent large responses), i.e., a distribution with high kurtosis [26]. We quantified the sparsity of a population code using sample kurtosis:(2)where is the fourth central moment at time , is the population variance at time , and is the population mean at time .
For both and , the basic problem was to find an ensemble of STRFs that maximized the respective objective function subject to constraints that (1) bounded the amplitude of the filter responses and (2) minimized redundancy among the learned ensemble. This was achieved by enforcing the responses have unit variance and be mutually uncorrelated, i.e., where is the Kroenecker delta function (see Methods); we refer to these as response constraints. These constraints ensured that the responses had a bounded magnitude and that the STRFs did not all converge to the same solution.
We optimized both the sustained objective and sparsity objective using an ensemble of natural stimuli comprising speech, animal vocalizations, and ambient outdoor sounds. Each ensemble of filters was initialized at random using zero-mean, unit variance Gaussian noise, and each STRF covered from 0–250 ms in time and 62.5–4000 Hz along the tonotopic axis.
For the sustained objective, we considered a wide range of correlation intervals from very brief () to very long (). Examples of emergent STRFs for are shown in Figure 2A. For the spectro-temporal patches shown, red and blue colors indicate that the presence of energy in a particular spectro-temporal region yields excitatory and inhibitory responses, respectively. We observe a variety of STRFs that are highly localized, sensitive to narrowband spectral and temporal events, oriented, and some that are seemingly noise-like and not convergent to any particularly interesting shape. Importantly, such observations about these basic STRF classes align with those made in a number of previous physiological studies (see, e.g., [23], [24], [27]). Moreover, coverage of the STRFs appears to span the full time-frequency space. These results suggest that the sustained firing objective may underlie part of the coding strategy used by central auditory neurons.
Shown in Figure 2B are examples of emergent STRFs obtained by optimizing the sparsity objective. Indeed, this particular objective yields STRFs that are highly localized and sparsely distributed, with sensitivity to bandlimited spectral and temporal events. While both objective criteria yield noisy STRFs, it is clear that the sparse ensemble is much more noisy, with a less extensive coverage of the basic sound classes as observed with the sustained ensemble.
Since the information-bearing components of natural sounds vary concurrently across multiple timescales, it was expected that the structure of STRFs learned under the sustained objective would vary with the correlation interval . Indeed, inspection of the sustained ensembles for a range of suggested the presence of a number of latent classes whose membership varied smoothly from short to long correlation intervals. To quantify variations in population diversity over ecologically relevant timescales, we performed unsupervised clustering of the emergent STRFs and studied how class membership changed with objective function and correlation interval.
We pooled STRFs from the sparse ensemble and from the sustained ensembles for 10, 25, 50, 125, 250, 500, 1000, and 2000 ms, yielding a total of 3600 STRFs. We then applied normalized spectral clustering to discover latent classes among the pooled STRFs. In general, spectral clustering algorithms require an affinity matrix that specifies pairwise similarities between the objects being clustered. Viewing this affinity matrix as an undirected graph, spectral clustering finds a partition of the graph into groups whose elements have common similarity with one another. A natural measure of similarity between STRFs can be derived from the two-dimensional cross-correlation between pairs of spectro-temporal patches. Such a measure is similar to that considered by Woolley et al. [28] and is desirable since it does not depend on subjective choice of spectro-temporal features to use for clustering. In this work, we defined the measure of similarity between pairs of STRFs as the absolute value of the maximum value of the two-dimensional cross-correlation matrix; we used absolute value since we wished to group similar STRFs regardless of whether they were excitatory or inhibitory. Furthermore, as the STRFs tended to be distributed with a variety of phases in the input space, we considered cross-correlations for arbitrary time-frequency shifts (see Methods for details).
Results obtained using normalized spectral clustering of the emergent ensembles into nine classes are shown in Figure 3. In the center panel of the figure, a stacked bar chart illustrates the the percentage of STRFs at a particular assigned to one of nine classes. Different segment colors correspond to each of the nine classes, and segment width is proportional to the number of STRFs assigned to that class. Surrounding the bar chart are examples from six classes that best illustrate how diversity varies with , namely noisy, localized, spectral, complex, temporal, and directional classes. These labels are qualitative descriptors of each class and not quantitative assessments of the time-frequency characteristics of each category.
Inspection of the cluster groupings reveal rich structural variations over a wide range of correlation intervals. In particular, the STRFs labeled according to the noisy class are found to dominate the sparse ensemble, with a large presence in the sustained ensemble for . Membership in this class drops for between 10 and 125 ms, and begins to increase at 125 ms. We also observe that short correlation intervals (10, 25, and 50 ms) have a large concentration of localized STRFs, with membership dropping with increasing . While the temporal class holds relatively steady across the sustained ensembles, we find that membership in the directional, complex, and spectral classes varied smoothly across . In general, we find that ensemble diversity is maximized for (max. entropy of 3.08 bits), but the overall trends suggest rich ensemble structure between 10 and 250 ms, which is notably in the range of the timescales of natural sounds [29], . This is further supported by the increasing presence of noisy STRFs for large correlation intervals (1000 and 2000 ms).
In addition to studying structural variations in the shapes of the emergent STRFs, it is also of interest to examine the structure of the STRF outputs in response to natural sounds. In particular, we sought to address the extent to which enforcing sustained responses does indeed yield responses that persist over time. We defined the neuron to be significantly “active” when its firing rate exceeded 1 standard deviation over time. While this is not meant to be a precise measure of a neuron's activation (since, for instance, the firing rate is not used to modulate a Poisson spike generation process), such a measure nevertheless quantifies and characterizes a strong versus weak ensemble response to natural stimuli.
Shown in Figure 4A are the distribution of activation times for individual neurons for ensembles of 10 and 125 ms in response to a held-out set of natural stimuli. The neurons are shown sorted according to decreasing median activation time, and the interquartile ranges of activation time are indicated by the shaded regions. We observed that the most diversity in median activation times across ensembles occurred in approximately the top 10% of the most persistent neurons. To summarize these observations, we considered the distribution of median activation times of the top 10% of neurons with most persistent responses (i.e., the top 40 neurons); these distributions are illustrated as boxplots in Figure 4B.
As noted previously with the clustering results, shorter values favor mostly localized and noisy STRFs and consequently it was expected that activations would be brief. Interestingly, however, we observe that with increasing , median activations peak between 50 and 500 ms and fall off for large despite the STRFs being optimized to promote sustained responses over long intervals. This overall trend aligns with the previous clustering results that demonstrate how population diversity is maximized over intervals corresponding to timescales that predominate natural stimuli. The STRFs corresponding to the top 10% most persistent responses for are shown in Supplementary Figure 1, and we find that they generally have a spectral tuning, but are fairly narrowband and localized.
Additionally, we considered the responses of the top 40 most persistent responses obtained using the sparsity objective function; the distribution of median activations is in the first column of Figure 4B. We find that the sparse ensemble yields responses most similar to those for short .
How do the emergent STRFs learned under the sustained firing objective compare to those observed in physiological studies? Broadly speaking, we find that the emergent STRFs share many of the trends with biological receptive fields typically observed in animal models. We explored this issue by comparing our model ensembles with a set of 1586 STRFs recorded from awake, non-behaving ferret primary auditory cortex using TORC [31] and speech stimuli [27], [32] (see Methods for more details). Where applicable, we also compared our results with reported results from anesthetized ferrets by Depireux et al. [23] and cats by Miller et al. [24] in the literature.
Illustrative examples of the types of STRFs found in the neural data are shown in Figure 5. In particular, we find neural STRFs that are qualitatively similar those found in the localized, complex, noisy, and directional clusters shown earlier in Figure 3. Because the temporal and spectral sampling rates used in our model are higher than those used in the physiological data, we did not find good matches with the temporal and spectral classes.
To visualize the overlap between the spectro-temporal modulation coverage of the neural and model STRFs, we used the ensemble modulation transfer function (eMTF). The eMTF is derived by averaging the magnitude of the 2D Fourier Transform of each neuron in a given ensemble, and jointly characterizes modulations in time (rate, in Hz) and in frequency (scale, in cyc/oct). We first applied normalized spectral clustering to the neural STRFs to obtain nine clusters. Next, we computed the eMTF for each cluster, extracted isoline contours at the 65% level, and overlaid these curves on the eMTF of the model STRFs for . These results are shown in Figure 6 and illustrate the overlap between the model and neural data, particularly at the “edges” of the neural STRF modulations. While the overlap is not complete, it is clear that the modulation spectra of each ensemble are not disjoint. Moreover, the model eMTF suggests a general ensemble sensitivity to relatively fast modulations; this point is explored further in a later section (“Emergent STRFs capture spectro-temporal modulation statistics of stimulus”).
To better characterize the relationship between the neural and model data, we employed a statistical comparison of the distribution of the two datasets. If the models truly generated STRFs similar to those in physiological studies, then one might expect a nearest-neighbor (NN) similarity distribution akin to one derived from the neural ensemble we considered. We computed the symmetric KL-divergence between each of the model and within-physiology NN similarity distributions (shown in Supplemental Figure 2). We found that the sustained-response (presented here) and sustained-shape (presented later in this paper) distributions had KL divergences of 0.80 and 0.85, respectively, whereas the sparse distribution had a KL distance of 1.05. KL typically measures the expected number of bits required to code samples from one distribution using codes from the other. While these numbers are difficult to assess in absolute terms, they give a sense of how the different model optimizations and constraints compare to each other. These numbers reveal that the sustained ensembles are similarly comparable to the physiology, whereas the sparse ensemble has a somewhat worse match. Of course, caution must be taken with these numbers because the set of neural STRFs we analyzed represent only a subset of mappings that likely exist in central auditory areas.
Next, we measured a variety of parameters from the neural and model STRFs (for ) that more fully characterized the extent of spectro-temporal coverage and modulation sensitivity of the ensembles (see Methods), the results of which are summarized in Figure 7.
Based on the distribution of directionality indices, shown in panel (A), we observe that the model STRFs are largely symmetric, with the majority of neurons having no preference for upward or downward moving input stimuli (mean0). As indicated by the tails of this distribution, however, a subset of neurons have a strong directional preference. This agrees with the neural STRFs, and similar observations have been made in MGB and primary auditory cortex of cats by Miller et al., as well as in measurements by Depireux et al. from primary auditory cortex of ferrets. Furthermore, panel (B) illustrates that a large number of model STRFs are fairly separable, with a peak in the separability index (SPI) distribution around 0.10 and an average value of 0.26. This trend aligns with values reported in the literature by Depireux et al. in measurements from ferret auditory cortex (mean of approx. 0.25). However, it is worth noting that this low level of separability is not uniformly reported across physiological studies of receptive field of mammalian auditory cortex. For instance, the physiological data analyzed in the current study (examples of which are shown in Figure 5) do yield a higher average SPI (mean = 0.37).
The temporal modulation statistics of the model STRFs, as quantified by best rate (BR), also align generally with results reported from mammalian thalamus and cortex. In panel (C) we observe a broad, bandpass distribution of best rates, with an average of 23.9 Hz. Reported physiological results from Miller et al. show similarly broad ranges of temporal tuning with preferences around 16 Hz and 30 Hz range for cortex and thalamus, respectively. The neural STRFs we analyzed show a somewhat slower tuning, with an average BR of 9.5 Hz. Furthermore, in panel (D), we computed the normalized average rate profile from the model STRFs. We observe a peak at 7.8 Hz, with an upper 6-dB cutoff of 34.4 Hz. Here we find a close overlap with the rate profile computed from the neural STRFs as well as with average profile results as reported by Miller et al. (peak at 12.8 Hz; upper 6-dB cutoff at 37.4 Hz).
The spectral modulation statistics of the model STRFs, as quantified by best scale, are generally faster than those reported from studies of thalamic and cortical nuclei. The distribution of best scales shown in panel (E) is bandpass with a wide range of slow to fast spectral coverage, with an average tuning of 1.40 cyc/oct. The neural STRFs, in contrast, are tuned to much slower scales (mean = 0.47 cyc/oct). Similarly, results from Miller et al. in MGB indicate a generally slower tuning (0.58 cyc/oct), whereas measurements from cortical neurons, while having a similarly wide range of tunings as with the model, indicate a slower average value of 0.46 cyc/oct and an upper cutoff of approx. 2 cyc/oct.
Finally, the ensemble average scale profile, shown in panel (F), is bandpass and exhibits a peak at 0.7 cyc/oct with an upper 6-dB cutoff of 2.9 cyc/oct. The neural STRFs, however, are much slower with peak at 0.2 cyc/oct and an upper cutoff of 1.9 cyc/oct. This is similar to observations from MGB by Miller et al., where they reported that the ensemble average scale profile is generally low-pass, with average scale profile peaks and upper 6-dB cutoffs at 0 cyc/oct and 1.3 cyc/oct, respectively, with similar observations in cortex.
In summary, while we cannot map the emergent STRFs to any exact synapse, they nevertheless reflect the general processing characteristics of various stations along in the central auditory pathway. There is good alignment with the neural STRFs and reported results in mammalian MGB and primary auditory cortex with respect to directional sensitivity and spectro-temporal separability. The temporal modulation statistics of the emergent sustained STRFs appear to be most similar to those measured from thalamus and cortex. Furthermore, the model STRFs are generally faster with regard to spectral modulations than those measured from thalamus and cortex.
To explore the relationship between STRFs optimized to promote sustained responses and those that explicitly maximize population sparsity, we compared the average responses of the sustained ensemble for with the sparse ensemble. Specifically, we used the converged STRFs to analyze a held-out set of natural stimuli, computed a histogram of the population responses at each time, and computed the average histogram across the entire test input (see Methods). Since the sparse ensemble was optimized to yield a highly kurtotic firing rate distribution, it was of interest to examine the shape of the distribution when promoting sustained responses.
Results comparing the average histograms of sustained versus sparse responses is shown in Figure 8, with log-probabilities shown on the vertical axis to emphasize differences between the tails of the distributions. The main observation is that both the sustained and sparse ensembles have distributions that have long tails and are are highly peaked around a firing rate of zero. For reference, we show the average histograms obtained by filtering the stimulus through the first 400 principal components of the stimulus (see Supplemental Figure 3) as well as through a set of 400 random STRFs; a zero-mean, unit variance Gaussian distribution is also shown. Therefore, despite promoting temporally persistent responses, the sustained responses yield a population response that is not altogether different from an ensemble that explicitly maximizes kurtosis. Interestingly, this observation was also made by Berkes and Wiscott in the context of complex cell processing in primary visual cortex (see Sec. 6 of [33]).
Finally, we sought to explore the consequences of relaxing the constraint that the responses be mutually uncorrelated. Rather than directly constrain the responses, we considered constraints to the shapes of the model STRFs. This was achieved by solvingi.e., we require the STRFs to form an orthonormal basis. So long as the stimuli are bounded, this set of constraints meets our requirements that (1) the output of the STRFs be bounded and (2) we minimize redundancy in the learned ensemble. As before, the optimization is described in the Methods. We consider an ensemble size of STRFs initialized at random. Examples of shape-constrained STRFs that optimize the sustained objective function for are shown in Figure 9. Again, we observe STRFs that are bandpass, localized, oriented, and sensitive to a variety of spectral and temporal input. However, there was an apparent difference between the speed of the spectro-temporal modulations and those from STRFs learned subject to the response constraints.
It is well known that natural sound ensembles are composed largely of slow spectro-temporal modulations [29], [30], [34]. However, the emergent STRFs learned subject to response constraints appear to be tuned to relatively fast spectral and temporal modulations, whereas the STRFs learned subject to shape constraints appear to have a broader tuning. To further examine how both sets of constraints jointly capture and are related to the spectro-temporal modulations observed in stimulus, we compared the average 2D modulation profile of the stimulus to the eMTFs derived from both sets of constraints.
An interesting view of how the emergent STRFs capture the spectro-temporal modulations of the stimulus is illustrated in Figure 10 for . Shown is the average 2D modulation profile of the stimulus overlaid with a single isoline contour (at the 65% level) of the eMTFs learned subject to response (thick red lines) and shape constraints (thick black lines). We also show the constellation of BR versus BS for each ensemble (indicated by ‘’ and ‘’ for response and shape constraints, respectively). As implied by the contours, the response constraints yield STRFs that follow the spectro-temporal “edge” of the stimulus, while the shape constraints explicitly capture most of the “slowness” of the stimulus. As mentioned previously, the response constraints effectively force the temporal response of the sustained ensemble to be sparse, which consequently results in highly selective STRFs that tend to be tuned to fast modulations. Nevertheless, they implicitly capture the spectro-temporal extent of the stimulus. Moreover, since the shape constraints effectively force the STRFs to form a basis that spans the input space, this results in neurons that explicitly capture the slow modulations of the stimulus. Similar observations were made across the range of , and for each case it was clear that the spectro-temporal modulations of the stimulus are fully captured by the combination of both sets of constraints.
In this paper, we considered a framework for studying how choice of a sustained firing versus sparse coding objective affects the shapes of model spectro-temporal receptive fields in central auditory areas. The sparse coding objective considered here, namely that of maximizing population kurtosis, yields STRFs that are mostly noisy. Those that do converge are generally highly localized. In contrast, enforcing the sustained firing objective subject to the same response constraints yields richly structured ensembles of STRFs whose population diversity varies smoothly with the correlation interval . Of course, the observed structural variations are necessarily biased due to construction of the stimulus. Nevertheless, this diversity, as revealed by the results of the unsupervised clustering, paired with the responses of the most persistent STRFs, supports the notion that sustained neural firings are preferred in the range of timescales predominant in natural sounds. While we do not necessarily attribute the emergent sustained STRFs to any particular synapse in the auditory pathway, we instead note that the observed filters exhibit general similarities to physiological observations made in auditory thalamus and cortex.
We also observed that enforcing the sustained firing objective with response constraints yields an ensemble firing rate distribution that is similar, on average, to one where population sparsity was explicitly enforced. This supports the proposal that the two coding objectives are not necessarily at odds, and that in some sense a sustained firing objective yields “sparsity for free.” Of course, the sustained firing and sparse coding objectives could be quantified in many different ways (see, e.g., Hashimoto [35] and Carlson et al. [11]), but the present study is a promising step in understanding their relationship in the central auditory system from a computational perspective.
Finally, to explore the consequences of relaxing the constraint that the responses be mutually uncorrelated, we explored an alternative set of orthonormality constraints on the sustained firing objective. While still minimizing a notion of redundancy, we observed that the emergent ensembles are generally slower, potentially better capturing the slow spectro-temporal modulations known to be present in natural sounds. This experiment further demonstrated the utility of the considered framework for directly addressing questions about coding schemes and various sets of constraints in representing sound in central auditory areas.
The combination of shape and response constraints on the sustained objective function yield STRF ensembles that appear to jointly capture the full range of spectro-temporal modulations in the stimulus. However, the distinct differences in MTF coverage illustrate the tradeoff between redundancy and efficiency in sensory representations. In particular, the shape constraints yield STRFs that are somewhat akin to the first few principal components of the stimulus (see Supplemental Figure 3). This is not surprising given that the objective function defines a notion of variance of linear projections, the component vectors of which are constrained to form an orthonormal basis. However, since the responses are not strictly enforced to be uncorrelated, orthonormality imposed on the filter shapes does not necessarily reduce redundancy in the resulting neural responses.
In contrast, the response constraints yield STRFs that are highly selective to the input and are thus comparatively “fast” in the modulation domain. This representation can be thought of as more efficient since at any given time only a few neurons have a large response. However, while the shapes of individual STRFs fail to explicitly capture the slow spectro-temporal modulations predominant in natural sounds, it instead appears that the ensemble MTF of the response-constrained STRFs collectively forms a contour around the high-energy modulations of the stimulus that implicitly capture its spectro-temporal extent.
Is this contouring of the average modulation spectrum of natural sounds something performed by the auditory system? The neural STRFs we considered certainly had an eMTF that reflects a tuning to slower modulations near the MTF origin. However, there is some evidence that the auditory system uses an “edge”-sensitive, discriminative modulation profile for analyzing sound. Woolley et al. [36], in an avian study, showed that the eMTF of neurons from Field L (the avian A1 analog) has a bandpass temporal modulation profile (at low scales) that facilitates a discriminative tuning of temporal modulations among classes of natural sounds. Nagel and Doupe [37] have also shown examples of avian Field L STRFs that orient themselves near the spectro-temporal “edge” of the stimulus space. Moreover, Rodriguez et al. [38], in a study of mammalian IC neurons, showed that neural bandwidths can scale to better capture fast, but less frequently occurring, modulations. In light of these observations, the modulation profiles observed from the sustained STRFs for both response and shape constraints are consistent with the notion that the auditory system makes an explicit effort to capture all modulations present in natural sounds: fast, feature-selective, and consequently discriminative modulations, as well as frequently occurring slow modulations.
The notion that sustained neural firings form part of the neural representation of sensory systems is not limited exclusively to the auditory modality. In fact, the sustained firing objective considered in this paper is related to a broad class of sensory coding strategies referred to collectively under the temporal slowness hypothesis. This concept proposes that the responses of sensory neurons reflect the time-course of the information-bearing components of the stimulus—which are often much slower with respect to the fast variations observed in the stimulus—and may therefore reflect invariant aspects of the sensory objects in the environment. Examples of early neural network models exploring slowness as a learning principle were considered by Földiák [39], Mitchison [40], and Becker [41]. More recently, a number of computational studies, particularly in vision, have established slowness as a general sensory coding strategy and have revealed relationships with a number of general machine learning techniques. Here we outline the connections between the sustained firing criterion considered in this study and previous work.
Our definition of the sustained firing objective, , was adapted from a notion of temporal stability proposed by Hurri and Hyvärinen termed temporal response strength correlation (TRSC) [18]. This study considered modeling of simple cells in primary visual cortex, and their objective function was defined as(3)for a single fixed . By maximizing subject to the decorrelation constraints , they showed the emergence of spatial receptive fields similar to those observed in simple cells in primary visual cortex. It is clear that the objective functions and are equivalent for a single time step, but the main difference between the two is that we sought to enforce temporal stability over a time interval , rather than between two distinct times and . Interestingly, optimization of the TRSC objective was shown by Hyvärinen to yield a solution to the blind source separation problem [42], suggesting perhaps that in the auditory domain, such a criterion may underlie separation of overlapping acoustic sources.
The sustained firing objective is also related to a well-known model of temporal slowness known as slow feature analysis (SFA) [20]. The computational goal of SFA is to find a mapping of an input that extracts the slow, and presumably more invariant, information in the stimulus. Briefly, for an input , linear SFA finds mappings that minimize(4)subject to , , and . Note that the input is not necessarily the raw stimulus but could represent a non-linear expansion of the input, akin to applying a kernel function in a support vector machine [43]. Therefore, SFA finds a mapping of the input that varies little over time and whose outputs are bounded and mutually uncorrelated. In the visual domain, Berkes and Wiskott found that SFA could explain a variety of complex cell phenomena in primary visual cortex such as the emergence of Gabor-like receptive fields, phase invariance, various forms of inhibition, and directional sensitivity [33]. Similar to our study, they also found the emergence of a sparse population code based on SFA. More importantly, however, they established a link between SFA at the level of complex cells and , which in turn links to the sustained firing objective explored in our study. Specifically, they showed that when a complex cell output is expressed as a quadratic form [35], [44], the SFA objective could be written as(5)which is equivalent to maximizing (and thus for a single time-step) plus cross-correlation terms. As noted by Berkes and Wiskott, this relationship suggests that sustained firing rates at the level of simple cells are modulated as part of a hierarchical cortical processing scheme in primary visual cortex. Given the increasing understanding of such hierarchical circuits in the auditory system [45], the possibility that sustained firing rates are varied as part of a higher-order processing strategy in primary auditory areas is an exciting prospect worth further exploration.
Other important relationships exist between SFA and a number of general machine learning principles. Blaschke et al. [46] established a relationship between SFA and independent component analysis, a widely used method for blind source separation (see, e.g., [47]). Klampfl and Maass [48] showed that under certain slowness assumptions about the underlying class labels in observed data, SFA finds a discriminative projection of the input similar to Fisher's linear discriminant. Furthermore, SFA has links to methods for nonlinear dimensionality reduction: Creutzig and Sprekeler [49] described the link between SFA and the information bottleneck whereas Sprekeler [50] showed a connection between SFA and Laplacian eigenmaps.
In summary, the temporal slowness hypothesis forms a sound basis for learning a representation from data with rich temporal structure. Slowness as a learning principle has also been shown to explain the emergence of simple and complex cell properties in primary visual cortex. As described above, the sustained firing principle considered in this paper has fundamental links to SFA, which in turn is related to a number of general machine learning strategies. To the best of our knowledge, ours is the first thorough study that establishes a link between the temporal slowness hypothesis and an emergent spectro-temporal representation of sound in central auditory areas.
The ensemble modulation coverage results are particularly interesting since it is widely thought that “slow” spectro-temporal modulations carry much of the message-bearing information for human speech perception. Furthermore, it is known in the speech processing community that features that capture slow temporal [51] and joint spectro-temporal modulations [52], [53] are important for noise-robust automatic speech recognition. The observed contouring effect resulting from the sustained firing criterion may thus reflect a mechanism to detect the spectro-temporal “edges” of the message-bearing components of the stimulus, and possibly contribute to a noise-robust representation of sound. We have recently considered this principle and have demonstrated that 2D bandpass filters derived from eMTF contours learned from a speech-only stimulus yield state-of-the-art noise-robust acoustic features for automatic speech recognition [54]. Moreover, it is possible that the contour level may be chosen adaptively as a function of ambient signal-to-noise ratio to better capture variations in the high-energy modulations of the stimulus. Also, since the emergent STRFs capture general spectro-temporal patterns that characterize the stimulus, it is possible that ensembles of STRFs could be learned in various speech-plus-noise scenarios to perhaps better characterize noise-corrupted acoustic environments. Such hypotheses can be readily verified experimentally and may have practical impact to automated sound processing systems in noisy acoustic environments.
Finally, the framework considered in this paper can be extended in a number of ways. For instance, to address the linearity limitation of the STRF, it is worthwhile to consider a model based on a linear-nonlinear cascade [55]. As mentioned earlier, the auditory pathway is necessarily hierarchical, and warrants consideration of hierarchical computational models. Indeed, recent physiological evidence also indicates that the representation becomes increasingly complex and nonlinear as one moves from away thalamo-recipient layers in primary auditory cortex (for a review, see [45]). Finally, a recent computational study in vision by Cadieu and Olshausen [56] proposes a hierarchical generative model that explicitly unifies notions of sparse coding and temporal stability. In particular, a two-layer network learns a sparse input representation whose activations vary smoothly over time, whereas a second layer modulates the plasticity of the first layer, resulting in a smooth time-varying basis for image sequences. One can imagine that such a framework could be extended to spectro-temporal acoustic stimuli.
An ensemble of natural sounds comprising segments of speech, animal vocalizations, and ambient outdoor noises was assembled for use as stimuli. Two sets were generated, one for training and one for evaluating the response characteristics of the STRFs. Phonetically balanced sentences read by male and female speakers were used [57]. Examples of animal vocalizations included barking dogs, bleating goats, and chattering monkeys [58]. The ambient sounds included, for example, babbling creeks and blowing wind, and other outdoor noises. The speech utterances were approximately three seconds each and comprised 50% of the stimulus. The animal vocalizations and ambient sounds formed the remaining 50% of the stimulus (25% each), were broken into three-second segments, and were windowed using a raised cosine window to avoid transient effects. Finally, segments from each class were downsampled to 8 kHz, standardized to be zero-mean and unit variance, and randomly concatenated to yield a waveform approximately three minutes in overall length, i.e., 90 seconds of speech, 45 seconds of animal vocalizations, and 45 seconds of ambient outdoor noises.
We used a computational model of peripheral processing to account for the transformation of a monaural acoustic stimulus to a joint time-frequency representation in the auditory midbrain; this representation is referred to as an auditory spectrogram [59], [60]. The auditory spectrogram represents the time-varying spectral energy distribution on the (logarithmic) tonotopic axis, and accounts for the physiology of inner hair cell transduction and filtering on the auditory nerve, enhanced frequency selectivity in the cochlear nucleus via a lateral inhibitory network, and the loss of phase locking to stimuli observed in midbrain nuclei. The specific model details have been presented previously and as such we forego a detailed description here, except to note that we sampled the log-frequency axis over six octaves with ten equally spaced channels per octave, with a short-term integration interval of 5 ms, i.e., we obtained a 60 channel spectral vector every 5 ms. An example auditory spectrogram is shown for a segment of speech in Figure 1A.
To quantify the relationship between a spectro-temporal stimulus and its corresponding response in central auditory areas, we used the spectro-temporal receptive field. Such a functional characterization of a neuron is useful for identifying the components of the stimulus to which it is most sensitive. An STRF models the linear transformation of a time-varying spectro-temporal input to an instantaneous firing rate, i.e.,(6)where is an LTI filter that defines the STRF, is a spectro-temporal stimulus, and is the average firing rate. Without loss of generality, we assume . Observe that the mapping represents convolution in time and integration across all frequencies, and we can interpret the STRF as a matched filter that acts on the input auditory spectrogram.
For discrete-time signals and filters, and assuming that has a finite impulse response, we can express Eq. 6 compactly in vector notation as(7)where are column vectors denoting the stimulus and filter, respectively [61]. Furthermore, to express the response of an ensemble of neurons, we concatenate the STRFs into a matrix and write(8)
From the stimulus auditory spectrogram, we extracted 250 ms spectro-temporal segments once every 5 ms. Each segment was stacked columnwise into a vector where (i.e., 50 vectors/segment 60 channels). A total of 30 k spectro-temporal vectors were extracted from the stimulus. We subtracted the local mean from each segment and scaled each vector to be unit norm [18], and note that this pre-processing was also applied to the test stimulus used for evaluating the STRF response characteristics. Finally, each spectro-temporal input patch was processed by the ensemble of STRFs to yield a population response . Figure 1B illustrates the procedure for obtaining stimulus vectors and response vector .
To constrain the responses of the STRFs to have unit variance and be mutually uncorrelated, we first note that the individual constraints can be written aswhich can then be compactly expressed as an ensemble constraint(9)where denotes the sample covariance matrix and is the identity matrix. Since is real-symmetric, it is unitarily diagonalizable as , where is a matrix of (columnwise) eigenvectors with corresponding eigenvalues along the diagonal of . Substituting this decomposition into Eq. 9, we obtainedwhere . By recasting the constraints, we can rewrite the original matrix of STRFs as and consequentlywhere corresponds to a whitening of the input acoustic data, i.e., has a spherical covariance matrix. For computational efficiency, we reduced the dimensionality of the input using a subset of the principal components of the stimulus, i.e.,where and , , are the matrices of eigenvalues and eigenvectors, respectively, that captured 95% of the variance of the input. In this work, we found . Therefore, the core problem we wished to solve is:(10)where corresponded to either the sustained firing or sparse coding objective function.
To optimize this nonlinear program, we used the gradient projection method due to Rosen, the basic idea of which is as follows [62], [63]. Let denote the update to the matrix of (rotated and scaled) STRFs , let be a learning rate, and let be an integer used to adjust the learning rate. Assume is a matrix with orthonormal columns that is a feasible solution to the problem in Eq. 10. We updated via gradient ascent as follows:(11)where is a projection of the gradient update so that satisfies the orthonormality constraint required in Eq. 10. If the update was such that , we set and recomputed the projected gradient update, repeating until was non-decreasing. Finally, learning ceased when the relative change between and fell below a threshold or a maximum number of iterations were reached; in our experiments, we stopped learning for or a maximum number of 30 iterations. Upon convergence, the desired STRFs were obtained using . Note that for the case of the sustained firing objective, was formed from the sum of independent terms, allowing us to directly sort the emergent STRFs according to their contribution to the overall objective function; such a sorting was not possible for the sparsity objective.
Of course, the above procedure required a suitable projection , and one was derived as follows [64]. In general, for a matrix , we wish to find a matrix with orthonormal columns that minimizesIntroducing a symmetric matrix of Lagrange multipliers , and recalling that , we sought to find a stationary point of the LagrangianComputing the (elementwise) partial derivative of w.r.t. and setting it to we obtained [65]Observing thatwe have thatAssuming had full column rank, then an optimal orthogonal matrix that minimized that can be used for the projection in Eq. 11 was found as(12)
Finally, to optimize a given objective function subject to the STRFs being orthonormal, i.e., , we solveHere we can again use Rosen's projected gradient method in Eq. 11 along with the projection defined in Eq. 12, but the only difference from before is that it does not require pre-whitening of the stimulus.
We first characterized the emergent STRFs based on parameters that described their individual spectro-temporal and modulation tuning.
Next, we considered measures that characterized a variety of ensemble-based spectro-temporal and modulation properties.
To summarize the spectro-temporal modulations present in the natural sound stimulus, we averaged the magnitude of the 2D Fourier transform of 250 ms patches (non-overlapping) of the auditory spectrogram.
The optimization procedure resulted in a set of richly structured patterns that suggested the presence of a number of latent classes whose membership varied with both choice of objective function and correlation interval . To quantify these variations, we applied the normalized spectral clustering algorithm of Ng et al. [66].
We defined the similarity between a given pair of STRFs and by computing the normalized 2D cross-correlation matrix for arbitrary shifts in time and frequency and selecting the maximum of the absolute value of this matrix, i.e.,whereImportantly, the absolute value of the cross correlation was used here since we wished to group STRFs regardless of whether they were excitatory or inhibitory. Next, we pooled all STRFs we sought to cluster and constructed a pairwise similarity matrix . Viewing as a fully connected graph with edge weights specified by , spectral clustering finds a partitioning of the graph into groups such that edges between groups have low similarity whereas edges within a group have high similarity.
Defining the degree matrix where and unnormalized graph Laplacian , the normalized spectral clustering algorithm is as follows:
We clustered the STRFs initially into 12 groups. While this number was necessarily an arbitrary choice, it was found to sufficiently capture variations in population diversity with . However, we found that (i) three of the resulting clusters could be reasonably labeled as noisy, whereas (ii) two of the resulting clusters could be reliably labeled as localized; merely reducing the number of initial classes did not merge the clusters, but instead blurred distinctions among the other major categories we sought to study. We interpreted noisy patterns as those with no obvious spectro-temporal structure and not indicative of any subset of the stimulus.
Merging of the initial 12 classes was achieved by computing the average of STRFs from the initial class labels and ranking the classes in descending order. Indeed, the three noisy classes had the highest average and consequently resulted in a group with average greater than 0.5. Similarly, the localized STRFs were typically highly spherical and sorting the initial clusters by resulted in the two localized classes to be ranked highest. Consequently, we grouped these two clusters that had an average of greater than 0.69. This resulted in a final cluster count of nine classes.
We obtained ensembles of neural STRFs estimated using TORC [31] and speech stimuli [27], [32]. There were 2145 TORC and 793 speech STRFs, and each STRF was pre-processed to cover 110 ms in time (sampling rate = 100 Hz) and span 5 octaves in frequency (sampling rate = 5 cyc/oct). For the spectral clustering analysis, we subsampled the TORC set by randomly selecting 793 STRFs and combined them with the speech STRFs, yielding a total of 1586 STRFs in the neural data set. In this way, the neural data analysis was not biased towards one stimulus type or the other.
|
10.1371/journal.pntd.0001662 | Early Prediction of Treatment Efficacy in Second-Stage Gambiense Human African Trypanosomiasis | Human African trypanosomiasis is fatal without treatment. The long post-treatment follow-up (24 months) required to assess cure complicates patient management and is a major obstacle in the development of new therapies. We analyzed individual patient data from 12 programs conducted by Médecins Sans Frontières in Uganda, Sudan, Angola, Central African Republic, Republic of Congo and Democratic Republic of Congo searching for early efficacy indicators.
Patients analyzed had confirmed second-stage disease with complete follow-up and confirmed outcome (cure or relapse), and had CSF leucocytes counts (CSFLC) performed at 6 months post-treatment. We excluded patients with uncertain efficacy outcome: incomplete follow-up, death, relapse diagnosed with CSFLC below 50/µL and no trypanosomes. We analyzed the 6-month CSFLC via receiver-operator-characteristic curves. For each cut-off value we calculated sensitivity, specificity and likelihood ratios (LR+ and LR−). We assessed the association of the optimal cut-off with the probability of relapsing via random-intercept logistic regression. We also explored two-step (6 and 12 months) composite algorithms using the CSFLC.
The most accurate cut-off to predict outcome was 10 leucocytes/µL (n = 1822, 76.2% sensitivity, 80.4% specificity, 3.89 LR+, 0.29 LR−). Multivariate analysis confirmed its association with outcome (odds ratio = 17.2). The best algorithm established cure at 6 months with < = 5 leucocytes/µL and relapse with > = 50 leucocytes/µL; patients between these values were discriminated at 12 months by a 20 leucocytes/µL cut-off (n = 2190, 87.4% sensitivity, 97.7% specificity, 37.84 LR+, 0.13 LR−).
The 6-month CSFLC can predict outcome with some limitations. Two-step algorithms enhance the accuracy but impose 12-month follow-up for some patients. For early estimation of efficacy in clinical trials and for individual patients in the field, several options exist that can be used according to priorities.
| Because Human African trypanosomiasis is fatal, it is crucial for the patient to determine if curative treatment has been effective. Unfortunately this is not possible without a 24-month laboratory follow-up, which is problematic and largely unaccomplished in the field reality. Studies that assessed early indicators have used small cohorts, yielding limited statistical power plus potential bias because of including patients with equivocal outcome. We tackled this problem by pooling a large dataset which allowed for selecting cases providing strictly unequivocal information, still numerous enough to produce sound statistical evidence. We studied predictors based on the CSF leucocytes count, a laboratory technique already available in the field, evaluating their predictive power at 6 and 12 months post-treatment. We found a predictor at 6 months (10 leucocytes/µL of CSF) that has sub-optimal accuracy but may be valuable in some particular situations, plus two-step algorithms at 6 and 12 months that offer sufficient confidence to shorten the patients' follow-up. Until better biomarkers are identified, these findings represent a significant advance for this neglected disease. Benefits are foreseen both for patients and for overburdened treatment facilities. In addition, research for new treatments can be accelerated by using early predictors.
| Human African trypanosomiasis (HAT) or sleeping sickness, caused by Trypanosoma brucei gambiense (most common form, West and Central Africa) and rhodesiense (East and Southern Africa), is fatal unless treated. After infection, the disease progresses from the easily treatable haemolymphatic first stage to the meningoencephalitic second stage, when parasites invade the central nervous system.
Patients who receive treatment can not be considered cured immediately, because the parasite may remain viable, redeveloping fully the disease many months later. A long post-treatment follow-up period is thus required to assess cure [1]. This follow-up time is fixed at 24 months by convention, although in comparative clinical trials it is considered acceptable to measure the efficacy at 18 months [2]. Follow-up consists of control visits generally every 6 months when lymph, blood and cerebrospinal fluid (CSF) are examined. The detection of trypanosomes in any body fluid unequivocally identifies a relapse. Unfortunately, parasites are often not detected early enough to allow for timely re-treatment, plus many patients do not adhere to this demanding and invasive follow-up schedule. To better detect the relapses and avert the risk for serious sequelae or death, the variation in number of white blood cells (WBC) in CSF is widely used as a proxy marker of relapse. Other markers of relapse are under investigation and not in routine field use.
Because most HAT patients are located in remote rural areas, the post-therapeutic follow-up is particularly challenging: poverty, distance, bad roads, lack of transportation, subsistence priorities, displacement (sometimes conflict related), add to the fear of the lumbar puncture. As a result, patients' compliance with follow-up decreases markedly after the first assessment at 6 months [3].
Such long follow-up is a handicap not only for routine patient management but also for therapeutic efficacy studies [4], and particularly when a sequence of clinical studies is required (e.g. dose-finding studies). Some time can be saved when a given investigational treatment is assumed to have insufficient efficacy due to early failures surpassing a pre-defined threshold. However, when the cumulative failure rate is below that threshold, the risk of subsequent final outcome (cure or relapse) can not be predicted.
Research on ways of shortening the follow-up is scarce. One study suggests that HAT patients with <5 CSF leucocytes/µL at 6 months are at low risk of relapse (negative predictive value >0.93, n = 146) [5] and that at 6 and 12 months, patients with ≥50 and ≥20CSF leucocytes/µL, respectively, are at high risk. Another study tested an algorithm combining 6 and 12 months CSF exams on a cohort of 206 treated patients showing 97.8% specificity and 94.4% sensitivity to predict relapse [6]. Considering that these promising findings originated from relatively small cohorts, recruited each time in one single centre (Bwamanda and Mbuji Mayi, DRC, respectively), and that confirmed and unconfirmed efficacy outcomes (lost to follow-up, deaths during follow-up, etc) were mixed in the assessment via assumptions, further research is needed on larger datasets and with more restrictive selection criteria.
To meet this goal, we consolidated individual-patient data from 12 sites in Uganda, Sudan, Angola, Central African Republic, Republic of Congo and Democratic Republic of Congo where Médecins Sans Frontières (MSF) had conducted HAT programs, and we selected patients with confirmed diagnosis, confirmed stage, complete follow-up (thus confirmed outcome), and meeting a restrictive, laboratory-confirmed definition of relapse, so as to maximize information certainty. Our analysis aimed at identifying early efficacy indicators using the CSF leucocytes count at 6 and 12 months after treatment.
The study received ethical clearance from the Médecins Sans Frontières International Ethical Review Board (Geneva, Switzerland). All data analyzed were anonymized from the start.
Using a large pooled dataset from routine MSF gambiense HAT control programs, we selected patients with confirmed second-stage disease and having received second-stage treatment, who completed their follow-up (minimum 22 months) until confirmation of an outcome (cured or relapsed) and who had a CSF leucocytes count performed at 6 months post-treatment. We considered 22 months as complete follow-up because in practice patients coming for control at 22–23 months are not asked to come again at 24 months.
Second stage was defined by the finding of trypanosomes in blood, lymph nodes or CSF, with ≥20 leucocytes/µL in CSF.
We excluded patients who (i) had missing or incoherent data on key variables, or (ii) died during treatment or follow-up, or (iii) were diagnosed with relapse before 6 months or later than 36 months post treatment, or (iv) for the first analysis only: relapsed at 6 months.
Individuals who relapsed before 6 months were excluded because they do not contribute to the objectives of this analysis, and those relapsing after 36 months because they are less certainly distinguishable from reinfections.
Cure was defined as absence of trypanosomes in all body fluids and < = 20 leucocytes in CSF at ≥22 months post-treatment; and relapse as trypanosomes detected in any body fluid or ≥50 CSF leucocytes/µL anytime [7]. Patients diagnosed with relapse without meeting this definition were excluded. We kept the patients who continued on follow-up despite having ≥50 CSF leucocytes/µL and who had a confirmed outcome later (either cure or relapse).
The strict inclusion criteria aimed at strengthening the validity of the results by focusing on patients that provide unequivocal information, using the advantage of having a large cohort.
We defined tolerance windows for each planned follow-up visit: 6 (5–9); 12 (10–16); 18 (17–21); and 24 (≥22) months [2].
Melarsoprol treatment included the following regimens: one series of 10 daily injections; 2 or 3 series of 3 injections; and 3 series of 4 injections. Eflornithine included series of either 7 or 14 days, all at 400 mg/kg/day divided in 4 infusions per day. Combination treatment included melarsoprol-eflornithine, nifurtimox-eflornithine and melarsoprol-nifurtimox co-administrations.
We used the Wilcoxon test to compare CSF leucocytes between different groups of patients. We plotted the evolution of CSF leucocytes (median, IQR) during the follow-up, overall and by treatment received.
Patients selected were 1822 for the first analysis and 2190 for the second analysis (Figure 1) and had been diagnosed between September 1995 and February 2006. Throughout this time period the same diagnostic tools were used. The largest portion of the cohort was from the centers of Omugo, Northern Uganda (44%) and Ibba, Southern Sudan (21%), as these two sites achieved higher follow-up compliance by investing specific resources. Baseline characteristics are shown in Table 1.
At pre-treatment, the CSF leucocytes count was not different between the 1460 patients who cured (median 137.5 cells, IQR 65–274) and the 362 who later relapsed (132 cells, IQR 53–270) (Wilcoxon test p = 0.15), whereas at 6 months it was significantly higher among patients who later relapsed (29.5 cells, IQR 11–78) than in patients who cured (4 cells, IQR 2–9) (p<0.001). This difference increased at 12 and 18 months, as expected (Figure 2). The difference was observed in all treatment groups, except at 6 and 12 months post-treatment in patients receiving drug combinations, emerging from 18 months onwards.
The evolution of the CSF leucocytes count was similar in naïve (first-time treated) and non-naïve patients throughout the post-therapeutic follow-up (data not shown).
The ROC analysis showed that the absolute CSF leucocytes count at 6 months was at least as good a predictor of outcome (AUC 0.84) as the percent reduction (AUC 0.81). The latter being also the least practical (requiring a bedside calculation involving the initial laboratory results), we did not explore it further. The CSF leucocytes count at 6 months showed the best trade off between sensitivity and specificity at cut-off values of 10 to 13 leucocytes/µL. The best accuracy was obtained with a cut-off at >10 leucocytes/µL which predicted relapse with 76.2% (95%CI, 71.84–80.65%) sensitivity, 80.4% (95%CI, 78.37–82.45%) specificity, 3.89 (95%CI, 3.45–4.38) LR+, 0.29 (95%CI, 0.26–0.32) LR− (Table 2). The positive predictive value was 0.49 (95%CI, 0.45–0.53) and the negative predictive value 0.93 (95%CI, 0.92–0.94).
The multivariate analysis confirmed, after adjustment on treatment, age and sex, that the six-months CSF leucocytes count, with a cut-off at 10 cells, was very strongly associated with the risk of relapse (odds ratio = 17.2, 95%CI, 12.6–23.5).
Table 3 shows the performance of the two-steps algorithms when tested with our large dataset (n = 2190) of selected patients with laboratory-confirmed outcome. In the first line we show the results reported by Mumba et al. [6] on a smaller cohort (“algorithm 5-50-20”).
The same algorithm in our dataset predicted relapse with 87.4% sensitivity (95%CI, 85–90), 97.7% specificity (95%CI, 97–98), LR+ of 37.84 (95%CI, 26.4–54.3) and LR− of 0.13 (95%CI, 0.11–0.16). It wrongly classified as cured (false negatives) 87/1945 patients (4.5%; 95%CI, 3.6–5.5). Two thirds (66.4%; 95%CI, 64.4–68.4%) of the patients followed-up were already classified at 6 months.
The algorithms 5-40-20 and 5-40-15 also performed well, with confidence intervals overlapping the algorithm 5-50-20. The 5-30-15 algorithm was slightly more sensitive but less specific. The proportion of patients classified as cured who later relapsed ranged from 3.5 to 4.5% in all tested algorithms. The portion of the cohort classified at 6 months ranged from 66.4 to 74.1%, leaving the rest to be classified at 12 months. Of the algorithms tested, the 5-50-20 appeared as the best overall with the highest LR+ and a proportion of false negatives not significantly different from the other algorithms (Figure 3).
The CSF leucocytes count at 6 months showed a good prognostic value for final efficacy outcome. However, a small proportion of patients was wrongly classified. Translated into field patient management, those wrongly classified as relapsed would be unnecessarily re-treated, sometimes with toxic drugs (e.g. melarsoprol if first-line treatment was eflornithine or eflornithine-nifurtimox) and, more importantly, patients wrongly classified as cured would be at risk of death due to HAT relapse. A two-step algorithm, at 6 and 12 months, provided a better classification tool featuring an excellent ability to predict relapses with a lower misclassification rate.
At 6 months, the CSF leucocytes cut-off at 10/µL had the best trade-off between positive and negative likelihood ratios. This indicator can rule out relapse at 6 months post-treatment with a good degree of confidence (0.93 negative predictive value), but its ability to identify true relapses is sub-optimal.
Other cut-off values may be of interest for decision-making in the context of clinical trials, e.g. to continue or suspend enrolment of new participants based on 6-months data (patients already enrolled would always benefit of complete follow up).
It is important to underline that the relapse rate was 19.9% in our dataset (first analysis), which is much higher than the relapse rate reported with the new and increasingly used nifurtimox-eflornithine combination therapy (NECT) [10], [11]. Because positive and negative predictive values depend on the relapse rate, the lower the relapse rate is, the lower will be the positive predictive value for relapse, but on the other hand the negative predictive value will be higher, increasing the confidence on the prediction of cure. Reporting the likelihood ratios allows to make abstraction of this phenomenon, since they are independent from relapse rates. Likelihood ratios, both LR+ and LR− are one of the best ways to measure diagnostic accuracy. In medicine, a test is generally regarded as valuable when the LR+ is >5 or the LR− is <0.2 [9], [12].
Because the CSF leucocytes count at 6 months alone remains insufficiently accurate for outcome determination, we evaluated various two-step algorithms at 6 and 12 months, following the model published by Mumba et al [6].
All tested algorithms performed well, but the 5-50-20 algorithm showed the highest specificity (97.7%) and LR+ (37.8). The sensitivity (87.4%), LR− (0.13) and proportion of patients falsely declared as cured (4.5%) were statistically comparable to the other tested algorithms (table 3, figure 3). Two-third of patients (66.4%) could be classified at 6-months post-treatment. Algorithms that would increase this proportion (such as 5-20-20, 5-20-15 or 5-20-10 that classify 74.1% of patients at 6-months) could be particularly interesting in settings with poor follow-up compliance beyond the first visit at 6 months.
Our findings therefore confirm that the diagnostic algorithm 5-50-20 performs well to predict post-treatment outcome, allowing for a shorter follow-up period.
Other algorithms can be applied depending on the setting and priority objectives, e.g. clinical trials or individual patient management in settings with poor follow-up compliance such as in conflict areas.
In all cases, patients who are declared cured early by using these predictors should be encouraged to come for control if symptoms reappear later.
Early determination of outcome presents several key advantages for HAT control programs: first, it cuts down on uncured patients remaining infective until eventually detected or dying; second, it reduces the workload and costs of follow-up; third, it facilitates the monitoring of treatment effectiveness. For some patients it is life-saving or preventive of serious sequelae, for most others it reduces the burden of complying with follow-up schemes. For clinical studies it accelerates acquisition of results and decreases costs.
One major strength of this study was the restrictive selection criteria, which minimized information bias that is typically present in HAT studies: most cohorts include important proportions of patients with uncertain or unknown efficacy outcome, due to the difficulties in completing the patients' follow-up.
Another strength was the large sample size, which increases the precision of the findings.
Finally, the statistical methods used, in particular the analysis by logistic regression with a random intercept controlling the inter-site heterogeneity.
A weakness arose from the nature of the data used, collected by field routine programs, which is generally of lower quality than data collected prospectively within planned studies.
Another weakness arises from the reference used for “true outcome”: a composite definition based on the presence of trypanosomes or a CSF leucocytes count ≥50. The predictors studied are also based on the CSF leucocytes count (at an earlier time) and are therefore not independent from the outcome measurement. In particular when the predictor includes the same value (CSF leucocytes count ≥50, such as in the algorithm 5-50-20) the specificity is to some extent over-estimated.
The marker at the center of our analysis, the CSF leucocytes count, is subject to measurement error, being a manual laboratory technique. However, this particular laboratory exam is regarded as crucial for the patient and it has been the object of great attention in the MSF sites that were included in this study. Internal quality control was implemented in all field laboratories, through blinded double and triple CSF leucocytes counts, showing good levels of consistency in the results (authors' direct field observation, data not published). To our knowledge there are no published works to shed more light into this issue.
The timing of the follow-up assessments was treated via the consolidation of the visit dates into time “tolerance” windows, which are arbitrary groupings (we followed conventional windows) [2]. This interval censoring is an imperfect way of capturing the timing of events: for example what we treat as the “6 months” leucocytes count in reality happened anywhere between 5 and 9 months, with an uneven spread that tends to concentrate after the 6-months date. This field data distribution can be assumed to correspond well with the reality of the routine programs, but it will fit less the temporal distribution in clinical trials that usually have intensive follow-up of patients.
This study provides robust evidence on the value of the CSF leucocytes count to predict, at 6 and 12 months, the efficacy outcome of second-stage T. b. gambiense HAT treatment.
For decision-making on individual patients followed-up in the field, our findings confirm the good performance of the two-steps algorithm using cut-off values of 5-50-20 leucocytes/µL. Other algorithms can be used depending on the setting.
For the early estimation of efficacy in clinical trials, several options are revealed, both in one step at 6 months and in two steps at 6 and 12 months.
|
10.1371/journal.pgen.1006362 | Dbo/Henji Modulates Synaptic dPAK to Gate Glutamate Receptor Abundance and Postsynaptic Response | In response to environmental and physiological changes, the synapse manifests plasticity while simultaneously maintains homeostasis. Here, we analyzed mutant synapses of henji, also known as dbo, at the Drosophila neuromuscular junction (NMJ). In henji mutants, NMJ growth is defective with appearance of satellite boutons. Transmission electron microscopy analysis indicates that the synaptic membrane region is expanded. The postsynaptic density (PSD) houses glutamate receptors GluRIIA and GluRIIB, which have distinct transmission properties. In henji mutants, GluRIIA abundance is upregulated but that of GluRIIB is not. Electrophysiological results also support a GluR compositional shift towards a higher IIA/IIB ratio at henji NMJs. Strikingly, dPAK, a positive regulator for GluRIIA synaptic localization, accumulates at the henji PSD. Reducing the dpak gene dosage suppresses satellite boutons and GluRIIA accumulation at henji NMJs. In addition, dPAK associated with Henji through the Kelch repeats which is the domain essential for Henji localization and function at postsynapses. We propose that Henji acts at postsynapses to restrict both presynaptic bouton growth and postsynaptic GluRIIA abundance by modulating dPAK.
| To meet various developmental or environmental needs, the communication between pre- and postsynapse can be modulated in different aspects. The release of presynaptic vesicles can be regulated at the steps of docking, membrane fusion and endocytosis. Upon receiving neurotransmitter stimuli from presynaptic terminals, postsynaptic cells tune their responses by controlling the abundance of different neurotransmitter receptors at the synaptic membrane. The Drosophila NMJ is a well-defined genetic system to study the function and physiology of synapses. Two types of glutamate receptors (GluRs), IIA and IIB, present at the NMJ, exhibit distinct desensitization kinetics: GluRIIA desensitizes much slower than GluRIIB does, resulting in more ionic influx and larger postsynaptic responses. By altering the ratio of GluRIIA to GluRIIB, muscle cells modulate their responses to presynaptic release efficiently. However, how to regulate this intricate GluRIIA/GluRIIB ratio requires further study. Here, we describe a negative regulation for dPAK, a crucial regulator of GluRIIA localization at the PSD. Henji specifically binds to dPAK near the postsynaptic region and hinders dPAK localization from the PSD. By negatively controlling dPAK levels, synaptic GluRIIA abundance can be restrained within an appropriate range, protecting the synapse from unwanted fluctuations in synaptic strengths or the detriment of excitotoxicity.
| Coordinated action and communication between pre- and postsynapses are essential in maintaining synaptic strength and plasticity. Presynaptic strength or release probability of synaptic vesicles involves layers of regulation including vesicle docking, fusion, and recycling, as well as endocytosis and exocytosis. Also, how postsynapses interpret the signal strength from presynapses depends largely on the abundance of neurotransmitter receptors at the synaptic membrane [1, 2]. During long-term potentiation, lateral diffusion of extrasynaptic α-amino-3-hydroxy-5-methyl-4-isoxazolepropionic acid receptor (AMPAR) to synaptic sites is accelerated [3, 4] and the exocytosis of AMPAR is enhanced near the postsynaptic density (PSD), causing an accumulation of synaptic receptors [5, 6]. In contrast, under the long-term depression condition, synaptic AMPAR is reduced by hastened endocytosis [7, 8]. While molecular mechanisms are proposed to play roles in regulating and fine-tuning postsynaptic glutamate receptor (GluR) abundance in plasticity models, the developmental regulation of GluR abundance at the synaptic surface still needs to be elucidated.
Synapses at the Drosophila neuromuscular junction (NMJ) use glutamate as the neurotransmitter, and have properties reminiscent of mammalian central excitatory synapses [9, 10]. Homologous to vertebrate AMPAR and kainate receptors, Drosophila GluR subunits assemble as tetramers to gate ion influx [11]. Each functional receptor contains essential subunits (GluRIIC, GluRIID and GluRIIE) and either GluRIIA or GluRIIB; therefore, synaptic GluRs can be classified according to their subunit compositions as either A- or B-type receptors [12–17]. These two types of receptors exhibit distinct developmental and functional properties. Newly-formed PSDs tend to accumulate more GluRIIA channels, while the IIA/IIB ratio becomes more balanced when PSDs mature [18]. In addition, GluRIIB channels have much faster desensitization kinetics, which results in smaller quantal size than GluRIIA channels [12]. Therefore, the synaptic composition of these two types of GluRs greatly influences the postsynaptic interpretation of neuronal activities. The Drosophila homolog of p21-activated kinase (dPAK) regulates GluRIIA abundance at the PSD; GluRIIA receptor clusters at the postsynaptic membrane are strongly reduced in dpak mutants [19]. However, overexpression of dPAK in postsynapses is not sufficient to increase GluRIIA cluster size, suggesting that dPAK activity in regulating GluRIIA abundance is tightly controlled.
Ubiquitination and deubiquitination play critical roles in regulating synaptic functions [20–24]. In loss-of-function mutants for highwire, a gene encoding a conserved E3 ubiquitin ligase, NMJs overgrow, producing supernumerary synaptic boutons [25, 26]. This phenotype is duplicated by overexpression of the deubiquitinating enzyme Fat facets (Faf) in presynapses [27]. These studies underline the importance of balanced ubiquitination in synapse formation and function. Cullin-RING ubiquitin ligases (CRLs) are large protein complexes that confer substrate ubiquitination [28, 29]. Importantly, CRLs promote ubiquitination through substrate receptors that provide specific recognition of substrates for ubiquitination. The BTB-Kelch proteins are suggested to be the substrate receptors for Cul3-scaffolded CRLs [30–32].
In this study, we identified a BTB-Kelch-containing protein, Henji, also known as Dbo [33], which regulates NMJ growth and synaptic activity by restricting the clustering of GluRIIA. Synaptic size of henji mutants was significantly expanded, as viewed under transmission electron microscopy (TEM). Immunostaining for dPAK and GluRIIA also suggests larger areas of PSDs in the absence of Henji, and the intensity of each fluorescent punctum becomes stronger, indicating abnormal accumulation of these PSD proteins. By genetically reducing one gene dosage of dpak in henji mutants, GluRIIA accumulation and abnormal bouton morphology was suppressed. In contrast, reducing the gluriia gene dosage in henji mutants restored bouton morphology but failed to suppress dPAK accumulation. Thus, Henji regulates bouton morphology and GluRIIA clustering levels likely through a control of dPAK. Interestingly, while overexpression of dPAK, either constitutively active or dominantly negative, had no effects on GluRIIA clustering, overexpression of these dPAK forms in henji mutants modulated GluRIIA levels, indicating that Henji limits the action of dPAK to regulate GluRIIA synaptic abundance. Henji localized to the subsynaptic reticulum (SSR) surrounding synaptic sites, consistent with the idea that Henji functions as a gatekeeper for synaptic GluRIIA abundance.
We are interested in how ubiquitination might regulate synaptic function through controlling specific synaptic proteins. As putative substrate receptors of Cul3-based E3 ubiquitin ligases, each BTB-Kelch protein could recognize one or multiple synaptic proteins to regulate their abundance and thus their synaptic functions [31]. The Drosophila genome encodes 16 BTB-Kelch proteins (S1 Fig) and those with available RNAi or P-element insertion lines were examined for NMJ morphological abnormality. By immunostaining for horseradish peroxidase (HRP) and Synapsin to reveal NMJ morphology, we found the P-element insertion (PBac{PB}dboc04604) in the CG6224 locus induced satellite boutons (see below). Whereas CG6224 is known to encode Dbo [34], given the supernumerary bouton morphology, we named this mutant “henji”, meaning “very crowded” in Mandarin, and this name is used in this study.
To study Henji function in vivo, we generated mutants by P-element excision in the annotated CG6224/dbo locus (S2A Fig). Two excision mutants with truncation of the shared promoter region between CG6224 and CG6169 were lethal. Addition of the CG6169 genomic transgene rescued the lethality of these deletions, which are named henji1 and henji8 with disruption specifically in henji expression (S2A Fig). Indeed, the henji mRNA expressions analyzed by reverse transcription PCR (RT-PCR) showed lower levels in all three henji mutants used in this study, with a medium level in the PBac{PB}dboc04604 insertion line (named as henjiP in this study), a low level in henji1, and an almost undetectable level in henji8 (S2B Fig). As the protein translation start site was deleted in henji8, we conclude that henji8 is a null allele, henji1 a strong loss-of-function allele, and henjiP a hypomorphic allele.
At wild-type (WT) NMJs, each bouton is connected to adjacent boutons through linear or bifurcated branches. However, at henji NMJs, multiple smaller boutons emerged from large parental boutons (Fig 1B, insets). These smaller boutons, defined as satellite boutons, usually resulted from more than three buds emanating from a parental bouton [35, 36]. Satellite boutons were rarely found at WT NMJs, but were a prominent feature in all henji mutants we had examined, including henji1/1 and henji1/8 (Fig 1B). The number of boutons with normal size, however, was either slightly reduced in henji1/8 or remained normal in henji1/1 (Figs 1C and S2C), suggesting that the formation of satellite boutons is not at the expense of normal boutons.
A bouton houses tens of synaptic sites where the presynaptic active zones (AZs) opposes the postsynaptic PSDs. We first examined PSD structure and PSD-localized GluR clusters in henji mutants by co-immunostaining with antibodies against PSD-specific dPAK and the GluRIIA subunit (Fig 1D). In WT, dPAK localized as well-separated puncta. In the henji1/8 mutant, the area of individual puncta was expanded and the immunofluorescent intensity was enhanced (Fig 1D). When normalized to co-stained HRP, the dPAK immunofluorescent intensity and punctum size were increased while the density of dPAK puncta was normal, as compared to WT control (Fig 1E, upper bar graphs). As dPAK is required for PSD formation and regulates GluRIIA cluster formation [19], the increase in dPAK levels and patch size suggests a possible enlargement of the PSD that houses GluR clusters. Consistently, GluRIIA immunopositive puncta increased in both intensity and size in the henji1/8 mutant (Fig 1D and 1E). The increases in dPAK and GluRIIA immunointensities were also detected in henji1/1 (Fig 2A and 2B). GluRIIB immunostaining signals in henji1/8, however, showed no significant difference in the intensity to WT control (S2D Fig). These results suggest that Henji regulates the dPAK level at PSDs and specifically confines GluRIIA cluster size. At NMJs lacking henji, the increase in GluRIIA but not GluRIIB abundance leads to a shift in the synaptic GluRIIA/GluRIIB ratio.
To examine if henji is responsible for the defects observed in henji mutants, we generated a genomic rescue construct in which GFP was fused to Henji at the N-terminus, and the GFP-henji transgene is driven by the endogenous henji promoter. When introduced into henji1/8, GFP-henji restored dPAK and GluRIIA to near WT levels (Fig 1D). The intensities of dPAK and GluRIIA immunofluorescent signals showed no significant difference to WT controls (Fig 1E). Thus, the lack of henji is the cause for the augmented dPAK and GluRIIA levels at NMJs.
The increased PSD size in henji mutants prompted us to examine the opposing AZ in presynapses. Bruchpilot (Brp), an essential component of the T-bar structure within AZs [37, 38], was expressed in a normal pattern and intensity at the henji1/8 NMJ (S3A Fig). Each presynaptic Brp punctum matched an enlarged dPAK patch in postsynapses, showing a characteristic pattern between pre- and postsynapses. Compared to control, Brp punctum from the henji mutant was unaltered in the intensity, density, and size (S3A Fig, bar graphs). Also, the levels and patterns of the SSR protein Discs large (Dlg), the cell adhesion molecule Fasciculin II (FasII), and the microtubule-associated protein Futsch at henji NMJs were indistinguishable to WT controls (S3B Fig).
The specific alterations in dPAK and GluRIIA expressions at henji mutant NMJs suggest that Henji functions in postsynapses. To determine the functional site of Henji, we performed a rescue experiment with tissue-specific GAL4 drivers to induce UAS-henji expression. As shown, henji1/1 also displayed higher levels of dPAK and GluRIIA in postsynapses (Fig 2A and 2B). When expressed in the henji1/1 postsynapse by muscular C57-GAL4, the intensities of both dPAK and GluRIIA puncta at NMJs were suppressed to WT levels (Fig 2A and 2B). In addition, supernumerary satellite boutons in henji mutants were also suppressed by postsynaptic expression of henji (Fig 2B, bottom panel). In contrast, presynaptic expression of UAS-henji using neuronal elav-GAL4 failed to suppress any of these phenotypes (Fig 2A and 2B). Thus, henji is required in postsynapses to regulate postsynaptic dPAK and GluRIIA abundance and presynaptic bouton growth.
With the requirement for henji in postsynapses, we examined Henji localization at NMJs. We raised antibodies against Henji, which failed to reveal any specific signal in immunostaining. Also, the GFP-henji transgene that was tagged with GFP failed to show any detectable expression level. These results suggest that Henji might be expressed at very low levels. We took advantage of the GFP-henji transgene that also includes a UAS for GAL4-induced expression. When muscular C57-GAL4 was added, GFP-Henji showed localization near the synaptic region (Fig 2C, third row). The postsynaptic-enriched pattern of Henji is specific, as the expression of cytosolic GFP showed diffuse staining in muscle cells without any particular pattern (Fig 2C, second row). Presynaptically expressed GFP-Henji by neuronal GAL4 drivers also displayed diffuse signals in terminal boutons and axonal tracts (Fig 2C, bottom two rows, respectively). To further examine the postsynaptic-enriched expression, co-immunostaining with Dlg was performed. GFP-Henji localized to the SSR and extended slightly outward as compared to Dlg immunostaining (Fig 2D). The postsynaptic localization of Henji suggests a direct mechanism for Henji to regulate the abundance of dPAK and GluRIIA.
Considering the evident GluR compositional shift towards elevated GluRIIA levels in the henji mutants, we addressed whether synaptic transmission is also affected by performing electrophysiological recordings. The amplitude of evoked junctional potential (EJP) did not show any defect (Fig 3A). However, the postsynaptic response to spontaneous neurotransmitter release (quantal size), as assessed by measuring the miniature EJP (mEJP) amplitude, was strongly elevated in the henji1/1 mutant (Fig 3B), consistent with an increase in the GluRIIA/GluRIIB ratio. The frequency of mEJP remained normal (Fig 3C). The quantal content, representing the number of effective synaptic vesicles released upon a nerve stimulus, was calculated as the ratio of EJP/mEJP amplitudes. We found that quantal content values decreased significantly in the henji1/1 mutants as compared to WT control (Fig 3D). Similarly, we also detected similar elevation of mEJP and normal EJP in henji1/P, leading to a reduction of the quantal content, as compared to the henji1/+ heterozygous control (Fig 3F–3H, compare first two bars). To further confirm the reduction of the quantal content, failure analysis was performed, which showed decreased release probability in the henji1/1 mutant (Fig 3E). These data suggest that GluRIIA accumulates in the henji mutant, causing an elevation in postsynaptic responses. However, homeostatic mechanisms might tune down presynaptic release to reduce the quantal content, thereby maintaining a normal EJP output.
We then tested whether the enhanced mEJP amplitude in the henji mutant is caused by the absence of Henji in postsynapses. Muscle expression of henji suppressed the mEJP amplitude, both in the henji1/P mutant and WT background, whereas neuronal expression did not (Fig 3G). As Henji plays a role in suppressing the GluRIIA level in postsynapses (Fig 2A and 2B), the elevation of the GluRIIA level is consistent with the enhancement of mEJP in the henji mutant.
The increase of dPAK and GluRIIA patches may be associated with an expansion of the synaptic size in henji mutants. To examine this possibility, ultrastructures of boutons were analyzed by TEM. Cross-sections of boutons showed electron-dense membrane regions, representing the matching sites between presynaptic AZ and postsynaptic PSD (Fig 4A, within two arrows). In presynapses, synaptic vesicle-docked T-bars located within AZs, while in postsynapses, membranous SSR enwraps the bouton. We found that the electron-dense membrane region was expanded in henji mutants (enlarged panels in Fig 4A). Quantification indicated that the length of the electron-dense membrane region significantly increased in the henji1/1 and henjiP/P mutants (Fig 4B). Moreover, as the bouton perimeter did not differ significantly between henji mutants and WT control (S1 Table), the synaptic membrane accounted for a larger proportion of the total membrane region in the lack of henji (Fig 4B, right panel). These analyses indicate that the synaptic membrane region is expanded in the henji mutants.
Given the elevation of synaptic dPAK levels in the henji mutant, we tested whether reducing the dpak gene dosage could have an effect on henji mutant phenotypes. When the dpak6 null allele was introduced into the henji1/8 mutant background, GluRIIA abundance was suppressed (Fig 5A). Similarly, both kinase-dead dpak3 and Dock-interaction-disrupted dpak4 alleles [39] also suppressed GluRIIA abundance in the henji1/8 mutant, suggesting that these functional domains are critical for dPAK to regulate GluRIIA abundance (Fig 5A). Satellite boutons in the henji1/8 mutant were also suppressed by dpak6 and, to a lesser extent, dpak3 and dpak4 (Fig 5B). In removing one copy of the dpak6 null allele in the henji1/8 mutant, dPAK was indeed reduced to near the WT level (S4 Fig), consistent with the idea that the reduction of the dPAK level is able to suppress henji phenotypes. These genetic suppressions suggest that the upregulated dPAK level contributes to GluRIIA accumulation and abnormal bouton morphology in the henji mutant.
It has been shown that dPAK regulates synaptic GluRIIA abundance; in the dpak mutant, the GluRIIA level was reduced at the PSD [19]. We then examined the epistatic relationship between henji and dpak mutants. In dapk3/6 larvae that survived to late larval stages, the GluRIIA level was greatly reduced (Fig 5C), consistent with the previous report [19]. In contrast, the GluRIIA level was enhanced in the henji1/1 mutant (Fig 2A). In the henji1/1 dapk3/6 double mutant, GluRIIA was reduced to the level similar to that of the dapk3/6 single mutant (Fig 5C). Thus, in the absence of dpak, the GluRIIA level fails to be upregulated in the henji mutant, suggesting that dpak functions downstream of or in parallel to henji.
With the upregulated GluRIIA level in the henji mutant, we also examined any suppression effect by reducing GluRIIA in a henji mutant background. Introducing one copy of a gluriia deletion allele in the henji mutant strongly suppressed the satellite bouton phenotype (Fig 5D, arrowheads). However, dPAK abundance was not suppressed by reducing a gluriia gene dosage (Fig 5E). Taken together, these data support a model whereby Henji restricts postsynaptic GluRIIA abundance by downregulating the dPAK levels. In henji mutants, accumulated GluRIIA induces abnormal bouton growth in the retrograde direction, resulting in satellite bouton morphology.
To further understand the role of Henji on postsynaptic regulation, we generated N-terminal GFP-tagged deletion constructs ΔBTB, ΔBACK and ΔKelch that truncate one of the three conserved domains of Henji (Fig 1A). These Henji truncations and full-length control were expressed in postsynapses for rescuing henji1/8 mutant phenotypes. As expected, full-length Henji when expressed in postsynapses suppressed the elevated dPAK and GluRIIA in henji1/8 while ΔKelch failed to do so, suggesting that the Kelch repeats region is essential for Henji function in postsynapses (Fig 6A). Surprisingly, ΔBTB and ΔBACK significantly suppressed elevated intensities of both dPAK and GluRIIA in the henji1/8 mutant. These results suggest that both BTB and BACK domains are dispensable for Henji to function in postsynapses. To further address functional domains of Henji in postsynapses, full-length and truncations of Henji were overexpressed in postsynapses, and synaptic abundances of dPAK and GluRIIA were assessed. Significantly, full-length Henji when overexpressed in postsynapses caused reductions in dPAK and GluRIIA levels at NMJs, suggesting that Henji is sufficient to promote downregulation of dPAK and GluRIIA levels (Fig 6B). Unlike the full-length Henji, truncating any of the three domains failed to downregulate dPAK and GluRIIA levels. Instead, ΔBTB produced a dominant-negative effect by inducing dPAK and GluRIIA accumulations while ΔBACK and ΔKelch had no effects (Fig 6B). Therefore, the Kelch repeats seems to be the most critical domain of Henji to regulate the postsynaptic abundance of dPAK and GluRIIA.
We then examined the domain requirement for Henji in postsynaptic localization. Full-length and truncated Henji transgenes were expressed in muscles of the henji1/8 mutant, and the protein localization was detected by GFP immunostaining. By co-staining with Dlg, full-length Henji localized to postsynaptic SSR (Fig 7A). Similarly, absence of the BACK domain (ΔBACK) still retained proper synaptic localization of Henji. Lacking the BTB domain also retained some localization signals at the postsynapses. Finally, lacking the Kelch repeats (ΔKelch) completely abolished Henji postsynaptic localization (Fig 7A). These analyses suggest that the Kelch repeats region are essential for proper postsynaptic localization of Henji.
Considering the role of Kelch repeats in Henji postsynaptic function and localization, we tested whether Kelch repeats and dPAK interact physically. Indeed, both Flag-tagged full-length and Kelch repeats of Henji co-immunoprecipitated with Myc-tagged dPAK, providing evidence of physical interactions between Henji and dPAK (Fig 7B, indicated by arrows). Taken together, these results indicate that the Kelch repeats that bind to dPAK are required for Henji localization and function to control dPAK and GluRIIA postsynaptic abundances.
While dPAK is required for postsynaptic localization of GluRIIA, how dPAK functions to regulate GluRIIA abundance still remains unknown [19]. A constitutive-active form of dPAK that is membrane-tethered failed to increase GluRIIA abundance at PSDs, hinting another layer of regulation on dPAK. We tested whether Henji plays the critical role on limiting dPAK activity in GluRIIA regulation. To investigate this possibility, we generated Myc-tagged dPAK of WT, constitutive-active (CA), or dominant-negative (DN). The CA form contained a phospho-mimic T583E point mutation at the first autophosphorylation residue [40–42]. The DN form contained three point mutations, with H91L and H94L disrupting binding and activation by Cdc42 and Rac, and K459R eliminating kinase activity. Western blot analysis showed that, when driven by Tub-GAL4, WT, CA, and DN were expressed at similar levels (S5 Fig). All three transgenes were overexpressed postsynaptically and GluRIIA abundance was quantified. We found that, as previously reported [19], overexpression of dPAK failed to alter GluRIIA abundance regardless of the activation status (Fig 8A).
To examine whether Henji confers the extra layer of regulation on dPAK activation, dPAK transgenes were overexpressed postsynaptically in the henji1/1 mutant background. Overexpression of dPAK WT did not alter the already enhanced GluRIIA level in the henji1/1 mutant, which may reflect a constraint in dPAK activation such as the requirement of CDC42/Rac1 in dPAK activation [40–42]. Interestingly, overexpression of the CA form further enhanced GluRIIA synaptic intensity in the henji1/1 mutants (Fig 8B), a phenotype that was not detected when dPAK CA was overexpressed in the henji+/+ background (Fig 8A). Also, dPAK DN suppressed GluRIIA intensity in the henji1/1 mutant, a phenotype that was not detected in the henji+/+ background, either (Fig 8A and 8B). Taken together, we propose that Henji limits the action of dPAK to regulate GluRIIA abundance by adding another layer of regulation to conventional phosphorylation-mediated dPAK activation.
Finally, we examined the localization of dPAK by Myc immunostaining. Overexpressed dPAK proteins of WT, CA and DN forms showed dispersed weak puncta in muscle cells without forming specific patterns (Fig 8C). When overexpressed in the absence of henji activity, Myc-positive puncta became brighter and accumulated around the synaptic region. Some of the puncta were co-labeled with GluRIIB, representing specific accumulation at PSDs (Fig 8D). Taken together, these analyses suggest that Henji functions to limit dPAK from localization at postsynaptic sites, which is important for modulating GluRIIA abundance.
Here, we show that Henji functions at the postsynapse to regulate synaptic development and function at the NMJ. The PSD area is expanded and GluRIIA clusters abnormally accumulate at the PSD. We provide genetic evidences to support that the elevation of GluRIIA synaptic abundance is at least partially caused by a corresponding accumulation of dPAK in henji mutants. We also show that Henji is sufficient to downregulate dPAK and GluRIIA levels and the Kelch repeats of Henji play the most critical role in this process. Henji tightly gates dPAK in regulating GluRIIA abundance, as dPAK enhances GluRIIA cluster abundance only when Henji is absent. Therefore, we have identified a specific negative regulation of dPAK at the postsynaptic sites that contributes to the PSD formation and GluR cluster formation at the NMJ (Fig 9).
PAK proteins transduce various signaling activities to impinge on cytoskeleton dynamics. Through kinase activity-dependent and -independent mechanisms, PAK regulates not only actin- and microtubule-based cytoskeletal rearrangement but also the activity of motors acting on these cytoskeletal tracks [43–46]. In mammalian systems, PAKs participate in many synaptic events including dendrite morphogenesis [47, 48], neurotransmitter receptor trafficking [49, 50], synaptic strength modulation [51], and activity-dependent plasticity [52]. Pathologically, PAK dysregulation also contributes to serious neurodegenerative diseases [53], such as Alzheimer disease [54, 55], Huntington’s disease [56, 57] and X-linked mental retardation [58–60].
At Drosophila NMJs, dPAK has divergent functions; loss of dpak causes a dramatic reduction in both Dlg and GluRIIA synaptic abundance [19], but the underlying molecular mechanisms have not been revealed. Our data show that Henji functions to restrict GluRIIA clustering but has no effect on Dlg levels (S3B Fig), suggesting that Henji regulates one aspect of dPAK activities, probably via the SH2/SH3 adaptor protein Dock [19]. Alternatively, Henji may function to limit dPAK protein levels locally near the postsynaptic region, rendering its influence on GluRIIA clustering, while dPAK that regulates Dlg may localize outside of the Henji-enriched region. Supporting this idea, Henji is specifically enriched around the SSR region instead of dispersed throughout the muscle cytosol (Fig 2D). Moreover, ectopic Myc-dPAK localized at the postsynapse only when henji was mutated (Fig 8C and 8D), indicating that Henji regulates dPAK postsynaptic localization.
The interaction with Rac, Cdc42, or both triggers autophosphorylation and subsequent conformational changes of PAK, resulting in kinase activation. The myristoylated dPAK that has been shown to be active in growth cones [39] failed to enhance GluRIIA abundance at the NMJ [19]. This result shows that dPAK is necessary to regulate GluRIIA synaptic abundance, but is itself tightly regulated at the synaptic protein level or the kinase activity. Indeed, we provide evidence to show specific negative regulation of dPAK by Henji; overexpression of dPAK CA that could not enhance GluRIIA abundance in WT larvae further increased the already enhanced GluRIIA levels in the henji mutant (Fig 8A and 8B). Similar to the CA form, the DN form also showed no effect on GluRIIA when simply overexpressed in the WT background, but exhibited strong suppression of GluRIIA in the henji mutant background (Fig 8A and 8B). Thus, regardless of the possible conformational differences between the CA and DN forms, Henji appears to confer a constitutive negative regulation of dPAK at postsynapses, suggesting a tight control that could be at subcellular localization. In contrast to CA and DN forms, activation of dPAK requires binding to Rac1 and Cdc42, and subsequent protein phosphorylation. This additional layer of regulation may serve as a limiting factor rendering dPAK WT from recruiting GluRIIA to PSDs regardless in WT or henji mutant background.
The structural feature suggests that Henji could function as a conventional substrate receptor of the Cul3-based E3 ligase complex. At Drosophila wing discs, Dbo functions as a Cul3-based E3 ligase to promote Dishevelled (Dsh) downregulation [61]. Similar to the henji alleles, we confirmed that the dbo [Δ25.1] allele and dbo RNAi were competent to induce dPAK and GluRIIA accumulation at the postsynapse (S6A Fig). In our immunoprecipitation experiment, we detected Henji and dPAK in the same complex (Fig 7B, lane 3), and dPAK also forms a complex with the C-terminal substrate-binding Kelch-repeats region (lane 4). However, we did not detect any notable or consistent increase in Henji-dependent dPAK poly-ubiquitination in both S2 cells and larval extracts. Also, the Cul3-binding BTB domain of Henji seems dispensable in the suppression of dPAK levels in henji mutants (Fig 6A). Importantly, Cul3 knockdown in muscle cells failed to cause any accumulation of GluRIIA and dPAK at the NMJ (S6A Fig). Sensitive genetic interaction between henji and Cul3 failed to induced dPAK and GluRIIA accumulation (S6B Fig). Dbo functions together with another BTB-Kelch protein Kelch (Kel) to downregulate Dsh [61]. However, Kel negatively regulates GluRIIA levels without affecting dPAK localization at the postsynaptic site (S6C Fig). This data argues that Kel functions in a distinct pathway to Henji in postsynaptic regulation of GluRIIA. Taken together, we found no direct evidence to support that dPAK is downregulated by Henji through ubiquitination-dependent degradation. Alternately, Henji could bind dPAK near the postsynaptic region and this interaction may block the recruitment or localization of dPAK onto postsynaptic sites (Fig 9). Under this model, dPAK is less restricted and has a higher propensity to localize at postsynaptic sites in the absence of Henji, resulting in synaptic accumulation of dPAK and GluRIIA expansions.
As many synaptic events require rapid responses, local regulation of protein levels becomes crucial in synapses. To achieve accurate modulation, certain synaptic proteins should be selectively controlled under different developmental or environmental contexts. Indeed, emerging evidence shows that various aspects of synapse formation and function are under the control of the ubiquitin proteasome system (UPS), including synapse formation [62, 63], morphogenesis [64], synaptic pruning and elimination [65, 66], neurotransmission [67–69], and activity-dependent plasticity [21, 70]. In particular, the membrane abundance of postsynaptic GluR that modulates synaptic function can be regulated by components of the UPS. When Apc2, the gene encoding Drosophila APC/C E3 ligase, is mutated, GluRIIA shows excess accumulation but the molecular mechanism was not elucidated [71]. Similarly, loss of the substrate adaptor BTB-Kelch protein KEL-8 in C. elegans also results in the stabilization of GLR-1-ubiquitin conjugates [72]. However, no evidence shows direct ubiquitination and degradation of GLR-1 by KEL-8. Also, absence of the LIN-23-APC/C complex in C. elegans affects GLR-1 abundance at postsynaptic sites without altering the level of ubiquitinated GLR-1. Therefore, GLR-1 receptor endocytosis and recycling or ubiquitination and degradation of GLR-1-associated scaffold proteins are proposed to be the underlying mechanism for E3 ligase regulation [23, 67]. In mammals, endocytosis of AMPAR can be influenced by poly-ubiquitination and degradation of the prominent postsynaptic scaffold protein PSD-95 [21, 73].
In this study, we describe a novel regulation by the BTB-Kelch protein Henji on synaptic GluRIIA levels. By limiting GluRIIA synaptic levels, Henji modulates the postsynaptic output in response to presynaptic glutamate release. In the absence of Henji, quantal size is elevated (Fig 3B and 3G), coinciding with an increase in the postsynaptic GluRIIA/GluRIIB ratio (Figs 1D and 1E and S2D). In a previous study, increases in the GluRIIA/GluRIIB ratio by overexpressing a GluRIIA transgene in the muscle or by reducing the gene copy of gluriib promote NMJ growth, but co-expression of both GluRIIA and GluRIIB did not alter the bouton number [74]. Combined with our findings, those data provide a link between an increased GluRIIA-mediated postsynaptic response and bouton addition at NMJs. However, satellite boutons were not detected following GluRIIA overexpression [74]. One possibility is that satellite boutons are considered as immature boutons [36, 75] and their appearance may indicate the tendency for NMJ expansion, as in the case of excess BMP signaling [75]. Failure to become mature boutons may be caused by the lack of cooperation with other factors such as components of the presynaptic endocytic pathway [76], actin cytoskeleton rearrangement [77–79] or neuronal activity [36]. We found no significant alterations in endocytosis and the BMP pathway in the henji mutant (S7 Fig). Nevertheless, we cannot rule out that Henji may modulate other presynaptic events that are defective in henji mutants to interfere with bouton maturation (S1 Table).
w1118 was used as the WT control and to backcross all henji alleles described in this study. Flies of all genotypes were reared at 25°C for experiments. We performed P-element-mediated imprecise excision to generate henji1 and henji8 alleles (S2A Fig). Plasmids of UAS-Flag-henji (full-length), UAS-Flag-Kelch (533–623 a.a.), UAS-GFP-henji (full-length), UAS-GFP-ΔBTB (delete 1–167 a.a.), UAS-GFP-ΔBACK (delete 168–276 a.a.), UAS-GFP-ΔKelch (delete 305–623 a.a.), UAS-Myc-dpak WT (full-length), UAS-Myc-dpak CA (point mutation T583E) and UAS-Myc-dpak DN (triple mutations H91L, H94L and K459R) were constructed using the Gateway System into the pUAST vector (Invitrogen and Drosophila Genomics Resource Center, DGRC). The genomic rescue transgene GFP-henji was constructed by fusing GFP to the ATG codon of henji cDNA and driven by the putative promoter region containing the genomic sequence between CG6169 ATG and henji ATG. C57-GAL4, elav-GAL4, kelDE1, and dpak6 were from Bloomington Drosophila Stock Center (BDSC). RNAi lines for Cul3 (109415) and gbb (5562R) [80] were from Vienna Drosophila RNAi Center (VDRC) and National Institute of Genetics (NIG), respectively. Mutant strains that have been described are dpak3 and dpak4 [39], dbo[Δ25.1] and dbo RNAi [61], and GluRIIA-GFP rescuing gluriia and gluriib double mutants [81].
Primary antibodies used were: mouse anti-Dlg (4F3, 1:100, Developmental Studies Hybridoma Bank, DSHB), mouse anti-GluRIIA (1:100, DSHB), rabbit anti-dPAK (1:1000)[82], rabbit anti-GluRIIB (1:1000)[14], mouse anti-Brp (1:100, DHSB), mouse anti-FasII (1:100, DSHB), mouse anti-Futsch (1:100, DSHB), chicken anti-GFP (1:100; Abcam Co.), mouse anti-Myc (9E10, 1:100, Santa Cruz Co.), rabbit anti-pMAD (1:250)[83] and rabbit or goat anti-HRP conjugated FITC, TRITC and Cy5 (Jackson ImmunoResearch Laboratories). Secondary antibodies used were anti-rabbit or -mouse Cy3 and Cy5 (Jackson ImmunoResearch Laboratories). Muscles, though not shown in figures, were revealed by staining with FITC-conjugated phalloidin (1:1000; Sigma Co.).
A2 to A6 segments of NMJ4s and A3 segments of NMJ6/7s of wandering third instar larvae were analyzed. Larvae were dissected in cold calcium-free HL3 saline (70 mM NaCl, 5 mM KCl, 20 mM MgCl2, 10 mM NaHCO3, 5 mM trehalose, 115 mM sucrose, and 5 mM HEPES, pH 7.2) and larval fillets were fixed in 4% paraformaldehyde for 20 min and washed in PBT (0.03% triton-X-100) for 10 min three times. For GluRIIA and GluRIIB staining, larval fillets were fixed in Bouin’s fixative (Sigma Co.) for 5 min. Fixed fillets were incubated with primary antibodies overnight at 4°C, washed in PBT three times, and incubated with secondary antibodies for 2 hr at room temperature. Larval fillets were mounted in solution containing PBS with 87.5% glycerol and 0.22 M 1, 4-diaza-byciclo (2.2.2) octane (Dabco, Sigma Co.). Images were acquired via LSM 510 confocal microscopy (Carl Zeiss) using 40x water and 100x oil objectives. Images were processed by LSM5 image examiner (Carl Zeiss) and Adobe Photoshop Creative Suite, and further quantified by Image J for immunofluorescence intensities, punctum densities and cluster sizes. For quantification, Z-section images were projected for further processing. Immuno-positive regions were defined by using Image J in which threshold setting was used to eliminate background noise. Intensity of each synaptic protein was normalized to corresponding HRP intensity. Satellite bouton numbers were normalized to corresponding muscle areas. For image presentation, immunostaining images presented in figures represent single sections except the Futsch images in S3B Fig were from projection of Z sections. Unpaired Student t-test is used in calculating statistical significance.
Drosophila S2 cells were maintained in Schneider's medium (Thermo Fisher Scientific) at 25°C. S2 cells (5 x 106 cells in each 10 cm dish) were transfected with 1 μg DNA of individual constructs using Cellfectin (Invitrogen). S2 cells were collected and homogenized in RIPA lysis buffer (20 mM Tris-HCl, pH8.0, 150 mM NaCl, 5 mM EDTA, 1% Triton-X-100, 2 mM Na3VO4, 50 mM NaF and 1 mM PMSF, supplemented with protease inhibitor cocktail (Roche). Protein concentrations were calculated with the aid of protein assay (Bio-Rad Laboratories). For immunoprecipitation, Myc-tagged dPAK was co-transfected with either Flag-tagged Henji or the Kelch repeats domain of Henji into S2 cells. Cell lysates were incubated with beads coated with anti-Myc (9E10, Santa Cruz) and the immunoprecipitates were blotted with mouse anti-Flag antibody (1:1000, Sigma Co.). Antibodies used for immunoblotting were anti-dPAK (1:5000), anti-Myc (1:1000), and anti-α-tubulin (1:200000, Sigma Co.).
Larval fillets were dissected in cold calcium-free HL3 saline and subsequently fixed overnight in modified Trump’s universal fixative (4% paraformaldehyde, 1% glutaraldehyde in 0.2 M cacodylate buffer, pH 7.2). The PELCO BioWave® laboratory microwave system was used for subsequent steps. Samples were post-fixed with 1% aqueous osmium tetroxide in 0.2 M cacodylate buffer (pH 7.2) under 20 inHg vacuum. After stained with 2% uranyl acetate for 30 min at room temperature, samples were dehydrated in gradually-increasing ethanol concentrations (50%, 70%, 80% and 90%). Later, fillets were infiltrated in Spurr’s resin with gradual increases of concentrations (25%, 50%, 75% and 100%). Ultrathin sections, obtained by ultramicrotome (Leica), were further stained with uranyl acetate and lead citrate. Images were viewed by Tecnai G2 Spirit TWIN (FEI Company, Hillsboro, OR). Electron-dense region was determined by Gatan DigitalMicrograph. A line was drawn along the bouton membrane and spots with lowest intensity were picked and labeled as the boundary of electron-dense region. Bouton parameters were quantified by Image J.
For sample preparation, larvae were dissected with the segmental nerves cut close to the ventral ganglion region in cold modified calcium-free HL3.1 saline (70 mM NaCl, 5 mM KCl, 10 mM MgCl2, 10 mM NaHCO3, 5 mM trehalose, 115 mM sucrose, 5 mM HEPES, pH 7.2). Samples were then incubated in modified HL3.1 saline containing 0.8 mM CaCl2 for stimulation, and recordings were taken at room temperature. The two electrodes for voltage-clamping were filled with 3 M KCl and impaled in muscle 6 of the A3 segment. One microelectrode (15~20 MΩ) monitored the muscle membrane potential while the other (5~8 MΩ) delivered electric currents. 5-8V stimulation was given to stimulate the nerve. The muscle membrane potential was clamped at -60 mV. Without any stimulation on the segmental nerves, mEJPs within 100 sec were recorded. For evoking an EJP, the segmental nerve was stimulated by a suction electrode every 30 sec with pulse duration of 0.1 msec at the voltage two times that of the threshold. For failure analysis EJP is evoked in 0.2 mM [Ca2+], the failure rate was calculated by ln(n/N), with n the number of failure events, and N the total number of stimuli [16]. For high-frequency stimulation, the segmental nerve was stimulated at 13.3 Hz for 4 min in 2 mM [Ca2+] buffer. Data were digitized by a DigiData 1440 interface (Molecular Devices) at 50 kHz, and weak signals were filtered at 10 kHz, and analyzed by Clampfit10 (Molecular Devices).
|
10.1371/journal.pcbi.1005812 | Identification of immune signatures predictive of clinical protection from malaria | Antibodies are thought to play an essential role in naturally acquired immunity to malaria. Prospective cohort studies have frequently shown how continuous exposure to the malaria parasite Plasmodium falciparum cause an accumulation of specific responses against various antigens that correlate with a decreased risk of clinical malaria episodes. However, small effect sizes and the often polymorphic nature of immunogenic parasite proteins make the robust identification of the true targets of protective immunity ambiguous. Furthermore, the degree of individual-level protection conferred by elevated responses to these antigens has not yet been explored. Here we applied a machine learning approach to identify immune signatures predictive of individual-level protection against clinical disease. We find that commonly assumed immune correlates are poor predictors of clinical protection in children. On the other hand, antibody profiles predictive of an individual’s malaria protective status can be found in data comprising responses to a large set of diverse parasite proteins. We show that this pattern emerges only after years of continuous exposure to the malaria parasite, whereas susceptibility to clinical episodes in young hosts (< 10 years) cannot be ascertained by measured antibody responses alone.
| Understanding naturally acquired immunity against P. falciparum malaria is of fundamental importance for malaria control and elimination efforts. The identification of parasite antigens that could potentially be considered as vaccine targets often relies on prospective cohort studies where observed infection rates are related to measured immune responses. However, what is unknown, is how these population-level associations between antibody titres and protection from severe malaria can predict the risk of an infection for an individual. We therefore analysed three sets of cohort-based immune profiles using a machine learning approach in order to identify distinct immune signatures that are predictive of protection at the individual-level. Our results show that even statistically significantly associated responses fail to provide robust information about an individual’s risk of malaria and that machine learning approaches should be considered more prominently alongside traditional methods for analysing these complex and high dimensional datasets.
| Naturally acquired immunity to malaria is a complex and poorly understood process, by which individuals living in P. falciparum endemic areas develop protection against clinical and symptomatic infections over years of repeated exposure. Since the first experimental evidence demonstrating how passively transferred immunoglobulins from immune adults can dramatically reduce parasitaemia in infected recipients [1, 2] there has been a growing body of evidence that antibody (Ab) responses play an important role for parasite control and protective immunity. However, the unambiguous identification of the target antigens involved has been difficult, and even after decades of research there is still no strong consensus about which candidates could be considered as potential components of an anti-asexual stage vaccine.
Prospective cohort studies, in which individuals’ immune responses against panels of P. falciparum-specific antigens at time zero are related to their subsequent risk of developing clinical malaria, have frequently shown how responses to various antigens correlate with increased protection against clinical malaria in an age- and/or exposure dependent manner [3–14]. Proteins expressed by the merozoite life-stage of P. falciparum, such as the merozoite surface protein (MSP) or apical membrane protein (AMA), are often the focus of such studies, partially due to their higher sequence conservation compared to other immunogenic but highly polymorphic variant surface proteins (e.g. PfEMP1) that are expressed during the intra-erythrocytic life-stages of the parasite. The protective potentials of anti-merozoite antibodies have been confirmed in in vitro and animal studies, which led to those antigens now being considered as potential vaccine targets (see e.g. [15] for a review). However, their contribution to clinical immunity in a field-setting is yet to be quantified.
Small effect sizes and the difficulty in reliably quantifying previous exposure [16] makes the distinction between markers of exposure and markers of protective immunity problematic and has resulted in inconsistent and contradictory findings in the past [17]. More importantly, though, routine analytical approaches based on comparisons between population-level mean responses often fail to convey information about the robustness of the derived associations and how sensitive they are to even small changes in the observed data. The shortcomings of traditional statistical methods are highlighted when trying to predict individual-level protection from population-wide associations. In particular, when dealing with high dimensional data, where a vast number of combinations and interactions must be tested. Here, practitioners typically rely on univariate tests, whilst adjusting for common markers of exposure, thus ignoring potential interplay between different antigens.
Conversely, predictive modelling frameworks based on machine learning offer a systematic way to consider all possible combinations of immune responses against various antigens. These hypothesis-free approaches do not assume a priori functional relationships between the measured variables (e.g. Ab-levels) and the response (e.g. the risk of a clinical episode), and test whether these associations could be due to chance (i.e. the ubiquitous P-value). Instead, the outcome of interest is the predictive accuracy, i.e. the degree by which the model can predict the response at the level of the individual. They further provide a better understanding of the contribution of individual predictors towards model performance. Thus, machine learning techniques have become popular choices for the analysis of high dimensional datasets in biology and ecology (see e.g. [18–23]).
Here we used a random forests machine learning approach to analyse antibody profiles against panels of P. falciparum-specific antigens with the aim to identify signatures that are predictive of an individual’s protective status against clinical malaria episodes. Our results show that immune signatures that clearly distinguish clinically immune individuals can be found only when considering a broad set of antigens from individuals spanning a sufficiently wide age-range, whereas the responses taken from young cohorts are less likely to be informative of the individual’s susceptibility to malaria.
We analysed previously published data from three prospective cohort studies conducted in Kenya, Kenya/Tanzania and Mali, which we simply refer to here as KEN, KTZ and MAL, respectively. The datasets can be found as supporting information (S1 Data, S2 Data and S3 Data). The underlying studies are described in detail elsewhere [8, 12, 24] and summarised in Table 1, so here we just provide a brief overview.
The KEN dataset contains immune profiles for 286 individuals. However, in line with the original study [12], our analysis was performed on the subset of children who were parasite-positive at screening (N = 121, age = 1-10 years). Immune profiles are ELISA-based antibody titres against 36 P. falciparum-specific antigens, taken at the start of the transmission season, with host age and schizont extract reactivity used as exposure proxies. The response variable was incidence of a clinical malaria episode, defined as an axillary temperature of > 37.5°C, plus any parasitaemia for children less than 1 year, and an axillary temperature of > 37.5°C, plus parasitaemia > 2500/μl for individuals older than 1 year, during a 6-months follow-up.
The KTZ dataset is based on luminex-derived IgG levels against 46 individual PfEMP1 domains of 447 children (5-18 months old, mean = 11.4 months) living in Kilifi (Kenya) and Korogwe (Tanzania), taken from the placebo arm of the RTS,S malaria vaccine trial [25]. Individuals were followed for an average of 8 months with multiple samples taken over the time course, resulting in a total of 1269 immune profiles. The outcome of interest was the incidence of at least one clinical malaria episode, defined as an axillary temperature of > 37.5°C plus parasitaemia > 2500/μl. Age and bednet use were used as proxy variables for exposure risk.
The MAL dataset comprises protein microarray-based antibody reactivity of 186 individuals aged 2-25 years against a panel of 2320 P. falciparum-specific epitopes of the 3D7 line, representing 1204 unique proteins (∼ 23% of the P. falciparum proteome), taken before the start of the transmission season. The response variable was incidence of clinical malaria, defined as axillary temperature of > 37.5°C plus parasitaemia > 5000/μl, over an 8-months period of follow-up. Age and infection status (parasite positive or negative) at first screening were used as exposure proxy variables.
To identify predictive immune signatures underlying clinical protection we employed a random forests [26] machine learning approach, using the randomForest package [27] in R [28]. Each dataset (KEN, KTZ, MAL) was analysed separately. As input variables for our models we used the measured immune profiles and, where applicable, their respective exposure proxies. The response variable, i.e. the outcome to be predicted, was the incidence of clinical infections during follow-up as defined by the respective studies. For this work we classified individuals as susceptible if they had a recorded episode of clinical malaria within the specified time window, and protected otherwise. Unavoidably, protected individuals also included those who may have been uninfected. However, this scenario is probably minimised by the fact that these studies were conducted in villages of moderate to high transmission.
Random forests are able to deal with data sets where the number of predictors is larger than the sample size. However, when the number of features greatly outweighs the number of samples, such as in the MAL dataset (sample size N = 186, number of features M = 2320), feature selection is advised to remove uninformative variables and focus on the ones that exhibit sufficient predictive power [29, 30]. We start by first ignoring strongly linearly correlated responses (above a Pearson correlation coefficient of ρ = 0.8; which corresponds to around 36% of the original feature set), to avoid biasing the variable importance measures computed by the random forests [31]. Recall that only around half of the measured responses represent unique proteins (see Table 1). Note that these correlated variables are reintroduced in the interpretation stage if any of the features they are associated with have been selected for the final model. The remaining covariates undergo a rigorous supervised feature selection process, based on the mProbes [32] and xRF [33] algorithms, as follows:
A detailed layout of the random forests model fitting procedure for the MAL dataset can be found in S1 Text; S1 Script provides the R code to run our feature selection procedure.
We used two measures to assess and report on the models’ predictive accuracies: (i) the receiver operating characteristic (ROC) curve, which is generated by plotting the true positive rate against the false positive rate (i.e. the observed incidence against the false predicted incidence) at various threshold settings. The area under the curve (AUC) is a measure of predictive accuracy, with an AUC = 1 equating to zero error and an AUC = 0.5 equating to random guessing; and (ii) by means of a confusion matrix, which contrasts the instances of the predicted classes (protected or susceptible) against the observed classes. The misclassification-rates are based on the so-called out-of-bag (OOB) errors [27]. These are derived by iteratively testing the model’s performance against subsets of data left out during the fitting process (recall that each decision tree in a random forests is built on a bootstrapped sample of the original data).
OOB errors represent an estimate of the generalisation error, that is, how well the model would fare against previously unseen data. For the MAL dataset, the OOB error computed on a model fitted only on the selected features would be over-optimistic due to selection bias [34]. Instead, we use our feature selection algorithm inside a five-fold cross-validation loop. Within each fold, the model fitted using the selected features (for each fold the number of selected features may vary) is tested against the left out fold. We report the average AUC across all folds.
We first analysed two datasets comprising ELISA and Luminex-derived antibody profiles against P. falciparum—specific antigens obtained from prospective cohort studies in Kenya (KEN) and Kenya / Tanzania (KTZ) (see Material and methods). For each dataset, we used a random forests (RF) machine learning approach to predict individual-level protection against clinical immunity over a specified period of time based on (i) measured antibody levels (Ab), (ii) proxies for exposure (such as age or bednet use) (Exp), and (iii) all measured variables (Ab and Exp).
Fig 1 illustrates the outcome of this analysis by two measures of predictive accuracy: the receiver operator characteristic (ROC) curves (Fig 1A and 1C) and illustrated confusion matrices (Fig 1B and 1D). To our surprise, Ab-levels considered in these studies, including those against previously proposed vaccine targets, such as MSP1 or AMA1, are poor predictors of individual-level protection, with misclassification rates of up to 56%. A weak signal for protective immunity could be found in the KEN dataset using both antibodies and exposure variables. However, a null-model based solely on exposure proxies (age and schizont extract, blue line in Fig 1A) was equally predictive of an individual’s risk of malaria as the more complex model that also included their immune profiles (green line, Fig 1A). High misclassification was also found in the KTZ cohort data (Fig 1C and 1D), which we believe was mainly due to the very young host ages in this study (between 4-18 months at recruitment), where individuals were still experiencing their first malaria infections. Therefore, exposure did not contribute to the model’s predictive performance.
What these results demonstrate is that even in the case where some responses might show univariate statistically significant associations with protection at the population-level, their effect sizes are too small to be able to predict whether someone with elevated titres will be protected during the next transmission season or not. It is equally possible that the data was simply too limited with respect to the age-range and/or the specificity and number of (allelic) antigens considered in the respective assays to identify immune signatures that clearly distinguish clinically immune individuals.
The protein-microarray-based MAL dataset comprises a much broader set of antigens, representing over 1000 unique proteins, and contained individuals of two different age classes: children between the age of 2-10 years (mean = 5.8, N = 149), which were mostly classed as susceptible (117/149), and young adults between 18-25 years (mean = 20.8, N = 37), who had predominantly acquired a state of clinical protection (33/37 individuals remained symptom-free).
Fig 2A shows the mean antibody measures against all antigens stratified by either age or infection outcome. In both cases we find a consistent, qualitative shift in the immune profiles of protected and older individuals where sets of high-titre antibodies are further elevated and sets of low-titre responses further reduced. Importantly, and as shown in the beanplots [35] in Fig 2B, there is no notable difference in the mean reactivity between the various strata (in all three cases P > 0.2, Welch two sample t-test), implying that clinical protection within this dataset is not characterised by an increase in overall reactivity.
To elucidate the most predictive antigens at the individual-level in this dataset, where the number of immune responses (M = 2320) was much larger than the number of samples (N = 186), we performed feature selection inside a five-fold cross-validation loop (see Material & methods and S1 Text). We first fitted a model to all individuals (2-25 years) using all available predictors (2320 antigenic responses, age and parasite status at screening). The average AUC across all testing folds was 0.83 (Fig 3A), exhibiting very good discrimination between the protected and susceptible individuals. However, this predictive performance dropped considerably when the model was identified using only the children (2-10 years), average AUC across four of the five folds was 0.56 (Fig 3B). For one of the cross-validation folds, no predictors were retained by the feature selection algorithm, so no model could be built.
Fig 4 shows a heatmap of immunoreactivity using the set of antigens selected in at least one of the cross-validation folds, and a word cloud for the protein product description based on the selected features and any other proteins they are highly correlated with (ρ > 0.8). The set of predictive antigens consists of those with either an increasing or a decreasing response as individuals grow older and gain clinical protection. Interestingly, none of the 47 proteins which showed an increasing response with age were correlated with others. Whereas, 232 proteins were correlated with the 32 responses that exhibited a decreasing response with age. This temporal pattern, as well as the strong relationship between age and immune status, suggests that the change in the immune responses that allows us to distinguish susceptible from clinically immune individuals (at least as measured by the protein microarray) is taking shape through continuous exposure to the malaria parasite during childhood, but does not fully develop until early adulthood. In part, this explains why we see such an extensive degradation to model performance when only the children are considered, and why 79 antigens were selected for the model with all individuals, but only 9 when considering only the children. Moreover, whilst age was an important predictor for all individuals (coming up in all five-folds), it was never selected for the model based on only the children. The word cloud suggests that the responses that increase with age are related to surface variable antigens, whilst the ones that decrease consist of a number of conserved proteins of (currently) unknown function. S4 Data contains the complete list of selected antigens and their annotations.
Despite decades of intensive research, we still do not know which antigens are central in the induction of immune responses that protect against clinical malaria. Small effect sizes of individual responses and the often polymorphic nature of many immune targets, coupled with considerable inter-individual heterogeneity, are partially to blame for the lack of a clear relationship between a measured response and the level of protection against malaria it offers. Osier and colleagues [3, 12] have previously proposed that protection could in fact be due to the accumulation of responses against sets and or specific combinations of antigens in a threshold-dependent manner, which would certainly help to overcome the issue of individual responses having small effect sizes. However, the number of combinations and possible interactions between antigens that need to be tested and compared across studies to draw robust conclusions can easily become infeasible using standard statistical approaches. Mainly due to identifiability issues, where insufficient samples are available to estimate each model parameter. The predictive modelling framework based on machine learning described here offers a systematic approach to consider all possible combinations of measured antibodies and to extract the most distinguishing features from these high-dimensional datasets in a hypothesis-free way. Furthermore, as the outcome of this approach is the predictive accuracy, i.e. the degree by which the model can predict the response at the level of the individual, results are easily comparable across studies, in contrast to P-values, which are strongly dependent both on sample size and the chosen statistical test and/or model.
Our analysis of different sets of cohort data based on immune profiles against relatively limited sets of P. falciparum-specific antigens demonstrated that commonly assumed immune correlates and potential vaccine candidates (e.g. MSP-1, MSP-2, or AMA-1) are poor predictors of clinical protection in children. Apart from small effect sizes and the fact that in most studies only a limited set of target alleles are investigated, the age-range of individuals considered in these studies may also play a role in explaining our findings. That is, previous studies have shown that protection against life-threatening disease might be acquired early in life after only a few infections (see e.g. [36]). On the other hand, clinical protection is a more gradual process whereby the probability of a clinical episode declines slowly under repeated exposure. This may make the identification of immune signatures that are highly predictive at the individual-level problematic, unless the considered age ranges, and therefore the levels of cumulative exposure, are sufficiently broad.
What our results also point towards is that a diverse set of antigens must be considered to robustly identify predictive immune signatures. The distinct pattern that we found in the data based on protein-microarrays was characterised by an exposure-driven change in the responses to several surface-expressed and internal, conserved and polymorphic parasite proteins. Importantly, a small subset of antigens was sufficient to predict an individual’s risk of presenting with a symptomatic malaria infection during the following transmission season with a high degree of accuracy, at least within a set of individuals covering a suitably wide age-range. In contrast to previous studies, the difference in the immune profiles between the two phenotypes consisted of both increased and decreased responses to certain antigens, which intensified with age. Whereas those responses which showed increasing intensities with exposure contained various known immune targets, such as PfEMP1, the responses that decreased with age were mostly against proteins of unknown function. It is possible that these decreasing responses are an artefact of the microarray data as they mostly consisted of low-titre responses (see S1 Text), i.e. those with small signal-to-noise ratios. In their original analysis, Crompton et al. [8] only considered responses immunogenic if they were 2 SDs above the controls, whereas we considered all responses and are therefore more likely to pick up immunologically irrelevant features. Further investigations are therefore required to verify if and to what degree these responses relate to (protective) immunological pathways.
The fact that the predictive responses showed such a strong association with age also begs the question whether they are true targets of the protective responses or whether they simply mirror infection histories; the latter is probably more plausible for the internal rather than surface-expressed antigens. Finding reliable markers of previous/cumulative exposure is arguably one of the most fundamental problems in correctly identifying antibody-based correlates of protection. Not only are they important in assessing how many undetected infections get past passive and/or active surveillance during follow-up periods, but the ability to discern between cases where someone is truly protected, i.e. infected but without showing clinical symptoms, or simply has a lower exposure risk, would allow us to define reliable phenotypes. A more direct approach would be the inclusion of only those individuals with documented infections [37]. However, this relies on a much higher sampling frequency or reliable markers of recent infections. Age, homestead location and previous malaria incidence are common markers, but none of these directly quantify how often the immune system was challenged by a P. falciparum infection. Using a predictive framework, Helb and colleagues [21] recently established an alternative way for estimating recent exposure by identifying key antigens that were most predictive of days since the last P. falciparum infection and incidence of symptomatic malaria during the previous year. The immune signature identified in our analysis is not so much indicative of recent infections but more of repeated challenges over years of continuous exposure.
In order to make our findings and methodological approach relevant, not only for understanding the process of natural acquired immunity to malaria, but also with regards to future intervention measures, including vaccines, our results need to be validated against independent datasets. In the first instance these should involve replicate studies in similar transmission settings, including the follow up of individuals over successive transmission seasons to test for the robustness of the identified immune signatures. More importantly, though, is the issue of differentiating between individuals who are protected and those who did not get challenged. In the studies considered here this was not a major concern as they were all based in moderate to high transmission settings with most children experiencing a clinical episode over the study periods. However, in areas of lower transmission intensities this is a pressing concern. One obvious way around this would be the use of longitudinal cohort studies with active surveillance. Another, and much cheaper option would be to exploit data analytical approaches as employed by Helb and colleagues [21].
A good correlate of protection should be a universal biomarker reflecting an immune response that prevents the parasite from causing clinical and life-threatening disease. Our results need to be validated against independent datasets to investigate whether the predictive signatures we identified perform equally well in different malaria endemic settings and to test to what degree they truly capture functional immune mechanisms. In that respect we believe that predictive models offer clear advantages over univariate association analyses. Not only is predictive accuracy directly comparable between studies, but these frameworks also provide a systematic way to consider all putative correlates of protection whilst reducing the chances of false discoveries. With the advent of more detailed and complex big data sets in the field of malaria immuno-epidemiology these models should therefore be considered more prominently alongside standard statistical approaches in an attempt to unravel the complex interplay between exposure and infection outcome in P. falciparum malaria.
|
10.1371/journal.pgen.1004961 | Membrane Recognition and Dynamics of the RNA Degradosome | RNase E, which is the central component of the multienzyme RNA degradosome, serves as a scaffold for interaction with other enzymes involved in mRNA degradation including the DEAD-box RNA helicase RhlB. Epifluorescence microscopy under live cell conditions shows that RNase E and RhlB are membrane associated, but neither protein forms cytoskeletal-like structures as reported earlier by Taghbalout and Rothfield. We show that association of RhlB with the membrane depends on a direct protein interaction with RNase E, which is anchored to the inner cytoplasmic membrane through an MTS (Membrane Targeting Sequence). Molecular dynamics simulations show that the MTS interacts with the phospholipid bilayer by forming a stabilized amphipathic α-helix with the helical axis oriented parallel to the plane of the bilayer and hydrophobic side chains buried deep in the acyl core of the membrane. Based on the molecular dynamics simulations, we propose that the MTS freely diffuses in the membrane by a novel mechanism in which a large number of weak contacts are rapidly broken and reformed. TIRFm (Total Internal Reflection microscopy) shows that RNase E in live cells rapidly diffuses over the entire inner membrane forming short-lived foci. Diffusion could be part of a scanning mechanism facilitating substrate recognition and cooperativity. Remarkably, RNase E foci disappear and the rate of RNase E diffusion increases with rifampicin treatment. Control experiments show that the effect of rifampicin is specific to RNase E and that the effect is not a secondary consequence of the shut off of E. coli transcription. We therefore interpret the effect of rifampicin as being due to the depletion of RNA substrates for degradation. We propose a model in which formation of foci and constraints on diffusion arise from the transient clustering of RNase E into cooperative degradation bodies.
| Recent discoveries that two ribonucleases with major roles in mRNA degradation, RNase E of Escherichia coli and RNase Y of Bacillus subtilis, are localized to the inner cytoplasmic membrane suggest that spatial separation of transcription and mRNA degradation are general features of the bacterial cell. Here we show that RNase E rapidly diffuses over the entire inner membrane forming short-lived foci. Results of molecular dynamics simulations lead us to suggest that RNase E interacts with the lipid membrane by a novel mechanism permitting a high degree of translational freedom. We show that RNA substrate is necessary for the formation of RNase E foci and that formation of foci correlates with constraints on the diffusion of RNase E. We therefore propose that foci are degradation bodies in which several RNase E molecules engage an RNA substrate. The sequestration of the mRNA degradation machinery to the inner cytoplasmic membrane has important consequences for mRNA turnover. This organization likely favors formation of polyribosomes on nascent transcripts before they are exposed to the degradation machinery. Rapid diffusion of RNase E on the inner cytoplasmic membrane could be part of a scanning mechanism that facilitates recognition of cytoplasmic polyribosomes and cooperative degradation of mRNA.
| In Escherichia coli, Salmonella, and many other bacteria, RNase E makes critical contributions to general and regulated mRNA degradation [1, 2]. General mRNA degradation is the default turnover pathway, whereas regulated mRNA degradation is controlled by factors such as sRNA (small RNA) and the RNA binding protein Hfq [3, 4]. RNase E contains a large noncatalytic region that is the scaffold for the assembly of a multienzyme complex known as the RNA degradosome [5]. Recently, RNase E was shown to be localized to the inner cytoplasmic membrane by tagging with fluorescent protein [6, 7], a finding that has been corroborated for the native enzyme as well as other RNA degradosome components by proteomic analyses of the inner membrane [8, 9]. The association of RNase E with the membrane benefits organism fitness as indicated by the slow growth of strains bearing deletions or point mutations that disrupt membrane binding [6], so the interaction is likely to be functionally important. It has been postulated that membrane association physically separates sites of transcription from sites of mRNA degradation and thereby confers a time delay before the onset of decay of a transcript [10]. The general importance of the localization of RNase E has been underscored by the recent finding that RNase Y, a key ribonuclease of mRNA degradation in Bacillus subtilis, is also membrane-localized [11]. What makes this parallel especially striking is that RNase E and RNase Y share no common evolutionary ancestor and their functional analogy therefore arose through convergent evolution.
The basis for the interaction of RNase E with a phospholipid bilayer was established by the identification of a 15-residue MTS in the noncatalytic region that is necessary and sufficient for membrane localization [6]. The MTS, with the propensity to form an amphipathic α-helix, bears signature features conserved in RNase E homologs throughout the γ-Proteobacteria including the clustering of bulky aromatic residues on one face of the helix, small hydrophilic residues on the opposite face and the presence of basic residues flanking the hydrophobic core. Mutation of signature residues of the MTS of E. coli RNase E confirmed their importance for membrane localization in vivo and for interaction with protein-free phospholipid vesicles in vitro [6]. To elucidate the structural basis for RNase E recognition of the cytoplasmic membrane, we have undertaken molecular dynamics simulations with a realistic model of the E. coli inner membrane and the MTS peptide, and we have performed binding studies using a fluorescein-labelled derivative of this peptide. These analyses shed light on the geometry, energetics and dynamics of the interaction of the MTS with the lipid bilayer.
Recent reports suggest that RNase E and RhlB, which is a DEAD-box RNA helicase component of the RNA degradosome, form a membrane-associated cytoskeletal-like structure, and that RhlB localizes to the cytoskeleton independently of RNase E [7, 12, 13]. The question therefore arises how RNase E and RhlB interact in vivo, and to what degree RNase E and RhlB are free to diffuse on the inner cytoplasmic membrane. To address this question, we have conducted microscopy studies in live cells in which RNase E and RhlB were tagged with fluorescent protein. Under live cell conditions, the association of RhlB with the membrane depends on its interaction with RNase E. In addition, we observe that RNase E rapidly diffuses on the inner cytoplasmic membrane forming transient foci. The likely impact of the membrane dynamics of RNase E on its access to RNA substrates and the coordinated activities of the degradosome will be discussed.
Although there is evidence that the DEAD-box RNA helicase RhlB associates with RNase E through a direct protein-to-protein interaction [5, 14, 15], recent reports have suggested that RhlB by itself can form a membrane-associated cytoskeletal-like structure [12, 13]. We therefore explored the structural requirement for the localization of RhlB to the inner cytoplasmic membrane by constructing strains in which RNase E and RhlB were tagged at their C-terminal ends with mCherry and CFP (Cyan Fluorescence Protein), respectively. These constructs are functional single copy chromosome replacements in the NCM3416 background, which is a wild type E. coli K12 strain [16]. Additional constructs contain variants of RNase E-mCherry in which the MTS, protein Scaffold (Sca) or HBS (Helicase Binding Site) were deleted based on previous work mapping these sites [17].
Fig. 1 presents a gallery of micrographs showing images of strains expressing RhlB-CFP and RNase E-mCherry. In Fig. 1, a few cells were chosen from a large field (S1 Fig.). Cultures were grown to mid logarithmic phase in MOPS-glycerol-amino acids media at 30°C. Similar results were obtained in LB media and at 37°C. In the wild type strain (top panel), RNase E and RhlB are enriched in foci at the periphery of the cell. RNase E and RhlB do not co-localize in these images, which were made with a 4 s exposure time due to the weak RhlB-CFP fluorescence signal. The apparent lack of co-localization is likely due to rapid movement of RNase E under the live cell conditions used in these experiments (S2 Fig. and results below). In the second panel, RNase E ∆MTS is a variant in which the MTS has been deleted. Both RNase E and RhlB are delocalized from the periphery and the signal is cytoplasmic and diffuse. These results demonstrate that the membrane localization of RhlB depends on the MTS of RNase E. In the third panel, RNase E ∆Sca is a variant with a deletion of the scaffold, which interacts with RhlB, enolase and PNPase. Like the wild type protein, RNase E ∆Sca is localized in foci at the periphery of the cell, whereas RhlB is cytoplasmic and diffuse. These results demonstrate that the membrane localization of RhlB requires an interaction with the scaffold region of RNase E, either directly or indirectly via an interaction with another component of the RNA degradosome. In the fourth panel, RNase E ∆HBS is a variant of RNase E in which the binding site for RhlB has been deleted. RhlB is localized to the cell interior as in the RNase E ∆Sca construct. These results demonstrate that the localization of RhlB to the membrane depends on a direct protein-to-protein interaction with RNase E. Due to the limit of resolution of light microscopy, we cannot exclude the possibility that some molecules of RhlB-CFP remain membrane associated in the RNase E ∆Sca and ∆HBS constructs. We nevertheless conclude that most if not all of the membrane localization of RhlB-CFP depends on a protein interaction with RNase E and that RhlB by itself is not a membrane binding protein. Throughout the construction and imaging of these strains, as well as imaging in other strain backgrounds [6], we have not observed cytoskeletal-like structures that were reported in other studies [7, 12, 13].
To better understand the molecular basis for the interaction of the MTS with the phospholipid bilayer, and to determine what constraints, if any, this type of membrane association has on the diffusion of RNase E, we performed coarse-grain molecular dynamics simulations. We used a membrane model with realistic E. coli lipid composition, namely cardiolipin (CARD), dipalmitoylphosphatidylglycerol (PG) and dipalmitoylphosphatidylethanolamine (PE) in a ratio of approximately 1:2:6, respectively [18, 19]. We used a 21-residue sequence corresponding to the residues 565–585 of RNase E, which includes the 15-residue core of the MTS (Fig. 2A). This peptide was used in previous biophysical measurements of the interaction of the MTS with a phospholipid bilayer made with purified E. coli lipids [6]. In the coarse-grain simulation, the peptide starts in bulk solution in a helical conformation, and with time randomly encounters the phospholipid bilayer, whereupon it adheres to the membrane surface and then penetrates into the acyl interior (Fig. 2B). A molecular graphics movie of the simulation is provided (S1 Video). The simulation shows that the MTS inserts in the membrane with the helical axis oriented parallel to the bilayer plane and hydrophobic side chains buried deep into the acyl core of the lipids. The peptide interacts with approximately 50 phospholipids (S3 Fig.), in reasonable agreement with previous calorimetry measurements in which the binding isotherm was fitted to an interaction with approximately 40 phospholipids [6]. The helical conformation of the MTS was retained throughout the coarse-grain simulation. Atomistic simulations of the MTS without any restraints on the secondary structure revealed that the MTS is a stable helix in the environment of the phospholipid bilayer. This result is consistent with previous circular dichroism measurements showing that the peptide has greater propensity to form an α-helical conformation upon interaction with a phospholipid bilayer composed of E. coli lipids [6].
Molecular dynamics simulations were also undertaken with MTS variants that were previously studied experimentally [6]. These are a double substitution of phenylalanine with alanine (F574A/F575A); a substitution of phenylalanine with glutamic acid (F575E); and a proline insertion between residues F574 and F575. In the wild type peptide, the hydrophobic amino acid side chains penetrate into the acyl interior of the lipid bilayer, the small hydrophilic and basic amino acid side chains interact at the level of the glycerol moieties (Fig. 2C). The 3-residue N- and C-terminal extensions flanking the MTS interact at the level of the phosphate and ethanolamine moieties. Estimates of the energy of interaction (S4 Fig.) correlate with the depth of insertion of the hydrophobic amino acids into the membrane. In the F574A/F575A and F575E variants, the hydrophobic side chains protrude into the aqueous layer and the small hydrophilic and basic residues interact at the level of the ethanolamine moieties (Fig. 2C). In previous calorimetry experiments employing the corresponding peptides, no heat of interaction with lipid vesicles was detected with peptides corresponding to the F574A/F575A and F575E variants [6] suggesting that interaction of the hydrophobic side chains with the acyl core of the lipid bilayer drives the binding reaction. Microscopy of fluorescein-labelled peptides that correspond to the sequences in the molecular dynamics simulations confirms binding in vitro to vesicles made with purified E. coli lipids (Fig. 2D). The degree of fluorescence at the membrane diminishes with the F574A/F575A variant, in accord with the findings from the molecular dynamics simulations and the diminished membrane association of the corresponding RNase E variant in vivo [6]. The simulation with the proline insertion variant shows that it has a more favorable energy of interaction than wild-type (S4 Fig.), probably due to additional contacts with the inserted proline, which is buried deep in the acyl interior of the lipid bilayer (Fig. 2C). The results of the molecular dynamics calculations are in agreement with experimental work showing that full-length RNase E with a proline insertion at this position interacts with phospholipid vesicles in vitro and localizes to the cytoplasmic membrane in vivo [6]. The stability of the helix in the membrane and the estimated free energy explain why the proline insertion does not disrupt interaction with the phospholipid bilayer. The congruence between the previous biophysical measurements and the properties predicted by the molecular dynamics simulation validate the coarse-grain approach used here.
Since membrane composition and curvature can affect protein binding [20, 21], we asked whether these parameters are predicted to affect the interaction of the MTS with the phospholipid bilayer. The molecular dynamics simulations indicate that the MTS exhibits preferential interaction with anionic lipids (CARD and PG) in comparison to Zwitterionic lipids (PE) (S5 Fig.). This result suggests that the basic residues flanking the hydrophobic core of the amphipathic α-helix, which is a conserved feature of the MTS [6], form favorable electrostatic contacts with anionic lipids that help to stabilize the α-helix and/or increase the energy of interaction. Membrane composition could therefore affect the interaction of the MTS with the phospholipid bilayer. The molecular dynamics simulations presented here are based on planar lipid bilayers, but simulations of small lipid vesicles with curved surfaces composed of phosphatidlycholine (PC) or a realistic mixture of E. coli lipids gave similar results suggesting that the interaction of the MTS with the phospholipid bilayer is not sensitive to membrane curvature.
Diffusion can be important for the function of a membrane protein since it affects interactions with substrate and other proteins. In the coarse grain simulation using a 10 µs period, the rate of diffusion for the MTS on the membrane was predicted to be in the order of 103 µm2/s. We know of no other study of a bacterial peripheral membrane protein that can be used for comparison. The rate of diffusion of the membrane anchor of GRP1, which is a eukaryotic peripheral membrane protein, is 330-fold slower than the predicted rate of diffusion of the MTS [22]. The membrane anchor of GRP1 makes a specific high affinity interaction with phosphatidylinositol-3,4,5-trisphosphate (PIP3). Experimental work including molecular dynamics simulations suggests that the rate of diffusion of GRP1 on the membrane is limited by the frictional coefficient of PIP3. The much faster predicted rate of diffusion of the MTS of RNase E suggests a different mechanism of translocation. We propose that the MTS ‘glides’ in the phospholipid bilayer by making a large number of weak contacts that are rapidly broken and reformed. Although RNase E in the cell would likely have a slower rate of diffusion, our results suggest that it would have a high degree of translational freedom in the absence of interactions with other components such as membrane proteins or polyribosomes.
During the cell imaging work shown in Fig. 1 and previous work with the KSL2000/pVK207 strain, which expresses RNase E-YFP (Yellow Fluorescent Protein) [6], we noticed that the RNase E fluorescence signal in live cells rapidly fluctuated, giving the impression that foci of RNase E circulate on the periphery of the cell. Fixation of the cells with formaldehyde arrests this motion. To better define the distribution of RNase E in the cell, we first examined formaldehyde fixed cells (KSL2000/pVK207) by confocal microscopy (Fig. 3A). Deconvolution of this image shows that RNase E localizes to the periphery of the cell. We then used TIRFm to selectively excite RNase E-YFP in a thin layer adjacent to the coverslip. Fig. 3B shows wide field and TIRF images of two fields of live cells. The diagram in the lower right hand corner of each image indicates the plain of focus. In these images, an exposure time of 100 ms was used to minimize motion during the acquisition, which can result in apparent but artifactual polymer-like structures. No elongated polymeric or helical structures were observed under these high-speed imaging conditions. Rather, randomly distributed clusters of RNase E were observed. S2 Fig. shows two high-speed TIRFm images of live cells taken 4 s apart. An overlay of these images artificially colored red and green shows massive redistribution of RNase E in the 4 s time interval. Using an approach that has been previously used to track MreB movement [23–25], the distribution and directionality of RNase E movement was analyzed in the kymograms shown in Fig. 3C. A single cell was scanned along its long and short axes. Heat maps derived from the intensity of the fluorescent signal were accumulated to construct the kymograms. Inspection of the fixed cell shows that the distribution of RNase E on the membrane is unchanged over a period of 3 seconds, whereas the distribution rapidly fluctuates in live cells. The lack of any clear repeat pattern in the kymograms of the live cell suggests that the movement of RNase E is random; that is, strongly correlated movement should appear as tracks or waves in the kymograms.
In the KSL2000/pVK207 strain analyzed in Fig. 3, RNase E-YFP was expressed from a low copy number plasmid that complemented a deletion of the gene encoding RNase E on the chromosome [6]. Comparable results were obtained using the Kti164 strain in which RNase E-GFP was expressed from a single copy construct on the chromosome, but the GFP signal is less intense than the YFP signal. Experimental work suggests that this difference is due to the intrinsic relative brightness of the YFP and GFP constructs since the level of RNase E-YFP and RNase E-GFP, as determined by SDS-PAGE and Western blotting is comparable to wild-type RNase E levels ([6] and S6 Fig.). We interpret the results of epifluorescence microscopy, confocal microscopy and TIRFm as evidence for the rapid diffusion of RNase E over the entire inner membrane and the formation of short-lived foci containing multiple molecules of RNase E.
We wanted to know if the localization or diffusion of RNase E on the membrane depends on an energy source or forces generated by transcription or translation. We therefore analyzed the distribution of RNase on the membrane after treatment of the cells with carbonyl cyanide m-chlorophenyl hydrazone (CCCP), kanamycin or rifampicin. CCCP collapses the transmembrane proton gradient, while kanamycin and rifampicin are inhibitors of translation and transcription, respectively. Treatment of the cells with CCCP or kanamycin had no discernable effect on RNase E localization making it unlikely that the electrochemical gradient, ATP generation or translation influences the cellular distribution of RNase E (S7 Fig.). In contrast, after treatment with rifampicin the appearance of RNase E on the membrane is different as evidenced by the loss of foci and smooth distribution along the perimeter of the cell (Fig. 4A). The intensity of fluorescence along the membrane was measured by the line scans shown in Fig. 4B providing a method to quantify changes in the distribution of RNase E on the membrane. This analysis was applied to a field of cells to generate plots of average pixel intensity and variance (Fig. 4C). Average pixel intensity is a measure of the level of RNase E in the cell. The plot in Fig. 4C shows that the distribution of average pixel intensity is not affected by rifampicin treatment. This is expected since RNase E is a stable protein and rifampicin treatment is not predicted to change its level. In contrast, the average variance is clearly lower after treatment with rifampicin showing that the smooth distribution of RNase E along the perimeter of the cell (Fig. 4A) is a general property of a population of cells.
To further examine the effect of rifampicin, we used the intrinsic photobleaching that occurs in TIRFm to measure the diffusion of RNase E on the membrane. Briefly, since only a portion of the membrane is excited in TIRFm, the rate of photobleaching is related to the rate of diffusion of the fluorescent protein. If a fluorescent protein diffuses rapidly relative to the intrinsic rate of photobleaching, then the pool of bleachable molecules will be slowly depleted since individual molecules only spend a short time on the surface that is excited. If, on the other hand, diffusion is slow, then the pool of bleachable molecules will be depleted faster since individual molecules will spend a longer time on the surface that is excited. With the appropriate controls, it is possible to estimate a relative rate of diffusion by this technique [21]. Fig. 5A shows snapshots of RNase E distribution by TIRFm before and after treatment with rifampicin, Fig. 5B is a photobleaching time course and Fig. 5C is the quantification of the photobleaching experiment. Rifampicin treatment results in a diffuse distribution of RNase E with few if any intense foci as compared to the untreated control. Remarkably, rifampicin decreases the rate of photobleaching, indicating an increase in the rate of RNase E diffusion. To test if rifampicin treatment has a general effect on the diffusion of membrane proteins, we measured the photobleaching rate of the F1Fo ATP synthase before and after treatment with rifampicin (S8 Fig.). As the rate of diffusion of the F1Fo ATP synthase is not affected, we conclude that rifampicin specifically affects RNase E diffusion.
To explore whether RNA substrate is required to form RNase E foci, we exploited the ENS134 strain encoding bacteriophage T7 RNA polymerase [26, 27], which is insensitive to inhibition by rifampicin. In this strain, it is possible to inhibit E. coli RNA polymerase under conditions in which T7 RNA polymerase actively transcribes genes with a bacteriophage T7 promoter. The ENS134 strain has a chromosomal copy of the gene for T7 RNA polymerase under the control of an inducible lac promoter, and a chromosomal copy of the gene encoding lacZ under the control of a bacteriophage T7 promoter (Fig. 6A). The lacZ gene is fused to a tRNA gene followed by a transcription terminator. Transcription by T7 RNA polymerase results in high level synthesis of a lacZ-tRNA transcript. RNase E is necessary for degradation of the lacZ mRNA [27]. Maturation of the tRNA is predicted to require RNase E and RNase P considering the established pathway in E. coli [28]. In order to visualize RNase E in the ENS134 strain, we introduced pVK207, which is the low copy number plasmid encoding the RNase E-YFP fusion used in the work shown in Figs. 3, 4 and 5. Fig. 6B shows epifluorescence images of the ENS134/pVK207 strain. The control panel (-rifampicin, -IPTG) shows that RNase E-YFP distribution is comparable to what we observed in the KLS2000/pVK207 strain. Autoregulation of RNase E expression, which is predicted to down regulate expression of plasmid rne-yfp and chromosomal rne genes, results in a level of total RNase E (RNase E + RNase E-YFP) comparable to the normal RNase E level [6, 29, 30]. Since there are approximately 5 plasmid copies of rne-yfp for each chromosomal copy of rne, RNase E-YFP is the predominant form of RNase E in these cells.
Images in Fig. 6B show that cells treated first with IPTG and then rifampicin have a pattern of peripheral RNase E foci similar to cells that have not been treated with rifampicin (+IPTG, +rifampicin vs. +IPTG, ‑rifampicin). This result suggests that synthesis of the lacZ-tRNA transcript in the presence of rifampicin results in an interaction that stimulates the formation of RNase E foci. The effect of rifampicin on a field of cells was quantified (Fig. 6C). This result shows that changes in distribution of RNase E along the perimeter of the cell is a general property of the population of cells treated with rifampicin. The small reduction in variance after treatment with rifampicin (+IPTG, +rifampicin vs. +IPTG, ‑rifampicin) is likely due to inhibition of E. coli RNA polymerase. From these results we conclude that the effect of rifampicin on formation of RNase E foci is not a secondary consequence of the shut off of E. coli transcription. We next analyzed RNase E-YFP photobleaching. Fig. 6D shows that there is no effect on the rate of diffusion of RNase E if IPTG is added before rifampicin. This result suggests that synthesis of the lacZ-tRNA transcript in the presence of rifampicin results in an interaction that constrains the diffusion of RNase E. Taken together, these results show that there is a direct correlation between formation of foci and constraints on the diffusion of RNase E. Considering this experimental work, we propose that the formation of RNase E foci requires interaction with RNA substrate, that foci formation constrains the diffusion of RNase E, and that rifampicin acts on the foci indirectly by depleting the pool of RNA substrate.
Live cell microscopy shows that RNase E is located on the cytoplasmic membrane and that the DEAD-box RNA helicase RhlB is associated with RNase E, but neither protein forms cytoskeletal-like structures as reported earlier [12, 13]. This is the first report in which RhlB has been visualized directly in live cells as previous work employed indirect immunofluorescence. The YFP, GFP and CFP tags used here are A206K variants that have been reported to minimize dimerization [31]. The fusion proteins are present at levels comparable to wild type. We have used agarose pads that do not disturb the physiology of the cell, whereas cells immobilized on glass slides were imaged in the work showing cytoskeletal-like structures. Under the conditions used here, RNase E is highly mobile over the entire surface of the membrane. The movement is spontaneous since it does not appear to be driven by an energy dependent process suggesting that the motion of degradosome assembly is due to continuous buffeting by other macromolecules in the densely packed milieu of the cell including the membrane-associated cell wall synthesis machinery [21]. Recent biophysical studies of live bacterial cells suggest that particle size has a disproportionate influence on diffusion. Small particles including free ribosomes and multienzyme complexes can be treated as components in a liquid-like state whereas larger particles such as polyribosomes are apparently constrained in a solid-like state that requires metabolic activity for ‘mixing’ [32, 33].
Molecular dynamics simulations provide a conceptual framework for visualizing how the MTS anchors RNase E to the inner cytoplasmic membrane by a novel mechanism permitting a high degree of translational freedom. Although the interaction of an individual MTS with the membrane represents a weak force compared to integral membrane proteins with multiple trans-membrane segments, the cumulative effect of having four MTS elements spatially co-localized as a consequence of RNase E tetramerization should have a strong binding effect through chelate cooperativity [34], which is consistent with biochemical work showing that the solubilization and purification of RNase E requires high concentrations of nonionic detergent [35]. Considering previous measurements of the interaction of a peptide corresponding to the MTS with a phospholipid bilayer composed of E. coli lipids (Kd = 1.3 × 10-6 M) [6] and the effect of chelate cooperativity, we believe that RNase E should be fully membrane-bound, which is consistent with the confocal microscope image presented here. Membrane association could aid the organization of the RNase E tetramer and bring the noncatalytic C-terminal region, which interacts with the other components of the RNA degradosome, within a restricted hemisphere in which they may cooperate with the catalytic core of RNase E. The interaction of the MTS with the membrane is predicted to affect the local concentration of lipid species. The spatial co-localization of the four MTS elements in the RNase E tetramer could accentuate a preference for anionic lipids at the site of membrane docking. Cooperativity in the interaction between anionic lipids with basic residues flanking the hydrophobic core of four MTS elements could further increase avidity of RNase E for the membrane.
We have described a highly dynamic distribution on the membrane in which RNase E forms short-lived foci. Analysis of the dynamics of RNase E motion by TIRFm suggests that diffusion on the membrane is random with no indication of correlation with the long or short axis of the cell. This result excludes models in which RNase E moves along ‘tracks’ in the membrane or is constrained by the machinery involved in cell wall synthesis as has recently been described for MreB [23, 24]. Depletion of RNA substrates by inhibition of transcription results in disappearance of foci and an increase in the rate of diffusion of RNase E. Our results suggest that foci are sites of degradation in which several molecules of RNase E as well as other components of the RNA degradosome interact with an RNA substrate. RNase E foci could therefore have a function analogous to eukaryotic P-bodies and stress granules, which are ribonucleoprotein particles containing factors involved in translation inhibition and mRNA degradation [36, 37]. RNase E foci are nevertheless much smaller and much more short-lived than P-bodies or stress granules.
An important activity of RNase E is general and regulated mRNA degradation. An example of regulated mRNA degradation is the rapid response associated with quorum sensing in Vibrio cholera, which involves sRNA, Hfq and RNase E [38]. A caveat to such regulation is that the target mRNA finds its way to RNase E and is actively degraded. Recent work suggests that RNase E can interact with polyribosomes and sRNA/Hfq complexes [39, 40]. It can therefore be envisaged that RNase E associated with an sRNA/Hfq complex could interrogate a polyribosome as mRNA spools through the translational machinery providing a window of opportunity for degradation if the sRNA can match a recognition region in the transcript. General mRNA degradation, which is initiated in the absence of a regulatory factor, could also involve polyribosome interrogation with direct competition between translation re-initiation and cleavage by RNase E. In either case, a productive interaction would lead to the formation of ribosome-free mRNA facilitating the recruitment of additional RNase E molecules. Sequestration of the RNA degradosome to the two dimensional surface of the inner cytoplasmic membrane and rapid diffusion could increase cooperativity in general and regulated mRNA degradation as well as other processes involving RNase E such as tRNA maturation. It is also possible that this system has a quality control function in the degradation of defective transcripts that fail to form polyribosomes or other ribonucleoprotein complexes. Characterization of RNA substrates localized to RNase E clusters could help to identify cellular processes that are facilitated by the formation of cooperative degradation bodies.
Atomistic simulations were performed using the GROMACS simulation package, version 4.5.1 [41] with the GROMOS53a6 force-field [42, 43] and the SPC water model [44]. Simulations were performed in the NPT ensemble, with the Nosé Hoover thermostat [45, 46] with a time constant of 0.5 ps and the Parrinello—Rahman barostat [47] with a time constant of 5.0 ps used to maintain a temperature of 310 K and a pressure of 1 bar. Long-range electrostatic interactions were treated using the smooth particle mesh Ewald method and a long-range dispersion correction was applied to the energy and pressure beyond the cut-off. The neighbor list was updated every 5 steps during the simulations. All bonds were constrained using the LINCS algorithm [48] allowing a 2 fs timestep to be applied in all simulations. The protonation states of all titratable residues of the peptides were assigned using pH 7. Repeats of all of the simulations were performed using different randomly assigned starting velocities. Peptides were manually positioned in the bulk solvent just above the lipid headgroups of one leaflet of the pre-equilibrated DPPC lipid bilayers.
Within the GROMACS simulation package, the MARTINI force-field was used for the lipids and a variant of this was used to represent the interactions between lipid and protein [49]. The peptide secondary structure was retained using weak restraints between backbone particles to represent hydrogen-bonds, as has been shown to work well for other peptides [50]. The different bilayers (E. coli membrane models contained 14 cardiolipin, 88 DPPE, and 26 DPPG lipid molecules) were setup by substituting the appropriate lipids into pre-equilibrated DPPC bilayers. The non-bonded neighbor list was updated every 10 steps. The integration time step was 40 ps. All simulations were performed at constant temperature (310 K) and pressure (1 bar), using the Berendsen thermostat and barostat [51]. Lennard-Jones interactions were shifted to zero between 9 Å and 12 Å, and electrostatics were shifted to zero between 0 Å to 12 Å, with a relative dielectric constant of 15.
Strains and plasmids are listed in Table 1. The Kti series of strains was constructed in the NCM3416 background using the λ Red recombination system [52, 53]. PCR products (S1 Table) were transformed into NCM3416/pKD20. Cmr (Chloramphenicol resistant) transformants were selected at 30°C and then streaked at 42°C to eliminate pKD20. The constructs were purified by bacteriophage P1 transduction of the Cmr marker into NMC3416. The strains were then transformed with pCP20 and ampicillin resistant colonies selected at 30°C. To eliminate pCD20 and the Cmr cassette, transformants were streaked at 43°C on ampicillin and then tested for loss of the Cmr marker. The coding sequence of the fusion protein and flanking regions were then sequenced. All constructs are C-terminal fusions and the FRT scar is located downstream of the translation stop codon. For construction of the strains expressing RNase E-mCherry, RNase E-GFP and RhlB-CFP, DNA fragments generated by crossover PCR were transformed into NCM3416/pKD20. For construction of strains expressing RNase E(∆MTS)-mCherry, RNase E(∆Sca)-mCherry and RNase E(∆HBS)-mCherry, plasmids containing the mutant rne allele fused to mCherry and the Cmr cassette were constructed (S1 Table). These plasmids were then used as templates to produce PCR products to transform NCM3416/pKD20.
Deletions in RNase E variants are as follows (wild type RNase E coordinates): ∆MTS, 567–582; ∆scaffold, 702–1061; ∆HBS, 705–737. SDS-PAGE showed that the fusion proteins are stable and their levels are comparable to the wild type protein. For Western blotting, the transfer of RNase E-GFP and RNase E-mCherry is inefficient and not sufficiently reproducible to be quantified. We therefore visualized the fusion proteins directly by SDS-PAGE and Coomassie staining (S6A–S6B Fig.), which is possible because the fusions are amongst the largest proteins in the cell and they migrate as distinct bands. As it is difficult to raise specific high-titer antibodies against RhlB, we examined RhlB and RhlB-CFP using affinity purified polyclonal rabbit antibody as described [54] or antibody against CFP (S6C–S6D Fig.). These results show that RhlB-CFP is stable and that its level in the cell is comparable to wild type RhlB. Strains expressing defective RNase E variants grow slower than an isogenic wild type control and the defective variant is expressed at a higher level due to autoregulation [6, 29, 30]. Since the growth rate and RNase E levels of the strains expressing RNase E-mCherry, RNase E-GFP and RNase E-YFP are normal (S6 Fig. and [6]), we conclude that these fusion proteins are fully active. RhlB is not essential, but its activity can be tested in vivo since ∆rhlB in a pcnB- background results in the accumulation of mRNA degradation intermediates [55]. RhlB is part of a 3′ exonucleolytic mRNA degradation pathway involving RhlB and PNPase as components of the RNA degradosome. S9 Fig. shows a Northern blot probed for an mRNA degradation intermediate known to accumulate in the ∆rhlB, pcnB- background. A probe for 5S ribosomal RNA was included as a loading control. In the pcnB- background, the mRNA degradation intermediate accumulates in the rhlB-cfp strain, but the level is much higher in the ∆rhlB strain. We therefore conclude that the RhlB-CFP fusion is active in vivo albeit at a lower level than wild type RhlB.
Liquid cultures were grown at 30°C with vigorous aeration in LB or MOPS medium [56] supplemented with glycerol (0.5%) and amino acids (50 µg/ml each l-amino acid except tryptophane, tyrosine and phenylalanine). Microscope setups are listed in S2 Table. Bacteria were prepared for microscopy as described [57]. Aliquots (0.5–1.0 µl) from cultures grown to mid logarithmic phase were spotted on microscope slides covered with a thin layer of agarose (1.2% in water). After a few minutes to allow absorption of the cells, the agarose pad was covered with a slip and the slide immediately mounted on the microscope to take images. Preparation of the microscope slide and imaging was performed at room temperature. To chemically fix cells, aliquots from cultures grown to mid logarithmic phase were treated with formaldehyde (1%) for 10 min at 30°C with agitation and then quenched with glycine (100 mM). Images were analyzed using ImageJ [58, 59]. TIRFm photobleaching and quantitative analyses to determine relative diffusion rates were performed as described [21].
For the analysis of peptide binding, FITC-labelled PAQPGLLSRFFGALKALFSGGK (wild type) or PAQPGLLSRAAGALKALFSGGK (F574A/F574A variant) were mixed with liposomes and visualized by spinning disk confocal fluorescence microscopy. Liposomes were prepared from E. coli lipid extracts (Avanti Polar Lipids) using Octyl-Glucoside detergent dialysis method as described [57]. Liposomes were diluted to 1 mg/ml in 5 mM Tris pH 7.5, 150 mM KCl, and peptides were added at concentration of 100 ug/ml. 2 mg/ml BSA was included as unspecific protein. Incubation (5 min) and microscopy were performed at 30°C.
Not applicable. This study did not involve human participants, specimens or tissue samples, or vertebrate animals, embryos or tissues.
|
10.1371/journal.pntd.0003513 | Anopheles Midgut Epithelium Evades Human Complement Activity by Capturing Factor H from the Blood Meal | Hematophagous vectors strictly require ingesting blood from their hosts to complete their life cycles. Exposure of the alimentary canal of these vectors to the host immune effectors necessitates efficient counteractive measures by hematophagous vectors. The Anopheles mosquito transmitting the malaria parasite is an example of hematophagous vectors that within seconds can ingest human blood double its weight. The innate immune defense mechanisms, like the complement system, in the human blood should thereby immediately react against foreign cells in the mosquito midgut. A prerequisite for complement activation is that the target cells lack complement regulators on their surfaces. In this work, we analyzed whether human complement is active in the mosquito midgut, and how the mosquito midgut cells protect themselves against complement attack. We found that complement remained active for a considerable time and was able to kill microbes within the mosquito midgut. However, the Anopheles mosquito midgut cells were not injured. These cells were found to protect themselves by capturing factor H, the main soluble inhibitor of the alternative complement pathway. Factor H inhibited complement on the midgut cells by promoting inactivation of C3b to iC3b and preventing the activity of the alternative pathway amplification C3 convertase enzyme. An interference of the FH regulatory activity by monoclonal antibodies, carried to the midgut via blood, resulted in increased mosquito mortality and reduced fecundity. By using a ligand blotting assay, a putative mosquito midgut FH receptor could be detected. Thereby, we have identified a novel mechanism whereby mosquitoes can tolerate human blood.
| Mosquitoes are important vectors in the transmission of many human diseases. Their life cycle requires a blood meal to be completed. Ingested blood contains bioactive molecules belonging to the innate immune defense mechanisms against microbes, like the complement system, that can damage foreign cells. We have identified in this study a mechanism whereby mosquitoes can escape the damaging activity of the complement system in the ingested human blood. The mosquito midgut epithelial cell surface captured factor H, a natural regulator of the alternative pathway of complement activation, from the ingested blood. Consequently, the deposition of C3b, a key complement component, on the epithelial cell surface was impaired and cell death was avoided. Interfering with the complement regulatory activity of factor H by monoclonal antibodies, carried to the midgut via blood feeding, increased mosquito mortality and reduced fecundity. The putative Anopheles mosquito factor H binding proteins could be transmission blocking vaccine candidates targeting the malaria parasite carrying vectors.
| Mosquitoes can transmit important parasitic diseases such as malaria and filariasis and viral diseases such as yellow fever, dengue, Rift Valley fever and the West Nile virus. Anopheles, Aedes, Culex, Coquillettidia, Mansonia and Ochlerotatus species are the best known disease transmitting mosquitoes[1]. They all require a blood meal to obtain proteins from their hosts. Blood proteins are needed for the development and laying of eggs to complete the life cycles of the mosquitoes. Parasites and viruses carried in the host blood can therefore be transmitted to other individuals of the same host species and sometimes also to other species if the organisms can multiply inside mosquitoes and survive in the new hosts. Ingestion of host blood has been suggested to pose a danger to mosquitoes as a result of exposing the alimentary canal (AC) to bioactive molecules that normally exist in host blood as part of the host defense mechanisms against microbes. Likewise, other ingested blood-derived factors such as antibodies, hemoglobin-derived peptides, enzymes and signaling molecules could alter the physiology of hematophagous vectors (reviewed in[2]). The most immediate system that has been shown to be overcome by mosquitoes and other hematophagous vectors is the coagulation system[3]. Mosquitoes’ and ticks’ salivary molecules were found to inhibit blood clotting at the biting site. The injected saliva contained anti-coagulants that permitted smooth flow of blood from the skin of the host to the vector and prevented blockage of the blood sucking capillary[3].
The complement system is a host defense mechanism that could impose danger to disease vectors upon blood feeding. It is a cascade that attacks the surfaces of foreign cells[4]. Complement plays a central role in the innate immune response to combat microbial infections. There are three pathways to activate complement, the classical, the alternative and the lectin pathway. The classical pathway is triggered when C1 interacts with antibodies bound to their antigens. This results in the cleavage of C4 and C2 and the formation of the classical pathway C3-convertase, C4b2a, which cleaves C3 into C3b. The lectin pathway is activated when the mannan-binding lectin (MBL) or one of the three ficolins binds to sugar residues (Man, GalNAc or acetylated sugars) on target surfaces. The alternative pathway is initiated through a spontaneous cleavage of an internal thioester bond in C3. All three pathways converge in the cleavage of the C3 protein. This will direct phagocytosis of targets and leads ultimately to the formation of membrane attack complexes (MAC). MAC is a pore that can cause damage or lysis of the target cells.
Under normal circumstances, activation of complement is kept under tight control by the coordinated action of soluble (e.g. C1-INH, C4bp, factor H, factor I, S-protein and clusterin) and membrane-associated (DAF, CR1, MCP, CRIg and CD59) complement regulatory proteins[5]. The control of complement activity involves inhibition of assembly or dissociation of C3 convertases and the inhibition of the membrane attack complex formation[5]. Among the complement regulators, factor H (FH) plays an integral role in controlling complement activation. It regulates the positive feedback loop of the alternative pathway, which could otherwise lead to excessive activation of C3 and damage to nearby cells [6,7]. FH regulates complement activation on self-cells surfaces by possessing both cofactor activity for the factor I-mediated C3b cleavage, and decay accelerating activity to dissociate the C3-convertase, C3bBb. FH protects self surfaces because it binds to glycosaminoglycans (GAGs) and other polyanions that are generally present on host cell surfaces. Microbes usually lack these structures [6,7].
Given the importance of FH in controlling complement activation, several pathogens that are known to resist complement-mediated killing have been shown to express FH-binding proteins on their surfaces [5]. We hypothesized that this strategy could be used by the hematophagous vectors, as well, to avoid complement-mediated damage when exposed to blood. We therefore analyzed how complement is regulated in the midgut after a blood meal. We observed that complement remained active and could become activated in the mosquito midgut. This suggested that an evasion mechanism to avoid the detrimental consequences of complement activation must exist in the mosquito midgut cells. Accordingly, FH was found to bind to the luminal surface of the mosquito proventriculus (located between the esophagus and anterior midgut), the anterior midgut and the anterior-posterior region of the midgut. Complement activation on these surfaces was shown to be prevented. Taken together, complement activation was shown to occur in the mosquito midgut but the Anopheles mosquitoes were able to protect themselves at this critical site via a novel evasion mechanism.
Anopheles stephensi (Nijmegen strain) and Anopheles gambiae 4ARR (MR4 strain) adult mosquitoes were maintained in 20x20x20 cm gauze cages at 28°C, 80 ± 5% relative humidity, and a photo-scotophase of 12:12 light:dark. The mosquitoes had access to a 5% sucrose solution on a cotton pad. The larvae were reared in tap water on plastic trays and fed daily with Tetramin fish food. Pupae were collected daily and placed in adult cages for emergence. Adult mosquitoes were regularly fed on a blood meal that contained 1:1 human erythrocytes and normal human serum (NHS) using the glass membrane feeder for maintaining the mosquito rearing cycle.
Cold-anesthetized mosquitoes were dissected over a sterile glass slide containing a drop of PBS under a stereomicroscope. ACs (proventriculus, anterior and posterior midgut) or midguts (anterior and posterior midgut) were isolated and placed in PBS, LB-medium, PBS-EDTA or in a commercial ELISA dilution buffer.
To measure the extent of complement activation in the mosquito midguts post blood feeding (PBF), A. stephensi mosquitoes were allowed to feed on a human volunteer arm and midguts were dissected from a pair of mosquitoes at 10, 30 and 90 min PBF. Each isolated pair of mosquitoes was placed in 200 μl specimen diluent from the MicroVue C3a Plus EIA kit, macerated for 30 sec with the tip of 21G sterile needle and centrifuged at 16,000g for 5 min at 4°C. Supernatants were immediately frozen at -70°C until processing. The average serum content in the midgut of a BF mosquito was estimated to be about 1 μl, this value was used when estimating specimen dilution for C3a and C5b-9 ELISAs. C3a concentration (ng/ml) was measured using the MicroVue C3a Plus EIA kit according the manufacturer’s instructions. C5b-9 concentration (AU/ml) was measured as previously described[8].
To study the effect of complement in the mosquito midgut on bacterial counts (viability) PBF, A. stephensi mosquitoes were allowed to feed on a blood meal that contained 1:1 human erythrocytes and either NHS or HIS. At 30, 90 and 24 h PBF, 5 mosquitoes representing each condition and time point were dissected and the isolated midguts were placed in 1 ml of LB-medium. The midguts were then macerated with 21G sterile needles and the sample tubes were vortexed for 30 sec. The standard plate counting method that involves serial dilutions, plating and counting of bacterial colonies was used under aerobic conditions to determine the number of live bacteria per midgut. Midguts from 5 sugar-fed mosquitoes were treated similarly to estimate the initial microbial load before blood feeding. A bacterial colony morphologically representing the majority of the colonies growing from either feeding on NHS or HIS based blood meal was isolated and analyzed for serum sensitivity, when incubated in 20% NHS or HIS for 1 h as previously described[9]. The identities of both bacterial strains were assessed by 16S ribosomal DNA PCR as previously described[10].
A. stephensi and A. gambiae mosquitoes were allowed to feed on a human volunteer arm. A. stephensi female mosquitoes were also fed on a blood meal that contained normal mouse serum (containing active complement) and mouse erythrocytes in a membrane glass feeder and blood-fed mosquitoes were left to rest for 2 h PBF. Blood- and sugar-fed (BF and SF) mosquitoes were then cold-anesthetized and placed in PBS, pH 7.4, containing 4% paraformaldehyde for 2 h followed by the dehydration and the clearing steps prior to embedding in paraffin wax. Each block was cut into 4-μm thick sagittal sections and representative slides were stained with HE to check the quality of the mosquito sections. Prior to immunofluorescence staining, the sections were dewaxed in xylene 3x5 min, rehydrated in a series of ethanol baths (100, 95 and 70%) for 3 min each and rinsed twice in distilled H2O for 5 min each. Dewaxed slides made from BF and SF mosquitoes were then treated with 1% BSA-PBS for 30 min to prevent nonspecific protein binding. BSA-PBS was then replaced by 1:500 dilution of goat polyclonal anti-human FH (A312, Quidel, San Diego), rabbit polyclonal anti-human C3c (A0062, Dako), goat polyclonal anti-human C5 (A306, Quidel, San Diego) or murine monoclonal anti-human SC5b-9 (A239, Quidel, San Diego) antibodies. After a 1-hour incubation at room temperature (RT) in a humidified chamber, the slides were washed 3x10 min in PBS-0.05% Tween 20 (PBS-T). Binding of primary antibodies was then detected by overlying the slides with 1:1000 dilution of Alexa Fluor 488 Donkey anti-goat IgG, goat anti-rabbit IgG or goat anti-mouse IgG (Invitrogen), respectively. Cell nuclei were stained with 30 nM DAPI (4′,6-diamidino-2-phenylindole) included in the diluted secondary antibody. Following 1 h of incubation in a dark humidified chamber at RT, slides were washed 3x10 min in PBS-T. Cover slips were mounted on slides using Mowiol-based antifading medium and kept at 4°C for at least 30 min before examining with the Olympus BX51 fluorescence microscope. Images were captured using the Olympus DP70 camera with the help of DP controller software. Slides of sugar-fed mosquitoes were treated similarly and served as negative controls for binding of the primary antibodies to mosquito proteins in the immunofluorescence assays. For colocalizing bound FH with carbohydrates present on the surface of the midgut epithelium, Alexa Fluor 594 conjugated concanavalin-A (25 ug/ml) was combined with Alexa Fluor 488 Donkey anti-goat IgG that detects binding of the primary antibody to FH. Following 1 h of incubation in a dark humidified chamber at RT, slides were washed 3x5 min in PBS-T. Cover slips were mounted on slides using Mowiol-based antifading medium and kept at 4°C for at least 30 min before examining with Leica TCS CARS SP8 confocal microscope.
DNA fragmentation, possibly indicative of apoptosis or cell death, was studied using fluorometric TUNEL kit (TACS 2 TdT-Fluor In Situ Apoptosis Detection Kit, TREVIGEN, Gaithersberg). Briefly, A. stephensi mosquito sections on dewaxed slides were treated with proteinase K for 15 min at RT, washed 2x in deionized water for 2 min each and immersed in labeling buffer for 5 min. Sections were then covered with the labeling reaction containing TdT dNTP, Mn2+, TdT enzyme and labeling buffer and incubated at 37 C for 1h. Labeling reaction was then stopped in stop buffer for 5 min and the slides were washed 2x in deionized water for 5 min each at RT. Sections were then covered with Strep-Alexa 488 conjugate solution and incubated for 20 min at RT followed by washing 2x in PBS for 2 min each and mounted afterwards with Mowiol mounting medium. TUNEL-positive mosquito section slides were prepared by treating dewaxed slides with TACS-Nuclease to induce DNA fragmentation. After 30 min incubation at RT slides were washed 2x in PBS for 2 min each. Induced DNA fragmentation was then detected as previously described.
Blood meals that contained 1 mg/ml monoclonal anti-human FH antibody (131X) and 1:1.1 human erythrocytes and either NHS or HIS were pre-incubated at RT for 20 min prior to A. stephensi mosquito feeding. Mosquitoes were also fed on NHS and erythrocytes in the same ratio in a third experimental condition. Cold-anesthetized blood-fed mosquitoes were separated and placed in new cages. Dead mosquitoes were counted daily for 7 days PBF. On day 8, egg collectors were placed inside mosquito cages and the laid eggs were counted the following day. The average number of blood-fed mosquitoes from 3 biological repetitions of NHS+131X, HIS+131X and NHS test conditions was 31.8±11.3, 34.5±7.6 and 31.7±7.6, respectively. In a fourth experiment 10 blood-fed mosquitoes from each test condition were collected 2 h PBF and placed in 4% paraformaldehyde for 2 h followed by the dehydration and the clearing steps prior to embedding in paraffin wax. Sagittal sections were then prepared as described above and analyzed for the presence of apoptotic/dead cells using the TUNEL assay described above and fluorescence microscopy.
The kinetics of complement C3 activation/degradation in the mosquito midgut following blood feeding was assessed by Western blot analysis. A. stephensi mosquitoes were allowed to feed on a human volunteer arm and midguts were dissected at 10, 20, 30 and 90 min from a single mosquito for each time point. Individual midguts were collected in 200 μl of PBS+10 mM EDTA and macerated for 30 sec with 21G needle and centrifuged at 16,000g for 5 min at 4°C. Supernatants diluted 1:1 with a reducing SDS-PAGE sample buffer were run on 10% SDS-PAGE gels, and separated proteins were transferred to a nitrocellulose membrane and probed with 1:5000 dilution of rabbit polyclonal anti-human C3c. Binding of the primary antibody was visualized by incubating the membrane with 1:20,000 dilution of goat anti-rabbit IgG coupled to HRP (NEF812, Perkin Elmer life Sciences, Inc.). The blot was then developed by the enhanced chemiluminescent (ECL) Western blot analysis system-based detection according to the manufacturer’s instructions (GE Healthcare, Life Sciences, Bucks, UK) followed by exposure to Super RX film (Fujifilm). Possible deposition of C3b or the presence of its inactivation products, iC3b or C3d, on the luminal surface of the proventriculus and the midgut epithelium was similarly investigated using ACs isolated from 6 mosquitoes that were fed on NHS+human erythrocytes using the glass membrane feeder. ACs were macerated in PBS+protease inhibitor cocktail (Roche) using a 21G needle and washed 4x with PBS-T that contained the protease inhibitor cocktail. After a wash with PBS the AC tissues were pelleted by centrifugation at 16,000g for 5 min at 4°C. Proteins were then extracted with 90 μl 1x reducing SDS-PAGE sample buffer and a volume equivalent to 1 AC was run on a gel for C3b, iC3b, C3c or C3d detection. An equivalent number of mosquitoes that were normally fed on 5% sucrose was treated in parallel and served as a control. Binding of FH to the AC epithelium was assessed similarly using the former sample set with 1:5000 dilution of goat polyclonal anti-human FH as the primary WB antibody and a 1:20,000 dilution of rabbit anti-goat coupled to HRP (Dako, P0160) as the secondary antibody. Loading control lanes were included in WB analysis, when appropriate. They were probed with the MRA-258 monoclonal antibody (MR4, the BEI Resources Repository, NIAID, NIH) that recognizes ≈150 kDa protein in the mosquito midgut extract.
Possible FH binding protein(s) in the midgut membrane extract were detected by using the ligand blot analysis[11]. Briefly, membrane proteins were extracted by incubating the A. stephensi mosquito midguts in the membrane extraction buffer (containing the protein inhibitor cocktail) from the ProteoJET membrane protein extraction kit (Fermentas, K0321) for 2 h at 4°C with constant shaking. Extracted membrane proteins were recovered by centrifugation at 16,000g for 15 min at 4°C. A volume equivalent to 1 midgut per lane was then loaded onto 10% SDS-PAGE gel under non-reducing conditions. Resolved membrane proteins were transferred to nitrocellulose membranes and overlaid after a blocking step with 3% non-fat milk for 1h with either 20% HIS (a source of FH) or PBS (a negative control for FH binding) and incubated overnight at 4°C. After 5 washes in PBS-T, the membranes overlaid earlier with HIS were further incubated in either PBS (a negative control for binding of the secondary antibody) or with 2 μg/ml of the monoclonal anti-human FH antibody (131X). After a 1-hour incubation at RT and 5 washes in PBS-T, the membranes were incubated in a 1:10,000 dilution of goat anti-mouse IgG coupled to HRP (NFH822, Perkin Elmer life Sciences, Inc.). After a 1-hour incubation at RT and 5 washes in PBS-T, the blots were developed as described above.
Data were analyzed with JMP 11 (SAS Institute Inc.). Paired Student’s t-test was used to calculate the statistical significance of complement activation on the bacterial load in the mosquitoes midgut and the in-vitro serum sensitivity of midgut bacteria and of blocking of FH activity with antibodies on egg count. To calculate the statistical significance of blocking FH activity with antibodies on mosquito survival, life tables were constructed for each experimental condition, and survival curves were analyzed by using the Kaplan-Meier log-rank analysis. Asterisks in figures indicate the different p values: *, p < 0.05; **, p < 0.01; ***, p < 0.001. All experiments were repeated at least three times. Results from different repetitions of experiments were pooled together and are presented as the mean and, when appropriate, representative images are shown. In all figures, error bars depict standard errors of the mean.
In the current study we tested whether components of the complement system in the blood ingested by Anopheles mosquitoes could injure midgut cells, and if not so, how do the midgut cells manage to survive the potent cytotoxic activity of complement? We first tested whether the human complement system in the ingested blood becomes activated in the mosquito midgut. Therefore, A. stephensi mosquitoes were allowed to feed on a human volunteer arm. This was followed by measuring the concentrations of the anaphylatoxin C3a and the soluble terminal complement complex (TCC, also known as SC5b-9) as indicators of complement activation. For this, the blood bolus serum samples were isolated from BF mosquito midguts. As shown in Fig. 1A the complement system became strongly activated in the mosquito midgut. The peak level of activation was found 10 min PBF as shown by a 100-fold increase in the concentration of C3a, when compared to the basal level (zero time) in the human volunteer serum (Fig. 1A). Moreover, the C3a level dropped to only 7-fold at 30 min PBF and reached the basal level at 90 min PBF (Fig. 1A). A more drastic indication of complement activation was observed, when the SC5b-9 level was measured in the bolus serum samples. The SC5b-9 concentration increased to 150-fold at 30 min PBF, when compared to the basal level in the volunteer serum sample (Fig. 1A). SC5b-9 level went down to 100-fold at 90 min PBF (Fig. 1A). Both C3a and SC5b-9 measurements in blood bolus serum samples indicated that complement was strongly activated in the mosquito midgut. The different kinetics of the two complement activation products are expected because they are raised at different steps of complement activation and, as a multimolecular complex SC5b-9 takes a longer time to form. The clearance rates of C3a and SC5b-9 are also different, C3a being more rapidly cleared.
To support the assumption that C3a was produced as a result of complement activation we analyzed whether C3 was converted into C3b and further to iC3b in the same sample set by Western blotting (WB) analysis. The WB data showed that more than 50% of the initial C3 had become converted to iC3b in the mosquito midgut 60 min PBF (Fig. 1B). This indicated that C3 had become activated to C3a and C3b and C3b was further inactivated to iC3b. Consequently, we also measured the residual complement activity that remained in the mosquito midgut at 10, 30 and 90 min PBF. The results showed a complete loss of the alternative and lectin pathway activities in just 10 min PBF (Fig. 1C). At the same time point about 74% of the classical pathway activity was lost. At 90 min approximately 20% of classical pathway activity could be still detected (Fig. 1C). Since the residual complement activity measurements for the classical pathway were based on quantifying the levels of SC5b-9 resulting from activating complement in the test samples, some active C3 must have been remained in the samples. Taken together, these results indicated that the complement system in human blood became activated in the fluid phase in the mosquito midgut.
As the initial experiments showed that complement became strongly activated in the midguts of mosquitoes that were fed naturally with human blood, we next asked whether this activation would impose any threat to cells in the mosquito midgut. As a surrogate marker for the ability of complement to kill cells in the mosquito midgut we chose first to study effects on the midgut microbiota. Here, A. stephensi mosquitoes were allowed to feed artificially either on normal human serum (containing functional complement) or heat-inactivated serum (containing non-functional complement) in addition to human erythrocytes under both conditions. Mosquitoes were dissected at 30 min, 90 min and 24 h PBF and the viable bacterial contents of the midguts were estimated by a plate colony counting procedure. As expected, feeding mosquitoes on NHS decreased the bacterial count in midguts to 34% and 17% of what was recorded for mosquitoes fed on sugar or HIS, respectively, when dissection took place at 30 min PBF (Fig. 2A). At the same time point bacterial count showed two-fold increase in midguts of HIS-fed mosquitoes compared to sugar-fed mosquitoes. This simply shows that feeding on blood containing active complement reduces bacterial count, while feeding on non-active complement containing blood meal increased midgut bacterial load. At 90 min PBF the bacterial counts in midguts of mosquitoes fed on NHS dropped to 5% relative to feeding on HIS (Fig. 2A). At 24 h PBF the difference between bacterial counts in NHS and HIS feeding conditions was not drastic, but remained significant as the decrease in bacterial count in NHS was only 37% of what was recorded in HIS (Fig. 2A). The increasing number of bacterial count at the 24 h time point in midguts of mosquitoes fed on NHS relative to the former time points suggested that serum resistant bacterial strains or just the remaining surviving bacteria might have overgrown. Therefore, morphologically dominant colonies from both NHS and HIS conditions were selected and tested for their ability to resist complement-mediated killing. The bacteria were identified to belong to the genus Enterobacter (E. cloacae) in both cases. Enterobacter isolates from both conditions were, found to be sensitive to complement and behave similarly in their response to complement-mediating killing (Fig. 2B). Therefore, the likeliest explanation for bacterial overgrowth in midguts of mosquitoes fed on NHS after 24 h was the lack of sufficient complement activity to kill all midgut bacteria. Altogether, this data clearly shows that complement activation in the mosquito midgut has a detrimental effect on living cells in the midgut lumen.
In our previous experiments complement in the mosquito midgut appeared to drastically reduce the bacterial load in the midgut PBF. Mosquitoes’ mortality as a likely indicator of midgut damage does not vary upon feeding on a blood meal containing functional or non-functional complement (common laboratory observation) Thus, we hypothesized that mosquito midgut epithelium has developed evasion mechanisms to escape the deleterious effect of complement activation. Acquisition of natural soluble complement regulators (Cregs) from blood by microbial surfaces is one of the most common mechanisms of complement evasion[12]. A similar mechanism that involves acquisition of Cregs by midgut epithelial surface could also occur when midgut is full of host blood. We tested this hypothesis by feeding A. gambiae and A. stephensi mosquitoes naturally on a human volunteer arm and looking for binding of Cregs and other complement components to mosquito gut epithelium. Sagittal sections of BF A. gambiae (Fig. 3A and 3F) and A. stephensi (Fig. 4E, 4G and 4F) mosquitoes were made 2 h post feeding. Coronal sections of A. stephensi (Fig. 4A, 4B, 4C and 4D) were also prepared. Immunofluorescence staining with antibodies against human FH, the soluble regulator of the alternative pathway of complement activation, revealed binding of FH to the epithelium surface of the proventriculus (Fig. 3B and 3C), the anterior midgut (Fig. 3G and 3H) and the anterior-posterior midgut (Fig. 3G and 3H). Sections made from another blood-fed Anopheles mosquito species, A. stephensi, showed identical results (Fig. 4) indicating that this phenomenon is not species-specific. The specificity of the anti-FH binding to its target was verified by immunofluorescence assays using sections of sugar-fed A. stephensi mosquitoes that showed no binding of anti-FH to mosquito proteins (Fig. 5). Additionally, binding specificity of the secondary antibody, Alexa Fluor 488 Donkey anti-goat IgG, was also verified in using sections of blood-fed A. stephensi mosquitoes in the absence of goat anti-FH from the IFA assays (Fig. 6B).
Parallel immunofluorescence assays did not show any sign of deposition of C3, C5 or MAC on the epithelial surface of the mosquito’s midgut suggesting protection from complement attack (S1 Fig).
To further confirm binding of FH to the mosquito midgut epithelium A. stephensi mosquitoes were allowed to feed on a volunteer arm and their ACs were dissected 2 h PBF. After washing the ACs SDS-soluble material was run on SDS-PAGE gel for detecting bound complement proteins by Western blot (WB) analysis. In agreement with the immunofluorescence microscopy observations FH was detected (Fig. 7A) among the proteins that were eluted from the AC tissues of BF mosquitoes. In contrast, anti-FH antibody did not detect any proteins in the eluted material from the midguts of SF mosquitoes. Furthermore, no signs of deposition of C3b or its inactivation products on the AC epithelium (Fig. 7B) were observed, when the same preparations were used in WB analysis and probed with an antibody that detects human C3, C3b, iC3b and C3c. The functional activity of FH in the mosquito midgut was also assessed by analyzing the soluble content of the midguts of BF mosquitoes. WB analysis revealed FH-dependent degradation of C3b by FI into iC3b (Fig. 7C) in the midgut content that was collected 2 h PBF. iC3b was apparent by the presence of 46- and 43-kDa fragment bands of the processed α’ chain of C3b. The WB analysis also showed that the 68-kDa fragment of iC3b (Fig. 7C, lane 2) was completely cleaved into smaller fragments (Fig. 7C, lane 1) suggesting either further cleavage of iC3b to C3c+C3d,g by erythrocyte CR1 and factor I or the involvement of midgut proteolytic activity in the complement inactivation process in the midgut.
Altogether, our data indicates that the surfaces of the proventriculus and the midgut capture FH to inhibit C3 deposition that otherwise could lead to lysis of the target cells. Moreover, complement activation in the soluble phase in the blood bolus seemed to be efficiently controlled by factor I-mediated C3b degradation and midgut proteolytic activity.
Mouse serum containing active complement was also used to blood-feed A. stephensi mosquitoes to test for the ability of midgut epithelium to capture mouse FH from serum. Interestingly, immunofluorescence assays using goat anti-human FH that cross-reacts with mouse FH detected mouse FH on the surface of the midgut epithelium (Fig. 6A). FH signal was also more profound on the epithelium of the proventriculus and the anterior midgut as in the case of human FH binding to A. gambiae and A. stephensi midgut anterior midgut epithelium. Altogether, these data suggest that capturing FH by mosquito midgut epithelium from a blood meal could be a common mechanism utilized by blood-feeding mosquito species regardless of the host species.
To further characterize FH binding to the midgut epithelium, 2 h PBF A. stephensi mosquito sections were double stained for FH and Concanavalin A (ConA) binding to the midgut surface. Confocal microscopy images showed a thick glycocalyx layer covering the outer surface of luminal plasma membrane represented by the red fluorescence of ConA-Alexa 594 (Fig. 8E and H). FH was found to colocalize with approximately one quarter of the luminal side of the glycocalyx layer (Fig. 8I). FH binding also formed a thin layer lining the luminal surface of the glycocalyx (Fig. 8C, F and I). Most of bound FH also appeared to colocalize with the glyococalyx. This data shows the association of FH binding to the mosquito midgut glycocalyx layer confirming binding of FH to a protein or a glycoprotein integrated into the plasma membranes of midgut epithelial cells.
Binding of FH to the epithelial surface of the various compartments of the mosquito AC and the absence of C3b deposition prompted us to ask whether interfering with this binding would result in detrimental consequences. Therefore, A. stephensi mosquitoes were fed on blood meals that contained NHS alone or the monoclonal antibody, 131X, which functionally inactivates FH and either NHS or HIS. Mosquito mortality was first recorded at 48 h PBF and increased to 7% in the NHS/anti-FH-fed mosquitoes as compared with 1% and 2% in the NHS- and HIS/anti-FH-fed mosquitoes, respectively (Fig. 9). Mortality increased to 18% and 3% within 72 h in NHS/anti-FH- and HIS/anti-FH-fed mosquitoes, respectively. Mortality remained unchanged in the case of NHS-fed mosquitoes at the same time point. Highest mortality rates were observed at day 5 PBF in NHS/anti-FH, NHS- and HIS/anti-FH-fed mosquitoes being 22%, 5% and 6%, respectively (Fig. 9). No increase in mortality was reported until the last observation at day 7. The overall mortality was significantly different (Kaplan-Meier log-rank test, p<0.001) between NHS/anti-FH-and either NHS- or HIS/anti-FH-fed mosquitoes. To determine the effect of blocking FH activity in the midgut on mosquito fecundity, mosquito eggs were collected at day 8 PBF from NHS-, NHS/anti-FH- and HIS/anti-FH-fed surviving mosquitoes. The average number of eggs per mosquito was reduced to 28±17 eggs in the NHS/anti-FH-fed mosquitoes as compared with 44±15 and 48±27 eggs in the NHS- and HIS/anti-FH-fed mosquitoes, respectively (Fig. 9). From a subsequent identical experiment sagittal sections of mosquitoes were prepared 2 h PBF. Sections from six mosquitoes were then analyzed for the presence of apoptotic cells as a result of neutralizing FH with the 131X monoclonal antibody. TUNEL assays showed signs of cell death presented as green fluorescent nuclei (Fig. 10E and F) in midgut epithelium in one mosquito out of six in the 131X-neutralized FH group (131X+NHS group). No signs of cell death were detected in the absence of the 131X anti-FH antibody in sections of seven mosquitoes (Fig. 10B and C) comprising the control group (NHS group). Altogether, these results suggest that blocking FH-mediated complement evasion in the mosquitos has deleterious effect on the mosquito’s survival and the fecundity of those mosquitoes that managed to survive complement-mediated damage.
To shed light on the interaction of FH with mosquito midgut proteins, we initiated a search for the potential FH receptor(s). The ACs of A. stephensi mosquitoes that had never fed on blood were dissected and membrane proteins were extracted. The extracted proteins were run onto SDS-PAGE gel under non-reducing conditions and transferred to nitrocellulose sheets for ligand blot analysis to detect binding of FH to mosquito proteins in vitro. As shown in Fig. 11 two mosquito proteins of approximately 40 and 100 kDa were found to be the candidate receptors for binding FH.
In this study we observed binding of the major complement inhibitor factor H to the alimentary canal epithelium of the Anopheles mosquitoes. Apparently, this interaction protects the mosquito’s epithelium from complement-mediated damage and could provide a target for a transmission blocking vaccine-induced immunity. On the other hand, we found that complement becomes activated in the fluid phase in the mosquito midgut. As a consequence, complement activation was able to significantly reduce the number of the bystander bacteria in the midgut, a phenomenon that could help Plasmodium parasites to survive in the mosquito midgut. Further experiments showed that blocking of FH activity by a monoclonal antibody that interferes with its function and boosts complement activation had a detrimental effect on mosquito survival and fecundity.
Earlier studies, with a particular interest in the resistance of Plasmodium gametes to complement activity, reported long-lasting persistence of complement components or activity in the mosquito midgut post blood feeding. Margos et al [13] observed that rat complement components necessary to initiate the alternative pathway (factor B, factor D, and C3) as well as C5 were present for several hours following the ingestion of P. berghei-infected rat blood. In a more recent study, Simon et al [14] have shown that the concentration (OD values) of C3a, a marker for complement activation, was 2.3-fold higher in mosquito midgut at 1 h PBF (artificial feeding) compared to control. In the current study, mosquito’s natural feeding on a human volunteer arm showed that complement became strongly activated in the mosquito midgut. C3a concentration (ng/ml) was 100-fold higher within 10 min PBF and dropped to 7-fold at 30 min PBF compared to the basal level in the human volunteer serum. The sharp decrease in C3a level was probably due to its rapid binding to receptors and/or to anionic surfaces because of its cationic nature. Leukocytes present in the blood meal carry receptors for C3a (C3aR). SC5b-9 is a more stable complement activation product than C3a and the result of the full activation sequence. Therefore, the SC5b-9 level peaked after the C3a peak, as expected from the kinetics of complement activation. This was followed by a slight decrease in the SC5b-9 level indicating that no more activation was taking place at 90 min PBF. The presence of a considerable amount of C3 in the mosquito midgut at 90 minutes PBF suggested that some C3 activity was still left. The residual activity assays, however, showed complete loss of the alternative pathway activity at 10 min PBF, whereas, about 20% of the classical pathway activity still remained at 90 min PBF in an assay format that detects the ability to generate SC5b-9 complexes. This finding suggested that the classical and the terminal pathway components were partially active and functional C3 was involved in the measured residual activity. In contrast, in the alternative pathway a rate limiting factor was consumed or, alternatively, the pathway had become specifically inhibited. An earlier study showed that soluble molecules from the midgut of the mosquito Aedes aegypti inhibited complement activation and C3b deposition in vitro via the alternative pathway by 52% and via the classical pathway by only 24%[15]. Altogether, the current data suggests that some mechanisms of complement inhibition operate in the mosquito midgut, and that while complement becomes activated for some time in the fluid phase the mosquitoes do not seem to suffer from complement attack.
The potential ability of complement in the mosquito midgut to cause damage to cells was recognized from the significant decrease in the number of viable bacteria post mosquito feeding on NHS-supplemented blood as compared with feeding on HIS-supplemented blood. Complement-mediated killing of the midgut bacteria was taken as an indirect indication that this could have been the fate of the midgut epithelium unless there was an evasion mechanism that allowed the epithelial cells to escape complement-mediated damage. Complement-mediated killing of midgut bacteria could also lead to consumption of complement components that could otherwise target the midgut epithelium. The lack of damage to mosquito cells was supported by the absence of nuclear chromatin fragmentation, a sign of cell death, in the midgut epithelial cells. In agreement with our postulation, a previous study in which the bug Triatoma brasiliensis was fed by the forced feeding procedure to bypass mixing blood with the saliva that contains a complement inhibitor showed significant signs of midgut cell death as early as 1 h PBF [15]. This finding supports a possible role for complement inhibitors of hematophagous vectors in protecting midguts from host complement-mediated injury. Additionally, mosquitoes fed on mice immunized with mosquito ACs homogenate have shown a significant reduction in mosquito survival compared to control groups [16]. A similar study has also shown impairment of the development of Plasmodium spp. inside the Anopheles spp. midgut and, thereby, a block in parasite transmission [17]. It is likely that mosquito killing in these studies was mediated by antibody-mediated activation of the classical complement pathway, for which less inhibitory activity was found to operate in the mosquito midgut in our study. On the other hand, antibodies neutralizing an inhibitor of the alternative pathway would enhance complement killing of the midgut cells. In addition to complement-mediated damage, mosquito killing could also be due to interference of anti-AC antibodies with vital cell functions on the mosquito midgut surface.
The observed drastic complement-mediated reduction of bacterial count in the mosquito midgut also argues for a special attention of studies focusing on understanding the association of midgut microbiota with the malaria parasite survival in the mosquito midgut [18,19]. Natural blood feeding could thus greatly reduce the number of midgut bacteria by complement-mediated killing, as opposed to artificial feeding that is usually based on heat-inactivated serum. Shaping the diversity of gut microflora of hematophagous species was also reported for the Hirudo medicinalis, a medicinal leech, as active complement in the ingested blood limited the gut microflora to only the complement-resistant bacterial strains[20].
Using an indirect immunofluorescence assay FH was detected to bind to the mosquito proventriculus and midgut. C3, C5 and MAC were absent from the ACs epithelial surface indicating the protective activity of the cell-bound FH. Factor H binding is thus a potential complement evasion mechanism developed by the mosquitoes to protect their ACs from complement-mediated damage. Furthermore, a monoclonal antibody directed against FH that is known to inhibit its function significantly increased the mortality rate in blood-fed mosquitoes within 72 h from about 1% in NHS-supplemented blood meal to 18% when NHS was supplemented with anti-FH. The relatively low level of mortality, albeit significant, in the presence of the anti-FH monoclonal antibody could have resulted from insufficient quantities of the monoclonal antibody, rapid consumption of active complement or mosquito inhibitors of complement activation. Complement can become spontaneously activated, and this process is accelerated in the presence of an anti-factor H antibody.
In our attempts to identify mosquito’s FH-binding molecules we confirmed this interaction by the ligand blotting assays and detected two molecules with molecular masses of about 100 and 40 kDa as potential FH binding molecules. The identities of these molecules are the subject of ongoing work. It is also possible that, in addition to specific proteins, mosquito cell surface glycosaminoglycans or other polyanions could play a role in FH binding to mosquito AC cell surfaces. Exploiting soluble complement regulators, particularly FH, from the host by many pathogenic bacteria has been well described [5]. Factor H binding has been reported also to fungi, viruses and parasites [5,14]. The current study is to our knowledge the first one to describe the utilization of a similar complement evasion strategy by a hematophagous vector. Previous reports on complement evasion or inhibition by hematophagous vectors were almost exclusively from studies on ticks. Those studies identified other unique strategies to block complement activation. For example, a salivary protein, OmCI, from the Ornithodoros moubata tick, the vector of human relapsing fever caused by Borrelia duttoni, was shown to specifically bind and inhibit C5, thereby preventing activation of the terminal complement pathway[21]. Another salivary protein, TSLPI from Ixodes scapularis tick, vector of the Lyme disease caused by Borrelia burgdorferi, was shown to interfere with the lectin pathway of complement activation by preventing MBL binding to its ligand[22]. TSLPI was also shown to be beneficial to B. burgdorferi as it resulted in impaired neutrophil phagocytosis and chemotaxis and reduced killing of Borrelia [22]. A third strategy that was shown to be utilized by ticks to block complement activation was binding of tick salivary proteins such as Isac, Irac-1 and-2, and Salp20 to properdin and displacing it from the alternative pathway C3 convertase resulting in its inhibition[23–25]. So far, no analogous anticomplement strategies have been identified in mosquitoes, except the one in the current study. However, mosquito antihemostatic molecules such as anophelin from Anopheles spp. [26,27], anti-factor Xa from Aedes aegypti[28] and alboserpin from Aedes albopictus[29] that target human coagulation system were already identified in mosquito’s saliva.
Molecules involved in complement inhibition from hematophagous vectors are of potential interest to generate an anti-vector vaccine that could interfere with the lifespan of the disease vectors or with infectivity of the pathogen. Studies to envisage the possible use of arthropod vector proteins as anti-vector vaccine already exist[30], but none of them have attempted to use a complement regulator-neutralizing vaccine so far. Thus, mosquito midgut antigens involved in essential biological processes such as in complement inhibition would serve as more specific candidates for similar studies. Moreover, identifying complement inhibitors from hematophagous vectors could also provide pharmacological agents to treat diseases, where complement activation is known to play a role[31].
In conclusion, we have shown in this study that the complement system becomes immediately activated in the mosquito after ingestion of human blood while, at the same time, the mosquito AC surface molecules captured FH from the blood meal and inhibited the deposition of C3b on the midgut epithelium. The initial complement activation that occurred in the blood bolus in the midgut was able to kill midgut bacteria that were not resistant to complement. On the other hand, acquisition of FH by the midgut epithelial cells contributed to mosquito’s survival against the innate immune system in the ingested blood meal. Interfering with the complement regulatory activity of FH in the mosquito midgut increased mosquito mortality and reduced fecundity. The putative Anopheles mosquito FH binding proteins could be transmission blocking vaccine candidates targeting the malaria parasite carrying vectors.
|
10.1371/journal.pgen.1006088 | Msa1 and Msa2 Modulate G1-Specific Transcription to Promote G1 Arrest and the Transition to Quiescence in Budding Yeast | Yeast that naturally exhaust their glucose source can enter a quiescent state that is characterized by reduced cell size, and high cell density, stress tolerance and longevity. The transition to quiescence involves highly asymmetric cell divisions, dramatic reprogramming of transcription and global changes in chromatin structure and chromosome topology. Cells enter quiescence from G1 and we find that there is a positive correlation between the length of G1 and the yield of quiescent cells. The Swi4 and Swi6 transcription factors, which form the SBF transcription complex and promote the G1 to S transition in cycling cells, are also critical for the transition to quiescence. Swi6 forms a second complex with Mbp1 (MBF), which is not required for quiescence. These are the functional analogues of the E2F complexes of higher eukaryotes. Loss of the RB analogue, Whi5, and the related protein Srl3/Whi7, delays G1 arrest, but it also delays recovery from quiescence. Two MBF- and SBF-Associated proteins have been identified that have little effect on SBF or MBF activity in cycling cells. We show that these two related proteins, Msa1 and Msa2, are specifically required for the transition to quiescence. Like the E2F complexes that are quiescence-specific, Msa1 and Msa2 are required to repress the transcription of many SBF target genes, including SWI4, the CLN2 cyclin and histones, specifically after glucose is exhausted from the media. They also activate transcription of many MBF target genes. msa1msa2 cells fail to G1 arrest and rapidly lose viability upon glucose exhaustion. msa1msa2 mutants that survive this transition are very large, but they attain the same thermo-tolerance and longevity of wild type quiescent cells. This indicates that Msa1 and Msa2 are required for successful transition to quiescence, but not for the maintenance of that state.
| In spite of the many differences between yeast and humans, the basic strategies that regulate the cell division cycle are fundamentally conserved. In this study, we extend these parallels to include a common strategy by which cells transition from proliferation to quiescence. The decision to divide is made in the G1 phase of the cell cycle. During G1, the genes that drive DNA replication are repressed by the E2F/RB complex. When a signal to divide is received, RB is removed and the complex is activated. When cells commit to a long term, but reversible G1 arrest, or quiescence, they express a novel E2F/RB-like complex, which promotes and maintains a stable repressive state. Yeast cells contain a functional analog of E2F/RB, called SBF/Whi5, which is activated by a similar mechanism in proliferating yeast cells. In this study, we identify two novel components of the SBF/Whi5 complex whose activity is specific to the transition to quiescence. These factors, Msa1 and Msa2, repress SBF targets and are required for the long term, but reversible G1 arrest that is critical for achieving a quiescent state.
| The need to stop proliferation and remain in a protected quiescent state is universally conserved and is just as important to yeast as it is to human cells. Failure to enter, or unscheduled exit from quiescence results in uncontrolled proliferation and cancer in humans, and death in unicellular organisms [1]. Most cells enter quiescence from G1. As such, there must be regulators in G1 cells capable of recognizing stop signals when they arise and provoking a stable but reversible halt to S phase. The regulatory strategy that controls the G1 to S transition in cycling cells is well understood and its basic framework is highly conserved from yeast to humans [2]. Studies of yeast have provided many insights into this process, but little is known about the cell cycle regulators that give rise to quiescent yeast cells. We have identified a pair of related transcription factors that play a critical role in halting the cell cycle in G1, specifically during the transition to quiescence. Like the highly conserved quiescence-specific complexes of higher eukaryotes [3–5], these factors repress transcripts that promote the G1 to S transition and enable yeast cells to enter the quiescent state.
In rapidly growing yeast cells, as in higher cells, the G1 to S transition is tightly controlled by two consecutive waves of cyclin expression. Cln3 is expressed at the M/G1 boundary and initiates the transition by binding and activating the cyclin-dependent kinase (Cdk). The critical target of Cln3/Cdk is Whi5, which represses SBF. SBF is a transcription factor complex that includes Swi6 and its DNA binding partner Swi4. Cln3 phosphorylates and releases Whi5 from the complex, thus enabling SBF to activate late G1-specific transcription of the G1 cyclins CLN1 and CLN2 and other genes that promote the G1 to S transition [6–8]. The G1 cyclin/Cdk complexes then phosphorylate Sic1 and target it for degradation. Once Sic1 is degraded, the B type cyclin/Cdk complexes that are bound and inhibited by Sic1 are released, allowing them to phosphorylate and activate the DNA replication machinery and S phase ensues.
Swi6 also associates with a second DNA binding protein, Mbp1, which is related to Swi4 and binds to a similar but distinct DNA sequence [9]. This complex, referred to as MBF, also confers late-G1 specific transcription on many genes involved in DNA replication and repair. These genes are regulated by Nrm1-dependent negative feedback [10]. Nrm1, itself a late-G1 transcript, accumulates in S phase, binds MBF complexes and represses transcription through S and G2/M.
This wave of late G1 transcription is critical for the timing and fidelity of DNA replication. If the G1 to S transition is accelerated by ectopic expression of Swi4 or the G1 cyclin Cln2, there are checkpoint proteins, including Mec1 and Rad53, that detect replication stress and become essential for delaying S phase and promoting DNA repair [11, 12]. This is in part accomplished by the direct phosphorylation of Nrm1 by Rad53, which releases it from the MBF complex and allows DNA replication and repair genes to be activated [13].
The transition from logarithmic growth to quiescence involves a stable but reversible cell cycle arrest in G1. Our previous studies have shown that this transition begins with a lengthening of G1, which is initiated before the diauxic shift (DS), when all the glucose has been taken up from the media [14]. The cell divisions that follow are highly asymmetric and the physical growth of those cells slows, resulting in a dramatic shift in the cell size of the population [15]. To explore the mechanism of this stable but reversible G1 arrest associated with quiescence, we have assessed the roles of known regulators of the G1 to S transition in rapidly growing cells. In wild type cells, Rad53 plays a role in the transition to quiescence, and it becomes essential if the G1 to S transition is driven by Cln3 overproduction. A second checkpoint gene, Rad9, is not required during this transition. Rad9 responds to DNA damage, while Rad53 responds to both DNA damage and replicative stress. This indicates that the latter, replicative stress, is the likely signal for Rad53 activation during the transition to quiescence [14]. The G1 arrest is maintained in post-diauxic cells by Xbp1, which is induced to high levels and represses CLN3 along with over 800 other genes [14]. Xbp1 recruits the histone deacetylase, Rpd3, which plays a unique and prominent role in the transcriptional repression that takes place in quiescent cells [16]. Rpd3 is targeted to at least half the budding yeast promoters, where it affects global nucleosome repositioning, histone deacetylation and a 30-fold global repression of transcription [16].
In this paper, we report the role of two proteins, Msa1 and Msa2 in the early transcriptional regulation that promotes G1 arrest and the transition to quiescence. Msa1 and Msa2 are two related proteins that were identified by mass spectrometry to be associated with SBF and MBF complexes [17]. Mutations in these proteins have mild phenotypes in rapidly growing cells [17–19], but we find that Msa1 and Msa2 are both important during the transition to quiescence. Each single mutant survives this transition, but the msa1msa2 double mutant fails to G1 arrest and loses viability rapidly. When paired with rad53-21, we observe a G1 arrest defect with the single mutants, especially msa2. These data indicate that Msa1 and Msa2 are both important regulators that promote G1 arrest during the transition to quiescence and cells rely on the Rad53 checkpoint function when either protein is missing. We have carried out RNA deep sequencing with the single and double msa mutants as cells transition from log phase to quiescence and find that they have significant impact on the expression of both MBF and SBF target genes, specifically in post-diauxic cells. In many cases, both Msa1 and Msa2 are required to repress SBF targets and activate MBF targets, and their effects are not additive. This suggests that they both play critical roles in the regulation of these late G1-specific transcripts. In other contexts, either Msa protein is sufficient to perform their regulatory function in post-DS cells. Chromatin immunoprecipitation (ChIP) of candidate targets show binding of both Msa1 and Msa2 to their targets, and show stronger binding in post-DS cells. The post-diauxic regulation of these genes by Msa1 and Msa2 is likely to be important for a normal transition from proliferation to quiescence.
Quiescent cells can be purified from stationary phase cultures due to their high density [20]. These quiescent (Q) cells, by definition, are in a stable, long-lived, but reversible arrest, and they have G1 DNA content, suggesting that they exit the cell cycle from G1. It follows that regulators that promote the G1 to S transition, might disrupt entry into quiescence and those that prolong G1 might facilitate it. This is clearly true of the activator Cln3, which in excess reduces Q cell yield and when absent increases Q cell yield [14] and Table 1).
We have carried out a survey of mutants that affect the length of G1 and find that there is a good correlation between the length of G1 during logarithmic growth and quiescent cell yield (Fig 1A and Table 1). The longer cells spend in G1, the more efficient is their transition into the quiescent state. This is consistent with the view that cells normally enter quiescence from the G1 phase of the cell cycle. This survey also shows that most of the known regulators of the G1 to S transition during rapid growth are not required for Q cell formation.
The SBF transcription factor drives the transcription of many genes in late G1, which play important activatory roles in the G1 to S transition. Hence, it would be a likely target of negative regulation as cells enter the stable G1 arrest associated with quiescence. If so, loss of Swi4 or Swi6 activity might promote the transition to quiescence as we’ve seen with loss of Cln3 (Table 1). However, we find that both swi4 and swi6 mutants suffer significant loss of viability as they are grown from logarithmic (log) phase to stationary phase (SP). Though not considered an essential gene, swi6 mutants grow very slowly (Fig 1B). Only 60% of the cells are viable during the logarithmic phase of growth and that drops to 20% after seven days (Fig 1C). SWI4 is an essential gene in the W303 strain, but that lethality is suppressed by SSD1 [21]. SSD1swi4 cells show a similar slow growth and loss of viability pattern. Both swi6 and swi4 mutants undergo the diauxic shift (DS) late, at about half the cell number that their respective wild type cells undergo the DS (Fig 1B). However, the optical densities of these mutant cultures are comparable to wild type at the DS (OD550 5 to 6), indicating that cell mass is the key variable for the timing of this transition. Both mutants are larger and more heterogeneous than wild type cells based on light scattering (S1 Fig). Dead cells predominate, based on dye exclusion (Fig 1C) and the accumulation of cell debris on the left margin of their flow cytometry profiles (S1 Fig). These data indicate that normal growth control in response to nutrient limitation requires the activities of both of these key regulators of G1.
Wild type cells that have entered quiescence can be purified based on their density in percoll gradients [20]. No such high density cells can be purified from swi6 cultures (yellow dot in Fig 1A and 1D). This indicates that Swi6 is critical for the transition to quiescence. In contrast, about half the swi4 SSD1 cells become dense, but these cells are three times the size of wild type Q cells (Fig 1E) and they include both live and dead cells (Fig 1F). The live high density swi4 cells suffer a further three-fold loss of viability over the course of an 80 day incubation in water compared to wild type cells which drop very little (Fig 1F). The high density swi4 cells also recover very slowly upon re-feeding (Fig 1G). We conclude that the longevity of the dense swi4 cells is compromised. By all these criteria, swi6 and swi4 mutants are defective in both the log phase of growth and the transition into and out of quiescence.
We have also assayed mutants of other known components of the G1 transcription complexes. The DNA binding component of MBF, Mbp1, is not required for G1 arrest (S1 Fig), viability (Fig 1B and 1C) or for Q cell production (Table 1). Stb1, a component of both SBF and MBF [22–25], is also not required for Q cell production (Table 1). Cells lacking Whi5, which binds and inhibits SBF [6, 7, 25], undergo more cell divisions than wild type (Fig 2A) and they significantly delay, but finally achieve 80% G1 arrest after 48 hours of growth (Fig 2B). The whi5 mutant produces almost wild type levels of Q cells (Fig 2C), and these cells are identical in size to wild type (Fig 2D). The whi5 Q cells also have a comparable, if not somewhat longer life span (Fig 2E). This indicates that this SBF repressor plays a role in achieving efficient G1 arrest, but it has no detectable role in the maintenance of quiescent cells. However whi5 Q cells show a 30 minute delay in recovery from the quiescent state (Fig 2F). This is the opposite of what is seen with G1 cells purified by elutriation, where whi5 accelerates the transition to S phase and produces smaller cells [6, 7]. These observations suggest that the late G1-specific SBF transcription complex of Swi4 and Swi6 plays a critical role in the transition to quiescence, but that its regulation as cells enter and exit quiescence may involve novel partners other than Whi5.
The whole genome duplication that S. cerevisiae underwent [26] gave rise to a Whi5-related protein, which was originally identified as a high copy suppressor of rad53 lethality (SRL3 [27].) More recently, Srl3 was shown to bind to SBF in response to DNA damage [28] and to regulate the nuclear localization of Cln3 [29]. SRL3 transcription is induced by DNA damage and many other forms of stress [30–32]. This led us to determine whether Srl3 (also known as Whi7) plays a redundant role with Whi5 in the transition to quiescence. Fig 2 shows that loss of Srl3 causes a modest further delay of G1 arrest, but only when Whi5 is also missing. Its most striking phenotype is the delay of budding as srl3 and srl3whi5 Q cells re-enter the cell cycle upon re-feeding.
These observations led us to consider two other known components of SBF and MBF transcription complexes. Msa1 and Msa2 are also related proteins that arose from the whole genome duplication. They were initially found by tandem affinity purification and multidimensional protein identification technology (MudPIT) to be associated with both SBF and MBF [17]. Msa1 was also identified as a high copy suppressor of three temperature-sensitive DNA replication mutants [18]. Genome-wide transcript analyses of rapidly growing cells indicated that Msa1 has both an activating and a repressing role at a small and diverse set of target genes during the log phase of growth [18]. The fifty genes identified in that study that are both bound and regulated by Msa1 are mostly involved in glucose metabolism, cell wall organization or ribosomal structure. MSA1 is an ECB-driven transcript that peaks at the M/G1 boundary, like CLN3 [33, 34]. Msa1 binds to both SBF and MBF-regulated promoters and has a modest impact on the timing of late G1 transcription and budding in log phase cells [17]. This suggests that Msa1 performs an activatory function at these promoters during the log phase of growth. However, excess Msa1 leads cells to accumulate in G1 and S phase [35], suggesting that Msa1 either represses cell cycle progression directly, or that its presence in excess is activating the DNA damage or replication stress checkpoint. Msa1 also binds to Dbf4, which is the regulatory subunit of the Cdc7/Dbf4 kinase required for DNA replication and for activation of the replication stress checkpoint [36]. More recently, it was shown that the Hog1 kinase phosphorylates Msa1 during osmotic stress, and may play a role in delaying S phase under these conditions [19]. We have not observed a Hog1-dependent effect on the production of quiescent cells (Table 1).
The Msa2 protein sequence is highly conserved compared to that of Msa1. Despite the tight cross-species conservation of Msa2, almost nothing is known about its role in cells. Like, Msa1, it associates with SBF and MBF [17]. MSA2 is an MBF target [37], which is transcribed in late G1, and induced by DNA damage and other forms of stress [13]. Msa1 and Msa2 have also been found to form a distinct activatory complex with Ste12 and Tec1 on the FLO11 and MSB1 promoters [38]. These genes are involved in cell adhesion and pseudohyphal development and the msa1msa2 double mutant is adhesion-defective. This led the authors to conclude that the Msa proteins may play a role in coordinating cell division with development.
Our data indicate that both Msa1 and Msa2 are critical for cell division arrest and growth arrest as cells transition to quiescence. In prototrophic W303, loss of either Msa1 or Msa2 or both has no effect on the fraction of time the cells spend in G1 when they are growing logarithmically. The fraction of log phase cells that are in G1 in the single and double msa mutants is comparable to wild type (Table 1 and S1 Fig). However, as these cultures increase in cell number, the double mutant stops dividing at about half the cell density of wild type cultures (Fig 3A). With wild type cells, the percent of cells in G1 triples as they approach the diauxic shift [14] and Fig 3B). The single msa mutants are slightly delayed if at all in this response compared to wild type cells. In contrast, the msa1msa2 double mutant shows a slow accumulation of G1 cells that plateaus at 60% after 18 hours of growth (Fig 3B). After seven days of growth, the msa1msa2 cells show very low heterogeneous DNA content (Fig 4A) and most of the cells are dead (Fig 4B). Not surprisingly, the msa1msa2 Q cell yield is also low (red dot in Fig 1A).
The failure of the double mutant to arrest in G1 and its loss of viability over this time course suggests that Msa1 and Msa2 play redundant roles in halting cell cycle progression specifically during the transition to quiescence. However, in checkpoint-deficient cells, carrying the rad53-21 mutation [39], Msa1 and Msa2 are both required for efficient G1 arrest in response to nutrient consumption (Fig 3C) We have previously shown that the Rad53-mediated replication stress checkpoint plays a role during the transition to quiescence [14]. Just as in rapidly cycling cells [11, 12, 40], Rad53 function is essential for restraining cells in G1 and achieving quiescence when the transition to S is driven prematurely by excess Cln3 [14] and Table 1). If the Msa proteins are also important for G1 arrest, we expected that their absence would also exacerbate the rad53-21 phenotype, and this is exactly what we observe. As noted previously [14], checkpoint-deficient rad53-21 cells do not achieve the full G1 arrest observed with wild type cells after 48 hours of growth (Fig 3C). The additional loss of either Msa1 or Msa2 mutants has a more extreme phenotype. These double mutants are almost as defective in cell cycle arrest after 48 hours of growth as the msa1msa2 mutant (Fig 3C). This indicates that when Rad53 is not present to reinforce the arrest, Msa1 and Msa2 are both required to efficiently halt cell cycle progression. msa2rad53-21 has the most extreme defect. It is an unstable strain and we were unable to construct the msa1msa2rad53 triple mutant. However, in contrast to msa1msa2, most of the msa1rad53-21 and msa2rad53-21 cells achieve G1 arrest after seven days in culture (Fig 4A) and retain viability (Fig 4B).
Our previous work shows that shortly after the diauxic shift, wild type cells undergo a dramatic shift in cell size, due to a highly asymmetric cell division [15]. Fig 4C shows the cell size distribution of wild type cells during logarithmic growth compared to that of cells after the diauxic shift (18hr) and after seven days in culture. The asymmetric cell division of wild type cells gives rise to daughters that are about 14 femtoliters (fL) in volume. These cells slowly increase in volume to about 20 fL and never attain the 40–60 fL volume observed in log phase cultures. To see if the Msa proteins are required for this asymmetric cell division, we plotted the modal cell size of the single and double mutants as they grew from log to stationary phase (Fig 4D). All five strains undergo asymmetric cell division, but the msa2rad53-21 and msa1msa2 cells continue to enlarge. msa1rad53-21 has an intermediate phenotype. Interestingly, the msa2rad53-21 cells increase to the same size as msa1msa2 cells, but do not lose viability over this time course. They also produce nearly wild type levels of Q cells (Table 1). This is correlated with and may be explained by the fact that the majority of the msa1rad53 and msa2rad53 cells eventually attain G1 arrest after seven days of growth (Fig 4A). Despite their large size, the viable cells in G1 purify as Q cells from those seven day cultures (Figs 4A and 5D) We conclude that Msa1 and Msa2 both contribute to efficient cell division arrest and cell growth arrest as nutrients become limiting. Rad53 checkpoint function provides critical backup in restraining cell cycle progression and growth when either Msa protein is absent. However, Msa1 and Msa2 have critical overlapping roles in achieving full G1 arrest and maintaining viability during the transition to quiescence.
To see if the survivors of this transition achieve a protective quiescent state, we purified the high density cells from a seven day old culture of msa1, msa2 and msa1msa2. Fig 5A and 5B compare the size of these cells from the starting log phase culture and the purified high density Q cells. Both msa mutants are smaller than wild type cells during logarithmic growth, but the msa1msa2 cells are clearly larger and more heterogeneous than the single mutants. The small size of the single mutants could indicate that they spend less time in G1, but this is not born out by their flow cytometry profiles, which look like wild type (S1 Fig). The msa1msa2 cells also have a flow cytometry profile very similar to the wild type profile during logarithmic growth (Fig 4A). This suggests that their overlapping function is not important during rapid growth. After seven days of growth to stationary phase, we obtained wild type yields of high-density cells from the msa1 and msa2 mutants (61 ±3% and 57 ±2% respectively.) However, the double mutant produces a lower, more variable yield of 34 ±6% (Table 1). The dense msa1msa2 cells are also very large (Fig 5B), consistent with their inability to cease cell growth (Fig 4D). Despite their large size, one third of the msa1msa2 cells achieve the density characteristic of Q cells. To see if these high-density cells attained other features of quiescent cells, we tested their thermo-tolerance (Fig 5C) and their longevity (Fig 5D). The dense fraction of msa1msa2 cells is comprised of 75% viable cells (Fig 5D), compared to only 30% viable cells found in the seven day old cultures (Fig 4B). Interestingly, these high-density cells have the same thermo-tolerance and longevity of wild type quiescent cells (Fig 5C and 5D.) We conclude that Msa1 and Msa2 are critical for the efficient transition into quiescence, but the cells that survive this transition achieve at least some of the protective features of Q cells.
Another feature of quiescent cells is their rapid and synchronous return to the cell cycle upon re-feeding [20]. Fig 5E shows the typical 90 minute delay, followed by a highly synchronous cell division, that we observe when Q cells are transferred from water to rich media. The single msa mutants are clearly delayed, and the msa1msa2 cells lag considerably longer. The starting population of msa1msa2 Q cells is about 15% budded, but these budded cells are likely dead, based on their phase dark appearance in the microscope and their failure to progress. The unbudded population begins to bud after 135 minutes. These delays show that both Msa1 and Msa2 are important for an efficient transition out of quiescence, but they are not required for this transition because eventually nearly all the live cells bud.
The failure of the msa1msa2 mutant to arrest in G1 and the association of these proteins with both SBF and MBF led us to ask if known MBF and/or SBF targets were deregulated in these mutants as they transition from log phase to quiescence. As discovered previously [18], msa1 and msa2 mutants have minimal impact on transcription during the log phase of growth. However, they have a significantly greater influence on transcription after the diauxic shift (S2 Fig). To see if MBF and SBF targets are affected, we looked for known MBF and/or SBF targets [37], which were negatively or positively affected by msa1, msa2 or the double mutant. Using a cutoff of 1.8-fold, we found that about half the known SBF and MBF targets were affected by msa1 or msa2. Fig 6 shows the levels of these MBF and SBF target transcripts in the msa single and double mutants in log phase cells and in cells that have undergone the diauxic shift, expressed as a ratio of mutant over wild type. With the exception of YOX1 and MNN1, none of these transcripts are significantly affected by the msa mutants during the log phase of growth. However, after the diauxic shift, we find that MBF targets, and targets of both MBF and SBF are primarily under-represented, and those that are only SBF targets are primarily elevated in the mutants. We also find that the impact of msa1 and msa2 is similar at these promoters and their effects are not additive. This indicates that both Msa1 and Msa2 are required to regulate these MBF and SBF targets in post-diauxic cells. The fact that loss of both Msa proteins has about the same effect at these targets as loss of either one suggests that both Msa proteins are critical components of the same pathways that confer this regulation. It is also worth noting that while only about half the known targets of these late G1-specific transcription factors meet the 1.8-fold threshold in either single mutant, many others are affected in the same way but to a lesser extent (S1 Table). We predict that other transcriptional regulators, mRNA stability or mRNA sequestration are likely to be variables that complicate the extent of de-regulation we observe.
The set of MBF targets whose activation in post-diauxic cells require both Msa1 and Msa2 include SLD2 and a number of other genes involved in DNA replication. In fact, 22 of the 54 MBF targets (p value = 10−9) most repressed in the msa single mutants are involved in DNA metabolism and 18 are involved in DNA repair (p value = 10−10.) This may help explain why galactose-induced overproduction of Msa1 suppresses sld2 and some other DNA replication defects [18]. SLD2 is also an MSA2 activated gene, but Msa2 over-expression does not suppress sld2. This asymmetry could be explained if MSA2, itself an MBF target, is also over-expressed upon galactose-induction of Msa1. This would result in high levels of both Msa proteins, which would activate SLD2 and other DNA replication genes. In contrast, high levels of Msa2 would not be expected to induce high levels of Msa1, and since both are required for activation of these replication genes, high Msa2 alone would not have the same suppressing effect.
The next most enriched class of MBF targets that are activated by both Msa1 and Msa2 are genes involved in sister chromatid cohesion (SMC1, SMC3, IRR1, PDS5, MCD1 and CSM3.) It is known that genome-wide cohesion occurs in response to a single double strand break [42, 43]. Moreover, this break-induced cohesion prevents the loss of the unbroken chromosomes, indicating that it serves a purpose beyond repair of the single double strand break [43]. The co-activation of cohesion genes as cells enter quiescence brings up the intriguing possibility that cohesion may protect and/or compact the genome in quiescent cells. It has recently been shown that quiescent cells have uniquely compact chromatin in which all the telomeres are in a tight cluster in the center of the nucleus [44, 45].
At the SBF target promoters, where both Msa1 and Msa2 are required for the repression, the histone transcripts are among the most affected. All eight core histone transcripts are highly elevated in the msa mutants (Fig 6 and S1 Table). HHT1 is not pictured in Fig 6 because it missed the cutoff for being an assigned SBF target [37]. In addition, the linker histone HHO1 and the H2A variant HTZ1, both barely missed our cutoff for inclusion on Fig 6, each being about 1.7-fold elevated above wild type in post-diauxic msa2 cells. There are 29 targets that meet or exceed the 1.7-fold mRNA level increase and 10 of them are histones. The only other histone, CSE4, which is centromere-specific, is unaffected. Histone expression is tightly controlled by multiple mechanisms that are not entirely understood despite decades of investigation [46]. To confirm the role of SBF in transcription of these genes, we assayed the HTA1 promoter activity through the cell cycle in swi4, mbp1 and swi4mbp1 cells. S3 Fig shows that Swi4 contributes to the cell cycle-specific activation of this promoter and Mbp1 has much less effect.
Our data indicate that Msa1 and Msa2 both make independent contributions to the post-diauxic regulation of many MBF and SBF targets. Loss of these activities is not additive, suggesting that they participate in the same pathways of regulation at these promoters. However, the extreme phenotype is only observed in the double mutant. This suggests some redundancy of function. Redundancy is also suggested by the sequence similarity between Msa1 and Msa2. To address this question, we looked across the genome for transcripts that were mis-regulated in the msa1msa2 double mutant, but not mis-regulated to the same extent in either single mutant in the post-diauxic time point. Table 2 lists all the transcripts across the genome that are mis-regulated 1.7 fold or more in msa1msa2 but less so in msa1 or msa2 after the diauxic shift. Most of these transcripts also show additivity, in that the single mutants mis-regulate in the same way, but to a lesser extent. This can be readily seen in the bar graph in S4 Fig. What is striking from this genome-wide survey is that of the 47 transcripts most down-regulated in the double mutant, one-quarter are already known MBF and/or SBF targets (Table 2). This further supports the view that regulating MBF and SBF activities are the critical functions of Msa1 and Msa2 in post-diauxic cells. Overall, eleven of the 47 genes are cell cycle genes, and all but three of these are involved in chromosome segregation and/or the establishment of polarity in the cell division process. Others affect a diversity of processes.
There is a smaller set of transcripts that are elevated specifically in the double mutant (Table 3). Here again, the single mutants typically also elevate the transcript levels, but to a lesser extent. None of these transcripts are known MBF and/or SBF targets, nor do they show significant enrichment for cell cycle regulation. Rather, seven of these 30 genes respond to stress, and four are meiosis-specific. Another four genes (TIR1, 3,and 4 and DAN1) are specifically expressed during, and required for anaerobic growth [48, 49]. Three of the most elevated transcripts are involved in pyrimidine biosynthesis. There is no reason to think that these are direct targets of Msa1 or Msa2. Rather we suspect that failure to properly initiate the transition to quiescence indirectly results in the ectopic expression of genes involved in other developmental pathways. This ectopic expression could contribute, indirectly, to the loss of viability of the double mutant.
It is unclear which of the mis-regulated transcripts listed in Table 2 and/or 3 might be responsible for the loss of viability observed in the double mutant. Indeed, there is no reason to assume that a single member of either class is causing loss of viability in the msa1msa2 mutant. For example, seven of the down-regulated genes are essential for viability (bold in Table 2) and the reduced expression of any or all of these genes could be deleterious. Included among the up-regulated transcripts is an uncharacterized gene (YLR162W), which is known to cause growth arrest and apoptosis when over-expressed [50].
The down-regulation of MBF targets in the absence of Msa activity could be due to direct binding and activation by the Msa/MBF complex, or it could be due to indirect effects on other regulators. We note that NRM1 is slightly elevated and RAD53 is substantially down-regulated in the msa mutants (S1 Table). Rad53 is known to phosphorylate and release the negative regulator Nrm1 from MBF complexes in the presence of DNA damage [13, 51]. Similarly, the high levels of SBF target transcripts in the msa mutants may be due to direct repression by the Msa/SBF complex, or it could reflect indirect effects on other regulators that are mis-expressed in the Msa mutants. For example, Swi4, the DNA binding component of SBF is up-regulated in the msa mutants (Fig 6). To see if Msa regulation of these targets is direct or indirect, we assayed binding of Msa1 and Msa2 to a set of SBF and MBF targets by chromatin immunoprecipitation. Fig 7A shows a survey of Msa binding to thirteen of these genes, including eleven from Fig 6 and two from Table 2 (TOS6 and CSI2.) We see robust binding signals for both Msa proteins on SBF targets (S) and on promoters containing both SBF and MBF binding sites (B). Binding is weaker to the MBF targets (M). In most cases, the binding signals are higher in the post-diauxic time points (16 and 24 hours.) To confirm the relatively weak binding to MBF targets, we asked if the binding was dependent upon Mbp1, which is the DNA binding component of MBF [9]. Fig 7B shows that the binding observed at three MBF target promoters in post-diauxic cells is Mbp1-dependent. Fig 7C shows that the binding of Msa1 does not depend on the presence of Msa2, or vice versa, at both MBF and SBF targets. These data are most consistent with the direct binding of Msa1 and Msa2 to both classes of late G1-specific promoters. The Msa-dependent regulation of these transcripts is likely to be important for preventing entry into S phase as cells respond to a waning nutrient supply and enter quiescence.
The transition from proliferation to quiescence involves a stable but reversible cell cycle arrest in G1. It follows that the transcriptional regulators that drive the G1 to S transition have to either be eliminated or reprogrammed. Perhaps because of the need for rapid reversibility of this arrest, budding yeast utilize their E2F-like complex of Swi4 and Swi6 (SBF) as a platform for a novel form of regulation that involves Msa1 and Msa2 and is initiated as glucose levels drop. The regulation conferred by Msa1 and Msa2 is critical for the cell cycle arrest, cell growth arrest and viability of cells as they transition to quiescence. Other known SBF regulators (Stb1, Whi5, Srl3) do not play a significant role in the transition into quiescence, but Whi5 and the related Srl3 protein seem to accelerate the reversal of quiescence when nutritional conditions improve. This is opposite their roles as negative regulators in exponentially growing cells [6, 7, 29]. One possible explanation is that Whi5 (and/or Srl3) may displace the Msa proteins from SBF as an early step in the recovery phase. Such an exchange would maintain repression of SBF targets, but would make their activation responsive to the increased cyclin levels that accompany the transition to S phase.
The Msa1 and Msa2 proteins were identified by two very different strategies. The MudPIT analysis used to identify Whi5 and Nrm1 as components of the SBF and MBF complexes [7] also identified two related proteins that were named MBF- and SBF- associated (Msa) proteins [17]. Msa2 had been shown to interact with Swi6, the common component of SBF and MBF [52], and both MSA genes were known to be transcribed in a cell cycle-specific manner. MSA1 (YOR066W) is transcribed at the M/G1 boundary in a Yox1/Mcm1-dependent manner [34, 53], and MSA2 is a late G1-specific transcript [54]. Both proteins are also expressed only in G1 and they undergo cell cycle-specific modifications in growing cells [17, 18]. Msa1 is among the handful of proteins whose nuclear localization is regulated by cyclin-dependent kinase activity and is G1-specific, like Whi5 [55]. Both proteins bind to SBF and MBF target promoters, specifically during G1 in cycling cells, and that binding is Swi4- and Mbp1-dependent, respectively [17].
Msa1 was also identified as a high copy suppressor of three temperature sensitive DNA replication mutants: drc1-1/sld2, dbp11-1 and pol2-12 [18]. Interestingly, Msa1 over-expression had a deleterious effect on other DNA replication genes (cdc6-1 and cdc7-1) and other cell cycle regulators (cdc28 and cdc14-1.) Being aware of the previous study showing the interaction of Msa1 with the SBF and MBF transcription complexes [17], these authors carried out genome-wide chromatin immunoprecipitations and transcript microarrays to identify Msa1 targets. They found 50 genes that were both bound and regulated by Msa1 in cycling cells. These genes affect all aspects of cell growth, but showed no clear connection to DNA replication. This left the mystery of Msa1’s role in DNA replication unresolved. They did identify about 60 MBF and/or SBF targets as binding sites for Msa1, but most were not Msa1 regulated in cycling cells, just as we observe. Our study has shown that there are many DNA replication genes that are activated by Msa proteins, including SLD2, that could explain the suppression of DNA replication defects, but this only occurs in post-diauxic cells.
One thing that these studies, as well as the study implicating Msa1 in osmoregulation of the cell cycle [19], have in common is the relatively mild phenotypes observed for the msa single and double mutants in cycling cells. Altering the levels of Msa1 causes modest changes in the timing of late G1-specific transcription [17], and the initiation of S phase [18]. Further, no synergistic effects in the msa1msa2 double mutant were reported in cycling cells. We observe similarly mild phenotypes for these mutants during the log phase of growth. This is surprising, considering the many layers of regulation that are exerted upon these proteins in cycling cells, and the critical roles their putative targets play in the G1 to S transition. However, we observe strong deleterious effects of these mutants after the diauxic shift, when cells are preparing to shift from proliferation to quiescence. The Msa proteins are critical during this transition, and one clear effect that they have is in the reprogramming of SBF and MBF activity. Our data suggest that this reprogramming is important for entry into and recovery from quiescence.
We know that the Msa proteins are produced, localized to the nucleus and have the capacity to bind SBF and MBF targets specifically during the G1 phase of every cell cycle. However, only after the cells receive a signal of nutrient limitation are the Msa proteins able to influence the activity of most of these transcription complexes. We propose that the purpose of this tight G1-specific regulation in cycling cells is to ensure that these proteins are present, in the G1 nucleus, to respond immediately to these environmental signal(s), to modulate late G1 transcription and to promote G1 arrest and cell growth arrest. In this way, cells in other phases of the cell cycle would continue to progress and only the G1 cells would initiate cell cycle arrest. This may also explain why there is a correlation between the length of G1 in cycling cells and the ability to enter the quiescent state (Fig 1A).
The majority of the Msa1- and Msa2-dependent regulation we observe fits into one of two patterns. In many cases, loss of either Msa1 or Msa2 disrupts regulation, and loss of both is not additive, indicating that both Msa proteins are required in the same pathway of regulation. In other cases, we see maximum deregulation in the double mutant, which suggests some redundancy. However, in most of these cases, both of the single mutants also de-regulate but to a lesser extent. Though there are exceptions (Tables 2 and 3), the bulk of the evidence suggests that both Msa proteins are required at most promoters. Consistent with this, if we eliminate Rad53 checkpoint function, we see deleterious effects of the single msa1 or msa2 mutants that are qualitatively similar but less severe than that of the msa1msa2 mutant. We conclude that Msa1 and Msa2 have undergone substantial functional divergence, but there is a set of critical targets at which either Msa protein can regulate to a sufficient extent to promote survival during the transition to quiescence.
We have shown that Mbp1 is required for Msa binding at three MBF target sites, but deletion of Mbp1 does not interfere with the transition to quiescence (Fig 1B and Table 1). In contrast, the principle SBF components: Swi4 and Swi6, are required for a normal transition to quiescence. This makes it most likely that the critical targets of Msa regulation that promote G1 arrest are among the SBF targets that they regulate. However, more complicated scenarios are possible. Swi4 and Mbp1 have similar DNA binding domains and similar binding sites [2], and there are instances in which an SBF binding site in G1 becomes an MBF binding site in S phase [56]. These and other complexities make it difficult to guess which of these transcripts could play a critical role in promoting a stable G1 arrest and thereby be responsible for the loss of viability of the double mutant. Further work will be required to determine how Msa1 and Msa2 activity is modulated by nutritional cues, how they achieve this regulation and which of their direct or indirect targets are responsible for the G1 arrest that occurs as cells transition to quiescence.
Despite the lack of physical similarity at the protein sequence level, there are striking parallels between the transcriptional regulation that promotes the G1 to S transition in yeast and mammalian cells [2]. Like SBF and MBF, there are E2F protein complexes that activate transcription in G1 and promote S phase. These complexes are inactive in early G1 due to the binding of repressors (Whi5 and RB), which recruit histone deacetylases to their target genes. In both cases, activation requires removal of the repressors by cyclin-dependent kinases. This enables them to induce transcription of their target genes, many of which are also conserved (e.g. cyclins, replication proteins and histones.) With this work, we extend this conservation of strategy to the transcriptional regulation that promotes the transition from G1 to quiescence. In higher eukaryotes, entry into quiescence depends on the formation of novel E2F complexes that serve to repress these same target genes [4]. This so-called DREAM complex of DP, RB-like, E2F and MuvB was first identified in worms and flies [3, 57] and later found to perform a similar function in human cells [4]. It is assembled on E2F target genes to repress transcription in cells entering quiescence and disruption of these complexes drives cells back into the cell cycle [5]. Msa1 and Msa2 perform a similar function, by binding SBF and MBF complexes and reprogramming their activities. They are not required in cycling cells, but they are critical for the transition to quiescence. In their absence, cells fail to arrest in G1 and lose viability. Interestingly, Msa1and Msa2 do not significantly affect the longevity of cells that successfully enter quiescence, but they are required for efficient entry to and exit from quiescence. It will be interesting to determine how cell cycle re-entry from quiescence differs from the G1 to S transition in cycling cells and which of their targets are rate limiting for this transition.
All yeast strains used in this study are isogenic with BY6500, the prototrophic version of W303 [58], unless otherwise indicated. Strain numbers are provided in Table 1 or in figure legends. The 5xCLN3 was created by integrating four additional copies of CLN3 at different marker loci [14] with the integrating vectors, pRS303-306 [59]. The W303 SSD1 was created as described [58]. All the deletions were made using the Longtine deletion vectors [60] unless otherwise indicated. The checkpoint deficient rad53-21 mutant [39] was crossed with the prototrophic W303 (BY6500) and then crossed with the deletion strains as listed in Table 1. The myc-tagged strains were constructed using pFA6a-13Myc-KanMX6 [60].
Viability was monitored by Live/Dead FungaLight (Invitrogen, Grand Island, NY) and colony formation as described. Cell size and cell number was measured on a Z2 Beckman Coulter Counter (Beckman Coulter, Brea, CA.) Growth assays were all carried out in triplicate at 30°C in rich media with 2% glucose (YEPD) with 200 rpm aeration on platform shakers. Growth from log to stationary phase (log to SP) was followed by starting two equivalent 25 ml cultures at an OD600 of .02, ten hours apart, from the same culture maintained in log phase. The first culture is sampled at 8, 10, 12, 24, 28 and 48 hours. The second culture is used for the 14, 16, 18, 20 and 38 hour time points. At each time point, samples for cell count, cell size, and flow cytometry were taken. The diauxic shift was determined by the absence of glucose in the media. Glucose levels were determined using glucose detection strips (GLU 300, Precision Labs, Inc. West Chester, OH.) To follow seven days of growth, cultures were inoculated as above. Samples were taken for the zero time point 5 hours after inoculation, then daily, for assaying cell number, viability and colony forming units. Number of trials averaged for these figures is shown in parentheses.
Quiescent cells were purified after seven days of growth as described above, by centrifugation through a percoll gradient [15, 20]. Typically 200 OD600 units of cells are loaded onto a 25 ml gradient and Q yield is calculated as the percentage of OD600 units that sediment to the bottom 9 ml of the 25 ml gradient. The high density Q cells are washed and maintained in water. Longevity of the Q cells was monitored in triplicate from 13 ml suspensions of Q cells in water inoculated to an OD600 of 1.0 and incubated with aeration at 30°C. We see no acidification of the water after 300 days of incubation under these conditions with wild type cells. To monitor longevity in the non-dividing state, samples are taken from these Q cell suspensions at two week intervals for cell count, cell size, cell viability and colony forming units. Before sampling, these suspensions are weighed and water is added to replace loss due to evaporation. Thermo-tolerance of the high density Q cell fraction was assayed in triplicate starting with Q cells in water at an OD600 of 1.0. 50μl of these cells were transferred to a .5 ml PCR tube and incubated for 10 minutes at the specified temperature. These were chilled, diluted and plated for colony forming units. Q cell re-entry into the cell cycle was followed in triplicate by transferring 10 OD600 units of Q cells in one ml into 25 ml of YEPD, sampling at 15 minute intervals and counting percent of budded cells from a total of 200 cells for each time point. Number of trials averaged for these figures is shown in parentheses.
Flow cytometry was carried out as in [15]. DNA content was quantified by staining with Sytox Green and the percent of cells in G1 was determined using the cell cycle module of FlowJo V9.6.4. As we have shown [15], cells transitioning to the quiescent state undergo asymmetric cell divisions and fortify their cell walls. These events give rise to heterogeneity in the flow cytometry profile. In particular, the G1 peak splits into three peaks, which must be added together to obtain the total number of cells in G1 (Fig 4A and S1 Fig). Our plots all report the percent of live cells that are in G1. Dead cells and cell debris, which accumulate in swi4, swi6, msa1msa2, and msa2rad53 cultures pile up on the left margin of the DNA fluorescence histograms. Number of trials averaged for these figures is shown in parentheses.
RNA sampling, collection and paired end Next-Generation RNA sequencing was carried out as described [14]. mRNA expression levels following polyA selection were assayed by using the HiSeq 2500 next generation sequencing system from Illumina [61] in the Fred Hutchinson Cancer Research Center Genomics Core Facility. Sequences were aligned to the reference genome W303 using the Tophat2 application [62], then counted with HTSeq [63]. Differential expression between samples was measured using the DESeq package from Bioconductor [64]. Ratios of expression in mutant versus wild type were then computed from the normalized read counts. Two biological replicates were generated and averaged for this analysis. These data for all SBF and MBF targets is provided as (S1 Table).
Cells carrying Msa1 or Msa2 tagged with the myc epitope or non-tagged controls were collected from log phase cells, or from cells that had passed the diauxic shift as indicated by the lack of glucose in the media. Proteins were cross-linked to DNA as described [65]. Frozen cell pellets were resuspended in lysis buffer (50 mM HEPES pH 7.6, 140 mM NaCl, 1% Triton X-100, 0.1% Na deoxycholate, 1mM EDTA, 1 mM PMSF, 1 μg/ml aprotinin, leupeptin, and pepstatin A). Cells were broken with glass beads in a Mini Beadbeater-8 (BioSpec Products, Bartlesville, OK) three times for 30 seconds on the Homogenize setting. After a 15-min centrifugation the supernatant was discarded and the pellet (chromatin fraction) was resuspended in the initial volume of lysis buffer. The DNA was fragmented to ∼500 base pairs with a Sonifier Cell Disrupter (Heat-Systems-Ultrasonics, Inc., Plainview, NY), sonicating at setting 3 for 10 seconds 5 times with a one minute ice rest between sets. After clarification, immunoprecipitation was performed with 3 × 109 cells of chromatin, the monoclonal anti-c-MYC antibody 9E10 (Roche Applied Science, Indianapolis, IN) and protein A sepharose CL-4B beads (GE Healthcare, Pittsburg, PA) rolling overnight at 4°C. Immune complexes were washed twice with 1 ml of lysis buffer, 1 ml of lysis buffer with 250 mM NaCl, 1 ml of ChIP wash buffer (10 mM Tris pH 8.0, 250 mM LiCl, 0.75% NP-40, 0.75% Na deoxycholate, 1 mM EDTA), and 1 ml of Tris-EDTA. DNA-protein cross-linking was reversed in 100 μl 1% SDS/Tris-EDTA at 65°C overnight. DNA was cleaned up with 50μg RNase A at 37°C for one hour then 300μg Proteinase K at 50°C for one hour. DNA was purified on Purelink PCR Purification columns (Invitrogen, Grand Island, NY) according to the manufacturer's instructions. PCR reactions (5 min 95°C, 26 times [1 min 94°C, 1 min 55°C, 1 min 72°C], 10 min 72°C, hold 4°C) were performed using HotStarTaq Plus DNA Polymerase (QIAGEN, Hilden, Germany) on 1 μl of 1/1000 eluted input (chromatin) and 1 μl of eluted immunoprecipitation. Sequences of the primers used to detect binding are available upon request. PCR fragments were separated on a 2% agarose gel and visualized by ethidium bromide.
All the demultiplexed FASTQ RNA sequence files are available from the National Center for Biotechnology Sequence Read Archive from accession SRP068917.
|
10.1371/journal.ppat.1007613 | Transmission phenotype of Mycobacterium tuberculosis strains is mechanistically linked to induction of distinct pulmonary pathology | In a study of household contacts (HHC), households were categorized into High (HT) and Low (LT) transmission groups based on the proportion of HHC with a positive tuberculin skin test. The Mycobacterium tuberculosis (Mtb) strains from HT and LT index cases of the households were designated Mtb-HT and Mtb-LT, respectively. We found that C3HeB/FeJ mice infected with Mtb-LT strains exhibited significantly higher bacterial burden compared to Mtb-HT strains and also developed diffused inflammatory lung pathology. In stark contrast, a significant number of mice infected with Mtb-HT strains developed caseating granulomas, a lesion type with high potential to cavitate. None of the Mtb-HT infected animals developed diffused inflammatory lung pathology. A link was observed between increased in vitro replication of Mtb-LT strains and their ability to induce significantly high lipid droplet formation in macrophages. These results support that distinct early interactions of Mtb-HT and Mtb-LT strains with macrophages and subsequent differential trajectories in pathological disease may be the mechanism underlying their transmission potential.
| Mycobacterium tuberculosis (Mtb), the bacteria that causes tuberculosis (TB), is spread through the air from infected patients to their close contacts. In a household contact (HHC) study of patients with TB, we found that in some households a larger proportion of contacts were infected with Mtb compared to other households. We categorized the households into High (HT) and Low (LT) transmission groups. The Mtb strains obtained from the TB patients of the HT and LT households were designated Mtb-HT and Mtb-LT, respectively. In this study, we investigated in a mouse model of TB, the mechanistic basis for the variability in transmission of Mtb from infected patients to their close contacts. We found that C3HeB/FeJ mice infected with Mtb-LT strains developed diffused inflammatory lesions characteristic of granulocytic tuberculous pneumonia. In stark contrast, a significant number of mice infected with Mtb-HT strains developed caseating granulomas, a lesion type with high potential to cavitate. In conclusion, we report that the segregation of Mtb strains into high and low transmission phenotype is mechanistically linked to their ability to induce distinct pulmonary pathology.
| Mtb is one of the most successful pathogens known; yet mechanisms underlying variability in transmission remain poorly understood. Pioneering work from Riley and colleagues [1] using TST conversion in guinea pigs that were exposed to air from a TB ward as a measure of infectiousness, found significant variability in infectiousness among untreated patients with drug susceptible Mtb. Despite having comparable sputum positivity, only 8 of the 61 patients in the TB ward were found to transmit infection [1]. Using a similar model of air sampling analysis in test animals, a great degree of variability in infectiousness was also reported for HIV-infected patients with drug susceptible and resistant TB [2–4]. Subsequent studies demonstrated that some patients transmit their infection to large numbers of contacts whereas other patients transmit rarely or not at all (even after controlling for factors such as extent of disease in the index case and length of exposure) [5–10]. Variability in transmission could result from differences in the variability of infectious aerosols produced during coughing by patients with pulmonary tuberculosis [11].
Mycobacterium tuberculosis Complex (MTBC) comprises of six human adapted lineages and a recently discovered lineage 7 [12]. Whole genome sequence data show that the lineages differ in their content of SNPs, small insertion and deletions, large genomic deletions, large duplications and insertion sequences [12]. Several studies have addressed whether Mtb genotypic diversity is associated with diversity in clinical outcome [reviewed in [12, 13]]. Overall, lineage 2 was found to be highly associated with virulence and transmission in different ethnic populations [14–20]. However, other studies did not find a higher fitness for Beijing strains of lineage 2 [21–23]. For example, in a cohort study of patients with TB and their HHC in The Gambia where M. africanum is endemic, there was no difference in transmission between M. africanum and Mtb or between the MTBC lineages [24]. In another population-based study in Montreal, Canada, four Mtb lineages were identified- Euro-American (lineage 4), Beijing (lineage 2), Indo-Oceanic (lineage 1) and East African-Indian (EAI-lineage 3) [25]. In contrast to previous studies, higher transmissibility of the Beijing lineage was not observed in this study. However, the EAI lineage was associated with lower rates of TB transmission, as measured by positive TST among close contacts of pulmonary TB cases [25]. The Beijing genotype consists of a number of sub-lineages and so the discrepant findings could be due to the prevalence of different sub-lineages in the different study populations, as suggested previously [12]. For example, in a study conducted in Western Cape, South Africa, significantly higher transmission was found to be linked to recently evolved sub-lineages of the Beijing strain family than to other sub-lineages, indicating that strains within individual lineages have acquired distinct transmissibility traits [17]. Similarly, a study from China also concluded that transmissibility was dissimilar among the Beijing sub-lineages [26]. Consistent with the idea that there is heterogeneity within a lineage, studies in an animal model of transmission reported that Beijing genotype strains exhibited various degrees of virulence phenotype and transmissibility [26]. It has been suggested that the heterogeneic immune response and virulence of the Beijing strains may be due to their differential engagement of the innate Toll-like receptors (TLR) [27].
The high transmissibility and prevalence of some sub-lineages belonging to the Beijing phenotype raises the question of whether the type of immune response elicited by some strains in the lineage gives them a selective advantage over other genotypes. In this regard, strains within the highly prevalent “modern” Beijing genotype have been linked to low inflammatory cytokine responses [28–30] and faster growth rate in vitro [31, 32] compared to strains within the “ancient” Beijing genotype. By selecting representative strains from the Modern (lineage 2, lineage 3 and lineage 4) and ancient lineages (lineage 1, lineage 5 and lineage 6), Portevin et al. [30] extended the analysis of macrophage cytokine responses to all the major phylogenetic lineages. They found that overall, strains from the modern lineages induced lower levels of pro-inflammatory cytokines when compared with strains representing ancient lineages. Furthermore, two strains, HN878 (lineage 2), a member of the Beijing family that caused several outbreaks of TB in Texas [33, 34] and Mtb strain CH (lineage 3) that caused a large outbreak of TB in Leicester, UK [35] also exhibited a low inflammatory phenotype. In contrast, CDC1551 (lineage 4) that also caused a large number of TB infections in a small rural community in Tennessee [36], induced a robust inflammatory response and was less virulent in mice [37]. The finding with CDC1551, together with several studies showing lack of association of “modern” lineages with disease presentation [38–41] or transmission [21–24], indicates an incomplete understanding of bacterial factors that favor transmission success.
Many factors, including source infectiousness [42], cough aerosols [43] and nature and proximity of contact [44] affect transmission. To address pivotal questions pertaining to transmissibility of different Mtb strains, we conducted a study that included 731 household contacts (HHC) of 124 infectious TB patients, and found marked heterogeneity in Mtb transmission within households [42]. Index cases (with pulmonary TB disease) and their respective households were categorized into high (HT) and low (LT) transmission groups based on the proportion of HHC with a positive tuberculin skin test. The Mtb strains from HT and LT index cases of the households were designated Mtb-HT and Mtb-LT, respectively. The goal of this study was to explore the role of Mtb strain in the observed differences in transmission and the mechanistic basis by which the epidemiologically characterized Mtb-HT and Mtb-LT strains, though belonging to lineage 4, have diversified in their transmission profile. The C3HeB/FeJ mouse model was employed to examine the role of Mtb strain in the dichotomous transmission outcomes of the HT and LT households.
Although there are disparities in susceptibility to infection and disease manifestations that exist between humans and animals, experimental models such as C3HeB/FeJ mice that present with lung pathology more typical of human TB disease [45–49] are useful for hypothesis-driven research aimed at understanding TB immunopathogenesis. Using the C3HeB/FeJ mouse model, we found that Mtb-HT and Mtb-LT strains induced distinct growth pattern and pathological disease that fit their transmission phenotype. We also found increased lipid biogenesis in Mtb-LT infection of macrophages compared to Mtb-HT infection suggesting that this difference in host response may be a factor in the divergence in infection outcome between the two groups of animals.
As previously reported, the HHC study only included crowded (≥3 HHC) dwellings with an intense (≥3 weeks of cough) and homogeneous (sputum AFB ≥2) infectious exposure, and classification of households as “high” or “low” transmission was based on TST positivity of contacts at the end of enrollment [42]. Briefly, the percentage of contacts from a given household that had a TST≥ 10mm of induration either at baseline or by 8–12 weeks was used to determine the transmission category of the household. If there were ≥70% TST positivity in HHCs, the active TB disease patient or “index case” was considered “High Transmission” (HT) and if there was ≤40% contacts that were TST positive, the index case was considered “Low Transmission” (LT) [42]. 293 TB patients were screened and 124 index cases were enrolled. The Mtb strain isolated from HT index case was designated Mtb-HT and from LT index case was designated Mtb-LT.
From the panel of HT and LT strains, three strains from each group were randomly picked, blinded to any patient or household contact details. Based on subsequent examination of the index case and household characteristics, it is reasonable to assume that the six randomly selected strains are representative of the larger groups of HT and LT strains. All of the HT and LT strains belong to lineage 4. The index case and household characteristics of the six strains are described in Table 1. Total TST positivity was 100% for the HT index cases and 14–40% for the LT cases (Table 1). All of the six isolates were collected from the Vitória Metropolitan area. Vitória, Brazil is the capital city of Espírito Santo, a state with a TB incidence rate of 38/100,000 and very low HIV prevalence in TB cases (<2%) and the general population (<1%). While migration into this region is limited, there is a high level of inter-municipality mobility amongst inhabitants. We previously showed that there was no relationship between the HT and LT transmission phenotypes and the presence of RFLP/spoligotyping clusters in the community [50]. As shown in Table 1, the six strains were unique isolates or belonged to different clusters. Based on RFLP, Mtb-HT1 and Mtb-LT1 are different even though they both fall within the LAM9 sub-lineage. This sub-lineage is the most common spoligotyping sub-lineage in Brazil. Based on radiographic appearance, HT1 and HT3 index cases had far advanced lung disease and HT2, LT1, LT2 and LT3 index cases had moderately advanced lung disease. Chest X-rays of the TB patients showed that HT1, HT2 and HT3 index cases had cavitations, whereas only patient LT1 presented with cavitary disease (Table 1). For in vitro assays, seven additional strains isolated from HT and LT index cases were tested (S1 Table).
In separate infections, previously grown stocks of the six Mtb clinical strains described in Table 1 were used to infect C3HeB/FeJ mice. Bacterial burden was monitored at different time-points following aerosol infection of the mice (Fig 1). We observed that despite similar inoculum size, lung bacterial burden in mice infected with two of the three Mtb-LT strains, Mtb-LT1 and Mtb-LT2, were significantly higher than all three HT strains at all time points sampled (Fig 1). Mtb-LT3 produced higher bacterial growth at week 2 post infection but was not significantly different from Mtb-HT strains at later time points (Fig 1). These changes in bacterial growth were also reflected in extra-pulmonary dissemination to the spleen and liver (S1 Fig). A repeat experiment with Mtb-HT1 and Mtb-LT1 also showed that CFU in the lungs, mediastinal lymph nodes and spleen at 12 weeks of infection was significantly higher in Mtb-LT1 infected mice compared with Mtb-HT1 infected mice (S2 Fig).
To delineate whether host genotype contributed to the differences between Mtb-HT and Mtb-LT strains, we also infected inbred C57BL/6 and BALB/c mice with Mtb-HT1 and Mtb-LT1 strains. In both genotypes of mice, we observed that compared to Mtb-HT1, Mtb-LT1 infected C57BL/6 and BALB/c mice showed significantly higher CFU in the lungs (S3A Fig).
Next, we evaluated the composition of the cellular infiltrates to the lungs of 4 week-infected C3HeB/FeJ mice. We observed that all three strains of Mtb-LT infected C3HeB/FeJ mice had significantly higher number of total viable cells in the lungs as compared to the mice infected with the three Mtb-HT strains (Fig 2A). Further analysis of the cellular composition of the recruited cells showed that, compared to Mtb-HT1, all three Mtb-LT infected mice had significantly increased numbers of CD8+ T cells and CD11b+CD11c+ recruited macrophages while B220+ B-cells, CD11bhiLy6G+ neutrophils were increased in Mtb-LT1 and Mtb-LT2 infections. CD4+ T cells and CD11b-CD11c+ alveolar macrophages were significantly enhanced only in Mtb-LT1 infection (Fig 2B). There was no significant increase in any cellular subsets in Mtb-HT2 and Mtb-HT3 infections compared to Mtb-HT1 (Fig 2B). The increase in neutrophil numbers in Mtb-LT1 infection was confirmed in immunohistochemical staining of lung sections. A large number of Ly6G+ neutrophils were observed in the granulomatous lesions of four week Mtb-LT1 infected mice compared to Mtb-HT1 infected mice that had sparse neutrophils in the granuloma (Fig 2C). Consistent with increased cellular recruitment, evaluation of lung homogenates showed an overall increase in TNF, IL-1β, IL-6, IL-17 and KC (CXCL1) in Mtb-LT infected mice at four weeks following infection compared with Mtb-HT infected mice (S4 Fig). These data indicate that Mtb-LT strains induce an overall strong inflammatory response, admittedly, though, this could be an indirect effect as a result of the increased bacterial burden in these mice.
H&E staining of paraffin-embedded lung sections revealed that there were striking differences in the types of granulomatous lung lesions that were formed between animals infected with Mtb-HT and Mtb-LT strains. Although early on, there were no clear differences in lung pathology between the Mtb-HT and Mtb-LT infected mice (Fig 3A and 3B, top panel), at four weeks post-infection Mtb-HT infected mice had well-defined, circumscribed lesions in otherwise normal parenchyma. In contrast, diffuse granulomatous inflammation was observed in Mtb-LT infected animals (Fig 3A and 3B, middle panel). This is consistent with the enhanced cytokine and chemokine induction, and neutrophil accumulation in Mtb-LT infected mice. During the chronic stages of infection (weeks 8–12), well-formed granulomas in the lungs of Mtb-HT infected mice were seen to expand, whereas Mtb-LT infected mice showed widespread tissue destruction (Fig 3A and 3B, bottom panel).
At the latest time point, caseating granulomas in the HT-infected mice were well-defined, circumscribed lesions surrounded by a thin layer of fibroblasts and collagen, in otherwise normal parenchyma (Fig 4A). Many of these lesions were in close proximity to intact airways (Fig 4C), spilling bacterial and necrotic debris into the airway lumen. Surrounding the necrotic center were acid fast organisms (Fig 4D) and collagen-rich cell debris (Fig 4B and 4E) enriched with foamy macrophages within a collar of fibroblasts (Fig 4B). Many of the granulomas in Mtb-HT infected mice appeared as small aggregates of cells (Fig 5A), dominated by lymphocytes (Fig 5B), minimal inflammation (Fig 5C) and containing few intracellular bacteria (Fig 5D). In contrast to the discrete granulomas observed in mice infected with Mtb-HT strains, there was diffuse lung pathology in mice infected with the Mtb-LT strains (Fig 6A). Bacteria-laden macrophages expanded to fill alveolar spaces, with no evidence of containment of the inflammatory process. In addition, multiple small lymphoid aggregates were present throughout the lung as well as foci of acute neutrophilic inflammation (Fig 6B). Besides the widespread destruction of lung parenchyma, fluid and inflammatory cells were present in remaining airways (Fig 6C). Of note, throughout the course of infection, we observed 20% (6 of 30 mice) and 12% (3 of 25 mice) mortality in Mtb-LT1 and Mtb-LT2, respectively. This can be attributed to the replacement of alveolar surfaces by the rapidly progressing inflammatory process seen in these animals. Although there were far more bacilli present in the lungs of Mtb-LT infected mice, AFB staining demonstrated that these were primarily intracellular (Fig 6D).
In addition, within the granulomatous lesion of Mtb-LT infected animals we found foamy macrophages and lipid clefts similar to those found in atheromatous plaques in patient with hypercholesterolemia (S5 Fig). These solid lipid crystals, dissolved by the solvents used in tissue processing to leave clefts, are a source of inflammation as well as mechanical injury [51]. To further confirm that the differences in the pattern of disease pathology is not confounded by examining a single lobe for histopathology, we performed another infection with Mtb-HT3 (n = 4) and Mtb-LT3 (n = 4) and analyzed multiple lung lobes from 12 week-infected animals. In line with earlier experiments, all of the Mtb-HT3 infected animals developed discrete granulomas and 2 of the 4 showed caseating granulomas (S6A Fig) in one lobe while 100% of Mtb-LT3 infected mice exhibited diffused inflammatory pathology that was present in all lung lobes (S6B Fig). Together, these data confirm that Mtb-HT and Mtb-LT infected mice develop strikingly different pathological disease.
Combined histological evaluations of 12 and 16 week-infected lungs revealed that only a proportion of Mtb-HT infected animals developed caseating granulomas (Table 2). The rest of the mice exhibited discrete granulomas with the potential to progress to caseous necrotic granulomas. Of significance, none of the Mtb-HT infected animals at either 12 or 16 week of infection exhibited diffused inflammation. In contrast, all of the Mtb-LT1 and Mtb-LT2 infected animals developed diffused inflammation with no evidence of caseating or discrete granulomas in any of the mice. Notably, despite having lung bacterial numbers similar to Mtb-HT infected mice, 8 of 9 mice of the Mtb-LT3 infected animals developed diffused inflammation and none of these infected mice had caseating granulomas (Table 2). If we consider the two pathological outcomes, caseating granulomas and diffused inflammation between Mtb-HT1-3 and Mtb-LT1-3 as a primary endpoint, a two-sided Fisher’s exact test yields a p value of p<0.0001. This suggests that the 3 Mtb-HT and 3 Mtb-LT strains have significantly different pathological outcomes.
In parallel, we also evaluated the lung pathological response in C57BL/6 and BALB/c animals that were infected with Mtb-HT1 and Mtb-LT1 strains. Quantification of granulomatous inflammation in both mouse genotypes established that at 12-weeks post infection, animals infected with Mtb-LT1 strain had significantly higher lung area involvement as compared with Mtb-HT1 infected animals (S3B Fig). Between the genotypes, BALB/c mice infected with Mtb-LT1 had significantly more lung area involvement at this time point (S3B Fig).
Central to the intracellular growth of Mtb is its ability to induce foamy macrophage by triggering the accumulation of lipid droplets, which are composed of triglycerides and cholesteryl esters [52, 53]. Lipid droplets thus provide a nutrient source to intracellular Mtb and promote successful replication of the pathogen in the host [54, 55]. We argued that the difference in the growth of Mtb-LT and Mtb-HT in vivo in mice is due to their differential ability to induce lipid droplet formation. MH-S cells, a mouse lung alveolar macrophage cell line, were infected individually in vitro at an MOI of 10 with the 3 Mtb-HT and 3 Mtb-LT strains studied so far and with an additional 7 each of Mtb-HT and Mtb-LT strains. Based on flow cytometric analysis, cells infected with Mtb-LT strains had a strong signal for the LipidTOX dye (a reagent that stains neutral lipid droplets) and overall significantly high MFI as compared to cells infected with Mtb-HT strains (Fig 7A). Representative images from confocal microscopy supported the flow cytometry observations (Fig 7B). These data suggest that induction of lipid droplets may be differentially regulated by Mtb-HT and Mtb-LT strains.
Mepenzolate bromide (MPN) was shown previously to reduce lipid droplet formation in Mtb-infected macrophages by targeting the anti-lipolytic G protein-coupled receptor GPR109A which resulted in enhanced TAG turnover [52]. Consistent with lipid droplets serving as a nutrient source for Mtb, growth of Mtb in vitro in macrophages was reduced in the presence of MPN [52]. Therefore, to determine if there was a link between lipid droplet formation and increased Mtb-LT growth, bone marrow macrophages were infected with the three Mtb-HT and 3 Mtb-LT strains, with or without MPN, and intracellular bacterial growth was determined at day 7 following infection. Firstly, we found that the intracellular growth of Mtb-LT strains was significantly higher than Mtb-HT strains. Furthermore, MPN treatment selectively decreased intracellular bacterial growth of only Mtb-LT strains (Fig 7B). However, bacterial growth in liquid culture of all three Mtb-HT and Mtb-LT1 and Mtb-LT2 strains was significantly inhibited by MPN. Mtb-LT3 had a slow growth rate in vitro and addition of MPN did not further reduce growth. (S7A Fig). Although, the target of MPN in Mtb remains unclear, nonetheless, these data indicate that in vitro growth characteristics of Mtb-HT and Mtb-LT strains are not predictive of its intracellular growth pattern in macrophages, the latter we propose being more relevant to in vivo infections and further that lipid droplet formation is a factor in the enhanced intracellular growth of the Mtb-LT strains. Signaling via GPR109A activates inflammatory [56] and anti-inflammatory pathways [57]. We, therefore, tested whether the level of TNF was altered in MPN-treated macrophages and found no significant difference in the production of the cytokine between macrophages with or without MPN treatment (S7B Fig). Overall, these data indicate that Mtb-LT strains induce significantly higher lipid droplets to promote their intracellular growth.
In this study, Mtb clinical strains were carefully characterized on the basis of epidemiological data into high and low transmission groups and studied to gain insight into the pathogenic mechanisms leading to their transmission phenotype. The experimental results demonstrate that Mtb-HT and Mtb-LT strains exhibit differential bacterial growth and lung pathology in genotypically similar hosts. Our results also demonstrate early modulation of lipid biogenesis by Mtb-LT strains which could be the likely mechanism dictating the differential outcome of infection between Mtb-HT and Mtb-LT-infected animals. Data arising from this study also advance the C3HeB/FeJ mouse model as a tractable system to identify bacterial determinants that interact with the host immune response to cause the type of pathological disease that enables different extent of bacterial transmission.
Previous studies have shown that C3HeB/FeJ mice infected with Mtb Erdman or H37Rv develop caseating granulomas and exhibit granulocytic tuberculous pneumonia referred to as Type I and Type II lesions, respectively [48]. Interestingly though, the same Mtb strain induced both lesion types. The novel observation in this study is that the granulomas evolved to caseation characterized by fibrous encapsulation with central liquefaction necrosis similar to Type 1 lesions only in Mtb-HT infected animals. Although, caseating granulomas did not develop in all of the Mtb-HT infected mice, nonetheless, none of them presented with diffused inflammatory pathology. In contrast, all of the Mtb-LT-infected mice rapidly developed diffused inflammatory pathology, but none presented with caseating granulomas. The bacterial burden in mice infected with Mtb-LT3, although similar to Mtb-HT at later time points of infection, was nonetheless significantly higher than the three Mtb-HT strains at week 2 of infection. Of note, the Mtb-LT3 exhibited a pathological response that was very similar to the two other Mtb-LT strains. This suggests that the early events leading to differences in bacterial growth in Mtb-HT and Mtb-LT-infected animals may dictate the divergence in pathological response. The genetic differences in Mtb-HT and Mtb-LT strain and the ensuing specific interactions with the host may be driving the development of predominantly one or the other lesion types in the infected hosts. Consistent with this idea, a recent study in guinea pigs also found that strains that do not transmit disease caused more inflammatory pathology compared to high transmission strains [58].
The difference in bacterial replication and inflammation between the Mtb-HT and Mtb-LT strains in the C3HeB/FeJ mice was also recapitulated in two other mouse genotypes, indicating that bacterial factors likely contribute to the differential pathological outcome of Mtb-HT and Mtb-LT infections. A comparative genomics study of a panel of 19 clinically and epidemiologically characterized isolates of Mtb found that those with greater genome deletions caused significantly less pulmonary cavitation suggesting that pathogenicity is linked to bacterial factors [59]. That bacterial genotype contributes to the transmission phenotype of the host is also indicated by the finding that a large sequence polymorphism in a gene encoding molybdopterin oxyreductase was associated with clustering [60] and an Mtb strain associated with a large outbreak in the UK harbored an insertion in an intergeneic region Rv2815-2816c [61]. Whole genome sequencing and evolutionary convergence analysis of 100 strains either least or most likely to be transmitted revealed that five Mtb genes were shared by the transmissible strains. Importantly, the Mtb strains with mutations variably affected monocyte and lymphocyte cytokine production and neutrophil generation of reactive oxygen species [62]. Together, these findings indicate that differences in bacterial factors could regulate the extent of TB transmission occurring in a host. However, future studies should determine if host genetics synergize with bacterial factors to enhance transmission. Diversity Outbred (DO) [63, 64] and Collaborative Cross (CC) mice are highly heterogeneous populations that provide a tractable experimental system to model the genetic diversity of the human outbred population. The outcome of Mtb infection in DO [65] and CC [66] mice was highly varied, ranging from resistance to high susceptibility, and was associated with a diverse range of pathological responses. Of note, the susceptible mice exhibited necrotizing tuberculous pneumonia, similar to the pathological response of C3HeB/FeJ mice infected with Mtb-LT. In the DO and CC mice, the same strain induced different pathological responses whereas in our study strain variation contributed to the different pathological disease outcome. An integrated approach that combines Mtb-HT and Mtb-LT infections in heterogenous DO and CC mice will provide a powerful means to investigate whether transmissibility is the combinatorial effect of host and strain genetic diversity.
Mtb-HT and Mtb-LT strains induce distinct immunopathological responses in the susceptible host and thereby create an environmental context that we posit is differentially permissive to transmission between the susceptible host and an exposed individual. The finding that there are higher levels of Mtb-LT organisms in granulocytic pneumonitis in mice, if directly relevant to humans would not necessarily equate to their greater abundance in infectious aerosols. Rather, it is bacterial replication in area of caseation necrosis in granulomas that is associated with cavity formation and also perhaps differential survival in aerosols that lead to increased transmission. The C3HeB/FeJ mice are disease susceptible, however, a limitation of the model is that not all Mtb-HT infected animals developed caseating necrotic granulomas, and furthermore, none of the mice developed cavitary disease, a key pathological feature of human TB. A reason for why the Mtb-HT infected mice did not develop cavitary disease may be because mice lack a functional ortholog of human MMP1 that causes matrix destruction in TB [67]. However, pulmonary cavitation in C3HeB/FeJ mice was detected after aerosol infection with Mtb by serial computed tomography (CT) imaging [68]. Together, these findings suggest that further refinement of the model such as crossing MMP-1 transgenic mice with C3HeB/FeJ or the B6.C3H-sst1 mice [69] may provide a superior mouse model to study TB pathogenesis and transmission.
The formation of lipid-filled foamy macrophages is a hallmark of Mtb infection [53]. In mycobacteria infected cells, lipid droplets are found in close apposition to the phagosome [70]. Lipid droplet and phagosome interaction leads to engulfment of mycobacteria into the lipid droplet, providing the microbe unrestricted access to host lipids [70, 71]. Accumulation of triacylglycerol-rich lipid bodies has been shown to aid mycobacterial survival and host neutral lipids can further be stored within the bacilli as intracytoplasmic lipid inclusions, thus acting as an energy source and enhancing bacterial growth in granulomatous lesions [54, 55]. Thus lipid droplets play a prominent role in sustaining successful Mtb survival and replication in the host. What bacterial factors from Mtb-LT strains activate the GPR109A to induce the rapid accumulation of lipid droplets in macrophages awaits clarification, nonetheless, the findings from this study show that Mtb-LT exploits this host signaling pathway to its advantage. A recent study argues that lipid droplet formation is not a bacterially driven process during infection with the laboratory derived Mtb Erdman, but instead is dependent on IFNγ, and is not a source of host lipids for the pathogen [72]. Our findings that Mtb-HT induce significantly less lipid droplets and that their intracellular growth is not affected by MPN, suggests that, like Mtb Erdman, lipid droplets may not be a source of lipids for Mtb-HT. However, lipid droplet formation by Mtb-LT and their decreased growth in the presence of MPN indicates that lipid droplets specifically contribute to the enhanced intracellular growth of Mtb-LT strains. The data we present show that Mtb-LT strains engage the GPR109A pathway that is known to suppress TAG turnover to enhance lipid droplet accumulation. However, Mtb can also induce host cell lipid synthesis [73] and use triacylglycerol to accumulate lipid droplets [74]. Future investigations should explore these possibilities. Additionally, whether other factors, such as ability to survive aerosol stress also contributes to the overall transmissibility of an Mtb strain needs further inquiry.
From the standpoint of bacterial dynamics, how Mtb-LT strains survive and propogate in the community remains an interesting question. In fact, the low fitness for transmissibility of Mtb-LT strains may portend a future decline in prevalence in the community unless the strains show increased propensity to progress to disease. It is possible as well that Mtb-LT transmission is enhanced in hosts with co-morbidities, such as HIV infection, diabetes or malnutrition and in them rapid transmission to disease maintains the strains in the community.
The overall findings from this study are that Mtb-HT and Mtb-LT strains belonging to the same lineage differ in their interaction with the host immune system leading to different trajectories in bacterial growth and in the development of disease pathology. The divergence in disease pathology is likely the underlying cause of differences in infectiousness of the source case. However, since the current findings are based on a small sample size, larger confirmatory studies are required for broader inference that individual strains have biological properties that induce different pathological response that affects their transmission potential. Bearing in mind this limitation, the planned next stage of this work is to define an in vitro immune phenotype correlating with the in vivo growth and pathology to provide a high throughput screening method for validation of the current findings in a large panel of Mtb-HT and Mtb-LT strains. Furthermore, findings from ongoing studies of whole genome sequencing and metabo-lipidomic profiling of a large panel of HT and LT strains will uncover key genes and bacterial factors responsible for the dichotomy in pathogenesis. In the future, this can be translated into transmission interventions that target the bacteria.
The household contact study from which the Mtb strains were derived was approved by the Comite de Ética em Pesquisa do Hospital Universitário Cassiano Antonio de Morais, and the Institutional Review Boards of Rutgers University Biomedical Health Sciences-Newark (RBHS) (formerly UMDNJ) and Boston University School of Medicine. Written informed consent and assent in Portuguese were obtained from all study participants as per the consent procedure approved by IRBs from all participating institutions.
All animal experiments described in this study conform to the Rutgers University Biomedical Health Sciences-Newark (RBHS) and Institutional Animal Care and Use Committee (IACUC) Guidelines as well as NIH and USDA policies on the care and use of animals in research and teaching. Efforts were taken to ensure minimal animal pain and suffering and when applicable, approved anesthesia methods were employed for the same.
Clinical strains of Mtb were first grown on Lowenstein Jensen (LJ) growth media. Bacterial colonies picked from LJ slants were cultured in 7H9 media until mid-log phase. Cultures were then centrifuged at 400 RPM for 5 minutes, causing clumps to settle down in the bacterial pellet. The culture supernatant was collected, mixed with a final concentration of 20% glycerol, and was stored in 1 mL aliquots at -80°C. The stock titer was determined by plating 10-fold serial dilutions on Middlebrook 7H11 selective medium (Difco by BD, Franklin Lakes, NJ) and by counting the bacterial colonies 15–20 days later. The bacteria were passaged in vitro only twice to minimize any phenotypic/genotypic changes that might occur in the growing cultures.
6–8 weeks old female C57BL/6J, BALB/c and C3HeB/FeJ mice were purchased from the Jackson Laboratory (Bar Harbor, ME, USA). Bacterial stocks were generated as described above. Mice were exposed for 40 minutes to nebulized bacteria at a density optimized to deliver a standard low dose of around 50–120 CFU (unless otherwise indicated) using Glass-Col Full Body Inhalation Exposure. For all infections, the actual infection dose was determined by plating total lung homogenates from a minimum of 3 mice on Middlebrook 7H11 plates at 24 hours after aerosol exposure. Lungs, spleens, mediastinal lymph nodes and livers were harvested and homogenized at indicated time points post-infection. Total CFU per organ was determined by plating 10-fold serial dilutions on Middlebrook 7H11 plates, which were counted after 28–35 days of incubation at 37°C.
Post-mortem, lungs of Mtb-infected mice were perfused with sterile PBS and fixed in 4% paraformaldehyde for seven days, followed by paraffin embedding. For histopathological analysis, 5- to 7-μm sections were cut and stained using a standard H&E protocol. Leica SCN400 F whole-slide scanner (Experimental Pathology Research Lab, NYU Langone Health) was used for scanning histological sections and images were analyzed using Aperio ImageScope. Stereoscopic images were obtained using Act-1 software from Nikon. For quantitation of granulomatous inflammation in the lung section, Image-Pro Discovery Software was used to create a grid overlay onto each photomicrographs of H&E stained lung section and numbers of points hitting areas of granulomatous infiltration were counted. Masson’s trichrome staining [75] was carried out by NJMS Histology core. For visualization of acid-fast bacilli (AFB), tissue sections were stained using the Ziehl-Neelsen method.
For immunohistochemical detection of Ly6G+ cells, tissue samples were de-paraffinized with xylene and rehydrated with ethanol gradations and water. The samples were subjected to heat-induced antigen retrieval by microwave warming using 10 mM citrate buffer (pH 6.0). Endogenous peroxidase activity was blocked using 0.3% hydrogen peroxide and then subsequently blocked with 1× PowerBlock (BioGenex). PBS containing 0.05% Tween-20 was used to wash tissues in between steps. For each sample, serial sections were incubated with the primary anti-mouse Ly6G antibody (clone 1A8; Biolegend) at a 1∶250 dilution or with isotype control (Rat IgG2a, κ; BioLegend) at the same concentration. Sections were subsequently incubated with biotinylated secondary antibody (1∶100 Vector Laboratories). Streptavidin horseradish peroxidase (BioGenex) was used to label the secondary antibody for immunodetection by DAB chromogen (BioGenex). After counterstaining with Mayer’s hematoxylin (BioGenex), the samples were dehydrated with ethanol gradations, dipped in xylene, and mounted using Cytoseal-60 (ThermoFisher). Histopathological evaluations were performed with blinding to the identity of the strain.
Single-cell lung suspensions were prepared and cell viability was determined using Trypan-blue exclusion method. For surface staining, approximately 1 million cells were washed and resuspended in FACS buffer (PBS + 2% fetal calf serum (FCS) and 0.09% sodium azide) containing a cocktail with the appropriate concentrations of specific fluorochrome-conjugated monoclonal antibodies. Isotype controls were included for each. Cells were first incubated with LIVE/DEAD Fixable Aqua Dead Cell stain. Directly conjugated fluorochrome labeled antibodies were used for the following cell-surface markers: anti-mouse CD4-V450 (clone RM4-5; BD Horizon), anti-mouse CD8-AF488 (clone 53–6.7; BD Pharmingen), anti-mouse B220-PECF594 (clone RA3-6B2; BD Pharmingen), anti-mouse CD11b-PE (clone M1/70; BD Pharmingen), anti-mouse Ly6G-PECy7 (clone 1A8; BD Pharmingen), anti-mouse CD11c-AF700 (clone HL3; BD Pharmingen) and anti-mouse Ly6C-PerCPCy5.5 (clone HK1.4; eBiosciences). Following surface staining, samples were fixed in 4% paraformaldehyde for 30 minutes and then acquired on a LSRII flow cytometer (BD Biosciences). Analysis was performed using FlowJo software (Tree Star, Inc.). Gating was based on fluorescence minus one (FMO) controls.
For detection of lipid bodies, infected cells were fixed with Cytofix/Cytoperm solution (BD) for 20 minutes. Then, cells were washed twice with Perm/Wash buffer (BD) and cells were then resuspended in PBS solution of HCS LipidTOX Deep Red Neutral Lipid stain (ThermoFisher Scientific). Following this incubation step, cells were washed twice with PBS and resuspended in PBS for flow cytometric analysis.
Adherent MH-S cells were grown on coverslips (Fisherbrand) placed in 6-well plates (Corning). Following Mtb infection, cells were fixed in 4% formaldehyde, washed and then stained with PBS solution of HCS LipidTOX Deep Red Neutral Lipid stain (ThermoFisher Scientific), followed by 300 nM DAPI nuclear stain solution (ThermoFisher Scientific). Coverslips were then washed three times with PBS and then mounted on Super frost/Plus microscope slides using Molecular Probes Slowfade Light antifade medium. Nikon A1RS confocal microscope was used to acquire images and quantification of signal intensity was performed using ImageJ software and Nikon imaging software, Nikon Elements 4.5.
BMDMs were prepared as described previously [76]. On day 7, BMDMs were plated in 96 well plate (Corning) at a cell density of 0.08 x 106 cells/well in 200μL of D10 media [antibiotic free DMEM media (Mediatech, Inc.) containing 10% defined FBS (HyClone Laboratories, Logan, UT)] and supplemented with 2% conditioned medium from L-cells. Cells were infected in replicates of 5 with 3 MOI of three Mtb-HT and three Mtb-LT strains for 4 hours. Wells were then washed 4 times with PBS + 1% BCS and cells were untreated or treated with 100nM of MPN. Infected cells were maintained in D-10 media supplemented with 2% conditioned medium from L-cells. At day 7, post-infection cells were washed with serum-containing PBS and then lysed with sterile water. Total CFU was determined by plating 10-fold serial dilutions on Middlebrook 7H11 plates, which were counted after 21 days of incubation at 37 °C. 48-hour culture supernatants were harvested for measuring TNF levels.
Lung lysates from Mtb infected mice were treated with 2X protease inhibitor (ThermoFisher Scientific) at the time of collection and supernatants from infected BMDMs were sterilized using 0.22μm. Ultrafree-MC centrifugal filter (EMD Millipore). ELISA Ready-Set-Go kit was used or IL-17 (eBioscience). For all ELISA, colorimetric analyses were used to calculate protein concentration levels (Molecular Devices, Softmax Pro). For TNF, IL-1β, IL-6 and KC, a multi-analyte detection system that incorporates electro-chemiluminescence based readout was used (MesoScale Discovery, Rockville, MD, USA). Pre-coated 10-spot MULTI-SPOT plates with capture antibodies were purchased (catalog # K15048D). The assays are based on the principle of electrochemiluminescence (ECL) sandwich ELISA. The calculations to establish calibration curves and determine analyte concentrations were carried out using the MSD DISCOVERY WORKBENCH analysis software.
All statistical analyses were performed using Graph Pad Prism software. For analysis of two groups, the unpaired t-test was used. For greater than two groups, One- or Two- way ANOVA with Bonferroni’s correction was used. In all cases, p value <0.05 was considered to be statistically significant.
|
10.1371/journal.pgen.1007731 | Slower environmental change hinders adaptation from standing genetic variation | Evolutionary responses to environmental change depend on the time available for adaptation before environmental degradation leads to extinction. Explicit tests of this relationship are limited to microbes where adaptation usually depends on the sequential fixation of de novo mutations, excluding standing variation for genotype-by-environment fitness interactions that should be key for most natural species. For natural species evolving from standing genetic variation, adaptation at slower rates of environmental change may be impeded since the best genotypes at the most extreme environments can be lost during evolution due to genetic drift or founder effects. To address this hypothesis, we perform experimental evolution with self-fertilizing populations of the nematode Caenorhabditis elegans and develop an inference model to describe natural selection on extant genotypes under environmental change. Under a sudden environmental change, we find that selection rapidly increases the frequency of genotypes with high fitness in the most extreme environment. In contrast, under a gradual environmental change selection first favors genotypes that are worse at the most extreme environment. We demonstrate with a second set of evolution experiments that, as a consequence of slower environmental change and thus longer periods to reach the most extreme environments, genetic drift and founder effects can lead to the loss of the most beneficial genotypes. We further find that maintenance of standing genetic variation can retard the fixation of the best genotypes in the most extreme environment because of interference between them. Taken together, these results show that slower environmental change can hamper adaptation from standing genetic variation and they support theoretical models indicating that standing variation for genotype-by-environment fitness interactions critically alters the pace and outcome of adaptation under environmental change.
| Adaptation under environmental change is expected to depend on the time available for the sequential fixation of mutations, but also on standing variation in genotype-by-environment fitness interactions. In the later circumstances, some genotypes might be initially favored but then disfavored and overtaken by other genotypes that are better at more extreme environments if they were not lost by genetic drift or founder effects in the meantime. We addressed this idea with experimental evolution in Caenorhabditis elegans populations with standing variation for genotype-by-environment fitness interactions and by developing a model to describe natural selection on extant genotypes during experimental evolution. We find that under slower environmental change, the genotypes that are initially selected are not the best at the most extreme environments, and as a consequence, that these best genotypes can be lost due to genetic drift and founder effects. We further find that the longer polymorphism is maintained the more likely that selective interference will reduce the best genotypes to low frequencies and increase the chances for their loss through genetic drift. We conclude that under slower environmental change adaptation will be deterred if populations can only rely on standing genetic variation.
| With human activities contributing to climate change [1], it has become urgent to pinpoint the ecological and evolutionary mechanisms by which natural populations adapt at different rates of environmental change. It is generally accepted that lower rates of environmental change allow more time for beneficial mutations to appear, to be selected, and, as a consequence, to promote adaptation and rescue populations before environmental degradation leads to their extinction [2–6]. Experimental evolution results from studies with microbes that depend on de novo mutation support the idea that slower environmental change facilitates adaptation [7–10]. Unlike microbial experimental evolution, however, most species in nature have small populations, are genetically structured by geography, breeding mode or reproduction system, and might have long generation times. In all these cases, adaptation to changing environments will likely depend on standing genetic variation, and less so on de novo mutation [11, 12].
Adaptation to changing environments from standing genetic variation is conditional on how each extant genotype performs within the environments that may be encountered in the near future (Fig 1A). Depending on the shape of these “fitness reaction norms” [3, 10, 13], and previous evolutionary history responsible for standing genotype frequencies [5, 14], natural selection may initially favor genotypes at intermediate challenging environments that are not necessarily the best at the more extreme environments. In other words, adaptation will depend on standing genotype-by-environment (GxE) fitness variance [11], until de novo mutations or new recombinant genotypes escape genetic drift and start to be selected upon [15].
Understanding the population genetic dynamics of adaptation from standing genetic variation to changing environments has only been recently formalized using a moving polygenic trait optimum model [11]. In contrast to evolution from de novo mutation, one of the strongest conclusions from ref. [11] was that slower environmental change can restrict adaptation when evolving populations depend on standing genetic variation. One reason for this is that the maintenance of standing genetic variation for longer periods can result in reduced fitness variance and thus reduced rates of adaptation [16, 17] (Stage II in Fig 1A). Another reason is that since all populations are finite, and may suffer bottlenecks in the novel environments, the longer it takes to reach the most extreme environments (Stages I and II in Fig 1A) the more probable it is that the best genotypes are lost by genetic drift or by founder effects and thus be unavailable when populations reach the extreme environments (Stage III in Fig 1A). Because of standing variation for GxE fitness interactions, this later process will be more pronounced if the best genotypes at the extreme environments are initially selected against in the less extreme environments (Stage I in Fig 1A). In general, whether or not a population has standing genetic variation is expected to greatly affect the tempo and mode of adaptation in changing environments.
Given a fixed amount of standing genetic variation, and assuming no input of de novo mutation or new recombinant genotypes during evolution, we here experimentally test how the rate of environmental change affects adaptation. We investigate whether slower environmental change can constrain adaptation because of the loss of extant genotypes that would perform best in the most extreme environments. These genotypes may be lost during the period of environmental change via genetic drift or founder effects. To this end we performed experimental evolution under a sudden or gradual environmental change (Fig 1B), using populations of the nematode Caenorhabditis elegans with standing genetic variation and where individuals can only reproduce by self-fertilization. In this situation, we expect that asexual population genetic dynamics will be followed and that they will depend on standing GxE fitness variance. At several periods we collected genome-wide single-nucleotide polymorphism (SNP) data and used these to infer the fitness reaction norms of the genotypes that were present in the ancestral population as well as their expected frequency changes during experimental evolution due to selection. We performed a second set of evolution experiments where we test for the repeatability of adaptation in the most extreme environment to show that genetic drift and founder effects during prior gradual evolution can lead to the loss of the best genotypes and impact selection efficacy.
As previously reported in ref. [18], we performed experimental evolution for 50 generations in the nematode C. elegans under different rates of change in the NaCl (salt) concentration that individuals experience from early larvae to adulthood (Fig 1B). In one regime, populations were suddenly placed in high salt concentration conditions (305 mM NaCl) and then maintained in this environment for 50 generations (see Materials and Methods and S1 Text section 1.1). In another experimental evolution regime, populations faced gradually increasing salt concentrations for 35 generations, being thereafter maintained in constant high salt for an extra 15 generations. For the “sudden” regime, 4 replicate populations undergoing independent evolution were followed, while for the “gradual” regime we followed 7 replicate populations (S1 Table). All populations were derived from a single ancestor population adapted for 140 generations to lab conditions (25 mM NaCl), after initial hybridization of several wild isolates [19], that has abundant genetic diversity (expected SNP heterozygosity of ~0.3, for 1 SNP per kbp on average, presumably maintained in excess by balancing selection at overdominant loci, see [20, 21]) but where individual hermaphrodites reproduce exclusively by self-fertilization (obtained by the introgression of a male-killing mutant to the lab adapted population, cf. [18]). Hermaphrodites are expected to be mostly homozygous throughout their genome before the start of experimental evolution in changing salt environments [18, 22]. Except for salt concentrations, the same life-cycle of discrete and non-overlapping generations at stable census population sizes of 104 hermaphrodites at the time of reproduction were maintained as during lab adaptation. Lab adaptation occurred under partial self-fertilization and outcrossing, with an estimated effective population size of the order of 103 [20]. With exclusive self-fertilization this number should be halved to at least 500 [22]. A control regime with 3 replicate populations was also maintained at the 25 mM NaCl conditions of lab adaptation. Given exclusive self-fertilization, the expected effective population sizes, and the time span of experimental evolution, de novo mutation or new recombinants from standing genetic variation should not contribute much to adaptation to high salt concentrations [11, 23].
During experimental evolution, we measured the frequency of biallelic SNPs obtained from genotyping hermaphrodites in the ancestral population and the evolved populations at generations 10, 35 and 50 (Figs 1B and 2A). All replicate populations from the control and sudden regimes were genotyped, while from the gradual regime we genotyped 4 out of the 7 replicates (see Materials and Methods and S1 Text section 1.6 for the genotyping protocol, and S1 Fig for SNP density and sample sizes). SNPs were chosen based on the known diversity present in the 140 generation lab adapted population [21], and the expected genetic distance between them [24](S1A Fig). Given the limited amount of genomic DNA per hermaphrodite to perform whole-genome genotyping at a high density, we chose to genotype each hermaphrodite only in a pair of chromosomes (C. elegans is diploid with six similarly-sized chromosomes, for a genome of 100 Mbp), with the objective of sampling haplotypes at relatively low frequencies (Fig 2A).
With self-fertilization and complete linkage disequilibrium, the number of observed chromosome-wide haplotypes (CWH) should be similar to the number of observed “region-wide” haplotypes (RWH), each defined by a pair of homozygous chromosomes in each hermaphrodite (Fig 2A). We estimate, however, that linkage disequilibrium is not complete since when we take the data from all populations and time points into consideration between 5% and 21% more RWHs than CWHs are found, depending on the region (Fig 2B), with the majority of them being at low frequencies across the dataset (S2 Fig). We estimate that the ancestral population must have segregated at least 212 RWHs (the minimum observed number), and by extrapolation at least the same number of whole-genome haplotypes although many more are possible (see S1 Text section 1.11).
To facilitate computation, we grouped minor frequency region-wide haplotypes (RWHs) in each replicate population into a single class (see S1 Text section 1.10, and S2 and S3 Figs). We find that the majority of RWHs in the ancestral population are quickly selected against under all experimental evolution regimes (Fig 3). By generation 50, all populations are dominated by a single RWH in each of the two-chromosome regions. Populations faced with a sudden change in the first generation followed by constant high salt (305 mM NaCl) consistently show a single haplotype sweeping and nearing fixation by generation 50. In contrast, populations faced with a gradual increase in salt until generation 35 showed a different haplotype initially sweeping but then reverting in frequency when they were kept in the target high salt environment for another 15 generations. Control populations also show that a single RHW per genomic region sweeps through them. This is the same haplotype as that found during initial gradual evolution, suggesting continued lab adaptation under exclusive self-fertilization independently of salt [18].
Since experimental evolution occurred under exclusive self-fertilization and we assume complete homozygosity, the fitness reaction norms of genome-wide haplotypes, here defined as “lineages”, are the key variables for describing selection in changing salt environments, and thus the eventual outcome of adaptation to the extreme high salt environment (Fig 1). The non-monotonic RWH frequency dynamics observed in the gradual populations in particular (Fig 3) can be explained by the crossing of fitness reaction norms somewhere along the salt gradient but, conceivably, also by negative frequency-dependent selection among segregating lineages [15].
To detect selection in changing environments, we adapted standard population genetics modeling [25] to infer the fitness reaction norms of segregating lineages and their expected frequency dynamics (S1 Text section 1.8). We model a single additive multi-allelic locus in effectively asexual populations, and thus do not specifically account for dominance or epistasis. We further modeled deterministic environmental and population genetics (i.e., there are no genotype frequency changes and no genotype extinction/fixation due to random environmental fluctuations or finite population sizes), with discrete non-overlapping generations and viability selection.
We consider that the environment faced in a given generation is represented by a single environmental value x, in our case corresponding to the NaCl concentration (Fig 1B). A population is composed of G genome-wide lineages, with the fitness reaction norm for lineage k being described by λk(x) (Fig 4), corresponding to the expected absolute number of live offspring produced under environment x. Selection is defined by the per generation growth multiplier (growth rate) of each lineage relative to mean population fitness–with the frequency of each lineage expected to follow a deterministic logistic frequency trajectory [25, 26]. Our model allows for any parameterization of the fitness reaction norms although we only investigate linear and quadratic functions.
To infer the frequency dynamics of the lineages during experimental evolution, we developed a maximum-likelihood model that estimates the parameters describing the fitness reaction norms of these lineages (S1 Text section 1.9). For this, we rely on genotyping data, consisting of the number of each RWH observed when genotyping the populations in various time-points. Since the model is parameterized on absolute fitness, we also rely on fitness data, which serves to properly scale the estimated parameters.
Inference is done in several steps, illustrated in Fig 4. We first sample the lineages that likely compose the ancestral population, taking the sample sizes and estimated RWHs frequencies into consideration, since the true lineage identities and their starting frequencies are unknown (S4 Fig and S1 Text section 1.11). We then estimate the parameters for the fitness reaction norms of the various RWHs constituting a lineage (each determined by the combination of sampled RWHs, assuming linkage equilibrium among them), and define that the lineage parameters are the sum, in log space, of their constituent RWH parameters (Fig 4). The final likelihood depends on the probability of observing the mean ancestral population fitness (in low and high salt, see below) and the observed RWH time series (for all populations and regimes), given the sampling done to identify the lineages and their frequencies in the ancestral population.
We initially modelled linear fitness reaction norms and found that two lineages dominate the population genetic dynamics. The measured RWH frequency dynamics (Fig 3) are consistent with a single lineage sweeping through the sudden populations (Fig 5), which we label L28 (see below). In contrast to the sudden populations, the gradual populations had an initial increase of a lineage other than L28 (labeled L11), but then started to be overtaken by L28 after the 15 generations of high salt (Fig 5; see S5–S7 Figs for detailed frequency dynamics of major constituent RWHs in each genomic region in all replicate populations and regimes). L11 clearly shows a non-monotonic trajectory in the gradual populations, initially being positively selected and later being negatively selected. Under all experimental evolution regimes, a few other lineages are predicted to also explain population genetic dynamics, although these lineages do not regularly approach a frequency above 15% at any period (S8 Fig: see, for example, lineages 13 and 20 in the sudden and gradual populations, or lineage 470 in the control populations). We reach the same conclusions regarding RWHs (S9 Fig) and lineage (S10 Fig) frequency dynamics when we used a quadratic parameterization for the reaction norms.
S11 and S12 Figs show the expected dynamics of the mean and the variance in population fitness under linear and quadratic models, respectively. Under the linear model, adaptation to intermediate salt conditions in the gradual regime results in a great loss of fitness variance. At the same time, mean population fitness also decreases, a result that is consistent with the existence of an adaptive “lag load”, cf. [2, 11], since L28 is for most periods not being selected. In the sudden regime, mean population fitness strictly increases while L28 is being positively selected. Under the quadratic model, dynamics are more idiosyncratic in the gradual regime, but mean population fitness decreases to similar levels as in the linear model, and then recovers at a similar pace.
We measured the ancestral population absolute fitness as the growth rate over one generation at 25 mM and 305 mM NaCl to help with the inference of fitness reaction norms (see previous section, S1 Text section 1.3). We first sought to validate the analysis by measuring the ancestral population absolute fitness at an intermediate salt concentration (225 mM NaCl). Results show that there is a large difference between the expected and observed fitness values at 225 mM NaCl (Fig 6A), although they are intermediate to 25 mM and 305 mM NaCl. The discrepancy between observed and expected fitness values was anticipated since our inference at 225 mM NaCl was only informed by the observed RWHs frequencies in the gradual populations at generation 25 (Fig 1B and S5–S7 Figs).
More directly, we sought to validate the analysis by measuring the fitness reaction norms of the two lineages (L11 and L28) that appear to dominate the population genetic dynamics during experimental evolution (Fig 5 and S8 Fig). Using whole-genome sequencing data on 100 lineages derived from two gradual populations at generation 50 (as reported in [21]), we identified those corresponding to L28 and L11 (S13 Fig and S2 Table). Our model predicts that the linear or quadratic fitness reaction norms of these two lineages cross between 200–250 mM NaCl (Figs 5A and S10A). To test this prediction, we revived L28 and L11 from frozen stocks and assayed their absolute fitness at 25 mM, 225 mM and 305 mM NaCl. Absolute fitness was measured as the growth rate over two generations under non-competitive conditions. We find a close agreement with the model in that the lineages’ reaction norms cross at about 225 mM NaCl (Fig 6B and 6C), even if the observed values are higher than the predicted ones. At 25 mM NaCl there is a larger difference between observed and expected fitness values than at other salt concentrations. Differences between observed and expected fitness values can be explained by the low frequency of these two lineages in the ancestral and control populations (those that experienced 25 mM NaCl), and the gradual populations when at 225 mM NaCl by generation 25. Supporting our interpretation, the observed fitness values at 305 mM NaCl closely match the expected fitness values (Fig 6B and 6C), particularly for the L28 lineage. In this case the inference was mostly informed by the lineages segregating in the sudden populations, as they always experienced this salt concentration during experimental evolution (S8 Fig).
It is possible that non-transitive interactions between standing genetic variation, in particular because of negative frequency-dependence, could in part also explain the discrepancies between observed and expected absolute fitness values in the ancestral population and the L28 and L11 lineages. To test for this possibility, we conducted head-to-head competitive (relative) fitness assays between L28 and L11 (S1 Text, section 1.4). In these competition assays, performed for 2 consecutive generations, both lines were initially placed at 1:1 ratios at the usual population sizes, noting that these frequency ratios between L28 and L11 were never realized during experimental evolution (Fig 5B). The results from the competition assays are qualitatively similar to those under non-competitive conditions (Fig 6D, compare with Fig 6C). Non-transitive interactions between L28 and L11 therefore do not appear to be significant in explaining differences between observed and inferred fitness values.
Besides the uncertainty in estimating the frequency of segregating lineages, the discrepancy between observed and expected ancestral and lineage fitness can be explained by how well the parameterization of the reaction norms is done. For example, in the linear model variance in fitness as a function of salt levels must be strictly correlated while in the quadratic model the extra parameter allows the variance in fitness to differ between salt levels. Since power to infer fitness at low salt is generally weak, predictions with the quadratic model will necessary be less precise.
So far, our experiments and modeling demonstrate that the population genetic dynamics under different rates of environmental change are contingent on the GxE fitness variance present in the ancestral population. We found that lineage L28 is the best genotype in high salt, and therefore–assuming no de novo mutation or recombinants–adaptation can be hindered under slow rates of environmental change if the loss of this lineage by genetic drift or founder effects is more probable than under fast rates of environmental change (see Introduction and Fig 1A). From our model, we expect the L28 frequency in the ancestral population and in the gradual populations at generation 35 to be negligible (Fig 5). The model assumes deterministic frequency dynamics and infinite population sizes and thus cannot be verified with the experimental data, since, for example, some of the replicate gradual populations could have lost L28 by the time they reached generation 35. Although some of the region-wide haplotypes constituting the L28 lineage are observed in the 4 replicate gradual populations at generation 35 (S5–S7 Figs), they not only are at low frequencies but could also be detected as part of other lineages (S8 Fig and S2 Table).
To address if genetic drift and founder effects can be implicated in the loss L28 under slower rates of environmental change, we revived frozen stocks from the 7 replicate gradual populations at generation 35 (Figs 1B and 7A), and performed a new set of evolution experiments at two different population size regimes, 104 and 2·103, for 30 generations in constant high salt (Fig 7A, see Materials and Methods and S1 Text section 1.5). In this second set of experiments, we refer to each of the 7 gradual populations as ancestrals #1–7 (S1 Table).
Two main factors, prior genetic drift or selection, could lead to differences in the adaptive responses observed from each of the new ancestral populations, as well as between population size regimes. First, the best high salt lineage determined from the first set of evolution experiments, L28, may have been lost by genetic drift before the second set of experiments started. The freezing and reviving process of the populations could also have resulted in L28 loss; in this case a population size bottleneck would cause a founder effect for the second set of experiments. The second factor is that the efficacy of selection on the best lineages should be lower because of stronger genetic drift in small populations [22].
At two time points during this second set of evolution experiments, a small number of SNPs across the genome were genotyped in pools of individuals, chosen to maximize the ability to distinguish lineage L28 in large samples (S14 and S15 Figs). We then calculated the probability of a L28 sweep under the inferred fitness reaction norms of segregating lineages found above, given the pooled genotyping data (S1 Textsection 1.12, S16 and S17 Figs). Under our genotyping protocol and analysis, an L28 sweep does not imply its fixation during the time frame of experimental evolution–indeed, with deterministic dynamics we predict that after 30 generations at high salt L28 frequency would only reach 50% (Fig 5)–nor are we able to determine if lineages other than L28 sweep through the populations.
We found that the evolutionary responses from the 7 ancestral populations fell into four distinct categories (Fig 7). The first category demonstrates the consequences of a founder effect on adaptation to high salt since the L28 lineage did not sweep through any population (Fig 7B and S16 and S17 Figs). Yet, from at least the first two ancestrals, another unidentified lineage appears to have responded more rapidly at large population sizes than at small population sizes, a result indicating higher selection efficacy at larger population sizes. From the third ancestral, we can only conclude that the L28 lineage was lost before starting the second set of experiments.
A second category of responses more directly illustrates the effects of genetic drift on adaptation (Fig 7C). From the fourth ancestral, the L28 lineage swept rapidly in the high population size regime, while at smaller population sizes the response was more restricted and L28 probably lost during evolution in high salt.
The third category also illustrates the effects of genetic drift to adaptation, although in the opposite sense (Fig 7D). From the fifth and sixth ancestrals, we find that the L28 lineage swept in a fraction of the populations, but exclusively in those with small population sizes. This seemingly puzzling result can be explained if one postulates that, together with L28, the ancestor populations segregated at a relatively high frequency other unidentified lineages that were almost as fit as L28 in high salt (such as lineage 13, see S8 Fig). Interference at large population sizes could have transiently kept L28 at a lower frequency than that expected [27], possibly promoting its extinction by genetic drift [28, 29]. In contrast, at small population sizes the loss of high fit lineages (or maintenance at very low frequencies) by genetic drift might have in turn freed selection to sometimes favor L28 unconstrained.
Despite population size regime, all populations derived from the seventh ancestral showed rapid sweeping of L28 (Fig 7E). This indicates that L28 was initially at a relatively high frequency in the seventh ancestor, when compared to the fourth ancestor (where L28 rapidly sweep only in the large population). At generation 35 of gradual experimental evolution, we observed that one of the constituent region-wide haplotypes of L28 was present in the seventh ancestor, while absent in the fourth ancestor, in line with the expected initial frequency differences in L28 between them (S5–S7 Figs). However, without further genome-wide SNP sampling at high densities and sizes, we cannot assess how prior gradual evolution impacted L28 frequency for continued experimental evolution in constant high salt (S16 and S17 Figs).
Adaptation to extreme environments under different rates of environmental change is expected to depend on ancestral GxE fitness variance and thus on the shape of fitness reaction norms and relative frequencies of extant genotypes. Ignoring the input from de novo mutations and of new recombinant genotypes during adaptation, our experiments in a gradually changing environment show that genotypes initially favored by selection are later selected against when they are overtaken by better genotypes as the environment becomes more extreme. Further, because adaptation to intermediate environments during gradual evolution decreases the frequencies of the genotypes that are most adapted to the extreme environments, these best genotypes can be lost before populations reach such extreme environments.
During the last decade there has been a substantial effort in the development of inference methods to detect selection on DNA sequence diversity during experimental evolution [30–32], although no prior work has explicitly dealt with changing environments. Without directly assaying fitness of each individual genotype, our approach allowed us to infer the distribution of standing GxE fitness variance, inference of both genotype frequencies and the genotypic effects across an environmental (salt) gradient. Based on the inferred distribution we could have predicted the outcome of selection under any rate of environmental change, although we only explored the experimental evolution regimes that were actually performed. Future studies could therefore investigate how different distributions of ancestral GxE fitness variance–in the amount of diversity and shape of reaction norms–determine the loss of genetic variance during environmental change and, for example, the mean population fitness lag load [2, 11]. Our preliminary results indicate that independently of the specific parameterization of fitness reaction norms, slower environmental change transiently results in maladaptation and ultimately delays adaptation. Reaction norms with more flexible parameterizations, however, seem to generate complex fitness variance dynamics, presumably because genotypes favored at early periods can become neutral at other periods and then again positively selected at later periods. For example, the loss of fitness variance at intermediate salt levels is more pronounced under linear than quadratic functions, although by generation 50 the mean population fitness is actually higher under the linear than the quadratic model. Despite our approach allowing for arbitrary parameterizations of the reaction norms, one can of course argue that the decision to model particular reaction norm shapes should first hinge on an understanding of individual development and physiology in the relevant environments.
The most obvious limitation of our inference method is that population size was not included as a parameter and thus we could not account for the effects of genetic drift. Such extension of the model would allow explicit predictions about the loss of genetic variance with variable population sizes and thus the probability of extinction to deteriorating environments, an especially important problem in the context of changing environments, e.g., [33]. An approach by Nené and colleagues [34], focused on the case of evolution of new haplotypes in a population via mutation and positive selection in a constant environment, could perhaps be adapted to detect selection in changing environments with stochasticity. They developed a phenomenological "delay-deterministic” model where an "effective" mutation rate was conditioned on the current frequency of the focal haplotype, with a given threshold mutation rate being parameterized to mimic the effects of genetic drift. Under a limited set of simulated data, the addition of the delay term to their deterministic model better reproduced the frequency dynamics and produced better estimates of selection coefficients. We anticipate, however, that methods explicitly accounting for stochasticity, for example Bayesian models estimated using MCMC techniques, will be necessary in order to manage computational constraints and allow for hypothesis testing and model fit evaluation.
Another future extension of our approach should be to apply it to outcrossing haploids and diploids. The model could be adapted to account for mating and recombination by finding genomic regions at high to complete linkage disequilibrium during the relative short periods of experimental evolution and treating them as we did here the “region-wide haplotypes” (RWH). But expanding our model with recombination represents a considerable challenge since it requires characterizing the degree of polygenicity for fitness [21] and whether or not accounting for dominance and epistasis is necessary [35]. With selection on new genotypes generated by recombination, as with de novo mutation, adaptive rates may increase if it takes longer for a population to reach extreme environments. The net effect of loss of genotypes during adaptation to intermediate environments and the production of new genotypes by recombination is not immediately clear [11], and thus reconciling experimental evolution results that depend on standing genetic variation with and without recombination with those where adaptation occurs from de novo mutation is a major future task.
Experimental evolution studies with microbes that depend on de novo mutation suggest that adaptive gains become smaller with each mutational event, and therefore that adaptation involves diminishing-returns epistasis for fitness [36, 37]. Microbial experiments further indicate that slower environmental change allows more time for the exploration of mutational “space” and the possibility to fix mutations at intermediate environments that predispose subsequent fixation of additional mutations at more extreme environments. Such outcomes should depend on the empirical fitness relationship between alleles related by single mutational steps [38, 39]. In the study of Gorter and colleagues [10], under some stressors, slow environmental change retarded adaptation but not the fitness gains in the most extreme environments. In the study of Lindsey and colleagues [40] the populations that survived a sudden environmental change had higher fitness than those that survived a more gradual change, suggesting, just as in our experiments from standing genetic diversity, a key role of GxE fitness interactions. In changing environments, GxE fitness interactions appear to be sufficient to explain adaptation to extreme environments when evolution occurs from standing genetic variation (without recombination), while both GxE interactions and epistasis are important when evolution occurs from de novo mutation. Clearly, we were unable to determine if epistasis played a role in adaptation since, by definition, there was only selection between non-recombining genotypes.
Little theoretical work has focused on understanding the population genetics of adaptation from standing genetic variation in changing environments. An exception is the study by Matuszewski and colleagues [11], which explored the distribution of fitness effects of fixed alleles starting from standing variation, and with mutational input, under a moving trait under stabilizing selection and epistasis for fitness. The trait was modeled as polygenic with additive interactions between alleles (effectively a biallelic infinite-site and continuum of alleles model), with recombination rates following a Poisson distribution and cross-overs a uniform position along the genome. Matuszewski and colleagues found that populations facing a fast environmental change show larger trait changes than those facing a slow environmental change, due to increases in both the expected number of fixations and the expected trait effect per allele substitution. Although they did not analyze situations of an abrupt environmental change under complete linkage (no recombination), as in our sudden evolution experiments, they nonetheless predicted a higher number of fixations under faster environmental change, and that adaptation would be deterred under slower environmental changes. Matuszewski and colleagues further found that while fast environmental change eliminates sets of de novo mutations, it also helps to keep standing genetic variation until it can be picked up by selection. On the other hand, under slow environmental change, most large effect alleles are already eliminated by genetic drift (or stabilizing selection) before they could contribute to adaptation.
Although the mathematical assumptions of the model of Matuszewski and colleagues do not closely match our experimental conditions, some of their predictions are consistent with the results obtained. We found that slower environmental change allows populations to maintain more genotypes for longer than faster environmental change, and that this can compromise adaptation. Besides loss by genetic drift, one reason for compromised adaptation is that when fitness reaction norms cross, the fitness variance is reduced and adaptive rates diminished [16]. Previous demography, form of selection and degree of environmental variability will determine standing levels of genetic variation and thus from where along the environmental gradient adaptation will ensue [5] (Fig 1). If the population has already exhausted standing GxE variance, then the rate of environmental change will not affect the loss of relevant genotypes simply because they are not present in the population.
Selective “interference” is yet another process that could in part determine hindered adaptation under slower environmental change, and that could also explain why the best genotype was not favored at high population sizes in some of the high salt continued evolution experiments. In this scenario, since slower environmental change can promote the maintenance of polymorphism for longer periods, it is possible that reduced selection efficacy on the best genotypes kept them at low frequencies and caused in turn their loss by genetic drift before populations reached the most extreme environment. With recombination, interference between the best genotypes should be diminished [28, 29, 41], and hence adaptation to the extreme environments will probably not be constrained as when there is limited recombination. Selective interference has been theoretically and empirically studied for microbial evolution experiments in constant environments where the mutational supply is high enough for competing asexual lineages to interfere with each other and retard fixation of the best mutations [42, 43]. Other findings posit an important role for interference and stochasticity in maintaining the long-term standing genetic variation in sexual organisms [29, 41], in particular those reproducing by self-fertilization and with greatly reduced effective recombination rates [44, 45]. However, the importance of interference between beneficial genotypes and stochasticity in promoting their loss in changing environments remains to be explored.
Understanding the outcome of selection in changing environments is complicated because the historical sequence of population genetic changes, recombination and mutational input will determine the way populations respond later in evolution. Since our experimental design used fixed standing genetic variation, with little opportunity for new mutations or recombinants, we were able to examine in isolation the interaction between the sequence of environmental change and the ancestral variation in fitness reaction norms. We demonstrated that under gradual environmental change the genotypes most adapted to the extreme environments do not rise to high frequency during the early periods at less extreme environments. This then opens the door for stochastic loss of genotypes by genetic drift and founder effects, as revealed by our continued evolution experiments. Ultimately, the combination of these processes results in greater adaptation under faster environmental change.
All populations employed are ultimately derived from a hybrid population of 16 wild isolates [19], followed by 140 generations of laboratory domestication to a 4-day non-overlapping life-cycle under partial self-fertilization (self-fertilization) at census sizes of N = 104 at the time of reproduction [15, 19], and introgression and homozygosity of the xol-1(tm3055) sex determination mutant allele at high populations sizes for 16 generations to generate an ancestral population capable of reproduction only by self-fertilization [18]. Experimental evolution in changing environments has been previously reported (Fig 1B, [18]). Large samples of the ancestral population were revived from frozen samples [46], expanded in numbers and first larval staged (L1s) individuals seeded at the appropriate densities to three regimes. The “sudden” regime was characterized by the same conditions to which previous lab-adaptation occurred, except that the NGM-lite media (US Biological) where worms grew was supplemented with NaCl (305 mM) from the start and for 50 generations (4 replicate populations; S1 Table). For the “gradual” regime plates were supplemented with increasing concentrations of NaCl from 33 mM at generation 1 to 305 mM NaCl at generation 35 and onwards until generation 50 (7 replicate populations). A “control” regime was maintained in the ancestral environmental conditions without any salt supplement (3 replicates). Individual hermaphrodites from the ancestor population and generation 10, 35 and 50 from the sudden, gradual and control populations were handpicked for genotyping.
All 7 replicate populations from the gradual regime at generation 35 were revived from frozen stocks, expanded in numbers for two generations, and then split into two regimes: large population sizes of N = 104 and small population sizes of N = 2·103 at the time of reproduction (Fig 7A). From each of the 7 gradual populations at generation 35, one replicate was maintained at large population sizes and three replicates were maintained at small population sizes. All populations were kept at constant 305 mM NaCl for 30 generations. Over 103 L1s were collected per population at generation 15 and 30, for pool-genotyping.
The ancestral population was thawed from frozen stocks and individuals reared for two generations at 25 mM NaCl before they were exposed to the three salt treatments: 25 mM, 225 mM or 305 mM NaCl (Fig 6). Following the usual culture protocol during experimental evolution, on the third generation, five Petri dishes per NaCl treatment were seeded each with 103 L1s. These five plates constituted one technical replicate, and there were four for each salt treatment. After 66 h, individuals were harvested and exposed to a 1 M KOH:5% NaOCl solution (to which only embryos survive). After 16 h, debris was removed and the total number of live L1s estimated by repeated sampling of small volumes.
Statistical analysis was done based on the log-transformed per-capita L1-to-L1 growth rate values, using a linear model with the assay environment as a categorical variable. For this, the assay environment for the i-th measurement is denoted as Ei, and given by: Ei = 0, for 25 mM NaCl; Ei = 1, for 225 mM NaCl and Ei = 2, for 305 mM NaCl. In this way, the 25 mM NaCl is taken as the reference environment. The model then takes the form:
ξi=β(Ei)=β0+β1I(Ei,1)+β2I(Ei,2)
where I(Ei,j) is the indicator function:
I(Ei,j)={1,ifEi=j0,otherwise
and β0, β1 and β2 are coefficients to be estimated.
The data was analyzed in R [47], using the following formula to specify the model in the lm function:
log(growthRate) ~ saltTreatment
Least-square estimates of the expected log-growth rates in each of the three assay environments were then obtained using the R package lsmeans [48]. Note that for inferring fitness reaction norms only the ancestral fitness estimates at 25 mM and 305 mM were used (see below).
During experimental evolution in changing environments, one lineage (whole-genome haploid haplotype) swept through the sudden populations, while another lineage was initially sweeping though the gradual populations when they were at intermediate salt concentrations (Fig 5). From two gradual populations at generation 50, we derived in [21], by repeated single hermaphrodite self-fertilization for >10 generations, 100 “lines” which were the whole-genome sequenced. Comparing the >300k SNPs in the lines with the 761 SNPs collected during experimental evolution (see below), we identified lines L28 and L11 as representatives of the lineages predicted to explain the experimental population dynamics (S11 Fig and S2 Table).
We also conducted absolute fitness assays for L28 and L11 (Fig 6), in a similar manner and replication as for the ancestral population, except that L1-to-L1 growth rate data were collected for two generations. Statistical analysis was done based on the log-transformed per-capita L1-to-L1 growth rate values, with the value obtained for the i-th measurement denoted as ξi. Since the data were gathered over two generations, we accounted for the potential presence of transgenerational effects by using a mixed effects model [49], with environment, lineage and a transgenerational component as fixed effects, and assay block (defined by when the lineages were revived from frozen stocks) as a random effect:
ξi=β(Ei,Li)+α(Li)gi+γ(Bi)
where Ei denotes the assay environment (Ei = 0, for 25 mM NaCl; Ei = 1, for 225 mM NaCl and Ei = 2, for 305 mM NaCl), Li denotes the line (L11 or L28; Li = 0, for L28; Li = 1, for L11), gi corresponds to the transgenerational component (described below) and Bi is the assay block (Bi ∈ {1,2,3}). Ei, Li and Bi are categorical variables, while gi is a continuous variable. In this model, the transgenerational value gi is given by:
gi=(ci−25305−25)(ti−1)
where ci is the NaCl concentration, in mM, and ti ∈ {1,2} is the generation assayed. The various terms of the model correspond to: i) β(Ei,Li), the statistical interaction between environment and line; ii) α(Li), the line-dependent transgenerational effect; and iii) the intercept-based effect of block.
The data was analyzed in R [47], using the following formula to specify the model in function lmer from package lme4 [49]:
log(growthRate) ~ saltTreatment * line + line * tGenComp + (1 | block)
With the R package lsmeans [48] being then used to obtain estimates of interest:
~ saltTreatment * line
pairwise ~ line | saltTreatment
In both cases, the estimates obtained do not include contributions of transgenerational effects (by evaluating the model at gi = 0, via parameter tGenComp = 0).
L28 and L11 were further assayed in head-to-head competitions (Fig 6D). Lineages were revived and reared for two generations at 25 mM NaCl before they were set up at three NaCl concentrations: 25 mM, 225 mM and 305 mM. On the third generation, L1 larvae from the two lineages were mixed in 1:1 ratio, at a density of 103 L1s in each of two Petri dishes per replicate assay. Each replicate assay was maintained for two generations. At both the assay generations, L1 samples were collected for pool-genotyping of single nucleotide polymorphisms (SNPs). Assays were performed in three blocks, with 3 replicates per salt concentration in each of two blocks, and 4 replicates in the third block. The data for analysis was based on the L28 and L11 SNP frequency values obtained after doing calibration curves where the ratio of both lines was known (S12 Fig). For analysis, the estimated frequencies for L28 were forced to be in the interval (0.005, 0.995).
To estimate relative fitness we calculated the selection coefficients of L28 with respect to L11, for the three assay environments considered, using a mixed effects model per SNP [49]. Each model included salt treatment and generation as fixed effects, and replicates as a random effect:
yi=β0+α(Ei)ti+γ(Ri)
where yi is the logarithm of odds-ratio of the L28 allele:
yi=log(pi/(1−pi))
where Ei denotes the assay environment (Ei = 0, for 25 mM NaCl; Ei = 1, for 225 mM NaCl and Ei = 2, for 305 mM NaCl), ti denotes the generation, and Ri is the replicate (Ri ∈ {1,2,⋯,30}).
The data was analyzed in R [47], using the following formula to specify the model in function lmer from package lme4 [49]
log(OdssRatioL28Allele) ~ generation : saltTreatment + (1 | replPop)
The selection coefficients in each of the three assay environments were obtained via the point estimates for the corresponding parameters of the model.
Individual L4 genomic DNA was prepared with the ZyGEM prepGEM Insect kit following [20]. A total of 925 biallelic SNPs across the genome were assayed by iPlex Sequenom MALDI-TOF methods [50]. We chose the SNPs known to segregate in the lab adapted population, following [21]. Due to the limited amount of genomic DNA, each individual was assayed for two of the six C. elegans chromosomes, each pair of chromosomes being referred to as a region (chromosomes I and II: region 1; III and IV: region 2; V and VI: region 3). 64 L4s from the ancestral population and 16 L4s from each of the evolved populations at generations 10, 35 and 50 were sampled per region (3 replicate control populations, 4 replicate sudden, 4 replicate gradual).
Quality control was based on discarding SNPs with a high frequency of heterozygous calls, SNPs with a high frequency of genotyping failures (> 30%), and individuals in which many SNPs failed genotyping (> 25%). The 761 SNPs that passed quality control were imputed into chromosome-wide haplotypes using fastPHASE [51].
Genomic DNA from pooled samples was prepared using the Qiagen Blood and Tissue kit, and genotyped for 84 SNPs in chromosomes I, IV and V, using the iPlex Sequenom methods in 3 technical replicates for each SNP assay. In parallel, pooled gDNA was prepared to calibrate SNP L28 allele frequencies when mixed with L11 or the ancestor population at several known proportions (8–14 technical replicates each). After quality control, we retained 29 SNPs, 18 of which differentiating L28 and L11. We interpolated expected L28 frequencies from the calibration curves (S14 Fig), using Levenberg-Marquardt algorithm in R package minpack.lm [52]. For the principal component analysis of the matrix containing the frequency of the alternative alleles in each sample (Fig 7), the function prcomp in R was used.
We model an asexual population of a haploid organism, and consider deterministic environmental and population dynamics, discrete non-overlapping generations and viability selection, with the only environmentally-relevant variable being the NaCl concentration. We assume an infinite population size, such that any given lineage (genome-wide haploid haplotype) never goes extinct, and that there are no density- or frequency-dependencies, and that transgenerational effects are absent.
Following for example ref. [25], a population is composed of G lineages, such that the frequency of the k-th lineage in generation t + 1, denoted by gk(t+1), is given by:
gk(t+1)∝λk(x(t+1))gk(t)
[1]
where x(t) is the environment value faced in generation t, and λk(x) the expected number of live offspring produced by lineage k when faced with the environment x. The function λk(x) thus defines the fitness reaction norm for lineage k.
Following the genotyping setup, the genome is divided into L non-overlapping regions, and we refer to the haplotype in a region as a region-wide haplotype (RWH). A “lineage” k is described by a tuple Sk, indicating the RWHs in each region, such that Sk = (lk,1,lk,2,⋯,lk,L), and where lk,i is the RWH located in region i in lineage k. We assume that the fitness reaction norm of a lineage is an additive function of the fitness reaction norms of the RWHs in that lineage such that:
ξk(x)=log(λk(x))=log(λ(x|Θ,Sk))=∑l∈Skf(x|θl),f(x|θl)∈R
[2]
where Θ is a vector of parameters for the region-wide haplotypes, θl the parameters for RWH l, and f(x|θl) the parametric function describing the fitness reaction norm for a single RWH. We here consider f(x|θl) to be a linear f(x|θl) = alx + bl, such that θl = (al,bl) or quadratic function f(x|θl) = alx2 + blx + cl, such that θl = (al,bl,cl) of the environmental value x.
Given genotyping data at H time-points plus the ancestral, we consider distinct epochs of the experimental evolution, evaluated at generations T0,T1,⋯,TH (such that T0 = 0, T1 = 10, T2 = 35 and T3 = 50). To denote the epoch to which a certain variable corresponds, a superscript inside square brackets is used. For a single population, the frequency of lineage k in epoch h, denoted by gk[h], follows from the frequencies of the lineages in the previous epochs:
gk[h]∝exp(∑t=1+Th−1Thξk(x(t)))gk[h−1],h=1,2,⋯,H
[3]
where x(t) is the environment faced in generation t.
The ancestral population, consisting of G lineages, is described by two variables: A = (S1,S2,⋯,SG), corresponding to the RWHs present in each lineage; and g[0]=(g1[0],g2[0],⋯,gG[0]), specifying the frequency of each lineage (such that ∑k=1Ggk[0]=1).
For inferring the lineage fitness reaction norms, λk(x), we consider that A and g[0] are known. Since this is not the case in the analysis of the experimental data, we sample the pair (A,g[0]), given the experimental data, and then estimate the RWH parameters Θ, repeating these two steps multiple times (sections 1.7.6 and 1.7.7 of the S1 Text).
Under the population genetics model used, all replicate populations within a single evolutionary regime c have the same dynamics of the lineage frequencies gk[h]. Let Xc=(Xc[1],Xc[2],⋯,Xc[H]) denote the sequence of environmental values in regime c, where Xc[h]=(x(t1[h]),x(t2[h]),⋯,x(tTh−Th−1[h])),ti[h]=i+Th−1. Inference is framed in a maximum likelihood context, with contributions from each evolutionary regime, given the fitness and genotyping data. We consider without loss of generality that fitness and genotyping data are available for all epochs T0,T1,⋯,TH for each regime. The case in which data is available only for certain epochs is treated by evaluating the corresponding likelihood function only for those epochs. The S1 Text details how the input data, at the level of the replicate populations, is converted to that at the level of each regime.
Let Wc=(Wc,1,Wc,2,⋯,Wc,NE) denote the fitness data on regime c, with NE assay environments, with xm being the environmental value, and ϕc,m[h] the observed population-averaged fitness value of a population from regime c in epoch h in the m–th assay environment. We assume a log-normal model for noise in the observed values ϕc,m[h]. The log-likelihood for the RWH parameter vector Θ given the fitness data on regime c is then:
LW(Θ|Wc,Xc,A,g[0])∝−∑h=0H∑m=1NElog2(1ϕc,m[h]∑k=1Gλk(xm)gk[h])
[4]
Let Dc=(Dc[1],Dc[2],⋯,Dc[H]) be the genotyping data on regime c (note that it does not include the data on the ancestral), such that Dc[h]=(nc,l1[h],nc,l2[h],⋯,nc,lM[h]), where nc,l[h] is the number of copies of RWH l that were observed in epoch h in regime c. Then, the log-likelihood given the genotyping data on regime c is given by:
LD(Θ|Dc,Xc,A,g[0])∝∑h=1H∑lnc,l[h]log(∑k=1GI(l,Sk)gk[h])
[5]
where I(l,Sk) is an indicator function, equal to 1 if lineage k has RWH l, or equal to 0 otherwise.
Considering all evolutionary regimes C, the log-likelihood is then obtained by combining Eqs [4] and [5]:
∑c∈CLW(Θ|Wc,Xc,A,g[0])+LD(Θ|Dc,Xc,A,g[0])
[6]
Model fitting is then performed by maximizing Eq [6], using a gradient-based optimization algorithm, starting from random initial conditions.
All data and code for analysis has been archived in Dryad.org: doi:10.5061/dryad.76n6f7c. The archive consists of the following sets of files, each with a README.md for instructions on setting up the analysis and running the code: 1) input_data-genotp_data_NaCl.zip: raw genotyping data on the initial NaCl experiment (50 generations); 2) analysis_code-genotp_data_NaCl.zip: R code for preparing and summarizing the genotyping data on the initial NaCl experiment in changing environments. 3) input_data-growth_rate_data_NaCl.zip: raw growth-rate data on the ancestral population and the lines L28 and L11; 4) analysis_code-growth_rate_data_NaCl.zip: R code for the analysis of the growth-rate data. This is necessary for inference of the RWH parameters and the lineages, since the inference relies on fitness data on the ancestral population. 5) analysis_code-inferring_RWH_params.zip: R code for the inference of RWH parameters and the lineages, given the genotyping data during the NaCl experiment and the fitness data on the ancestral. 6) analysis_results.zip: the overall results of the analysis in the paper, which was the source for the figures; 7) input_data-genotp_data_NaCl_continuation.zip: raw genotyping data for the second set of experiments (30 generations); 8) analysis_code-genotp_data_NaCl_continuation.zip: R code for the analysis of the data on the second set of experiments.
|
10.1371/journal.ppat.1006112 | Early Antiretroviral Therapy Is Associated with Lower HIV DNA Molecular Diversity and Lower Inflammation in Cerebrospinal Fluid but Does Not Prevent the Establishment of Compartmentalized HIV DNA Populations | Even when antiretroviral therapy (ART) is started early after infection, HIV DNA might persist in the central nervous system (CNS), possibly contributing to inflammation, brain damage and neurocognitive impairment. Paired blood and cerebrospinal fluid (CSF) were collected from 16 HIV-infected individuals on suppressive ART: 9 participants started ART <4 months of the estimated date of infection (EDI) (“early ART”), and 7 participants started ART >14 months after EDI (“late ART”). For each participant, neurocognitive functioning was measured by Global Deficit Score (GDS). HIV DNA levels were measured in peripheral blood mononuclear cells (PBMCs) and CSF cell pellets by droplet digital (dd)PCR. Soluble markers of inflammation (sCD163, IL-6, MCP-1, TNF-α) and neuronal damage (neurofilament light [NFL]) were measured in blood and CSF supernatant by immunoassays. HIV-1 partial C2V3 env deep sequencing data (Roche 454) were obtained for 8 paired PBMC and CSF specimens and used for phylogenetic and compartmentalization analysis. Median exposure to ART at the time of sampling was 2.6 years (IQR: 2.2–3.7) and did not differ between groups. We observed that early ART was significantly associated with lower molecular diversity of HIV DNA in CSF (p<0.05), and lower IL-6 levels in CSF (p = 0.02), but no difference for GDS, NFL, or HIV DNA detectability compared to late ART. Compartmentalization of HIV DNA populations between CSF and blood was detected in 6 out of 8 participants with available paired HIV DNA sequences (2 from early and 4 from late ART group). Phylogenetic analysis confirmed the presence of monophyletic HIV DNA populations within the CSF in 7 participants, and the same population was repeatedly sampled over a 5 months period in one participant with longitudinal sampling. Such compartmentalized provirus in the CNS needs to be considered for the design of future eradication strategies and might contribute to the neuropathogenesis of HIV.
| Human Immunodeficiency virus (HIV) enters the central nervous system (CNS) early after infection and provides the basis for the development of neurocognitive impairment and potentially the establishment of latent reservoirs. Early initiation of antiretroviral therapy reduces HIV reservoir size in the periphery, but no previous study has assessed whether this strategy can also affect the HIV reservoir in the CNS. In this study, we prospectively collected and evaluated cerebrospinal fluid (CSF) and peripheral mononuclear blood cells (PBMC) from a cohort of 16 HIV-infected participants on suppressive antiretroviral therapy (ART) who started ART early (<4 months) and late (>14 months) after the timing of HIV infection. We found that early ART initiation was associated with lower molecular diversity of HIV DNA and lower levels of inflammatory markers in CSF in comparison to late ART start. We also found evidence of compartmentalized HIV DNA populations between the CSF and blood in the majority (75%) of the participants with available paired sequences, including two (66%) participants from the early ART group. Such compartmentalized provirus in the CNS will be important for the design of future eradication strategies and could contribute to the neuropathogenesis of HIV.
| Human Immunodeficiency Virus (HIV) invades the central nervous system (CNS) early during the course of infection [1,2] providing the foundations for neurocognitive impairment (NCI) and potentially establishing a latent reservoir [3,4]. Newly infected individuals typically have homogeneous HIV populations in blood [5,6] that evolve during untreated infection to generate diverse viral variants [2,7,8]. Compartment-specific selective pressures can subsequently lead to the emergence of unique HIV populations in different anatomical sites during the course of infection, including the CNS [2,7,9–11], the genital tract [12], and other tissues [13,14]. HIV RNA variants can be sequestered from blood into the CNS early after infection (within 2–6 months) and give rise to a separate HIV RNA population in the cerebrospinal fluid (CSF) [2,8], which remains genetically distinct from blood throughout the course of infection. Overall, these observations suggest that the CNS can be permissive for HIV replication from a very early period after HIV infection. The presence of compartmentalized HIV variants within the CNS has important implications: (1) compartmentalization of HIV RNA in CNS has been associated with greater inflammation and worse neurocognitive outcomes [15–17] and, (2) independent replication of HIV within the CNS might hinder HIV eradication efforts by providing a distinct reservoir of HIV persistence different from that found in peripheral CD4+ T cells. This has been suggested by previous observations reporting differential emergence of drug resistance mutations between CSF and blood during antiretroviral therapy (ART) failure [18–20].
Combination ART has markedly reduced the incidence of HIV-associated dementia [21,22]. However, the true impact of early ART initiation on HIV-associated neurocognitive impairment is still under investigation [23]. While the viral replication and evolution of HIV RNA in the CNS has been extensively studied even during early HIV infection [2,8,24,25], little is known about the HIV DNA populations persisting in this anatomic compartment during the earliest phase of HIV infection, and especially during suppressive ART. Similar to blood [26,27], initiation of ART during early HIV infection might limit the diversification of HIV DNA within the CNS, affecting the size and molecular diversity of the HIV reservoir, preventing inflammation, and limiting brain damage. But these features have not been evaluated yet for the CNS. Our study used a unique set of samples from a well-characterized cohort of HIV-infected individuals followed longitudinally from early HIV infection to investigate the effects of early ART initiation on the size and molecular and phylogenetic characteristics of the HIV DNA populations while on long-term suppressive ART. Additionally, since chronic inflammation has been associated with HIV persistence [28], we evaluated the effects of early ART on selected inflammatory markers in blood and CSF supernatant.
Study participants (n = 16) were all HIV-infected males with a median age of 41 years (Inter Quartile Range [IQR]: 32.5–52.5) selected among participants of the San Diego Primary Infection Resource Consortium (SD PIRC). At baseline (pre ART), the median plasma viral load was 176,000 HIV RNA copies/μl (IQR: 40,287–515,900). Participants achieved viral suppression after a median of 76 days (IQR: 47–256) ART start and remained undetectable during the entire follow-up (median of 3.5 viral load measurements per participant, median of 168 days between visits, median % of time-points with suppressed HIV RNA during follow-up 100%). Participants received ART for a median duration of 2.6 years (IQR: 2.2–3.7) and had suppressed levels of HIV RNA in blood plasma (<50 copies/ml) and in the CSF supernatant (at single copy level) at the time of sample collection. Six out of sixteen participants were on a protease inhibitor (PI)-based ART regimen, 6/16 were on a non-nucleoside reverse transcriptase inhibitor (NNRTI)-based regimen and 4/16 on an Integrase Strand Transfer Inhibitor (INSTI)-based regimen, all in combination with two nucleoside reverse transcriptase inhibitors (NRTI).
While we recruited participants with early and late ART initiation according to study design, the exact categorization (<4 months or >14 months) was performed retrospectively to participant enrollment, but a priori to any molecular data generation or interpretation. The “early ART group” (n = 9) started ART within a median of 1.8 months from estimated date of infection (EDI) (IQR: 1.5–3) while the “late ART group” (n = 7) started ART within a median of 17.2 months from EDI (IQR: 14.8–30.9). Detailed demographic and clinical characteristics of the study population are summarized in Table 1. No significant differences between the early and late ART groups were observed for any recorded demographic or clinical characteristics (p>0.2). Paired CSF and blood samples were obtained at baseline from all 16 participants. Two participants (both belonging to the late ART group) agreed to donate CSF and blood at a second (T0338 and T0366) and a third (T0366) longitudinal visit. These additional samples were obtained 5 and 3 months from the first evaluation and 2 months from the second evaluation, respectively.
Overall, this study comprised 16 participants with baseline samples (9 early ART and 7 late ART) and 3 extra time points from 2 participants (both belonging to the late ART group). Among the 16 baseline samples, we detected HIV DNA from 6 CSF cell pellet samples (37.5%) by ddPCR and amplified the HIV partial env gene (C2V3, HXB2 coordinates 6,928–7,344) in 8 CSF cellular samples (50%) by nested PCR (Summarized in Supplementary S1 Table). For the purpose of our study, we considered as “positive” any CSF sample with detectable HIV DNA by either ddPCR or nested PCR (or both). This resulted in 10 HIV DNA positive CSF samples at baseline (62.5%, 5 in the early ART and 5 in the late ART group) and 6 undetectable (negative for both ddPCR and nested PCR). Of the 3 extra time point samples (longitudinal), we detected HIV DNA from one CSF cellular sample by ddPCR (T0338 TP2) but we were able to amplify C2V3 env in all 3 CSF cellular samples (T0338 TP2 and T0366 TP2 and TP3).
Of note, only 5 samples (out of the 13 with detectable HIV DNA) had consistent detection of HIV DNA by ddPCR and nested PCR across both aliquots. This inconsistency across aliquots is not surprising because of the low number of infected cells which increases the proportional impact of unequal cell numbers across the two separate aliquots during processing. Also, the dilution of lysates before the ddPCR droplet generation may have significantly reduced the sensitivity of the ddPCR assay.
When comparing the two groups, HIV DNA was detected in 5 out of 9 CSF cell pellet samples tested as part of the early ART group and in 5 of 7 in the late ART group, but this difference was not statistically significant (55% versus 71%, relative risk 0.78, p = 0.63); HIV DNA was detected in all but one (93.8%) of the 16 PBMC samples.
To further characterize the HIV DNA population, we sequenced partial env from CSF cell pellets (n = 8) and PBMCs (n = 14) at baseline. For two participants, we also obtained partial env sequences from one additional time-point (T0338 and T0366). Detailed characteristics of the viral sequences are provided in supplementary S2 Table (for PBMC) and S3 Table (for CSF cell pellets).
Overall, participants in the early ART group presented a lower molecular diversity of the CSF HIV DNA population, as compared to the late ART group (Fig 1; Median: 0.9% versus 2.5%, p = 0.11). In contrast, no difference in molecular diversity was observed in the PBMC HIV DNA population between the two ART groups (Fig 1, Median: 2.1% versus 2.5%, p = 0.26). The CSF/PBMC diversity ratio was 0.58 (range: 0.31–0.69) for the early ART group and 0.84 (range: 0.33–1.06) late ART group (p = 0.12).
Next, we used a mixed-effects model where baseline viral diversity was predicted by log-transformed time to ART from EDI as a continuous variable to evaluate its association with percentage of diversity (Fig 2). We observed a higher percentage of diversity among participants with the longer time to ART from EDI, collapsed across blood and CSF (b = 0.36, p = 0.04, η2p = 0.28). When evaluating the compartments separately, this association was significant in CSF (p = 0.05, η2p = 0.22), but not in blood (p = 0.08, η2p = 0.19). Diversity was significantly higher in PBMC than in CSF by 0.8% (p = 0.02, η2p = 0.31), regardless of time to ART. We also included five covariates (age, peak viral load, CD4, CD8, and CD4/CD8 ratio) separately in the model to examine their potential effects on diversity and the association between time to ART and diversity. None of the covariates was significantly associated with diversity (all p-values>0.1, all η2p<0.05) while the association between time to ART start and diversity remained consistently significant (p-values<0.05).
The average number of input HIV DNA templates from CSF cells into the first round PCR reaction was estimated using the number of HIV DNA and RPP30 copies (based on our ddPCR data). The median HIV DNA copies per million cells among HIV-positive CSF cell samples was 2,701 copies/million cells (IQR: 1,119–4,526). The median number of CSF cells for each ddPCR reaction (estimated by RPP30) was 2,340 (IQR: 1120.5–2700 cells). After adjusting for the different volumes (5 μl for ddPCR and 10μl for nested PCR) and the dilution factor, we estimate that the average calculated HIV template input was 22 copies of HIV DNA (range: 4–64) per reaction. It should be noted that these levels are likely an under-estimate, given the inherent dilution with the ddPCR methods, as described in the method section and above.
To further evaluate if the low HIV DNA input for the sequencing reaction influenced our measures of molecular diversity, we performed additional sensitivity analyses based on our baseline model described above. We first assessed the potential impact of HIV DNA copies on diversity measures by including log-transformed HIV DNA levels (measured in blood and CSF when available) into our model; we found no statistical evidence that the number of HIV DNA copies was associated with any bias in molecular diversity (p = 0.21, η2p = 0.10). Second, to take into account the lack of consistency across aliquots, we compared diversity measures between cases with consistent versus inconsistent detectability across aliquots (assuming that cases with ddPCR+/nested PCR+ will have higher HIV DNA levels compared to cases with ddPCR-/nested PCR+) and we did not find a significant difference (p = 0.46, η2p = 0.04). While the ability to detect a significant effect in our sensitivity analysis was surely limited by the small sample size, this analysis suggests that the effect size of our primary predictor (time from EDI to ART, η2p = 0.28) on molecular diversity of partial env was greater than the effect sizes of each covariate, including the number of template HIV DNA copies (η2p = 0.10) and the number of positive aliquots (η2p = 0.04). Finally, to test the consistency of the diversity measures across blood and CSF, we performed a correlation analysis, and found that molecular diversity in CSF pellets was significantly associated with molecular diversity in PBMC (Pearson r = 0.78, p = 0.02), strongly supporting the validity of our conclusion and measurements within the context of all the aforementioned limitations.
Paired HIV DNA sequences (partial env) from CSF cell pellets and PBMCs were obtained for 8 participants, 3 from the early ART group and 5 from the late ART group. Two individuals (both from the late ART group) had additional HIV DNA sequences from a second time-point available (obtained 3 and 5 months from the first evaluation, respectively). One individual had a third time-point (2 months apart). Compartmentalization was assessed using three distinct methods: distance-based FST test with and without collapsed haplotypes and tree-based Slatkin-Maddison (SM) test. Applying our conservative definition (i.e. significant compartmentalization for all three methods), we observed a significant genetic compartmentalization between the HIV DNA populations sampled from CSF cells and PBMCs in 6 of 8 participants, including 2 individuals in the early ART group (T0104 and T0430) (Table 2). Of note, the Fst estimates were congruent between both distance-based approaches, with and without collapsed haplotypes (Kendall τ test p<0.01).
Maximum likelihood (ML) phylogenetic trees were created to evaluate the structure of the HIV DNA populations for participants with paired env sequences from CSF cells and PBMCs (Fig 3 and Fig 4).
Tree topologies revealed the presence of monophyletic HIV DNA populations in CSF for 7 participants (Figs 3 and 4, indicated with an asterisks). Two (T0104 and T0430) of the six individuals with evidence of well-segregated viral populations in the CSF were part of the early ART group. The same monophyletic CSF virus population was sampled from longitudinal CSF pellets over a period of 5 months for the one individual with a second time-point (T0338; Fig 4, see asterisk).
Next, we investigated the effect of early ART on inflammatory markers and a marker of neuronal damage. In our cross-sectional analysis (including baseline samples), participants from the early ART group had lower levels of interleukin (IL)-6 (Fig 5 and Table 1, p = 0.03) and tumor necrosis factor (TNF)-α (Fig 5 and Table 1, p = 0.02) in CSF compared to participants from the late ART group. ART groups did not differ for any of the other soluble inflammatory markers in CSF (sCD163 and MCP-1) or blood (sCD163, IL-6, TNF-α and MCP-1) or for neurofilament light (NFL) in CSF (p>0.1; Table 1).
We also used the time to ART as a continuous variable to evaluate its association with the levels of the four cytokines. We observed higher IL-6 levels among participants with the longest time to ART start from EDI, collapsed across blood and CSF (b = 0.19, p = 0.02, η2p = 0.16). When evaluated separately, this association was significant in CSF (p = 0.02, η2p = 0.16), but not in blood (p = 0.54, η2p = 0.01). Again the five covariates were included in the model to control for their potential effects. The CD4/CD8 ratio was significantly negatively correlated with IL-6 levels (b = -0.37, p = 0.05, η2p = 0.12), while the other four were not correlated (all p-values>0.1, all η2p<0.07). Regardless of the covariate included in the model, the association between time to ART and IL-6 remained consistently significant (p-values<0.05).
Since IL-6 levels and HIV DNA diversity showed a similar, positive association with time to ART, we performed an additional mediation analysis to test the hypothesis that time to ART might have influenced diversity through its effect on IL-6 levels (Fig 6). While the direct effect of time to ART on diversity was still significant (p = 0.02), its indirect effect through IL-6 levels was not (p = 0.52), suggesting that IL-6 is unlikely the main mechanism connecting shorter timing of ART initiation to lover HIV DNA diversity.
To cure HIV, all forms of viral persistence should be considered, including viral reservoirs in different tissues and anatomical compartments [2,17,29–31]. Strong evidence supports that HIV can independently replicate in the CNS during untreated infection [2,11,32] and that the virus can establish a latent reservoir in this anatomic compartment [33,34], which may be distinct from the one in circulating CD4+ T cells. The exact timing of HIV compartmentalization within the CNS is uncertain but likely occurs soon after infection in at least some individuals [2,25]. Similarly to the periphery [35–38], we hypothesized that initiation of ART during early HIV infection would reduce the size and diversity of the viral reservoir within the CNS. To test this hypothesis, we evaluated a unique cohort of 16 HIV-infected individuals with known EDI who were sampled while receiving long-term ART and with sustained HIV RNA suppression. As previously described [39], we were able to detect HIV DNA in cells collected from the CSF, even in participants who started ART during early HIV infection (within 4 months of EDI). We did observe that early ART was associated with less molecular diversity of HIV DNA in both CSF cells and PBMC compared to late ART. Molecular diversity was not associated with age, peak viral load, CD4, CD8 and CD4/CD8 ratio.
Interestingly, although early ART initiation was associated with lower molecular diversity of provirus, most participants presented evidence of genetic compartmentalization of HIV DNA within the CSF (including 2 out of the 3 participants from the early ART group). Seven participants had a clear monophyletic population of HIV DNA in the CSF. Overall, our results are consistent with previous studies reporting the presence of compartmentalized HIV RNA in CSF of HIV-infected people very early after infection [2,25]. The detection of viral compartmentalization does not necessarily imply that the populations in CSF and in blood are completely segregated, but instead, distinct subpopulations can occur in each compartment. This can occur in two different ways. First, HIV RNA populations can be sequestered from blood and populate the CNS early after infection, giving rise to a HIV RNA population within the CSF that remains genetically distinct from blood throughout the course of infection [2]. Alternatively, HIV RNA can enter the CNS early and evolve over time as a consequence of isolated replication and differential selection pressures, creating a genetically complex population within the CNS [2]. Overall, these observations suggest that the CNS compartment is permissive for HIV replication in at least a subset of persons from a very early period after infection and likely originates a distinct reservoir from that found in the blood; however, it is noted in our study that we do not know if any of these HIV DNA sequences represented replication competent proviruses.
Another open question is the cellular source of this genetically distinct HIV DNA isolated from CSF cells. In our study, we were not able to determine the exact cellular source of the HIV DNA due to technical limitations and the nature of the samples. It is possible that this genetically distinct HIV DNA population detected in CSF might be carried by macrophages or T cells into the CSF or that T-cells circulating in CSF could get infected through contact with HIV-infected macrophages residing in the brain tissue in proximity to the brain vessels [40]. Alternatively, this HIV DNA population might be originating from CD4+ T cells circulating in the CSF after crossing the blood brain barrier but this seems less likely, since HIV-infected CD4+ T cells trafficking from the periphery into the CNS should present an equilibrated viral population in comparison to blood, especially in the setting of suppressive ART. Alternatively, unrecognized isolated HIV replication within the CNS during the period before our study visit might be responsible for our observations. Unfortunately, we did not collect longitudinal CSF samples in time points previous our baseline study visits, as part of the study design.
The novelty of our study derives from the fact that we evaluated the HIV DNA populations from cells circulating in CSF and we demonstrated the presence of compartmentalized monophyletic HIV DNA populations in CSF from HIV-infected participants receiving suppressive ART, including two participants who started ART during primary infection. Both participants with longitudinal sampling showed sustained compartmentalization at all time-points, and the same monophyletic population was repeatedly sampled from CSF over a period of 5 months in one participant.
Despite several technical limitations (described below), our findings are important for the design of future eradication strategies and also to improve our understanding of HIV pathogenesis in the CNS. In fact, the presence of compartmentalized HIV populations has been associated with neurocognitive impairment [15, 41]. Several studies reported associations between circulating HIV DNA levels in blood and neurocognitive impairment with and without ART [42–46]. While this observation might hold true also for HIV DNA in CNS, this has not been consistently reported especially in the setting of suppressive ART. One previous study [3], found higher levels of HIV DNA in brain tissue from people with HIV encephalitis and moderate neurocognitive impairment compared to HIV-positive controls dying without neurologic symptoms. However, this study was limited since it included autopsy material from people dying with advanced disease and variable ART exposure. Likely due to limitations in samples size and the fact that people treated early during HIV infection have overall less neurological complications, we did not find associations between HIV DNA levels and neurological impairment.
Our study also evaluated the effect of early ART initiation on selected inflammatory biomarkers in CSF and blood. Increased inflammation has been extensively reported in the CNS during HIV and was often associated with neurocognitive impairment [47–49] even during suppressive ART [49,50]. In our study, the early ART group presented significantly lower levels of IL-6 and TNF-α in CSF (but not in blood) compared to the late ART group. We also explored the possible effect of IL-6 on molecular diversity and no mediation effect was observed. These data further support the concept that early ART initiation reduces the levels of at least some inflammatory mediators in CSF.
This study has several limitations. First of all, even though we were able to collect the volumes of CSF necessary to recover sufficient cells by the lysis buffer protocol, the detection of HIV DNA from CSF has been challenging due the low number of cells typically present in CSF in the absence of neurological symptoms and when HIV is suppressed. The low number of input cells might increase the potential for error related to sampling bias, could possibly amplify the number of false positive events from the ddPCR assay and could affect our diversity and compartmentalization analysis. To partially evaluate its impact, we performed multiple sensitivity analysis to address a possible bias in our analysis. Although we acknowledge that the small samples size has limited our statistical evaluation, our primary predictor of interest (i.e. the time to ART initiation) appears to have a greater effect on molecular diversity than the assay-related covariates. Further, we significantly elevated the threshold of compartmentalization detection and specifically included computational tests to increase robustness against significant errors in frequency estimation. Template input was particularly low in some (but not all) CSF samples, which could negatively impact our capacity to find unique clades within the CSF: assuming we are simply resampling the most common variants, we are more likely to find that CSF sequences fall within better sampled blood variants. In contrast, despite the possible sampling bias in CSF, we were still able to observe monophyletic CSF variants at baseline in several participants. Also, the reproducibility of the phylogenetic trees with similar variants sampled across longitudinal CSF samples for one participant, suggests that our sequences are likely informative and not substantially affected by random error or sequencing bias. Despite this, and the fact that we are analyzing only a partial region of env gene (~400 bp), we found differences in molecular diversity of the HIV DNA populations in CSF between the early and late ART groups.
Another limitation of the analysis is the lack of randomization for the timing of ART initiation, which might introduce some unrecognized biases in our study design. For example, people with more symptomatic infection (including the presence of neurologic symptoms, which were not tested as part of our study) will start ART earlier and might also be more likely to present compartmentalized HIV populations. The small sample size also limited our statistical power. Even though some comparisons did not reach statistical significance, effect sizes were medium to large in some cases, supporting that the study was underpowered to answer these questions.
Another limitation is inherent in all CSF studies: CSF only approximates events in the brain. Despite this, CSF has provided many important insights into brain events in HIV and other diseases [51]. A high degree of HIV DNA compartmentalization within the CSF suggests that the sampled HIV DNA is originating from brain tissue, but it could also reflect a population of cells that preferentially migrate into CSF from blood. This will need to be evaluated in future studies using larger cohorts and post-mortem brain tissues. Finally, in this study, we were also not able to determine if the HIV DNA population sampled in the CSF is replication competent.
Despite these limitations, our data provide a unique perspective by analyzing HIV DNA populations sampled using CSF prospectively collected from a unique cohort of individuals who started ART and with known EDI. Our study supports the idea that initiation of ART during early infection may limit the diversity of HIV populations and inflammation in CNS. Future studies may want to evaluate the CSF HIV DNA populations in bigger cohorts and include longitudinal assessments prior and after initiation of ART to characterize dynamics of the CNS as a HIV reservoir. Moreover, future studies need to assess the CNS replication competent HIV DNA populations. The presence of unique HIV DNA populations within the CSF during ART might be relevant for future eradication strategies.
The study was approved by the Institutional Review Board at the University of California. All adult participants (age ≥ 18 years) provided written informed consent. No children were included in this study.
Study participants were selected among HIV-infected men who enrolled in the SD PIRC between 2001 and 2012 and were still engaged in follow-up [52]. All SD PIRC participants are recruited during primary infection and followed with longitudinal blood drawn. Per protocol, visits occur at weeks 1, 2, 4, 8, 12, and 24, and then every 24 weeks thereafter. The date of infection is estimated for each participant following an established algorithm (summarized in supplementary S4 Table) [36]. Although early ART initiation is encouraged for all SD PIRC participants, implementation is based on participants’ personal decision, primary care physician input and following the current ART guidelines at the time of recruitment. Participants started ART between 2003 and 2012. Selection criteria for this study were: (1) HIV-infected males recruited during primary infection, (2) started ART during follow-up early or later during HIV-infection, (3) reached undetectable HIV RNA in blood plasma (<50 HIV RNA copies/ml) and remained undetectable during follow-up until the time of baseline CSF collection (based on our longitudinal viral loads and participant self-report) [53]. None of the participants had evidence of other inflammatory neurologic disorders or pleocytosis.
Participants were divided in early ART versus late ART groups as follow: 9 were included in the early ART group (≤4 months from estimated date of infection [EDI]) and 7 in late ART group (>14 months from EDI).
Paired blood and CSF samples were collected from each HIV-infected participant cross-sectionally. A subset of 2 participants provided a second pair of samples (3 and 5 months after their first evaluation, respectively) and one participant provided a third pair of samples (2 months thereafter).
We designed our study to maximize cellular recovery by collecting 40 ml of CSF fluid by lumbar puncture. Following standard procedures at the HIV Neurobehavioral Research Center (HNRC), the LPs were performed using atraumatic needle by an experienced physician. None of our study participants reported any complication following the CSF collection. From this larger volume, we obtained a CSF cell pellet and split it into two separate aliquots. Cell pellet lysates (containing HIV DNA) were used for ddPCR and for C2V3 env nested PCR as described below (see supplementary S1 Fig).
CSF supernatant was used to measure levels of selected markers of inflammation and neuronal damage (described below) and to measure HIV RNA by Aptima HIV RNA assay (Hologic), after concentrating 5 ml of supernatant (with single copy sensitivity). The CNS penetration effectiveness (CPE) index for the most recent ART regimen was determined as previously described [54]. For all participants, blood CD4+ T-lymphocytes were measured by flow-cytometry (CLIA certified local laboratory). Levels of HIV RNA in blood plasma were quantified by the Amplicor HIV Monitor Test (Roche Molecular Systems Inc.).
For each participant, neurocognitive functioning was assessed using a standardized clinical battery of seven ability areas consistent with Frascati recommendations for neuroAIDS research [55] and summarized using the validated global deficit score (GDS) [56].
The levels of selected markers of monocyte activation (sCD163), general inflammation (IL-6) and (TNF-α) and monocyte trafficking monocyte chemoattractive protein (MCP)-1 as well as brain damage (NFL chains were measured in all participants. Enzyme-linked immunosorbent assay (ELISA) was used to quantify the levels of sCD163 (Trillium Diagnostics, Brewer, ME, USA) from blood plasma and CSF, and NFL in CSF (Uman Diagnnostics, Sweden). Electrochemiluminescence multiplex assay (Meso Scale Diagnostics, Rockville, MD, USA) was used to quantify the levels of IL-6, TNF-α and MCP-1 in CSF supernatant and blood plasma. All assessments were performed according to the manufacturer’s procedures.
Genomic DNA was extracted from 5 million PBMC for each participant (QIAmp DNA Mini Kit, Qiagen, CA) per manufacturer's protocol. Genomic DNA was also extracted from 1 (out of 2) aliquot of cell pellets obtained from 20 mL of CSF (in average, there were 34,000 white blood cells/aliquot, range: 20,000–60,000) using direct lysis as previously described [22, 23]. Levels of HIV DNA (pol gene region: HXB2 coordinates 2536–2662) were measured in triplicate by (dd)PCR [57]. Briefly, 5 μL of 1:2 diluted CSF lysates or 1000 ng of DNA from PBMC per replicate was digested with BANII enzyme (New England Biolabs) prior to ddPCR. Reactions were performed with the following cycling conditions: 10 minutes at 95°C, 40 cycles consisting of a 30 second denaturation at 94°C followed by a 60°C extension for 60 seconds, and a final 10 minutes at 98°C. For DNA from CSF cell pellets, we used 5 μL (diluted 1:2) of lysate per replicate. A 1:10 dilution of the digested DNA was used for host cell RPP30 (ribonuclease P30) ddPCR and cycled with same parameters described above. Copy numbers were calculated as the mean of the three PCR replicates measurements and normalized to one million of cells (PBMC or CSF cells) as determined by RPP30 levels. The limit of detection of the ddPCR assay for HIV DNA using the same primer-probe set was previously described as 0.7 copies per million of cells [57]. The detected number of RPP30 copies in each ddPCR reaction was used to estimate the number of cells per aliquot of CSF cellular pellet.
We amplified the HIV-1 env C2-V3 (HXB2 coordinates 6928–7344) region from DNA extracted from CSF cellular pellets and PBMC by nested PCR using specific primers [58]. Sequencing was performed using 454 GS FLX Titanium (454 Life Sciences, Roche, Branford, Connecticut, USA). Read (FASTA) and quality score files produced by the 454 instruments were further analyzed using a purpose-built bioinformatics pipeline [25–27]. The pipeline is available at https://github.com/veg/HIV-NGS and the key steps were summarized briefly bellow:
Raw data were filtered by removing sequences of low quality (q-score of less than 15) using the Datamonkey analysis tool [59] and aligned to a subtype B reference sequence [60]. High-quality reads were retained and aligned to HXB2 as a reference sequence (without generation of contigs) using an iterative codon-based alignment procedure implemented in Datamonkey. A Bayesian Dirichlet mixture of multinomials probabilistic model was used to distinguish sequencing error from true low-frequency variants (posterior probabilities of ≥99.99%). For PBMC, we obtained a median of reads of 16927.5 [13725, 23106.5] and for CSF, we obtained a median of reads of 16198 [9590, 20157.5]. All sets of representative reads were screened for evidence of recombination using GARD [29], APOBEC signatures, hypermutations and frame-shifts as part of our pipeline procedure. All sequences were screened for in-house cross-contamination using BLAST [61]. Identical sequence reads were clustered, allowing identification of non-redundant sequences. A minimum of 10 identical sequence reads were clustered into haplotypes, and the proportion of reads in each haplotype was provided. Hence, the output consists of a list of representative haplotypes and their relative frequencies. The average number of HIV DNA haplotypes recovered from the CSF is 21 (range: 11–29), while 27 (range: 9–46) haplotypes were recovered from blood. For each sample, we computed the mean of all pairwise Tamura-Nei 93 distances between reads with at least 100 overlapping base pairs to quantify nucleotide diversity [62].
Viral compartmentalization was first assessed by the Fst approach defined as FST=1−πIπD, where πI is the estimate of mean pairwise intra-compartment genetic distance (TN93) [28], and πD is its inter-compartment counterpart [63]. Both quantities were computed by comparing all reads from blood and CSF compartments, subject to the requirement that they share at least 150 aligned nucleotide positions. The large number of pairwise comparisons (107−109) was handled computationally using an efficient implementation of the TN93 distance calculator (github.com/veg/tn93), which achieves a throughput of 107 distances/second on a modern multi-core desktop. Subsequently, to guard against inference of compartmentalization by skewing of allelic frequencies due to PCR amplification and other biases, we recomputed FST by discarding copy number counts for read clusters (i.e. each cluster was counted as having only one sequence), i.e. all haplotypes are assigned a relative weight of 1. Statistical significance of both tests was derived via 1,000 population-structure randomization/permutation test. Finally, we performed a second tree-based Slatkin-Maddison (SM) test for compartmentalization [64]. Conservatively, we defined a CSF sample as compartmentalized only if all of the following tests were consistent and significant: (1) distance based FST test, (2) sensitivity test FST with collapsed haplotypes and (3) tree- based SM test.
Viral haplotypes were realigned using MUSCLE [65], piped to FastTree 2 [66] for maximum likelihood trees reconstruction, and subjected to codon-based (MG94) phylogenetic analyses in HyPhy [67].
Statistical differences between groups (early versus late ART initiation) were examined using linear mixed-effects models with individuals included as random intercepts. The time-to ART variable was dichotomized or log transformed, and outcome variables were rank-transformed when appropriate. When residual variance differed by a specific factor in analyzing untransformed outcomes, we allowed heterogeneous variances across levels of that factor. Differences for sparse variables were detected by Fisher exact test. Whenever possible, partial η2 (η2p) was provided as a measure of the strength of association. Statistical analyses were performed using the R statistical language ver 3.3 [68] and the nlme package [69].
|
10.1371/journal.pgen.1003883 | The Drosophila eve Insulator Homie Promotes eve Expression and Protects the Adjacent Gene from Repression by Polycomb Spreading | Insulators can block the action of enhancers on promoters and the spreading of repressive chromatin, as well as facilitating specific enhancer-promoter interactions. However, recent studies have called into question whether the activities ascribed to insulators in model transgene assays actually reflect their functions in the genome. The Drosophila even skipped (eve) gene is a Polycomb (Pc) domain with a Pc-group response element (PRE) at one end, flanked by an insulator, an arrangement also seen in other genes. Here, we show that this insulator has three major functions. It blocks the spreading of the eve Pc domain, preventing repression of the adjacent gene, TER94. It prevents activation of TER94 by eve regulatory DNA. It also facilitates normal eve expression. When Homie is deleted in the context of a large transgene that mimics both eve and TER94 regulation, TER94 is repressed. This repression depends on the eve PRE. Ubiquitous TER94 expression is “replaced” by expression in an eve pattern when Homie is deleted, and this effect is reversed when the PRE is also removed. Repression of TER94 is attributable to spreading of the eve Pc domain into the TER94 locus, accompanied by an increase in histone H3 trimethylation at lysine 27. Other PREs can functionally replace the eve PRE, and other insulators can block PRE-dependent repression in this context. The full activity of the eve promoter is also dependent on Homie, and other insulators can promote normal eve enhancer-promoter communication. Our data suggest that this is not due to preventing promoter competition, but is likely the result of the insulator organizing a chromosomal conformation favorable to normal enhancer-promoter interactions. Thus, insulator activities in a native context include enhancer blocking and enhancer-promoter facilitation, as well as preventing the spread of repressive chromatin.
| Insulators are specialized DNA elements that can separate the genome into functional units. Most of the current thinking about these elements comes from studies done with model transgenes. Studies of insulators within the specialized Hox gene complexes have suggested that model transgenes can reflect the normal functions of these elements in their native context. However, recent genome-wide studies have called this into question. This work analyzes the native function of an insulator that resides between the Drosophila genes eve and TER94, which are expressed in very different patterns. Also, the eve gene is a Polycomb (Pc) domain, a specialized type of chromatin that is found in many places throughout the genome. We show that this insulator has three major functions. It blocks the spreading of the eve Pc domain, preventing repression of TER94. It prevents activation of TER94 by eve regulatory DNA. It also facilitates normal eve expression. Each of these activities are consistent with those seen with model transgenes, and other known insulators can provide these functions in this context. This work provides a novel and convincing example of the normal role of insulators in regulating the eukaryotic genome, as well as providing insights into their mechanisms of action.
| A variety of regulatory elements have evolved in higher eukaryotes to regulate gene expression. Cis-regulatory modules (CRMs, or enhancers) are bound by DNA-binding transcription factors that coordinately recruit coactivators and corepressors. Enhancers communicate with basal promoters at least in part through a looping out of intervening DNA, allowing them to act over large distances along a chromosome, or even in trans, with a promoter on another chromosome [1]–[4]. Enhancer activities are regulated by the chromatin environment, which is “managed” by both the enhancers themselves and other DNA elements such as Polycomb-group response elements (PREs) [5]–[8]. Further coordination of these activities is provided by elements such as insulators that affect chromosomal organization and conformation. Insulators harbor activities that can limit the range of action of enhancers and repressive chromatin, as well as facilitate long-range enhancer-promoter communication, depending on context [9]–[12].
Insulators typically show “barrier” function that prevents the spread of heterochromatin, as well as enhancer blocking activity, in model transgene assays [9]–[12]. Pairs of insulators can interact with each other to generate chromosomal loops between them. This has been postulated to create distinct functional domains that somehow prevent enhancer-promoter cross-talk between domains.
Repressive chromatin structures include heterochromatin and Polycomb (Pc) chromatin, which constitutes a form of epigenetic transcriptional memory, stabilizing developmental fate choices, among other functions. Pc chromatin is maintained through the recruitment of Pc-group (PcG) gene products to PREs [5]–[8]. PREs can extend their influence outward to produce Polycomb domains that encompass multiple regulatory regions within a gene or a gene complex [13]–[17]. PREs can also synergize with each other in trans [18], and in some cases facilitate long-range enhancer-promoter communication [19]. Both Pc domains and mammalian X-inactivation involve the histone modification H3K27me3, catalyzed by Pc-repressive complex 2 (PRC2) [20]–[22].
The functions of PREs and insulators have been studied within Drosophila Hox genes [23]–[25]. There, functional chromatin domains are flanked by insulators, so that all the enhancers and PREs within a domain are coordinately regulated. Enhancers acting early in development (“initiators”) are spatially regulated to determine whether a domain will be active or not throughout the rest of development. They do this by inactivating PREs, so that where initiators are active, later-acting enhancers can also be active. The main effect of deleting insulators in this context is to extend the influence of initiators to inactivate PREs in the adjacent domain, which allows its later-acting enhancers to be inappropriately active. However, phenotypic details suggest that in some cells, repressive chromatin may spread instead [26].
Genome-wide chromatin immunoprecipitation (ChIP) analysis of the locations of insulator binding proteins show a wide range of binding patterns [27]–[40]. In Drosophila, insulator proteins include dCTCF (CCCTC-binding factor), Mod(mdg4)67.2, Su(Hw) (Suppressor of hairy wing), CP190 (centrosomal protein 190), BEAF32 (boundary element-associated factor 32), and Zw5 (Zeste-white-5). Recent genome-wide studies also implicate the mitotic spindle protein Chromator [41] and the nuclear lamina [30], [37] in insulator function. In mammals, CTCF is associated with most known insulators [10], [12], [42]–[44]. CTCF functions in the regulation of β-globin [45], [46] and the imprinted Igf2 and H19 loci [47]–[49]. Based on recent genome-wide studies, it has been suggested that insulator proteins bind at many sites that do not function as predicted by model transgene assays [34], [36], [38]. Transgenic dissection in a native context can help to determine their normal functions.
The even skipped (eve) locus is a well-defined Pc domain based on genome-wide analysis [13]–[17], and is regulated by PcG genes [50]–[54]. An insulator flanks its well-characterized regulatory region, which includes the eve PRE at its 3′ end [51], [55]. Thus, this insulator is in a position to separate both positive and negative eve regulatory elements from the constitutively expressed neighboring gene TER94, and/or to prevent ectopic activation of eve by TER94 enhancers. This insulator was shown to have 3 distinct activities in model transgene assays. In addition to enhancer blocking, it causes homing of P-element transgenes to the endogenous eve neighborhood, for which it was nicknamed Homie (Homing insulator at eve). Furthermore, from within a several megabase region flanking endogenous eve, it causes long-range interactions of transgenic promoters with endogenous eve enhancers [55]. Genome-wide analysis showed that most known insulator proteins bind to the Homie region [27], [33].
Homie shares properties with other insulators based on model transgene assays and, like many other putative insulators, is situated close to both a transcription start site (TSS) and a PRE. Thus, understanding Homie's function in its native context can illuminate many of the mysteries that surround this enigmatic group of regulatory elements. In order to investigate its native function, we constructed a transgenic eve-TER94 locus that mimics the normal regulation of both genes. Using this artificial locus, we show that Homie functions as a PRE blocker to protect TER94 from repression due to spreading of the eve Pc domain. Heterologous insulators and PREs can substitute for Homie and the eve PRE, suggesting that limiting the range of PRE action is an important function of insulators generally. Homie also prevents the eve enhancers from activating TER94 in specific tissues. Furthermore, Homie facilitates normal eve expression by augmenting communication between the eve promoter and its 3′ enhancers, likely through a chromosomal looping mechanism.
Insulators are generally considered to have two major functions. First, they can shield promoters from the effects of distal enhancers. Second, they can block the spread of repressive chromatin. Here, we investigate the roles that these activities play in the normal functions of an insulator (Homie) located between the ubiquitously expressed TER94 gene and the highly patterned eve gene. We find that blocking the spread of repressive Polycomb chromatin by Homie is critical for normal TER94 promoter expression. In addition to exhibiting the canonical insulator activities in a near-native context, we find that Homie facilitates certain aspects of normal eve expression, and we present a model for how this occurs.
In order to analyze the function of the eve 3′ insulator Homie, we employed a pseudo-locus that contains all the regulatory DNA necessary for normal expression of both eve [56]–[59] and the 3′ adjacent gene TER94 [60]–[62]. This transgene extends from −6.4 to +11.3 kb relative to the eve TSS, from the 5′-most enhancer of eve to the 3rd exon of TER94. In addition to all of the eve enhancers, this region contains a characterized PRE [51] located just upstream (on the eve side) of Homie [55]. On the other side of Homie is the TER94 promoter and TSS, which are sufficient for ubiquitous expression, augmented by enhancers in the TER94 introns (data not shown). The eve coding region was replaced with lacZ coding DNA, and the 3rd exon of TER94 was fused with the EGFP coding region (Figure 1A). In this study, we make repeated use of a version of recombinase-mediated cassette exchange (RMCE) [63] that allows modified transgenes to be inserted in either orientation at pre-defined chromosomal landing sites. All aspects of transgene expression were consistent for both orientations and at multiple landing sites, with a few minor exceptions (as noted below).
In embryos, TER94 RNA is present ubiquitously at early blastoderm, and begins to fade around stage 10. Most of this RNA is maternally derived, but there is a ubiquitous zygotic contribution as well (see below). At stage 10 and later, strong expression is also observed throughout the brain and central nervous system (CNS) [55]. TER94-GFP expression from our transgene simulates endogenous TER94 expression (Figure 1B, “intact t'gene”). Although the level of expression varies somewhat with chromosomal location, the relative behavior of modified transgenes was consistent at each chromosomal location (compare Figure 1B and Figure S1).
Deletion of Homie caused a severe loss of early, ubiquitous expression driven by the TER94 promoter (Figure 1B “ΔHomie”, Figure S1). When the eve PRE was deleted in addition to Homie, the ubiquitous expression in embryos returned (Figure 1B “ΔHomie ΔPRE”, Figure S1). Deletion of the eve PRE alone did not affect the expression pattern (Figure 1B “ΔPRE”). These results show that the loss of ubiquitous expression from the TER94 promoter caused by deletion of Homie depends on the presence of the PRE. So, one function of Homie is to protect TER94 from PRE-dependent repression.
We also note that when Homie is removed, expression in an eve-like pattern is seen (Figure 1B “ΔHomie”, Figure S1A). This indicates that without Homie, eve enhancers can access the TER94 promoter. We investigate this effect further below.
Early, ubiquitous expression of TER94 comes from maternally deposited RNA, based on its early appearance and the fact that TER94 is expressed strongly in developing oocytes [60]–[62]. This was confirmed by staining for transgene expression in the absence of a maternal contribution, which is much weaker at early stages than the maternally derived signal (Figure S2 “intact t'gene”; compare to Figure 1B, Figure S1). Since TER94-GFP RNA is deposited maternally, we examined expression in ovaries. TER94 mRNA is present in both the germline, including nurse cells, and somatic epithelial follicle cells [60]–[62]. No eve expression in ovaries has been reported. In our transgenic lines, strong TER94-GFP expression was seen at all stages of oogenesis (Figure 2 “intact t'gene”, Figure S4) in both germline and somatic epithelial cells (Figure S3 “intact t'gene”). However, the level depended to some extent on chromosomal location (compare Figure 2 and Figure S4). In each case, expression was severely repressed when Homie was deleted (Figure 2 “ΔHomie”, Figure S4). As was seen in embryos, it was restored when the PRE was also deleted (Figure 2 “ΔHomie ΔPRE”, Figure S4). These data confirm that in ovaries, Homie is required for TER94 promoter activity, due to its blocking of PRE-dependent repression.
Since most of the ubiquitous TER94-GFP RNA seen in early embryos is maternally derived, we tested whether zygotic expression from a paternally-derived transgene is affected by Homie deletion. In this assay, non-transgene-carrying (yw) female flies are crossed with transgene-carrying males, so that there is no maternal GFP RNA in the progeny. Two chromosomal locations were analyzed. In both cases, GFP expression was reduced when Homie was deleted (Figure S2).
Because of the relatively low level of expression, we quantified GFP RNA using RT-PCR. Embryos from three timed collections were analyzed: 2–3 hr. (stages 5–6) and 4–6 hr. (stages 9–11) after egg deposition, and stages 13–15. The effect of Homie deletion paralleled those described above for both ovaries and embryos, in that expression was repressed. However, unlike in ovaries, when both Homie and the PRE were deleted, TER94-GFP expression remained repressed at all stages examined (Figure S2 and data not shown). This is consistent with the idea, confirmed below, that another PRE in the eve locus substitutes in embryos (but not in ovaries) for the eve 3′ PRE. In fact, the eve promoter-proximal region has PRE-like properties [51] (see Discussion).
Are the functions of Homie seen in our assays unique, or are they shared among insulators? In order to test this, we replaced Homie with other known insulators. As a negative control, a >500 bp stretch of λ phage DNA was tested. It had no effect on repression of the TER94 promoter by the eve PRE (Figure 3A “ΔHomie”). In contrast, other characterized Drosophila insulators can substitute for Homie to block repression. gypsy (Figure 3A, B), Fab-7, scs' (Figure 3A), and Fab-8 (Figure 3B) each prevented TER94 promoter repression. Although in one orientation, scs did not work (Figure 3A, “+ scs”), it did work in the opposite orientation (Figure 3A, “+ scs(inv)”). Fab-8 and gypsy showed a minor directionality in their effectiveness (not shown). Restoration of GFP expression is somewhat weaker for scs' and scs(inv) than for gypsy and Fab-7, indicating that they only partially block eve PRE action. Despite differences in efficiency, blocking of PRE action in this context is a shared property of insulators.
In order to test whether the repression of TER94 by the eve PRE is due to some unusual property associated with this PRE, we replaced it with other known PREs. We tested both the bxd PRE and an en PRE for the ability to substitute for the eve PRE in ovaries, in the context of a Homie-deleted transgene. In both cases, repression was seen at a comparable level to that seen with the eve PRE (Figure 3C, compare to Figure 3A,B “ΔHomie”), indicating that TER94 repression is due to a property shared by PREs.
We also tested whether Homie can prevent repression by heterologous PREs. To so this, we replaced the eve PRE with either the bxd PRE or the en PRE. In both cases, Homie blocked their action on the TER94 promoter, and the resulting GFP expression was like that of the wild-type transgene (Figure 3D). This shows that Homie can block repression by a variety of PREs. Taken together, these results suggest that insulators block PRE-dependent repression generally. Thus, the commonly occurring arrangement of PREs flanked on one side by insulators [31] is likely to function to provide a sharp transition in chromatin structure.
The eve locus is a Pc domain, associated with both Polycomb and the characteristic histone modification H3K27me3 [13]–[17]. We asked whether the repression of TER94-GFP in the ΔHomie transgene is accompanied by spreading of this Pc domain over the TER94-GFP promoter. Indeed, in ovaries, we found that H3K27me3 was increased in the TER94-GFP region when Homie was deleted (Figure 4A,B). Additionally removing the PRE reversed this effect almost completely (Figure 4A,B), indicating that spreading of H3K27me3 depends on the eve PRE. Thus, when TER94-GFP is repressed, H3K27me3 is increased, and when this repression is reversed, H3K27me3 levels return to normal. This suggests that Pc domain spreading is likely to be responsible for the repression.
In embryos, as in ovaries, H3K27me3 spreads into the TER94-GFP region when Homie is deleted (Figure 4C,D). However, in contrast to ovaries, additionally removing the PRE does not reverse the effect (Figure 4C,D), suggesting that there is redundancy between this PRE and other PREs in embryos. This redundant activity may be provided by the eve upstream promoter region [51], or by uncharacterized PREs within the eve locus. Again, recalling that the eve PRE is redundant in embryos for repression of TER94-GFP in the absence of Homie (Figure S2), there is a striking correlation between spreading of the Pc domain and repression of the TER94 promoter.
Intriguingly, when Homie is deleted, the loss of ubiquitous TER94-GFP expression is accompanied by weak expression in an eve pattern (Figure 1B “ΔHomie”, Figure S1A). With the intact transgene, early stripe expression of TER94-GFP driven by eve enhancers might be obscured by early ubiquitous expression, so we cannot rule out that eve enhancers are working on the TER94 promoter at early stages. In fact, eve-like stripe expression from the transgenic TER94 promoter is seen at one chromosomal landing site when the intact transgene is heterozygous and paternally derived, so that there is no maternal contribution (Figure S2B “intact t'gene”). However, eve-like mesodermal, CNS, and anal plate ring (APR) expression seems clearly to be caused by deletion of Homie (Figure 1B “ΔHomie”, stages 11 and 15; Figure S1A “ΔHomie”, stage 13), because with the intact transgene, ubiquitous expression in these tissues is low, yet no such eve-like expression is seen. Furthermore, these later-stage aspects of eve expression are not seen with a paternally derived, intact transgene (Figure S2). Therefore, the data suggest that one of Homie's functions is to prevent interaction between the TER94 promoter and eve enhancers. Accompanying the recovered ubiquitous expression when the PRE is also deleted, expression in an eve pattern is lost (Figure 1B “ΔHomie ΔPRE”, stages 11 and 15; Figure S1A “ΔHomie ΔPRE”, stage 13). This loss of mesodermal, CNS, and APR expression of TER94-GFP caused by additional deletion of the PRE indicates that the PRE not only represses ubiquitous TER94 promoter activity, but also facilitates communication between the eve enhancers and the TER94 promoter in the absence of Homie (see Discussion).
We then tested whether Homie affects eve promoter activity. To do this, we monitored transgenic lacZ expression, which is driven by the eve promoter (Figure 5, Figure S5A,B). When Homie is removed, there is a reduction in expression driven by enhancers located 3′ of the eve coding region. Interestingly, these are the eve enhancers located between Homie and the eve TSS. Comparing “ΔHomie” with the intact transgene at stage 5 (Figure 5 left column), we see that stripes 1, 4, 5, and 6 are weakened relative to stripes 2, 3, and 7. A similar reduction of expression is seen at later stages, where mesodermal, CNS, and APR expression are weakened by deletion of Homie (Figure 5 middle and right columns). This effect is seen at all transgene landing sites tested (Figure 5A, Figure S5A,B), although it varies in strength with the direction of transgene insertion (data not shown). Despite these differences, we consistently see significant disruptions of normal eve expression when Homie is removed.
It seemed possible that the effects of removing Homie on eve promoter activity were caused by the relief of enhancer blocking, which then might allow eve enhancers access to the TER94 promoter. The resulting promoter competition might reduce eve expression. Alternatively, removing Homie might cause the loss of a chromosome conformation that favors eve enhancer-promoter interactions. This possibility is suggested by the ability of Homie to promote the activation by endogenous eve enhancers of a transgenic eve promoter located up to several megabases away [55]. To distinguish between these possibilities, we performed two sets of experiments. First, we tested whether expression from the TER94 promoter occurs in a pattern that matches the loss of expression from the eve promoter when Homie is deleted. We found that this is not the case. Rather, TER94 expression in an eve striped pattern does not show a difference among the stripes (Figure 1B ΔHomie, compare to Figure 5). Second, we directly tested the promoter competition hypothesis by deleting the TER94 promoter in addition to Homie. Removing the potentially competing promoter did not restore normal eve expression (Figure 5B). While we cannot rule out competition with endogenous promoters, promoter competition seems unlikely to be the primary cause of the eve pattern disruptions that result from removal of Homie. The facilitation of eve promoter activity by Homie may therefore be due to its ability to organize specific chromosomal loops, possibly with the eve promoter (see Discussion).
Interestingly, heterologous insulators are able to restore normal eve enhancer-promoter interactions to different degrees (Figure S5C and data not shown), roughly in parallel to their abilities to restore PRE blocking (Figure 3A,B). For example, gypsy restores normal eve promoter activity, while scs does not (Figure S5C). The abilities of heterologous insulators to perform this function may be due to interactions between them and a region of the eve locus that normally interacts with Homie.
The eve 3′ insulator, Homie, was shown previously to have three activities: P-element transgene homing, enhancer blocking, and facilitation of long-range enhancer-promoter communication between endogenous eve enhancers and a transgenic promoter [55]. We sought to address how these activities relate to Homie's normal function. Both eve and TER94 are essential genes, and eve is highly dose-dependent, making it problematic to manipulate the endogenous locus. Therefore, we constructed a transgene that contains these genes in their normal configuration. Both the eve and TER94 coding regions were replaced with reporter genes to monitor promoter activity. This transgene simulates the expression pattern of both genes, when inserted at several different chromosomal sites. We used this system to manipulate both Homie and the nearby PRE, to assess their normal functions.
A major finding of this study is that Homie is required to prevent PRE-dependent repression of the TER94 promoter. Removal of Homie causes a near-complete loss of the normally ubiquitous TER94 promoter activity. Although Homie is close to the TER94 promoter, its removal does not affect the promoter directly. Rather, removing Homie allows eve enhancers to drive the TER94 promoter in an eve pattern (Figure 1). Furthermore, additional removal of the nearby PRE restores ubiquitous expression. This restoration is complete in some instances (e.g., Figures 1, 2), although it is incomplete in others (e.g., with a paternally-derived transgene in embryos, Figure S2). A simple explanation for the lack of complete restoration in some circumstances is that PRE activity varies in different tissues, and the eve 3′ PRE is partially redundant with other PREs at some times in development.
Ubiquitous expression of TER94 in early embryos, as well as some of the later ubiquitous CNS expression, is due to maternally loaded RNA. Consistent with this, expression in ovaries is robust, and, like early embryonic expression, is strongly repressed without Homie (Figures 2, S4). Accompanying repression in both ovaries and embryos, trimethylation of H3K27 at TER94-GFP is strongly increased when Homie is removed (Figure 4). Thus, without Homie, the eve Pc domain spreads into the adjacent gene, apparently shutting down expression.
Homie is bound in vivo by most known insulator binding proteins, including Su(Hw), CP190, Mod(mdg4)67.2, BEAF32, CTCF, and GAF [27], [33]. In a previous study, depletion of CTCF by RNAi in a cultured cell line caused a reduction in H3K27me3 levels throughout the eve locus [36]. The authors suggested that depleting CTCF altered the activity of insulators flanking eve, which led to a decrease in H3K27me3. In contrast, we found that deletion of Homie did not cause a significant reduction in H3K27me3 levels in the eve-lacZ region of our pseudo-locus, either in embryos or in ovaries (Figure 4). There could be several possible reasons for this discrepancy, including the cell types assayed, and indirect effects of depleting CTCF.
With removal of Homie, the spreading of H3K27me3 in ovaries is reversed by deletion of the PRE (Figure 4A,B). However, in embryos, this spreading is only partially reversed (Figure 4C,D). A simple explanation for this is that additional PRE activity within the eve locus comes into play in embryos. Consistent with this, the eve promoter-proximal region has PRE-like properties. It causes pairing-sensitive silencing of mini-white in transgenes that carry it [51], a property associated with most known PREs. Furthermore, it has consensus binding sites for several PRE-associated DNA binding proteins [8], [51], and it shares with the eve 3′ PRE the ability to support positive epigenetic maintenance of enhancer activity from embryos to larvae within eve-positive neurons [51].
Perhaps the clearest evidence for redundant PRE activity within the eve locus is that the level of H3K27me3 at the eve-lacZ coding region is not significantly reduced when the 3′ PRE is deleted. This is true in both embryos and ovaries. In contrast, spreading of the Pc domain into TER94 in ovaries requires the 3′ PRE (Figure 4). Our data are consistent with the idea that PREs are the nucleation point for spreading of the H3K27me3 mark, and that PRE activity is regulated, so that PREs are differentially active in different tissues.
Furthermore, because there may be a dynamic balance between active and repressive chromatin, maintaining a boundary between them may have different requirements at different chromosomal locations, and at different times in development. Insulators that are not required to maintain a boundary in one cell type may be required for that function in other cells, or at specific times in development, as previously suggested [34]. One reason for such differences may be regulated PRE activity.
In some cases, spreading of repressive chromatin can be stopped by an active promoter [11], [64]. In the case of the TER94 promoter, although it is robustly expressed, particularly in ovaries, this is not sufficient to stop the spreading of H3K27me3 in the absence of an insulator. This contrasts with the suggestion from recent genome-wide studies in both cultured cells and Drosophila that insulator protein function is generally not required to prevent spreading of H3K27me3 into active genes, or to maintain most normal gene expression [34], [36], [38]. Because many insulator proteins bind to overlapping sets of sites, it is likely that there is considerable redundancy in their function. Thus, knocking out any one of them may not reveal the full function of a majority of their binding sites.
It is intriguing that the eve locus is a Pc domain with well-defined boundaries that flank its extensive regulatory regions. Within chromosomal domains of the Drosophila bithorax complex (BX-C), active enhancers prevent the establishment of repressive Pc-dependent chromatin in early embryos. Conversely, in tissues where such repressive chromatin has been established, such as in parts of the CNS and imaginal discs, later-acting enhancers are repressed [25]. Do similar mechanisms operate within the eve locus? Extensive dissection of eve regulatory DNA has not identified enhancers that can drive expression outside the normal eve pattern, arguing against such a close analogy with the BX-C. However, in PcG mutants, eve is ectopically expressed throughout the late-stage embryonic CNS [50], [54], showing that PcG genes do negatively regulate eve, as they do the Hox complexes.
In our previous studies of eve PRE activity, we found that in a transgenic context, both the 3′ PRE and the PRE-like eve promoter region could facilitate positive maintenance of an eve CNS enhancer from embryonic to larval stages, as well as prevent ectopic expression in cells that normally do not express eve [51]. Unlike maintenance elements [65] in the BX-C, the eve 3′ PRE was found to require the DNA binding PcG protein Pleiohomeotic rather than Trithorax-group members for positive maintenance [51]. In this study, we also see evidence of a positive effect of the eve 3′ PRE on enhancer activity. In this case, it facilitates TER94-GFP expression in an eve pattern when Homie is removed (Figure 1B: eve-like mesodermal and CNS expression are seen when Homie is removed, but are not seen when both Homie and the PRE are removed). One possible explanation for this is that eve enhancers have evolved to function within a Pc domain, and they may be better able to activate the TER94 promoter when the Pc domain spreads over it. In this view, PREs facilitate both the on state and the off state, yet the chromatin may be differently modified in the two cases. This model is similar to the “integration model” proposed for how heterochromatin can have a positive effect on the expression of genes that normally reside within it [66].
Homie sits adjacent to the eve 3′ PRE, an arrangement that is reminiscent of some boundaries in the BX-C. The mammalian homologs of eve, evx1 and evx2, are located at the 3′ end of the HOX-A and HOX-D clusters, respectively, suggesting that the ancestral eve locus was part of a Hox cluster [67]. Consistent with conservation of the eve insulator-PRE relationship, recent studies identified an enhancer-blocking activity between evx2 and Hoxd13 [68], and a PRE in the HOX-D cluster [69]. The presence of a PRE near an insulator, with a promoter on the other side, may indicate a functionally important boundary between active and repressive chromatin domains.
Previous studies showed that within the BX-C, neither scs nor gypsy could functionally replace Fab-7 [70], indicating that there are different classes or strengths of insulators. In these cases, the primary effects of insulator deletion was ectopic activation, due to early acting enhancers (“initiators”) “turning off” PREs throughout a chromatin domain delineated by insulators [26], [71]. In our system, the major effect of insulator deletion is the spreading of the eve Pc domain, reminiscent of the shielding of transgenic reporter genes from repressive effects at some insertion sites [18], [72]–[75]. Despite the differences in normal function, BX-C insulators can replace Homie in our assay, indicating some degree of universality in insulator function as a PRE blocker. However, our assays did reveal differences in effectiveness in carrying out this function. Specifically, scs' showed slightly weaker activity than either gypsy, Fab-7, or Fab-8, while the activity of scs was highly orientation-dependent (Figure 3).
Deletion of Homie results in expression of TER94-GFP in an eve pattern. In fact, the eve early embryonic stripe enhancers may access the TER94 promoter even when Homie is present, because with a paternal-only transgene, we sometimes see eve-like stripe expression from TER94-GFP (Figure S2B “intact t'gene”). However, at later stages of embryogenesis, we do not see eve-like expression in either the mesoderm, CNS, or APR unless Homie is deleted. Therefore, one of Homie's functions is to prevent communication between the TER94 promoter and eve enhancers.
Deletion of Homie, but not deletion of the PRE, also reduced eve-lacZ expression driven by the eve 3′ enhancers (Figures 5, S5). We considered the possibility that because the TER94 promoter has access to eve enhancers in the absence of Homie, the resulting promoter competition might reduce eve promoter activity. However, in ΔHomie lines where we see TER94 expressed in eve stripes, there is no apparent bias in expression toward the 3′ enhancers (Figures 1B, S1A), arguing against this possibility. Furthermore, at later embryonic stages, eve promoter activity is reduced when both Homie and the PRE are removed (in mesoderm, CNS, and APR, which are all the tissues where eve is expressed at these stages, Figure 5A), but this is not accompanied by TER94-GFP expression in an eve pattern (Figures 1B, S1A). Finally, when the TER94 promoter is removed along with Homie, pattern disruptions persist (Figure 5B). While we cannot rule out competition with other promoters in the genome, these lines of evidence together suggest that promoter competition is unlikely to be responsible for this effect.
A second possible explanation for the reduction in eve 3′ enhancer-promoter communication when Homie is deleted is that a 3-dimensional (3-D) conformation that allows the eve promoter to better access the 3′ enhancers is stabilized by the presence of Homie. One possible conformation is a loop between the eve promoter region and Homie (Figure 6). Although we have not tested this directly, evidence consistent with this model is that activation of promoters, including the eve promoter, by downstream Gal4 binding sites can be facilitated by heterologous insulators in a model transgene assay [76]. This possible pairing of Homie with the eve promoter region would result in a loop that would bring the 3′ enhancers in closer proximity to the promoter. Such a model is similar to that proposed for the 3-D organization of regulatory regions upstream of the Abd-B gene [25]. If such loops are anchored to large clusters of insulator proteins, perhaps within insulator bodies, this may serve as a 3-D barrier that separates distinct chromatin domains, and occludes interactions between regulatory elements located on opposite sides of the insulator. At the same time, otherwise distant elements can be brought closer together, facilitating specific enhancer-promoter contacts, particularly if those elements are brought to the same side of the 3-D barrier.
The activities of Homie and the eve PRE are largely interchangeable with those of other insulators and PREs, respectively, in our assay system. Previous studies showed that Homie and the eve PRE have the canonical properties of insulators and PREs when tested in other contexts [51], [55]. Thus, our results are likely to be applicable to many such elements throughout the genome. In particular, a common function of insulators is likely to be to limit the action of PRE-dependent repressive chromatin.
Genome-wide studies using RNAi to knock down specific insulator proteins suggested that insulators may not typically be required in their normal context either to block enhancer-promoter cross-talk or to prevent the spread of repressive chromatin [34], [36]. Our results suggest that Homie is critically important in its normal context for just such activities, functionally separating the loci on either side. Importantly, other insulators function in place of Homie. This suggests that the activities of insulators defined in model transgene assays do in fact correspond to their normal functions. In particular, as with Homie and the TER94 promoter, the tendency of insulator proteins to cluster just upstream of promoters suggests that one of their typical functions is to shield basal promoters from the effects of upstream CRMs, especially PREs. Further, our finding that insulators facilitate enhancer-promoter communication in this context suggests that their ability to organize chromosomal conformations that augment appropriate transcription is also likely to be a common mode of endogenous insulator function.
The eve-TER94 locus construct (“intact t'gene” in figures) was created as follows (detailed sequence coordinates are given in Figure S6). DNA from −6.4 kb to +166 bp relative to the eve TSS was fused to the lacZ coding region. The 3′ end of the lacZ coding region was fused to DNA from +1.3 to +11.4 kb, which includes the eve poly-A signal, and extends into the 3rd exon of TER94. This was joined with the EGFP coding region, followed by the poly-A signal of α–tubulin. The entire construct was placed between two inverted attB sequences [63], [77]. The following deletions were then made in this construct: from +8.4 to +9.2 kb for ΔPRE, from +8.4 to +9.7 kb for ΔHomie ΔPRE, and from +9.2 to +9.7 kb for ΔHomie. To test promoter competition between eve and TER94, DNA from −7.4 to +8.6 kb relative to the eve TSS was used, with the eve coding region replaced by that of lacZ, as described above. This construct does not contain the TER94 promoter.
Replacements of Homie with either heterologous insulators or phage λ DNA were created using the ΔHomie construct, and adding DNA fragments corresponding to gypsy [78], Fab-7 [79]–[82], Fab-8 [83], [84], scs [85], [86], scs' [85], [86], or λ DNA (see Figure S6 for details). For testing repression activity of heterologous PREs, either the engrailed 181PRE [87] or the bxd PRE [88] were inserted into the ΔHomie ΔPRE construct at the site of deletion. For testing Homie activity against these PREs, either the en PRE or the bxd PRE were inserted into the ΔPRE construct at the site of deletion.
All transgenic lines were made using φC31 recombinase-mediated cassette exchange (RMCE) [63]. Three alternative attP target sites were used, at cytological locations 95E5, 74A2, and 30B5. The direction of each insertion was determined by PCR. Both directions were analyzed if obtained. Some variations with insertion site were found, as described in Results.
Embryos were collected at time points described in figure legends, and subjected to in situ hybridization using DIG-labeled anti-sense RNA probes against either lacZ or GFP. Expression patterns were visualized by alkaline phosphatase-conjugated anti-DIG with BCIP and NBT as substrates (Roche Applied Science).
GFP expression was detected by fluorescence microscopy in ovaries dissected from 1–2 day-old females. In some cases, expression was also detected using anti-GFP antibody staining (Roche Applied Science), analyzed by confocal microscopy (Zeiss) of material in DAPI-containing mounting medium.
Ovaries were dissected from 2–3 day-old females. Fifty ovaries were cross-linked in 1.8% formaldehyde in PBS for 10 min. After sonication so as to produce a peak near 500 bp in the DNA fragment size distribution, isolated chromatin was immunoprecipitated with anti-H3K27me3 (EMD Millipore), and with rabbit IgG (Jackson ImmunoResearch) as a negative control. Precipitated chromatin samples were collected using ProteinG magnetic beads (EMD Millipore). Immunoprecipitated DNA samples were dissolved in 20–50 µl TE, and 1 µl was used for each PCR reaction. Either duplicate or triplicate samples were analyzed by real-time PCR (Life Technologies, StepOnePlus), using SYBR Green Master Mix with ROX dye (Roche Applied Science). Data were analyzed with StepOne software (Life Technologies), using the standard curve method. Standard deviations were calculated using Excel software (Microsoft). Embryo ChIP analysis was described previously [51], except that results were quantified by real-time PCR, as described above for ovary analysis.
Specific ChIP signals were determined by subtracting the average non-specific IgG signal from the average α-H3K27me3 signal, with standard deviations combined by adding. Errors bars for specific signals relative to that of endogenous eve were determined by adding the relative errors in quadrature; that is, by taking the sum of the squares of the relative standard deviations (the standard deviations divided by their respective averages) to give the square of the relative standard deviation of the ratio.
The following primers were used: TCCAGTCCGGATAACTCCTTGAAC and TGTAGAACTCCTTCTCCAAGCGAC for the endogenous eve coding region, TGAAGCCACCGCGTGGTATTCTTA and TTTGGACATGATCTCCGGTCCGTT for the endogenous TER94 coding region, GCTGTGCCGAAATGGTCCATCAAA and TACTGACGAAACGCCTGCCAGTAT for the transgenic eve-lacZ coding region, and GGGCACAAGCTGGAGTACAACTACAA and TGGCGGATCTTGAAGTTCACCTTG for the transgenic TER94-GFP coding region.
Total RNA was purified from either five pairs of ovaries from 2–3 day-old females or 10–20 µl of dechorionated embryos for each data point, using an RNA purification kit (Roche Applied Science). RNA was eluted in 50–100 µl elution buffer and stored at −80°C. cDNA was synthesized using the Transcriptor first strand cDNA synthesis kit (Roche Applied Science), and quantified by real-time PCR as described above. A constitutively expressed RNA, RpL32 (a.k.a. RP49), was used to normalize GFP RNA levels. The primers listed above for TER94-GFP were used for GFP, and AAGCCCAAGGGTATCGACAACAGA and TGCACCAGGAACTTCTTGAATCCG were used for RpL32.
|
10.1371/journal.ppat.1000111 | Pathogenesis of Listeria-Infected Drosophila wntD Mutants Is Associated with Elevated Levels of the Novel Immunity Gene edin | Drosophila melanogaster mount an effective innate immune response against invading microorganisms, but can eventually succumb to persistent pathogenic infections. Understanding of this pathogenesis is limited, but it appears that host factors, induced by microbes, can have a direct cost to the host organism. Mutations in wntD cause susceptibility to Listeria monocytogenes infection, apparently through the derepression of Toll-Dorsal target genes, some of which are deleterious to survival. Here, we use gene expression profiling to identify genes that may mediate the observed susceptibility of wntD mutants to lethal infection. These genes include the TNF family member eiger and the novel immunity gene edin (elevated during infection; synonym CG32185), both of which are more strongly induced by infection of wntD mutants compared to controls. edin is also expressed more highly during infection of wild-type flies with wild-type Salmonella typhimurium than with a less pathogenic mutant strain, and its expression is regulated in part by the Imd pathway. Furthermore, overexpression of edin can induce age-dependent lethality, while loss of function in edin renders flies more susceptible to Listeria infection. These results are consistent with a model in which the regulation of host factors, including edin, must be tightly controlled to avoid the detrimental consequences of having too much or too little activity.
| Like any organism, fruit flies respond to invading microorganisms by mounting an immune defense. Many aspects of the immune defense in fruit flies are similar to the inflammatory response in mammals, including the harmful effects of a sustained response against persistent pathogenic infections. We found in the past that mutations in the gene wntD cause flies to succumb more easily to Listeria monocytogenes infections, apparently by losing an element of control over the inflammatory response. How does the wntD gene work? In this paper, we have identified genes that may mediate the susceptibility of wntD mutants to lethal infection. These genes include eiger, a homolog of the mammalian TNF gene, and a previously uncharacterized gene called edin (elevated during infection). Edin is expressed excessively in wntD mutant flies, and its expression also correlates with the level of pathogenesis induced by two different strains of Salmonella typhimurium. In its own right, overexpression of the edin gene can induce lethality, while losing edin function renders flies more susceptible to Listeria infection. Our results support a model in which the regulation of host factors, including edin, must be tightly controlled to avoid the detrimental consequences of having too much or too little activity.
| Drosophila has an effective innate immune system to combat infection. This response relies heavily on the Toll and Immune deficiency (Imd) pathways, both of which utilize NF-κB related transcription factors as central mediators of signaling: Dorsal and Dorsal-related immunity factor (Dif) in the case of Toll, and Relish (Rel) in the case of Imd (reviewed in [1]–[3]).
The Toll and Imd pathways have largely been characterized with respect to their role in the humoral immune response, a branch of immunity that is triggered through recognition of microbial molecular signatures by upstream components of both the Imd and Toll pathways and subsequent nuclear translocation and activation of the cognate NF-κB factor(s). The activation of these transcription factors leads to transcription of hundreds of genes following infection [4]–[6]. The most studied are the antimicrobial peptide (AMP) genes, which are transcribed in the fat body, leading to secretion of these peptides into the circulating hemolymph (reviewed in [7]).
In addition to its role in AMP regulation, the Toll pathway is also known to participate in two other branches of immunity: the deposition of melanin and the cellular immune response [8]–[12]. The cellular response in particular has become of increasing interest, as studies of Drosophila immunity progress beyond the characterization of acute responses to non-pathogenic bacteria to those involving chronic infections that eventually kill the fly [13]–[16]. Many of these model infections proceed intracellularly within the phagocytic cells of the circulating hemolymph, thereby shielding the bacteria from the action of circulating AMPs. This provides a convenient model system for studying the molecular interactions between pathogens and their hosts, including the processes that eventually lead to the host's demise.
One principle that has been understood in mammals for decades, and seems to also be true in Drosophila, is that an immune response can be both beneficial and detrimental to a host. Indeed, the same signals that are critical to containing a localized infection will kill the host if uncontrolled [17]. One such signal is Tumor Necrosis Factor (TNF), which is both necessary to fight local infections of many organisms and sufficient to induce lethal septic shock if released systemically [18],[19]. Homologous processes may also occur in Drosophila; loss of function mutations in the TNF family member eiger result in prolonged survival during infection with Salmonella typhimurium [14],[20]. Thus Drosophila offers an appealing genetic system to uncover host genes that may have dual effects during the immune response, mediating deleterious consequences to both the pathogen and the host itself.
Previously, we reported evidence that flies mutant for the Wnt family member wntD have a defective immune system and succumb prematurely to infection with the gram-positive, lethal bacteria Listeria monocytogenes [21]. Given that WntD acts as a feedback inhibitor of Toll-Dorsal signaling during embryonic development [21],[22], we presented a model in which wntD mutants exhibit a hyperactivated immune system, including the overexpression of specific Dorsal target genes that are deleterious to the flies' health. Here, we extend those observations by using Affymetrix oligonucleotide arrays to examine the whole genome transcriptional profiles of wntD mutants prior to and following infection with L. monocytogenes. We examine two groups of candidate mediators of the decreased survival of wntD mutants, and provide evidence that one of those genes, edin (elevated during infection; synonym CG32185), could be a novel effecter of pathogenesis.
In order to gain insight into the processes that are misregulated in wntD mutants and that may contribute to their susceptibility to L. monocytogenes infection, we collected RNA from wntD and control flies under two conditions: naïve and 24 hours following infection with L. monocytogenes. This time point was chosen because we had observed significant mortality of wntD mutants between 24 and 48 hours under these infection conditions, and hypothesized that misregulation of genes causally involved in this mortality would be seen most clearly at the beginning of this time window [21].
Previously, we showed that wntD mutants exhibit elevated expression of the AMP Diptericin prior to and following infection with the non-pathogenic bacterium Micrococcus luteus, while the AMP Drosomycin is expressed in wntD mutants at levels indistinguishable from wild type [21]. To test the idea that wntD mutants have a hyper activated basal immune system on a more global scale, we used our array data to look at the correlation between each gene's response to infection in wild type (log2(infected controls/uninfected controls)) and its level of misregulation in wntD mutants prior to infection (log2(uninfected wntD/uninfected controls)). As shown in Figure 1, the top thirteen genes most induced by infection all showed higher levels of expression in uninfected wntD mutants compared to uninfected controls. Of these thirteen genes, seven showed an average of greater than 2-fold difference between mutants and controls and had p-values less than 0.025 (Figure 1 and Table 1). This set of genes was comprised of the novel immunity gene edin, IM23, AttD, AttB, AttA, DiptB, and Def, all of which are known to be induced by infection under various conditions [5],[23],[24]. It is worthwhile noting, however, that several known immune-regulated genes that were strongly induced by infection in our study showed no significant difference between wntD mutants and controls, including CG6639, CecB, TotM and Dros (Figure 1 and data not shown).
Overall, the correlation coefficient (r) for these data sets was 0.14, with a p-value<0.0001. Calculating the coefficient of determination (r2) suggests that approximately 2% of the variation within the data can be explained by the correlation between the two data sets. This corresponds to approximately 235 genes, a plausible number given previous studies have indicated that about 400 genes are significantly regulated by infection [4]. In a similar analysis looking at the misregulation of immune genes in wntD mutants following infection, no significant correlation was observed (data not shown). As is evident from the cluster analysis presented below and the data in Table 1, a subset of immune-induced genes were expressed more highly in wntD mutants following infection, but many of the most highly induced immunity genes were not significantly different between wntD mutants and controls, and some were expressed at lower levels in the mutants. This may have resulted from a lack of sensitivity from the array at these high levels of expression, saturation of the signaling processes leading to induction of expression, or dominant negative effects of activated Dorsal on the activity of other NF-κB proteins.
To identify genes as candidate mediators of wntD mutants' infection sensitivity, cluster analysis was used [25]. Hierarchical clustering revealed several distinct groups of genes that showed correlation in their expression patterns across the four different conditions. However, two related clusters of genes were selected for further analysis based on the following rationale: the expression of genes actively contributing to pathogenesis will most likely be elevated following infection, and genes within this group that might be implicated in the more rapid lethality seen in wntD mutants would be expressed higher in these mutants. The average expression level under each condition for the two selected clusters (Clusters A and B) are shown in Figure 2. The clusters differ in that Cluster A shows a greater overall change in response to infection than does Cluster B (Figure 2).
Cluster A includes a number of known targets of infection, including several AMPs (Table S1). While it is certainly possible that several of these are contributing to pathogenesis in the fly, one uncharacterized gene in particular stood out based on its levels of expression. Confirmed by quantitative RT-PCR, edin shows strong induction by Listeria infection (∼45 fold), and dramatically higher levels of expression in infected wntD mutants versus infected controls (∼7.5 fold) (Figure 2B). Furthermore, only a 1.7 fold difference was seen between mutants and controls prior to infection, illustrating synergy between Listeria infection and the absence of wntD function on the regulation of edin.
Cluster B is composed of genes that show less dramatic changes in response to infection, but are still elevated in wntD mutants versus controls (Figure 2A, Table S2). It seems likely that this set of genes would include those that are regulated by processes aside from those sensing acute infection (Toll, Imd), and may include both mediators and markers of pathogenesis. Interestingly, this cluster includes the gene eiger, a TNF homolog known to mediate disease processes following Salmonella and Mycobacterium infections [14],[20]. In this case, using quantitative RT-PCR, we see a statistically significant elevation of eiger expression only in infected wntD mutants (Figure 2C).
The edin gene is predicted to encode a secreted protein 115 amino acids in length (http://flybase.bio.indiana.edu/.bin/fbidq.htmlFBgn0052185). The gene has homologs in other insects, but not in other Phyla. (Figure 3). Furthermore, no known conserved domains were identified in Edin or its putative ortholog in Drosophila pseudoobscura and secondary structure prediction failed to identify any similar proteins or motifs based folding patterns (data not shown).
To answer the question of whether edin misregulation in wntD mutants is specific to infection with Listeria, wntD and control flies were injected with the non-pathogenic gram-positive bacteria Micrococcus luteus. Analysis of Edin expression levels prior to and following infection were monitored using quantitative RT-PCR (Figure 4A). The results are strikingly similar to those seen for Listeria infection; expression of edin is elevated 1.7-fold in wntD mutants compared to controls prior to infection, and 8-fold following infection. Again, a synergistic relationship is seen between infection and the presence of the wntD mutation. The smaller magnitude of edin induction seen in response to M. luteus compared to Listeria (∼10 fold versus ∼45 fold in wild-type flies) may be explained by the shorter timecourse of infection (5 hours versus 24 hours), a smaller bacterial load at the time of assay, or intrinsic differences between the two species of bacteria.
The strong regulation of edin in response to bacterial challenge raises the question of whether its transcription is regulated by the Toll and/or Imd pathways. To investigate this possibility, the induction of edin was monitored in genetic backgrounds each containing a loss of function mutation for a component in one of the pathways (Figure 4B). Mutations in imd reduced the expression of edin following infection to approximately 25% of that seen in wild type. This indicates that the Imd pathway participates in edin regulation, but is not strictly required for its induction following infection. By contrast, loss of function mutations in the Toll ligand spatzle did not reduce the transcriptional induction of edin, and in fact resulted in higher than normal levels of expression. This has been seen for other genes (such as diptericin) that do not have a strong requirement for Toll signaling, and could be due to increased survival of the bacteria in these mutants (data not shown; [4]). Levels of edin were slightly elevated (4-fold) in naïve flies carrying a dominantly activated allele of Toll in the absence of infection (Toll10b; Figure 4B). These data indicate that Toll signaling may be sufficient to induce low levels of edin expression, but is not required for its expression.
In order to investigate whether Edin plays an essential role in disease progression, we knocked down its expression using two independently made UAS-driven RNA interference (RNAi) constructs. Edin expression was knocked down using the fat body driver Lsp2-Gal4 to ablate its activity in a major immune tissue. Edin knockdown flies displayed increased sensitivity to Listeria monocytogenes, with flies dying significantly faster than all controls (p<0.001) (Figure 5). This demonstrates that edin is required for an effective host response against Listeria infection. Interestingly, bacterial loads in edin knockdown flies were not significantly different from controls (data not shown). This places edin among several previously identified genes that affect a fly's endurance during Listeria infection rather than its ability to combat bacterial growth [26]. While the mechanism for this effect is unknown, we hypothesize that knockdown of edin expression alters the physiology of the fly in a way that makes it more susceptible to Listeria pathogenesis.
Immunity can be a double-edged sword that has to be regulated precisely to help defend against infection while limiting damage to the body. Overexpression of genes misregulated during an immune response led us to edin and we found that it is required for fly survival during an L.monocytogenes infection. Next, we thought it was of great interest to determine whether Edin expression contributed to pathology. We first looked for more evidence that Edin was associated with pathology under different circumstances. We compared the expression of edin following infection of wild-type flies with wild-type Salmonella typhimurium or a SPI1, SPI2 mutant strain of Salmonella that has decreased pathogenicity [14]. As shown in Figure 4C, edin was expressed at significantly higher levels during the course of a wild-type Salmonella infection compared to the less pathogenic strain at both time points tested. The more dramatic difference was seen later in infection, when edin transcript levels were over 5-fold higher in flies infected with wild-type Salmonella (Figure 4C). These data add more correlative evidence that Edin is associated with pathogenesis.
Do edin expression levels affect survival? To answer this question, we first overexpressed edin using the UAS/Gal4 system. Two different insertions of the p-element carrying UAS-edin resulted in varied levels of expression, with one insert (19-3) causing expression levels ∼100 fold over wild type when combined with actin-Gal4, and the other overexpressing edin over 500 fold (Figure 6A). We observed that the higher level of expression resulted in significant levels of lethality prior to and following eclosion (Figure 6B,C). Flies strongly overexpressing edin survived to adulthood at a frequency less than 50% of expected, compared to 111% for the lower expresser. The value greater than 100% can most likely be attributed to non-specific deleterious effects of carrying the CyO balancer. The average lifespan of those flies surviving to adulthood was also significantly reduced in the context of strong overexpression of edin (Figure 6C). Given that wntD mutants infected with L. monocytogenes displayed similar levels of expression to the strong insertion of UAS-edin (about 350 fold over uninfected wild-type flies; Figure 2B), it is possible that edin expression is contributing to the rapid mortality of these mutants.
Taken together with the observation that edin loss of function mutants show increased sensitivity to L.monocytogenes, these data support a model in which edin expression must be tightly controlled during a host response to infection: moderate induction is essential to an effective response, but uncontrolled, high levels of expression become detrimental to the host animal.
The idea that an elevated immune response could be detrimental to an infected host is at first unintuitive. However, it is well established that, like most other biological processes, proper regulation and containment of the immune response is critical to an animal's viability. In mammals, LPS-triggered TNF release at a site of injury/infection is critical to mobilize the immune and inflammatory processes required to fight the infection, but in the rare cases when this reaction becomes uncontrolled and systemic, the shock will rapidly kill the host [17]. Studies in the fly have shown that genetic removal of a TNF-like molecule called Eiger increases flies' longevity during some infections, but decreases it during others [14],[20]. Thus eiger appears to be a double-edged sword – necessary for fighting some infections, but not without a cost to the host. Similarly, flies carrying Tl10b mutations, which dominantly activate the Toll pathway, die more rapidly from Drosophila X virus infection, despite lower viral loads [27],[28], and over-activation of the IMD pathway has a negative impact on larval survival during bacterial infection [28]. These results imply that both the Toll and IMD pathways activate the transcription of genes that have a deleterious effect on a fly's survival during pathogenic infection, one of which could well be eiger. In light of these findings, the observation that wntD mutants die more quickly from Listeria infection, while hyperactivating immune genes, is less surprising. Furthermore, this phenotype is suppressed by loss of dorsal, implying that Dorsal is actively regulating processes that decrease the fly's survival [21].
We presented two experiments that compared the expression profiles of flies undergoing two different levels of pathogenesis: wntD versus control flies following L. monocytogenes infection, and wild-type S. typhimurium versus a SPI1, SPI2 mutant strain. In both cases the gene edin was strongly elevated in the flies closer to death. In comparing wntD mutant versus control flies following Listeria infection, RNA samples were taken 1 day after infection, shortly before the mutants exhibit a sharp decrease is survival [21]. Expression of edin was about 8-fold higher in the wntD mutants. Similarly, at 7 days post Salmonella infection, flies infected with wild type have begun to die, while those infected with a SPI1,SPI2 mutant strain will live for several more days despite carrying dramatically higher loads of bacteria [14]. In this case, we observed a 5-fold elevation in edin expression in the flies beginning to die. Thus, high edin expression is correlated with increased pathogenesis, although a causal relationship is not established by these data.
Two results strongly suggest that edin induction is not downstream of pathogenesis. First, edin expression is elevated following infection with M. luteus, a non-pathogenic bacterium, and is more strongly induced in wntD mutants (Figure 2A). These data demonstrate that pathogenesis is not required for edin expression. Second, the Imd pathway appears to play a significant role in regulating edin, and this pathway is acutely induced upon recognition of bacterial moieties and does not strictly depend on pathogenesis [29]–[31].
Could Edin play a causal role upstream of pathogenesis? The induction of edin during M. luteus infection without any demonstrable pathogenesis suggests that the amount of Edin produced during this infection is not sufficient to elicit pathogenesis. However, these levels are approximately 5-fold lower than those seen for Listeria infection and persist for less than a day (data not shown), in contrast to the chronic induction during infection with Listeria or Salmonella. Furthermore, the lethality induced by strong chronic overexpression of edin using the UAS/Gal4 system implies that this gene can induce processes detrimental to a fly's survival that could be affecting viability during persistent infections. Though Edin can be shown to cause pathology when overexpressed, it is difficult to produce clean evidence that this occurs during infection, because the overexpression of many genes can cause pathology; therefore it remains a suggestion.
Edin shows several characteristics consistent with it being an AMP. First, it is strongly induced by infection; edin was the second most highly induced gene in wild-type flies following L. monocytogenes infection, and the most highly induced gene in wntD mutants. Second, edin is predicted to encode a short peptide and a processed form has been observed circulating in the hemolymph of infected flies [23]. However, edin also displays properties that would make it unique among AMPs, suggesting that it may be more broadly affecting physiology, perhaps in a cytokine-like role similar to that of eiger. For instance, the expression of this gene is required for normal survival following L. monocytogenes infection. While necessity for the signaling pathways controlling AMP expression is well documented, this is the first case of an individual putative AMP being necessary to fight infections {Ferrandon, 2007 #329}. This requirement during infection, combined with the toxicity observed upon overexpression suggests that Edin may be a powerful component of the immune response that must be tightly regulated to optimize survival. Further analysis of edin and other genes that are differentially regulated during pathogenesis could provide interesting clues into the complicated and evolving nature of the host-pathogen interaction.
The construction of wntD mutants was described previously [21]. Any reference to wntD mutant is the genotype yw; wntDKO1. References to ‘wild type’ refer to yw; +/+; +/+ or w1118; +/+; +/+ if so noted. pP[UAS-edin] was constructed by amplifying the edin open reading frame using PCR, and cloning this fragment into the Xba-1 site of pPUAST [32]. UAS-RNAi(edin)2 was created at the VDRC (transformant 14289). UAS-RNAi(edin)1 was generated by PCR amplification of the complete cDNA with XbaI sites at both 5′ and 3′ ends. This fragment was subcloned into the pWIZ vector [33] in two sequential cloning steps on either side of a small intron in a 3′to 5′/5′to 3′ orientation. Expression of the double-stranded RNA is under the control of the UAS promoter and is transformed into a snapback hairpin upon splicing of the small intron. Flies carrying expression constructs were created using standard p-element transformation techniques.
All injections were done using male flies aged one week post eclosion. A culture of Listeria monocytogenes was diluted to an OD(600) of 0.1, and a 25 nL volume was injected abdominally using a pulled glass needle as previously described [15]. Groups of 20 flies of each genotype were injected in an alternating manner to control for variability over time. Flies were maintained on non-yeasted, standard dextrose medium at 25°C, 65% relative humidity, and survival was monitored daily. Micrococcus luteus and Salmonella typhimurium was injected as described for L. monocytogenes. For experiments on the regulation of edin, flies of different genetic backgrounds were injected with a mixture of M. luteus, L. monocytogenes, and E. coli, each at an OD(600) of 0.1.
Groups of 6 flies were collected, crushed in 150 µl of Trizol reagent, and RNA was extracted according to the manufacturer's recommendations. 1 µl RNA was used for subsequent reverse transcription using the ThermoScript RT-PCR system (Gibco BRL), following the manufacturer's instructions and using a random hexamer as primer. Quantitative PCR was preformed in a LightCycler (Roche), using the LightCycler FastStart DNA MasterPLUS SYBR green I kit (Roche) and following the manufacturer's recommendations.
Primers used for PCR were as follows:
Groups of 30 yw;wntDKO1 or yw flies (some previously infected with Listeria monocytogenes as described above) were collected in 1.5 mL microfuge tubes. Each experiment was done in triplicate, for 12 total samples. Conditions were: yw uninjected, yw;wntDKO1 uninjected, yw 24 hours post Listeria infection, yw;wntDKO1 24 hours post Listeria infection. Flies were crushed in 1 mL Trizol reagent, and RNA was isolated using the manufacturer's recommendations. 15 µg of each RNA sample was then used for cDNA synthesis, which was done using the one cycle cDNA synthesis (Affymetrix) and following the manufacturer's recommendations. cRNA was also synthesized using the manufacturer's protocol, and 20 ug was used for the subsequent fragmentation step. cRNA was hybridized to Affymetrix Drosophila Genome 2.0 arrays by the Stanford Protein and Nucleic Acid Biotechnology Facility (http://pan.stanford.edu). Arrays were analyzed using the Affymetrix GCOS software to produce normalized values for each probe set on each array.
Clustering was performed on a dataset in which genes were included only if they were marked as “present” by GCOS in all 3 samples of at least one condition. Clustering was done using Cluster 3.0 for Mac OS X (http://bonsai.ims.u-tokyo.ac.jp/mdehoon/software/cluster/software.htm). Parameters used for clustering were: Data was log transformed and genes were centered. Data was filtered to include only genes where the difference between the highest and lowest values was greater than or equal to 1 (representing a two-fold change or greater). Hierarchical clustering was performed using the centroid linkage algorithm. Clusters were viewed using Java Treeview software (http://genetics.stanford.edu/alok/TreeView/). Gene identities and annotations shown in Tables S1 and S2 were retrieved using the Netaffx analysis webpage (http://www.affymetrix.com/analysis/index.affx).
|
10.1371/journal.pbio.1000123 | Direct Binding of pRb/E2F-2 to GATA-1 Regulates Maturation and Terminal Cell Division during Erythropoiesis | How cell proliferation subsides as cells terminally differentiate remains largely enigmatic, although this phenomenon is central to the existence of multicellular organisms. Here, we show that GATA-1, the master transcription factor of erythropoiesis, forms a tricomplex with the retinoblastoma protein (pRb) and E2F-2. This interaction requires a LXCXE motif that is evolutionary conserved among GATA-1 orthologs yet absent from the other GATA family members. GATA-1/pRb/E2F-2 complex formation stalls cell proliferation and steers erythroid precursors towards terminal differentiation. This process can be disrupted in vitro by FOG-1, which displaces pRb/E2F-2 from GATA-1. A GATA-1 mutant unable to bind pRb fails to inhibit cell proliferation and results in mouse embryonic lethality by anemia. These findings clarify the previously suspected cell-autonomous role of pRb during erythropoiesis and may provide a unifying molecular mechanism for several mouse phenotypes and human diseases associated with GATA-1 mutations.
| Red blood cell production, or erythropoiesis, proceeds by a tight coupling of proliferation and differentiation. The earliest erythroid progenitor identifiable possesses remnant stem cell characteristics as it both self-renews and differentiates. Each progenitor gives rise to more than 10,000 cells, including secondary progenitors. Yet, during the next stage of differentiation, much of this renewal capability is lost, and terminal erythroid differentiation progresses in a stepwise manner through several stages separated by a single mitosis. The transcription factor GATA-1 is essential for erythroid differentiation because it induces the expression of all the known erythroid-specific genes. Here, we show that GATA-1 directly interacts with proteins that are central to the process of cell division: the retinoblastoma protein pRb and the transcription factor E2F. Specifically, E2F becomes inactivate after engaging in a GATA-1/pRb/E2F tricomplex. Another erythroid transcription factor, termed FOG-1, is able to displace pRb/E2F from this complex in vitro upon binding to GATA-1. We hypothesize that the liberated pRb/E2F can then be the target of subsequent regulation to ultimately release free E2F, which triggers cell division. The physiological role of this new pathway is evidenced by transgenic mouse experiments with GATA-1 mutants unable to bind pRb/E2F, which result in embryonic lethality by anemia.
| With more than 100 billion red blood cells generated every day, the erythroid lineage has the largest quantitative output of cell production in adult mammals. This impressive capability requires a pattern of cell proliferation closely related to that of embryonic cells followed by ultimate inhibition of cell division, when terminal erythroid differentiation is completed. Yet, the putative molecular pathways that coordinate cell proliferation and erythroid differentiation remain obscure. The transcription factor GATA-1 is essential for erythroid differentiation as it transactivates all the known erythroid-specific genes upon binding to specific DNA motifs [1],[2]. GATA-1 also exerts a repressive action on a subset of genes [3], and its overexpression inhibits cell proliferation [4],[5]. The cofactor Friend-of-GATA-1 (FOG-1) binds to GATA-1 [6] and modulates its activity on given target genes, and mice deficient in either GATA-1 [7],[8] or FOG-1 [9] die from severe anemia. Perturbation of the cell proliferation machinery also commonly results in lethal fetal anemia, as seen in mice defective in pRb [10]–[12], the three cyclins D together [13], more than one E2F members or Cdk4/6 [14],[15]. With respect to the role of pRb in erythropoiesis during development, conflicting views persist as to its cell-autonomous (intrinsic) or nonautonomous (extrinsic) nature, the latter involving the accessory contribution of macrophages [16] or even the placenta [17] as the primary cause for embryonic lethality. Yet, other studies support the existence of a cell-autonomous component for pRb in erythropoiesis, although the underlying molecular pathways remain unknown [18]–[21]. Particularly puzzling is the phenotypic paradox of mutations of the GATA-1 gene that result in the synthesis of an N-terminally truncated GATA-1 protein (GATA-1s) [22]. In a family of patients with an inherited mutation of the GATA-1 gene that results in GATA-1s expression, a severe anemia occurs [23]. In contrast, patients with the Down syndrome (trisomy 21) are prone to cellular selection of acquired somatic GATA-1 mutations that produce GATA-1s and result in preleukemic myeloproliferative disorders [24].
Here, we provide evidence of a direct physical interaction between pRb/E2F-2 and GATA-1 and document its physiological significance. We have also uncovered a potentially novel function for FOG-1 as a regulator of pRb for the control of cell proliferation. This direct interplay between GATA-1, FOG-1, pRb, and E2F sheds a new light on a constellation of mouse phenotypes and human syndromes that implicate mutations of the Rb or GATA-1 genes.
By examining the coding sequence of human GATA-1 (hGATA-1), we have identified an LNCME motif located at amino acid positions 81 to 85. This sequence exactly matches the consensus LXCXE motif present in many cellular or viral pRb-binding proteins [25],[26]. Peptidic alignment of all known GATA-1 orthologs shows the presence of the LXCXE or its variant LXSXE in all species (Figure 1A). Serine is structurally similar to cysteine but contains a hydroxyl (-OH) group in place of the thiol (-SH) group. Neither of these two motifs (LXCXE and LXSXE) is present in any of the other members of the GATA family (i.e., GATA-2 to -6; unpublished data). To probe for the functionality of this putative Rb-binding domain in hGATA-1 in vivo, NIH-3T3 cells, which do not express GATA-1, were transduced with retroviral vectors that encode either mGATA-1 or hGATA1. Cross coimmunoprecipitation (co-IP) analysis with GATA-1 and pRb (p110) specific antibodies (Abs) confirmed the hypothesis that pRb is associated with h/mGATA-1 in this reconstituted model (Figure S1A), thus suggesting that both LXCXE and LXSXE motifs are functional within the GATA-1 proteins for pRb binding. To demonstrate that pRb/hGATA-1 interaction is specifically dependent upon the LNCME motif, NIH-3T3 cells were transduced with retroviral vectors that express either (1) hGATA-1, (2) a naturally occurring N-terminally truncated form of hGATA-1, referred to as hGATA-1s (“s” for “short” [24]), which is initiated at methionine 84 within the LNCM84E motif, or (3) hGATA-1 bearing two amino acid substitutions within the LNCME motif (LNCME to LNGMK; referred to as hGATA-1Rb−) (Figure 1A). Cross co-IP with GATA-1– and pRb-specific Abs showed that only wild-type (wt) hGATA-1 interacts with pRb, whereas hGATA-1Rb− and hGATA-1s do not (Figure 1B), thus establishing that the LNCME motif in hGATA-1 is required for direct association of pRb to GATA-1. In all the aforementioned co-IP experiments, potential immunoglobulin (Ig) artifacts were ruled out by analyzing IgG isotype controls for all the Abs used and for each GATA-1 variant expressed, under the same experimental conditions (Figure S1B).
Because pRb activity is regulated by phosphorylation, we analyzed the pRb phosphorylation status within the GATA-1/pRb complex in erythroid (UT7) and nonerythroid (NIH3T3) cells. There are 16 possible CDK phosphorylation sites (Ser/Thr-Pro motifs) in pRB. By western blot analysis, Rb proteins segregate into two distinctly clustered groups of migrating bands: the more slowly migrating correspond to the so-called “hyperphosphorylated” forms, whereas the faster migrating are the “hypophosphorylated” forms, without defined species phosphorylated on specific residues among the 16 always present. Co-IP with an anti–GATA-1 Ab followed by western blot analysis with an anti-pRb Ab indicated that only the hypophosphorylated (“p”) forms of pRb were involved in the GATA-1/pRb complex, while one could detect both hypophosphorylated and (“pp”) hyperphosphorylated forms of pRb by IP of the original sample with an anti-pRb antibody (Figure 1B). Because available pRb antibodies are notoriously fastidious at differentiating hypo- from hyperphosphorylated pRb with optimal clarity, an additional experiment was performed by means of antibodies to phospho-pRb. Co-IP was first performed with specific phospho-pRb antibodies (Pser780, Pser807/811, or Pser795) before resolution by western blot analysis reacted with a GATA-1 antibody. We found that GATA-1 was not precipitated under these conditions (Figure 1C).
Because hypophosphorylated pRb is known to bind members of the E2F family and actively suppress their transcriptional activity, whereas it has been recently reported that only E2F-2—among all the E2F members present in end-stage (CD71+, TER119+) fetal liver erythroid cells—directly interacts with pRb (p110) [20], we investigated whether E2F-2 is present within the GATA-1/pRb complex. To this end, we purified CD71+ erythroid cells (>90% TER119+) from embryonic day (E)12.5 mouse fetal liver. We first established that all Abs used were efficient and specific by immunoprecipitation (IP) and western blot analysis and that signals were not Ig artifacts (IgG isotype controls), especially for GATA-1 proteins that migrate at levels similar to heavy Ig chains (50 kDa) (Figure S2B). Results of co-IP experiments demonstrated that GATA-1, pRb, and E2F-2 do exist as a tricomplex in nuclear extracts from purified E12.5 mouse fetal liver erythroid cells (Figure S2C and S2D).
To further characterize the molecular composition of the GATA-1/pRb/E2F-2 tricomplex, we performed a co-IP assay with an Ab specific for E2F-2 followed by western blot analysis reacted with Abs specific for either pRb, GATA-1, or FOG-1, the known partner of GATA-1. Whereas GATA-1 and pRb proteins could be identified within the E2F-2 immunoprecipitate, FOG-1 was not found (Figure 1D). As we previously demonstrated that GATA-1 is directly associated with pRb and that GATA-1 is not known to bind directly to E2F-2, we can thus surmise that two pRb complexes can possibly coexist in nuclear extracts from fetal liver erythroid cells: pRb/E2F-2 and GATA-1/pRb/E2F-2. To probe whether the presence of E2F-2 is required for GATA-1/pRb complex formation, we first performed an immunodepletion of E2F-2 followed by western blot analysis of the unbound fraction with GATA-1 or pRb specific Abs (Figure S2E). Results indicate that both GATA-1 and pRb could be detected in the unbound fraction (Figure 1D and S2D). The E2F-2 immunodepleted fraction was then submitted to secondary co-IPs with Abs specific for either pRb, GATA-1, or FOG-1. Secondary IP precipitates were resolved by western blot analysis with Abs specific for either E2F-2, pRb, GATA-1 or FOG-1. An association between pRb and GATA-1 could not be detected in the E2F-2 unbound fraction (Figure 1D).
From these series of experiments, we can conclude that (1) a GATA-1/pRb/E2F-2 tricomplex, in which pRb is hypophosphorylated, is present in late-stage erythroid cells from mouse fetal livers, (2) the GATA-1 LXCXE motif is required for direct association of pRb/E2F-2 to GATA-1, (3) GATA-1 and pRb do not form a bicomplex in the absence of E2F-2, (4) only E2F-2, but no other form of E2F present in these erythroid cells, allows measurable GATA-1/pRb/E2F tricomplex formation, and (5) FOG-1 is excluded from the GATA-1/pRb/E2F-2 tricomplex.
Before addressing the physiological relevance of these findings, we first probed the apparent paradox that FOG-1 is known to be a key cofactor of GATA-1 in erythroid cells, whereas we have now established that FOG-1 is excluded from the GATA-1/pRb/E2F-2 tricomplex. Although the amino acid residues of GATA-1 known to interact with FOG-1 lie within a zinc finger domain located outside the LNCME motif [27],[28], we investigated whether FOG-1 could, nevertheless, regulate the formation of the tricomplex. To this end, we generated pure populations of NIH3T3 cells that express hGATA-1 after retroviral transfer, in which we subsequently transiently transfected various amounts of a plasmid that expresses human FOG-1 (hFOG-1). We chose NIH3T3 cells because they constitutively express E2F proteins, including E2F-2, although they do not naturally express GATA-1 or FOG-1 in the absence of transfection. Co-IP assays showed that hFOG-1 was able, in a dose-dependant manner, to prevent hGATA-1 from forming a complex with pRb (Figure 2A). To assess whether a direct protein–protein interaction between GATA-1 and FOG-1 is required to dissociate pRb from GATA-1, we expressed relevant mutants of hGATA-1 and hFOG-1 in NIH3T3 cells following an identical experimental approach. We made use of the hGATA-1V205G mutant, which comprises a point mutation that substitutes a glycine for a valine at codon 205. This mutation, naturally found in patients and previously studied in mice, is associated with severe dyserythropoietic anemia and thrombocytopenia [28],[29]. It was previously established that this mutation causes the disruption of GATA-1 binding to FOG-1 and, as a consequence, a failure of expression of genes dependent on GATA-1 for their transactivation as well as the lack of repression of the c-Myc and GATA-2 genes [30]. We thus expressed hGATA-1, hGATA-1Rb−, or hGATA-1V205G in NIH-3T3 cells by retroviral transfer followed by transient transfection of a plasmid that expresses hFOG-1. Co-IP revealed that hFOG-1 was unable to dissociate hGATA-1V205G from pRb in contrast to hGATA-1 (Figure 2B). As a corroborating demonstration, we used a compensatory hFOG-1 mutant (hFOG-1S706R) that is able to bind to hGATA-1V205G and rescue the erythroid maturation of GATA-1−/− cells expressing hGATA-1V205G [30]. Co-IP assays showed that hFOG-1S706R was able to trigger a partial dissociation of hGATA-1V205G from pRb (Figure 2B). The residual association of hGATA-1V205G to pRb in the presence of hFOG-1S706R is likely to result from the lower affinity of hFOG-1S706R for hGATA-1V205G compared to that of hFOG-1 for hGATA-1 (Figure 2B and [30]).
Altogether, we thus conclude that FOG-1 is capable of displacing pRb/E2F from GATA-1 in vitro by direct GATA-1/FOG-1 interaction. An equilibrium model between GATA-1, pRb, E2F-2, and FOG-1 is proposed (Figure 2C).
As GATA-1 overexpression is known to inhibit cell growth [4],[5], we then set out to study the putative role of the GATA-1/pRb/E2F complex on cell proliferation. Because erythroid cells naturally express both GATA-1 and FOG-1, thus making it difficult to dissect the respective contribution of each of the factors and to decouple their effects on proliferation per se versus terminal erythroid cell differentiation, we first chose to establish a reconstituted cellular model after forced expression of wt and mutant GATA-1 proteins together with their corresponding FOG-1 interacting partners in the nonerythroid NIH3T3 cell line. We monitored, every day for 4 d, the growth of NIH-3T3 cells transduced with retroviral vectors that express either hGATA-1 or hGATA-1Rb−. Because hGATA-1 and hGATA-1Rb− proteins were expressed at similar levels in transduced cells (Figure 1B) with undistinguishable transcriptional activity (Figure S3), this assay becomes relevant. Whereas hGATA-1 expression impaired NIH-3T3 cell proliferation, as previously reported [5], we found that hGATA-1Rb− had no effect on it (Figures 3A and S4). Because pRb/E2F has been shown to be necessary to the G1/S transition [25],[26], we investigated whether the cellular distribution of phases of the cell cycle would be altered by the GATA-1/pRb/E2F association. NIH3T3 cells were transduced with retroviral vectors constitutively expressing either hGATA-1 or hGATA-1Rb−. Transduced NIH3T3 (eGFP-positive) cells were then starved in 1% serum (cell synchronization) for 72 h, and cell cycle progression was studied after stimulation in 10% serum at Day 2. As previously described [5], expression of GATA-1 blocked the cell cycle at the G1/S transition, whereas hGATA-1Rb− expression did not (Figure 3B).
To clarify the role of pRb in the observed GATA-1–dependent impairment of cell growth, we knocked down pRb expression in the transduced NIH-3T3 cells by transient transfection of a small interfering RNA (siRNA) directed against pRb. Consistent with our expectation, GATA-1–mediated inhibition of cell growth was abrogated for 2 d before cells ultimately stopped proliferating as the inhibitory effect of the siRNA vanished (Figure S5A). pRb knock-down by siRNA was concurrently assessed by western blot analysis, which indicated that complete inhibition of pRb expression occurred only during the first 2 d after siRNA transfection (Figure S5B) without interference with the expression of p107 (Figure S5C). When the same experiment was performed with the human osteogenic sarcoma SAOS-2 cell line, which does not express endogenous GATA-1 or pRb, neither retrovirally expressed hGATA-1 nor hGATA-1Rb− had an effect on cell proliferation (Figure S5D). These results together establish that direct interaction of pRb to GATA-1 through its LXCXE motif is central to the antiproliferative effect of GATA-1.
We then focused on the putative role of FOG-1 in this process. NIH-3T3 cells constitutively expressing hGATA-1 after retroviral transduction were transiently transfected with increasing amounts of a hFOG-1 expression plasmid and cell proliferation monitored for 5 d. hFOG-1 relieved the growth inhibitory effect of hGATA-1 for NIH-3T3 cells in a dose-dependent manner (Figure 3C). When the same experiment was performed with the non–FOG-1–interacting mutant hGATA-1V205G instead of hGATA-1, hFOG-1 was unable to counteract the inhibition of NIH-3T3 cell proliferation mediated by hGATA-1V205G at the displacing dose of FOG-1 established for hGATA-1 (Figure 3D). When hFOG-1S706R was combined with hGATA-1V205G in a similar experiment, rescue of cell proliferation was observed albeit only partially (Figure 3D), consistent with the lower affinity of hFOG-1S706R for hGATA-1V205G compared to that of hFOG-1 for hGATA-1 (Figure 2B and [30]).
To establish the relevance of these finding for erythroid cells, we made use of the erythroid cell line G1E, which is blocked at the proerythroblast stage and derives from mouse embryonic stem (ES) cells with complete biallelic inactivation of the endogenous GATA-1 genes (GATA-1−/−) [31]. G1E and NIH3T3 cells were transduced with retroviral vector constitutively expressing either hGATA1, hGATA-1Rb−, hGATA-1V205G, or hGATA-1V205GRb−. The hGATA-1V205G mutant and the double hGATA-1V205GRb− mutant were included in this study in an effort to decouple effects on cell proliferation per se versus terminal erythroid cell differentiation, which is known to be dependant upon GATA-1/FOG-1 interaction and would here confuse data interpretation since G1E cells naturally express FOG-1. In agreement with the conclusion we reached in reconstituted NIH3T3 cells, the various GATA-1 forms behaved similarly in G1E and NIH3T3 cells (i.e., inhibition of cell proliferation with hGATA1 and hGATA-1V205G, as opposed to maintenance of cell proliferation with hGATA-1Rb− and hGATA-1V205GRb−) (Figure 4A and 4B). However, the rate of cell proliferation was lower in the G1E cell population in the presence of the hGATA-1Rb− mutant as compared to G1E cells expressing the double hGATA-1V205GRb− mutant and NIH3T3 cells expressing either of these single or double mutants, in agreement with findings discussed in the subsequent section that the hGATA-1Rb− mutant induces partial terminal erythroid differentiation of the G1E cell population through GATA-1/FOG-1 complex formation. Taken together, these findings establish a novel function for FOG-1 and point to the role of the GATA-1/FOG-1 association as a potential molecular rheostat to regulate the inhibition of cell proliferation induced by GATA-1. However, G1E cells transduced with a retroviral vector that expresses GATA-1s hyperproliferate (Figure 4C). This suggests that hGATA-1Rb−, and hGATA-1s are not functionally equivalent.
Since pRb has been implicated in the differentiation processes of several cell types [16],[26], in addition to its role in cell cycle control, we then investigated whether the interaction between pRb and GATA-1 is involved in the differentiation of erythroid cells. To this end, we returned to the erythroid cell line G1E. Retrovirus-mediated expression of hGATA-1 in G1E cells induced terminal erythroid differentiation of cells positive for coexpression of vector-encoded eGFP, as evidenced by May-Grünwald-Giemsa and benzidine staining for identification and scoring of the various red blood cell precursors (Figure 5A and 5B). In contrast, forced expression of hGATA-1Rb− dramatically altered the distribution of erythroid precursors towards the more immature elements, although the initiation of erythroid differentiation was not impaired (Figure 5A and 5B). The same results were obtained with another erythroid cell line, referred to as GAK14 (M. Yamamoto, unpublished data), which was independently developed from mouse GATA-1.05 ES cells (Figure S6A). To determine more precisely the cell composition of G1E-transduced (eGFP positive) cells, erythropoietic maturation was assessed 6 d after transduction by cytofluorometry for c-Kit, CD71, and TER119 expression. Although the complete loss of c-Kit expression in both hGATA-1– and hGATA-1Rb−–transduced G1E cells indicated that erythroid differentiation was initiated normally in either case, quantification of CD71−TER119hi (41% vs. 0%) and CD71hiTER119− (2% vs. 72%), respectively, showed that terminal erythroid differentiation was impaired in hGATA-1Rb− G1E cells (Figure 5C). GATA-1Rb− G1E cells were able to progress to an intermediate (CD71hiTER119lo) stage like hGATA-1 G1E cells (25% vs. 16%, respectively, Figure 5C), but hGATA-1 G1E cells differentiated further. A similar conclusion was reached by morphological identification and scoring of the various red blood cell precursors of the retrovirally transduced GAK14 cell line (Figure S6B).
To study the potential effects of the interplay between GATA-1, pRb/E2F, and FOG-1 upon the differentiation of erythroid cells, we analyzed by flow cytometry the phenotypes of G1E cells expressing hGATA-1, hGATA-1V205G, or hGATA-1V205GRb− after retroviral vector transduction (Figure 5C). Whereas hGATA-1–expressing cells became c-Kit−, CD71+, and TER119+, both hGATA-1V205G and hGATA-1V205GRb− expressing cells failed to express TER119 and continued to express c-Kit. These data indicate that it is the GATA-1/FOG-1 association, but not the GATA-1/pRb/E2F complex, which is required for the down-expression of c-Kit during the early stage of proerythroblast differentiation.
To assess the putative physiological effects of the protein–protein interaction between GATA-1 and pRb in a whole animal, we performed a transgenic complementation rescue assay [32]. Because the GATA-1 gene is located on the X chromosome, which is randomly inactivated in every cell, neither homozygous GATA-1−/GATA-1− females nor hemizygous GATA-1−/Y males are viable. We thus made use of mice that carry a mutant GATA-1 allele, referred to as GATA-1.05, which expresses only 5% of GATA-1 wt levels. As previously shown, male GATA-1.05/Y mice die in utero from impaired hematopoiesis, whereas heterozygous female mice (GATA-1.05/X) spontaneously recover shortly after birth from embryonic/fetal and neonatal anemia [32]. In preparation for crossing experiments with these female mice, we generated transgenic (Tg) lines of mice that express either hGATA-1Rb− or its wt counterpart hGATA-1 under the control of transcriptional regulatory sequences referred to as GATA-1 hematopoietic regulatory domain (HRD). HRD-driven transcription is known to recapitulate the endogenous GATA-1 gene expression pattern in both primitive and definitive erythroid lineages of Tg mice [1]. We focused on two of the hGATA-1Rb−–expressing lines (Tg Lines 2 and 5), because expression of the hGATA-1Rb− transgene was at a level equivalent to that of the endogenous mouse GATA-1 gene (Figure 6A). Mouse Tg line 8 was also analyzed as an example of supraphysiologic expression of the hGATA-1Rb− transgene (level 280% that of the endogenous mouse GATA-1 gene). Males of these Tg mice were then mated with heterozygous GATA-1.05/X females, and their male progeny, referred to as 1.05/Y+hGATA-1Rb− and 1.05/Y+hGATA-1, studied. The expected and the observed pup numbers of XGATA-1.05/Y mice harboring the transgene are indicated in Figure 6B. The first observation we made is that the control hGATA-1 transgene could fully rescue GATA-1.05/Y male mice from lethality. Although some of the 1.05/Y+hGATA-1 male embryos showed a slightly anemic appearance at E13.5, most were indistinguishable from wt littermate embryos (Figure 6C). Importantly, their erythropoiesis caught up during late gestation, and viable pups were born with the expected Mendelian distribution (15 rescued out of 133 pups). In contrast, 1.05/Y+hGATA-1Rb− males from either Tg Line 5 (0 rescued out of 45 pups) or Tg Line 2 (0 rescued out of 47 pups) suffered from profound anemia and small size at E12.5–E13.5 leading to embryonic lethality for all animals around E15.5 of development (Figure 6C). In contrast and as expected from previous studies of other GATA mutants, mice from Tg Line 8, which expressed supraphysiologic levels of hGATA-1Rb−, were indistinguishable from wt littermate embryos. This observation underscores the importance of studying the phenotypic effects of mutant GATA transgenes within a physiologic range of expression for this type of transgenic mouse approach (see Discussion).
Flow cytometry analysis for c-Kit and TER119 expression of liver cells from 1.05/Y+hGATA-1Rb− male E13.5 embryos generated from either Lines 2 or 5 showed a decrease in the proportion of mature erythroid cells and late erythroid precursors with concurrent increase in that of early precursors, comparatively to wt littermate embryos (24% vs. 81%− for TER119high c-Kitlow and 30% vs. 8% for TER119− c-Kithigh, respectively), indicating a severe delay in erythroid maturation (Figure 7A). No significant delay was observed in 1.05/Y+hGATA-1 embryos at E13.5 (unpublished data). The liver of E15.5 embryos also contained abnormal erythroid precursors with lobulated nuclei (Figure 7B). Staining of fetal liver sections for the proliferating cell nuclear antigen (PCNA) revealed an increased number of mitotic cells (Figure 7C). Red blood cells harvested from the intracardiac space of E15.5 GATA-1Rb− embryos showed defective erythroid terminal differentiation illustrated by a significant proportion of nucleated cells with defects in nuclear condensation (Figure 7D) and a mild increase in the number of apoptotic cells (Figure 7E). Furthermore, analysis of CD71-positive cells showed a mild accumulation of cells in the S phase of the cell cycle (Figure 7F). These features share striking similarities to those observed in the fetal liver cells of pRb−/− mice [33] and with our data obtained with transduced G1E and GAK14 cell lines (Figures 5 and S6, respectively).
We have shown here that (1) direct protein–protein interaction between GATA-1 and pRb is required for normal terminal erythroid differentiation in vitro and in vivo, (2) in late-stage erythroid cells from mouse fetal liver, GATA-1 associates, through a LXCXE motif, with hypophosphorylated pRb engaged with E2F-2 to form a GATA-1/pRb/E2F-2 tri-complex, (3) GATA-1 and pRb do not form a bicomplex in the absence of E2F-2, (4) the GATA-1/pRb/E2F-2 tricomplex inhibits the proliferation of erythroid precursors (5) the association of GATA-1 to pRb/E2F alters G1-to-S phase progression and (6) FOG-1 is capable of displacing pRb/E2F-2 in vitro from the GATA-1/pRb/E2F-2 tricomplex. The physiological relevance of this latter observation needs to be further investigated.
A mechanism by which GATA-1/pRb/E2F-2 tricomplex induces cell cycle arrest during erythroid differentiation is likely to be, at least in part, by sequestering E2F within a complex in which pRb is in its hypophosphorylated form. FOG-1 is able to displace GATA-1 away from the complex, and this may then allow E2F-2 to be liberated upon subsequent phosphorylation of pRb. In parallel, the GATA-1/FOG-1 complex exerts its transcriptional effects as part of the program of erythroid differentiation, as exemplified by the extinction of c-Kit expression. Fine-tuning by exogenous inducers is likely to operate, as pRb and GATA-1 are regulated by phosphorylation [34]. This study does not yet address the possibility that the GATA-1/pRb/E2F-2 tricomplex possesses specific transcriptional activity in addition to the mere sequestration of E2F-2. For instance, GATA-1/pRb/E2F-2 may activate or inhibit E2F target genes transcription by (1) displacing other pRb-associated LXCXE proteins, (2) increasing gene transactivation through the GATA-1 transactivator domain, (3) acting as a protein platform for other associated GATA-1 proteins (e.g., acetyltransferase P300/CBP, PU.1, EKLF). Preliminary chromatin immunoprecipitation (ChIP) assays, using a GATA-1–specific Ab for ChIP, indicate that known E2F targets genes such as Cdc6 [35] can be detected (unpublished data). Transcriptional studies and cDNA microarrays analysis would sort through these hypotheses, as previously conducted for other Rb-associated proteins that include Id2 [36].
For the past several years, conflicting reports have sparked an active debate as to whether the fatal anemia observed in Rb−/− mice was of extrinsic (cell-nonautonomous) [16],[17] and/or intrinsic (cell-autonomous) [18],[19] origin. During the revision of this manuscript, two studies have added to our understanding of the role of pRb during erythroid differentiation. The Walkley and Orkin laboratories have gathered convincing evidence for a cell-type–intrinsic and cell-autonomous role of pRb in erythropoiesis, using conditional Rb inactivation restricted to the erythroid lineage [21]. Another study from the Macleod laboratory has shed light on the interplay between pRb and E2F-2 during erythropoiesis by showing that the concurrent loss of E2F-2 and pRb is surprisingly capable of rescuing terminal erythroid maturation to Rb null red cells [20]. With the double mutants, erythroid precursors resume normal cycle cell arrest in S phase and concurrent terminal erythroid maturation, although the compensatory mechanisms remain unknown [20]. Our study brings a molecular basis for the existence of an intrinsic component by establishing a direct physical link between pRb/E2F-2 and the master transcription factor of erythroid differentiation GATA-1.
Our findings also provide an explanation for the lack of full rescue of GATA-1–deficient mice by the related factors GATA-2 and GATA-3 when expressed in lieu of GATA-1, since GATA-2 and GATA-3, which share extensive functional similarities with GATA-1 [37], do not contain an LXCXE motif. The variable degree of severity of the observed phenotypes between various studies is likely to reflect the levels of expression obtained for the GATA-2 and GATA-3 factors according to the transgenic approach used, ranging from embryonic lethality by lack of erythropoiesis to viability with anemia at the adult stage [32],[37]. A similar observation has been reported with other GATA-1 mutants overexpressed in the same transgenic mouse model [38], indicating the importance of comparing GATA factors within the same range of transgene expression levels in these studies [39]. So is our observation with the supraphysiologic expression level of the hGATA-1Rb− Tg Line 8, and it is likely that pRb-independent pathways are also affected in the case of GATA overexpression. For instance, supraphysiologic GATA-1 expression levels may disturb the optimal GATA-1 to Gfi-1B cellular ratio, thus deregulating the antiapoptotic Bcl-x(L) factor [40], or alter microRNA (miRNA) levels during erythroid maturation [41].
In humans, a severe anemia has been recently described in a family of patients with an inherited splicing mutation of the GATA-1 gene that results in exon skipping and expression of an N-terminally deleted GATA-1 protein (GATA-1s) [23]. GATA-1s is translated from a downstream ATG initiation codon that encodes Met84 in the full-length GATA-1. Here, we show that this mutation disrupts the association of GATA-1s to pRb because Met84 is located within the LNCM84E motif, thus deleting it from the GATA-1s form. However, our data also suggest that GATA-1s and hGATA-1pRb− are not functionally equivalent and, thus, that it is likely that the N-terminal moiety of GATA-1 has an additional Rb-independent role. hGATA-1s may thus result in a phenotype less severe than hGATA-1pRb− possibly because a double loss of function may elicit a syndrome of lesser severity than a single loss, as observed with the double-null mice E2F-2−/− pRb−/− [20] or Id2−/− pRb−/− [16]. A caveat to the interpretation of these experiments is that the mutation in the LXCXE motif may concurrently disrupt some other critical yet unknown function of GATA-1. These observations may also help understand the specific pre-leukemic syndrome associated with acquired somatic GATA-1 mutations, which also result in GATA-1s expression, in patients with the Down syndrome (trisomy 21) [24].
The interplay uncovered here between factors that regulate cell cycle and transcription factors key to the differentiation of the red blood cell lineage should be considered for the interpretation of pRb, GATA-1, or FOG-1 mutant mice and corresponding mutations associated with specific human syndromes. These findings may also form the basis for the discovery of similar pathways in other tissues during both ontogenesis and homeostatic cell production throughout life.
Human GATA-1Rb− cDNA was generated by PCR and sequenced. This cDNA was then subcloned in the retroviral vector Migr. Human FOG-1 cDNA was subcloned in pRSVcDNA3 (Invitrogen). GATA-1V205G, GATA-1V205GRb−, and FOG-1S706R mutants were obtained using the QuickChange Site-Directed Mutagenesis Kit (Stratagene) with respective, previously described, plasmids as templates. Construct sequences were confirmed by DNA sequencing. Retroviral production and cell transduction were performed as previously described [34]. Two days after transduction, GFP-positive cells were sorted by flow cytometry (Epics Altra; Beckman Coulter).
NIH-3T3 were maintained at low population doubling and density. NIH-3T3 and SAOS-2 cells were transfected using lipofectamine 2000 transfection reagent (Invitrogen). Uptiblue reagent (Interchim) was used for cell proliferation assays with fluorimetric excitation at 560 nm and reading at 590 nm (Typhoon 9400; Amersham Bioscience). G1E cells were grown as previously described [34]. Cytospin samples were stained with May-Grünwald-Giemsa to assess and score the various stages of erythroid differentiation. Hemoglobinization was evidenced by benzidine staining. Erythroid cells from mouse fetal livers were obtained from E12.5 C57BL/C embryos.
For nuclear extract preparation, cells were washed once with PBS and incubated for 10 min at 4°C in buffer A (10 mM HEPES [pH 7.6], 3 mM MgCl2, 10 mM KCl, 5% glycerol, 0.5% NP-40) containing phosphotyrosine phosphatase inhibitors (1 mM Na2VO4), phosphoserine/threonine phosphatase inhibitors (20 mM NaF, 1 mM sodium pyrophosphate, 25 mM β-glycerophosphate), and proteinase inhibitors. After centrifugation, nuclear pellets were resuspended in buffer A containing 300 mM KCl. and incubated for 30 min at 4°C. After centrifugation, nuclear extracts were quantified by BCA staining (Pierce), half-diluted with buffer A and subjected to IP using the appropriate Ab: hGATA-1 C-terminal region epitope C-20 (Santa Cruz Biotechnology), mGATA-1 C-terminal region epitope, M-20 (sc-1234; Santa Cruz Biotechnology), pRb (N° 554136; BD Pharmingen), E2F-2 (sc-633; Santa Cruz Biotechnology), FOG-1 (sc-9361; Santa Cruz Biotechnology), Nph (sc-6013R; Santa Cruz Biotechnology), or with nonimmune corresponding (control) Abs, respectively, normal goat IgG (sc-2028), normal mouse IgG (sc-2025), and normal rabbit IgG (sc-2027) (Santa Cruz Biotechnology). Beads were then washed five times with buffer A and resuspended in 1× Laemmli buffer. Western blot analyses were performed as described [34] with primary Abs against GATA-1 (N1, Cat. N° sc-266; Santa Cruz Biotechnology), FOG-1 (M-20, Cat. N° sc-9361; Santa Cruz Biotechnology), pRb (Cat. N° 554136; BD Pharmingen; and human c-terminal pRb epitope C15, Cat. N° sc-50; Santa Cruz Biotechnology) or phospho-specific pRbs (pRbPser780, pRbPser870/811, and pRbPser795 PhosphoPlus Rb antibody Kit Cat. N° 9300; Cell Signaling Technology).
Wt and mutant human GATA-1 cDNAs were cloned 3′ to the mouse GATA-1 HRD [32]. DNA fragments were purified from vector sequences and transgenic mice generated by DNA microinjection into fertilized BDF1 eggs using standard procedures. The GATA-1.05 allele was monitored by PCR using primers corresponding to the neomycin-resistance gene in the original GATA-1.05 targeting vector.
Whole embryos were fixed in 4% formaldehyde solution at 4°C for 16 h followed by embedding in paraffin. PCNA staining was performed using a kit from Zymed Laboratories.
Phycoerythrin-conjugated anti-mouse TER119 (TER119-PE; BD Pharmingen, cat N°553673), APC-conjugated anti-mouse CD117 (c-Kit) (BD Pharmingen, cat N°553356), biotin-conjugated anti-mouse CD71 (BD Pharmingen, cat N°557416) and streptavidin-PE-PC5 secondary antibodies (BD Pharmingen) were used for surface labeling of cells. Flow cytometry was performed using FACSCalibur, and data were analyzed with the Cell Quest Pro software.
Cells were pelleted, fixed in 70% ethanol, and then resuspended at a concentration of 106 cells per milliliter in PBS containing 5 mM EDTA and 5 µg/ml Hoechst 33342 (Molecular Probes–Invitrogen). The cells were then analyzed with a LSRII cytometer (BD Biosciences) equipped with both the DIVA and the Flowjo 8.8.3 software using the Dean-Jett-Fox algorithm.
|
10.1371/journal.ppat.1002260 | Host Phylogeny Determines Viral Persistence and Replication in Novel Hosts | Pathogens switching to new hosts can result in the emergence of new infectious diseases, and determining which species are likely to be sources of such host shifts is essential to understanding disease threats to both humans and wildlife. However, the factors that determine whether a pathogen can infect a novel host are poorly understood. We have examined the ability of three host-specific RNA-viruses (Drosophila sigma viruses from the family Rhabdoviridae) to persist and replicate in 51 different species of Drosophilidae. Using a novel analytical approach we found that the host phylogeny could explain most of the variation in viral replication and persistence between different host species. This effect is partly driven by viruses reaching a higher titre in those novel hosts most closely related to the original host. However, there is also a strong effect of host phylogeny that is independent of the distance from the original host, with viral titres being similar in groups of related hosts. Most of this effect could be explained by variation in general susceptibility to all three sigma viruses, as there is a strong phylogenetic correlation in the titres of the three viruses. These results suggest that the source of new emerging diseases may often be predictable from the host phylogeny, but that the effect may be more complex than simply causing most host shifts to occur between closely related hosts.
| Emerging infectious diseases such as SARS, HIV and swine-origin influenza have all been recently acquired by humans from other species. Understanding the reasons why parasites jump between different host species is essential to allow us to predict future threats and understand the causes of disease emergence. Here we ask how host-relatedness might determine when host-shifts can occur in the most important group of emerging diseases—RNA viruses. We show that the relationship between host species is the primary factor in determining a virus's ability to persist and replicate in a novel host following exposure. This can be broken down into two components. Firstly, species closely related to the virus's natural host are more susceptible than distantly related species. Secondly, independent of the distance effect, groups of closely related host species have similar levels of susceptibility. This has important implications for our understanding of disease-emergence, and until now the only large-scale studies of viruses have been correlative rather than experimental. We also found groups of related species that are susceptible to these viruses but are distantly related to the natural hosts, which may explain why viruses sometimes jump between distantly related species.
| A major source of emerging infectious diseases are host shifts, where the parasite originates from a different host species. In humans, HIV [1], influenza [2] and Plasmodium [3] have all been recently acquired from other species. Host shifts can also have devastating effects on wildlife; for example Ebola epidemics have resulted in marked declines in some primate populations [4] and canine distemper virus has jumped from dogs into Serengeti lions and caused considerable mortality [5]. As we have come to realise that the sources of human, domestic animal or crop pathogens are likely to be from wild species [6], [7], understanding what causes these parasite host shifts to occur has become increasingly important.
For a host shift to occur, the new host must first be exposed to the parasite, the parasite must then be able to replicate in the new host, and finally there must be sufficient onward transmission in the new host for the infection to spread in the population [6]. Exposure is clearly important in determining whether a host shift occurs, and some cases of disease emergence have followed changes in the geographic range of species that have brought parasites in contact with new hosts [5], [8], [9], [10]. However, once exposure has occurred, the factors that determine whether the pathogen can replicate in a new host are poorly understood.
One factor that can potentially affect whether a parasite can replicate in a new host species is host relatedness — parasites may be more likely to replicate in species closely related to the original host [11], [12], because closely related hosts will tend to present a more similar environment to the parasite. Parasites must evade an elaborate array of host defences and rely on the host for their physiological needs, and this will result in specialised adaptations [13], [14]. These adaptations have in turn resulted in some extremely specialised parasites that are only able to survive in a narrow range of similar host species [15]. If this is the case, host shifts may occur most frequently between closely related species.
Here we use a new analytical approach to analyse host shifts, which allows us to separate two different ways in which the host phylogeny might affect the ability of a parasite to infect a new host species. The first of these, which we term the ‘distance effect’, reflects the fact that the chances of successful infection may be higher in species that are more closely related to the natural host. However, it is also likely that related species share similar levels of susceptibility independently of how related they are to the natural host, a process that we term the ‘phylogenetic effect’. These are statistically and biologically distinct phenomena. The distance effect will result in the expected susceptibility of new hosts declining as they become less related to the natural host. In contrast, the phylogenetic effect will have no effect on the expected susceptibility with distance from the natural host. However, it will result in distantly related species often having very different levels of susceptibility from the natural host, as it results in the variance in susceptibility increasing among more distantly related species.
The two effects may generate very different patterns of host switching. The distance effect would result in most host shifts infecting species closely related to the natural host. In contrast, the phylogenetic effect might mean that host clades distantly related to the natural host are susceptible to a parasite, and this could cause parasites to jump between distantly related species.
Previous research has examined the distance effect only. While there is evidence that parasites most often shift between related hosts from correlative studies of parasite-incidence in wild animals (e.g. [16]), experimental evidence has been surprisingly rare. Cross-infection experiments using plants and fungi [17], [18], Drosophila and nematode worms [19], and beetles and Spiroplasma bacteria [20] have all found that the ability of a parasite to establish an infection declines as a novel host's relatedness to the natural host declines.
The extent to which host relatedness influences host switching varies between different groups of parasites, and it has been suggested that RNA viruses may be particularly prone to jump between distantly related hosts [21]. Reviewing emerging viral diseases in vertebrates, Parrish et al [21] observed that “Spillover or epidemic infections have occurred between hosts that are closely or distantly related, and no rule appears to predict the susceptibility of a new host.” Viruses are more likely than other groups of parasites to be shared between distantly related primates [16], and many human diseases that have been recently acquired from other species are RNA viruses [22]. The ability of certain viruses to infect distantly related hosts may result from the use of conserved host receptors to enter cells [23], [24], or the existence of hosts that do not posses broad resistance mechanisms to that type of parasite [25], [26]. However, some studies have found evidence for the importance of the host phylogeny; rabies virus strains have higher rates of cross species transmission between closely related host species in the wild [27] and primate lentivirus phylogenies show signs of preferential switching between closely related hosts [28].
To explore this question we have conducted a large cross-infection experiment in which three sigma viruses were injected into 51 different species of Drosophilidae. Sigma viruses are a clade of rhabdoviruses (RNA viruses with single-stranded negative-sense genomes), which infect various species of Diptera [29], [30]. They are normally vertically transmitted [31], [32], leading to extreme specialisation on just a single host species. However, the sigma virus of Drosophila melanogaster (DMelSV) will replicate in a range of different dipteran hosts [33], and differences between the host and virus phylogenies show that sigma viruses have switched between distantly related host lineages during their evolution [30]. Here we find that the host phylogeny explains most of the variation in the ability of sigma viruses to replicate in novel hosts, with both the distance and phylogenetic effects being large. These results not only allow us to explore the different ways in which the host phylogeny may affect host switching, but they are also, to our knowledge, the first study to experimentally test the effect of host genetic distance on infection success in RNA viruses — the most important source of emerging diseases.
We measured the ability of three Drosophila sigma viruses to persist and replicate following injection into 51 fly species sampled from across the phylogeny of the Drosophilidae (Figure 1). The three viruses were DAffSV, DMelSV and DObsSV, which naturally occur in D. affinis, D. melanogaster and D. obscura respectively [29].
We extracted DAffSV, DMelSV and DObsSV from infected stocks of D. affinis, D. melanogaster and D. obscura. To clear these stocks of any bacterial or other viral infections, they were aged for at least 20 days, before collecting embryos [32] and de-chorionating them in ∼2.5% w/v sodium hypochlorite solution for one minute [31]. The embryos were then rinsed in distilled water and placed onto clean food. To collect flies infected with a sigma virus, the adults were exposed to 100% CO2 at 12°C for 15 mins and the paralysed individuals were retained [31], [32], [34]. These were frozen at −80°C to rupture cells, homogenised in Ringer's solution [35] (2.5 µl/fly), and then briefly centrifuged twice, each time retaining the supernatant. This was passed through Millex PVDF 0.45 µM and 0.22 µM syringe filters (Millipore, Billerica, MA, USA) to remove any remaining host cells or bacteria, before being stored in aliquots at −80°C.
Stocks of each fly species were kept in half pint bottles of staggered ages, and each day freshly eclosed flies were sexed, males were removed, and females were aged at 18°C for 3 days on agar medium (recipe in Text S1) before injection. At the same time we stored remaining flies in ethanol for wing size measurements. The food medium, rearing temperature and whether each species was composed of single or multiple lines can be found in Table S1.
Female flies were injected with 69 nl of the virus extract intra-abdominally using a Nanoject II micro-injector (Drummond scientific, Bromall, PA, USA). Half the flies were frozen immediately in liquid nitrogen as a reference sample to control for relative dose size, and the rest were kept on agar medium at 18°C for 15 days to allow the virus to replicate before being frozen in liquid nitrogen. The day 15 time-point was chosen based on pilot time-course data, and we note that the change in viral titre includes a decline in the virus following injection, followed by a growth/replication phase (Figure S1). Frozen flies were then homogenised in Trizol reagent (Invitrogen Corp, San Diego, CA, USA). Based on quantitative reverse-transcription PCR (qRT-PCR), the dose of the three viruses was similar (with a maximum of a 1.6x difference between viruses).
The injections were carried out over a period of 18 days, with the aim of completing 3 biological replicates for each virus per fly species (3 replicates each of the day 0 and day 15 treatments). The virus (DAffSV, DMelSV or DObsSV) was rotated on a daily basis, whilst treatment (frozen immediately or on day 15) and the injection order of fly species were randomised each day. On average we injected and quantified viral titre in a pool of 10 flies per replicate (range of across species means = 5–15). Out of the 153 fly-virus combinations, 126 had 3 biological replicates, 24 had 2 biological replicates and 3 had 1 biological replicate.
Wolbachia endosymbionts have recently been shown to provide resistance to a range of positive sense RNA viruses [36], [37], [38], [39]. Although it does not affect the replication of DMelSV (L. Wilfert and M. Magwire, unpublished data), we nonetheless tested each species for Wolbachia using PCR primers that amplify the wsp gene [40].
We also checked that the body size of the different species did not affect our results. To do this, we measured wing length, which is commonly used as a body size measure in Drosophila and strongly correlates with thorax length [41], [42]. Wings were removed from ethanol-stored flies, photographed under a dissecting microscope and the length of the IV longitudinal vein from the tip of the proximal segment to where the distal segment joins vein V [43] was measured (relative to a standard measurement) using ImageJ software (v1.43u) [44].
Viral titres were estimated using qRT-PCR. To ensure that we only amplified viral genomic RNA and not messenger RNA, the PCR primers were designed to amplify a region spanning two different genes. The copy-number of viral genomic RNA was expressed relative to the endogenous control housekeeping gene RpL32 (Rp49). We designed different RpL32 primers specific for each species. First, we sequenced the RpL32 gene from all of the species (we were not able to amplify RpL32 from Drosophila busckii, see Table S2). We then designed species-specific primers in two conserved regions (Table S3).
Total RNA was extracted from our samples using Trizol reagent, reverse-transcribed with Promega GoScript reverse transcriptase (Promega Corp, Madison, WI, USA) and random hexamer primers, and then diluted 1∶4 with DEPC treated water. The qRT-PCR was performed on an Applied Biosystems StepOnePlus system using a Power SYBR Green PCR Master-Mix (Applied Biosystems, CA, USA) and 40 PCR cycles (95°C for 15 sec followed by 60°C for 1 min). Two qRT-PCR reactions (technical replicates) were carried out per sample with both the viral and endogenous control primers. Each qRT-PCR plate contained a standard sample, and all experimental samples were split across plates in a randomised block design. A linear model was used to correct for the effect of plate. We repeated any samples where the two technical replicates had cycle threshold (Ct) values more than 1.5 cycles apart after the plate correction.
To estimate the change in viral titre, we first calculated ΔCt as the difference between the cycle thresholds of the sigma virus qRT-PCR and the endogenous control. The viral titre of day 15 flies relative to day 0 flies was then calculated as 2−ΔΔCt, where ΔΔCt = ΔCtday0 – ΔCtday15, where ΔCtday0 and ΔCtday15 are a pair of ΔCt values from a day 0 biological replicate and a day 15 biological replicate for a particular species-virus combination. We used a dilution series to calculate the PCR efficiency of the three sets of viral primers and thirteen of the RpL32 primer combinations (covering 40 of the 51 Drosophila species). The efficiencies of the three virus primers were 95%, 97%, and 100%, (DAffSV, DMelSV and DObsSV) and the average efficiency of RpL32 primers across species was 106%, with all being within a range of 98–112%.
The host phylogeny was inferred using the COI, COII, 28S rDNA, Adh, SOD, Amyrel and RpL32 genes. We downloaded all the available sequences from Genbank, and attempted to sequence COI, COII, 28S rDNA, Adh and Amyrel in those species from which they were missing (details in Table S4). This resulted in sequence for all species for COI, COII and 28S and partial coverage for the other genes (50 out of 357 species-locus combinations were missing from the data matrix). The sequences of each gene were aligned using ClustalW (alignments and accession numbers are Datasets S1-S8 in supporting information). To reconstruct the phylogeny we used BEAST [45], as this allows construction of an ultrametric (time-based) tree using a relaxed molecular clock model. The genes were partitioned into 3 groups each with their own substitution and molecular clock models. The three partitions were: mitochondrial (COI, COII); ribosomal (28S); and nuclear (Adh, SOD, Amyrel, RpL32). Each of the partitions used a HKY substitution model (which allows transitions and transversions to occur at different rates) with a gamma distribution of rate variation with 4 categories and estimated base frequencies. Additionally the mitochondrial and nuclear data sets were partitioned into codon positions 1+2 and 3, with unlinked substitution rates and base frequencies across codon positions. Empirical studies suggest that HKY models with codon partitions are a good fit for most protein coding data sets [46]. A random starting tree was used, with a relaxed uncorrelated lognormal molecular clock and we used no external temporal information, so all dates are relative to the root age. The tree-shape prior was set to a speciation-extinction (birth-death) process. The BEAST analysis was run for 100 million MCMC generations sampled every 1000 steps (additionally a second run was carried out to ensure convergence). The MCMC process was examined using the program Tracer (v1.4) [47] to ensure convergence and adequate sampling. Trees were visualised using FigTree (v. 1.3) [48].
We used a phylogenetic mixed model to examine the effects of host relatedness on viral persistence and replication in a new host [49], [50], [51]. This framework allows (random) phylogenetic effects to be included in the model, with the correlation in phylogenetic effects between two host species being inversely proportional to the time since those two host species shared a common ancestor (following a Brownian model of evolution). In general, conclusions drawn from phylogenetic comparative methods that include a species term in the model seem to be robust to alternative (non-Brownian) evolutionary models [52].
We fitted the model using a Bayesian approach in the R package MCMCglmm [53, R Foundation for Statistical Computing, Vienna, Austria] and REML in ASReml [54]. The two methods gave similar results so we only report the Bayesian analysis (Figure S2). The model had the form:where yvhi is the viral titre of the ith biological replicate of host species h infected with virus v. is the intercept term for virus v, and can be interpreted as the viral replication rate in the species at the root of the phylogeny. dvh is the phylogenetic (patristic) distance between the original host of virus v and species h, and the associated regression coefficient () determines the degree to which viral replication rate of virus v changes as the phylogenetic distance increases. The random effect up:vh is the deviation from the expected viral replication rate for virus v in host h due to historical processes (i.e. the host phylogeny). The species random effect us:vh is the deviation from the expected viral replication rate of virus v in host h that is not accounted for by the host phylogeny. The residual is evhi, which included within-species genetic effects, individual and micro-environment effects and measurement/experimental error. The random effects (including the residual) are assumed to come from multivariate normal distributions with zero mean vectors (because they are deviations) and structured covariance matrices. Denoting as a vector of phylogenetic effects across species for virus v, and A as a matrix with elements ajk representing the proportion of time that species j and k have had shared ancestry since the root of the phylogeny:where is the variance of phylogenetic effects for DAffSV, and is the covariance between phylogenetic effects for DAffSV and DMelSV.
Similar distributions are assumed for species effects:where I is an identity matrix indicating that species effects are independent of each other. The posterior modes for were close to zero for viruses DAffSV and DObsSV and these were omitted from the model (except for the calculation of σ2p/(σ2p+ σ2s), see below).
The residuals are distributed as:The off-diagonal elements of Ve (i.e. the covariances) were set to zero since viruses were not replicated within biological replicates. In a Bayesian analysis prior probability distributions have to be specified for the fixed effects and the covariance matrices. As described in detail in the supporting materials (Text S1) we used several different priors to check if the results are sensitive to the choice of prior. The results presented were obtained using parameter expanded priors for the Vp and Vs matrices [53]. The P-values reported (PMCMC) correspond to 2pmin, where pmin is the smaller of the two quantities a) the proportion of iterations in which the posterior distribution is positive or b) the proportion of iterations in which the posterior distribution is negative. The 95% credible intervals (CI) were taken to be the 95% highest posterior density intervals. Marginal means of the posterior distribution are used as summaries of central tendency. Significance of the fixed effects was inferred if the 95% CI of the posterior distribution did not cross zero, and the P-values were equal to or less than 0.05.
We also checked whether several additional factors affected viral replication by repeating the analysis with these factors included in the model as fixed effects. There was no significant effect of wing size (an average of 33 measured per species, PMCMC = 0.50), the presence of the bacterial endosymbiont Wolbachia (Table S4, PMCMC = 0.51) or rearing temperature (PMCMC = 0.55). We also repeated the analysis with three outliers removed, so that the distribution of the residuals was not significantly different from normal according to an Anderson-Darling test (A = 0.61, P = 0.11). The parameter estimates were very similar to those obtained when including all the taxa (as reported in the results).
We measured the change in viral titre over 15 days for three sigma viruses each injected into 51 species of Drosophila, including their natural hosts (see Figure 1). In total we injected and quantified viral titre in 887 biological replicates (a total of 8762 flies). To investigate how the host phylogeny affects the ability of the virus to persist and replicate in the different species, we reconstructed the phylogeny of all 51 species using the sequences of seven different genes. The resulting tree broadly corresponds to previous studies [55], [56], with the close phylogenetic relationships being generally well supported and more ancient nodes were less well supported (Figure 1).
There are two ways in which the host phylogeny could affect the ability of the three viruses to infect new host species. First, the chances of successful infection may be higher in species that are more closely related to the natural host (the ‘distance effect’). Second, related species may share similar levels of susceptibility independently of how related they are to the natural host — an effect that we refer to as the ‘phylogenetic effect’. To separate these two processes we fitted a phylogenetic mixed model to our data.
All three viruses have greater viral titres in fly species that are more closely related to their natural host (Figure 2). If we assume that titres of all three viruses decline with genetic distance from their natural host at the same rate, then there is a significant negative relationship between titre and distance (slope: γ = −1.96; 95% CI = −3.66, −0.43; PMCMC = 0.022). If we instead allow the effect to differ between viruses, the negative effect of genetic distance from the natural host on replication is greatest for DObsSV (Figure 2; slope: γO = −4.03; 95% CI = −6.11, −0.94; PMCMC = 0.005), is smaller and only marginally non-significant for DAffSV (Figure 2; slope: γA = −1.82; 95% CI = −3.99, 0.37; PMCMC = 0.095), and not significant for DMelSV (Figure 2; slope: γM = −0.47; 95% CI = −3.06, 1.94; PMCMC = 0.692). These effects were still present when the natural host species were removed from the analysis (data not shown). Therefore, the rate at which viral titres decline with genetic distance of the new host from the natural host differs between the individual viruses.
There is also a strong influence of host phylogeny on viral replication that could not be explained by the distance of the novel host from the original host. The between-species variance consists of two components; σ2p, which is the variance that can be explained by the host phylogeny, and a species-specific component σ2s which cannot be explained by a Brownian-motion model of evolution on the host phylogeny. These statistics do not include the effects of the distance from the natural host, as this was included as a fixed effect in the model [57]. To assess the importance of the host phylogeny, we calculated the proportion of the between-species variance that can be explained by the phylogeny (σ2p/(σ2p+ σ2s), which is similar to Pagel's λ [58], [59] or phylogenetic heritability [50], [51]). The phylogeny explained almost all of the between-species variance in viral titre for DAffSV and DMelSV (σ2p/(σ2p+ σ2s) = 0.86, 95% CI = 0.53–1 and 0.91, 95% CI = 0.74–1, respectively), and most of the between-species variation for DObsSV (σ2p/(σ2p+ σ2s) = 0.72, 95% CI = 0.43–0.98). Therefore, most of the differences between species in viral titres can be explained either by the host phylogeny or the distance from the natural host.
Is it the distance from the natural host, or host phylogeny per se, that is most important in determining viral replication and persistence in a new host? To allow a direct comparison of these two effects, we calculated the expected amount of change in viral titre from the root to the tips of the tree that will result from the phylogenetic effect. This was done by taking the product of the standard deviation of the phylogenetic effect and , which is the mean of a folded zero-centred normal distribution, and is the predicted change under a Brownian model. This gave values of 2.15, 3.28 and 2.69 viral-titre-units for DAffSV, DMelSV and DObsSV respectively. These can be compared directly to the estimates described above of the amount of change in viral titre as the genetic distance from the natural host increases (−1.82, −0.47 and −3.70 viral-titre-units for DAffSV, DMelSV and DObsSV respectively). The time from the root to tip of the phylogeny has been estimated as ∼40 million years [60], so for every ∼40 million years travelled along the phylogeny, or from the natural host, we expect to see the above changes in viral titre. From these estimates it is clear that over this timescale the two processes are of similar importance for DAffSV and DObsSV, but that the host-phylogeny is more important than distance-from-the-original-host in determining the replication and persistence of DMelSV in a new host.
Differences between hosts in viral replication and persistence could either reflect differences in susceptibility to all three viruses (‘general susceptibility’), or the effects on the three viruses could be independent (‘specific susceptibility’). We found that most of the phylogenetic effect was caused by species differing in their level of general susceptibility, as there were strong phylogenetic correlations between viruses (Table 1). Furthermore, the correlation is not greater between the two viruses that naturally infect closely related hosts (DAffSV and DObsSV). Therefore, the phylogenetic effects mean that a given host species' susceptibility to one virus is strongly correlated to its susceptibility to another sigma virus, regardless of whether the virus originated from a closely or distantly related host.
The analysis above assumes that we have the correct phylogeny, but some of the relationships are poorly resolved (Figure 1). To check whether this affected our results, we repeated the analysis integrating over the posterior sample of trees generated during the phylogenetic analysis [61]. This was achieved by fitting the phylogenetic mixed model to 2000 different trees from the posterior sample (from 100,000 trees we used a burn-in of 30,000 trees and then used every 35th tree). This gave very similar results to our main analysis, suggesting that phylogenetic uncertainty does not affect our conclusions. We would note however, that σ2p is biased downwards whenever the tree is incorrect, and this bias is not removed by this procedure.
We found that the ability of three sigma viruses to persist and replicate in 51 different species of Drosophila is largely explained by the host phylogeny. The effect of phylogeny can be broken down into two components; not only did viral titres tend to decline with increasing genetic distance from the natural host, but there is also a tendency for related hosts to have similar titres, independent of the distance effect.
The decline in viral titres with increasing distance from the natural host suggests that the greater the change in the cellular environment, the less well adapted the virus is. This might be caused by changes in the cellular machinery used by the virus in its replication cycle, or the virus being less adept at avoiding or suppressing the immune response. Regardless of the causes of this effect, it suggests that successful host shifts may be more likely between closely related hosts [6]. A host shift requires the new host to be exposed to the pathogen, the virus to replicate sufficiently for an individual to become infected, and finally for there to be sufficient onward transmission for the infection to become established in the population. Our data suggests that the second step is most likely to occur between closely related hosts. It is possible that higher titres may also lead to greater onward transmission, as the titre of DMelSV in D. melanogaster correlates with the rate at which the virus is transmitted [31], [62]. Furthermore, it has also been reported that although DMelSV will replicate in a range of Drosophila, but it was stably transmitted only in the closely related Drosophila simulans and not the more distantly related Drosophila funebris [63]. However, viral titres should only be used with caution as a proxy for transmission rates, as many other factors may affect transmission rates, including trade-offs between replication and virulence [64].
There is tentative evidence that host shifts of sigma viruses occur most often between closely related species in natural populations. Although comparisons of Drosophila and sigma virus phylogenies show evidence of past host shifts, the host and virus phylogenies are more similar than expected by chance [30]. This may be the result of more frequent host switches between closely related species, as would be predicted by our results (although cospeciation would produce the same pattern and more data is required to confirm these findings).
This result is interesting because it has previously been questioned whether the genetic distance between host species plays an important role in predicting the source of host shifts, especially for RNA viruses [6], [22]. Indeed, some plant viruses can replicate in an enormous range of species; Cucumber mosaic virus can infect 1,300 plant species in over 100 families and Tomato spotted wilt virus can infect 800 plant species in 80 families [65]. The use of conserved receptors to enter host cells may be key to large potential host ranges in animals [23], [24], [66]. However, although a virus may be able to enter the cells of many different species, it then relies on numerous different components of the cellular machinery to replicate effectively, and this may make shifts to hosts that are distant from the natural host unlikely.
A factor that could lead to changes in host suitability across the phylogeny is selection for resistance to viruses. One reason to suspect that this may be important is that genes involved in antiviral immunity often evolve exceptionally rapidly in Drosophila [67], [68], [69], [70], and this may translate into rapid phenotypic changes in host susceptibility. If this process is driving the patterns that we see, then the observation that natural host-parasite combinations tend to be more susceptible would suggest that the viruses have been able to overcome these host defences, resulting in viruses that are well adapted to their natural hosts, rather than vice versa.
After accounting for the effect of distance from the natural host, the host phylogeny still explains most of the remaining variation in viral titre between species. This ‘phylogenetic effect’ means that that closely related host species have similar levels of resistance due to their non-independence as a result of common ancestry. Indeed, the most distantly related clade to all of the natural hosts examined (the Scaptodrosophila) have one of the highest viral titres (Figure 1). For two of the viruses (DAffSV and DObsSV), we found that this phylogenetic effect was of comparable importance to the effect of genetic distance from the natural host, and for the third virus (DMelSV) it was more important.
The phylogenetic effect and distance effects are statistically (and biologically) distinct phenomena. If we imagine two sister species (A and B) and an out-group (C) are infected with a virus originally from species A, there are two ways in which the host phylogeny could affect the ability of the viruses to infect the three species. Under a Brownian motion model of evolution we expect viral titre in species A to be more different to that in C than B. Importantly, however, we do not expect this difference to have a consistent sign, as it is only the magnitude of the difference that should be larger for species C. A second process is that as we move away from species A we may expect a systematic change in viral titre – either that the viral titre increases as we move to species B and then to species C, or alternatively a systematic decrease. We call this first effect – where the change does not have a predictable sign – a phylogenetic effect, and the second effect - where change does have a predictable sign – a distance effect.
The phylogenetic and distance effects may also generate distinct patterns of host switching (see Introduction). For example, our data regarding the phylogenetic effect imply that sigma viruses may more easily switch between infecting flies in the subgenus Sophophora and the distantly–related, but highly susceptible, Scaptodrosophila. However, the two distinct patterns may emerge from the same underlying evolutionary process. If related hosts have similar levels of susceptibility (i.e. the phylogenetic effect), and pathogens can only become established in the most susceptible hosts, then we would expect to see a decline in viral titre in species distantly related to the natural hosts (i.e. the distance effect).
The phylogenetic effect is mostly caused by variation in susceptibility to all three viruses (there is a strong phylogenetic correlation in the titres of the three viruses). Such patterns may arise if the common ancestors of different host clades have acquired or lost immune or cellular components that affect susceptibility to all sigma viruses. The frequent gain and loss of immune components is well-established, for example, Drosophila species in the obscura group have lost a type of blood cell (lamellocytes) that are found in other Drosophila, which means they are particularly susceptible to parasitoid wasps [26]. Similarly a class of antifungal peptides (drosomycins) are found only in the melanogaster group of Drosophila [71], [72] and components of antiviral RNAi pathways have lineage-specific distributions [73], [74]. Part of the phylogenetic effect could be explained by the evolutionary history of the viruses, for example if they have recently switched between host species and are still well-adapted to a previous host. The strong phylogenetic correlation between the three viruses we studied might seem surprising as these viruses are very different to one another at the sequence level (amino-acid identities are ∼20%–40% [29], [30]). However, even viruses which show no similarities at the sequence level often share elements of protein structure [75], [76], [77], and different rhabdoviruses are known to have similar modes of action (for example, infecting nervous tissue [31], [78]).
The strong phylogenetic effect that we found also has practical implications for comparative studies of resistance in different species. It means that observations on related species will not be independent, so it is essential to account for these effects in the analysis of comparative data [79]. For example, the decline in the resistance of novel hosts with genetic distance from the natural hosts that has been observed in some previous studies may be attributable to a phylogenetic effect, rather than distance itself.
In conclusion, our results show that the host phylogeny is an important determinant of viral persistence and replication in novel hosts, and therefore may also be an important influence on the source of new emerging diseases. The effect is more subtle than simply leading to a decline in infection success with genetic distance from the original host, because the strong phylogenetic effect may sometimes result in susceptible hosts being grouped in phylogenetically distant clades, allowing parasites to jump great phylogenetic distances. The importance of these phylogenetic effects on replication and persistence relative to factors affecting exposure and onward transmission requires further study if we are to understand how they affect a parasites ability to host shift in nature.
|
10.1371/journal.pntd.0004063 | Ergot Alkaloids (Re)generate New Leads as Antiparasitics | Praziquantel (PZQ) is a key therapy for treatment of parasitic flatworm infections of humans and livestock, but the mechanism of action of this drug is unresolved. Resolving PZQ-engaged targets and effectors is important for identifying new druggable pathways that may yield novel antiparasitic agents. Here we use functional, genetic and pharmacological approaches to reveal that serotonergic signals antagonize PZQ action in vivo. Exogenous 5-hydroxytryptamine (5-HT) rescued PZQ-evoked polarity and mobility defects in free-living planarian flatworms. In contrast, knockdown of a prevalently expressed planarian 5-HT receptor potentiated or phenocopied PZQ action in different functional assays. Subsequent screening of serotonergic ligands revealed that several ergot alkaloids possessed broad efficacy at modulating regenerative outcomes and the mobility of both free living and parasitic flatworms. Ergot alkaloids that phenocopied PZQ in regenerative assays to cause bipolar regeneration exhibited structural modifications consistent with serotonergic blockade. These data suggest that serotonergic activation blocks PZQ action in vivo, while serotonergic antagonists phenocopy PZQ action. Importantly these studies identify the ergot alkaloid scaffold as a promising structural framework for designing potent agents targeting parasitic bioaminergic G protein coupled receptors.
| The parasitic infection schistosomiasis afflicts millions of people worldwide and is clinically treated using a single drug, praziquantel (PZQ). Despite the fact that PZQ has served as a stalwart anthelmintic for decades, the molecular basis of action of this clinical agent is poorly understood. This lack of mechanistic information impedes the rational design of alternative therapies and highlights the need for new approaches for studying the target(s) and effectors engaged by PZQ in vivo. Here, we exploit the predictive phenology between free-living planarian regenerative screens and parasitic neuromuscular physiology to reveal a broad efficacy of ergot alkaloids in phenocopying the action of PZQ. In planarian regenerative screens, data highlight structural features of the ergoline scaffold that yield specific regenerative effects to promote or inhibit head regeneration. Ergot alkaloids with efficacy in regenerative assays were also found to modulate the contractility of schistosomules. Overall, these data highlight a possible therapeutic potential of ergot alkaloids as antischistosomals and the action of PZQ as an ergomimetic.
| Schistosomiasis is a neglected tropical disease that infects over 200 million people worldwide, burdening economies with an annual loss of several million disability-adjusted life years [1–3]. The disease is caused by parasitic flatworms of the genus Schistosoma and treatment is largely reliant on a single drug—praziquantel (PZQ), used clinically for over 30 years [4–6]. PZQ is a synthetic tetracyclic tetrahydroisoquinoline that was initially developed by Merck while screening for compounds with tranquilizer properties, and arose from a compound that lacked sedative properties but was remarkably effective against parasitic flatworms [7,8]. PZQ has shown remarkable durability compared with other anthelmintics, but incidences of decreased PZQ efficacy have been reported in both the laboratory [9–11] and the field [12,13], raising concerns that PZQ-resistant strains of schistosomiasis may emerge especially as eradication initiatives increase distribution of this drug [4]. Development of alternative therapies to PZQ has been hampered by the fact that the mechanism of action of PZQ remains unresolved and rationally designed derivatives of PZQ typically prove less efficacious [7,14,15]. These longstanding roadblocks impair the iteration of next generation antischistosomals needed to counter the likely emergence of PZQ-resistant isolates [16]. Resolution of the pathways engaged by PZQ in vivo is therefore a key priority.
A fresh perspective toward this problem comes from the discovery of an unusual axis-duplicating effect of PZQ during regeneration of free-living planarian flatworms [17,18]. The striking phenotype of PZQ-evoked bipolarity (Fig 1A), coupled with the genetic tractability of this system for RNAi [19] and a retained predictive value against parasitic flatworms [20] establishes a novel platform for identifying relevant in vivo effectors of PZQ action based on the molecular phenology between these systems. In planarian regenerative screens, the bipolarizing efficacy of PZQ depends on a coupling of voltage-operated Ca2+ channels to bioaminergic signals [20], which likely regulate polarity signaling from flatworm muscle cells to coordinate regenerative outcomes [21].
Here, we apply genetic and pharmacological approaches to dissect our observation that activation of serotonergic signaling in the planarian Dugesia japonica blocks the bipolarizing ability of PZQ. This effect is shown to be unique to serotonin (5-HT), and highlights the importance of characterizing serotonergic receptors to identify 5-HT blockers that could potentiate, or phenocopy, PZQ action. Intriguingly, serotonergic screens highlight ergot alkaloids as a class of compounds that potently and penetrantly miscued planarian regeneration and schistosomule muscle function, with structure activity insight from active compounds highlighting modifications of the ergot scaffold predictive for flatworm efficacy. Based on these data, we contend that the ergot alkaloid scaffold merits further exploration to yield novel chemotherapeutics with selective efficacy against parasite musculature.
Exposure of regenerating planarian (D. japonica) trunk fragments to praziquantel (PZQ) yielded two-headed worms (Fig 1A, [17]), an effect never observed in the absence of drug exposure. This effect was dose-dependent (EC50 38±3.6μM, Fig 1B), with maximal doses being completely penetrant [17]. Strikingly, the bipolarizing action of PZQ was blocked by co-incubation with 5-HT, or the analogue O-methylserotonin (O-MT), but not by co-incubation with other bioaminergic neurotransmitters (Fig 1C). As this result suggests serotonergic signals functionally oppose PZQ action, we implemented genetic (Fig 2) and pharmacological strategies (Fig 3) to interrogate 5-HT signaling pathways in planarians.
To enable interrogation of 5-HT receptor function by in vivo RNAi, we generated a de novo transcriptome assembly for D. japonica (see Methods) to allow a comprehensive bioinformatic identification of 5-HT receptors in this system (Fig 2A). A total of 17 predicted serotonergic G protein coupled receptor (GpCR) sequences were identified based upon homology to previously identified sequences [22–24]. These putative 5-HT receptors clustered into three discrete clades (S1-, S4- and S7-like, Fig 2A) defined by homology with C. elegans serotonin receptors (SER1, SER4 & SER7) [24]. Previously identified planarian 5-HT receptors (5HTLpla1-4, DjSER-7, DtSer1 [22,23,25]) all localized within the Ser-7 clade.
To simplify nomenclature for these sequences, we assigned names to each receptor based on these three clades and transcript abundance within each grouping (from FPKM values, fragments per kilobase of transcript per million mapped reads), such that the most abundant transcript in the S7 clade was named S7.1 and the least abundant of the eight transcripts was designated as S7.8. Comparison of FPKM values for all these sequences revealed that S7.1 was the most abundantly expressed 5-HT receptor in this system (Fig 2B), accounting for ~40% of the total FPKM values assigned to all predicted serotonergic receptors. As the most abundant receptor, S7.1 had previously been cloned by degenerate PCR (5HTLpla4, DtSer1 [23,25]) and recently demonstrated to couple to cAMP generation [25]. Expression levels of S7.1 mRNA changed during regeneration [23], and we observed increased FPKM values for S7.1 at early regenerative timepoints (Fig A in S1 Text). Although prediction of the S1, S4 and S7-like sequences as serotonergic GpCRs is based on specific sequence features known to be important for 5-HT binding (see below), as well as overall homology to other serotonin receptors (Fig B in S1 Text), we do note that both the planarian receptor (S7.4, DjSER-7 [22]) and a schistosome S7-like receptor have been successfully deorphanized following heterologous expression and shown to respond to 5-HT [25,26].
Alignment of the planarian sequences with human bioaminergic GpCR sequences (Table 1) revealed conservation of key residues within the orthosteric binding pocket known to be important for ligand binding. With reference to molecular docking studies of 5-HT into crystal structures of human 5-HT1B (and 5-HT2B) receptors [27], these include (i) a salt bridge between the amino group of 5-HT and D3.32 in the 5-HT1B receptor (itself stabilized by Y7.43, Ballesteros & Weinstein numbering [28]), (ii) a hydrogen bond from T3.37 to the indole (N-H) hydrogen of 5-HT, and (iii) a hydrophobic cleft formed by contributions from (W6.48, F6.51, F6.52, C3.36 and I3.33). All these residues are well conserved in the planarian receptor sequences (Table 1). Notably, bioaminergic receptors that respond to different ligands (e.g. dopamine, adrenaline, histamine) present a more polar interface at resides 5.42 and 5.46, whereas human 5-HT sequences present no more than one polar residue ([27], Table 1). This feature has been suggested to facilitate interaction with the less polar indole group of 5-HT compared to the other bioaminergic transmitters [27]. The planarian sequences also conform to this principle with the combination of residues at this position being diagnostic of the three different clades of 5-HT receptor sequences (e.g. ‘AA’ for S7, ‘S/A, A/S for S4, ‘xT’ for S1, Table 1). Another notable feature of the planarian 5-HT groupings is receptor architecture, for example the spacing between these critical residues in helix 3 and helix 5, and helix 5 and helix 6 (Table 1) appears diagnostic of the different serotonergic clades. For example, the S7 clade exhibits a consistent spacing (~74 residues) between 3.37 and 5.42, and a shorter third intracellular loop between TM5 and TM6 compared with the other clades.
As S7 represented the most abundantly expressed clade of 5-HT GpCRs, we proceeded to perform RNAi against each individual receptor. First, we screened for effects on PZQ-evoked bipolarity. RNAi-mediated suppression of S7.1 potentiated the number of two-headed regenerants (82±6%) compared to the number of bipolar worms observed in control RNAi cohorts (59±5% at submaximal PZQ). Estimation of the effectiveness of knockdown of S7.1 transcripts was assessed by qPCR analysis in the same cohorts used for the regenerative assays. These assays revealed a decrease of 43±3% of S7.1 mRNA relative to controls (Fig 2C, inset).
Aside from the polarity effect on regenerating fragments, S7.1 RNAi also impaired the movement of intact worms. Planarians subject to S7.1 RNAi showed decreased mobility (Fig 2D), quantified by monitoring the distance traversed by S7.1 RNAi worms (61±4mm, average of 10 worms, n = 3 independent RNAi cohorts) compared with controls (104±7mm) over the same time period (2 mins). RNAi targeting other receptors in the S7 clade failed to yield a clear defect. We conclude knockdown of S7.1 modulated both regenerative polarity and motility outcomes.
Next, we employed a pharmacological approach to manipulate serotonergic signals by screening agents with known affinity for serotonergic receptors. While diverse classes of serotonergic blockers caused regenerative bipolarity, the penetrance was typically much lower than seen with PZQ (Table A in S1 Text). Results with ergot alkaloids were however of interest. Ergot alkaloids are a historically important class of compounds that realize their effects because of the close structural similarity of the ergoline scaffold to bioaminergic transmitters. Numerous ergot compounds yielded regenerative phenotypes, either phenocopying PZQ to promote bipolar (‘two-head’) regeneration or inhibiting head regeneration (‘no-head’, Fig 3A), all at doses lower than PZQ.
This broad efficacy of ergot alkaloids as a chemical class permitted structural-activity insight into features associated with specific polarity effects: for example, all ergots that caused bipolarity were either alkylated on the indole nitrogen or halogenated at the adjacent 2-position (Fig 3A). In contrast, all ergots that inhibited head regeneration lacked such modifications (Fig 3A). Structural studies have shown that the indole N1 hydrogen forms a key hydrogen bond with a conserved threonine residue T3.37 [27] within the orthosteric binding pocket of 5-HT GpCRs that is likely important for receptor activation [27,29]. This residue is also well conserved in the planarian 5-HT receptors sequences (Table 1). Disruption of this interaction by receptor mutagenesis interferes with 5-HT receptor activation by ergot alkaloids [29]. Similarly, alkylation of ergot derivatives at the N1 position also can cause decreased receptor activation yielding compounds that act as 5-HT receptor antagonists [30]. Therefore, this structural feature of the bipolarizing ergot compounds suggests they work through serotonergic blockade. This is consistent with observations that (i) structurally diverse 5-HT antagonists cause bipolarity (Table A in S1 Text), (ii) the ergots that inhibited head regeneration act as 5-HT agonists in other systems [31–33], (iii) other drugs that stimulate 5-HT signaling (8-OH DPAT and fluoxetine) also block head regeneration and PZQ action [20], and (iv) RNAi of tryptophan hydroxylase (TPH) to decrease 5-HT levels potentiates PZQ action [20]. Therefore, these data show that PZQ action mimics the bipolarizing ability of serotonergic blockers, and is opposed by 5-HT agonists.
The importance of identifying new drugs from planarian regenerative screens extends beyond basic science as planarian regenerative assays can predict the efficacy of compounds against parasitic worms [20]. Exploiting this phenology may assist discovery of new drug leads and targets for treating parasitic disease. Therefore, we were interested to assess whether the same set of compounds active in regeneration assays displayed activity against schistosomules, the immature form of parasitic schistosome flatworms that exist after penetration of host skin. Schistosomes display an endogenous contractile cycle permitting drug-evoked effects to be easily screened (paralysis versus stimulation of contractility, Fig 3B). In schistosome contractility assays, the compounds that caused planarian bipolarity all inhibited schistosomule motility (just like PZQ), whereas the compounds that inhibited planarian head regeneration caused the opposite effect, stimulating contractile activity (Fig 3B). Therefore, ergot alkaloids possess efficacy against schistosomules, with an action predictable by planarian polarity outcomes.
What is the molecular basis of this phenology between planarian regenerative polarity and schistosome motility? An appealing explanation relates to the recent identification of muscle cells as the coordinating nexus of positional signaling during planarian regeneration [21]. Specifically, a subepidermal population of myocytes was identified to coexpress all the relevant ‘position control genes’ known to regulate the planarian body plan, from which positionally appropriate transcriptional responses are engaged on injury [21]. This discovery is enlightening as it harbors the potential to rationalize a long literature on the effects of exogenous agents on regeneration dating back decades by suggesting that drugs which miscue regenerative patterning all possess a shared ability to modulate excitable cell physiology and perturb muscle function.
Therefore, we examined whether PZQ and the ergot alkaloids discovered to miscue polarity, impacted planarian motility. Acute incubation of intact worms with PZQ caused worms to adopt a spastic, curled morphology with inhibitory effects on worm motion (Fig 4A). This effect was dose-dependent (Fig 4B), with a concentration-dependence similar to that observed for the (longer term) polarity effect (EC50 = 38±3.6μM for bipolarity in trunk fragments versus EC50 = 23±2.4μM for mobility in intact worms, Figs 1B and 4B) and reversible following drug removal (Fig 4C). Just as observed with PZQ-evoked bipolarity, the immobilizing action of PZQ was also reversed by co-incubation with O-MT, but not other bioaminergic neurotransmitters (Fig 4C). Finally, each of the ergot compounds discovered to cause bipolarity (Fig 3A) also inhibited planarian mobility (Fig 4D), underscoring the association between polarity-miscuing drugs and the ability to perturb flatworm muscle function.
Finally, we returned to the fundamental observation of functional antagonism between serotonergic signals and PZQ action (Figs 1 and 4C). Does serotonergic activation modulate PZQ-evoked immobility in schistosomules? To address this, we examined the ability of exogenous O-MT to reverse PZQ-evoked effects on contractility (Fig 5A) and morphometry (Fig 5B and 5C). Addition of O-MT markedly ameliorated both the paralytic and compressed worm morphology resulting from PZQ exposure (Fig 5). Therefore, we conclude that PZQ action is functionally antagonized by serotonergic activation in schistosomule motility experiments, just as observed in planarian assays (Figs 1 and 4C).
The observations that serotonergic activation opposes, while serotonergic inhibition mimics PZQ action, reveal that serotonergic signaling is an important modulator of PZQ efficacy in vivo. We have previously suggested that PZQ engages dopaminergic pathways to subvert regeneration and it is noteworthy that both dopamine and serotonin regulate cAMP turnover with opposing effects on flatworm musculature [34–36]. Levels of cAMP change during regeneration [37] and cAMP is a known mediator of flatworm muscle contraction [38]. Therefore, this ‘functional antagonism’ model (Fig 6, [20]) envisages opposing Ca2+ entry pathways (Cav1A versus Cav1B) coupling to discrete bioaminergic neurotransmitters that differentially couple to cAMP within the excitable cell niche. Functional opposition of these bioaminergic systems are well evidenced in many systems [39].
Importantly, we demonstrate here that ergot alkaloids are efficacious modulators of planarian regeneration and motility (Figs 3 and 4). These two phenotypes are linked as surprisingly planarian polarity genes localize in a supepidermal population of muscle cells [21]. Indeed, ergot alkaloids have a well appreciated ability to modulate smooth muscle contraction based on their bioaminergic mimicry, a property that underpins several of their applications in the clinic. Beyond this ability to regulate muscle (including opposing effects on flatworm musculature [34–36]), dopamine and serotonin also are known regulators of Wnt signaling. D2Rs selectively associate with both β-catenin (to inhibit Wnt signals [40]) and Cav channels (to regulate their expression [41]). 5-HT is a well-established wounding signal [42], long range messenger involved in regenerative proliferation [43,44] and a reciprocally permissive cue for Wnt signaling [45,46]. Such associations provide precedence for coupling bioaminergic activity to the more established players of planarian regenerative signaling that localize with a myocyte population likely regulated by bioaminergic cues. Possibly all drugs that miscue regenerative polarity share such a commonality of action on the excitable cell niche.
Planarian regenerative screens hold predictive significance for discovering new drug leads and targets in parasitic flatworms [20]. Given the ease of performing drug screens in free living planarians compared to their parasitic cousins, this could be a fruitful source of novel therapeutic leads. 5-HT signaling in parasitic schistosomes is an appealing choice for therapeutic intervention given the dynamic expression of serotonergic gene products across the parasite life cycle [47–49] and a clearly evidenced role for bioaminergic signals in regulating muscle [34–36]. Parasite survival within the host requires worm muscle functionality: for example, muscle activity appears to be required for female pairing within the male gynecophoric canal, egg production and maintaining adult worm residency within the mesenteric vasculature. Paralytic agents such as PZQ have been proposed to act as antischistosomals by causing immobilized worms to shift from the mesenteric veins to the liver where they are eliminated [50]. Therefore, miscuing muscle function through bioaminergic cues is a promising route for drug intervention. Our data, revealing an ergomimetic quality to PZQ action, provide impetus for considering ergot alkaloids as potential drug leads for manipulating bioaminergic GpCRs to provide next generation antischistosomals [51]. Ergot alkaloids have been used clinically in a range of applications (migraine, obstetrics, Parkinson’s disease, diabetes), although owing to their broad GpCR binding profile they are often written off as problematic, ‘dirty’ compounds [30] and therefore often deliberately excluded from drug screens. However, this may be an oversight in the context of parasitic chemotherapy. Certain ergot compounds are penetrant and potent in planarian assays compared with PZQ. Further, the clear structural-activity principles emerging from our screen in free-living and parasitic worms (Fig 3) could illuminate structural differences in flatworm GpCR structure compared to their human hosts that may facilitate parasite targeting and mitigate host side effects. Based on our data from the planarian polarity and motility screens that are predictive of parasitic worm phenotypes, we contend that the ergot alkaloid scaffold merits further exploration by medicinal chemistry to identify novel chemotherapeutics with efficacy against parasite muscle.
A clonal line of Dugesia japonica (GI strain) was maintained at room temperature and fed strained chicken liver puree once a week [52]. Regenerative assays were performed using 5 day-starved worms in pH-buffered Montjuïch salts (1.6mM NaCl, 1.0mM CaCl2, 1.0mM MgSO4, 0.1mM MgCl2, 0.1mM KCl,1.2mM NaHCO3, pH 7.4 buffered with 1.5mM HEPES) and regenerative phenotypes archived using a Zeiss Discovery v20 stereomicroscope and a QiCAM 12-bit cooled color CCD camera [52]. Data were analyzed using two-tailed, unpaired t-tests, and presented as mean ± standard error of the mean from at least three independent assays.
Commercially available ergot alkaloids were sourced as follows: Sigma (bromocriptine, metergoline, nicergoline, ergotamine, dihydroergotamine); Tocris (LY215840, mesulergine, methylergometrine); THC Pharm (BOL-148, lysergol, elymoclavine). All other chemicals were from Sigma-Aldrich except where specified.
Total RNA from regenerating and intact D. japonica was harvested in Trizol and mRNA was purified by hybridization to oligo(dT) beads (Dynal). RNA-seq libraries were prepared according to the Illumina mRNA-Seq Sample Prep kit and Illumina TruSeq kit manufacturer protocols. Libraries were sequenced on Illumina HiSeq 2000 machines, producing 100bp paired end reads. Adapter sequences were trimmed and reads were passed through a sliding window quality filter (window size = 4, minimum average quality score = 25) using Trimmomatic version 0.22 [53]. Paired-end reads and singletons ≥ 50 bp in length were retained. Overlapping paired-end reads were merged using FLASH [54]. Surviving reads were combined and fed into the Trinity pipeline for de novo assembly [49]. Final assembly was carried out with a minimum k-mer coverage of 2 and the default k-mer size of 25. Complex graphs that proved unresolvable within a 6 hour window were manually excised to allow the assembly to proceed. The minimum contig or transcript length for both assembly pipelines was set to 100 nt. Candidate D. japonica 5-HT receptor sequences were selected based upon homology to receptors predicted in the planarian Schmidtea mediterranea [24]. Alignments were performed on predicted amino acid sequences in SeaView (version 4.5.1) using MUSCLE. Maximum likelihood phylogenies were generated using PhyML at 500 bootstrap replicates and visualized using FigTree (version 1.4.0).The depth of this resulting assembly proved comparable to transcriptomes generated for other planarian species [55,56], as well as the predicted open reading frames of the S. mediterranea genome [57], indicating that this resource is a reliable reference for the prediction and cloning of D. japonica gene products. The high level of coverage is evidenced by the fact that, of the 983 existing D. japonica nucleotide sequences manually cloned and deposited on NCBI, 982 are represented in our de novo assembly with a high degree of sequence identity. Reads were mapped onto the de novo assembly using RSEM [58] to obtain FPKM values reflecting transcript abundance. Sequences are provided as Supplementary material (Datasets A and B in S1 Text).
Total RNA was isolated from 50 starved, intact planarians using TRIzol and poly-A purified using a NucleoTrap mRNA mini kit. cDNA was synthesized using the SuperScript III First-Strand Synthesis System (Invitrogen). Gene products were amplified by PCR (LA Taq polymerase), ligated into pGEM-T Easy (Promega) for sequencing, and subcloned into the IPTG-inducible pDONRdT7 RNAi vector transfected into RNase III deficient HT115 E. coli. In vivo RNAi was performed by feeding [52], and a Schmidtea mediterranea six-1 (Smed-six-1) construct, which did not yield a phenotype in D. japonica, was used as a negative control. Cohorts of worms were fed bacterially expressed dsRNA targeting individual 5-HT receptors or the negative control over a total of five feeding cycles (three RNAi feedings separated by 1–2 days, followed by amputation, regeneration, two more RNAi feedings, followed by excision of trunk fragments for regenerative assays). Targeted sequences for RNAi are provided in Supplementary Materials (Dataset A in S1 Text). Knockdown was assessed by quantitative RT-PCR. Total RNA was isolated from 10 intact worms, treated with DNAse I (Invitrogen) and cDNA synthesized using oligo(dT) primers and the SuperScript III First-Strand Synthesis System. Samples (10 intact worms) were homogenized in Trizol to extract total RNA which was treated with DNAse I (Life Technologies) and 500ng were used for cDNA synthesis using random hexamers (SuperScript III First-Strand Synthesis System, Life Technologies). No RT controls were produced by using the same procedure but substituting DEPC-treated water for SuperScript RT enzyme. TaqMan qPCR reactions were performed using custom-designed TaqMan Gene Expression Assays (Applied Biosystems). Assays were designed for GAPDH (F’ GCAAAAGACTGTTGATGGACCAT, R’ CACGGAAAGCCATTCCAGTTATTTT, probe sequence CCTCTGCCATCTCGCC) and 5-HTR 7.1 (F’ CAATCTATCAAGGTTAGCTATTCCATTCGA, R’ GCTCCCACAACGATAATAAAAAATATAATCCC, probe sequence ACCAACCGGATATTTT) and cycled in a StepOnePlus Real-Time PCR System (Applied Biosystems) at 50°C/2min, 95°C/10min, 40 cycles of 95°C/15sec and 60°C/1min. 5-HTR 7.1 mRNA abundance was quantified by the ΔΔcT method relative to GAPDH.
Starved worms were exposed to drug / vehicle for five minutes, after which 10 animals were placed in drug-containing solution in the middle of a glass watchglass (50mm diameter, Fisher Scientific) centered over a LED backlit light (Edmund Optics, #83–873). Movement was captured using a digital video camera (Canon VIXIA HF R400) over a 2 minute period (30 frames per second). Representative images of this assay are displayed as minimal intensity z-projections (ImageJ) to provide a qualitative visual readout of experimental manipulations. The resulting videos were processed using custom written algorithms in Ctrax to track the motility of individual worms [59]. Motion was scored by quantifying total distance travelled (mm) over the fixed recording interval and averaged for the 10 worms in each assay. Errors in tracking were corrected using the Fix Errors Matlab Toolbox and descriptive statistics were computed using scripts in the Behavioral Microarray Matlab Toolbox and custom written algorithms in MATLAB.
Biomphalaria glabrata snails exposed to miracardia (NMRI Puerto Rican strain of Schistosoma mansoni) were obtained from the Biomedical Research Institute (Rockville, MD) and maintained at 26°C for 4 to 6 weeks. Isolation of matured cercaria and their transformation into schistosomules were performed as previously described [20]. For contractility assays, a custom written plugin (wrMTrck) in ImageJ was used to resolve schistosomule body length (major axis of an ellipse) over time following drug exposure (30min), as previously described [20]. For experiments on PZQ and 5-HT action on schistosomules, Basch media was made without 5-HT and drugs were added to the concentrations indicted.
|
10.1371/journal.pcbi.1000077 | Measuring Global Credibility with Application to Local Sequence Alignment | Computational biology is replete with high-dimensional (high-D) discrete prediction and inference problems, including sequence alignment, RNA structure prediction, phylogenetic inference, motif finding, prediction of pathways, and model selection problems in statistical genetics. Even though prediction and inference in these settings are uncertain, little attention has been focused on the development of global measures of uncertainty. Regardless of the procedure employed to produce a prediction, when a procedure delivers a single answer, that answer is a point estimate selected from the solution ensemble, the set of all possible solutions. For high-D discrete space, these ensembles are immense, and thus there is considerable uncertainty. We recommend the use of Bayesian credibility limits to describe this uncertainty, where a (1−α)%, 0≤α≤1, credibility limit is the minimum Hamming distance radius of a hyper-sphere containing (1−α)% of the posterior distribution. Because sequence alignment is arguably the most extensively used procedure in computational biology, we employ it here to make these general concepts more concrete. The maximum similarity estimator (i.e., the alignment that maximizes the likelihood) and the centroid estimator (i.e., the alignment that minimizes the mean Hamming distance from the posterior weighted ensemble of alignments) are used to demonstrate the application of Bayesian credibility limits to alignment estimators. Application of Bayesian credibility limits to the alignment of 20 human/rodent orthologous sequence pairs and 125 orthologous sequence pairs from six Shewanella species shows that credibility limits of the alignments of promoter sequences of these species vary widely, and that centroid alignments dependably have tighter credibility limits than traditional maximum similarity alignments.
| Sequence alignment is the cornerstone capability used by a multitude of computational biology applications, such as phylogeny reconstruction and identification of common regulatory mechanisms. Sequence alignment methods typically seek a high-scoring alignment between a pair of sequences, and assign a statistical significance to this single alignment. However, because a single alignment of two (or more) sequences is a point estimate, it may not be representative of the entire set (ensemble) of possible alignments of those sequences; thus, there may be considerable uncertainty associated with any one alignment among an immense ensemble of possibilities. To address the uncertainty of a proposed alignment, we used a Bayesian probabilistic approach to assess an alignment's reliability in the context of the entire ensemble of possible alignments. Our approach performs a global assessment of the degree to which the members of the ensemble depart from a selected alignment, thereby determining a credibility limit. In an evaluation of the popular maximum similarity alignment and the centroid alignment (i.e., the alignment that is in the center of the posterior distribution of alignments), we find that the centroid yields tighter credibility limits (on average) than the maximum similarity alignment. Beyond the usual interest in putting error limits on point estimates, our findings of substantial variability in credibility limits of alignments argue for wider adoption of these limits, so the degree of error is delineated prior to the subsequent use of the alignments.
| The study of genomics, and much of computational molecular biology, is about the inference or prediction of discrete, high-dimensional (high-D) unobserved variables, based on observed data. For example, in RNA secondary structure prediction, the challenge is to select a specific set of base pairs from a combinatorially large collection, as a prediction of the secondary structure of an RNA polymer, given its sequence. Similarly, in pathway inference, the challenge is to select a set of graph edges to connect genes or their products (nodes) from a combinatorially large collection of possible edge sets, based on gene expression or other data. Model selection problems for studying diseases stemming from mutlifactorial inheritance are becoming increasing common in the post-genome era. In these studies, the ultimate goal is to identify the combinations of genes responsible for inheritance components of disease etiology based on genetic and/or other post-genome data. In motif finding, the challenge is to select a single member of a large ensemble of possible combinations of motif sites in a set of sequences. Procedures that select the single best scoring solution, such as maximum similarity, maximum likelihood, maximum a-posteriori (MAP), or minimum free energy, dominate nearly all of these problems.
Sequence alignment is a typical example and is arguably the most important high-D discrete prediction problem for biology. Because it is the cornerstone capability used by a multitude of computational biology applications, we employ sequence alignment to make these general concepts concrete. Sequence alignment methods commonly focus on identifying the highest scoring alignment between two sequences, and assessing the statistical significance of this alignment [1]–[7]. Thus, alignment algorithms, heuristic [5], [8]–[10] (http://www.ncbi.nlm.nih.gov/BLAST/) and optimization [11] (http://fasta.bioch.virginia.edu/fasta_www2/) alike, typically report the selected alignment, and a statistical score that assesses how likely an alignment with a score as good or better could have emerged by chance, under a specified null distribution (commonly an E-value). While methods that assign the significance of alignments under a null distribution have been well studied, assessments of the uncertainty of a proposed alignment, defining the confidence in this alignment and assessing its overall reliability, have received considerably less attention.
Regardless of the alignment procedure employed, when a single alignment is chosen for the comparison of two (or more) sequences, it is a point estimate (or estimating alignment) selected from a large ensemble of all possible alignments. For example, two sequences of length m and n have possible local alignments, where k represents the number of matches in the alignment [11],[12]. This number grows rapidly with the length of the sequences being aligned; for example, two small sequences of only length 20 generate over 1029 possible local alignments. The question addressed here is: How, based on the available data, should we articulate the overall uncertainty of a selected estimating alignment (how well does it represent the large ensemble of possible solutions), and thus assess the reliability of this alignment?
The traditional approach to address the reliability of a single alignment is to evaluate the optimal alignment in the context of a set of near-optimal alignments. Near-optimal or suboptimal alignment analysis involves evaluating residue alignment consistency over the set of defined near-optimal alignments [12]–[17]. Specifically, the reliability of an alignment position (i,j) is assessed by comparing the score of the optimal alignment to the score of this alignment under the constraint that positions i and j do not align [14],[15]. More advanced methods have been proposed that determine reliability measures between residues aligned to both residues and gaps [17]. An alternative to computing near-optimal alignments, involving a single model that assigns probabilities to a specific residue pair, such as a pair Hidden Markov Model [7],[18],[19], can be derived and used to assess the reliability of individual aligned pairs.
With this in mind, these near-optimal alignment and model-based methods have offered significant improvements in reliability for tasks such as structural alignment. However, these methods are focused on delineating the reliability/uncertainty of the individual components of an estimated alignment, not the reliability of an estimated alignment in the context of the entire alignment space. There are methods to assess the accuracy of an alignment in the prediction of a ground-truth standard such as an alignment based on crystal structures [7], [18], [20]–[22]. But our focus here is on assessment of the reliability of an alignment based on its own characteristics, rather than the assessment of its accuracy in predicting an established reference. Toward this end, we describe a procedure for global assessment of the degree to which the members of the ensemble may depart from a selected estimate.
The introduction of probabilistic alignment methods [23]–[26] established the notion of sequence alignment as an inference procedure. For example, optimization-based alignment routines often search for the single alignment that is most probable among all those in the entire space of alignments. It is not surprising, given the immense size of the alignment space, that the most probable alignments, and thus all individual alignments, often have very small probabilities. This finding raises three questions:
We suggest the following answers to these questions:
To address these questions and test our proposed answers, we employ a Bayesian probabilistic approach. In the Methods section, we review some concepts on probabilistic alignments and distance measures, and then consider the distribution of the distances of the alignments in an ensemble from a proposed estimating alignment, including the quantiles and expected value of this distribution. We use the quantiles to identify credibility limits. The identification of credibility limits begs the question: What procedures can be developed to identify alignments with tight credibility limits? In an effort to achieve this goal, we employ statistical decision theory to find an estimation procedure that identifies the estimates with the minimum average distance from the posterior weighted ensemble; that is, the centroid. Centroid estimators, which were recently described by Carvalho and Lawrence [27], look promising to yield tight credibility limits because they minimize an average Hamming distance. Furthermore, we show that since popular procedures that select an estimate because it scores better than any other single solution (e.g., maximum likelihood, maximum similarity, maximum a-posteriori Viterbi solutions) are optimal under a zero/one-loss function, there is no principled reason to expect them to have tight credibility limits and, thus, to have high credibility. Below we compare the credibility limits for centroid alignments to those for maximum similarity alignments.
A statistical model that yields a probability distribution over an ensemble of solutions is essential for the characterization of uncertainty. Specifically, we are interested in using the data, in combination with any parameters that have been specified, to assign “posterior” probabilities to the members of the ensemble. We call these posterior probabilities because they are assigned after considering the implications of the data, the posterior weighted ensemble. Because in high-D settings it is often impossible to characterize the entire immense ensemble of solutions, it is common practice to employ representative samples from the posterior distributions to draw inferences or make predictions [28].
A probabilistic alignment model from which samples can be drawn can be described as follows. An alignment describes a set of aligned residues and associated insertion and deletion events. For a pair of sequences, and , let A be a matrix that characterizes an alignment whose (i,j)-entry is defined as:Without loss of generality, let I≤J. Because a residue cannot align with more than one other residue, two constraints must be satisfied, and . In addition, the alignment co-linearity constraint requires that Ai,j+Ak,l≤1, i≤k, l≤j. Let Θ be a matrix of residue pair similarities, such as one of the BLOSUM [29] or PAM [30] scoring matrices, and let Λ = (λo, λe) be the probability of opening and extending a gap, respectively.
Most sequence alignment methods optimize an objective function that can be described, based on a probabilistic model, as a log-likelihood [31],[32]. In traditional (frequentist) statistics, only the observed data, here R(1) and R(2), are seen as random variables, and the remaining terms are deterministic variables with perhaps unknown values. In maximum likelihood estimation, the values of these unknowns, which maximize the likelihood, are the maximum likelihood estimates. Typically, the user must set specific parameter values for the scoring matrix Θ0 and gap probabilities Λ0 to find the most probable alignment A* over all possible alignments:(1)This alignment is guaranteed to be the alignment that has the largest probability over all possible alignments, and with appropriate re-parameterization, it can also be shown to be the maximum similarity (MS) alignment [19].
To capture the entire alignment space in a probabilistic manner, the problem of alignment can be formulated as a Bayesian inference problem [19],[23],[26]. The Bayesian Algorithm for Local Sequence Alignment (BALSA) [24] describes such a probability model, the full joint distribution of all alignments, as the product of the likelihood and priors:Recursion can be employed to marginalize (i.e., sum out) over all possible alignments to obtain the marginal probability of the data in the two sequences, given only the defined scoring matrix, Θ0, and gap penalties, Λ0:The required sums are completed in an analogous manner to the Smith-Waterman recursion by essentially replacing the maximum function with a summation. The alignment parameters Θ and Λ can also be defined as random variables and marginalized over using Markov chain Monte Carlo (MCMC) sampling methods. In this application, to mirror common alignment practice, a specific scoring matrix (PAM 110) and gap-penalty parameters (gap opening = −14 and gap extension = −2) were selected as generic parameters used by sequence alignment algorithms. Now the probability of any single alignment can be computed as a posterior probability using the following Bayes formula:(2)Equation 2 is a ratio of the likelihood of the data and the alignment A* to the sum of these joint likelihoods over all alignments. It approaches a value of 1 when a single alignment dominates all others.
Given that the number of possible alignments for even small biopolymer sequences is immense, it is not feasible to calculate the probability of all alignments in a brute force manner. However, we can almost always use the recursive relationships that are fundamental to dynamic programming (DP) to draw guaranteed representative samples from the solution ensemble [19]. Because of the power of the recursions, such sampling procedures require no burn-in period to ensure that the samples are drawn from the equilibrium distribution, and these samples are independent of one another. Briefly, these algorithms use modified versions of the two fundamental steps of DP: the forward and back-trace recursions. In DP, the forward recursion finds the optimal value of the objective function (e.g., the best total alignment score) by using optimal solutions of subproblems to recursively build up to the best total score. In the sampling algorithm, we instead use an analogous recursion to build up to the sum over the entire ensemble of solutions. This sum finds the normalizing constant that assures that probabilities sum to one. In the back-trace step, instead of finding the solution that yields the optimal value of the objective function, we use an analogous recursion to sample solutions in proportion to their posterior probabilities. An important unappreciated fact is that for large ensembles, the accuracy of estimates based on a sample depends on the sample size only, and not on the size of the population [23]. Thus, a representative sample (i.e., a sample drawn in proportion to the probabilities of the unknowns) of even modest size, say 1000, can yield accurate estimates of unknowns, even if this sample is drawn from an ensemble of immense size. As we illustrate below, representative samples can be used to estimate credibility limits and define an ensemble centroid (EC) solution.
In this section, we describe procedures for finding credibility limits and mean distances for the sequence alignment problem. We begin by examining the distribution function of the distances of the ensemble members from a proposed estimate. Basic to this perspective are two concepts: 1) given the available data, the solution space is inherently uncertain; and 2) a proposed estimate is a point estimate (i.e., a single member of the ensemble) that is intended to represent the entire ensemble [33].
A simple measure of the difference between two members of a discrete ensemble (e.g., two possible alignments of a pair of sequences) is the Hamming distance. For two alignments, A(k) and A(m), of a pair of sequences, R(1) and R(2), of length I and J, the Hamming distance is simply the number of aligned positions that differ between A(k) and A(m), D(A(k), A(m)). For alignments, this distance is simply the sum of the differences in two binary matrices of size (I×J). When ensemble members are binary objects, the Hamming distances are also equal to distances on other scales [34]:(3)Using the metric in Equation 3, the distance between any proposed estimating alignment and the ensemble of alignments can be computed regardless of how one selects the estimating alignment. In this report, we compare the results of using two different estimating alignments: AM, the MS alignment, and AC, the EC alignment.
Specifically, let Di = D(Ai, Ax) be the distance of the ith member, Ai, of the ensemble from a proposed estimating alignment, Ax, where X is a categorical variable indicating the estimator (X∈[M, C]). We then rank the ensemble members by their distances from Ax, and let be the order statistics of these distances (i.e., the distances of the ensemble members from the estimating alignment) with the indices permuted to reflect their order in the distance ranking [35]. The distribution function of the distances is:(4)where d(1−α) is the (1−α)th quantile. Now the credibility limit at (1−α) is d(1−α). While higher-order DP recursions can be used to obtain these limits, they can also be quite reasonably estimated from a representative sample of even modest size by the following algorithm [35]:
The expected value of Di is(5)where An,i,j is 1 if i aligns with j in the nth member of the sample, and zero otherwise; qi,j is the marginal probability that An,i,j = 0; and pi,j is the marginal probability that An,i,j = 1. The required marginal probabilities can be estimated based on a sample, or when DP is available, they can be obtained using the forward- and back-trace algorithm described by Durbin et al. [19].
Hamming distances will, in general, be dependent on the lengths of the ensemble members. For example, in alignment, longer sequences will tend to return larger distances simply because the alignment matrix is larger. Thus, normalization is in order. For this normalization, we employ a normalization factor that uses maximum realized alignment lengths. Specifically, when calculating a credibility limit, the length of the estimating alignment (LE) is known, and the maximum length of an alignment in the ensemble is the length of the shorter of the two sequences (I). Thus, the maximum Hamming distances between an estimating alignment and the longest member of the ensemble is (LE+I). However, in our studies, we found that using this sum as a normalizing factor was misleading for cases in which the posterior space of alignments tended to be dominated by shorter local alignments. For example, the local alignments of the randomly shuffled sequences described in the Results section (see Figure 1) were dominated by short alignments. As a result, using (LE+I) as the normalizing constant in this case produced normalized distances that were not close to one, even when there were no base pairs in common between a sampled alignment and the estimating alignment. To adjust these differences, we used the length of the longest sampled alignment, LS, as the second term in our normalizing sum, and the normalizing distance between the estimating alignment Ax and the ith alignment in the sample is where S indicates the set of sampled alignments. Using this normalization factor yields normalizing distances with values between zero and one. A perfect match would yield an ND score of zero, and in the case where the longest sampled alignment has no base pairings in common with the estimating alignment, the ND score would be one. We define the credibility of the alignment at (1−α) to be ND(1−α).
Maximum similarity alignments, and the associated Viterbi alignments, have been the dominant alignment procedures for decades. In these procedures, an alignment output is typically the single alignment that has the maximum probability over all possible alignments. However, having the largest probability does not indicate that it represents the alignment space described by the billions (or more) possible alignments, except in the unusual event that this single alignment alone has high probability. In fact, the most probable alignment, the MS alignment, often has very small probability. For example, in this study, the probabilities of the MS alignments ranged from 10−37 to 10−249 for the alignments of the human/rodent pairs of gene and promoter sequences. Because it is the most probable alignment for a pair of sequences, all other alignments for that pair can be no more probable than the MS alignment. Thus, from a Bayesian prospective, any individual alignment represents the data only weakly at best.
As Carvalho and Lawrence [27] point out, procedures that identify the single, highest scoring alignment are optimal under a zero/one loss function. Accordingly, after the highest scoring alignments have been identified, all other alignments have a penalty of one (i.e., are all equally unimportant); thus, if no single alignment has a high probability mass then the expected loss will be large. As a result, with zero/one loss there is no reason for the optimal alignment to be positioned near any other member of the ensemble of alignments, therefore failing to garner support from any other member of the ensemble.
In contrast, centroid alignments garner information from the complete ensemble of alignments, because these alignments minimize the expected Hamming distance from the complete posterior weighted ensemble of alignments. Centroid alignments correspond directly to the reliable alignments of Miyazawa with a cut off 0.5 [26]. Reliable alignments are further described by Durbin et al. [19] and are elaborated on by Holmes and Durbin [34]. Furthermore, because these alignments minimize the average Hamming distance, we expect that they may yield tighter credibility limits than MS alignments. The alignment that is the centroid of the entire ensemble of alignments is called the EC alignment. These alignments meet the exclusive pairing and colinearity constraints of the alignment problem, but they do not necessarily meet the common requirement that a gap in one sequence cannot be followed by a gap in the other sequence. We compare the credibility limits of MS alignments and EC alignments below.
To assess the credibility measures and estimators described above, we examine the local alignments of sequences from (1) 20 orthologous genes between human and rodent, and (2) 24 orthologous genes between six species of Shewanella. All sequence pairs were evaluated using BALSA [24] with a PAM 110 scoring matrix, gap penalties of −14 and −2 for opening and extending a gap, respectively, and a sample size of 1000 to compute the estimated alignment distributions, credibility limits, and EC alignments.
The 20 orthologous genes for human/rodent are specifically up-regulated in human skeletal muscle tissue, and their upstream sequences have been used in previous studies to locate cis-regulatory modules [36]. The coding regions of the 20 human/rodent orthologous gene pairs were evaluated, as were the 20 sequence pairs that represent up to 3 kb of sequence upstream of the orthologous gene pairs. All sequence pairs were masked using RepeatMasker (http://www.repeatmasker.org/). For the local alignments of the 20 gene pairs and the 20 intergenic regions, we examined the credibility limits associated with two estimating alignments: the MS, and the EC. Specifically, we examined the 95% quantiles of the normalized distances (ND), computed based on the distances between these estimating alignments from the 1000 sampled alignments from the posterior alignment distribution. Figure 1 shows a scatter plot of the MS 95% credibility limits (MS ND95) versus the EC 95% credibility limits (EC ND95) for the local alignments of the genes and the intergenic regions. For contrast, the genes were randomly shuffled, and 95% credibility limits were defined for these non-related sequence pair alignments.
First, notice that the credibility limits for the gene sequence alignments are small, and the difference between the EC and MS is negligible. These genes are so highly conserved that the majority of the posterior distribution falls along a small set of paths with high probability, thus creating high correlation between the EC and MS. Alternatively, when the gene sequences are shuffled, the hyper-sphere surrounding 95% of the posterior distribution is very large because the probability of aligning any two residues is essentially random. This results in extremely large credibility limits with high deviation in the distance of the ensemble from the EC and MS. The intergenic regions are less conserved than the genes and, thus, are intermediate between these two extremes. Notice that the credibility limits are often surprisingly large, with normalized distances over 50% for 18 of the 20 MS alignments, and for 17 of the 20 EC alignments. This indicates that we have confidence in less than half the predicted aligned base pairs. As the plot shows, there is considerable variation in the credibility limits over the 20 examples when either the EC or MS limit is used. The credibility limits for the EC range from 29% of maximal to nearly 91%, while the MS limits range from 37% to almost 100% of maximal. This result highlights the need to report credibility limits for every sequence pair. We also see that for all but one of the sequence pairs, the MS credibility limits are greater than those for the EC. Furthermore, for 11 of the 20 upstream sequence pairs, the MS credibility limits were more than 600 base pairs larger than EC credibility limits. Thus while the differences in Figure 1 look modest, the MS credibility limits are often hundreds of base pairs larger than those of the EC estimators.
Taken together, the differences between the 20 MS normalized distances and 20 EC normalized distances in Figure 1 are significantly different (i.e., p<0.001, Wilcoxon Signed Rank test [37]). To offer further insight, we chose four alignments from the 20 to examine in more detail (Table 1); the results for all 20 pairs are in Table S1. In Figure 2, we show histograms of the distance of the 1000 sampled alignments from the two estimating alignments (MS, EC); in addition, the 95% quantile (ND95) for the EC and MS are shown as bars, and the values are given in Table 1. As Figure 1 indicates, pair (A) has the tightest credibility limits of all the promoter sequences. These tighter limits are a reflection of the fact that the ensemble of alignments is relatively close to the estimators; the 95th percentile alignment differs from the EC estimator by 270 of a possible 1556 base pairs that could potentially differ (ND95 = 0.29), while the MS is about 20% larger with an ND95 = 0.37. Of the 20 promoter sequence pairs, there are 11 in which the two credibility limits are markedly different (i.e., by more than 0.05). Figure 2D is another illustration of the characteristics of these 11 pairs for which the MS credibility limits are substantially larger than those of the EC, although for pair (D) the distance distributions have very little overlap, as well as large credibility limits. Figure 2C is representative of the remaining nine pairs, in which the posterior surface is quite flat, and the two credibility limits differ by less than 0.05. For the sequence pair shown in Figure 2C, the credibility limits for both estimators are large. Because the EC alignment is the nearest alignment to the mean [34], the large size of this limit for the EC alignment indicates that the alignments in the posterior distribution are widely dispersed over the ensemble. Also notice that in (B) and (C), the two distributions overlap substantially and have high ND95 values; for example, the alignment in Figure 2B shows a ND95 = 0.72 for the EC, and ND95 = 0.77 for the MS alignment. Because the centroid estimator is the closest feasible alignment to the mean, for this sequence pair the mean and the mode are close, as is typical of symmetric distributions [27].
We also examined the credibility limits for the MS and EC estimators for local alignments of orthologous pairs of intergenic regions (up to 500 bp upstream of orthologous genes) from six species of Shewanella for which full genome sequence data are available: 1) S. denitrificans OS217 (DENI), 2) S. loihica PV-4 (SPV4), 3) S. oneidensis MR-1 (SONE), 4) S. putrefaciens CN-32 (CN32), 5) Shewanella sp. MR-4 (SMR4), and 6) Shewanella sp. MR-7 (SMR7). We chose SMR4 as our base species, aligning orthologous sequences from each of the other five to the region from SMR4. Starting with SMR4, the species in order of increasing evolutionary distance are SMR4>SMR7>SONE>CN32>SPV4∼DENI. As before, we examined the 95% quantiles of the normalized distances, computed based on the distances between the estimating alignments and the sampled ensemble of alignments drawn from the posterior alignment distribution. Figure 3 shows a scatter plot of the MS ND95 versus the EC ND95 values for each of 24 randomly selected orthologous regions, for the pairwise comparison of SMR4 to each of the five species at varying evolutionary distances (120 total comparisons).
The two species SMR4 and SMR7 are very closely related, having been isolated from samples taken at different depths (5 m and 60 m, respectively) from a single location (latitude and longitude) in the Black Sea [38]. Thus, it is not surprising that even the intergenic regions are highly conserved and that the EC and MS exhibit tight credibility limits. Among the comparisons to species at increasing evolutionary distance, we observe increasing credibility limits. In fact, for many of the SMR4-DENI sequence pairs, the credibility limits are no better than expected for randomly shuffled sequence. While, on average, the credibility limits of a pair of species increase with increasing evolutionary distance, the figure also shows that the credibility limits of the alignments for a given pair of species vary greatly. For example, even though the credibility limits of most SMR4-DENI pairs are large (>0.8), there are sequence pairs from these two species that have credibility limits <0.3. The fact that there is wide variability in credibility limits for all of these pairs of species, except SMR4-SMR7, highlights the importance of assessing the reliability (credibility limits) of nearly all alignments. For example, there is a pair of SMR4-CN32 sequences whose alignment is very reliable (EC ND95 and MS ND95<0.05), but there are also three pairs whose alignments cannot be trusted (EC ND95 and MS ND95>0.6), and the remainder are scattered over the full range in between.
We further evaluated the findings shown in Figure 3 in the context of a single gene's orthologous upstream sequences. Often in evaluating promoter sequences across species it is unknown a priori which sequences it would be most beneficial to align. The tight credibility limits shown in Figure 4A and 4B indicate that when evaluating the promoter region of SMR4_0576, we would have confidence in the alignments with the orthologous region from SONE and CN32 (also with SRM7, data not shown). This is not the case for the orthologous regions from SPV4 and DENI. The high ND95 values for the EC and MS alignments indicate that alignment of SPV4 or DENI sequences would not contribute to a meaningful evaluation of the SMR4_0576 promoter region. Unfortunately, not all alignments of promoter regions from SMR4 with the promoter sequences of orthologous genes in SONE and CN32 are reliable. For example, as Figure 5 shows, the posterior distribution of the alignments of the SMR4_ 1557 promoter region with its CN32 ortholog is substantially more widespread and variable than the posterior distribution of alignments for the promoter region of SMR4_0576 with its orthologous region in CN32.
These findings of large differences in the reliability of alignments within species pairs have had a substantial practical impact on our studies of phylogenetic motif finding using these Shewanella species. Specifically, alignment of orthologous promoters can substantially increase the power of motif finding, if the alignments can be trusted. However, the findings shown in Figure 5 indicate that reliance on a single genome-wide measure of species distances is very frequently insufficient to assure that alignments of promoters from species pairs can be trusted. Thus, we are using credibility limits on a gene-by-gene and species-by-species basis to make decisions about which alignments can be trusted.
The use of heat maps or other means to visually illustrate confidence in the individual alignment of individual pairs of bases must accommodate a different feature for centroid alignments. Specifically, EC alignments have a feature not present in standard alignments, in that they allow stretches of sequence in the middle of an alignment to remain unaligned in a manner analogous to those regions at the ends of local alignments. That is, a residue in one sequence that cannot be reliably aligned with any single residue in the other sequence is excluded from the centroid alignment. Aligning any such residues to any bases in the other sequence would only increase the average distance of the centroid alignment from the posterior distribution of alignments. In addition, with probabilistic alignment, we return marginal probabilities of all residue pairs. Therefore, to display all the features of this alignment, we employ 1) a traditional dash to represent gaps, 2) a dot to represent residues that cannot be reliably aligned and are thus ignored in the alignment, and 3) a gradient color scheme (i.e., a heat map) to show the base pair alignment probabilities, where red indicates high probability for that residue pair, green indicates probabilities nearing 50%, and the ignored region is grayed out to further differentiate those residues for which the variability in alignments is too great to permit marginal pair probabilities of 0.5 or greater. Figure 6 gives an example of the heat map alignment display for a human/rodent intergenic sequence pair (the region upstream of the MYL2 gene). The red-to-green coloring of aligned regions allows quick distinction of areas of alignment of high versus low confidence.
Because prediction and estimation involve making inferences about unknown quantities based on the available data, they are inevitably uncertain. Thus, when a specific value is reported as a point estimate, it is common in many fields to simultaneously report a confidence limit or a credibility limit, which is the Bayesian analog. Such limits are all too often absent in computational biology. Here, to promote their broader adoption, we describe a method for estimating credibility limits and illustrate these concepts using sequence alignment. These credibility limits are derived from the empirical distribution function of the Hamming distance from the estimator to the members of the ensemble of solutions, or more accurately, a representative sample of the ensemble of solutions. The 95% credibility limit of a proposed estimate describes the posterior distribution by indicating the normalized Hamming distance containing 95% of the probability mass of the posterior distribution. The existence of these limits begs the question: What estimation procedure will yield tight credibility limits? We advocate the use of recently developed centroid estimators that minimize the expected Hamming distance to address this question.
While it is reasonable to expect centroid estimators to produce tighter credibility limits, it is not a guaranteed product of this procedure, because the centroid is the estimator that minimizes the average differences from the posterior ensemble, while the credibility limits are based on a quantile. Nevertheless, our finding of tighter credibility limits for EC alignments compared to MS alignments should come as no surprise, since the well-known zero/one loss risk associated with the latter estimators provides no principled reason to expect that such estimators will be near the center of the posterior distribution of alignments. On the other hand, centroid alignments, which are the alignment nearest to the multivariate mean of the posterior distribution, are centered in the posterior distribution [27].
Our findings of 1) high variability in the credibility limits in the alignments of promoter sequences of 20 human/rodent sequence pairs and 2) similar high variability among 4 of the 5 pairs of Shewanella species highlight the need for assessing the overall reliability of sequence alignments. Without such limits, there is little to distinguish alignments that vary greatly from one another in their reliability. Furthermore, our findings indicate that centroid estimators have promising potential to improve sequence alignment. For example, for over half of the human/rodent non-coding sequence pairs (each of ∼3000 bases) in our sample, the EC and MS alignments differ by more than 600 base pairs, and similar relative differences are observed in Shewanella alignments. While we report here on the credibility of nucleotide sequence alignments, they are equally applicable and valuable for protein sequence alignments.
In some discrete high-D inference problems, the posterior ensemble of solutions may not only be asymmetric, but also it may be multimodal, as has been reported for RNA secondary structures [39]. Since, in such a case no single point estimate can reasonably represent the posterior ensemble, class-specific estimates, with one for each distinct class, will be required. In these cases, samples associated with each class can be used to find credibility limits for the class estimates, and the overall credibility limits around these class-specific estimates can be identified based on distances to the nearest class estimate.
As mentioned above, the probabilistic model used is a Smith-Waterman recursive DP algorithm whose Viterbi alignment corresponded exactly to the MS alignment reported here. Thus, differences in credibility limits reported here are solely the result of the differences in the estimation procedures. In addition, the alignment that minimizes expected Hamming distance loss and also follows the requirement concerning adjacent gaps in the two sequences are available using a DP algorithm [19],[34]. However this alignment can only increase the average Hamming distance above that of the centroid.
While we believe this evidence supports reconsideration of the maximum scoring alignment paradigm, stronger evidence for reconsideration has been in the literature for over a decade. In 1995, Miyazawa [26] was the first to report what we now call centroid alignments [27]. In addition to his very insightful development of reliable alignments, he showed that these alignments are superior, using x-ray crystal structures of proteins as ground truth. Figure 7 (reproduced from Miyazawa's work [26], with permission of the author and Oxford Journals) shows that structural predictions based on reliable (centroid) alignments quite consistently produce lower root mean squared deviations than those based on maximum similarity alignments. Thus, from a practical biological prospective, there is already clear evidence in the literature that centroid alignments can be applied with advantage in the prediction of protein structures.
We also note that the time complexity of algorithms for obtaining centroid alignments and credibility limits is not different from those of more traditional optimization based methods. When recursions can be employed to obtain optimal solutions via DP, analogous recursions are frequently available for associated probabilistic models, and stochastic back-trace procedures can be employed to draw samples from the posterior ensemble of solutions [19]. In general, the time complexity for drawing these samples will be the same as that of the associated DP algorithm, and is set by the forward step of these algorithms. For example, in local sequence alignment, the most computationally intensive step is the forward-recursive step. For two sequences of length n and m, the time complexity is O(n*m) for both the optimization and Bayesian algorithms. Running times to obtain credibility limits in a recursive setting will generally be longer than times required to obtain optimal estimates because a back-trace step must be executed only once to obtain the optimal, while it must be employed multiple times to draw samples. However, this sampling will not generally greatly increase overall running times, because back-trace recursions are usually of a lower time complexity than their forward steps. For example, for local alignments the time complexity of the back-trace recursions is only O(min(n,m)). For problems not open to recursive solutions, MCMC algorithms are commonly employed, using procedures like simulated annealing. Credibility limits and centroids also can be obtained using MCMC sampling with run times that may be less than those for optimizations [27].
Some caveats are appropriate. In settings in which uncertainty is low, such as shown for the alignments of coding regions of human/rodent sequence pairs in Figure 1 and the promoter sequence pairs of very closely related species like Shewanella sp. MR-4 and MR-7 in Figure 3, credibility limits will likely be tight and not vary greatly among examples. Nevertheless, it would be reassuring to document this low variability by reporting credibility limits. While we have given principled arguments supporting our belief that centroid solutions should dependably have tighter credibility limits than optimization estimators, this advantage cannot be guaranteed. However, this trend was observed in both the human/rodent pairs and the Shewanella pairs. In our on-going work with Shewanella, we have found 1329 orthologous genes that were present in all six species and computed the 95% credibility limits for both the MS and EC, for all the promoters from SMR4 aligned with the orthologous sequences from each of the remaining 5 strains. The EC ND95 credibility limits were smaller than the MS ND95 limits in 6078 (91.55%) of these 6645 sequence pairs (i.e., p<1e-100, Wilcoxon Signed Rank test [37]).
In our comparison of centroid alignments to MS alignments, we focused on the alignment of individual pairs of sequences. However, we did not address how these two estimators would compare if we had available multiple pairs of sequences all drawn from a model with a single common “true” alignment. In the context of sequence alignment, such a situation would not be observed in nature because we know of no families of biological sequence pairs for which one can be confident that sequence pairs within this family all follow the same “true” alignment. For example, even for sequence pairs drawn from orthologous regions from clearly related species, alignments are likely to differ. This same absence of replicates, all of which are sampled from the same “true” value of the unknown, is expected for many, but not necessarily all, high-D discrete biological inference problems. Even when obtaining a large number of such biological replicates is possible in principle, such as a large number of biological replicates in a microarray study, obtaining them in practice is often prohibitively expensive. However, with advances in technology, this limitation may be overcome. When a substantial number of such replicate observations are available, the asymptotic properties of maximum likelihood estimates, such as consistency and asymptotic unbiasedness, can be brought to bare. In such cases, as sample size increases, the MS estimator will approach the true value, and the bias will tend toward zero. This reduction in bias might well counter-balance the higher variability (high credibility limits) reported here for individual sequence pairs.
The findings reported in this paper are for pairwise alignments. When multiple alignments are employed, we expect credibility limits to narrow because of the increased size of the data sets; however, we caution that the alignment space grows rapidly with increasing sequences in an alignment. Therefore, these limits may or may not shrink as quickly as expected. Furthermore, it is important to keep in mind that the credibility limits reported here are sampling estimates of true 95% quantiles, but with samples of 1000 the error bars on these estimates are 95%±1.35%. All the estimates in this work are based on a local probabilistic alignment model. While local alignment is the most common procedure, other probabilistic alignment procedures, or local alignments with other parameter settings [25],[26], may give varying results. As is common practice, all alignments here are given for a fixed set of parameters. Alignment parameters also can be estimated from the data; perhaps with such an approach, credibility limits could be smaller and more consistent, although this may not be the case because uncertainty of the parameter estimates would be introduced into the procedure.
Beyond the usual interest in putting error limits on point estimates, our findings of substantial variability in credibility limits of alignments argues for wider adoption of these limits, so that the degree of error is delineated prior to the subsequent use of the alignments. From a practical prospective, when credibility alignments are tight, those using these alignments in subsequent procedures can be confident in the input alignments and know the limited degree to which input alignment may vary. The absence of such limits may well lead to a false sense of confidence in subsequent findings, especially when credibility limits are wide, and/or seriously limit an investigator's ability to determine the source of difficulties or inconsistencies in subsequent procedures that depend on these unreliable alignments. In practice, knowing early in a study that alignments required for subsequent results are unreliable (i.e., have high credibility limits) might well lead an investigator to reconsider his/her plans. For example, in studies of phylogenetic tree reconstruction when it is known that input alignments are reliable, investigators' conclusions about phylogenetic relationships will be bolstered; whereas, prior knowledge that input alignments are unreliable will motivate serious investigators to revise their study design or, after the fact, permit reviewers to raise legitimate questions about the studies conclusions.
While the results presented here concern only sequence alignment, the procedures described are generally applicable to point estimates for high-D discrete spaces; this includes many major inference problems in computational biology, such as pathway prediction in systems biology, the prediction of phylogenetic trees, the reconstruction of ancestral states, the delineation of alternate splice forms, and prediction of RNA secondary structures. For any of these problems, the algorithm given in the Methods section “Credibility limits and means distance” can be employed to obtain ND95 values for any proposed estimate given a procedure for drawing samples from the posterior distribution. We caution that while the Hamming distance will be appropriate in many of these areas, it may not be as appropriate in some of these settings. Regardless of the distance measure used, the proposed procedure will return credibility limits for an estimator when a representative sample can be obtained. We believe the use of confidence or credibility limits is long overdue throughout the full spectrum of discrete high-D inference problems encountered in computational biology. These limits have a number of valuable uses, including gauging the degree by which solutions might depart from their estimated value, appraising the overall credibility of a prediction, and comparing the performance of alternative estimators in cases where a “gold standard” is not available. |
10.1371/journal.pcbi.1005078 | Higher-Order Synaptic Interactions Coordinate Dynamics in Recurrent Networks | Linking synaptic connectivity to dynamics is key to understanding information processing in neocortex. Circuit dynamics emerge from complex interactions of interconnected neurons, necessitating that links between connectivity and dynamics be evaluated at the network level. Here we map propagating activity in large neuronal ensembles from mouse neocortex and compare it to a recurrent network model, where connectivity can be precisely measured and manipulated. We find that a dynamical feature dominates statistical descriptions of propagating activity for both neocortex and the model: convergent clusters comprised of fan-in triangle motifs, where two input neurons are themselves connected. Fan-in triangles coordinate the timing of presynaptic inputs during ongoing activity to effectively generate postsynaptic spiking. As a result, paradoxically, fan-in triangles dominate the statistics of spike propagation even in randomly connected recurrent networks. Interplay between higher-order synaptic connectivity and the integrative properties of neurons constrains the structure of network dynamics and shapes the routing of information in neocortex.
| Active networks of neurons exhibit beyond-pairwise dynamical features. In this work, we identify a canonical higher-order correlation in network dynamics and trace its emergence to synaptic integration. We find that temporally coordinated firing preferentially occurs at sites of fan-in triangles—a synaptic motif which coordinates presynaptic timing, leading to greater likelihood of postsynaptic spiking. The influence of fan-in clustering leads to the surprising emergence of non-random routing of spiking in random synaptic networks. When synaptic weights are made artificially stronger in simulation, so that cooperative input is less crucial, dynamics are no longer dominated by fan-in triangles but instead more closely reflect the random synaptic network. Thus, the emergence of fan-in clustering in maps of synaptic recruitment is a collective property of individually weak connections in neuronal networks. Because higher-order interactions are necessary to shape the timing of presynaptic inputs, activity does not propagate uniformly through the synaptic network. Like water finding the deepest channels as it flows downhill, spiking activity follows the path of least resistance and is routed through triplet motifs of connectivity. These results argue that clustered fan-in triangles are a canonical network motif and mechanism for spike routing in local neocortical circuitry.
| Understanding any complex system requires a mechanistic account of how dynamics arise from underlying architecture. Patterns of connections shape dynamics in diverse settings ranging from electric power grids to gene transcription networks[1–5]. It is critical to establish how synaptic connectivity orchestrates the dynamics of propagating activity in neocortical circuitry, since dynamics are closely tied to cortical computation. For example, trial-to-trial differences in network dynamics[6–9] can be used to decode sensory inputs and behavioral choice[10,11]. It is particularly important to understand the transformation from connectivity to activity within local populations of neurons since this is the scale at which the majority of connections arise. Locally, neocortical neurons are highly interconnected, and their connectivity schemes are characterized by the prevalence of specific motifs[12]. At the level of local populations, functional coordination has been demonstrated in diverse ways, e.g. on the basis of active neurons[13,14] and their correlation patterns[15]. Yet predicting population responses on the basis of pairwise connections alone has proven to be difficult.
Establishing a mechanistic link between connectivity and dynamics in neocortical networks is intricate and non-trivial because individual neurons themselves are complex computational units[16–20]. Fundamentally, neurons are state dependent non-linear integrators of synaptic input[21–23]. When neurons in neocortex process information, they are generally subjected to numerous synaptic inputs which activate diverse receptors, and concomitant gating of voltage-dependent channels[24–26]. In consequence, neocortical neurons tend to operate in a high-conductance state, which lessens the impact of any one synaptic input[21,27]. Because inputs are weak individually, collective synaptic bombardments are necessary to depolarize a neuron to threshold for action potential generation. As a result, it is difficult to predict the flow of activity through a synaptic network based solely on knowledge of single connections, without the context of ongoing activity in the entirety of the system.
Network models are an important tool for linking synaptic connectivity to dynamics in neocortex because they enable precise measurement and manipulation of simulated connectivity. In this work, we generate networks comprised of leaky integrate-and-fire model neurons with naturalistic dynamics that mimic recordings from superficial neocortical layers. Despite random synaptic topology in the model network, we find that small-world topological organization emerges in maps of propagating activity. This paradoxical divergence of dynamics from synaptic connectivity is not explained by coactivity alone. Rather, recruitment preferentially occurs in a selective subset of active connected pairs.
In the model, activity is preferentially routed through clustered fan-in triangles, despite their statistical scarcity. Because they result in coordinated presynaptic timing, fan-in triangle motifs are particularly effective for spike generation. By comparison, among neurons converging on a common target but lacking presynaptic interconnectivity, presynaptic timing is less synchronous on average, and postsynaptic recruitment is less likely. Moreover, when we decrease the need for cooperative presynaptic action, by doubling synaptic weights in network models, the fan-in triangle motif becomes significantly less prevalent. We evaluate the prediction of our model using high speed two-photon imaging of emergent network activity ex vivo, in somatosensory cortex. We verify that propagating activity in real neuronal networks has small-world characteristics and elevated clustering, Decomposing this clustering, we discover that neocortical circuitry also manifests propagating activity that is dominated by the fan-in triangle motif. These results suggest a mechanistic account for the widespread findings of clustered activity in neuronal populations [14,28–31]. We suggest that clustered fan-in triangles are a canonical building block for reliable cortical dynamics.
Multineuronal dynamics are the computational substrate for sensation and behavior, implemented by synaptic architectures. Propagating multineuronal activity arises from three main sources: the underlying connectivity itself, recent network history, and the non-linear integrative properties of individual neurons. Here, multineuronal activity was modeled using conductance-based leaky integrate-and-fire neurons, stimulated with brief periods of Poisson input and recorded during self-sustained firing (Fig 1A). Model neurons were connected with heterogeneous synaptic weights drawn from a heavy-tailed distribution, in a random arrangement (Erdős-Rényi; pee = 0.2). Simulated dynamics were asynchronous, irregular, and sparse, with critical branching (see Methods).
A synaptic network was constructed for each simulation, consisting of excitatory model neurons and their synaptic connectivity. For each structural iteration of the model we generated three distinct maps of activity (and in two of the cases, multiplex connectivity and activity): a functional network, the active subnetwork, and a recruitment network (Fig 2). Edges in the functional network summarized network dynamics and represented frequency of lagged firing between every pair of nodes (with maximum interspike interval T = 25 ms; see Methods). The active subnetwork was a subgraph of the synaptic network and consisted of model neurons active at least once and all their interconnections (regardless of lagged firing relationships). Finally, the recruitment network was a subgraph of the functional network defined by its intersection with the synaptic network, to map the routing of activity through synaptic interactions. In this way, non-zero edges in the recruitment network linked synaptically connected nodes that also spiked sequentially in the interval T at least once. For T = 25 ms, 10.9 ± 3.52 excitatory presynaptic input spikes immediately preceded each postsynaptic spike (mean±std).
Surprisingly, although underlying synaptic connectivity was Erdős-Rényi (i.e. random), functional activity networks were small world (Fig 1B)[32]. To judge the small world character of these networks, global clustering coefficient and characteristic path were normalized by their respective abundances in density-matched Erdős-Rényi networks and combined as a quotient[33]. Comparison with density-matches was important given that sparseness itself results in enhanced smallworldness[34].
Functional networks were marked by significantly increased small world scores (functional network: 2.8±0.23; synaptic network: 1.0±0.035; n = 5, p = 0.0079, Wilcoxon rank-sum) resulting from increased clustering (function: 2.8±0.23; synaptic network: 1.0±0.035, n = 5, p = 0.0079), with characteristic path lengths similar to random-matches (function: 1.0±6.4x10-4; synaptic network: 0.99±0.033; n = 5, p = 0.69). The lag interval T was chosen to encompass important network timescales for synaptic plasticity and integration[35,36]. We also generated functional networks using intervals of 10 and 50 ms, which showed that the emergence of non-random features does not depend strongly on choice of T (functional network for T = 10ms: small world ratio 3.2±0.24, n = 5, p = 0.0079; functional network for T = 50ms: small word ratio 2.6±0.22, n = 5, p = 0.0079).
Given modest sampling conditions (e.g. binning near timescales of synaptic integration), functional relationships can indicate locations of probable synaptic recruitment[35]. However, a subset of edges in functional networks are 'false positives'—they reflect polysynaptic relationships and other combined statistical dependencies rather than monosynaptic connectivity and recruitment[35,37]. To determine whether these measurement artifacts were responsible for the statistical differences between functional and synaptic networks, we turned to recruitment networks. Pruned of false positives, recruitment networks were significantly more small world than functional networks constructed from the same activity (4.6±0.87; n = 5, p = 0.0079), with even shorter characteristic paths (recruitment: 0.65±0.072, n = 5, p = 0.0079 compared to function, Wilcoxon rank-sum) and a similar elevation in clustering (recruitment: 3.0±0.26; n = 5, p = 0.22). Thus, emergent statistical structure in the functional networks reflected coordinated timing among multiple synaptically connected neurons.
As demonstrated by non-random recruitment, i.e. clustering in the recruitment network, activity did not propagate homogeneously through the random topology. However, it remained a possibility that the seemingly non-random routing of activity was simply the byproduct of shared activity, without being selective on the basis of connectivity. As a control, the active subnetwork establishes the role of interactions among neurons with elevated firing rates (including pairs of neurons which never recruited one another within the interval T). Compared to functional networks, the corresponding active subnetwork exhibited reduced small world ratio (active network: 2.2±0.26, n = 5, p = 0.0159) and reduced clustering (1.3±0.041, p = 0.0079), despite somewhat shorter characteristic paths (0.60±0.055, n = 5, p = 0.0079).
If directed connections that never fired sequentially were pruned from the active subnetwork, it would attain the same binary topology as the recruitment network. Comparing the active network with the recruitment network, global clustering ratio was significantly increased (from 1.3±0.041 to 3.0±0.26, n = 5, p = 0.0079, Wilcoxon rank-sum). Thus, the select connections which were directly involved in propagation of spiking activity were more clustered than activated connections as a whole (Fig 1C).
We next evaluated whether neuronal pairs that never fired sequentially differed from those that did. Comparisons were performed between in-degree matched samples. Connected neurons that never fired in succession shared significantly fewer neighbors than those that did fire sequentially at least once (n = 500 pairs, p = 3.1 x 10−17, Wilcoxon rank-sum). In the model, activity was selectively routed through interconnected neighborhoods.
Connectivity within a triplet is the simplest way two nodes can share a common neighbor and be clustered. However, this measure fails to account for the direction of connection. Since direction is crucial in synaptic communication, we turned to a formulation which differentiates directed triangle motifs[38]. From the perspective of a reference postsynaptic neuron, clustered neighbors can be arranged into four kinds of three-edge triangle motifs: fan-in, fan-out, middleman, and cycle arrangements (Fig 3A). Taken in isolation, fan-in, middle-node, and cycle triangles are isomorphic to one another through rotation, i.e. dependent on labeling the reference node (which is necessary to compute local clustering). Measures of undirected clustering can be decomposed fractionally into these four components. Because the underlying model synaptic connectivity was random, none of the four triangle motifs were more prevalent than the others, and each contributed equally to synaptic clustering (Fig 3B). By contrast, in recruitment networks, fan-in triangle motifs were highly overrepresented (Fig 3C). The overrepresentation of fan-in triangle motif was also present in the functional network: for example, iterative Bayesian inference[35] was sensitive to asymmetric directed clustering in model activity (fan-in: 0.38±0.052, fan-out: 0.29±0.032, middleman: 0.19±0.016, cycle: 0.15±0.0076; mean±std, threshold at the 95th percentile).
To understand whether these higher order asymmetric features emerge from chance correlations tied to firing rates, we generated Poisson populations that were rate-matched on a neuron-by-neuron and trial-by-trial basis. This resulted in an inhomogeneous distribution of firing rates across all trails. Our Poisson null populations had identical expected spike counts as model activity in each 100ms bin but no synaptic interactions and no causal propagation of activity. Undirected clustering was significantly lower in iterative Bayesian maps of uncoupled Poisson rate-matched activity compared to connected network models (Poisson rate-match: 0.0052±3.6x10-4; simulated activity: 0.024±0.013; Wilcoxon rank-sum p = 0.036; n = 3), and the fan-in triangle motif was not elevated relative to other clustering patterns (Fig 3D). The Poisson populations demonstrated that elevated fan-in triangle motifs do not result trivially from the analysis procedure but instead are the result of synaptic interactions between neurons. Interestingly, we found that model neurons with high fan-out clustering were characterized by elevated firing rates (Fig 4A and 4B), but model neurons which comprised the fan-in triangle motif actually contracted towards low firing rates (Fig 4C and 4D). Fan-in triangles were more abundant in propagating activity than would be expected from their frequency in the synaptic network or component firing rates alone.
Like undirected clustering, the emergence of fan-in clustering in maps of propagating activity was robust to choice of T. Fan-in clustering was highly elevated in recruitment maps for T = 10 ms (undirected: 0.0068±0.0007; fan-in 0.011±0.0017; fan-out: 0.0028±0.0001; middle-node: 0.0068±0.0007; cycle: 0.0052±0.0004; mean±std for 5 simulations) and T = 50 ms (undirected: 0.019±0.0015; fan-in 0.027±0.0027; fan-out: 0.0077±0.0003; middle-node: 0.019±0.0013; cycle: 0.015±0.0007; mean±std for 5 simulations). Because of the different levels of sparseness in the numbers of connections these values should not be compared across values of T. Instead these analyses demonstrate that the over-representation of fan-in triangles is robust across a number of timescales.
To investigate the mechanism for overrepresentation of fan-in triangles in recruitment networks, we measured spike timing at their locations. The signature of fan-in triangle motifs is convergence from interconnected presynaptic neurons, a motif that could potentially facilitate cooperative summation of synaptic inputs. Consistent with this postulate, presynaptic neurons in fan-in triangle motifs were marked by increased probability of firing in the 10 ms prior to postsynaptic spiking (Fig 5A and 5B).
We next compared differences in presynaptic timing relationships at loci of fan-in triangle motifs compared to loci of simple convergence, to assess the role of presynaptic interconnectivity. For this analysis, random samples were obtained from epochs of coincident firing: 50 ms windows where every neuron in a triplet was active, centered on a spike in the postsynaptic reference neuron. To avoid confounds from juxtaposing multiple motifs, neuron triplets with any additional connections, including recurrent loops, were excluded for this specific analysis alone. As a result only fan-in triangles with exactly three interconnections were analyzed in this case. We found fan-in presynaptic neurons were stereotypically ordered in a manner consistent with the direction of their interconnection, resulting in an asymmetric distribution of intervals between their firing (Fig 5C). In addition to the temporal structure imposed by this asymmetry, mean absolute timing difference between presynaptic neurons in clustered fan-in motifs was modestly but significantly more temporally precise than were neurons in simple convergence motifs (13.5±10.2 ms compared to 14.9±10.7 ms; Wilcoxon rank-sum on mean-absolute timing difference, p = 0.0035, n = 1000 samples).
Moreover, we found that coincidence in fan-in triangle motifs occurred nearly twice as frequently as in motifs of simple convergence (1.9 ± 0.17 times more frequent, mean ± std; Wilcoxon rank-sum, p = 0.0079, n = 5 model datasets). Accounting for expected frequency of the two connection patterns in the underlying synaptic network, coincident activity is far more common at sites of fan-in triangles than at sites of simple convergence (linear regression: slope 3.0, y-intercept 0.00075, n = 5 simulations, r2 = 0.91, p = 0.011) (Fig 5D).
We postulated that clustering is efficacious for synaptic integration and examined whether the prevalence of clustering was predictive of postsynaptic membrane potentials. Pooling over all neurons and time bins, we binned the distribution of membrane voltages into segments that contained equal numbers of samples (Fig 6A). On average, because the model was active in the analyzed simulations, membrane voltages were depolarized from the resting equilibrium potential of -65 mV (median: -60.2 mV; lower quartile: -63.6 mV; upper quartile: -56.8 mV). To test our hypothesis, we generated functional networks that related recent presynaptic activity (within a 25 ms interval) to postsynaptic voltage (Fig 6B; see Methods), yielding one network for each division of the voltage distribution (Fig 6C). These networks can be viewed as reverse correlograms conditioned on postsynaptic voltage, and differed in the statistics of their topologies across different voltage regimes. At more negative membrane potentials, the active neurons which connected to the postsynaptic reference neuron (and accounted for its recent excitatory synaptic drive) were only modestly more clustered than random sparseness-matched controls. As the postsynaptic neuron depolarized, the presynaptic nodes driving that depolarization became increasingly clustered, peaking at the threshold for firing (Fig 6D). Characteristic paths were similar to random graphs at all subthreshold voltages. As a result of elevated clustering during membrane depolarization, small world ratios peaked at the most depolarized voltages corresponding to threshold for action potential generation. These data support the hypothesis that activity among clustered presynaptic neurons is particularly effective for recruiting the postsynaptic neuron to spike.
The statistical incongruence of function and synaptic connectivity indicates that spiking activity does not flow in an egalitarian fashion through the synaptic network. Instead, patterns of local clustering influence and direct where propagating activity occurs most frequently. That is, patterns of activity are shaped by higher-order patterns in synaptic connectivity and not just pairwise couplings. To further explore the dependence of activity flow on higher order synaptic connections we evaluated postsynaptic recruitment in a network model with a modest increase in mean synaptic strength. Synaptic connections were twice as strong on average compared to the network models used throughout the remainder of this study but remained too weak to drive spiking alone (Fig 7A). The two network designs did not differ in connection density. After synaptic weights were doubled, functional networks became more similar in topology to synaptic networks (small world ratio decreased; Wilcoxon rank-sum, p = 0.0079, n = 5) (Fig 7B). The double-strength models were less clustered (Fig 7C) (Wilcoxon rank-sum, p = 0.0079, n = 5), and exhibited longer average path lengths (Wilcoxon rank-sum, p = 0.0079, n = 5). Directed clustering was compared across the two families of models. Recruitment networks were analyzed with binary edges to control for their distinct mean synaptic weights. In addition to their decreased overall clustering, the fan-in triangle motif was significantly rarer in double-strength recruitment networks (Fig 7D) (from 0.030±0.0051 to 0.022±0.0025, p = 0.030, n = 6), while the fan-out triangle motif showed a small but significant increase in abundance (from 0.0040±2.0x10-4 to 0.0046±3.2x10-4, p = 0.0043, n = 6). Stronger presynaptic inputs reduced the need for extensive postsynaptic integration, allowing individual presynaptic cells to have a more independent impact on their postsynaptic partners. As a result, statistics of propagating activity were more faithful to underlying pairwise connections in the models with increased synaptic strength.
In model simulations, fan-in triangle motifs were abundant in maps of function and recruitment. We next evaluated whether the preponderance of fan-in triangle motifs would be robust to additional complexity in single-neurons and their connections. Unlike the simple model neurons that we used for simulation, real neurons are complex elements[16] and the connections between them are structured[12,39]. If clustered fan-in triangle motifs are a general feature of high-conductance nodes in a complex system, where coordinated inputs drive integration, the fan-in triangle will be overabundant in experimental dynamics. This postulate would be falsified if all directed clustering motifs were equally common in functional networks. To investigate, we analyzed high speed imaging data (20 Hz) of spontaneous circuit activity collected ex vivo in mouse somatosensory cortex (Fig 8A) (following [40]). We generated functional networks from the imaged experimental data using an iterative Bayesian approach which is robust to relatively small numbers of observations [33]. We then measured the prevalence of fan-in motifs in the functional topology (Fig 8B). Importantly, iterative Bayesian inference was not biased toward detection of fan-in triangle motifs, as demonstrated with rate-matched Poisson spiking (see Fig 3D).
Though imperfect indicators, functional weights probabilistically identify the likelihood of true monosynaptic excitatory connectivity[35]. As a result, expected error rate for inferred connections can be adjusted with a sliding threshold on functional weight. Stricter thresholds yield a more accurate approximation of the underlying recruitment network at the cost of restricted sampling. Using inferred recruitment networks, beginning at the top quartile of inferred weights, directed clustering was computed in five-percentile increments. Confidence intervals were obtained using bootstrap resampling under the assumption of a 30% false-positive rate. As confidence of synaptic connectivity increased, the fan-in triangle motif became increasingly abundant and fan-out triangles less so (Fig 8C). Differences between the two motifs were significant (threshold at 95th percentile, p = 4.8x10-34, n = 100 bootstrap resampled functional networks, Wilcoxon ranksum).
We next measured whether strong functionally coupled neurons were more spatially proximal than random pairs. We defined strong functional connections as those exceeding a 95% threshold on non-zero weights since previous work has indicated that these particular functional connections are more likely to reflect a causal synaptic connection[35]. We found that the median pairwise distance separating strong functionally connected cells was 249 μm, whereas randomly chosen pairs of neurons were separated by a median 263 μm (Wilcoxon-ranksum p = 0.0336, nfunctional = 638, nrandom = 10000). We then measured triplets of neurons with functional connections that form triangles to determine whether these neurons were more spatially proximal to one another than randomly chosen triplets of neurons. To investigate, proximity was quantified as the perimeter around the triangle formed by vertices at the spatial location of each neuron. Neurons in functional triangles with mutual connectivity and at least three functional connections were inscribed by perimeters of median length 807 μm, compared to median perimeter of 823 μm for randomly selected triplets that were unconstrained by direction and number of edges (Wilcoxon rank-sum p = 0.0097, ntriangles = 2556, nrandom = 10,000). Interestingly, triplets of neurons connected into arrangements of either simple divergence or simple convergence (i.e. neurons in wedges, lacking interconnectedness between the common neighbors), were even more distant, inscribed by a perimeter of median 839 μm (Wilcoxon rank-sum, ntriangles = 2556, nwedges = 14,882). Thus, clustered triplets (triangles) tended to be arranged significantly more locally than simple convergent or simple divergent triplets (wedges).
We then compared measures of clustering between the model, which was comprised of random connections, and the experimental data which almost certainly contained structured connectivity [12, 39] to evaluate how the measure of fan in and fan-out triangles depend on the underlying structural topology. To do so we used a measure of clustering propensity[41] which allowed us to make comparisons of networks which have very different connection densities. Clustering propensity (1-ΔCfan-in and 1-ΔCfan-out) results in a normalized value where 1 is extreme clustering as seen in lattices, and 0 indicates no clustering above that expected in Erdős-Rényi random networks. For the model, fan-in clustering was scored at 0.18 ± 0.019; and for the experimental data, fan-in clustering was scored at 0.20 ± 2.0x10-4 (Wilcoxon ranksum p = 1.74x10-4, nmodel = 5 simulations; ndata = 100 bootstrap samples). Thus, fan-in clustering was modestly but significantly more abundant in maps of propagating activity based on experimental recordings. We note that we compared thresholded graphs at the 80%-level (i.e. top 20% of non-zero edges) for this measure because the experimentally derived functional networks were not well-matched by regular lattices below this density.
Finally, we measured timing relationships among imaged active neurons. Reliable timing relationships were measured independent of other functional analyses, using cross-correlations on the normalized fluorescence traces (Methods). Presynaptic coactivity was assessed as the product of the two z-scored presynaptic traces and compared to postsynaptic fluorescence as a straightforward cross correlation. The resulting average cross-correlogram for fan-in triangles was stronger and more asymmetric than those measured from simple-convergence motifs (Fig 8D).
Thus, presynaptic activity in fan-in triangles was more predictive of postsynaptic firing than presynaptic activity in motifs of simple convergence. These results are consistent with fan-in triangles supporting coincident input and favoring reliable propagation of activity. Results from the model indicated that the fan-in triangle motif temporally coordinates presynaptic inputs, rendering them more capable of driving recipient neurons to threshold. Supporting our prediction of its fundamental importance for reliable recruitment, in acutely dissected neocortical tissue with more complex patterns of connectivity and intrinsic neuronal properties, we find a robust elevation of the same directed motif.
Using a model composed of random connections among leaky integrate-and-fire neurons with conductance-based synapses, we found that maps of propagating activity were structured and non-random. Small-world patterning in the dynamics emerged because a specific higher-order connection pattern was particularly effective for postsynaptic integration: convergence of synaptic input from connected neighbors. Synaptic connections between neighbors favored coincident timing of inputs onto their targets. This coincident activation led to efficient postsynaptic integration. As a consequence, clustering among active presynaptic cells tracked depolarization of model postsynaptic neurons. Thus, activity was preferentially routed through fan-in triangle motifs.
In experimental recordings of emergent activity in hundreds of neurons ex vivo, after mapping inferred recruitment patterns [33], we found that fan-in triangles were even more dramatically overrepresented than in the model. These results are contextualized by increasing recognition of non-random functional structure in networks of neurons: Rich club structure has been reported ex vivo and in vivo[31]. Clustered[30], small world functional networks[28], and nucleation of dynamics[29] have also been observed in neuronal cultures. Since cultured populations differ from neocortex in the details of their topological makeup, these findings across model systems further suggest that clustering in general and the fan-in triangle motif in particular may be a canonical feature of propagating activity among interconnected neurons. Despite differences in details of connectivity and neuronal intrinsic properties, dynamics are constrained by the requirement for coincident summation of individually weak inputs. Constraining dynamics with beyond-pairwise relationships can be helpful for cortical computation. Theoretical work has shown that non-uniform features of connection topology impact information transfer[42], and higher-order correlations were particularly impactful in low spike-rate regimes[43]. These complementary results from complex networks, statistical physics and network biology suggest that, by shaping feasible dynamics, the fan-in triangle motif could enhance information transfer from inputs to outputs.
We hypothesize that local circuits are organized around fan-in triangle motifs, promoting cooperative patterns of firing and stabilizing[44] the propagation of activity despite individually unreliable neurons. This canonical mechanism provides the coordination necessary to propagate signal despite weak synaptic connections. Indeed, reliable sequential firing was associated with number of fan-in triangles even after controlling for overall in-degree. Although clustering among fan-in triangles has not been tested directly until now, paired patch clamp recordings have shown that local neocortical circuitry is characterized structurally by abundant triplet motifs[12,39]. Our data and modeling suggest a functional consequence for a subset of these synaptic motifs: connected presynaptic neurons help establish coordinated timing among convergent inputs, leading to cooperative summation at the postsynaptic membrane. Such cooperativity has been shown to be one potential mechanism capable of generating spike trains that are consistent with experimental observations in vivo[45].
While there are certainly explicit developmental rules that govern neuron to neuron connectivity, our results suggest that higher-order connectivity need not require specification a priori. It could emerge autonomously if fan-in triangle motifs within a random network were stabilized and magnified during network development, e.g. by pruning non-recruiting connections through activity-dependent plasticity. Thus, higher-order synaptic motifs that are particularly effective for postsynaptic recruitment could potentially self-organize[46].
These results do not indicate a complete schism between synaptic connectivity and dynamics—one clearly depends on the other. However, their relationship is complicated by the integrative properties of single neurons. Synaptic integration constrains feasible dynamics, and distributed synaptic motifs route the propagation of activity. These interactions are a source of higher-order dynamical structure. The routing of information is coordinated by higher-order synaptic patterns and the context of ongoing activity because the routing of spikes is determined by relative timing and collective interactions.
Simulations were implemented using the Brian Brain Simulator[47]. Model populations consisted of 1000 excitatory neurons, 200 inhibitory neurons and 50 Poisson input units. Connection probabilities depended on source and target identity. For example, inhibitory-excitatory connections occurred with probability 0.25 (Pee = 0.2, Pei = 0.35, Pie = 0.25, Pii = 0.3).
Conductance based synaptic weights were drawn from a heavy-tailed distribution and assigned randomly[48,49]. Weights were drawn randomly from a lognormal distribution with mu = -0.64 and sigma = 0.51. These parameters are the mean and standard deviation of the corresponding normal curve. The resulting lognormal ensemble has expected mean of 0.60 and variance of 0.11, in multiples of the leak conductance. Connections from inhibitory to excitatory cells were scaled by a further 50% to simulate efficacious somatic contacts. A small tonic excitatory drive gt was supplied to all units to help stabilize sparse spiking. Synaptic bombardments induced exponentially shaped membrane conductances with leaky-integrate-and-fire summation. Conductance-based synapses are important for recapitulating synaptic integration in the high-conductance state[21,50]. We used sparse and randomly connected networks in which we did not impose any synaptic organization beyond cell-type dependent connection probabilities.
Trials began with 50 ms of activity in the input pool at 15 Hz, exciting the network via random input projections. After input units were silenced, the recording period began, and activity flowed through the network for 100 ms. Input units projecting to excitatory cells randomly and independently with probability 0.1. Every 100 trials (an epoch), new random projections were drawn from the input pool to the excitatory population, simulating a diversity of activity. Participation during a single input epoch totaled 64±0.98% of neurons (mean ± std), growing to encompass 85.5% of neurons when all sets of input projections were considered (i.e. over all epochs).
Excitatory reversal potential Ee was 0 mV, as was Et. Inhibitory reversal potential Ei was -90 mV. Reversal potential for leak current Eleak was -65 mV. Firing threshold was -48 mV, and post-spike reset was -70 mV. In addition to after spike hyperpolarization induced by the reset potential, a 1 ms absolute refractory period was imposed on model neurons. Leak conductance gleak was fixed at 0.20 mS. Tonic depolarizing conductance gt was equal in magnitude to the leak conductance. Membrane time constant τm was 20 ms; excitatory synaptic time constant τe was 10 ms; and inhibitory synaptic time constant τi was 5 ms. Additional description can be found in[35].
Spiking dynamics were compared to in vivo activity according to the following criteria: asynchrony[51] was measured with spike-rate correlations, by convolving spike times with a Guassian kernel of width σ = 3 ms. Among excitatory neurons in the recording period, mean correlation coefficient was 0.0019[50]. This asynchrony emerged in the presence of heterogeneous connection strengths, raising the possibility of combining stable propagation with rich internal dynamics[49,52]. Irregularity was measured with interspike-intervals, which were observed to have mean squared-coefficient of variation of 0.81, consistent with other reports of irregular activity[53]. To measure inter-spike intervals, model activity was stimulated with Poisson firing for 50 ms, then allowed to evolve for 950 ms in isolation. This procedure was repeated 100 times. Excitatory spiking activity was characterized by a median branching coefficient of 1.00 (for 10 ms bins), indicating near-critical dynamics[54–57]. Firing rates in the excitatory population during the recording period were 1.33 ± 3.15 Hz (mean ± std) consistent with findings in awake behaving mice[58]. Collective spiking generated spike-driven conductances that dwarfed the leak conductance, in keeping with definitions of high-conductance state[21].
Call the directed network of synaptic connections among excitatory neurons Esyn and the population of excitatory cells Ve. Construct the directed graph of synaptic connections:
Gstructural≡(Ve,Esyn)
To map functional relationships using lagged firing, define recent activity for neuron i at time t as firing at least once in the 25 ms preceding t.
More formally, we can define random variable Si representing the activity of neuron i such that
sit≡{2ifneuroniis firing at timet1ifneuroniis not firing but was active within the last25ms0otherwise}
In that case,
Eijlag≡P(sj=2|si>0)
The recruitment network encompassed synaptically connected neurons manifesting lagged patterns:
Eijomniscient≡{EijlagifEijsyn>00otherwise}
Grecruitment≡(Ve,Eomniscient)
Iterative Bayesian networks were measured with a heuristic optimization procedure, described further below and in [35], following [28].
Since shortest path measurements assume a cost matrix, edge weights were first inverted so strong connections were cheap and zero-weighted connections were infinitely costly. Shortest paths between all pairs were computed using Dijkstra’s algorithm. Mean path length was compared to sparseness-matched Erdős-Rényi graphs analyzed in the same way. Local clustering coefficients were computed using the neighbors of neighbors formulation[32] and aggregated as the mean over all neurons. Sparseness-matched Erdős-Rényi graphs were analyzed in the same fashion. Clustering score was the ratio of the actual mean to sparseness-matched null mean. Small-world topologies can be quantified as a ratio of ratios, clustering elevation divided by mean path length reduction[33].
Clustering was also investigated using a related definition, the number of connected undirected triangles as a fraction of all possible undirected triangles (transitivity formulation). Directed clustering was computed in the same way, using directed triangles instead of undirected[38].
To compare clustering between data and model networks, across connection densities that were very different, we followed the small-world propensity approach[41]. In that work, clustering levels ΔC are normalized as the fractional distance between density-matched lattice and random graphs. We termed this measure clustering propensity, expressing it as 1 – ΔC so that 1 signified extreme clustering and 0 signified no clustering beyond that expected at random. We made a straightforward extension to this approach to account for directed clustering, simply substituting directed triangle counts for undirected triangle counts, with appropriate normalizations[38]. Quantifications based on clustering propensity recapitulated our findings quantifying clustering as fractional abundance over random expectation.
For the set of voltage bins with lower bounds a and upper bounds b, construct one network for each bin k, where edge (i, j)k is quantifying the probability model neuron j will have postsynaptic potential Mj between ak and bk conditioned on presynaptic model neuron i being recently active. Recently active was defined as firing within 25 ms relative to postsynaptic voltage measurement. A final condition was imposed: that connected pairs also share a synaptic connection, a convenience of measurement unique to simulated networks.
Functional topologies were measured for simulations having typical synaptic weight distributions (n = 5) and for simulations where random draws from the synaptic weight distributions were scaled to double strength (n = 6). Ratios for global clustering, characteristic path, and smallworldness were quantified following[33], as above, on the two sets of weighted, symmetrized topologies. Directed clustering was measured following[38]. The directed clustering measurements were conducted on binary topologies to control for potential differences stemming from their different underlying mean synaptic weights.
All procedures were performed in accordance with and approved by the Institutional Animal Care and Use Committee at the University of Chicago. One juvenile mouse (postnatal day 14, of strain C57BL/6) was anesthetized by intraperitoneal injection of ketamine-xylazine and rapidly decapitated. The brain was dissected and placed in oxygenated, ice-cold artificial cerebrospinal fluid (Cut-ACSF; contents contain the following in mM: 3 KCl, 26 NaHCO3, 1 NaH2PO4, 0.5 CaCl2, 3.5 MgSO4 25 dextrose, and 123 sucrose). The brain was then sliced coronally using a vibratome (VT1000S; Leica) into 450 μm thick slices. These slices encompassed the mouse whisker somatosensory cortex. Slices were then transferred into 35°C oxygenated incubation fluid (Incu-ACSF; contents contain the following, in mM: 123 NaCl, 3 KCl, 26 NaHCO3, 1 NaH2PO4, 2 CaCl2, 6 MgSO4, 25 dextrose) for 30 min. Bulk loading of Ca2+ dye was then performed, via transfer of slices into a Petri dish containing ∼2 ml of Incu-ACSF and an aliquot of 50 μg Fura-2AM (Product code, Invitrogen, location) dissolved in 13 μl DMSO and 2 μl of Pluronic F-127 (Code, Invitrogen, location) (following [9]).
Throughout the duration of imaging, slices were continuously perfused with a standard ACSF solution (contents contain the following, in mM: 123 NaCl, 3 KCl, 26 NaHCO3, 1 NaH2PO4, 2 CaCl2, 2 MgSO4, and 25 dextrose, which was continuously aerated with 95% O2, 5% CO2). Visualization of Fura-2AM loaded neurons was performed via serial 5 min recordings, collected using the HOPS scanning technique (a suite of software and custom microscopy setup developed in-house, see [40]). This method allowed us to monitor action potential generation within individual neurons, by detecting contours of loaded cells from a raster image, then computing an efficient traveling salesman tour over those cell bodies. Our dwell time parameter was fixed at a value of 16 samples/cell/frame. Population framerate was 20 Hz, resulting in ~450 neurons sampled once every ~50 ms. Changes in emitted fluorescence were analyzed with a threshold-crossing approach. First, a signal-to-noise cutoff was implemented by measuring the ratio of the 99th percentile divided by the mean for the fluorescence trace of each cell. Cells exceeding 1.55 by this metric were retained for further analysis. Of the 444 sampled neurons, 189 exceeded our strict criterion on signal-to-noise (see Methods). Among these cells with clean fluorescent signals, instances of elevated firing were identified from excursions in the signal exceeding two-sigma, with inflection points more precisely identified by following these excursions backwards to the bin of their most recent median-crossing. The resulting binary vector identified high-probability periods of spiking activity across the imaged population[14,59].
Recurring timing relationships can be used to identify likely synaptic connections between individual pairs, particularly lagged firing near the timescale of synaptic integration. We used an iterative Bayesian inference algorithm to parse these lagged firing patterns[28,35]. The inference algorithm was initialized five times, and final weights were pooled as an average. The combined network was thresholded to isolate its strongest relationships. With increasing threshold, functional relationships became more precise in indicating true monosynaptic connectivity, and also more confidently overabundant in the fan-in triangle motif.
To understand the impact of mistaken inferences from a different perspective, independent of relationships between functional weight and true connectivity, bootstrap resampling was used to estimate how errors in inferred connectivity affected estimates of directed clustering measures. For an error rate of 30% estimated from simulated experimental constraints[35], differences in directed clustering were significant even after redacting possible false positives (100 bootstrap-resampled topologies; Fig 8C and 8E).
In a typical simulated network, the density of the recruitment network was 0.049, meaning only about one quarter of synaptic connections were a site of propagating activity. Since only those pairs are visible in patterns of lagged firing, the density of recruiting connections was shown for an additional definition of optimal performance (one potentially more appropriate for models with sparse firing).
Average cross-correlations were computed over a two-second sliding window using z-scored fluorescence traces. The first signal was computed as the product of two putative presynaptic fluorescence traces, as a simple score of their activity and/or coactivity. The second signal was the postsynaptic fluorescence trace. Their raw cross-correlation measures the timing offsets between putative presynaptic activity and postsynaptic firing. The functional relationships used to define fan-in triangle motifs versus simple convergence motifs inferred using iterative Bayesian inference, on the basis of single-frame lagged activity, measured in 50 ms bins.
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.